{
  "version": "https://jsonfeed.org/version/1.1",
  "title": "OpenClaw Chronicles",
  "home_page_url": "https://openclawchronicles.com",
  "feed_url": "https://openclawchronicles.com/feed.json",
  "description": "OpenClaw Chronicles covers OpenClaw releases, security alerts, migration guides, tutorials, and ecosystem news.",
  "icon": "https://openclawchronicles.com/icon-512.png",
  "favicon": "https://openclawchronicles.com/favicon.png",
  "authors": [
    {
      "name": "Cody",
      "url": "https://openclawchronicles.com/about/",
      "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
    }
  ],
  "language": "en-US",
  "banner_image": "https://openclawchronicles.com/assets/images/about-banner.jpg",
  "hubs": [
    {
      "type": "alternate",
      "url": "https://openclawchronicles.com/feed.xml",
      "mime_type": "application/rss+xml"
    }
  ],
  "items": [
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-14-guest-managed-worktrees/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-14-guest-managed-worktrees/",
      "title": "OpenClaw Guests Can Start Managed Worktrees",
      "summary": "OpenClaw guests can now select registered projects, choose a managed worktree, and start isolated sessions without path exposure or broader operator scopes.",
      "content_text": "OpenClaw merged a New Session UI fix that restores a useful collaboration path for guests. [PR #171811](https://github.com/openclaw/openclaw/pull/171811) lets guests who can create sessions select a registered or Gateway-cloned project, choose **New worktree**, and start an isolated session without broader operator scopes.\n\nThe bug was subtle: the project list correctly omitted host paths for guests, but the repository controller treated those path-redacted rows as idle. That hid the checkout control, even though creation by project ID already supported a managed worktree.\n\n## What Guests Can Do Now\n\nWith the fix, a guest can:\n\n- Select a registered project in New Session\n- Choose a managed worktree\n- Leave the base reference blank and let the Gateway choose the repository default\n- Start an isolated session using only session-scoped access\n\nThe request carries the existing project ID and worktree intent to `sessions.create`. It does not need to disclose a host path, and it does not add Gateway fields, scope grants, configuration, or path visibility.\n\nIf creation is refused, the UI preserves the selected project and isolation intent so the user can correct the problem or retry.\n\n## Why It Matters\n\nManaged worktrees are one of OpenClaw's stronger collaboration primitives. They let work happen in an isolated checkout rather than directly in the main project directory. For guests, that isolation is especially important: it allows contribution-style sessions without granting broad operator authority or exposing filesystem details.\n\nThe previous UI behavior created an odd mismatch. The backend could accept the safe project-ID-based creation path, but the UI did not expose the control when the project row had no host path. The fix aligns the interface with the existing Gateway behavior.\n\n## Path Privacy Stays Intact\n\nThe PR is careful about the boundary. The UI now distinguishes a path-redacted registered project awaiting creation-time validation from an inspected Git checkout. Clients with disclosed checkout paths keep the existing inspection flow. Guests without path access get the worktree option through project identity, not through filesystem disclosure.\n\nThat is the right shape for shared systems: give the user the capability they are authorized to use, but avoid revealing local implementation details they do not need.\n\n## Verification\n\nThe PR includes browser proof from an authenticated guest flow and regression coverage for project selection, worktree submission without a host path or forced reference, and preservation of isolation after an actionable creation refusal.\n\nIt also strengthens the sandbox-project integration to use real guest Gateway request admission and verify managed ownership while retaining the direct-access rejection. The focused candidate passed formatting, lint, affected type checks, and 138 tests across nine files. The two new UI cases fail on the baseline because the checkout control is hidden, which is exactly the bug being fixed.\n\nThe result is a cleaner guest workflow: registered project, new worktree, nonempty first message, running isolated session. No path leak required.",
      "content_html": "<p>OpenClaw merged a New Session UI fix that restores a useful collaboration path for guests. <a href=\"https://github.com/openclaw/openclaw/pull/171811\">PR #171811</a> lets guests who can create sessions select a registered or Gateway-cloned project, choose <strong>New worktree</strong>, and start an isolated session without broader operator scopes.</p><p>The bug was subtle: the project list correctly omitted host paths for guests, but the repository controller treated those path-redacted rows as idle. That hid the checkout control, even though creation by project ID already supported a managed worktree.</p><h2>What Guests Can Do Now</h2><p>With the fix, a guest can:</p><ul><li>Select a registered project in New Session</li><li>Choose a managed worktree</li><li>Leave the base reference blank and let the Gateway choose the repository default</li><li>Start an isolated session using only session-scoped access</li></ul><p>The request carries the existing project ID and worktree intent to <code>sessions.create</code>. It does not need to disclose a host path, and it does not add Gateway fields, scope grants, configuration, or path visibility.</p><p>If creation is refused, the UI preserves the selected project and isolation intent so the user can correct the problem or retry.</p><h2>Why It Matters</h2><p>Managed worktrees are one of OpenClaw's stronger collaboration primitives. They let work happen in an isolated checkout rather than directly in the main project directory. For guests, that isolation is especially important: it allows contribution-style sessions without granting broad operator authority or exposing filesystem details.</p><p>The previous UI behavior created an odd mismatch. The backend could accept the safe project-ID-based creation path, but the UI did not expose the control when the project row had no host path. The fix aligns the interface with the existing Gateway behavior.</p><h2>Path Privacy Stays Intact</h2><p>The PR is careful about the boundary. The UI now distinguishes a path-redacted registered project awaiting creation-time validation from an inspected Git checkout. Clients with disclosed checkout paths keep the existing inspection flow. Guests without path access get the worktree option through project identity, not through filesystem disclosure.</p><p>That is the right shape for shared systems: give the user the capability they are authorized to use, but avoid revealing local implementation details they do not need.</p><h2>Verification</h2><p>The PR includes browser proof from an authenticated guest flow and regression coverage for project selection, worktree submission without a host path or forced reference, and preservation of isolation after an actionable creation refusal.</p><p>It also strengthens the sandbox-project integration to use real guest Gateway request admission and verify managed ownership while retaining the direct-access rejection. The focused candidate passed formatting, lint, affected type checks, and 138 tests across nine files. The two new UI cases fail on the baseline because the checkout control is hidden, which is exactly the bug being fixed.</p><p>The result is a cleaner guest workflow: registered project, new worktree, nonempty first message, running isolated session. No path leak required.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-guest-managed-worktrees.png\" alt=\"OpenClaw Guests Can Start Managed Worktrees\" /></p>",
      "date_published": "2026-10-14T08:15:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Guests",
        "Start",
        "Managed",
        "Worktrees"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-guest-managed-worktrees.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-guest-managed-worktrees.png",
          "mime_type": "image/png",
          "title": "OpenClaw Guests Can Start Managed Worktrees cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-14-team-reports-pause-collection/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-14-team-reports-pause-collection/",
      "title": "OpenClaw Team Reports Can Pause Auto Collection",
      "summary": "OpenClaw Team Reports gained a schedule switch that pauses automatic collection while keeping stored reports, session views, and manual generation available.",
      "content_text": "OpenClaw's Team Reports plugin gained a useful operator control this morning: automatic collection can now be paused without disabling the rest of the reporting surface. [PR #171840](https://github.com/openclaw/openclaw/pull/171840) adds a default-on schedule setting for teams that want stored reports and manual generation to remain available while automatic collection is stopped.\n\nThe new setting is:\n\n```yaml\nplugins.entries.team-reports.config.schedule.enabled: false\n```\n\nWhen set to `false`, OpenClaw stops timer-based collection and catch-up admission. Stored reports, work-session views, and explicit manual day, week, or month generation continue to work.\n\n## Why This Control Was Needed\n\nBefore this patch, operators could disable intraday runs, but daily collection and catch-up behavior could still continue. That made it awkward to pause automatic Team Reports cleanly.\n\nThere are plenty of legitimate reasons to pause collection:\n\n- A team is auditing reporting scope or credentials\n- A workspace wants to reduce scheduled background work temporarily\n- A plugin owner wants to preserve reports while changing retention or configuration\n- An installation needs manual-only reporting during a migration window\n\nThe key distinction is that pausing collection should not make historical reports disappear, and it should not block an explicit manual report.\n\n## What OpenClaw Now Reports\n\nThe scheduler now owns the automatic-work decision directly. If scheduling is disabled, it skips automatic timer and catch-up admission, while manual runs continue to be accepted.\n\nStatus output also reports `scheduleEnabled`, and a disabled schedule has no `nextDue` entries. That gives operators a visible confirmation that automatic collection is off, rather than requiring them to infer it from the absence of future work.\n\nThe plugin manifest, configuration parser, README, and configuration guide were updated to describe the same control. The canonical plugin-config budget increases by one for the new default-true automatic-scheduling setting.\n\n## No Data Migration\n\nThis is source-only work. The PR says it changes no running Gateway, service configuration, stored reports, or external API data. Existing credentials and retention behavior are unchanged, and scheduling remains enabled by default.\n\nThat default is important. Most teams using Team Reports probably expect automatic collection to continue. The new control is for operators who explicitly want a pause state.\n\n## Verification\n\nThe regression coverage uses the real scheduler and SQLite store. It checks disabled scheduling across daily, weekly, and monthly rollover and restart paths; verifies that automatic collection and pruning do not run; confirms manual report generation still works; and preserves stored JSON and Markdown reads through registered Gateway methods.\n\nCorrected-head CI passed with config and docs checks, changed-extension tests, production and test types, and lint. The earlier run caught a missing generated baseline, which was fixed before the final pass.\n\nThe practical headline is simple: Team Reports now has a proper pause switch. Operators can stop background collection without losing the reporting surface they still need.",
      "content_html": "<p>OpenClaw's Team Reports plugin gained a useful operator control this morning: automatic collection can now be paused without disabling the rest of the reporting surface. <a href=\"https://github.com/openclaw/openclaw/pull/171840\">PR #171840</a> adds a default-on schedule setting for teams that want stored reports and manual generation to remain available while automatic collection is stopped.</p><p>The new setting is:</p><p>``<code>yaml<br />plugins.entries.team-reports.config.schedule.enabled: false<br /></code>`<code></p><p>When set to </code>false<code>, OpenClaw stops timer-based collection and catch-up admission. Stored reports, work-session views, and explicit manual day, week, or month generation continue to work.</p><h2>Why This Control Was Needed</h2><p>Before this patch, operators could disable intraday runs, but daily collection and catch-up behavior could still continue. That made it awkward to pause automatic Team Reports cleanly.</p><p>There are plenty of legitimate reasons to pause collection:</p><ul><li>A team is auditing reporting scope or credentials</li><li>A workspace wants to reduce scheduled background work temporarily</li><li>A plugin owner wants to preserve reports while changing retention or configuration</li><li>An installation needs manual-only reporting during a migration window</li></ul><p>The key distinction is that pausing collection should not make historical reports disappear, and it should not block an explicit manual report.</p><h2>What OpenClaw Now Reports</h2><p>The scheduler now owns the automatic-work decision directly. If scheduling is disabled, it skips automatic timer and catch-up admission, while manual runs continue to be accepted.</p><p>Status output also reports </code>scheduleEnabled<code>, and a disabled schedule has no </code>nextDue` entries. That gives operators a visible confirmation that automatic collection is off, rather than requiring them to infer it from the absence of future work.</p><p>The plugin manifest, configuration parser, README, and configuration guide were updated to describe the same control. The canonical plugin-config budget increases by one for the new default-true automatic-scheduling setting.</p><h2>No Data Migration</h2><p>This is source-only work. The PR says it changes no running Gateway, service configuration, stored reports, or external API data. Existing credentials and retention behavior are unchanged, and scheduling remains enabled by default.</p><p>That default is important. Most teams using Team Reports probably expect automatic collection to continue. The new control is for operators who explicitly want a pause state.</p><h2>Verification</h2><p>The regression coverage uses the real scheduler and SQLite store. It checks disabled scheduling across daily, weekly, and monthly rollover and restart paths; verifies that automatic collection and pruning do not run; confirms manual report generation still works; and preserves stored JSON and Markdown reads through registered Gateway methods.</p><p>Corrected-head CI passed with config and docs checks, changed-extension tests, production and test types, and lint. The earlier run caught a missing generated baseline, which was fixed before the final pass.</p><p>The practical headline is simple: Team Reports now has a proper pause switch. Operators can stop background collection without losing the reporting surface they still need.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-team-reports-pause-collection.png\" alt=\"OpenClaw Team Reports Can Pause Auto Collection\" /></p>",
      "date_published": "2026-10-14T08:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Team",
        "Reports",
        "Pause",
        "Auto",
        "Collection"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-team-reports-pause-collection.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-team-reports-pause-collection.png",
          "mime_type": "image/png",
          "title": "OpenClaw Team Reports Can Pause Auto Collection cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-14-usage-reports-partial-results/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-14-usage-reports-partial-results/",
      "title": "OpenClaw Usage Reports Now Return Partial Results",
      "summary": "OpenClaw usage reports now return safe partial totals and narrowing hints instead of errors when broad session usage requests exceed the fixed read budget.",
      "content_text": "OpenClaw merged a usage-reporting fix this morning that should make large team dashboards feel less brittle. [PR #171904](https://github.com/openclaw/openclaw/pull/171904) changes broad `sessions.usage` requests so they can return completed subtotals when the handler reaches its 15-second read budget, rather than failing the whole request.\n\nThat matters for teams with large session histories. Usage views are often most valuable when they answer broad questions: how much a group spent, which sessions consumed the most tokens, or whether a reporting period needs to be narrowed. Before this change, an over-budget request could produce an error instead of the partial information OpenClaw had already computed.\n\n## What Changed\n\nThe new behavior returns marked partial data with `partial.reason: \"time-budget\"`, a partial cache status, and the existing hint to narrow the request by date range or selected session. The overview can display that hint through its existing query warning path.\n\nThe important detail is that partial totals are explicitly incomplete. OpenClaw is not pretending a timed-out wide report is definitive. It is giving operators the safest available slice of the answer, while still pointing them toward a narrower query when precision matters.\n\nThe PR also keeps cancellation semantics intact. If a caller disconnects, OpenClaw still cancels instead of publishing a late response.\n\n## Why It Matters\n\nUsage reporting is one of those features that gets tested by real-world data volume. A small installation may never hit the timeout path. A busy team with thousands of sessions and retained runs can.\n\nThe PR's measurement fixture included:\n\n- 8,400 sessions\n- 10,080 retained runs\n- 168,000 assistant usage events\n- 20.16 million verified tokens\n- $201.60 in represented usage\n\nIn that fixture, eight concurrent cache-miss requests failed eight out of eight times before the change. After the patch, the same shape returned marked partial data eight out of eight times in roughly 15 seconds. Five of those responses returned full token totals; three retained the initial 256-session subtotal.\n\nThat is a practical reliability win: the operator gets a bounded answer and a next action instead of a blank wall.\n\n## The Implementation Shape\n\nThe fix separates the data-work budget from the caller and publication lifetime. Report preparation now reuses transcript statistics and rewrite-generation data already obtained during discovery, deleting a second statistics scan.\n\nOne initial 256-instance read preserves an early subtotal. Remaining instances use one read, and completed subtotals survive budget cancellation. Publication reads for missing authority facts are also bounded by the deadline; unverified rows stay excluded, and late reads release their resources.\n\nThe PR does not introduce a schema change or stored-format migration. It does add production code, because a safe partial response needs to preserve family subtotals, carry the memory allowance through the worker boundary, and avoid mixing incomplete data with verified totals.\n\n## Verification\n\nThe PR reports focused local coverage across the usage handler, worker I/O, and owner-attribution integration tests. New timeout assertions fail against the old handler and pass with the new partial-response path. Final broad CI passed on the merge head, including type checks, selected test shards, builds, UI gates, and baseline ratchets.\n\nThis is not claimed as a broad performance improvement. The author is careful about that: ordinary request timing stayed within observed drift. The win is narrower and more useful: when a usage report is too wide to complete inside the budget, OpenClaw now returns clear partial data instead of losing the report entirely.",
      "content_html": "<p>OpenClaw merged a usage-reporting fix this morning that should make large team dashboards feel less brittle. <a href=\"https://github.com/openclaw/openclaw/pull/171904\">PR #171904</a> changes broad <code>sessions.usage</code> requests so they can return completed subtotals when the handler reaches its 15-second read budget, rather than failing the whole request.</p><p>That matters for teams with large session histories. Usage views are often most valuable when they answer broad questions: how much a group spent, which sessions consumed the most tokens, or whether a reporting period needs to be narrowed. Before this change, an over-budget request could produce an error instead of the partial information OpenClaw had already computed.</p><h2>What Changed</h2><p>The new behavior returns marked partial data with <code>partial.reason: \"time-budget\"</code>, a partial cache status, and the existing hint to narrow the request by date range or selected session. The overview can display that hint through its existing query warning path.</p><p>The important detail is that partial totals are explicitly incomplete. OpenClaw is not pretending a timed-out wide report is definitive. It is giving operators the safest available slice of the answer, while still pointing them toward a narrower query when precision matters.</p><p>The PR also keeps cancellation semantics intact. If a caller disconnects, OpenClaw still cancels instead of publishing a late response.</p><h2>Why It Matters</h2><p>Usage reporting is one of those features that gets tested by real-world data volume. A small installation may never hit the timeout path. A busy team with thousands of sessions and retained runs can.</p><p>The PR's measurement fixture included:</p><ul><li>8,400 sessions</li><li>10,080 retained runs</li><li>168,000 assistant usage events</li><li>20.16 million verified tokens</li><li>$201.60 in represented usage</li></ul><p>In that fixture, eight concurrent cache-miss requests failed eight out of eight times before the change. After the patch, the same shape returned marked partial data eight out of eight times in roughly 15 seconds. Five of those responses returned full token totals; three retained the initial 256-session subtotal.</p><p>That is a practical reliability win: the operator gets a bounded answer and a next action instead of a blank wall.</p><h2>The Implementation Shape</h2><p>The fix separates the data-work budget from the caller and publication lifetime. Report preparation now reuses transcript statistics and rewrite-generation data already obtained during discovery, deleting a second statistics scan.</p><p>One initial 256-instance read preserves an early subtotal. Remaining instances use one read, and completed subtotals survive budget cancellation. Publication reads for missing authority facts are also bounded by the deadline; unverified rows stay excluded, and late reads release their resources.</p><p>The PR does not introduce a schema change or stored-format migration. It does add production code, because a safe partial response needs to preserve family subtotals, carry the memory allowance through the worker boundary, and avoid mixing incomplete data with verified totals.</p><h2>Verification</h2><p>The PR reports focused local coverage across the usage handler, worker I/O, and owner-attribution integration tests. New timeout assertions fail against the old handler and pass with the new partial-response path. Final broad CI passed on the merge head, including type checks, selected test shards, builds, UI gates, and baseline ratchets.</p><p>This is not claimed as a broad performance improvement. The author is careful about that: ordinary request timing stayed within observed drift. The win is narrower and more useful: when a usage report is too wide to complete inside the budget, OpenClaw now returns clear partial data instead of losing the report entirely.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-usage-reports-partial-results.png\" alt=\"OpenClaw Usage Reports Now Return Partial Results\" /></p>",
      "date_published": "2026-10-14T08:05:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Memory",
        "Usage",
        "Reports",
        "Return",
        "Partial",
        "Results"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-usage-reports-partial-results.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-14-usage-reports-partial-results.png",
          "mime_type": "image/png",
          "title": "OpenClaw Usage Reports Now Return Partial Results cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-13-telegram-caption-formatting/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-13-telegram-caption-formatting/",
      "title": "OpenClaw Fixes Telegram Rich Captions",
      "summary": "OpenClaw PR #171180 converts rich Telegram media-caption blocks into native caption formatting instead of leaking literal tags into chats.",
      "content_text": "OpenClaw merged [PR #171180](https://github.com/openclaw/openclaw/pull/171180), a Telegram formatting fix for rich media captions. With rich Telegram messages enabled, photo and document captions could display literal block tags instead of clean native formatting.\n\nThe broken cases were familiar rich-message shapes: lists, checkboxes, asides, and dividers. Instead of seeing a readable caption, users could see raw markup-like tags, compact items jammed together, doubled checklist markers, or lost list numbering.\n\nThat is the kind of polish bug that makes an otherwise capable agent feel rough at the exact moment it is sending something visible to another person.\n\n## What Changed\n\nPR #171180 moves rich caption handling to the shared caption owner. Authored rich blocks are projected into Telegram-friendly caption text while preserving generated numbering and indentation through Markdown rendering.\n\nThe fix also reuses OpenClaw's existing quote-aware HTML tokenizer and attribute parser. That matters because Telegram captions are constrained: the formatter needs to convert genuinely authored structure while preserving literal examples and escaped tags that the user meant to show as text.\n\nThe PR specifically preserves:\n\n- Enclosing list counters.\n- Compact list item separation.\n- Checklist markers without doubling ordinary bullets.\n- Native inline formatting.\n- Literal code and escaped tag examples.\n- Asides mapped to supported blockquote formatting.\n\nLong captions keep the existing behavior: when a media caption cannot carry the full rich text, OpenClaw can send a separate rich-text follow-up instead of truncating the content.\n\n## Why Telegram Captions Are Tricky\n\nTelegram media captions are not the same as ordinary text messages. They have length limits, supported entity types, and different rendering behavior around photos and documents. A renderer that works for a normal chat message can still produce awkward output when the same rich block is attached to media.\n\nThat is why this fix is scoped. It does not rewrite every Telegram rendering path. It handles media-caption projection where the bug occurred, and it keeps plain captions, code examples, and escaped markup stable.\n\nThe result should be captions that look authored rather than leaked from the renderer's internal representation.\n\n## Verification\n\nThe PR includes unusually concrete Telegram evidence. The author ran a focused formatter suite with 67 passing tests, then compared a baseline and candidate using Telegram Test Server and a TDLib userbot with `richMessages: true`.\n\nThe same results were observed for photos and documents. Compact lists became separate bullet lines, nested lists retained outer numbering and child indentation, checklists had exactly one checkbox marker per item, quotes became native blockquotes, and code examples stayed literal.\n\nThe PR also reports changed-file formatting and lint passing, plus a real-provider CLI gate returning `OPENCLAW_CAPTION_LIVE_OK`.\n\n## Why It Matters\n\nTelegram is one of the places where OpenClaw's output becomes immediately social. A malformed caption is not just ugly; it can make an agent look careless in a channel, group, or workflow where the message is meant to be consumed by humans.\n\nRich captions are especially important for media-heavy tasks: summaries with screenshots, generated images, document delivery, checklist-style status reports, and structured updates. Those messages need bullets, quotes, and checklist states to survive the trip into Telegram.\n\n## Bottom Line\n\nPR #171180 makes Telegram rich media captions much more trustworthy. OpenClaw now converts common rich blocks into native caption formatting, preserves literal examples when users mean them literally, and keeps long-caption handling intact.\n\nFor Telegram users, the visible difference should be simple: fewer raw tags, cleaner captions, and media messages that look like they were meant for the conversation they landed in.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/171180\">PR #171180</a>, a Telegram formatting fix for rich media captions. With rich Telegram messages enabled, photo and document captions could display literal block tags instead of clean native formatting.</p><p>The broken cases were familiar rich-message shapes: lists, checkboxes, asides, and dividers. Instead of seeing a readable caption, users could see raw markup-like tags, compact items jammed together, doubled checklist markers, or lost list numbering.</p><p>That is the kind of polish bug that makes an otherwise capable agent feel rough at the exact moment it is sending something visible to another person.</p><h2>What Changed</h2><p>PR #171180 moves rich caption handling to the shared caption owner. Authored rich blocks are projected into Telegram-friendly caption text while preserving generated numbering and indentation through Markdown rendering.</p><p>The fix also reuses OpenClaw's existing quote-aware HTML tokenizer and attribute parser. That matters because Telegram captions are constrained: the formatter needs to convert genuinely authored structure while preserving literal examples and escaped tags that the user meant to show as text.</p><p>The PR specifically preserves:</p><ul><li>Enclosing list counters.</li><li>Compact list item separation.</li><li>Checklist markers without doubling ordinary bullets.</li><li>Native inline formatting.</li><li>Literal code and escaped tag examples.</li><li>Asides mapped to supported blockquote formatting.</li></ul><p>Long captions keep the existing behavior: when a media caption cannot carry the full rich text, OpenClaw can send a separate rich-text follow-up instead of truncating the content.</p><h2>Why Telegram Captions Are Tricky</h2><p>Telegram media captions are not the same as ordinary text messages. They have length limits, supported entity types, and different rendering behavior around photos and documents. A renderer that works for a normal chat message can still produce awkward output when the same rich block is attached to media.</p><p>That is why this fix is scoped. It does not rewrite every Telegram rendering path. It handles media-caption projection where the bug occurred, and it keeps plain captions, code examples, and escaped markup stable.</p><p>The result should be captions that look authored rather than leaked from the renderer's internal representation.</p><h2>Verification</h2><p>The PR includes unusually concrete Telegram evidence. The author ran a focused formatter suite with 67 passing tests, then compared a baseline and candidate using Telegram Test Server and a TDLib userbot with <code>richMessages: true</code>.</p><p>The same results were observed for photos and documents. Compact lists became separate bullet lines, nested lists retained outer numbering and child indentation, checklists had exactly one checkbox marker per item, quotes became native blockquotes, and code examples stayed literal.</p><p>The PR also reports changed-file formatting and lint passing, plus a real-provider CLI gate returning <code>OPENCLAW_CAPTION_LIVE_OK</code>.</p><h2>Why It Matters</h2><p>Telegram is one of the places where OpenClaw's output becomes immediately social. A malformed caption is not just ugly; it can make an agent look careless in a channel, group, or workflow where the message is meant to be consumed by humans.</p><p>Rich captions are especially important for media-heavy tasks: summaries with screenshots, generated images, document delivery, checklist-style status reports, and structured updates. Those messages need bullets, quotes, and checklist states to survive the trip into Telegram.</p><h2>Bottom Line</h2><p>PR #171180 makes Telegram rich media captions much more trustworthy. OpenClaw now converts common rich blocks into native caption formatting, preserves literal examples when users mean them literally, and keeps long-caption handling intact.</p><p>For Telegram users, the visible difference should be simple: fewer raw tags, cleaner captions, and media messages that look like they were meant for the conversation they landed in.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-telegram-caption-formatting.png\" alt=\"OpenClaw Fixes Telegram Rich Captions\" /></p>",
      "date_published": "2026-10-13T23:15:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Telegram",
        "Rich",
        "Captions"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-telegram-caption-formatting.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-telegram-caption-formatting.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Telegram Rich Captions cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-13-standalone-plugin-startup/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-13-standalone-plugin-startup/",
      "title": "OpenClaw Starts Standalone Plugins Correctly",
      "summary": "OpenClaw PR #171330 fixes enabled standalone plugin files so their Gateway hooks register at startup and enforce runtime policy as expected.",
      "content_text": "OpenClaw merged [PR #171330](https://github.com/openclaw/openclaw/pull/171330), a P1 plugin fix for a confusing Gateway startup gap: an explicitly enabled standalone plugin file could appear loaded in CLI inspection while never registering its hooks with the running Gateway.\n\nThat is a nasty kind of failure. The operator sees a plugin as loaded, but the behavior the plugin is supposed to enforce never runs.\n\nThe fix makes standalone policy files behave like users would expect. If a file is enabled through `plugins.load.paths`, its hooks can register during Gateway startup. Disabled standalone files stay excluded, and packaged plugins keep their existing startup behavior.\n\n## What Broke\n\nThe PR explains the root of the problem: a standalone file has no manifest where it can declare startup activation. OpenClaw already creates a synthetic manifest for that file, but the missing startup intent meant the Gateway did not activate the hooks even though inspection could show the plugin as loaded.\n\nThat split between \"loaded\" and \"enforced\" matters most for policy plugins. A maintenance-access policy, approval hook, or local guardrail is useful only if the Gateway actually invokes it at runtime.\n\nPR #171330 updates the synthetic manifest path so enabled standalone files supply the startup metadata they need, while leaving enablement policy authoritative.\n\n## User Impact\n\nThe user-facing result is direct: enabled standalone policy files now enforce their hooks in the running Gateway.\n\nThe PR's reproduction used an authenticated `POST /tools/invoke` request for `session_status`. On the baseline, a plugin configured to block that call was visible in inspection but the invocation still returned HTTP 200. On the candidate, the same enabled standalone file returned HTTP 403 with `tool_call_blocked`.\n\nThe important controls also held:\n\n- A packaged plugin with `activation.onStartup: true` kept enforcing its hook.\n- An explicitly disabled standalone file stayed disabled.\n- Existing configuration, plugin files, state directories, databases, and user artifacts were preserved across an upgrade from `openclaw@2026.10.3` to the candidate build.\n\nThat last point is important because plugin-loading bugs often intersect with startup and migration paths. This fix does not introduce a new configuration format or ask users to reset plugin state.\n\n## Why It Matters\n\nStandalone plugin files are a power-user feature. They are useful for local policy, narrow integrations, development hooks, and quick operational guardrails that do not need a full packaged plugin.\n\nBut power-user features still need crisp semantics. If the CLI says a policy is loaded, operators reasonably expect the Gateway to honor it. A mismatch there can produce a false sense of protection, especially when the plugin is meant to block a tool call or enforce a local workflow boundary.\n\nPR #171330 closes that gap by aligning runtime startup with inspection.\n\n## Verification\n\nThe PR reports maintainer reproduction in a secretless Linux and Node 24 container, using built Gateway entrypoints and authenticated tool invocations. It also tested an existing-install upgrade from npm `openclaw@2026.10.3` using the same configuration and state.\n\nFocused planner and manifest-registry suites passed with 69 tests. Candidate build, changed-file formatting, and oxlint also passed. The production change is small: the PR reports three production lines added and no production lines removed, with broader test coverage around the behavior.\n\n## Bottom Line\n\nPR #171330 is a small but high-leverage plugin reliability fix. OpenClaw now starts enabled standalone plugin files in a way that matches what inspection reports, so Gateway hooks are not merely visible; they are active.\n\nFor operators using standalone policy files, that turns a quiet footgun into a predictable startup contract.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/171330\">PR #171330</a>, a P1 plugin fix for a confusing Gateway startup gap: an explicitly enabled standalone plugin file could appear loaded in CLI inspection while never registering its hooks with the running Gateway.</p><p>That is a nasty kind of failure. The operator sees a plugin as loaded, but the behavior the plugin is supposed to enforce never runs.</p><p>The fix makes standalone policy files behave like users would expect. If a file is enabled through <code>plugins.load.paths</code>, its hooks can register during Gateway startup. Disabled standalone files stay excluded, and packaged plugins keep their existing startup behavior.</p><h2>What Broke</h2><p>The PR explains the root of the problem: a standalone file has no manifest where it can declare startup activation. OpenClaw already creates a synthetic manifest for that file, but the missing startup intent meant the Gateway did not activate the hooks even though inspection could show the plugin as loaded.</p><p>That split between \"loaded\" and \"enforced\" matters most for policy plugins. A maintenance-access policy, approval hook, or local guardrail is useful only if the Gateway actually invokes it at runtime.</p><p>PR #171330 updates the synthetic manifest path so enabled standalone files supply the startup metadata they need, while leaving enablement policy authoritative.</p><h2>User Impact</h2><p>The user-facing result is direct: enabled standalone policy files now enforce their hooks in the running Gateway.</p><p>The PR's reproduction used an authenticated <code>POST /tools/invoke</code> request for <code>session_status</code>. On the baseline, a plugin configured to block that call was visible in inspection but the invocation still returned HTTP 200. On the candidate, the same enabled standalone file returned HTTP 403 with <code>tool_call_blocked</code>.</p><p>The important controls also held:</p><ul><li>A packaged plugin with <code>activation.onStartup: true</code> kept enforcing its hook.</li><li>An explicitly disabled standalone file stayed disabled.</li><li>Existing configuration, plugin files, state directories, databases, and user artifacts were preserved across an upgrade from <code>openclaw@2026.10.3</code> to the candidate build.</li></ul><p>That last point is important because plugin-loading bugs often intersect with startup and migration paths. This fix does not introduce a new configuration format or ask users to reset plugin state.</p><h2>Why It Matters</h2><p>Standalone plugin files are a power-user feature. They are useful for local policy, narrow integrations, development hooks, and quick operational guardrails that do not need a full packaged plugin.</p><p>But power-user features still need crisp semantics. If the CLI says a policy is loaded, operators reasonably expect the Gateway to honor it. A mismatch there can produce a false sense of protection, especially when the plugin is meant to block a tool call or enforce a local workflow boundary.</p><p>PR #171330 closes that gap by aligning runtime startup with inspection.</p><h2>Verification</h2><p>The PR reports maintainer reproduction in a secretless Linux and Node 24 container, using built Gateway entrypoints and authenticated tool invocations. It also tested an existing-install upgrade from npm <code>openclaw@2026.10.3</code> using the same configuration and state.</p><p>Focused planner and manifest-registry suites passed with 69 tests. Candidate build, changed-file formatting, and oxlint also passed. The production change is small: the PR reports three production lines added and no production lines removed, with broader test coverage around the behavior.</p><h2>Bottom Line</h2><p>PR #171330 is a small but high-leverage plugin reliability fix. OpenClaw now starts enabled standalone plugin files in a way that matches what inspection reports, so Gateway hooks are not merely visible; they are active.</p><p>For operators using standalone policy files, that turns a quiet footgun into a predictable startup contract.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-standalone-plugin-startup.png\" alt=\"OpenClaw Starts Standalone Plugins Correctly\" /></p>",
      "date_published": "2026-10-13T23:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Starts",
        "Standalone",
        "Plugins",
        "Correctly"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-standalone-plugin-startup.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-standalone-plugin-startup.png",
          "mime_type": "image/png",
          "title": "OpenClaw Starts Standalone Plugins Correctly cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-13-2026-10-3-release/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-13-2026-10-3-release/",
      "title": "OpenClaw 2026.10.3 Ships Update Fixes",
      "summary": "OpenClaw 2026.10.3 and beta 2026.10.5-beta.3 focus on safer updates, preserved credentials, restart recovery, and new model support for operators.",
      "content_text": "OpenClaw published two release signals after this morning's aggregation window: the stable [OpenClaw v2026.10.3 release](https://github.com/openclaw/openclaw/releases/tag/v2026.10.3) and the follow-up [v2026.10.5-beta.3 release](https://github.com/openclaw/openclaw/releases/tag/v2026.10.5-beta.3). npm also moved `latest` to `2026.10.3`, while the beta dist-tag advanced to `2026.10.5-beta.3`.\n\nThe stable release is the more important production story. Its changelog says OpenClaw 2026.10.3 fixes update blockers, preserves saved sign-ins and memory files during affected Doctor repairs, helps the Mac app connect during local setup, and lets conversations continue after compaction or interrupted provider replies.\n\nThat is not a flashy feature release. It is the kind of maintenance-heavy release that matters when OpenClaw is running as infrastructure.\n\n## What Stable Users Get\n\nThe v2026.10.3 changelog groups much of the work around update and maintenance recovery. Several fixes address cases where leftover recovery records, interrupted 2026.9.8 updates, abandoned npm updater copies, Windows path differences, pnpm service commands, and briefly busy state databases could block or confuse the next upgrade.\n\nFor operators, the pattern is clear:\n\n- Updates get better at clearing eligible stale state.\n- Doctor reports more useful next steps when maintenance is unfinished.\n- Recovery copies include more of the data that matters before repair.\n- Update checks are less likely to mistake their own bookkeeping for user-caused change.\n\nThe release also calls out a privacy and continuity fix: Doctor no longer discards saved OAuth profiles when an unreadable legacy auth file blocks cleanup. That is an important quality bar for repair tooling. A recovery command should avoid making credential loss worse while it is trying to help.\n\n## Mac, Models, and Messaging\n\nThe stable release is not only about updates. The Mac app gets fixes for eligible \"On this Mac\" setups that previously failed with \"Gateway rejected auth\", plus reconnection after trusted Gateway certificate renewal. On the provider side, Meta users get Muse Wonder support, MiniMax gets M3.1 Flash Preview support, and interrupted ChatGPT replies can continue from saved conversation state within existing retry limits.\n\nMessaging also receives practical reliability fixes. Channel tools keep working after the task that started or restarted a messaging account finishes, and recovery scanning avoids stealing a message that already has an active sender.\n\nThose are small details, but they are exactly the details that make chat-driven systems feel dependable.\n\n## What Beta 3 Adds\n\nThe beta release continues the same reliability arc, but with a broader preview surface. Its highlights include update paths that leave the Gateway running, recovery for stuck package activation journals, Doctor fixes that keep credentials and data, restart survival for interrupted turns and subagent work, and Telegram spool handling that prevents old updates from replaying as new instructions.\n\nThe beta also adds Meta Muse Wonder with tools, image input, streaming, reasoning controls, and a 1M-token context window, and restores a deprecated `createSessionHeaderLink` export for plugins built against OpenClaw 2026.10.1 and 2026.10.2.\n\n## Why This Release Matters\n\nOpenClaw's release stream has been heavy on update, Doctor, Gateway, and session durability work all week. v2026.10.3 is the stable consolidation point for a lot of that effort.\n\nIf you run OpenClaw casually, the headline is simple: fewer stuck updates and more resilient setup paths. If you run it as a long-lived agent host, the bigger story is stronger repair behavior around credentials, memory, plugin state, Mac app trust, and interrupted model replies.\n\nThe beta channel then previews the next layer: keeping Gateway services alive through more update shapes, protecting Telegram users from stale legacy spool replay, and extending restart recovery deeper into turns, automations, nodes, and subagents.\n\n## Bottom Line\n\nOpenClaw 2026.10.3 is a maintenance release with real operational weight. It moves stable users forward on update recovery, Doctor safety, Mac setup, provider continuity, and messaging reliability.\n\nThe beta 2026.10.5-beta.3 release shows where the next stable track is heading: more update paths that recover cleanly, fewer restart edge cases, and a wider model/provider surface without sacrificing the repair work that keeps running systems boring.",
      "content_html": "<p>OpenClaw published two release signals after this morning's aggregation window: the stable <a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.10.3\">OpenClaw v2026.10.3 release</a> and the follow-up <a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.10.5-beta.3\">v2026.10.5-beta.3 release</a>. npm also moved <code>latest</code> to <code>2026.10.3</code>, while the beta dist-tag advanced to <code>2026.10.5-beta.3</code>.</p><p>The stable release is the more important production story. Its changelog says OpenClaw 2026.10.3 fixes update blockers, preserves saved sign-ins and memory files during affected Doctor repairs, helps the Mac app connect during local setup, and lets conversations continue after compaction or interrupted provider replies.</p><p>That is not a flashy feature release. It is the kind of maintenance-heavy release that matters when OpenClaw is running as infrastructure.</p><h2>What Stable Users Get</h2><p>The v2026.10.3 changelog groups much of the work around update and maintenance recovery. Several fixes address cases where leftover recovery records, interrupted 2026.9.8 updates, abandoned npm updater copies, Windows path differences, pnpm service commands, and briefly busy state databases could block or confuse the next upgrade.</p><p>For operators, the pattern is clear:</p><ul><li>Updates get better at clearing eligible stale state.</li><li>Doctor reports more useful next steps when maintenance is unfinished.</li><li>Recovery copies include more of the data that matters before repair.</li><li>Update checks are less likely to mistake their own bookkeeping for user-caused change.</li></ul><p>The release also calls out a privacy and continuity fix: Doctor no longer discards saved OAuth profiles when an unreadable legacy auth file blocks cleanup. That is an important quality bar for repair tooling. A recovery command should avoid making credential loss worse while it is trying to help.</p><h2>Mac, Models, and Messaging</h2><p>The stable release is not only about updates. The Mac app gets fixes for eligible \"On this Mac\" setups that previously failed with \"Gateway rejected auth\", plus reconnection after trusted Gateway certificate renewal. On the provider side, Meta users get Muse Wonder support, MiniMax gets M3.1 Flash Preview support, and interrupted ChatGPT replies can continue from saved conversation state within existing retry limits.</p><p>Messaging also receives practical reliability fixes. Channel tools keep working after the task that started or restarted a messaging account finishes, and recovery scanning avoids stealing a message that already has an active sender.</p><p>Those are small details, but they are exactly the details that make chat-driven systems feel dependable.</p><h2>What Beta 3 Adds</h2><p>The beta release continues the same reliability arc, but with a broader preview surface. Its highlights include update paths that leave the Gateway running, recovery for stuck package activation journals, Doctor fixes that keep credentials and data, restart survival for interrupted turns and subagent work, and Telegram spool handling that prevents old updates from replaying as new instructions.</p><p>The beta also adds Meta Muse Wonder with tools, image input, streaming, reasoning controls, and a 1M-token context window, and restores a deprecated <code>createSessionHeaderLink</code> export for plugins built against OpenClaw 2026.10.1 and 2026.10.2.</p><h2>Why This Release Matters</h2><p>OpenClaw's release stream has been heavy on update, Doctor, Gateway, and session durability work all week. v2026.10.3 is the stable consolidation point for a lot of that effort.</p><p>If you run OpenClaw casually, the headline is simple: fewer stuck updates and more resilient setup paths. If you run it as a long-lived agent host, the bigger story is stronger repair behavior around credentials, memory, plugin state, Mac app trust, and interrupted model replies.</p><p>The beta channel then previews the next layer: keeping Gateway services alive through more update shapes, protecting Telegram users from stale legacy spool replay, and extending restart recovery deeper into turns, automations, nodes, and subagents.</p><h2>Bottom Line</h2><p>OpenClaw 2026.10.3 is a maintenance release with real operational weight. It moves stable users forward on update recovery, Doctor safety, Mac setup, provider continuity, and messaging reliability.</p><p>The beta 2026.10.5-beta.3 release shows where the next stable track is heading: more update paths that recover cleanly, fewer restart edge cases, and a wider model/provider surface without sacrificing the repair work that keeps running systems boring.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-2026-10-3-release.png\" alt=\"OpenClaw 2026.10.3 Ships Update Fixes\" /></p>",
      "date_published": "2026-10-13T23:05:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "2026",
        "Ships",
        "Update",
        "Fixes"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-2026-10-3-release.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-2026-10-3-release.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.10.3 Ships Update Fixes cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-13-interrupted-run-history/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-13-interrupted-run-history/",
      "title": "OpenClaw Shows Interrupted Runs More Clearly",
      "summary": "OpenClaw PR #170705 restores commentary, activity groups, and an Interrupted state when a restarted run has no final assistant reply.",
      "content_text": "OpenClaw merged [PR #170705](https://github.com/openclaw/openclaw/pull/170705), a Control UI fix for a confusing transcript state after a Gateway restart interrupts a run between tool iterations.\n\nBefore the fix, an interrupted run could appear as a user message followed by one collapsed activity line. If tool display was off, it could look like there was no visible assistant-side activity at all. That hid saved commentary and made the run feel like it had silently vanished.\n\nThe new behavior keeps the useful history visible: commentary appears inline, tool activity groups remain available, and the run can show an `Interrupted` state when there is no terminal assistant reply.\n\n## What Changed\n\nThe core issue was classification. The chat normalizer treated any row containing a tool call as a tool row, even if that assistant row also carried narration. Grouping then folded the narration into a collapsed tool disclosure. With tools hidden, the thread builder skipped those rows entirely.\n\nPR #170705 separates narration from mixed assistant/tool rows so it can stay visible in assistant groups. It also marks an inactive run as interrupted when its last assistant row is a `toolUse` row and there is no terminal row.\n\nThe fix is presentation-only. The PR explicitly says transcript persistence and restart recovery are handled separately in other PRs. That distinction matters: this patch is about making already-saved history understandable in the UI.\n\n## Why It Matters\n\nAgent systems need good failure presentation. A run that was interrupted by a restart should not look empty, complete, or mysterious. Users need to know what happened, what was already said, and what activity was in progress.\n\nThis is especially important when tool calls are hidden. Hiding tool details should reduce noise, not erase the only visible evidence that the assistant had started work.\n\nBy preserving commentary and showing an interrupted state, the Control UI gives the user enough context to decide what to do next. They can continue, rerun, inspect activity, or simply understand why a final answer never appeared.\n\n## Verification\n\nThe PR includes a synthetic fixture shaped like production history: a user turn, tool rows, two commentary rows, and no terminal assistant row. Before the fix, commentary was missing. After the fix, commentary, activity, and the interrupted state survive reloads.\n\nReported verification includes:\n\n- A transcript-render regression covering both tool-shown and tool-hidden modes.\n- A mocked-Gateway end-to-end test where commentary, activity, and `Interrupted` survive reload.\n- 165 chat transcript, agent-run grouping, thread grouping, and thread tests passing.\n- UI typechecks, changed-file lint, format checks, and `git diff --check` passing.\n\n## Bottom Line\n\nPR #170705 makes interrupted agent history honest again. When a Gateway restart cuts a run short before a final assistant reply, OpenClaw can now show the commentary it already saved, keep the related activity accessible, and label the run as interrupted.\n\nThat is a small presentation fix with a big trust payoff: the UI no longer hides the trail just when users most need it.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/170705\">PR #170705</a>, a Control UI fix for a confusing transcript state after a Gateway restart interrupts a run between tool iterations.</p><p>Before the fix, an interrupted run could appear as a user message followed by one collapsed activity line. If tool display was off, it could look like there was no visible assistant-side activity at all. That hid saved commentary and made the run feel like it had silently vanished.</p><p>The new behavior keeps the useful history visible: commentary appears inline, tool activity groups remain available, and the run can show an <code>Interrupted</code> state when there is no terminal assistant reply.</p><h2>What Changed</h2><p>The core issue was classification. The chat normalizer treated any row containing a tool call as a tool row, even if that assistant row also carried narration. Grouping then folded the narration into a collapsed tool disclosure. With tools hidden, the thread builder skipped those rows entirely.</p><p>PR #170705 separates narration from mixed assistant/tool rows so it can stay visible in assistant groups. It also marks an inactive run as interrupted when its last assistant row is a <code>toolUse</code> row and there is no terminal row.</p><p>The fix is presentation-only. The PR explicitly says transcript persistence and restart recovery are handled separately in other PRs. That distinction matters: this patch is about making already-saved history understandable in the UI.</p><h2>Why It Matters</h2><p>Agent systems need good failure presentation. A run that was interrupted by a restart should not look empty, complete, or mysterious. Users need to know what happened, what was already said, and what activity was in progress.</p><p>This is especially important when tool calls are hidden. Hiding tool details should reduce noise, not erase the only visible evidence that the assistant had started work.</p><p>By preserving commentary and showing an interrupted state, the Control UI gives the user enough context to decide what to do next. They can continue, rerun, inspect activity, or simply understand why a final answer never appeared.</p><h2>Verification</h2><p>The PR includes a synthetic fixture shaped like production history: a user turn, tool rows, two commentary rows, and no terminal assistant row. Before the fix, commentary was missing. After the fix, commentary, activity, and the interrupted state survive reloads.</p><p>Reported verification includes:</p><ul><li>A transcript-render regression covering both tool-shown and tool-hidden modes.</li><li>A mocked-Gateway end-to-end test where commentary, activity, and <code>Interrupted</code> survive reload.</li><li>165 chat transcript, agent-run grouping, thread grouping, and thread tests passing.</li><li>UI typechecks, changed-file lint, format checks, and <code>git diff --check</code> passing.</li></ul><h2>Bottom Line</h2><p>PR #170705 makes interrupted agent history honest again. When a Gateway restart cuts a run short before a final assistant reply, OpenClaw can now show the commentary it already saved, keep the related activity accessible, and label the run as interrupted.</p><p>That is a small presentation fix with a big trust payoff: the UI no longer hides the trail just when users most need it.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-interrupted-run-history.png\" alt=\"OpenClaw Shows Interrupted Runs More Clearly\" /></p>",
      "date_published": "2026-10-13T08:15:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Shows",
        "Interrupted",
        "Runs",
        "More",
        "Clearly"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-interrupted-run-history.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-interrupted-run-history.png",
          "mime_type": "image/png",
          "title": "OpenClaw Shows Interrupted Runs More Clearly cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-13-control-ui-offscreen-reveals/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-13-control-ui-offscreen-reveals/",
      "title": "OpenClaw Unfreezes Offscreen Reply Navigation",
      "summary": "OpenClaw PR #170789 removes a redundant render loop that could freeze offscreen reply navigation and position-rail jumps in the Control UI.",
      "content_text": "OpenClaw merged [PR #170789](https://github.com/openclaw/openclaw/pull/170789), a focused Control UI fix for a renderer freeze in offscreen reply navigation and position-rail jumps.\n\nThe bug lived in a subtle loop. When the target message bubble was not mounted, `completeMessageReveal()` would scroll toward the virtual row and then unconditionally request another render. A prior change moved layout completion from a browser frame to a microtask, and that made the repeated render request capable of starving the native scroll notification that needed to advance the virtual range.\n\nIn plain English: the UI could get stuck trying to reveal a message that was not currently mounted.\n\n## What Changed\n\nThe fix removes the redundant render request. The first reveal still asks for an update when folded work needs expansion, but subsequent movement is owned by the actual mechanisms that can make progress:\n\n- Native scroll notifications.\n- Mounted rows.\n- Deferred projection settlement.\n- Existing target highlighting and cancellation behavior.\n\nThe accepted edge case is also clear. If the target is absent, OpenClaw now waits for a real scroll, row mount, or projection event instead of polling unchanged DOM.\n\nThat tradeoff matters because the old behavior looked more aggressive but could block the very signal it needed. The new behavior is less noisy and more faithful to the renderer's actual state.\n\n## Why It Matters\n\nReply navigation is one of those features users notice only when it fails. In a long agent conversation, jumping back to the source of a reply, moving through the position rail, or revealing an earlier collapsed message needs to feel boringly reliable.\n\nThe PR's baseline evidence captured failures in several original cases: deep standalone-work reply reveal, dark and light position-rail navigation, and reply-source navigation after reconnect. The candidate then passed those original cases in a focused hosted run.\n\nThis is not a broad UI rewrite. The PR reports a one-file change with only a small line delta. That is good news for a renderer bug: less surface area means less chance of moving unrelated layout or transcript behavior.\n\n## Verification\n\nThe PR cites a focused hosted run where all four original cases passed:\n\n- Dark and light position-rail navigation.\n- Deep earlier-message reveal.\n- Reconnect source navigation.\n\nIt also reports that changed-file formatting, linting, and `git diff --check` passed, with an independent Codex autoreview finding no actionable P0-P2 issues. Ordinary PR CI passed selected required jobs, including security, types, lint, architecture, and static gates.\n\n## Bottom Line\n\nPR #170789 is a classic small UI reliability patch: remove the extra render nudge, let the real scroll and mount events drive progress, and keep the existing navigation semantics intact.\n\nFor anyone using long OpenClaw conversations in the Control UI, the result should be smoother jumps to earlier replies and fewer moments where the interface seems to lose its place.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/170789\">PR #170789</a>, a focused Control UI fix for a renderer freeze in offscreen reply navigation and position-rail jumps.</p><p>The bug lived in a subtle loop. When the target message bubble was not mounted, <code>completeMessageReveal()</code> would scroll toward the virtual row and then unconditionally request another render. A prior change moved layout completion from a browser frame to a microtask, and that made the repeated render request capable of starving the native scroll notification that needed to advance the virtual range.</p><p>In plain English: the UI could get stuck trying to reveal a message that was not currently mounted.</p><h2>What Changed</h2><p>The fix removes the redundant render request. The first reveal still asks for an update when folded work needs expansion, but subsequent movement is owned by the actual mechanisms that can make progress:</p><ul><li>Native scroll notifications.</li><li>Mounted rows.</li><li>Deferred projection settlement.</li><li>Existing target highlighting and cancellation behavior.</li></ul><p>The accepted edge case is also clear. If the target is absent, OpenClaw now waits for a real scroll, row mount, or projection event instead of polling unchanged DOM.</p><p>That tradeoff matters because the old behavior looked more aggressive but could block the very signal it needed. The new behavior is less noisy and more faithful to the renderer's actual state.</p><h2>Why It Matters</h2><p>Reply navigation is one of those features users notice only when it fails. In a long agent conversation, jumping back to the source of a reply, moving through the position rail, or revealing an earlier collapsed message needs to feel boringly reliable.</p><p>The PR's baseline evidence captured failures in several original cases: deep standalone-work reply reveal, dark and light position-rail navigation, and reply-source navigation after reconnect. The candidate then passed those original cases in a focused hosted run.</p><p>This is not a broad UI rewrite. The PR reports a one-file change with only a small line delta. That is good news for a renderer bug: less surface area means less chance of moving unrelated layout or transcript behavior.</p><h2>Verification</h2><p>The PR cites a focused hosted run where all four original cases passed:</p><ul><li>Dark and light position-rail navigation.</li><li>Deep earlier-message reveal.</li><li>Reconnect source navigation.</li></ul><p>It also reports that changed-file formatting, linting, and <code>git diff --check</code> passed, with an independent Codex autoreview finding no actionable P0-P2 issues. Ordinary PR CI passed selected required jobs, including security, types, lint, architecture, and static gates.</p><h2>Bottom Line</h2><p>PR #170789 is a classic small UI reliability patch: remove the extra render nudge, let the real scroll and mount events drive progress, and keep the existing navigation semantics intact.</p><p>For anyone using long OpenClaw conversations in the Control UI, the result should be smoother jumps to earlier replies and fewer moments where the interface seems to lose its place.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-control-ui-offscreen-reveals.png\" alt=\"OpenClaw Unfreezes Offscreen Reply Navigation\" /></p>",
      "date_published": "2026-10-13T08:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Unfreezes",
        "Offscreen",
        "Reply",
        "Navigation"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-control-ui-offscreen-reveals.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-control-ui-offscreen-reveals.png",
          "mime_type": "image/png",
          "title": "OpenClaw Unfreezes Offscreen Reply Navigation cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-13-workspace-tools-continuations/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-13-workspace-tools-continuations/",
      "title": "OpenClaw Fixes Workspace Tools After Continuations",
      "summary": "OpenClaw PR #170791 fixes workspace tools after session continuation so resumed turns can keep using prepared project files without weakening stale-tool checks.",
      "content_text": "OpenClaw merged [PR #170791](https://github.com/openclaw/openclaw/pull/170791) this morning to fix a sharp reliability bug in resumed sessions: workspace tools could fail with `Session changed while preparing its project` after a child result or other internal continuation.\n\nThat sounds narrow, but it lands in a very real workflow. If an agent resumes after delegated work, the user still expects the prepared project workspace to be available. Before this fix, the resumed turn could inherit an expired workspace-readiness context from the previous chat run. The result was a tool surface that looked prepared but refused to operate when the agent tried to read or use project files.\n\n## What Changed\n\nThe PR changes how autonomous runs and chat dispatch handle workspace-readiness context:\n\n- Autonomous runs now start without inheriting the previous run's expired readiness context.\n- Chat dispatch installs its own context even when no new preparation is pending.\n- Existing stale-tool safety checks remain enforced.\n- Existing cancellation and workspace-preparation behavior remain in place.\n\nThe user-facing version is simpler: continuing a session after internal work should no longer break the workspace tools that the session is supposed to have.\n\n## Why This Matters\n\nOpenClaw leans heavily on continuation. Parent sessions hand work to child sessions, agent runs resume after internal events, and a lot of useful automation depends on keeping those boundaries crisp. A continuation bug in workspace preparation is easy to miss in happy-path testing because the original workspace setup may have succeeded.\n\nThe important part of PR #170791 is that it does not solve the problem by relaxing stale-tool checks. The PR body says old retained tools remain denied, and the new regression tests use real host-bound filesystem reads to verify that the fresh continuation works while stale access still fails.\n\nThat is the right shape for this class of bug. The fix restores legitimate continuity without widening the authority of old tool handles.\n\n## Verification\n\nThe PR reports a direct reproduction of the error after a child-result continuation. Both new successor-boundary regressions failed before the patch and passed after it.\n\nThe focused evidence includes:\n\n- 32 Gateway tests passing.\n- 16 existing workspace-readiness safety tests passing.\n- Targeted typecheck, lint, formatting, and diff checks passing.\n- An independent review with no actionable P0-P2 findings.\n\nThe PR also notes one unrelated product fixture that did not reach its early-prompt precondition, so the author does not overclaim that scenario. That restraint is useful: the merged fix is specifically about continuation-owned workspace readiness, not every possible startup edge case.\n\n## Bottom Line\n\nPR #170791 is a small Gateway fix with a concrete quality-of-life effect. If a session resumes after child work or another internal continuation, OpenClaw should keep the prepared workspace usable for the right run and continue rejecting stale tool access from the wrong one.\n\nThat is exactly the kind of continuity polish that makes long-running agent work feel less brittle.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/170791\">PR #170791</a> this morning to fix a sharp reliability bug in resumed sessions: workspace tools could fail with <code>Session changed while preparing its project</code> after a child result or other internal continuation.</p><p>That sounds narrow, but it lands in a very real workflow. If an agent resumes after delegated work, the user still expects the prepared project workspace to be available. Before this fix, the resumed turn could inherit an expired workspace-readiness context from the previous chat run. The result was a tool surface that looked prepared but refused to operate when the agent tried to read or use project files.</p><h2>What Changed</h2><p>The PR changes how autonomous runs and chat dispatch handle workspace-readiness context:</p><ul><li>Autonomous runs now start without inheriting the previous run's expired readiness context.</li><li>Chat dispatch installs its own context even when no new preparation is pending.</li><li>Existing stale-tool safety checks remain enforced.</li><li>Existing cancellation and workspace-preparation behavior remain in place.</li></ul><p>The user-facing version is simpler: continuing a session after internal work should no longer break the workspace tools that the session is supposed to have.</p><h2>Why This Matters</h2><p>OpenClaw leans heavily on continuation. Parent sessions hand work to child sessions, agent runs resume after internal events, and a lot of useful automation depends on keeping those boundaries crisp. A continuation bug in workspace preparation is easy to miss in happy-path testing because the original workspace setup may have succeeded.</p><p>The important part of PR #170791 is that it does not solve the problem by relaxing stale-tool checks. The PR body says old retained tools remain denied, and the new regression tests use real host-bound filesystem reads to verify that the fresh continuation works while stale access still fails.</p><p>That is the right shape for this class of bug. The fix restores legitimate continuity without widening the authority of old tool handles.</p><h2>Verification</h2><p>The PR reports a direct reproduction of the error after a child-result continuation. Both new successor-boundary regressions failed before the patch and passed after it.</p><p>The focused evidence includes:</p><ul><li>32 Gateway tests passing.</li><li>16 existing workspace-readiness safety tests passing.</li><li>Targeted typecheck, lint, formatting, and diff checks passing.</li><li>An independent review with no actionable P0-P2 findings.</li></ul><p>The PR also notes one unrelated product fixture that did not reach its early-prompt precondition, so the author does not overclaim that scenario. That restraint is useful: the merged fix is specifically about continuation-owned workspace readiness, not every possible startup edge case.</p><h2>Bottom Line</h2><p>PR #170791 is a small Gateway fix with a concrete quality-of-life effect. If a session resumes after child work or another internal continuation, OpenClaw should keep the prepared workspace usable for the right run and continue rejecting stale tool access from the wrong one.</p><p>That is exactly the kind of continuity polish that makes long-running agent work feel less brittle.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-workspace-tools-continuations.png\" alt=\"OpenClaw Fixes Workspace Tools After Continuations\" /></p>",
      "date_published": "2026-10-13T08:05:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Workspace",
        "Tools",
        "After",
        "Continuations"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-workspace-tools-continuations.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-13-workspace-tools-continuations.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Workspace Tools After Continuations cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-12-web-fetch-json-number-fidelity/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-12-web-fetch-json-number-fidelity/",
      "title": "OpenClaw Web Fetch Preserves JSON Numbers",
      "summary": "OpenClaw PR #170123 fixes web_fetch so models see exact JSON number tokens instead of rounded IDs, null infinities, or rewritten negative zero.",
      "content_text": "OpenClaw merged [PR #170123](https://github.com/openclaw/openclaw/pull/170123), a deceptively important fix for `web_fetch`: JSON numbers are now preserved as exact source tokens before the model sees them.\n\nThe bug was classic JavaScript footgun territory. When `web_fetch` parsed JSON and serialized it back for model-visible output, large identifiers and high-precision decimals could be rounded, very large exponents could become `null`, and negative zero could be rewritten as `0`.\n\nFor ordinary web pages, that might not matter. For agent work, it absolutely can.\n\n## Why JSON Number Fidelity Matters\n\nAgents increasingly fetch APIs, not just human-readable pages. Those APIs contain issue IDs, database IDs, payment IDs, timestamps, balances, prices, sequence numbers, and scientific values. A model does not need those values to be \"numerically convenient.\" It needs the exact token the server sent.\n\nThe PR's examples include identifiers such as `9223372036854775807` and `12345678901234567890`, high-precision decimals, `1e400`, and `-0`. Those are the kinds of values that expose the mismatch between JavaScript's default number behavior and a tool result that is supposed to preserve source data.\n\nIf an agent reads a large ID incorrectly, the failure can be subtle. A rounded identifier can still look plausible. A nullified exponent can look like missing data. A changed zero sign can matter in numeric protocols or tests. The dangerous part is not just that the value changes; it is that the change can be silent.\n\n## The Shape Of The Fix\n\nThe merged implementation keeps the existing two-space JSON formatting, size limits, truncation behavior, malformed JSON fallback, content-type classification, external-content wrappers, and non-JSON handling.\n\nThe change is focused on numeric token preservation. The PR uses the native JSON reviver source context with `JSON.rawJSON` for numbers, then keeps the existing serializer and indentation path. Runtime support was exercised across supported Node 24 and 26 branches, plus the repository-pinned OpenClaw Bun version.\n\nThat narrowness is good engineering. The visible output format should not churn just because one fidelity problem was fixed. Existing agents, tests, and truncation boundaries depend on stable formatting.\n\n## What Changes For Models\n\nThe model now receives the server's exact numeric tokens in JSON tool results. In the PR's recorded proof, main rounded or rewrote several problematic values, while the candidate preserved them. Ordinary JSON, HTML, oversized JSON truncation, and non-JSON extraction remained equivalent after expected normalization.\n\nThis is one of those fixes that improves trust in the tool layer. `web_fetch` is a bridge between the internet and the model. If that bridge quietly edits data, the model can produce confidently wrong follow-up actions.\n\n## A Small Fix With A Large Blast Radius\n\nThe PR is labeled P2, but its blast radius is broad because `web_fetch` is such a basic primitive. Any workflow that asks OpenClaw to inspect API responses, structured docs, registry metadata, analytics payloads, or machine-generated reports benefits from exact JSON number handling.\n\nFor users, there is no new setting to flip. The improvement is that fetched JSON should now be a more faithful representation of the source response, especially in the edge cases where \"close enough\" was never actually close enough.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/170123\">PR #170123</a>, a deceptively important fix for <code>web_fetch</code>: JSON numbers are now preserved as exact source tokens before the model sees them.</p><p>The bug was classic JavaScript footgun territory. When <code>web_fetch</code> parsed JSON and serialized it back for model-visible output, large identifiers and high-precision decimals could be rounded, very large exponents could become <code>null</code>, and negative zero could be rewritten as <code>0</code>.</p><p>For ordinary web pages, that might not matter. For agent work, it absolutely can.</p><h2>Why JSON Number Fidelity Matters</h2><p>Agents increasingly fetch APIs, not just human-readable pages. Those APIs contain issue IDs, database IDs, payment IDs, timestamps, balances, prices, sequence numbers, and scientific values. A model does not need those values to be \"numerically convenient.\" It needs the exact token the server sent.</p><p>The PR's examples include identifiers such as <code>9223372036854775807</code> and <code>12345678901234567890</code>, high-precision decimals, <code>1e400</code>, and <code>-0</code>. Those are the kinds of values that expose the mismatch between JavaScript's default number behavior and a tool result that is supposed to preserve source data.</p><p>If an agent reads a large ID incorrectly, the failure can be subtle. A rounded identifier can still look plausible. A nullified exponent can look like missing data. A changed zero sign can matter in numeric protocols or tests. The dangerous part is not just that the value changes; it is that the change can be silent.</p><h2>The Shape Of The Fix</h2><p>The merged implementation keeps the existing two-space JSON formatting, size limits, truncation behavior, malformed JSON fallback, content-type classification, external-content wrappers, and non-JSON handling.</p><p>The change is focused on numeric token preservation. The PR uses the native JSON reviver source context with <code>JSON.rawJSON</code> for numbers, then keeps the existing serializer and indentation path. Runtime support was exercised across supported Node 24 and 26 branches, plus the repository-pinned OpenClaw Bun version.</p><p>That narrowness is good engineering. The visible output format should not churn just because one fidelity problem was fixed. Existing agents, tests, and truncation boundaries depend on stable formatting.</p><h2>What Changes For Models</h2><p>The model now receives the server's exact numeric tokens in JSON tool results. In the PR's recorded proof, main rounded or rewrote several problematic values, while the candidate preserved them. Ordinary JSON, HTML, oversized JSON truncation, and non-JSON extraction remained equivalent after expected normalization.</p><p>This is one of those fixes that improves trust in the tool layer. <code>web_fetch</code> is a bridge between the internet and the model. If that bridge quietly edits data, the model can produce confidently wrong follow-up actions.</p><h2>A Small Fix With A Large Blast Radius</h2><p>The PR is labeled P2, but its blast radius is broad because <code>web_fetch</code> is such a basic primitive. Any workflow that asks OpenClaw to inspect API responses, structured docs, registry metadata, analytics payloads, or machine-generated reports benefits from exact JSON number handling.</p><p>For users, there is no new setting to flip. The improvement is that fetched JSON should now be a more faithful representation of the source response, especially in the edge cases where \"close enough\" was never actually close enough.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-web-fetch-json-number-fidelity.png\" alt=\"OpenClaw Web Fetch Preserves JSON Numbers\" /></p>",
      "date_published": "2026-10-12T23:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fetch",
        "Preserves",
        "Json",
        "Numbers"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-web-fetch-json-number-fidelity.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-web-fetch-json-number-fidelity.png",
          "mime_type": "image/png",
          "title": "OpenClaw Web Fetch Preserves JSON Numbers cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-12-worker-backed-session-reads/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-12-worker-backed-session-reads/",
      "title": "OpenClaw Moves Session Reads Into Workers",
      "summary": "OpenClaw PR #169622 routes bundled session reads through async workers, lowering Gateway thread pressure while preserving authority checks.",
      "content_text": "OpenClaw merged [PR #169622](https://github.com/openclaw/openclaw/pull/169622), a large Gateway refactor that moves bundled session reads and authority preparation through async workers instead of ordinary synchronous SQLite paths.\n\nThis is not flashy in the product-marketing sense. It is the kind of plumbing change that matters when an agent runtime has to keep chats, plugins, memory, cron, channels, and UI views responsive at the same time.\n\n## The Problem: Ordinary Reads Still Hit The Owner Path\n\nThe PR says ordinary session reads in Gateway routing, agent execution, sandbox policy, plugin operations, Memory SDK consumers, and status rendering still reached synchronous SQLite through owner and SDK APIs.\n\nIn a small local setup, that may be invisible. In a busy Gateway, those read paths become part of the latency budget for every turn. They also make it harder to reason about ownership, because a read that looks descriptive can still cross the same surface as effect-bearing state.\n\nThe merged change routes bundled callers through existing workers or prepared facts updated by committed write receipts. The Gateway remains the single database owner, but more consumers await the worker-backed view instead of touching synchronous session storage directly.\n\n## The Boundary Did Not Disappear\n\nThe interesting part is what the PR explicitly keeps. Final authority checks remain at execution, mutation, disclosure, approvals, credentials, and external publication. The change is not \"trust cached state forever.\" It is \"carry the prepared facts through the owning operation, then still check authority where real effects happen.\"\n\nThat distinction matters for OpenClaw's security model. Faster session reads are useful only if they do not let stale membership, revoked access, or changed policy leak through to a tool call, credential access, publication, or channel send.\n\nThe PR also keeps four named effect-boundary synchronous checks for cases where current state is the point of the check: Discord voice steering, ClickClack disclosure and retention, and external channel creation or publication.\n\n## What Users Should Notice\n\nMost users will not see a new command or setting from this merge. The expected user-facing effect is steadier responsiveness under load and fewer hidden stalls when multiple parts of OpenClaw are inspecting session state.\n\nThe PR's live measurement is careful about scope, but the direction is still notable. Its warmed shared-harness comparison reported lower all-thread SQL statements and executing requests per turn, with MAIN-thread SQL at zero in the candidate path.\n\nThat is the arc OpenClaw has been following for months: move work off the hot Gateway path, keep ownership explicit, and preserve security-sensitive checks at the moment an effect is about to happen.\n\n## Why It Belongs In The Release Story\n\nThis merge landed shortly after 2026.10.5-beta.2, and it fits the same theme as the beta notes. Recovery, session routing, tool approvals, memory, and channel delivery are all converging on a more explicit authority model.\n\nThe practical takeaway for operators is patience: infrastructure refactors like this are rarely visible as single features, but they make the visible features less brittle. Session reads becoming worker-backed is one more step toward a Gateway that can keep serving real-time chats while background maintenance, status views, plugins, and memory work happen around it.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/169622\">PR #169622</a>, a large Gateway refactor that moves bundled session reads and authority preparation through async workers instead of ordinary synchronous SQLite paths.</p><p>This is not flashy in the product-marketing sense. It is the kind of plumbing change that matters when an agent runtime has to keep chats, plugins, memory, cron, channels, and UI views responsive at the same time.</p><h2>The Problem: Ordinary Reads Still Hit The Owner Path</h2><p>The PR says ordinary session reads in Gateway routing, agent execution, sandbox policy, plugin operations, Memory SDK consumers, and status rendering still reached synchronous SQLite through owner and SDK APIs.</p><p>In a small local setup, that may be invisible. In a busy Gateway, those read paths become part of the latency budget for every turn. They also make it harder to reason about ownership, because a read that looks descriptive can still cross the same surface as effect-bearing state.</p><p>The merged change routes bundled callers through existing workers or prepared facts updated by committed write receipts. The Gateway remains the single database owner, but more consumers await the worker-backed view instead of touching synchronous session storage directly.</p><h2>The Boundary Did Not Disappear</h2><p>The interesting part is what the PR explicitly keeps. Final authority checks remain at execution, mutation, disclosure, approvals, credentials, and external publication. The change is not \"trust cached state forever.\" It is \"carry the prepared facts through the owning operation, then still check authority where real effects happen.\"</p><p>That distinction matters for OpenClaw's security model. Faster session reads are useful only if they do not let stale membership, revoked access, or changed policy leak through to a tool call, credential access, publication, or channel send.</p><p>The PR also keeps four named effect-boundary synchronous checks for cases where current state is the point of the check: Discord voice steering, ClickClack disclosure and retention, and external channel creation or publication.</p><h2>What Users Should Notice</h2><p>Most users will not see a new command or setting from this merge. The expected user-facing effect is steadier responsiveness under load and fewer hidden stalls when multiple parts of OpenClaw are inspecting session state.</p><p>The PR's live measurement is careful about scope, but the direction is still notable. Its warmed shared-harness comparison reported lower all-thread SQL statements and executing requests per turn, with MAIN-thread SQL at zero in the candidate path.</p><p>That is the arc OpenClaw has been following for months: move work off the hot Gateway path, keep ownership explicit, and preserve security-sensitive checks at the moment an effect is about to happen.</p><h2>Why It Belongs In The Release Story</h2><p>This merge landed shortly after 2026.10.5-beta.2, and it fits the same theme as the beta notes. Recovery, session routing, tool approvals, memory, and channel delivery are all converging on a more explicit authority model.</p><p>The practical takeaway for operators is patience: infrastructure refactors like this are rarely visible as single features, but they make the visible features less brittle. Session reads becoming worker-backed is one more step toward a Gateway that can keep serving real-time chats while background maintenance, status views, plugins, and memory work happen around it.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-worker-backed-session-reads.png\" alt=\"OpenClaw Moves Session Reads Into Workers\" /></p>",
      "date_published": "2026-10-12T23:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Memory",
        "Moves",
        "Session",
        "Reads",
        "Workers"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-worker-backed-session-reads.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-worker-backed-session-reads.png",
          "mime_type": "image/png",
          "title": "OpenClaw Moves Session Reads Into Workers cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-12-beta-2026-10-5-beta-2-release/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-12-beta-2026-10-5-beta-2-release/",
      "title": "OpenClaw 2026.10.5 Beta 2 Tightens Recovery",
      "summary": "OpenClaw 2026.10.5 beta 2 lands with safer updates, stronger Doctor recovery, tool approval hardening, and steadier Control UI turns.",
      "content_text": "OpenClaw published [v2026.10.5-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.10.5-beta.2) today, and it is a dense beta for people running long-lived Gateways, paired devices, plugins, and unattended update paths.\n\nThe release follows this morning's stable 2026.10.2 coverage, but it is not just another tag on the same work. The beta channel is collecting several boundary fixes around updates, Doctor recovery, tool approvals, subagent provenance, and Control UI durability.\n\n## Updates Keep More State Intact\n\nThe headline is update recovery. The release notes say `openclaw update` now settles stale handoff leases, completed receipts, and older recovery records instead of stopping on `update-recovery-pending`.\n\nThat matters because OpenClaw updates are not a simple binary swap. A Gateway may have webhook listeners, service heap limits, device state, plugin metadata, and running conversations in play. A conservative updater has to avoid making a bad state worse, but it also has to avoid stranding a healthy install because of stale bookkeeping.\n\nThis beta aims at that middle path. The notes call out Windows short-name path handling, preservation of service heap limits and webhook listeners, and fewer cases where accepted handoffs or successful updates are reported as failures.\n\n## Doctor Gets More Careful With Data\n\nDoctor also gets a meaningful recovery pass. The release notes describe preservation of installed plugin configuration when path-safety checks block a plugin, plus a stricter stance toward malformed UTF-8 root memory files: leave them alone rather than importing uncertain bytes into a repaired state.\n\nRecovery copies now capture Canvas and Memory Core files before migrations move them, including installs without configured agents. That detail is easy to skim past, but it is important operationally. Recovery tools should preserve evidence before they normalize state, especially when the broken install may be the only copy of a user's data.\n\n## Tool Approvals Get A Sharper Boundary\n\nThe security-sensitive item in the beta is tool approval hardening. The release says hooks can no longer change tool arguments after a trusted inline approval without requiring another approval, and transforming plugins must request approval for their final arguments.\n\nThe practical rule is simple: the thing the user approved should be the thing that runs. If a hook or plugin changes arguments after approval, that is a new execution shape and it needs its own approval boundary.\n\nThe migration target is documented in [`docs/plugins/hooks/tool-policy.md`](https://github.com/openclaw/openclaw/blob/main/docs/plugins/hooks/tool-policy.md), which is the right place for plugin authors to start before this behavior moves closer to stable.\n\n## Subagents And Control UI Become Less Fragile\n\nThe beta also improves subagent message provenance. Inter-session messages retain the `[Inter-session message]` safety envelope across initial delivery and replay, so parent-sent instructions are not later presented to a child session as direct user instructions.\n\nControl UI turns also keep running through ordinary pairing activity such as node heartbeats, other devices updating, and tab reconnects. Real revocations still stop affected work. That distinction is the whole point: routine device churn should not look like a security revocation, while actual revocation still needs to win.\n\n## Why This Beta Matters\n\nOpenClaw's recent release train has been about reducing weird failure modes at boundaries: update handoffs, device pairing, plugin transforms, memory repair, and inter-session messaging. Beta 2 continues that theme with a bias toward data preservation and explicit authority.\n\nFor most stable users, 2026.10.2 remains the safer channel. For beta testers and operators tracking the next recovery/security bundle, 2026.10.5-beta.2 is the tag to read closely.",
      "content_html": "<p>OpenClaw published <a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.10.5-beta.2\">v2026.10.5-beta.2</a> today, and it is a dense beta for people running long-lived Gateways, paired devices, plugins, and unattended update paths.</p><p>The release follows this morning's stable 2026.10.2 coverage, but it is not just another tag on the same work. The beta channel is collecting several boundary fixes around updates, Doctor recovery, tool approvals, subagent provenance, and Control UI durability.</p><h2>Updates Keep More State Intact</h2><p>The headline is update recovery. The release notes say <code>openclaw update</code> now settles stale handoff leases, completed receipts, and older recovery records instead of stopping on <code>update-recovery-pending</code>.</p><p>That matters because OpenClaw updates are not a simple binary swap. A Gateway may have webhook listeners, service heap limits, device state, plugin metadata, and running conversations in play. A conservative updater has to avoid making a bad state worse, but it also has to avoid stranding a healthy install because of stale bookkeeping.</p><p>This beta aims at that middle path. The notes call out Windows short-name path handling, preservation of service heap limits and webhook listeners, and fewer cases where accepted handoffs or successful updates are reported as failures.</p><h2>Doctor Gets More Careful With Data</h2><p>Doctor also gets a meaningful recovery pass. The release notes describe preservation of installed plugin configuration when path-safety checks block a plugin, plus a stricter stance toward malformed UTF-8 root memory files: leave them alone rather than importing uncertain bytes into a repaired state.</p><p>Recovery copies now capture Canvas and Memory Core files before migrations move them, including installs without configured agents. That detail is easy to skim past, but it is important operationally. Recovery tools should preserve evidence before they normalize state, especially when the broken install may be the only copy of a user's data.</p><h2>Tool Approvals Get A Sharper Boundary</h2><p>The security-sensitive item in the beta is tool approval hardening. The release says hooks can no longer change tool arguments after a trusted inline approval without requiring another approval, and transforming plugins must request approval for their final arguments.</p><p>The practical rule is simple: the thing the user approved should be the thing that runs. If a hook or plugin changes arguments after approval, that is a new execution shape and it needs its own approval boundary.</p><p>The migration target is documented in <a href=\"https://github.com/openclaw/openclaw/blob/main/docs/plugins/hooks/tool-policy.md\"><code>docs/plugins/hooks/tool-policy.md</code></a>, which is the right place for plugin authors to start before this behavior moves closer to stable.</p><h2>Subagents And Control UI Become Less Fragile</h2><p>The beta also improves subagent message provenance. Inter-session messages retain the <code>[Inter-session message]</code> safety envelope across initial delivery and replay, so parent-sent instructions are not later presented to a child session as direct user instructions.</p><p>Control UI turns also keep running through ordinary pairing activity such as node heartbeats, other devices updating, and tab reconnects. Real revocations still stop affected work. That distinction is the whole point: routine device churn should not look like a security revocation, while actual revocation still needs to win.</p><h2>Why This Beta Matters</h2><p>OpenClaw's recent release train has been about reducing weird failure modes at boundaries: update handoffs, device pairing, plugin transforms, memory repair, and inter-session messaging. Beta 2 continues that theme with a bias toward data preservation and explicit authority.</p><p>For most stable users, 2026.10.2 remains the safer channel. For beta testers and operators tracking the next recovery/security bundle, 2026.10.5-beta.2 is the tag to read closely.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-beta-2026-10-5-beta-2-release.png\" alt=\"OpenClaw 2026.10.5 Beta 2 Tightens Recovery\" /></p>",
      "date_published": "2026-10-12T23:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "2026",
        "Beta",
        "Tightens",
        "Recovery"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-beta-2026-10-5-beta-2-release.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-beta-2026-10-5-beta-2-release.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.10.5 Beta 2 Tightens Recovery cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-12-public-link-preview-cards/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-12-public-link-preview-cards/",
      "title": "OpenClaw Public Links Get Session-Specific Cards",
      "summary": "OpenClaw PR #169527 adds session-specific public link preview cards with bounded redaction, work facts, revocation fallbacks, and cached PNG rendering.",
      "content_text": "[PR #169527](https://github.com/openclaw/openclaw/pull/169527) gives OpenClaw public session links a more useful first impression: instead of every shared conversation unfurling with the same generic card, public sessions can now render preview images specific to the shared content.\n\nThe change is small in concept but broad in execution. Link previews are one of those surfaces that look like polish until you remember where they appear: Slack, Discord, messaging apps, social feeds, issue comments, docs, and internal team threads. A meaningful card can tell readers what they are about to open before they click.\n\n## What the Cards Show\n\nPublic sessions now get a 1200 by 630 preview with the session title, opening public request, and available session facts. If a worktree already has recorded change counts, the card can become a work receipt with branch information, diff totals, and recorded PR or check state.\n\nPrivate, unknown, revoked, and unavailable publications keep the generic card. That fallback is important because unfurls are copied, cached, and displayed in places OpenClaw does not control. The PR notes that successful previews may be cached for five minutes, and revocation cannot recall copies a recipient has already downloaded.\n\n## Redaction Stays at the Center\n\nThe preview route uses the existing opaque share-token contract and shares the public page's admission, secure-ingress, current-publication checks, redaction, and bounded representation cache.\n\nThe content boundary is intentionally narrow. Cards expose only public conversation text and permitted display or recorded work facts. They omit owner identities, email addresses, costs, model names, internal paths, and unknown fields. Cold worktree snapshots do not trigger fresh Git or GitHub reads just to render an unfurl.\n\nThat matters because preview cards are easy to underestimate. They are generated automatically, often by third-party clients, and can spread outside the original browser session. The right default is to make the preview helpful while treating it as a public artifact.\n\n## Rendering Moves Off the Gateway Thread\n\nThe implementation adds a pinned runtime dependency, `@resvg/resvg-js@2.6.2`, for font-aware SVG rasterization. Rasterization and font measurement run in a lifecycle-owned worker, so the Gateway boot path does not load the renderer.\n\nThe PR also bundles byte-verified OFL font files and license notices from Google Fonts sources, with maintainer-approved provenance review recorded separately. That is a lot of detail for preview cards, but it explains why the rendered output can be deterministic without pulling rendering work into the hot Gateway path.\n\n## The Safety Tests Are the Story\n\nThe proof section focuses on boundary behavior. A real loopback HTTP test covers document and PNG delivery, caching, private and revoked fallbacks, and a paused-render revocation case. Public reader tests cover oversized transcript rows and private or assistant preludes, while renderer tests check escaping, length caps, line placement, fallback behavior, metadata, and cache invalidation.\n\nThis is a user-facing feature, but the interesting part is the restraint. OpenClaw public links now get richer cards, yet the implementation keeps the preview tied to the same admission and redaction model as the public page. That is the right shape for sharing agent work in places where a link preview may travel farther than the original conversation.",
      "content_html": "<p><a href=\"https://github.com/openclaw/openclaw/pull/169527\">PR #169527</a> gives OpenClaw public session links a more useful first impression: instead of every shared conversation unfurling with the same generic card, public sessions can now render preview images specific to the shared content.</p><p>The change is small in concept but broad in execution. Link previews are one of those surfaces that look like polish until you remember where they appear: Slack, Discord, messaging apps, social feeds, issue comments, docs, and internal team threads. A meaningful card can tell readers what they are about to open before they click.</p><h2>What the Cards Show</h2><p>Public sessions now get a 1200 by 630 preview with the session title, opening public request, and available session facts. If a worktree already has recorded change counts, the card can become a work receipt with branch information, diff totals, and recorded PR or check state.</p><p>Private, unknown, revoked, and unavailable publications keep the generic card. That fallback is important because unfurls are copied, cached, and displayed in places OpenClaw does not control. The PR notes that successful previews may be cached for five minutes, and revocation cannot recall copies a recipient has already downloaded.</p><h2>Redaction Stays at the Center</h2><p>The preview route uses the existing opaque share-token contract and shares the public page's admission, secure-ingress, current-publication checks, redaction, and bounded representation cache.</p><p>The content boundary is intentionally narrow. Cards expose only public conversation text and permitted display or recorded work facts. They omit owner identities, email addresses, costs, model names, internal paths, and unknown fields. Cold worktree snapshots do not trigger fresh Git or GitHub reads just to render an unfurl.</p><p>That matters because preview cards are easy to underestimate. They are generated automatically, often by third-party clients, and can spread outside the original browser session. The right default is to make the preview helpful while treating it as a public artifact.</p><h2>Rendering Moves Off the Gateway Thread</h2><p>The implementation adds a pinned runtime dependency, <code>@resvg/resvg-js@2.6.2</code>, for font-aware SVG rasterization. Rasterization and font measurement run in a lifecycle-owned worker, so the Gateway boot path does not load the renderer.</p><p>The PR also bundles byte-verified OFL font files and license notices from Google Fonts sources, with maintainer-approved provenance review recorded separately. That is a lot of detail for preview cards, but it explains why the rendered output can be deterministic without pulling rendering work into the hot Gateway path.</p><h2>The Safety Tests Are the Story</h2><p>The proof section focuses on boundary behavior. A real loopback HTTP test covers document and PNG delivery, caching, private and revoked fallbacks, and a paused-render revocation case. Public reader tests cover oversized transcript rows and private or assistant preludes, while renderer tests check escaping, length caps, line placement, fallback behavior, metadata, and cache invalidation.</p><p>This is a user-facing feature, but the interesting part is the restraint. OpenClaw public links now get richer cards, yet the implementation keeps the preview tied to the same admission and redaction model as the public page. That is the right shape for sharing agent work in places where a link preview may travel farther than the original conversation.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-public-link-preview-cards.png\" alt=\"OpenClaw Public Links Get Session-Specific Cards\" /></p>",
      "date_published": "2026-10-12T08:06:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "Public",
        "Links",
        "Session",
        "Specific",
        "Cards"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-public-link-preview-cards.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-public-link-preview-cards.png",
          "mime_type": "image/png",
          "title": "OpenClaw Public Links Get Session-Specific Cards cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-12-node-pairing-fix/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-12-node-pairing-fix/",
      "title": "OpenClaw Node Pairing Fix Keeps Devices Flowing",
      "summary": "OpenClaw PR #169575 fixes node requests and Gateway-to-node events that could fail while another device pairing record was being written in Gateway state.",
      "content_text": "[PR #169575](https://github.com/openclaw/openclaw/pull/169575) fixes a subtle but high-impact node reliability problem in OpenClaw: one device's pairing write could interrupt active node requests or Gateway-to-node events for another device.\n\nThe merged PR describes failures such as `node pairing changed before request dispatch` and `node pairing state unavailable before dispatch`. Those errors could appear during routine pairing writes, including token verification, reconnect bookkeeping, presence updates, metadata changes, approvals, and revocations.\n\n## The User-Facing Symptom\n\nThe practical version is simple: connected nodes could stop receiving requests or silently drop events while a separate device was updating pairing state.\n\nThat is a nasty failure mode because the affected node is not necessarily the one being changed. A routine presence or reconnect write for device B could make device A's request path temporarily treat pairing authority as unknown. For anyone depending on OpenClaw nodes for distributed work, that is exactly the sort of edge case that feels random from the outside.\n\nThe fix keeps node requests and events flowing while other devices connect, reconnect, verify tokens, or update pairing metadata. At the same time, the PR preserves the important security behavior: revoking or re-approving a node still retires that node's live connection as before.\n\n## Why the Fence Came Down\n\nThe old implementation placed the in-memory pairing publication into an authority-unknown state for most pairing writes while the worker round trip completed. The PR argues that the fence was broader than necessary because stored rows remain authoritative until the database commit, and the first binding read after commit installs the worker's receipt.\n\nSo the change removes the broad `invalidatesAuthority` distinction. Failed-read blocking, unknown-settlement blocking, epoch checks, incomplete-row checks, receipt installation, and live transport retirement on revocation remain intact.\n\nThat is the key engineering line: the patch reduces false uncertainty without weakening the paths that actually protect authority changes.\n\n## The Tradeoff Is Explicit\n\nThe PR is unusually clear about the race it accepts. A Gateway-to-node send that lands in the tiny interval between commit and receipt post can use the pre-commit binding. The author frames that as linearizing before the write, because the caller has not yet been told the write succeeded, and revocation still retires the transport immediately after.\n\nIf the worker dies after commit but before posting its receipt, old bindings remain visible until the operation settles as unknown, at which point publication blocks until a fresh read. That is a crash-recovery shape rather than a silent permanent mismatch.\n\n## Proof Coverage\n\nThe evidence is broad for a PR of this size. A Gateway transport proof uses real node WebSocket clients and holds a presence write for one node while another node receives and completes `node.invoke` work. The same harness covers rejection behavior for token revoke, token rotate, unpair, and reapproval.\n\nThe infra regression suite also holds writes for token verification, connection records, presence, metadata, and token revocation, then checks that unrelated device authority remains readable. On the old code, those cases failed because the unrelated device could not get a current publication.\n\nThis is a strong operational fix: it keeps OpenClaw's distributed node system usable during ordinary churn, while leaving revocation and reapproval boundaries in place.",
      "content_html": "<p><a href=\"https://github.com/openclaw/openclaw/pull/169575\">PR #169575</a> fixes a subtle but high-impact node reliability problem in OpenClaw: one device's pairing write could interrupt active node requests or Gateway-to-node events for another device.</p><p>The merged PR describes failures such as <code>node pairing changed before request dispatch</code> and <code>node pairing state unavailable before dispatch</code>. Those errors could appear during routine pairing writes, including token verification, reconnect bookkeeping, presence updates, metadata changes, approvals, and revocations.</p><h2>The User-Facing Symptom</h2><p>The practical version is simple: connected nodes could stop receiving requests or silently drop events while a separate device was updating pairing state.</p><p>That is a nasty failure mode because the affected node is not necessarily the one being changed. A routine presence or reconnect write for device B could make device A's request path temporarily treat pairing authority as unknown. For anyone depending on OpenClaw nodes for distributed work, that is exactly the sort of edge case that feels random from the outside.</p><p>The fix keeps node requests and events flowing while other devices connect, reconnect, verify tokens, or update pairing metadata. At the same time, the PR preserves the important security behavior: revoking or re-approving a node still retires that node's live connection as before.</p><h2>Why the Fence Came Down</h2><p>The old implementation placed the in-memory pairing publication into an authority-unknown state for most pairing writes while the worker round trip completed. The PR argues that the fence was broader than necessary because stored rows remain authoritative until the database commit, and the first binding read after commit installs the worker's receipt.</p><p>So the change removes the broad <code>invalidatesAuthority</code> distinction. Failed-read blocking, unknown-settlement blocking, epoch checks, incomplete-row checks, receipt installation, and live transport retirement on revocation remain intact.</p><p>That is the key engineering line: the patch reduces false uncertainty without weakening the paths that actually protect authority changes.</p><h2>The Tradeoff Is Explicit</h2><p>The PR is unusually clear about the race it accepts. A Gateway-to-node send that lands in the tiny interval between commit and receipt post can use the pre-commit binding. The author frames that as linearizing before the write, because the caller has not yet been told the write succeeded, and revocation still retires the transport immediately after.</p><p>If the worker dies after commit but before posting its receipt, old bindings remain visible until the operation settles as unknown, at which point publication blocks until a fresh read. That is a crash-recovery shape rather than a silent permanent mismatch.</p><h2>Proof Coverage</h2><p>The evidence is broad for a PR of this size. A Gateway transport proof uses real node WebSocket clients and holds a presence write for one node while another node receives and completes <code>node.invoke</code> work. The same harness covers rejection behavior for token revoke, token rotate, unpair, and reapproval.</p><p>The infra regression suite also holds writes for token verification, connection records, presence, metadata, and token revocation, then checks that unrelated device authority remains readable. On the old code, those cases failed because the unrelated device could not get a current publication.</p><p>This is a strong operational fix: it keeps OpenClaw's distributed node system usable during ordinary churn, while leaving revocation and reapproval boundaries in place.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-node-pairing-fix.png\" alt=\"OpenClaw Node Pairing Fix Keeps Devices Flowing\" /></p>",
      "date_published": "2026-10-12T08:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Memory",
        "Node",
        "Pairing",
        "Keeps",
        "Devices",
        "Flowing"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-node-pairing-fix.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-node-pairing-fix.png",
          "mime_type": "image/png",
          "title": "OpenClaw Node Pairing Fix Keeps Devices Flowing cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-12-release-update-recovery/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-12-release-update-recovery/",
      "title": "OpenClaw 2026.10.2 Sharpens Updates and Recovery",
      "summary": "OpenClaw 2026.10.2 focuses on cleaner upgrades, Doctor recovery, provider-native compaction, secret redaction, and Gateway responsiveness for operators.",
      "content_text": "[OpenClaw 2026.10.2](https://github.com/openclaw/openclaw/releases/tag/v2026.10.2) landed this morning with a practical theme: make upgrades finish cleanly, make recovery less brittle, and keep long-running agent systems responsive under real-world pressure.\n\nThis is a stable release, not just a small patch note. The official changelog says the audited record covers 85 pull requests since `v2026.10.1`, and the highlighted areas are exactly the sort of maintenance work that matters when OpenClaw is already living inside daily workflows.\n\n## Update Paths Get Calmer\n\nThe headline fix is update reliability. The release notes call out the first upgrade from stable `2026.10.1`, successful in-place updates that previously ended in errors, and Windows, cloned-filesystem, and private-Node paths that could stall or roll back.\n\nThat means the release is less about a shiny new button and more about reducing the anxiety around keeping a Gateway current. If an update has ever left a managed service stopped, returned a confusing finalization error, or required manual cleanup, this is the cluster of fixes to watch.\n\nThe release also changes how finalized update success and suppressed channels are reported. They now surface as warnings instead of failures, which should reduce false alarms without hiding useful diagnostics.\n\n## Doctor Recovery Gets Broader\n\nDoctor also receives a wide set of recovery hardening. The changelog names residual legacy imports, stale Gateway locks, slow state inspection, memory reindex locks, removed-channel allowlists, and stale-receipt sessions as areas that should no longer stop recovery cold.\n\nThat is important because Doctor sits in the awkward but essential space between product and repair tool. It needs to be strict enough to protect state, but tolerant enough to move past old receipts, lock leftovers, and partially migrated systems. `2026.10.2` keeps pushing that balance toward fewer manual dead ends.\n\n## Provider-Native Compaction Becomes Default\n\nOpenAI and Anthropic API-key routes now default to provider-native compaction. The notes also say compaction can recover when the primary auth profile is cooling down.\n\nFor users, the expected benefit is more stable long-running sessions and fewer failures when context management crosses account, model, or provider boundaries. It is also a cleaner default: when the provider has native compaction behavior, OpenClaw can lean into it instead of carrying every path through the same fallback shape.\n\n## Redaction and Cache Stability Improve\n\nThe security-related highlight is secret redaction. Credentials that cross a chunk boundary in long tool output are now masked, and large-text redaction should avoid stalls and stack overflow behavior while running faster.\n\nThe release also spends real attention on Anthropic prompt-cache stability. Conversation checkpoints, enriched user prompts, tool results, and inter-session provenance are intended to stay byte-identical across follow-ups, replays, Gateway restarts, and subagent completions. In practice, that means cached prefixes should continue hitting instead of being invalidated by harmless representation drift.\n\n## Gateway Responsiveness Is a Theme\n\nThe final major bucket is Gateway responsiveness. The release lists fixes for Control UI reconnects, model changes with large plugins, dashboard title work, and catalog loading that could freeze the Gateway or hang new sessions.\n\nTaken together, `2026.10.2` is the kind of release that makes OpenClaw feel less dramatic in production. The work is mostly invisible when it succeeds: updates complete, recovery keeps moving, secrets stay masked, sessions keep context, and the Gateway does not stall while background systems do their jobs.\n\nFor operators, that is the story: upgrade for the boring wins.",
      "content_html": "<p><a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.10.2\">OpenClaw 2026.10.2</a> landed this morning with a practical theme: make upgrades finish cleanly, make recovery less brittle, and keep long-running agent systems responsive under real-world pressure.</p><p>This is a stable release, not just a small patch note. The official changelog says the audited record covers 85 pull requests since <code>v2026.10.1</code>, and the highlighted areas are exactly the sort of maintenance work that matters when OpenClaw is already living inside daily workflows.</p><h2>Update Paths Get Calmer</h2><p>The headline fix is update reliability. The release notes call out the first upgrade from stable <code>2026.10.1</code>, successful in-place updates that previously ended in errors, and Windows, cloned-filesystem, and private-Node paths that could stall or roll back.</p><p>That means the release is less about a shiny new button and more about reducing the anxiety around keeping a Gateway current. If an update has ever left a managed service stopped, returned a confusing finalization error, or required manual cleanup, this is the cluster of fixes to watch.</p><p>The release also changes how finalized update success and suppressed channels are reported. They now surface as warnings instead of failures, which should reduce false alarms without hiding useful diagnostics.</p><h2>Doctor Recovery Gets Broader</h2><p>Doctor also receives a wide set of recovery hardening. The changelog names residual legacy imports, stale Gateway locks, slow state inspection, memory reindex locks, removed-channel allowlists, and stale-receipt sessions as areas that should no longer stop recovery cold.</p><p>That is important because Doctor sits in the awkward but essential space between product and repair tool. It needs to be strict enough to protect state, but tolerant enough to move past old receipts, lock leftovers, and partially migrated systems. <code>2026.10.2</code> keeps pushing that balance toward fewer manual dead ends.</p><h2>Provider-Native Compaction Becomes Default</h2><p>OpenAI and Anthropic API-key routes now default to provider-native compaction. The notes also say compaction can recover when the primary auth profile is cooling down.</p><p>For users, the expected benefit is more stable long-running sessions and fewer failures when context management crosses account, model, or provider boundaries. It is also a cleaner default: when the provider has native compaction behavior, OpenClaw can lean into it instead of carrying every path through the same fallback shape.</p><h2>Redaction and Cache Stability Improve</h2><p>The security-related highlight is secret redaction. Credentials that cross a chunk boundary in long tool output are now masked, and large-text redaction should avoid stalls and stack overflow behavior while running faster.</p><p>The release also spends real attention on Anthropic prompt-cache stability. Conversation checkpoints, enriched user prompts, tool results, and inter-session provenance are intended to stay byte-identical across follow-ups, replays, Gateway restarts, and subagent completions. In practice, that means cached prefixes should continue hitting instead of being invalidated by harmless representation drift.</p><h2>Gateway Responsiveness Is a Theme</h2><p>The final major bucket is Gateway responsiveness. The release lists fixes for Control UI reconnects, model changes with large plugins, dashboard title work, and catalog loading that could freeze the Gateway or hang new sessions.</p><p>Taken together, <code>2026.10.2</code> is the kind of release that makes OpenClaw feel less dramatic in production. The work is mostly invisible when it succeeds: updates complete, recovery keeps moving, secrets stay masked, sessions keep context, and the Gateway does not stall while background systems do their jobs.</p><p>For operators, that is the story: upgrade for the boring wins.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-release-update-recovery.png\" alt=\"OpenClaw 2026.10.2 Sharpens Updates and Recovery\" /></p>",
      "date_published": "2026-10-12T08:00:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Memory",
        "2026",
        "Sharpens",
        "Updates",
        "Recovery"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-release-update-recovery.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-12-release-update-recovery.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.10.2 Sharpens Updates and Recovery cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-11-control-ui-avatar-misses/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-11-control-ui-avatar-misses/",
      "title": "OpenClaw Control UI Stops Repeated Avatar Misses",
      "summary": "OpenClaw PR #168640 consolidates profile avatar availability so missing user and workspace images stop causing repeated failed requests on reload.",
      "content_text": "OpenClaw merged [PR #168640](https://github.com/openclaw/openclaw/pull/168640), a Control UI performance fix for a small but noisy problem: repeated missing avatar and workspace icon requests on signed-in dashboard reloads.\n\nThe PR introduces a shared `ProfileDirectory` owner for user profile data in the Control UI. Instead of each consumer guessing whether an avatar might exist, the directory supplies avatar availability to mentions, owner chips, menus, and author avatars.\n\n## The Problem\n\nThe PR describes a signed-in dashboard deep-link reload that repeatedly requested missing avatars and icons. A synthetic reload measurement recorded about ten user-avatar 404s plus an agent-avatar and workspace-icon 404 per reload.\n\nThose misses are not catastrophic, but they add avoidable noise. They also make the UI do work for images the app already knows are absent.\n\nThe tricky part is preserving real profile photos. The UI needs to avoid speculative misses without hiding actual user images when profile data arrives later.\n\n## The Fix\n\n`ProfileDirectory` becomes the Gateway-scoped owner of `users.list` in the Control UI. It coalesces reads, follows canonical profile merges, retires stale scopes, and refreshes person cards when opened.\n\nProfiles with uploaded avatars render through versioned avatar URLs. Profiles without avatars render initials and do not issue speculative image requests. Definite misses are cached privately for a short window, while stale or unversioned requests and transient failures are not cached.\n\nThe PR also keeps agent identity and workspace icon behavior scoped: hydrated agent identity owns avatar absence, and workspace misses are cached only after resolving the workspace root.\n\n## User Impact\n\nFor users, this should mostly feel like absence of friction. Menus, mentions, owner chips, and author avatars still show the right identity state, but reloads should stop hammering missing image URLs.\n\nThe PR body includes screenshot-backed proof for mention behavior: no image requests while availability is unknown, one successful versioned photo request when a photo is advertised, and no requests for absent avatars. It also verifies that chip geometry stays stable across desktop and mobile-like widths.\n\nThat last detail matters. A performance fix that shifts labels or chips around would trade one polish issue for another. This one is scoped to request behavior and profile-directory ownership.\n\n## Why It Matters\n\nThis is not a headline feature, but it is a good example of Control UI maturation. The app is increasingly treating profile, presence, and avatar data as shared state with a real owner rather than as repeated local fetch guesses.\n\nThe immediate benefit is fewer failed requests. The longer-term benefit is a cleaner identity layer for the web UI: one directory, clearer availability signals, and less duplicated profile plumbing across Lit and Solid consumers.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168640\">PR #168640</a>, a Control UI performance fix for a small but noisy problem: repeated missing avatar and workspace icon requests on signed-in dashboard reloads.</p><p>The PR introduces a shared <code>ProfileDirectory</code> owner for user profile data in the Control UI. Instead of each consumer guessing whether an avatar might exist, the directory supplies avatar availability to mentions, owner chips, menus, and author avatars.</p><h2>The Problem</h2><p>The PR describes a signed-in dashboard deep-link reload that repeatedly requested missing avatars and icons. A synthetic reload measurement recorded about ten user-avatar 404s plus an agent-avatar and workspace-icon 404 per reload.</p><p>Those misses are not catastrophic, but they add avoidable noise. They also make the UI do work for images the app already knows are absent.</p><p>The tricky part is preserving real profile photos. The UI needs to avoid speculative misses without hiding actual user images when profile data arrives later.</p><h2>The Fix</h2><p><code>ProfileDirectory</code> becomes the Gateway-scoped owner of <code>users.list</code> in the Control UI. It coalesces reads, follows canonical profile merges, retires stale scopes, and refreshes person cards when opened.</p><p>Profiles with uploaded avatars render through versioned avatar URLs. Profiles without avatars render initials and do not issue speculative image requests. Definite misses are cached privately for a short window, while stale or unversioned requests and transient failures are not cached.</p><p>The PR also keeps agent identity and workspace icon behavior scoped: hydrated agent identity owns avatar absence, and workspace misses are cached only after resolving the workspace root.</p><h2>User Impact</h2><p>For users, this should mostly feel like absence of friction. Menus, mentions, owner chips, and author avatars still show the right identity state, but reloads should stop hammering missing image URLs.</p><p>The PR body includes screenshot-backed proof for mention behavior: no image requests while availability is unknown, one successful versioned photo request when a photo is advertised, and no requests for absent avatars. It also verifies that chip geometry stays stable across desktop and mobile-like widths.</p><p>That last detail matters. A performance fix that shifts labels or chips around would trade one polish issue for another. This one is scoped to request behavior and profile-directory ownership.</p><h2>Why It Matters</h2><p>This is not a headline feature, but it is a good example of Control UI maturation. The app is increasingly treating profile, presence, and avatar data as shared state with a real owner rather than as repeated local fetch guesses.</p><p>The immediate benefit is fewer failed requests. The longer-term benefit is a cleaner identity layer for the web UI: one directory, clearer availability signals, and less duplicated profile plumbing across Lit and Solid consumers.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-control-ui-avatar-misses.png\" alt=\"OpenClaw Control UI Stops Repeated Avatar Misses\" /></p>",
      "date_published": "2026-10-11T23:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Control",
        "Stops",
        "Repeated",
        "Avatar",
        "Misses"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-control-ui-avatar-misses.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-control-ui-avatar-misses.png",
          "mime_type": "image/png",
          "title": "OpenClaw Control UI Stops Repeated Avatar Misses cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-11-async-auth-credential-prep/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-11-async-auth-credential-prep/",
      "title": "OpenClaw Async Auth Moves Credential Reads Off Gateway",
      "summary": "OpenClaw PR #169111 moves credential preparation and provider capability selection to worker-backed async auth reads while preserving final authority checks.",
      "content_text": "OpenClaw merged [PR #169111](https://github.com/openclaw/openclaw/pull/169111), a large auth refactor that moves credential preparation and provider capability selection away from synchronous SQLite-backed Gateway helpers and onto worker-backed async auth reads.\n\nThat sounds internal, but it touches a sensitive part of the runtime: model, media, speech, realtime voice, TTS, setup, and provider credential preparation. The PR keeps existing Plugin SDK synchronous contracts available and deprecated, while adding documented async replacements for plugin authors.\n\n## What Changed\n\nThe PR uses the existing auth owner for worker reads and atomic writes, then propagates async preparation through bundled OpenClaw callers. Provider selection still respects explicit, profile, and environment precedence. The canonical loader keeps its existing canonical-versus-overlay behavior.\n\nSeveral paths now prepare credentials once, await the owner, and avoid reloading the same auth store during setup. The PR body calls out examples across native web search, xAI fallback, scoped OAuth refresh, media model assembly, prompt snapshot generation, and Discord voice.\n\nThe result is not \"auth is now async everywhere.\" Final credential, execution, disclosure, deletion, and effect-boundary checks remain at the point where the action actually happens. That distinction matters: async preparation reduces blocking work, while final authority checks still guard the real effect.\n\n## Why It Matters\n\nThe Gateway has been shedding synchronous reads across the session, auth, and storage stack. Credential preparation is a natural but risky place to continue that work because it sits near both performance and security boundaries.\n\nFor users, the expected benefit is steadier preparation for provider-heavy flows. For plugin authors, the important part is the migration path: synchronous SDK APIs are deprecated instead of silently changing behavior, and async peers are documented in the plugin migration docs.\n\nFor maintainers, the deeper win is ownership. Instead of scattered callers taking read snapshots and writing them back, the auth owner owns atomic mutations and worker-backed reads. That reduces duplicated read-modify-write shapes and makes future authority boundaries easier to reason about.\n\n## Guardrails Stay In Place\n\nThe PR is explicit about remaining boundaries. It does not change schema, retention policy, permission policy, or durability policy. It also names follow-up work: config health caller-chain migration, async placement initialization, live verbose callbacks, published settings, and physical-target auth batch rollback.\n\nThat kind of disclosure is useful. It keeps the story from becoming \"all sync auth is gone\" when the real progress is more precise: ordinary preparation reads are moving off the Gateway thread, while compatibility APIs and final authority reads remain where they are needed.\n\n## The Bigger Pattern\n\nTaken together with today’s release and other session-worker PRs, #169111 shows OpenClaw continuing a broad runtime cleanup: move ordinary preparation into owners and workers, preserve real effect checks, and give plugin authors migration paths instead of abrupt SDK breaks.\n\nThat is not flashy, but it is exactly the kind of work that makes agent infrastructure less brittle at scale.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/169111\">PR #169111</a>, a large auth refactor that moves credential preparation and provider capability selection away from synchronous SQLite-backed Gateway helpers and onto worker-backed async auth reads.</p><p>That sounds internal, but it touches a sensitive part of the runtime: model, media, speech, realtime voice, TTS, setup, and provider credential preparation. The PR keeps existing Plugin SDK synchronous contracts available and deprecated, while adding documented async replacements for plugin authors.</p><h2>What Changed</h2><p>The PR uses the existing auth owner for worker reads and atomic writes, then propagates async preparation through bundled OpenClaw callers. Provider selection still respects explicit, profile, and environment precedence. The canonical loader keeps its existing canonical-versus-overlay behavior.</p><p>Several paths now prepare credentials once, await the owner, and avoid reloading the same auth store during setup. The PR body calls out examples across native web search, xAI fallback, scoped OAuth refresh, media model assembly, prompt snapshot generation, and Discord voice.</p><p>The result is not \"auth is now async everywhere.\" Final credential, execution, disclosure, deletion, and effect-boundary checks remain at the point where the action actually happens. That distinction matters: async preparation reduces blocking work, while final authority checks still guard the real effect.</p><h2>Why It Matters</h2><p>The Gateway has been shedding synchronous reads across the session, auth, and storage stack. Credential preparation is a natural but risky place to continue that work because it sits near both performance and security boundaries.</p><p>For users, the expected benefit is steadier preparation for provider-heavy flows. For plugin authors, the important part is the migration path: synchronous SDK APIs are deprecated instead of silently changing behavior, and async peers are documented in the plugin migration docs.</p><p>For maintainers, the deeper win is ownership. Instead of scattered callers taking read snapshots and writing them back, the auth owner owns atomic mutations and worker-backed reads. That reduces duplicated read-modify-write shapes and makes future authority boundaries easier to reason about.</p><h2>Guardrails Stay In Place</h2><p>The PR is explicit about remaining boundaries. It does not change schema, retention policy, permission policy, or durability policy. It also names follow-up work: config health caller-chain migration, async placement initialization, live verbose callbacks, published settings, and physical-target auth batch rollback.</p><p>That kind of disclosure is useful. It keeps the story from becoming \"all sync auth is gone\" when the real progress is more precise: ordinary preparation reads are moving off the Gateway thread, while compatibility APIs and final authority reads remain where they are needed.</p><h2>The Bigger Pattern</h2><p>Taken together with today’s release and other session-worker PRs, #169111 shows OpenClaw continuing a broad runtime cleanup: move ordinary preparation into owners and workers, preserve real effect checks, and give plugin authors migration paths instead of abrupt SDK breaks.</p><p>That is not flashy, but it is exactly the kind of work that makes agent infrastructure less brittle at scale.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-async-auth-credential-prep.png\" alt=\"OpenClaw Async Auth Moves Credential Reads Off Gateway\" /></p>",
      "date_published": "2026-10-11T23:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Async",
        "Auth",
        "Moves",
        "Credential",
        "Reads"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-async-auth-credential-prep.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-async-auth-credential-prep.png",
          "mime_type": "image/png",
          "title": "OpenClaw Async Auth Moves Credential Reads Off Gateway cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-11-beta-2026-10-5-recovery-release/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-11-beta-2026-10-5-recovery-release/",
      "title": "OpenClaw 2026.10.5 Beta Puts Recovery First",
      "summary": "OpenClaw 2026.10.5-beta.1 focuses on safer updates, restart-safe conversations, Gateway speed, reliable turns, X, voice, model pickers, and ClawHub.",
      "content_text": "OpenClaw shipped [v2026.10.5-beta.1](https://github.com/openclaw/openclaw/releases/tag/v2026.10.5-beta.1) today, and this beta is less about one splashy feature than about hardening the places where real agent systems usually fray: updates, restarts, Gateway responsiveness, long-running turns, and provider catalogs.\n\nThe release landed after the morning aggregation window, so it is the headline item for the nightly run. It is also the first new GitHub release after the stable `v2026.10.1` train was covered, making it a clean new source rather than a recycled changelog.\n\n## Safer Updates\n\nThe top release theme is update recovery. The release notes say updates now finish across devices, filesystems, Windows elevation, and interrupted activations, while recovering from stale leases or changed evidence instead of leaving an install stranded.\n\nThat is the kind of infrastructure work users mostly notice when it fails. OpenClaw has been shipping a steady stream of update and Doctor repair changes over the past week; this beta bundles those strands into a single testable release.\n\nThe practical takeaway: if an update crosses a reboot, a privilege boundary, or a partially retired install state, the project is trying to make the updater recover explicitly instead of requiring manual cleanup.\n\n## Restart-Safe Conversations\n\nThe second major theme is conversation continuity. The release highlights interrupted turns, first messages during worktree preparation, deferred tasks, concurrent writes, failed publications, and committed session data surviving Gateway restarts.\n\nFor operators, that matters because OpenClaw sits in long-running workflows: channels, automations, coding tasks, subagents, and background jobs. A restart should not turn every in-flight operation into a mystery.\n\nThis beta continues the move toward durable handoffs. User input, committed session state, and deferred work are increasingly treated as recoverable data rather than process-local memory.\n\n## Gateway Performance Keeps Moving To Workers\n\nThe release also calls out a \"faster, steadier Gateway,\" with session, transcript, history, and metadata reads moving onto worker threads. That matches the larger direction visible in recent merged PRs: keep SQLite and heavy projection work away from the Gateway main thread whenever possible.\n\nThe release notes emphasize lower SQLite work per turn and startup maintenance that runs after readiness. In plain terms, busy Gateways should spend less time doing blocking housekeeping while users are waiting for chats, channel responses, or UI updates.\n\n## Turns, X, Voice, Models, And ClawHub\n\nThe beta includes several product-facing areas too:\n\n- Reliable turns and subagents: recovery from stalled streams, reused tool-call IDs, expired ChatGPT websockets, and expired delivery windows.\n- X channel: a new channel plugin for answering X mentions, including streaming, guest mode for repository questions, and cost limits.\n- Voice calls: per-call briefs, reports, live steering, callbacks, and voicemail detection.\n- Model pickers: provider recommendations, newly listed Claude, ChatGPT, and Ollama models, and more responsive catalog loading.\n- ClawHub in the Gateway: browsing and searching ClawHub plugins and skills directly from the Gateway, with plugin release selection tied to what the running OpenClaw supports.\n\n## Why This Beta Matters\n\nThis is a beta release, so cautious users may wait for a stable follow-up. But the shape is important: OpenClaw is investing in the reliability layer beneath agents, not just the visible agent features.\n\nThe best beta releases make the next stable release boring in a good way. `v2026.10.5-beta.1` is pointing in that direction: fewer stranded updates, fewer lost turns, fewer blocking reads, and more recoverable work when the Gateway or its surrounding services restart.",
      "content_html": "<p>OpenClaw shipped <a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.10.5-beta.1\">v2026.10.5-beta.1</a> today, and this beta is less about one splashy feature than about hardening the places where real agent systems usually fray: updates, restarts, Gateway responsiveness, long-running turns, and provider catalogs.</p><p>The release landed after the morning aggregation window, so it is the headline item for the nightly run. It is also the first new GitHub release after the stable <code>v2026.10.1</code> train was covered, making it a clean new source rather than a recycled changelog.</p><h2>Safer Updates</h2><p>The top release theme is update recovery. The release notes say updates now finish across devices, filesystems, Windows elevation, and interrupted activations, while recovering from stale leases or changed evidence instead of leaving an install stranded.</p><p>That is the kind of infrastructure work users mostly notice when it fails. OpenClaw has been shipping a steady stream of update and Doctor repair changes over the past week; this beta bundles those strands into a single testable release.</p><p>The practical takeaway: if an update crosses a reboot, a privilege boundary, or a partially retired install state, the project is trying to make the updater recover explicitly instead of requiring manual cleanup.</p><h2>Restart-Safe Conversations</h2><p>The second major theme is conversation continuity. The release highlights interrupted turns, first messages during worktree preparation, deferred tasks, concurrent writes, failed publications, and committed session data surviving Gateway restarts.</p><p>For operators, that matters because OpenClaw sits in long-running workflows: channels, automations, coding tasks, subagents, and background jobs. A restart should not turn every in-flight operation into a mystery.</p><p>This beta continues the move toward durable handoffs. User input, committed session state, and deferred work are increasingly treated as recoverable data rather than process-local memory.</p><h2>Gateway Performance Keeps Moving To Workers</h2><p>The release also calls out a \"faster, steadier Gateway,\" with session, transcript, history, and metadata reads moving onto worker threads. That matches the larger direction visible in recent merged PRs: keep SQLite and heavy projection work away from the Gateway main thread whenever possible.</p><p>The release notes emphasize lower SQLite work per turn and startup maintenance that runs after readiness. In plain terms, busy Gateways should spend less time doing blocking housekeeping while users are waiting for chats, channel responses, or UI updates.</p><h2>Turns, X, Voice, Models, And ClawHub</h2><p>The beta includes several product-facing areas too:</p><ul><li>Reliable turns and subagents: recovery from stalled streams, reused tool-call IDs, expired ChatGPT websockets, and expired delivery windows.</li><li>X channel: a new channel plugin for answering X mentions, including streaming, guest mode for repository questions, and cost limits.</li><li>Voice calls: per-call briefs, reports, live steering, callbacks, and voicemail detection.</li><li>Model pickers: provider recommendations, newly listed Claude, ChatGPT, and Ollama models, and more responsive catalog loading.</li><li>ClawHub in the Gateway: browsing and searching ClawHub plugins and skills directly from the Gateway, with plugin release selection tied to what the running OpenClaw supports.</li></ul><h2>Why This Beta Matters</h2><p>This is a beta release, so cautious users may wait for a stable follow-up. But the shape is important: OpenClaw is investing in the reliability layer beneath agents, not just the visible agent features.</p><p>The best beta releases make the next stable release boring in a good way. <code>v2026.10.5-beta.1</code> is pointing in that direction: fewer stranded updates, fewer lost turns, fewer blocking reads, and more recoverable work when the Gateway or its surrounding services restart.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-beta-2026-10-5-recovery-release.png\" alt=\"OpenClaw 2026.10.5 Beta Puts Recovery First\" /></p>",
      "date_published": "2026-10-11T23:00:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "Local Models",
        "2026",
        "Beta",
        "Puts",
        "Recovery",
        "First"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-beta-2026-10-5-recovery-release.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-beta-2026-10-5-recovery-release.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.10.5 Beta Puts Recovery First cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-11-memory-lancedb-profile-isolation/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-11-memory-lancedb-profile-isolation/",
      "title": "OpenClaw Isolates Memory LanceDB Profiles",
      "summary": "OpenClaw PR #168944 moves the default Memory LanceDB store under the active state directory for named profiles.",
      "content_text": "OpenClaw merged [PR #168944](https://github.com/openclaw/openclaw/pull/168944), a Memory LanceDB fix that changes where the plugin puts its default long-term memory store for named profiles and isolated state directories.\n\nBefore this patch, `memory-lancedb` resolved its default database path from the user's home directory at module load. That meant a named profile, custom `OPENCLAW_STATE_DIR`, relocated `OPENCLAW_HOME`, or isolated proof run could still read and write the real `~/.openclaw/memory/lancedb` store unless the user explicitly set `dbPath`.\n\n## What Changed\n\nThe default LanceDB path now resolves at config-parse time through OpenClaw's documented `openclaw/plugin-sdk/state-paths` helper. Without an explicit `dbPath`, the store lives at:\n\n- Default profile: `~/.openclaw/memory/lancedb`, unchanged.\n- Named profile: that profile's active state directory under `memory/lancedb`.\n- Custom state directory: `<stateDir>/memory/lancedb`.\n\nThe Doctor contract now uses the same helper with the host-supplied state directory, so Doctor does not accidentally target the home store from an isolated state root.\n\nExplicit `dbPath` values behave as before. No LanceDB row format, schema, migration marker, or storage file is changed.\n\n## Why It Matters\n\nProfiles are supposed to isolate state. A named profile used for testing, proof work, customer reproduction, or a separate agent persona should not silently share the default profile's long-term memory database.\n\nThe maintainer decision here is intentionally conservative: isolate, do not migrate. Existing memories remain where they were written. For the default profile, that means nothing moves. For named profiles that previously leaked into the shared home store, the new isolated store starts empty unless the operator deliberately points `dbPath` back to the old shared location.\n\nThat avoids copying default-profile memories into every profile on startup, which would preserve the very leak the PR is trying to remove.\n\n## Proof From The PR\n\nThe regression tests cover both configuration and Doctor behavior. `config.test.ts` verifies that the default path respects `OPENCLAW_STATE_DIR` even when it is set after module import, while preserving explicit `dbPath`. `doctor-contract-api.test.ts` verifies that migration detection uses the migration state directory instead of `HOME`.\n\nThe PR also ran memory-lancedb startup/import and plugin doctor-contract suites, extension typechecks, lint, formatting, and diff checks.\n\nThe real CLI proof is especially useful. Before the fix, `ltm list` under a scratch state configuration registered the plugin against `<HOME>/.openclaw/memory/lancedb` and created that directory. After the fix, the rebuilt runtime registered the plugin against `<scratch>/memory/lancedb` and did not create the home store.\n\nCompatibility checks seeded two memories into the old default store. The candidate still listed both memories under the default profile, showed zero memories under an isolated profile, and allowed a separate profile to opt back into the shared store with an explicit `dbPath`.\n\n## Bottom Line\n\nPR #168944 closes a memory-isolation gap in OpenClaw's LanceDB plugin. Default users keep their existing memory path, while named profiles and custom state directories stop borrowing the home profile's long-term memory by accident.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168944\">PR #168944</a>, a Memory LanceDB fix that changes where the plugin puts its default long-term memory store for named profiles and isolated state directories.</p><p>Before this patch, <code>memory-lancedb</code> resolved its default database path from the user's home directory at module load. That meant a named profile, custom <code>OPENCLAW_STATE_DIR</code>, relocated <code>OPENCLAW_HOME</code>, or isolated proof run could still read and write the real <code>~/.openclaw/memory/lancedb</code> store unless the user explicitly set <code>dbPath</code>.</p><h2>What Changed</h2><p>The default LanceDB path now resolves at config-parse time through OpenClaw's documented <code>openclaw/plugin-sdk/state-paths</code> helper. Without an explicit <code>dbPath</code>, the store lives at:</p><ul><li>Default profile: <code>~/.openclaw/memory/lancedb</code>, unchanged.</li><li>Named profile: that profile's active state directory under <code>memory/lancedb</code>.</li><li>Custom state directory: <code><stateDir>/memory/lancedb</code>.</li></ul><p>The Doctor contract now uses the same helper with the host-supplied state directory, so Doctor does not accidentally target the home store from an isolated state root.</p><p>Explicit <code>dbPath</code> values behave as before. No LanceDB row format, schema, migration marker, or storage file is changed.</p><h2>Why It Matters</h2><p>Profiles are supposed to isolate state. A named profile used for testing, proof work, customer reproduction, or a separate agent persona should not silently share the default profile's long-term memory database.</p><p>The maintainer decision here is intentionally conservative: isolate, do not migrate. Existing memories remain where they were written. For the default profile, that means nothing moves. For named profiles that previously leaked into the shared home store, the new isolated store starts empty unless the operator deliberately points <code>dbPath</code> back to the old shared location.</p><p>That avoids copying default-profile memories into every profile on startup, which would preserve the very leak the PR is trying to remove.</p><h2>Proof From The PR</h2><p>The regression tests cover both configuration and Doctor behavior. <code>config.test.ts</code> verifies that the default path respects <code>OPENCLAW_STATE_DIR</code> even when it is set after module import, while preserving explicit <code>dbPath</code>. <code>doctor-contract-api.test.ts</code> verifies that migration detection uses the migration state directory instead of <code>HOME</code>.</p><p>The PR also ran memory-lancedb startup/import and plugin doctor-contract suites, extension typechecks, lint, formatting, and diff checks.</p><p>The real CLI proof is especially useful. Before the fix, <code>ltm list</code> under a scratch state configuration registered the plugin against <code><HOME>/.openclaw/memory/lancedb</code> and created that directory. After the fix, the rebuilt runtime registered the plugin against <code><scratch>/memory/lancedb</code> and did not create the home store.</p><p>Compatibility checks seeded two memories into the old default store. The candidate still listed both memories under the default profile, showed zero memories under an isolated profile, and allowed a separate profile to opt back into the shared store with an explicit <code>dbPath</code>.</p><h2>Bottom Line</h2><p>PR #168944 closes a memory-isolation gap in OpenClaw's LanceDB plugin. Default users keep their existing memory path, while named profiles and custom state directories stop borrowing the home profile's long-term memory by accident.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-memory-lancedb-profile-isolation.png\" alt=\"OpenClaw Isolates Memory LanceDB Profiles\" /></p>",
      "date_published": "2026-10-11T08:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Isolates",
        "Lancedb",
        "Profiles"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-memory-lancedb-profile-isolation.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-memory-lancedb-profile-isolation.png",
          "mime_type": "image/png",
          "title": "OpenClaw Isolates Memory LanceDB Profiles cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-11-update-heap-limit-preservation/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-11-update-heap-limit-preservation/",
      "title": "OpenClaw Preserves Heap Limits During Updates",
      "summary": "OpenClaw PR #168889 helps update, Doctor, recovery, and verification children inherit effective service heap controls.",
      "content_text": "OpenClaw merged [PR #168889](https://github.com/openclaw/openclaw/pull/168889), a high-priority update fix for large agent databases and managed Gateway services with custom heap limits.\n\nThe bug showed up during updates. A managed Gateway could be running with a sufficient Node heap limit, but the update child processes used for candidate rehearsal, activation Doctor, recovery, and verification could lose those heap controls. In another edge case, an empty service `NODE_OPTIONS` could erase a caller-supplied heap override.\n\nThe result was a confusing failure mode: the service had enough memory, but the updater children could run out of heap while Doctor checked a large database.\n\n## What Changed\n\nThe update service-state reader now projects the effective command's heap flags into the captured environment used by update children. It uses the existing heap parser rather than inventing a new configuration surface.\n\nService command-line heap flags keep precedence over `NODE_OPTIONS`. If the service has an empty `NODE_OPTIONS` value, OpenClaw now preserves the caller's heap controls while still clearing inherited preload and debugger flags as intended.\n\nThe fix does not add a new option, dependency, database migration, or update marker. It improves how existing service facts are carried into existing update phases.\n\n## Why It Matters\n\nUpdate reliability depends on matching the runtime conditions that operators already set for their Gateway. If a service needs a larger heap to manage a big agent database, the update path should not silently forget that constraint at the moment it runs Doctor or recovery.\n\nThe practical benefit is strongest for installations with:\n\n- Large agent databases.\n- Managed Gateway services using explicit heap flags.\n- Update rehearsals that run Doctor before activation.\n- Recovery or verification phases launched as child processes.\n\nThere is also an important limit: this protection must exist in the updater that is already installed before the update starts. An older updater cannot acquire this behavior from the candidate it has not safely activated yet.\n\n## Proof From The PR\n\nThe PR includes real child-process assertions that `NODE_OPTIONS` carries the expected heap control. All four targeted cases assert that the final old-space flag is `--max-old-space-size=160`, including service-argv precedence and clearing an inherited missing preload.\n\nFocused tests passed under both Node 26.11.1 and Bun 1.4.2. The broader unchanged-production evidence also ran 1,258 selected tests across 54 files, plus changed-file checks, import-cycle checks, and relevant lint/typecheck work.\n\nThe PR is careful not to overclaim. It explicitly says no live 7 GB database or operator Gateway was used, and it does not speculate on unrelated database-maintenance causes. The implemented fix is the heap-control inheritance path.\n\n## Bottom Line\n\nPR #168889 makes OpenClaw updates respect the service memory posture operators already chose. That should reduce update-time Doctor heap failures on larger installations without changing configuration or database behavior.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168889\">PR #168889</a>, a high-priority update fix for large agent databases and managed Gateway services with custom heap limits.</p><p>The bug showed up during updates. A managed Gateway could be running with a sufficient Node heap limit, but the update child processes used for candidate rehearsal, activation Doctor, recovery, and verification could lose those heap controls. In another edge case, an empty service <code>NODE_OPTIONS</code> could erase a caller-supplied heap override.</p><p>The result was a confusing failure mode: the service had enough memory, but the updater children could run out of heap while Doctor checked a large database.</p><h2>What Changed</h2><p>The update service-state reader now projects the effective command's heap flags into the captured environment used by update children. It uses the existing heap parser rather than inventing a new configuration surface.</p><p>Service command-line heap flags keep precedence over <code>NODE_OPTIONS</code>. If the service has an empty <code>NODE_OPTIONS</code> value, OpenClaw now preserves the caller's heap controls while still clearing inherited preload and debugger flags as intended.</p><p>The fix does not add a new option, dependency, database migration, or update marker. It improves how existing service facts are carried into existing update phases.</p><h2>Why It Matters</h2><p>Update reliability depends on matching the runtime conditions that operators already set for their Gateway. If a service needs a larger heap to manage a big agent database, the update path should not silently forget that constraint at the moment it runs Doctor or recovery.</p><p>The practical benefit is strongest for installations with:</p><ul><li>Large agent databases.</li><li>Managed Gateway services using explicit heap flags.</li><li>Update rehearsals that run Doctor before activation.</li><li>Recovery or verification phases launched as child processes.</li></ul><p>There is also an important limit: this protection must exist in the updater that is already installed before the update starts. An older updater cannot acquire this behavior from the candidate it has not safely activated yet.</p><h2>Proof From The PR</h2><p>The PR includes real child-process assertions that <code>NODE_OPTIONS</code> carries the expected heap control. All four targeted cases assert that the final old-space flag is <code>--max-old-space-size=160</code>, including service-argv precedence and clearing an inherited missing preload.</p><p>Focused tests passed under both Node 26.11.1 and Bun 1.4.2. The broader unchanged-production evidence also ran 1,258 selected tests across 54 files, plus changed-file checks, import-cycle checks, and relevant lint/typecheck work.</p><p>The PR is careful not to overclaim. It explicitly says no live 7 GB database or operator Gateway was used, and it does not speculate on unrelated database-maintenance causes. The implemented fix is the heap-control inheritance path.</p><h2>Bottom Line</h2><p>PR #168889 makes OpenClaw updates respect the service memory posture operators already chose. That should reduce update-time Doctor heap failures on larger installations without changing configuration or database behavior.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-update-heap-limit-preservation.png\" alt=\"OpenClaw Preserves Heap Limits During Updates\" /></p>",
      "date_published": "2026-10-11T08:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Memory",
        "Preserves",
        "Heap",
        "Limits",
        "During",
        "Updates"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-update-heap-limit-preservation.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-update-heap-limit-preservation.png",
          "mime_type": "image/png",
          "title": "OpenClaw Preserves Heap Limits During Updates cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-11-worker-inventory-shutdown/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-11-worker-inventory-shutdown/",
      "title": "OpenClaw Fixes Worker Inventory Shutdowns",
      "summary": "OpenClaw PR #168862 lets Gateway shutdown drain worker tunnels and transfers cleanly when a worker inventory retires first.",
      "content_text": "OpenClaw merged [PR #168862](https://github.com/openclaw/openclaw/pull/168862), a Gateway reliability fix for shutdowns that intersect with retired worker inventories.\n\nThe issue was narrow but operationally annoying. A Gateway could be otherwise finished shutting down, yet still fail because the worker inventory had already retired before the tunnel cleanup path could ask it for the final state. A related reconciliation path could also reject after shutdown had already started.\n\n## What Changed\n\nThe fix adds a typed retirement error for worker inventories and handles it at two shutdown boundaries. Tunnel shutdown now treats an already retired inventory as acceptable only around the fallback inventory lookup, while still awaiting and reporting real remote cleanup failures. Reconciliation also treats retirement as expected only after its existing stop flag has been set.\n\nThat distinction matters. OpenClaw is not hiding every worker-environment failure during shutdown. It is recognizing the specific case where the inventory owner has already crossed the retirement boundary and the cleanup path should finish draining retained tunnels and workspace transfers instead of failing the whole shutdown.\n\nThe PR also simplifies nearby tunnel status and live-owner code, but the behavioral change stays focused. No database schema, configuration, lifecycle state, or memory limit is added.\n\n## Why It Matters\n\nGateway shutdown is a trust surface. When users stop a Gateway, they need retained worker tunnels and workspace transfers to be cleaned up predictably. A false shutdown failure can make an operator wonder whether cleanup actually happened, whether they need to restart, or whether a workspace transfer was left in an uncertain state.\n\nThis PR makes the shutdown path more tolerant of one expected race:\n\n- Worker inventory retirement can happen before the last cleanup lookup.\n- Retained tunnels and transfers still drain.\n- Ordinary inventory, teardown, and remote cleanup failures remain visible.\n- A running Gateway with a retired inventory still needs separate recovery.\n\nThat last point keeps the fix honest. It repairs the shutdown boundary, not every possible worker inventory failure.\n\n## Proof From The PR\n\nThe PR records red/green regression evidence before the handling fix. The two retirement regressions failed before the patch while their ordinary-error counterparts still passed, confirming that the repair targeted the expected edge rather than muting broad errors.\n\nThe final proof covered `node-worker-tunnel` and `service-lifetime` suites on isolated AWS. The PR reports 26 tunnel tests and 27 service-lifetime tests passing, along with type-aware lint, export collision checks, line-cap checks, formatting, core typecheck, and Knip scans.\n\nThere was also a follow-up after CI found a missing shared retirement-error module in the PR wrapper source inventory. The final integration preserved the worker-file behavior, resolved conflicts with landed byte-reader work, and kept the refreshed P0/P1 review clean.\n\n## Bottom Line\n\nPR #168862 makes OpenClaw Gateway shutdowns less brittle around retired worker inventories. Operators should see fewer false shutdown failures while still getting signal when real worker cleanup fails.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168862\">PR #168862</a>, a Gateway reliability fix for shutdowns that intersect with retired worker inventories.</p><p>The issue was narrow but operationally annoying. A Gateway could be otherwise finished shutting down, yet still fail because the worker inventory had already retired before the tunnel cleanup path could ask it for the final state. A related reconciliation path could also reject after shutdown had already started.</p><h2>What Changed</h2><p>The fix adds a typed retirement error for worker inventories and handles it at two shutdown boundaries. Tunnel shutdown now treats an already retired inventory as acceptable only around the fallback inventory lookup, while still awaiting and reporting real remote cleanup failures. Reconciliation also treats retirement as expected only after its existing stop flag has been set.</p><p>That distinction matters. OpenClaw is not hiding every worker-environment failure during shutdown. It is recognizing the specific case where the inventory owner has already crossed the retirement boundary and the cleanup path should finish draining retained tunnels and workspace transfers instead of failing the whole shutdown.</p><p>The PR also simplifies nearby tunnel status and live-owner code, but the behavioral change stays focused. No database schema, configuration, lifecycle state, or memory limit is added.</p><h2>Why It Matters</h2><p>Gateway shutdown is a trust surface. When users stop a Gateway, they need retained worker tunnels and workspace transfers to be cleaned up predictably. A false shutdown failure can make an operator wonder whether cleanup actually happened, whether they need to restart, or whether a workspace transfer was left in an uncertain state.</p><p>This PR makes the shutdown path more tolerant of one expected race:</p><ul><li>Worker inventory retirement can happen before the last cleanup lookup.</li><li>Retained tunnels and transfers still drain.</li><li>Ordinary inventory, teardown, and remote cleanup failures remain visible.</li><li>A running Gateway with a retired inventory still needs separate recovery.</li></ul><p>That last point keeps the fix honest. It repairs the shutdown boundary, not every possible worker inventory failure.</p><h2>Proof From The PR</h2><p>The PR records red/green regression evidence before the handling fix. The two retirement regressions failed before the patch while their ordinary-error counterparts still passed, confirming that the repair targeted the expected edge rather than muting broad errors.</p><p>The final proof covered <code>node-worker-tunnel</code> and <code>service-lifetime</code> suites on isolated AWS. The PR reports 26 tunnel tests and 27 service-lifetime tests passing, along with type-aware lint, export collision checks, line-cap checks, formatting, core typecheck, and Knip scans.</p><p>There was also a follow-up after CI found a missing shared retirement-error module in the PR wrapper source inventory. The final integration preserved the worker-file behavior, resolved conflicts with landed byte-reader work, and kept the refreshed P0/P1 review clean.</p><h2>Bottom Line</h2><p>PR #168862 makes OpenClaw Gateway shutdowns less brittle around retired worker inventories. Operators should see fewer false shutdown failures while still getting signal when real worker cleanup fails.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-worker-inventory-shutdown.png\" alt=\"OpenClaw Fixes Worker Inventory Shutdowns\" /></p>",
      "date_published": "2026-10-11T08:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.483Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Memory",
        "Fixes",
        "Worker",
        "Inventory",
        "Shutdowns"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-worker-inventory-shutdown.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-11-worker-inventory-shutdown.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Worker Inventory Shutdowns cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-10-memory-flush-token-estimates/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-10-memory-flush-token-estimates/",
      "title": "OpenClaw Memory Flushes Now Use Token Estimates",
      "summary": "OpenClaw PR #83178 lets memory checkpoints run for local and compatible model servers even when provider usage metadata is missing.",
      "content_text": "OpenClaw merged [PR #83178](https://github.com/openclaw/openclaw/pull/83178), a memory-flush fix for model servers that omit token usage metadata.\n\nBefore this change, pre-compaction memory checkpoints could be skipped when a provider reported missing or zero usage, even when transcript-based compaction already knew the session was under context pressure. That was especially relevant for local and OpenAI-compatible model servers, where usage accounting is less consistent than hosted APIs.\n\n## What Changed\n\nMemory flushing now shares OpenClaw's existing transcript estimator with compaction. When provider usage is unavailable, the memory checkpoint can use transcript-estimated pressure to decide whether it should run.\n\nThe PR is careful about what gets persisted. Estimated pressure can trigger the checkpoint, but it does not overwrite provider-reported session status. Only provider-anchored totals are persisted as fresh usage. The estimate stays transient and scoped to the decision that needs it.\n\nThe implementation also reuses the transcript reader's already-read accounting snapshot, keeps read fences and cancellation behavior intact, and counts assistant output once through the shared projection path.\n\n## Why It Matters\n\nMemory checkpoints are most useful when a conversation is approaching compaction pressure. If a local model server omits usage metadata, OpenClaw should not have to choose between pretending pressure is unknown and skipping useful memory work.\n\nThis fix makes memory behavior more provider-tolerant without turning estimates into fake authoritative usage. That distinction is important. Operators still see honest usage freshness, while the runtime gets enough signal to preserve memory before the context window gets tight.\n\nThe practical effect is strongest for:\n\n- Local Ollama-style setups routed through compatible adapters.\n- Model servers that return zero or missing usage fields.\n- Sessions with long transcripts and small context windows.\n- Users relying on memory checkpoints before auto-compaction.\n\n## Real Local-Model Proof\n\nThe PR includes an isolated Gateway proof using Ollama `qwen2.5:7b` through an OpenAI-compatible loopback adapter that deliberately removed response usage metadata.\n\nTwo synthetic chat inputs were run before and after the fix. The first turn returned normally. The second reached compaction pressure with an estimated 21,719 tokens against a 6,144-token memory-flush threshold.\n\nBefore the fix, the memory flush had no token count and skipped. After the fix, OpenClaw used the transcript estimate, dispatched a real Qwen maintenance turn, completed the checkpoint in 7.06 seconds, and persisted a successful memory-flush marker. The source sessions still retained `totalTokens: 0` and `totalTokensFresh: false`, preserving the distinction between estimated pressure and provider-reported usage.\n\nThe oversized second chat still hit the same independent context-overflow failure on both versions, but the repaired memory checkpoint completed before that failure.\n\n## Bottom Line\n\nPR #83178 gives OpenClaw a better fallback when providers do not report token usage. Memory checkpoints can now fire based on transcript pressure, while session status stays honest about whether usage came from the provider.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/83178\">PR #83178</a>, a memory-flush fix for model servers that omit token usage metadata.</p><p>Before this change, pre-compaction memory checkpoints could be skipped when a provider reported missing or zero usage, even when transcript-based compaction already knew the session was under context pressure. That was especially relevant for local and OpenAI-compatible model servers, where usage accounting is less consistent than hosted APIs.</p><h2>What Changed</h2><p>Memory flushing now shares OpenClaw's existing transcript estimator with compaction. When provider usage is unavailable, the memory checkpoint can use transcript-estimated pressure to decide whether it should run.</p><p>The PR is careful about what gets persisted. Estimated pressure can trigger the checkpoint, but it does not overwrite provider-reported session status. Only provider-anchored totals are persisted as fresh usage. The estimate stays transient and scoped to the decision that needs it.</p><p>The implementation also reuses the transcript reader's already-read accounting snapshot, keeps read fences and cancellation behavior intact, and counts assistant output once through the shared projection path.</p><h2>Why It Matters</h2><p>Memory checkpoints are most useful when a conversation is approaching compaction pressure. If a local model server omits usage metadata, OpenClaw should not have to choose between pretending pressure is unknown and skipping useful memory work.</p><p>This fix makes memory behavior more provider-tolerant without turning estimates into fake authoritative usage. That distinction is important. Operators still see honest usage freshness, while the runtime gets enough signal to preserve memory before the context window gets tight.</p><p>The practical effect is strongest for:</p><ul><li>Local Ollama-style setups routed through compatible adapters.</li><li>Model servers that return zero or missing usage fields.</li><li>Sessions with long transcripts and small context windows.</li><li>Users relying on memory checkpoints before auto-compaction.</li></ul><h2>Real Local-Model Proof</h2><p>The PR includes an isolated Gateway proof using Ollama <code>qwen2.5:7b</code> through an OpenAI-compatible loopback adapter that deliberately removed response usage metadata.</p><p>Two synthetic chat inputs were run before and after the fix. The first turn returned normally. The second reached compaction pressure with an estimated 21,719 tokens against a 6,144-token memory-flush threshold.</p><p>Before the fix, the memory flush had no token count and skipped. After the fix, OpenClaw used the transcript estimate, dispatched a real Qwen maintenance turn, completed the checkpoint in 7.06 seconds, and persisted a successful memory-flush marker. The source sessions still retained <code>totalTokens: 0</code> and <code>totalTokensFresh: false</code>, preserving the distinction between estimated pressure and provider-reported usage.</p><p>The oversized second chat still hit the same independent context-overflow failure on both versions, but the repaired memory checkpoint completed before that failure.</p><h2>Bottom Line</h2><p>PR #83178 gives OpenClaw a better fallback when providers do not report token usage. Memory checkpoints can now fire based on transcript pressure, while session status stays honest about whether usage came from the provider.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-memory-flush-token-estimates.png\" alt=\"OpenClaw Memory Flushes Now Use Token Estimates\" /></p>",
      "date_published": "2026-10-10T23:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Local Models",
        "Flushes",
        "Token",
        "Estimates"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-memory-flush-token-estimates.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-memory-flush-token-estimates.png",
          "mime_type": "image/png",
          "title": "OpenClaw Memory Flushes Now Use Token Estimates cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-10-migration-state-owner-guard/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-10-migration-state-owner-guard/",
      "title": "OpenClaw Guards Migration Imports During Gateway Runs",
      "summary": "OpenClaw PR #168419 refuses migration imports while a live Gateway owns local state, reducing stale-state risk during migrations.",
      "content_text": "OpenClaw merged [PR #168419](https://github.com/openclaw/openclaw/pull/168419), a migration safety fix that refuses provider loading and imports while another Gateway already owns the selected local state.\n\nThe problem is easy to miss until it bites: migration commands can write local state outside the running Gateway. If that happens while the Gateway is live, the process that users are actually talking to can keep stale in-memory assumptions after an import changes files underneath it.\n\n## What Changed\n\nThe fix wraps migration provider loading and apply work with OpenClaw's existing local-state owner. If a Gateway is already running against the selected state root, the migration command now fails early with a clear refusal instead of planning or mutating state.\n\nWhen the Gateway is stopped, the command takes exclusive offline ownership through the import, rollback, and provider cleanup path. Memory imports that run inside the owning Gateway keep their existing route.\n\nThe PR deliberately avoids adding a new migration RPC layer, config-loading interception, schema change, or broad compatibility surface. It uses the ownership machinery that already exists and applies it at the migration boundary.\n\n## Why It Matters\n\nMigrations are one of the most sensitive pieces of an agent runtime. They often touch account data, imported transcripts, memory stores, provider configuration, or local files that other processes assume are stable.\n\nThe failure mode here is not only \"two writers at once.\" It is \"one writer succeeds while the visible Gateway keeps operating with old facts.\" That can lead to confusing follow-up failures, partial recovery, or cleanup that looks complete to one process and incomplete to another.\n\nThe new guard makes the operational rule explicit:\n\n- If a live Gateway owns the state, migration commands stay out.\n- If the Gateway is stopped, the command owns the offline import until cleanup is done.\n- If nested roots or uncertain child cleanup appear, the command preserves the existing physical lease behavior rather than silently admitting later writes.\n\nThat is a conservative design, and for state migration, conservative is the right shape.\n\n## Proof From The PR\n\nThe PR body records focused provider and command tests, changed-file checks, Madge, Knip, a worker ratchet, SDK comparison, and multiple review rounds. It also includes fresh regression evidence for two important cases.\n\nOne regression proved that a provider could change the selected root during planning and still be admitted before the repair. Another proved that uncertain child cleanup could leave later mutation admitted when it should have been refused. Both assertions are retained after the fix.\n\nThe final landing used a coordinator-authorized inherited-CI exception for an unrelated cron assertion that was already failing on main. The PR documents that exception rather than presenting the whole CI run as green.\n\n## Bottom Line\n\nPR #168419 tightens OpenClaw's migration boundary around the real owner of local state. Migration imports now wait their turn instead of racing the live Gateway, which should make import, rollback, and cleanup behavior easier to reason about when operators move data between providers.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168419\">PR #168419</a>, a migration safety fix that refuses provider loading and imports while another Gateway already owns the selected local state.</p><p>The problem is easy to miss until it bites: migration commands can write local state outside the running Gateway. If that happens while the Gateway is live, the process that users are actually talking to can keep stale in-memory assumptions after an import changes files underneath it.</p><h2>What Changed</h2><p>The fix wraps migration provider loading and apply work with OpenClaw's existing local-state owner. If a Gateway is already running against the selected state root, the migration command now fails early with a clear refusal instead of planning or mutating state.</p><p>When the Gateway is stopped, the command takes exclusive offline ownership through the import, rollback, and provider cleanup path. Memory imports that run inside the owning Gateway keep their existing route.</p><p>The PR deliberately avoids adding a new migration RPC layer, config-loading interception, schema change, or broad compatibility surface. It uses the ownership machinery that already exists and applies it at the migration boundary.</p><h2>Why It Matters</h2><p>Migrations are one of the most sensitive pieces of an agent runtime. They often touch account data, imported transcripts, memory stores, provider configuration, or local files that other processes assume are stable.</p><p>The failure mode here is not only \"two writers at once.\" It is \"one writer succeeds while the visible Gateway keeps operating with old facts.\" That can lead to confusing follow-up failures, partial recovery, or cleanup that looks complete to one process and incomplete to another.</p><p>The new guard makes the operational rule explicit:</p><ul><li>If a live Gateway owns the state, migration commands stay out.</li><li>If the Gateway is stopped, the command owns the offline import until cleanup is done.</li><li>If nested roots or uncertain child cleanup appear, the command preserves the existing physical lease behavior rather than silently admitting later writes.</li></ul><p>That is a conservative design, and for state migration, conservative is the right shape.</p><h2>Proof From The PR</h2><p>The PR body records focused provider and command tests, changed-file checks, Madge, Knip, a worker ratchet, SDK comparison, and multiple review rounds. It also includes fresh regression evidence for two important cases.</p><p>One regression proved that a provider could change the selected root during planning and still be admitted before the repair. Another proved that uncertain child cleanup could leave later mutation admitted when it should have been refused. Both assertions are retained after the fix.</p><p>The final landing used a coordinator-authorized inherited-CI exception for an unrelated cron assertion that was already failing on main. The PR documents that exception rather than presenting the whole CI run as green.</p><h2>Bottom Line</h2><p>PR #168419 tightens OpenClaw's migration boundary around the real owner of local state. Migration imports now wait their turn instead of racing the live Gateway, which should make import, rollback, and cleanup behavior easier to reason about when operators move data between providers.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-migration-state-owner-guard.png\" alt=\"OpenClaw Guards Migration Imports During Gateway Runs\" /></p>",
      "date_published": "2026-10-10T23:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Guards",
        "Migration",
        "Imports",
        "During",
        "Gateway"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-migration-state-owner-guard.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-migration-state-owner-guard.png",
          "mime_type": "image/png",
          "title": "OpenClaw Guards Migration Imports During Gateway Runs cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-10-2026-10-1-stable-npm-release/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-10-2026-10-1-stable-npm-release/",
      "title": "OpenClaw 2026.10.1 Reaches npm Latest",
      "summary": "OpenClaw 2026.10.1 is now the npm latest release, bundling agent, update, plugin startup, Codex, local model, and channel fixes.",
      "content_text": "OpenClaw 2026.10.1 is now the `latest` npm release. The package metadata changed at 2026-10-10 22:45 UTC, with npm reporting `2026.10.1` as latest, `2026.10.1-beta.2` as beta, and `2026.8.35` as extended-stable.\n\nThere is not a GitHub Releases page for the tag yet, but the signed Git tag `v2026.10.1` exists and points at a changelog refresh commit. That makes this a package-first stable ship: the installable npm package is live, while the richer release page may still be catching up.\n\n## What Shipped\n\nThe refreshed changelog expands the audited release record from 377 to 396 unique PRs and adds several practical fix groups. The biggest themes are agent continuity, update recovery, installed-plugin startup performance, local model behavior, and channel reliability.\n\nHighlights from the changelog include:\n\n- Agents and subagents keep completed results through memory-saving turns, follow-up messages, reasoning-only tails, and small-context auto-compaction recovery.\n- Update and Doctor paths settle prepared activations, avoid blocked rollbacks when the Gateway is stopped, and stop repeating refused update advice while recovery is still pending.\n- Gateways with installed plugins avoid repeated plugin Doctor captures and regain the fast Windows startup path for large plugin installs.\n- Codex, ChatGPT, Ollama, Gemma 4, and managed llama.cpp paths pick up fixes for final-token loss, rejected cached tool continuations, context errors, Tool Search, trailing tool-call text, and memory indexing.\n- Telegram, Bun state ownership, Docker owner recovery, memory-wiki status, global npm installs, and local browser terminal color support all get targeted reliability fixes.\n\n## Why The npm Signal Matters\n\nOpenClaw users often install or update through npm before reading a release page. A new `latest` tag means default install commands now resolve to 2026.10.1, not the earlier 2026.9.9 stable build.\n\nThat matters for operators who track stable rather than beta. The beta channel had already surfaced 2026.10.1 work earlier this week, but npm `latest` is the broad distribution switch. It is the version ordinary package consumers are most likely to receive.\n\nThe package metadata also includes the tarball and integrity string:\n\n- Package: `openclaw@2026.10.1`\n- Tarball: `https://registry.npmjs.org/openclaw/-/openclaw-2026.10.1.tgz`\n- Integrity: `sha512-C072/lxy83KP1nMCnJbipr4O+EtIR9XXbjzQFM+ol7VE5ci2MkKk5XTxfSG70jxpEVaGFeXcEmObHxR/ZHLdHw==`\n\n## What To Watch\n\nBecause the GitHub release page is not published yet, treat npm and the signed tag as the primary public artifacts for now. The changelog commit is still useful, though: it names the added fix buckets and the PR ranges behind them.\n\nThe strongest user-facing fixes are clustered around reliability under pressure. If you run local models, large plugin installs, Windows Gateways, Docker-backed Gateways, Telegram, or frequent update/rollback flows, 2026.10.1 is the stable release to pay attention to.\n\n## Bottom Line\n\nOpenClaw 2026.10.1 has crossed the line from beta train to npm latest. The release is broad rather than flashy: fewer headline features, more hardening across the places real Gateway operators tend to feel pain.",
      "content_html": "<p>OpenClaw 2026.10.1 is now the <code>latest</code> npm release. The package metadata changed at 2026-10-10 22:45 UTC, with npm reporting <code>2026.10.1</code> as latest, <code>2026.10.1-beta.2</code> as beta, and <code>2026.8.35</code> as extended-stable.</p><p>There is not a GitHub Releases page for the tag yet, but the signed Git tag <code>v2026.10.1</code> exists and points at a changelog refresh commit. That makes this a package-first stable ship: the installable npm package is live, while the richer release page may still be catching up.</p><h2>What Shipped</h2><p>The refreshed changelog expands the audited release record from 377 to 396 unique PRs and adds several practical fix groups. The biggest themes are agent continuity, update recovery, installed-plugin startup performance, local model behavior, and channel reliability.</p><p>Highlights from the changelog include:</p><ul><li>Agents and subagents keep completed results through memory-saving turns, follow-up messages, reasoning-only tails, and small-context auto-compaction recovery.</li><li>Update and Doctor paths settle prepared activations, avoid blocked rollbacks when the Gateway is stopped, and stop repeating refused update advice while recovery is still pending.</li><li>Gateways with installed plugins avoid repeated plugin Doctor captures and regain the fast Windows startup path for large plugin installs.</li><li>Codex, ChatGPT, Ollama, Gemma 4, and managed llama.cpp paths pick up fixes for final-token loss, rejected cached tool continuations, context errors, Tool Search, trailing tool-call text, and memory indexing.</li><li>Telegram, Bun state ownership, Docker owner recovery, memory-wiki status, global npm installs, and local browser terminal color support all get targeted reliability fixes.</li></ul><h2>Why The npm Signal Matters</h2><p>OpenClaw users often install or update through npm before reading a release page. A new <code>latest</code> tag means default install commands now resolve to 2026.10.1, not the earlier 2026.9.9 stable build.</p><p>That matters for operators who track stable rather than beta. The beta channel had already surfaced 2026.10.1 work earlier this week, but npm <code>latest</code> is the broad distribution switch. It is the version ordinary package consumers are most likely to receive.</p><p>The package metadata also includes the tarball and integrity string:</p><ul><li>Package: <code>openclaw@2026.10.1</code></li><li>Tarball: <code>https://registry.npmjs.org/openclaw/-/openclaw-2026.10.1.tgz</code></li><li>Integrity: <code>sha512-C072/lxy83KP1nMCnJbipr4O+EtIR9XXbjzQFM+ol7VE5ci2MkKk5XTxfSG70jxpEVaGFeXcEmObHxR/ZHLdHw==</code></li></ul><h2>What To Watch</h2><p>Because the GitHub release page is not published yet, treat npm and the signed tag as the primary public artifacts for now. The changelog commit is still useful, though: it names the added fix buckets and the PR ranges behind them.</p><p>The strongest user-facing fixes are clustered around reliability under pressure. If you run local models, large plugin installs, Windows Gateways, Docker-backed Gateways, Telegram, or frequent update/rollback flows, 2026.10.1 is the stable release to pay attention to.</p><h2>Bottom Line</h2><p>OpenClaw 2026.10.1 has crossed the line from beta train to npm latest. The release is broad rather than flashy: fewer headline features, more hardening across the places real Gateway operators tend to feel pain.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-2026-10-1-stable-npm-release.png\" alt=\"OpenClaw 2026.10.1 Reaches npm Latest\" /></p>",
      "date_published": "2026-10-10T23:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "Local Models",
        "2026",
        "Reaches",
        "Latest"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-2026-10-1-stable-npm-release.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-2026-10-1-stable-npm-release.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.10.1 Reaches npm Latest cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-10-interrupted-request-continuation/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-10-interrupted-request-continuation/",
      "title": "OpenClaw Recovers Interrupted Requests on Continue",
      "summary": "OpenClaw PR #168279 lets agents use bounded previews of interrupted inputs when users ask them to continue later.",
      "content_text": "OpenClaw merged [PR #168279](https://github.com/openclaw/openclaw/pull/168279), a recovery fix for sessions where the user asks an agent to continue after the original request was interrupted before execution.\n\nThe visible bug was frustrating: the Control UI could still show the interrupted request, but the agent did not receive enough context to continue from it. A user could type \"continue\" and see the system behave as if the original work had vanished.\n\n## What Changed\n\nOpenClaw now adds a bounded historical snapshot of retained interrupted input into the existing turn-context path. The snapshot is scoped to the physical session and limited to three previews of up to 4,000 characters from the latest retained-input page.\n\nThat wording matters. The fix does not replay old input, restore expired permissions, or expose queued, cancelled, hidden, or context-excluded messages. It gives the agent enough historical conversation data to understand what \"continue\" refers to while keeping custody and authorization boundaries intact.\n\nThe final version also tightened account handling. Interrupted previews now use the existing authorized history writer's target rather than caller-owned memory or an unknown account. Execution revalidates that same history capability for resumed turns and fresh recovery.\n\n## Why It Matters\n\nLong-running agent work often crosses messy boundaries: browser refreshes, Gateway restarts, interrupted turns, delayed recovery, and a human coming back later with a short follow-up. \"Continue\" only works if OpenClaw can bridge what the UI remembers with what the model is allowed to see.\n\nWithout that bridge, retained pending inputs become a display artifact instead of recoverable context. With it, built-in and CLI agent turns can understand the interrupted request in a constrained way.\n\nThe PR's contract is careful:\n\n- Current request and transcript bytes stay unchanged.\n- Retained inputs are previewed, not replayed.\n- Cancelled and context-excluded messages remain excluded.\n- Cross-account or unverifiable-account cases do not leak previews.\n- Revoked authority prevents resumed transport dispatch.\n\nThat combination is the heart of the fix: useful recovery without making interruption state a shortcut around normal access rules.\n\n## Evidence From The PR\n\nThe PR includes a real SQLite accepted-input regression. Before the fix, the production prompt-assembly boundary produced empty continuation context. After the repair, it exposed the interrupted request without consuming or replaying it.\n\nThe final proof covered cross-account and unverifiable-account leak cases, resumed dispatch after authority revocation, raw prepared input capture, and same-account previews. The focused suite passed 53 tests across four files, along with production types, test-type groups, lint, SQLite-worker guards, test-race checks, mock guards, and independent review.\n\nThere was also a later correction for CLI media-suffix ordering, keeping interrupted-request context before an existing media snapshot so media progress remains in the private context suffix.\n\n## Bottom Line\n\nPR #168279 makes OpenClaw more dependable when work is interrupted midstream. If the UI still has a retained request and the session authority still checks out, agents get bounded context to continue the work instead of losing the thread.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168279\">PR #168279</a>, a recovery fix for sessions where the user asks an agent to continue after the original request was interrupted before execution.</p><p>The visible bug was frustrating: the Control UI could still show the interrupted request, but the agent did not receive enough context to continue from it. A user could type \"continue\" and see the system behave as if the original work had vanished.</p><h2>What Changed</h2><p>OpenClaw now adds a bounded historical snapshot of retained interrupted input into the existing turn-context path. The snapshot is scoped to the physical session and limited to three previews of up to 4,000 characters from the latest retained-input page.</p><p>That wording matters. The fix does not replay old input, restore expired permissions, or expose queued, cancelled, hidden, or context-excluded messages. It gives the agent enough historical conversation data to understand what \"continue\" refers to while keeping custody and authorization boundaries intact.</p><p>The final version also tightened account handling. Interrupted previews now use the existing authorized history writer's target rather than caller-owned memory or an unknown account. Execution revalidates that same history capability for resumed turns and fresh recovery.</p><h2>Why It Matters</h2><p>Long-running agent work often crosses messy boundaries: browser refreshes, Gateway restarts, interrupted turns, delayed recovery, and a human coming back later with a short follow-up. \"Continue\" only works if OpenClaw can bridge what the UI remembers with what the model is allowed to see.</p><p>Without that bridge, retained pending inputs become a display artifact instead of recoverable context. With it, built-in and CLI agent turns can understand the interrupted request in a constrained way.</p><p>The PR's contract is careful:</p><ul><li>Current request and transcript bytes stay unchanged.</li><li>Retained inputs are previewed, not replayed.</li><li>Cancelled and context-excluded messages remain excluded.</li><li>Cross-account or unverifiable-account cases do not leak previews.</li><li>Revoked authority prevents resumed transport dispatch.</li></ul><p>That combination is the heart of the fix: useful recovery without making interruption state a shortcut around normal access rules.</p><h2>Evidence From The PR</h2><p>The PR includes a real SQLite accepted-input regression. Before the fix, the production prompt-assembly boundary produced empty continuation context. After the repair, it exposed the interrupted request without consuming or replaying it.</p><p>The final proof covered cross-account and unverifiable-account leak cases, resumed dispatch after authority revocation, raw prepared input capture, and same-account previews. The focused suite passed 53 tests across four files, along with production types, test-type groups, lint, SQLite-worker guards, test-race checks, mock guards, and independent review.</p><p>There was also a later correction for CLI media-suffix ordering, keeping interrupted-request context before an existing media snapshot so media progress remains in the private context suffix.</p><h2>Bottom Line</h2><p>PR #168279 makes OpenClaw more dependable when work is interrupted midstream. If the UI still has a retained request and the session authority still checks out, agents get bounded context to continue the work instead of losing the thread.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-interrupted-request-continuation.png\" alt=\"OpenClaw Recovers Interrupted Requests on Continue\" /></p>",
      "date_published": "2026-10-10T08:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "Recovers",
        "Interrupted",
        "Requests",
        "Continue"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-interrupted-request-continuation.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-interrupted-request-continuation.png",
          "mime_type": "image/png",
          "title": "OpenClaw Recovers Interrupted Requests on Continue cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-10-windows-plugin-startup-fix/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-10-windows-plugin-startup-fix/",
      "title": "OpenClaw Fixes Windows Plugin Startup Stalls",
      "summary": "OpenClaw PR #168313 fixes a Windows Gateway regression where installed plugins could delay startup by several minutes.",
      "content_text": "OpenClaw merged [PR #168313](https://github.com/openclaw/openclaw/pull/168313), a Windows Gateway startup fix for machines with installed plugins.\n\nThe regression was serious enough to block release validation for `2026.10.5-beta.1`. According to the PR, Windows installer-fresh lanes were timing out because plugin source capture could spend minutes copying installed plugin files before the Gateway bound its port. Affected candidates included `2026.10.1` and `2026.10.5-beta.1`.\n\n## What Changed\n\nThe slowdown came from a guarded clone path introduced during plugin source capture hardening. On Windows, that guarded clone skips native copies and path-admission caching, making it much slower for ordinary plugin files. With installed plugins present, the startup path could stretch from a normal startup into a multi-minute wait.\n\nOpenClaw now uses the existing pinned-descriptor copy path again for small ordinary files on Windows. Large files and native files still keep the guarded clone path, preserving the safety boundary where it matters most.\n\nThe PR also moves the small-copy guard tests onto Windows, so the intended behavior is covered on the platform that regressed.\n\n## Why It Matters\n\nGateway startup is one of the few flows where a performance bug can feel like a total outage. If the process has not bound yet, the user does not see a slightly slower feature; they see OpenClaw failing to start, reconnect, or pass release checks.\n\nThis fix matters most for Windows users who install plugins and then rely on OpenClaw to recover cleanly after updates, restarts, or fresh installs. The PR reports affected candidates taking four to ten minutes to bind, while passing runs bind in about 42 seconds.\n\nThat difference is not cosmetic. It decides whether the release lane sees a working Gateway or a timeout.\n\n## Release Context\n\nPR #168313 is a forward-port to `main` of an equivalent fix that already landed on the `release/2026.10.1` branch in PR #168256. That matters because the release branch had already proven the same shape of repair, while `main` still carried the regression.\n\nThe user-facing impact is intentionally narrow:\n\n- Windows Gateways with installed plugins start without multi-minute source-capture stalls.\n- Small ordinary plugin files use the faster pinned-descriptor copy path.\n- Large and native files retain guarded clone behavior.\n- No schema, configuration, or plugin manifest change is involved.\n\n## Evidence From The PR\n\nThe PR cites the failing cross-OS installer-fresh release lane and identifies the last known passing run. It also reports a local macOS pass for `src/plugins/plugin-source-file.copy.test.ts`, while the Windows proof comes from CI and the release-validation lane after the backport.\n\nBecause the issue was platform-specific, that Windows release-lane evidence is the important part. The fix is not a broad plugin rewrite; it restores the correct copy strategy for the Windows startup case that broke.\n\n## Bottom Line\n\nPR #168313 removes a release-blocking Windows startup stall for plugin-heavy OpenClaw installs. For users, the visible result should be simple: the Gateway gets back to binding promptly instead of disappearing into minutes of plugin file capture.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168313\">PR #168313</a>, a Windows Gateway startup fix for machines with installed plugins.</p><p>The regression was serious enough to block release validation for <code>2026.10.5-beta.1</code>. According to the PR, Windows installer-fresh lanes were timing out because plugin source capture could spend minutes copying installed plugin files before the Gateway bound its port. Affected candidates included <code>2026.10.1</code> and <code>2026.10.5-beta.1</code>.</p><h2>What Changed</h2><p>The slowdown came from a guarded clone path introduced during plugin source capture hardening. On Windows, that guarded clone skips native copies and path-admission caching, making it much slower for ordinary plugin files. With installed plugins present, the startup path could stretch from a normal startup into a multi-minute wait.</p><p>OpenClaw now uses the existing pinned-descriptor copy path again for small ordinary files on Windows. Large files and native files still keep the guarded clone path, preserving the safety boundary where it matters most.</p><p>The PR also moves the small-copy guard tests onto Windows, so the intended behavior is covered on the platform that regressed.</p><h2>Why It Matters</h2><p>Gateway startup is one of the few flows where a performance bug can feel like a total outage. If the process has not bound yet, the user does not see a slightly slower feature; they see OpenClaw failing to start, reconnect, or pass release checks.</p><p>This fix matters most for Windows users who install plugins and then rely on OpenClaw to recover cleanly after updates, restarts, or fresh installs. The PR reports affected candidates taking four to ten minutes to bind, while passing runs bind in about 42 seconds.</p><p>That difference is not cosmetic. It decides whether the release lane sees a working Gateway or a timeout.</p><h2>Release Context</h2><p>PR #168313 is a forward-port to <code>main</code> of an equivalent fix that already landed on the <code>release/2026.10.1</code> branch in PR #168256. That matters because the release branch had already proven the same shape of repair, while <code>main</code> still carried the regression.</p><p>The user-facing impact is intentionally narrow:</p><ul><li>Windows Gateways with installed plugins start without multi-minute source-capture stalls.</li><li>Small ordinary plugin files use the faster pinned-descriptor copy path.</li><li>Large and native files retain guarded clone behavior.</li><li>No schema, configuration, or plugin manifest change is involved.</li></ul><h2>Evidence From The PR</h2><p>The PR cites the failing cross-OS installer-fresh release lane and identifies the last known passing run. It also reports a local macOS pass for <code>src/plugins/plugin-source-file.copy.test.ts</code>, while the Windows proof comes from CI and the release-validation lane after the backport.</p><p>Because the issue was platform-specific, that Windows release-lane evidence is the important part. The fix is not a broad plugin rewrite; it restores the correct copy strategy for the Windows startup case that broke.</p><h2>Bottom Line</h2><p>PR #168313 removes a release-blocking Windows startup stall for plugin-heavy OpenClaw installs. For users, the visible result should be simple: the Gateway gets back to binding promptly instead of disappearing into minutes of plugin file capture.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-windows-plugin-startup-fix.png\" alt=\"OpenClaw Fixes Windows Plugin Startup Stalls\" /></p>",
      "date_published": "2026-10-10T08:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Windows",
        "Plugin",
        "Startup",
        "Stalls"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-windows-plugin-startup-fix.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-windows-plugin-startup-fix.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Windows Plugin Startup Stalls cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-10-ollama-thinking-levels/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-10-ollama-thinking-levels/",
      "title": "OpenClaw Maps Ollama Thinking Levels Correctly",
      "summary": "OpenClaw PR #168166 sends Ollama models only the thinking values they advertise, keeping local reasoning controls aligned with discovery.",
      "content_text": "OpenClaw merged [PR #168166](https://github.com/openclaw/openclaw/pull/168166), a local-model routing fix that makes Ollama thinking controls respect the exact values each model reports during discovery.\n\nThe issue sat in a subtle place: OpenClaw already fetched `/api/show` metadata from Ollama, but it dropped the model's thinking descriptor before request time. That meant native Ollama requests could receive values a model did not advertise. The PR body calls out two concrete examples: Qwen could receive the string `high` even though it advertised boolean thinking, while GPT-OSS could receive `false` even though it advertised tiered values like `low`, `medium`, and `high`.\n\n## What Changed\n\nOpenClaw now caches native thinking mappings from Ollama discovery and carries them through the model catalog, setup, runtime projection, and final native stream request. The shared thinking ladder still drives the user-facing choices, but the final payload is shaped for the specific model.\n\nIn practice, that means:\n\n- Boolean models receive boolean thinking values.\n- Graded models receive supported effort strings.\n- Unsupported Minimal and XHigh controls are exposed only when advertised.\n- Models that cannot disable thinking map Off to the model's lowest supported value.\n- Explicit `params.think` and `params.thinking` remain respected until runtime selections intentionally override them.\n\nThe OpenAI-compatible Ollama route keeps its older behavior and does not receive native boolean mappings.\n\n## Why It Matters\n\nLocal model providers are becoming less uniform. One Ollama model may represent reasoning as a simple on/off switch, while another exposes a ladder of effort levels. A single generic value can look reasonable in OpenClaw and still be wrong for the model behind the request.\n\nThis fix makes the model picker and request layer behave more like a contract with discovery data. If the model says it supports `low`, `medium`, and `high`, OpenClaw sends those. If it says thinking is boolean, OpenClaw sends `true` or `false`. That is a small but important reliability win for people running local models through OpenClaw rather than a hosted API.\n\nIt also keeps saved settings safer. The PR preserves mandatory-thinking floors for models that require some reasoning setting, while avoiding surprise changes to existing stored configuration.\n\n## Evidence From The PR\n\nThe maintainers tested against Ollama 0.40.2 with a local Gateway profile, a `models.list` refresh, and real `chat.send` messages containing `/think <level>` followed by a strict pong request. A local proxy captured native `/api/chat` requests.\n\nThe proof covered `qwen3:8b`, advertised as boolean thinking, and `gpt-oss:20b`, advertised with low, medium, and high effort values. All eight tested requests completed. The PR also notes there were zero `/api/show` calls after catalog refresh, so turns reuse cached discovery rather than adding request-time discovery overhead.\n\nRegression tests cover descriptor validation, mapping, projection, native-stream behavior, legacy transport isolation, malformed descriptors, mandatory-thinking floors, and the formerly rejected XHigh profile entry.\n\n## Bottom Line\n\nPR #168166 makes OpenClaw's Ollama integration less guessy. Local reasoning controls now follow the model's advertised contract, which should reduce odd request failures and make advanced local-model settings feel more predictable.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/168166\">PR #168166</a>, a local-model routing fix that makes Ollama thinking controls respect the exact values each model reports during discovery.</p><p>The issue sat in a subtle place: OpenClaw already fetched <code>/api/show</code> metadata from Ollama, but it dropped the model's thinking descriptor before request time. That meant native Ollama requests could receive values a model did not advertise. The PR body calls out two concrete examples: Qwen could receive the string <code>high</code> even though it advertised boolean thinking, while GPT-OSS could receive <code>false</code> even though it advertised tiered values like <code>low</code>, <code>medium</code>, and <code>high</code>.</p><h2>What Changed</h2><p>OpenClaw now caches native thinking mappings from Ollama discovery and carries them through the model catalog, setup, runtime projection, and final native stream request. The shared thinking ladder still drives the user-facing choices, but the final payload is shaped for the specific model.</p><p>In practice, that means:</p><ul><li>Boolean models receive boolean thinking values.</li><li>Graded models receive supported effort strings.</li><li>Unsupported Minimal and XHigh controls are exposed only when advertised.</li><li>Models that cannot disable thinking map Off to the model's lowest supported value.</li><li>Explicit <code>params.think</code> and <code>params.thinking</code> remain respected until runtime selections intentionally override them.</li></ul><p>The OpenAI-compatible Ollama route keeps its older behavior and does not receive native boolean mappings.</p><h2>Why It Matters</h2><p>Local model providers are becoming less uniform. One Ollama model may represent reasoning as a simple on/off switch, while another exposes a ladder of effort levels. A single generic value can look reasonable in OpenClaw and still be wrong for the model behind the request.</p><p>This fix makes the model picker and request layer behave more like a contract with discovery data. If the model says it supports <code>low</code>, <code>medium</code>, and <code>high</code>, OpenClaw sends those. If it says thinking is boolean, OpenClaw sends <code>true</code> or <code>false</code>. That is a small but important reliability win for people running local models through OpenClaw rather than a hosted API.</p><p>It also keeps saved settings safer. The PR preserves mandatory-thinking floors for models that require some reasoning setting, while avoiding surprise changes to existing stored configuration.</p><h2>Evidence From The PR</h2><p>The maintainers tested against Ollama 0.40.2 with a local Gateway profile, a <code>models.list</code> refresh, and real <code>chat.send</code> messages containing <code>/think <level></code> followed by a strict pong request. A local proxy captured native <code>/api/chat</code> requests.</p><p>The proof covered <code>qwen3:8b</code>, advertised as boolean thinking, and <code>gpt-oss:20b</code>, advertised with low, medium, and high effort values. All eight tested requests completed. The PR also notes there were zero <code>/api/show</code> calls after catalog refresh, so turns reuse cached discovery rather than adding request-time discovery overhead.</p><p>Regression tests cover descriptor validation, mapping, projection, native-stream behavior, legacy transport isolation, malformed descriptors, mandatory-thinking floors, and the formerly rejected XHigh profile entry.</p><h2>Bottom Line</h2><p>PR #168166 makes OpenClaw's Ollama integration less guessy. Local reasoning controls now follow the model's advertised contract, which should reduce odd request failures and make advanced local-model settings feel more predictable.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-ollama-thinking-levels.png\" alt=\"OpenClaw Maps Ollama Thinking Levels Correctly\" /></p>",
      "date_published": "2026-10-10T08:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.482Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Local Models",
        "Maps",
        "Ollama",
        "Thinking",
        "Levels",
        "Correctly"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-ollama-thinking-levels.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-10-ollama-thinking-levels.png",
          "mime_type": "image/png",
          "title": "OpenClaw Maps Ollama Thinking Levels Correctly cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-9-cron-outage-alert-hold/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-9-cron-outage-alert-hold/",
      "title": "OpenClaw Calms Cron Alerts During Outages",
      "summary": "OpenClaw PR #167957 delays cron repair requests and alerts while provider-outage retries are still pending.",
      "content_text": "OpenClaw merged [PR #167957](https://github.com/openclaw/openclaw/pull/167957), a cron reliability change that makes scheduled automations less noisy during short provider outages.\n\nThe issue was timing. A recurring automation that hit a temporary provider failure, such as DNS trouble, a refused connection, a 5xx response, overload, or a rate limit, could spend its owner repair request or send a failure alert while OpenClaw's own retry ladder was still pending. If the retry succeeded shortly after, the user had already been interrupted.\n\n## What Changed\n\nOpenClaw's cron outcome path now records when a transient provider failure still has a retry pending. That signal is passed into the failure-notification flow, which holds the repair request or alert for that cycle.\n\nThe hold is bounded. It lasts only while the transient retry budget remains, including the quick retry ladder and ordinary backoff when a short-interval job's next slot arrives first. The PR describes the quick retry cadence as 30 seconds, one minute, and five minutes.\n\nThe hold also has important exclusions. OpenClaw does not delay notifications when there is no next run, when a one-shot job is retired after retries, when a job is disabled, or when the failure is caused by the job itself. Cron watchdog timeouts, helper-script failures, script timeouts, and command-job timeouts keep their existing repair timing.\n\n## Why It Matters\n\nAutomation alerts are useful only when they arrive at the right moment. Too early, and they train users to ignore noise. Too late, and they hide real breakage. This change is a pragmatic middle ground for provider outages: let the built-in retry plan breathe first, then ask for help if the failure persists.\n\nFor owners of scheduled OpenClaw jobs, the behavior should now be calmer:\n\n- Short provider outages can recover silently.\n- Repair requests are not spent while a retry is already queued.\n- Failure alerts still arrive if the retry budget is exhausted.\n- Job-owned failures are still surfaced promptly.\n- No config key, schema migration, or stored-state change is required.\n\nThis matters for any installation that depends on scheduled summaries, reminders, checks, or outbound updates. Provider blips happen; the automation system should not treat every transient network stumble as a durable broken job.\n\n## Evidence From The PR\n\nThe PR reports unit coverage for owned, unowned, hourly, weekly, every-minute, disabled, one-shot, and no-future-slot cases. In the DNS-failure cases, owned and unowned jobs sent no repairs or alerts through failure three, then sent one repair or alert after failure four if the outage persisted.\n\nThe live proof used an isolated Gateway and a mock provider that dropped TCP connections for the job's model requests. On the fixed branch, an owned hourly job produced no owner repair or alert through the first three failures, then sent one repair after the fourth. On main, the same setup sent a repair after failure two and an alert after failure three.\n\n## Bottom Line\n\nPR #167957 makes OpenClaw cron jobs a little more patient in exactly the cases where patience helps. It does not hide permanent failures, and it does not soften job-caused failures. It simply waits for OpenClaw's own transient retry budget before waking the owner.\n\nThat should make automation alerts feel more like useful signals and less like turbulence reports.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167957\">PR #167957</a>, a cron reliability change that makes scheduled automations less noisy during short provider outages.</p><p>The issue was timing. A recurring automation that hit a temporary provider failure, such as DNS trouble, a refused connection, a 5xx response, overload, or a rate limit, could spend its owner repair request or send a failure alert while OpenClaw's own retry ladder was still pending. If the retry succeeded shortly after, the user had already been interrupted.</p><h2>What Changed</h2><p>OpenClaw's cron outcome path now records when a transient provider failure still has a retry pending. That signal is passed into the failure-notification flow, which holds the repair request or alert for that cycle.</p><p>The hold is bounded. It lasts only while the transient retry budget remains, including the quick retry ladder and ordinary backoff when a short-interval job's next slot arrives first. The PR describes the quick retry cadence as 30 seconds, one minute, and five minutes.</p><p>The hold also has important exclusions. OpenClaw does not delay notifications when there is no next run, when a one-shot job is retired after retries, when a job is disabled, or when the failure is caused by the job itself. Cron watchdog timeouts, helper-script failures, script timeouts, and command-job timeouts keep their existing repair timing.</p><h2>Why It Matters</h2><p>Automation alerts are useful only when they arrive at the right moment. Too early, and they train users to ignore noise. Too late, and they hide real breakage. This change is a pragmatic middle ground for provider outages: let the built-in retry plan breathe first, then ask for help if the failure persists.</p><p>For owners of scheduled OpenClaw jobs, the behavior should now be calmer:</p><ul><li>Short provider outages can recover silently.</li><li>Repair requests are not spent while a retry is already queued.</li><li>Failure alerts still arrive if the retry budget is exhausted.</li><li>Job-owned failures are still surfaced promptly.</li><li>No config key, schema migration, or stored-state change is required.</li></ul><p>This matters for any installation that depends on scheduled summaries, reminders, checks, or outbound updates. Provider blips happen; the automation system should not treat every transient network stumble as a durable broken job.</p><h2>Evidence From The PR</h2><p>The PR reports unit coverage for owned, unowned, hourly, weekly, every-minute, disabled, one-shot, and no-future-slot cases. In the DNS-failure cases, owned and unowned jobs sent no repairs or alerts through failure three, then sent one repair or alert after failure four if the outage persisted.</p><p>The live proof used an isolated Gateway and a mock provider that dropped TCP connections for the job's model requests. On the fixed branch, an owned hourly job produced no owner repair or alert through the first three failures, then sent one repair after the fourth. On main, the same setup sent a repair after failure two and an alert after failure three.</p><h2>Bottom Line</h2><p>PR #167957 makes OpenClaw cron jobs a little more patient in exactly the cases where patience helps. It does not hide permanent failures, and it does not soften job-caused failures. It simply waits for OpenClaw's own transient retry budget before waking the owner.</p><p>That should make automation alerts feel more like useful signals and less like turbulence reports.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-cron-outage-alert-hold.png\" alt=\"OpenClaw Calms Cron Alerts During Outages\" /></p>",
      "date_published": "2026-10-09T23:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Calms",
        "Cron",
        "Alerts",
        "During",
        "Outages"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-cron-outage-alert-hold.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-cron-outage-alert-hold.png",
          "mime_type": "image/png",
          "title": "OpenClaw Calms Cron Alerts During Outages cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-9-ollama-tool-search-fix/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-9-ollama-tool-search-fix/",
      "title": "OpenClaw Fixes Ollama Tool Search Calls",
      "summary": "OpenClaw PR #167933 repairs Ollama tool-call routing when Tool Search is enabled, preserving canonical tool history.",
      "content_text": "OpenClaw merged [PR #167933](https://github.com/openclaw/openclaw/pull/167933), a P1 Ollama compatibility fix for a sharp edge between local models, Tool Search, and native tool calls.\n\nThe failure mode was specific but nasty. With Tool Search enabled, Ollama could turn valid native calls such as `exec` into malformed `tool_call` invocations. The model intended to call one tool, but the transport parser latched onto an envelope-like name first, causing the turn to fail before the real tool could run.\n\n## What Changed\n\nOpenClaw now aliases parser-sensitive tool names at the Ollama transport boundary. The important detail is where that aliasing lives: execution, policy decisions, catalog IDs, stored history, and replay keep the canonical OpenClaw tool names. Only the wire shape offered to Ollama changes.\n\nThat keeps the repair narrow. OpenClaw does not rewrite the user's system prompt, does not mutate core prompt sections, and does not store alias names as if they were the original tools. The adapter translates definitions, assistant replay, named tool results, selectors, streaming events, final messages, and result consumers across the same boundary.\n\nThe PR calls out several parser-sensitive names that need protection, including `tool_call`, `tool_calls`, `TOOL_CALLS`, `function`, `name`, `arguments`, and short substrings such as `ls` that can collide with known envelope formats.\n\n## Why It Matters\n\nOllama is one of the most important local-model routes for OpenClaw users. Tool Search is also a natural fit for local models because it lets the assistant discover a smaller tool surface instead of carrying every possible tool definition all the time.\n\nWhen those two features collide, the result can feel arbitrary: a simple request like running `echo hello` fails even though the model produced something close to a valid tool call. This fix keeps local model workflows usable without weakening OpenClaw's internal tool identity rules.\n\nThe user-facing impact is direct:\n\n- Qwen3:8b and Qwen2.5:7b can execute real Gateway tools with default Tool Search.\n- Ollama OpenAI-compatible chat-completions routing gets the same repair.\n- Stored session history keeps canonical tool names.\n- Policy and approval checks continue to see the real OpenClaw tool identities.\n- No session migration or operator configuration is required.\n\nThe PR is also careful about limits. This is a compatibility repair for known envelope collisions, not a universal recovery system for arbitrary malformed tool intent.\n\n## Evidence From The PR\n\nThe PR includes live Gateway probes against Ollama 0.40.1 with isolated state, default Tool Search, and local models. On main, Qwen3:8b and Qwen2.5:7b misparsed native wrapper calls for the `echo hello` prompt. On the fixed branch, both produced one `exec` call, returned the actual tool result, and delivered a visible reply with zero failures.\n\nDeferred Tool Search smokes also passed for `agents_list`, and a Llama3.2 OpenAI-compatible route successfully executed `exec`. The PR reports 80 passing alias and plugin-boundary tests, a passing runtime build, and a clean scoped review through P2.\n\n## Bottom Line\n\nPR #167933 makes OpenClaw's Ollama path more dependable for tool-using local agents. The fix respects the boundary between provider wire quirks and OpenClaw's canonical tool system, which is exactly where this kind of compatibility code belongs.\n\nFor users running local models with Tool Search, this should turn a confusing parser failure into ordinary tool execution.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167933\">PR #167933</a>, a P1 Ollama compatibility fix for a sharp edge between local models, Tool Search, and native tool calls.</p><p>The failure mode was specific but nasty. With Tool Search enabled, Ollama could turn valid native calls such as <code>exec</code> into malformed <code>tool_call</code> invocations. The model intended to call one tool, but the transport parser latched onto an envelope-like name first, causing the turn to fail before the real tool could run.</p><h2>What Changed</h2><p>OpenClaw now aliases parser-sensitive tool names at the Ollama transport boundary. The important detail is where that aliasing lives: execution, policy decisions, catalog IDs, stored history, and replay keep the canonical OpenClaw tool names. Only the wire shape offered to Ollama changes.</p><p>That keeps the repair narrow. OpenClaw does not rewrite the user's system prompt, does not mutate core prompt sections, and does not store alias names as if they were the original tools. The adapter translates definitions, assistant replay, named tool results, selectors, streaming events, final messages, and result consumers across the same boundary.</p><p>The PR calls out several parser-sensitive names that need protection, including <code>tool_call</code>, <code>tool_calls</code>, <code>TOOL_CALLS</code>, <code>function</code>, <code>name</code>, <code>arguments</code>, and short substrings such as <code>ls</code> that can collide with known envelope formats.</p><h2>Why It Matters</h2><p>Ollama is one of the most important local-model routes for OpenClaw users. Tool Search is also a natural fit for local models because it lets the assistant discover a smaller tool surface instead of carrying every possible tool definition all the time.</p><p>When those two features collide, the result can feel arbitrary: a simple request like running <code>echo hello</code> fails even though the model produced something close to a valid tool call. This fix keeps local model workflows usable without weakening OpenClaw's internal tool identity rules.</p><p>The user-facing impact is direct:</p><ul><li>Qwen3:8b and Qwen2.5:7b can execute real Gateway tools with default Tool Search.</li><li>Ollama OpenAI-compatible chat-completions routing gets the same repair.</li><li>Stored session history keeps canonical tool names.</li><li>Policy and approval checks continue to see the real OpenClaw tool identities.</li><li>No session migration or operator configuration is required.</li></ul><p>The PR is also careful about limits. This is a compatibility repair for known envelope collisions, not a universal recovery system for arbitrary malformed tool intent.</p><h2>Evidence From The PR</h2><p>The PR includes live Gateway probes against Ollama 0.40.1 with isolated state, default Tool Search, and local models. On main, Qwen3:8b and Qwen2.5:7b misparsed native wrapper calls for the <code>echo hello</code> prompt. On the fixed branch, both produced one <code>exec</code> call, returned the actual tool result, and delivered a visible reply with zero failures.</p><p>Deferred Tool Search smokes also passed for <code>agents_list</code>, and a Llama3.2 OpenAI-compatible route successfully executed <code>exec</code>. The PR reports 80 passing alias and plugin-boundary tests, a passing runtime build, and a clean scoped review through P2.</p><h2>Bottom Line</h2><p>PR #167933 makes OpenClaw's Ollama path more dependable for tool-using local agents. The fix respects the boundary between provider wire quirks and OpenClaw's canonical tool system, which is exactly where this kind of compatibility code belongs.</p><p>For users running local models with Tool Search, this should turn a confusing parser failure into ordinary tool execution.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-ollama-tool-search-fix.png\" alt=\"OpenClaw Fixes Ollama Tool Search Calls\" /></p>",
      "date_published": "2026-10-09T23:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Local Models",
        "Fixes",
        "Ollama",
        "Tool",
        "Search",
        "Calls"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-ollama-tool-search-fix.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-ollama-tool-search-fix.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Ollama Tool Search Calls cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-9-chatgpt-model-picker-availability/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-9-chatgpt-model-picker-availability/",
      "title": "OpenClaw Tightens ChatGPT Model Availability",
      "summary": "OpenClaw PR #167975 keeps ChatGPT-only users from selecting pro OpenAI models their signed-in account does not actually list.",
      "content_text": "OpenClaw merged [PR #167975](https://github.com/openclaw/openclaw/pull/167975), a focused OpenAI model-catalog fix for users who sign in with a ChatGPT account but do not also have an OpenAI API key configured.\n\nThe problem was easy to miss until a user picked the wrong model. The default model list could show `openai/gpt-5.4-pro` and `openai/gpt-5.5-pro` for a ChatGPT-only account even when that account's live model listing did not include those models. Selecting one of them then failed later, after the UI had already implied the model was available.\n\n## What Changed\n\nOpenClaw now records the exact model IDs returned by a successful ChatGPT account listing as private catalog provenance. That list includes hidden rows when the account is entitled to them, but it is not projected publicly through the normal provider outcomes view.\n\nThe picker then uses that account-specific list when deciding whether a model should be available through the selected ChatGPT profile. If a model can be served by both a subscription route and an API-key route, but the signed-in ChatGPT account did not list it, OpenClaw no longer treats the subscription route as usable. The model remains available when a working non-subscription credential, such as an OpenAI API key, can serve it.\n\nThe PR also keeps the check scoped to the selected profile. One account's model listing does not narrow or widen another account's picker. When OpenClaw does not have a ready listing for a profile, it keeps the existing behavior rather than inventing a denial.\n\n## Why It Matters\n\nModel pickers are trust surfaces. If the UI says a model is available, users expect the next request to work or at least fail for a runtime reason, not because the picker offered something the account could never access.\n\nThis is especially relevant as OpenClaw supports more provider routes, account types, hidden entitlement rows, and plugin-discovered catalog entries. A single model ID can exist in several places, but the availability decision still has to respect the credential that will actually handle the request.\n\nFor ChatGPT-only users, the practical result is cleaner:\n\n- Unlisted pro models stop appearing as usable default choices.\n- Hidden models that the account actually lists remain available.\n- Users with an OpenAI API key still see API-key-served pro models.\n- Unavailable models show a neutral unavailable state instead of misleading sign-in guidance.\n\nThe change is intentionally limited to the picker. If an operator manually types an unlisted ID, the request can still reach OpenAI and fail with OpenAI's own response.\n\n## Evidence From The PR\n\nThe PR includes a real ChatGPT account comparison on a loopback Gateway. Before the fix, the default provider view showed 11 rows for a ChatGPT-only profile: eight listed rows plus `gpt-5.4-pro`, `gpt-5.5`, and `gpt-5.5-pro`. After the fix, the view showed nine rows: the eight listed rows plus `gpt-5.5`.\n\nThe API-key control case still showed both pro models as present and available. Regression tests cover unlisted dual-route IDs, hidden listed IDs, API-key fallback, unavailable listings, and multi-account profile isolation.\n\n## Bottom Line\n\nPR #167975 makes OpenClaw's OpenAI picker more honest about what a ChatGPT account can actually run. It is not a flashy feature, but it removes a confusing failure path right where users make their model choice.\n\nFor anyone using ChatGPT sign-in without an API key, this should make the default model list feel less optimistic and more accurate.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167975\">PR #167975</a>, a focused OpenAI model-catalog fix for users who sign in with a ChatGPT account but do not also have an OpenAI API key configured.</p><p>The problem was easy to miss until a user picked the wrong model. The default model list could show <code>openai/gpt-5.4-pro</code> and <code>openai/gpt-5.5-pro</code> for a ChatGPT-only account even when that account's live model listing did not include those models. Selecting one of them then failed later, after the UI had already implied the model was available.</p><h2>What Changed</h2><p>OpenClaw now records the exact model IDs returned by a successful ChatGPT account listing as private catalog provenance. That list includes hidden rows when the account is entitled to them, but it is not projected publicly through the normal provider outcomes view.</p><p>The picker then uses that account-specific list when deciding whether a model should be available through the selected ChatGPT profile. If a model can be served by both a subscription route and an API-key route, but the signed-in ChatGPT account did not list it, OpenClaw no longer treats the subscription route as usable. The model remains available when a working non-subscription credential, such as an OpenAI API key, can serve it.</p><p>The PR also keeps the check scoped to the selected profile. One account's model listing does not narrow or widen another account's picker. When OpenClaw does not have a ready listing for a profile, it keeps the existing behavior rather than inventing a denial.</p><h2>Why It Matters</h2><p>Model pickers are trust surfaces. If the UI says a model is available, users expect the next request to work or at least fail for a runtime reason, not because the picker offered something the account could never access.</p><p>This is especially relevant as OpenClaw supports more provider routes, account types, hidden entitlement rows, and plugin-discovered catalog entries. A single model ID can exist in several places, but the availability decision still has to respect the credential that will actually handle the request.</p><p>For ChatGPT-only users, the practical result is cleaner:</p><ul><li>Unlisted pro models stop appearing as usable default choices.</li><li>Hidden models that the account actually lists remain available.</li><li>Users with an OpenAI API key still see API-key-served pro models.</li><li>Unavailable models show a neutral unavailable state instead of misleading sign-in guidance.</li></ul><p>The change is intentionally limited to the picker. If an operator manually types an unlisted ID, the request can still reach OpenAI and fail with OpenAI's own response.</p><h2>Evidence From The PR</h2><p>The PR includes a real ChatGPT account comparison on a loopback Gateway. Before the fix, the default provider view showed 11 rows for a ChatGPT-only profile: eight listed rows plus <code>gpt-5.4-pro</code>, <code>gpt-5.5</code>, and <code>gpt-5.5-pro</code>. After the fix, the view showed nine rows: the eight listed rows plus <code>gpt-5.5</code>.</p><p>The API-key control case still showed both pro models as present and available. Regression tests cover unlisted dual-route IDs, hidden listed IDs, API-key fallback, unavailable listings, and multi-account profile isolation.</p><h2>Bottom Line</h2><p>PR #167975 makes OpenClaw's OpenAI picker more honest about what a ChatGPT account can actually run. It is not a flashy feature, but it removes a confusing failure path right where users make their model choice.</p><p>For anyone using ChatGPT sign-in without an API key, this should make the default model list feel less optimistic and more accurate.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-chatgpt-model-picker-availability.png\" alt=\"OpenClaw Tightens ChatGPT Model Availability\" /></p>",
      "date_published": "2026-10-09T23:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Tightens",
        "Chatgpt",
        "Model",
        "Availability"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-chatgpt-model-picker-availability.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-chatgpt-model-picker-availability.png",
          "mime_type": "image/png",
          "title": "OpenClaw Tightens ChatGPT Model Availability cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-9-image-worker-burst-capacity/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-9-image-worker-burst-capacity/",
      "title": "OpenClaw Image Workers Get Burst Capacity",
      "summary": "OpenClaw PR #167676 lets image bursts use two bounded workers while keeping Control UI file reads on their own admission path.",
      "content_text": "OpenClaw merged [PR #167676](https://github.com/openclaw/openclaw/pull/167676), a worker-pool change that gives image-heavy bursts more breathing room without making operators tune new settings.\n\nThe issue was straightforward: concurrent image transforms queued behind a single worker. At the same time, Control UI asset reads could wait for shared compute permits that were occupied by unrelated work. That combination could make image generation feel slower and make simple UI reads wait behind heavier tasks.\n\n## What Changed\n\nImage bursts can now use two workers inside the existing shared CPU budget. The second image worker is not permanent capacity. Surplus image workers retire after five idle seconds, while the first usable worker keeps its normal idle lifecycle.\n\nControl UI reads also get clearer separation. They use their existing bounded file pool independently of compute admission, so file reads do not have to wait behind occupied shared compute permits in the same way.\n\nThe PR says no operator configuration or migration is needed. The worker-pool retirement owner handles surplus deadlines and promotion. A worker that is already stopping still counts against capacity until native exit, but it no longer qualifies as the retained idle worker.\n\n## Why This Matters\n\nOpenClaw now has more workflows that touch image and file processing: generated media, previews, uploaded files, Control UI assets, and plugin outputs. A single-worker image lane can be fine for isolated tasks, but bursts quickly expose the queue.\n\nThe important part of this change is that it improves burst behavior without leaving extra workers sitting around indefinitely. That is a healthier tradeoff for self-hosted and local-first installations, where memory and CPU budget matter as much as peak speed.\n\nOperators should expect the practical effect to show up in a few places:\n\n- Image bursts can complete faster when multiple transforms arrive together.\n- Control UI file reads are less tied to shared compute pressure.\n- Surplus worker capacity retires after idle time.\n- No new configuration or schema migration is required.\n\nThis is not a rewrite of worker execution. The existing dispatch, FIFO admission, and native settlement machinery still own execution and cleanup.\n\n## The Measured Result\n\nThe PR reports a Linux benchmark with compiled production entrypoints processing two 1536x1536 JPEG encodes, six 64x64 JPEG encodes, and 24 uncached 256 KiB assets plus sidecars per burst. Warm results covered 12 bursts, and encoded-image plus asset byte digests matched across every comparison.\n\nWarm image queue-wait p99 fell from 209.99 ms to 103.54 ms. Spaced image queue-wait p99 fell from 204.36 ms to 108.89 ms. The complete warm mixed workload improved from 2448.42 ms to 1287.70 ms, reported as 1.90x faster.\n\nThere is a tradeoff: warm ordinary file queue-wait p99 increased from 13.90 ms to 25.76 ms, and cold file queue-wait p99 increased from 112.99 ms to 153.02 ms in that four-CPU workload. But when two shared compute permits were held, file queue-wait p99 improved sharply from 207.04 ms to 19.01 ms.\n\nIdle worker heap stayed essentially flat after about six idle seconds, at 47.20 MiB before and 47.23 MiB after. That supports the core design: add short-lived burst capacity, then retire it.\n\n## Bottom Line\n\nPR #167676 is a focused throughput improvement for OpenClaw's media and file-heavy paths. It lets image work scale modestly during bursts, keeps Control UI reads from being dragged through the wrong queue, and avoids a permanent resource increase.\n\nFor users leaning on generated media or asset-heavy plugin workflows, this should make peak moments feel less clogged.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167676\">PR #167676</a>, a worker-pool change that gives image-heavy bursts more breathing room without making operators tune new settings.</p><p>The issue was straightforward: concurrent image transforms queued behind a single worker. At the same time, Control UI asset reads could wait for shared compute permits that were occupied by unrelated work. That combination could make image generation feel slower and make simple UI reads wait behind heavier tasks.</p><h2>What Changed</h2><p>Image bursts can now use two workers inside the existing shared CPU budget. The second image worker is not permanent capacity. Surplus image workers retire after five idle seconds, while the first usable worker keeps its normal idle lifecycle.</p><p>Control UI reads also get clearer separation. They use their existing bounded file pool independently of compute admission, so file reads do not have to wait behind occupied shared compute permits in the same way.</p><p>The PR says no operator configuration or migration is needed. The worker-pool retirement owner handles surplus deadlines and promotion. A worker that is already stopping still counts against capacity until native exit, but it no longer qualifies as the retained idle worker.</p><h2>Why This Matters</h2><p>OpenClaw now has more workflows that touch image and file processing: generated media, previews, uploaded files, Control UI assets, and plugin outputs. A single-worker image lane can be fine for isolated tasks, but bursts quickly expose the queue.</p><p>The important part of this change is that it improves burst behavior without leaving extra workers sitting around indefinitely. That is a healthier tradeoff for self-hosted and local-first installations, where memory and CPU budget matter as much as peak speed.</p><p>Operators should expect the practical effect to show up in a few places:</p><ul><li>Image bursts can complete faster when multiple transforms arrive together.</li><li>Control UI file reads are less tied to shared compute pressure.</li><li>Surplus worker capacity retires after idle time.</li><li>No new configuration or schema migration is required.</li></ul><p>This is not a rewrite of worker execution. The existing dispatch, FIFO admission, and native settlement machinery still own execution and cleanup.</p><h2>The Measured Result</h2><p>The PR reports a Linux benchmark with compiled production entrypoints processing two 1536x1536 JPEG encodes, six 64x64 JPEG encodes, and 24 uncached 256 KiB assets plus sidecars per burst. Warm results covered 12 bursts, and encoded-image plus asset byte digests matched across every comparison.</p><p>Warm image queue-wait p99 fell from 209.99 ms to 103.54 ms. Spaced image queue-wait p99 fell from 204.36 ms to 108.89 ms. The complete warm mixed workload improved from 2448.42 ms to 1287.70 ms, reported as 1.90x faster.</p><p>There is a tradeoff: warm ordinary file queue-wait p99 increased from 13.90 ms to 25.76 ms, and cold file queue-wait p99 increased from 112.99 ms to 153.02 ms in that four-CPU workload. But when two shared compute permits were held, file queue-wait p99 improved sharply from 207.04 ms to 19.01 ms.</p><p>Idle worker heap stayed essentially flat after about six idle seconds, at 47.20 MiB before and 47.23 MiB after. That supports the core design: add short-lived burst capacity, then retire it.</p><h2>Bottom Line</h2><p>PR #167676 is a focused throughput improvement for OpenClaw's media and file-heavy paths. It lets image work scale modestly during bursts, keeps Control UI reads from being dragged through the wrong queue, and avoids a permanent resource increase.</p><p>For users leaning on generated media or asset-heavy plugin workflows, this should make peak moments feel less clogged.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-image-worker-burst-capacity.png\" alt=\"OpenClaw Image Workers Get Burst Capacity\" /></p>",
      "date_published": "2026-10-09T08:03:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Image",
        "Workers",
        "Burst",
        "Capacity"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-image-worker-burst-capacity.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-image-worker-burst-capacity.png",
          "mime_type": "image/png",
          "title": "OpenClaw Image Workers Get Burst Capacity cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-9-worktree-removal-maintenance/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-9-worktree-removal-maintenance/",
      "title": "OpenClaw Speeds Up Managed Worktree Removal",
      "summary": "OpenClaw PR #167695 moves repository pack repair out of individual worktree removal and preserves snapshots during missing-object recovery.",
      "content_text": "OpenClaw merged [PR #167695](https://github.com/openclaw/openclaw/pull/167695), a managed-worktree performance and recovery change focused on large repositories and partial clones.\n\nThe bug was not flashy, but it was expensive. Removing one managed worktree could rebuild the shared repository's entire multi-pack index. On large partial clones, that meant every removal could pay a repository-wide maintenance cost. Snapshot creation could also spend another two minutes implicitly fetching missing objects through inconsistent commit-graph state.\n\n## What Changed\n\nOpenClaw now keeps pack repair and consolidation under the background maintenance owner instead of performing that work during each individual removal.\n\nThat shift gives removal a cleaner job: preserve the checkout when required objects are missing, keep the existing ownership and timeout rules, and let explicit fetch or repository repair complete the path before retry. Native object validation, pending recovery pins, force and capacity-eviction snapshot-loss policy, and timeout budgets remain intact.\n\nThe PR also applies a no-fetch policy to snapshot inventory, ordinary snapshots, and exact-state verification through one operation-owned environment. That matters because implicit Git fetches during snapshotting can turn a local cleanup path into a slow network-dependent operation.\n\n## Why Operators Should Care\n\nManaged worktrees are part of the machinery that makes agent workspaces practical. They let OpenClaw prepare, isolate, snapshot, and clean up repository state without treating every task as a brand-new clone.\n\nWhen cleanup is slow, it can show up as delayed workspace retirement, retained disk pressure, or sluggish recovery after failed or abandoned work. When cleanup tries to repair repository-wide Git structures inline, one local removal can inherit the cost of the entire shared repository.\n\nThis change separates those concerns:\n\n- Individual removals no longer rebuild pack indexes.\n- Background maintenance owns pack repair and consolidation.\n- Snapshot operations avoid implicit object fetches.\n- Missing-object cases preserve the checkout for explicit repair and retry.\n- No schema migration or operator configuration is required.\n\nThat is a practical improvement for installations with large repositories, retained snapshots, or many short-lived managed worktrees.\n\n## The Measured Result\n\nThe PR reports real `ManagedWorktreeService.remove` runs on a synthetic partial clone with 4,700 indexed packs, 2,405,376 filler objects, and 20,000 tracked files. Across 30 samples, removal p50 fell from 1,946 ms to 1,372 ms, a 29.5% reduction. Warm-sample empirical p99 dropped from 2,193 ms to 1,647 ms.\n\nThe clearest operational number is multi-pack index writes: 30 per-removal MIDX writes before, zero after. After five bounded maintenance passes, both revisions consolidated 4,730 packs down to one, which supports the intended ownership split: maintenance still does the repair work, but removal no longer does it every time.\n\nThe PR is also transparent about tradeoffs. In an intentionally unindexed run, p50 got slower until maintenance built a MIDX. Cold-inclusive p99 was dominated by worker bootstrap, so the PR does not claim a cold-start win.\n\n## Bottom Line\n\nPR #167695 makes OpenClaw's managed-worktree cleanup less eager and more predictable. Removal stops acting like repository maintenance, snapshots stop fetching behind the operator's back, and missing-object recovery gets a safer explicit retry path.\n\nFor anyone running OpenClaw against large or partial-clone repositories, that is a meaningful cleanup-path improvement.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167695\">PR #167695</a>, a managed-worktree performance and recovery change focused on large repositories and partial clones.</p><p>The bug was not flashy, but it was expensive. Removing one managed worktree could rebuild the shared repository's entire multi-pack index. On large partial clones, that meant every removal could pay a repository-wide maintenance cost. Snapshot creation could also spend another two minutes implicitly fetching missing objects through inconsistent commit-graph state.</p><h2>What Changed</h2><p>OpenClaw now keeps pack repair and consolidation under the background maintenance owner instead of performing that work during each individual removal.</p><p>That shift gives removal a cleaner job: preserve the checkout when required objects are missing, keep the existing ownership and timeout rules, and let explicit fetch or repository repair complete the path before retry. Native object validation, pending recovery pins, force and capacity-eviction snapshot-loss policy, and timeout budgets remain intact.</p><p>The PR also applies a no-fetch policy to snapshot inventory, ordinary snapshots, and exact-state verification through one operation-owned environment. That matters because implicit Git fetches during snapshotting can turn a local cleanup path into a slow network-dependent operation.</p><h2>Why Operators Should Care</h2><p>Managed worktrees are part of the machinery that makes agent workspaces practical. They let OpenClaw prepare, isolate, snapshot, and clean up repository state without treating every task as a brand-new clone.</p><p>When cleanup is slow, it can show up as delayed workspace retirement, retained disk pressure, or sluggish recovery after failed or abandoned work. When cleanup tries to repair repository-wide Git structures inline, one local removal can inherit the cost of the entire shared repository.</p><p>This change separates those concerns:</p><ul><li>Individual removals no longer rebuild pack indexes.</li><li>Background maintenance owns pack repair and consolidation.</li><li>Snapshot operations avoid implicit object fetches.</li><li>Missing-object cases preserve the checkout for explicit repair and retry.</li><li>No schema migration or operator configuration is required.</li></ul><p>That is a practical improvement for installations with large repositories, retained snapshots, or many short-lived managed worktrees.</p><h2>The Measured Result</h2><p>The PR reports real <code>ManagedWorktreeService.remove</code> runs on a synthetic partial clone with 4,700 indexed packs, 2,405,376 filler objects, and 20,000 tracked files. Across 30 samples, removal p50 fell from 1,946 ms to 1,372 ms, a 29.5% reduction. Warm-sample empirical p99 dropped from 2,193 ms to 1,647 ms.</p><p>The clearest operational number is multi-pack index writes: 30 per-removal MIDX writes before, zero after. After five bounded maintenance passes, both revisions consolidated 4,730 packs down to one, which supports the intended ownership split: maintenance still does the repair work, but removal no longer does it every time.</p><p>The PR is also transparent about tradeoffs. In an intentionally unindexed run, p50 got slower until maintenance built a MIDX. Cold-inclusive p99 was dominated by worker bootstrap, so the PR does not claim a cold-start win.</p><h2>Bottom Line</h2><p>PR #167695 makes OpenClaw's managed-worktree cleanup less eager and more predictable. Removal stops acting like repository maintenance, snapshots stop fetching behind the operator's back, and missing-object recovery gets a safer explicit retry path.</p><p>For anyone running OpenClaw against large or partial-clone repositories, that is a meaningful cleanup-path improvement.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-worktree-removal-maintenance.png\" alt=\"OpenClaw Speeds Up Managed Worktree Removal\" /></p>",
      "date_published": "2026-10-09T08:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Speeds",
        "Managed",
        "Worktree",
        "Removal"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-worktree-removal-maintenance.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-worktree-removal-maintenance.png",
          "mime_type": "image/png",
          "title": "OpenClaw Speeds Up Managed Worktree Removal cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-9-queue-starvation-fix/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-9-queue-starvation-fix/",
      "title": "OpenClaw Bounds Queue Starvation Under Load",
      "summary": "OpenClaw PR #167683 prevents steady foreground traffic from indefinitely delaying background work while preserving existing queue limits.",
      "content_text": "OpenClaw merged [PR #167683](https://github.com/openclaw/openclaw/pull/167683), a queue fairness fix for busy systems where foreground work can keep arriving faster than lower-priority work gets a chance to run.\n\nThe change targets a specific failure mode: continuous foreground arrivals could indefinitely postpone queued announcements and background work. Both the per-lane selector and shared-capacity arbiter previously used strict priority, so age did not eventually help an older queued item break through.\n\n## What Changed\n\nThe new queue policy gives foreground work a head start, but not an unlimited one. According to the PR, foreground work retains a 15-second head start over normal work and a 30-second head start over background work. After that window, newer arrivals cannot keep overtaking older work forever.\n\nThat distinction matters. OpenClaw is not removing priority from the scheduler. A foreground task can still be favored when it first arrives, and running tasks are not preempted. The fix is about bounding overtaking so that older work eventually gets admitted under sustained load.\n\nThe implementation uses one comparator for FIFO lane heads and eligible capacity-group heads. It orders them by monotonic enqueue time plus the priority head start, then breaks ties by global enqueue sequence. The PR also removes duplicate arbitration policy and per-priority length counters.\n\n## Why It Matters\n\nQueue fairness is one of those runtime details that users usually notice only when it fails. A system can feel healthy during short bursts and still behave badly under a long stream of interactive work. Background announcements, scheduled work, and ordinary lower-priority tasks may look stuck even though the process is alive.\n\nThis PR narrows that gap. It keeps the existing concurrency and reservation model, preserves cancellation ownership, and avoids adding timers or promotion queues. The `queueAhead` value now represents the full enqueue-time backlog instead of a strict-priority rank that would become misleading once aging is involved.\n\nFor operators, the important takeaway is simple:\n\n- Foreground work still gets first response.\n- Background work is no longer starved indefinitely by constant arrivals.\n- Running work is not interrupted.\n- The queue contract is easier to reason about because the same comparator drives both lane and capacity-group ordering.\n\n## The Measured Result\n\nThe PR includes a 60-second load rig through the actual queue API. In that rig, eight background announcements were queued at the start while foreground clients kept replenishing work. A second 64-job background burst arrived at 15 seconds.\n\nThe reported announcement wait p99 dropped from 60.097 seconds to 30.175 seconds. The 64-job background burst wait p99 dropped from 45.296 seconds to 30.216 seconds. Peak active tasks stayed at 32, while peak RSS and JavaScript heap were slightly lower in the candidate run.\n\nForeground wait stayed effectively unchanged across the full run, with p99 moving from 100.96 ms to 101.09 ms. During the burst window, foreground p99 rose to 300.70 ms, which is the expected tradeoff when aged backlog briefly takes admission precedence.\n\nThe PR is careful not to overclaim. The load rig used synthetic work through the real queue API, not production provider requests. It proves the scheduler behavior, not every possible overload scenario.\n\n## Bottom Line\n\nPR #167683 is a small runtime change with a large operational effect. It keeps OpenClaw responsive to foreground activity while preventing lower-priority work from being pushed back forever.\n\nFor high-traffic Gateways, that is the kind of fairness fix that makes scheduled jobs, announcements, and background maintenance feel less mysterious when the system is busy.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167683\">PR #167683</a>, a queue fairness fix for busy systems where foreground work can keep arriving faster than lower-priority work gets a chance to run.</p><p>The change targets a specific failure mode: continuous foreground arrivals could indefinitely postpone queued announcements and background work. Both the per-lane selector and shared-capacity arbiter previously used strict priority, so age did not eventually help an older queued item break through.</p><h2>What Changed</h2><p>The new queue policy gives foreground work a head start, but not an unlimited one. According to the PR, foreground work retains a 15-second head start over normal work and a 30-second head start over background work. After that window, newer arrivals cannot keep overtaking older work forever.</p><p>That distinction matters. OpenClaw is not removing priority from the scheduler. A foreground task can still be favored when it first arrives, and running tasks are not preempted. The fix is about bounding overtaking so that older work eventually gets admitted under sustained load.</p><p>The implementation uses one comparator for FIFO lane heads and eligible capacity-group heads. It orders them by monotonic enqueue time plus the priority head start, then breaks ties by global enqueue sequence. The PR also removes duplicate arbitration policy and per-priority length counters.</p><h2>Why It Matters</h2><p>Queue fairness is one of those runtime details that users usually notice only when it fails. A system can feel healthy during short bursts and still behave badly under a long stream of interactive work. Background announcements, scheduled work, and ordinary lower-priority tasks may look stuck even though the process is alive.</p><p>This PR narrows that gap. It keeps the existing concurrency and reservation model, preserves cancellation ownership, and avoids adding timers or promotion queues. The <code>queueAhead</code> value now represents the full enqueue-time backlog instead of a strict-priority rank that would become misleading once aging is involved.</p><p>For operators, the important takeaway is simple:</p><ul><li>Foreground work still gets first response.</li><li>Background work is no longer starved indefinitely by constant arrivals.</li><li>Running work is not interrupted.</li><li>The queue contract is easier to reason about because the same comparator drives both lane and capacity-group ordering.</li></ul><h2>The Measured Result</h2><p>The PR includes a 60-second load rig through the actual queue API. In that rig, eight background announcements were queued at the start while foreground clients kept replenishing work. A second 64-job background burst arrived at 15 seconds.</p><p>The reported announcement wait p99 dropped from 60.097 seconds to 30.175 seconds. The 64-job background burst wait p99 dropped from 45.296 seconds to 30.216 seconds. Peak active tasks stayed at 32, while peak RSS and JavaScript heap were slightly lower in the candidate run.</p><p>Foreground wait stayed effectively unchanged across the full run, with p99 moving from 100.96 ms to 101.09 ms. During the burst window, foreground p99 rose to 300.70 ms, which is the expected tradeoff when aged backlog briefly takes admission precedence.</p><p>The PR is careful not to overclaim. The load rig used synthetic work through the real queue API, not production provider requests. It proves the scheduler behavior, not every possible overload scenario.</p><h2>Bottom Line</h2><p>PR #167683 is a small runtime change with a large operational effect. It keeps OpenClaw responsive to foreground activity while preventing lower-priority work from being pushed back forever.</p><p>For high-traffic Gateways, that is the kind of fairness fix that makes scheduled jobs, announcements, and background maintenance feel less mysterious when the system is busy.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-queue-starvation-fix.png\" alt=\"OpenClaw Bounds Queue Starvation Under Load\" /></p>",
      "date_published": "2026-10-09T08:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Bounds",
        "Queue",
        "Starvation",
        "Under",
        "Load"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-queue-starvation-fix.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-9-queue-starvation-fix.png",
          "mime_type": "image/png",
          "title": "OpenClaw Bounds Queue Starvation Under Load cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-8-provider-models-after-signin/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-8-provider-models-after-signin/",
      "title": "OpenClaw Shows Provider Models Right After Sign-In",
      "summary": "OpenClaw PR #167460 makes known provider models appear immediately after sign-in, even before live model discovery finishes.",
      "content_text": "OpenClaw merged [PR #167460](https://github.com/openclaw/openclaw/pull/167460), fixing a frustrating model-picker delay after provider sign-in.\n\nThe bug was simple from a user's perspective: sign in to a provider, then wait while the picker still showed no models. OpenClaw already knew some of those models from the provider plugin manifest, but the UI could remain empty until live model discovery finished. If that discovery took many seconds, the new account looked unusable for longer than necessary.\n\n## The New Behavior\n\nAfter this change, OpenClaw shows the provider models it already knows from the plugin manifest as soon as sign-in completes. When the account's live listing arrives, that live data replaces the temporary manifest-backed rows.\n\nThat replacement is important. The manifest can tell OpenClaw what the provider generally supports, but the live account listing may add account-only models or remove models the signed-in account cannot actually use.\n\nThe failure path is also better. If the first live listing fails, the known manifest rows remain visible. Signing out still removes the rows, and a listing that started before sign-out is fenced so it cannot bring stale models back.\n\n## Why It Was Broken\n\nThe PR explains that quick catalog publication after an auth change only seeded static rows for providers named in config. A provider that had just been signed in could get nothing until live discovery settled.\n\nAPI-key sign-in made the problem sharper because it rewrites `auth.profiles`. That config reload rebuilt the quick catalog without the newly signed-in provider.\n\nOpenClaw now captures manifest rows for every provider with current credentials. If a provider has env or stored credentials, it is treated as signed in for this purpose. At Gateway startup, those providers can show manifest rows until live discovery publishes a more exact account listing.\n\n## No Extra Runtime Loading\n\nThe implementation uses already-loaded manifest metadata. The PR says it adds no network calls and no plugin runtime loading for this quick path. Ordering, replace mode, visibility filters, allowlists, and auth-generation fencing remain unchanged.\n\nThat is the right tradeoff. The picker becomes useful sooner without pretending the manifest is the final source of truth.\n\n## The Proof\n\nThe regression coverage used a real Gateway with a fixture provider plugin whose model listing could be held or failed. Before the fix, the expected known row was missing while listing was held. After the fix, known models appeared while the listing was pending, stayed visible after an initial listing failure, disappeared on sign-out, and were replaced by live account rows once discovery succeeded.\n\nThe PR also reports isolated Gateway checks in a headless browser against the Control UI picker. The fixed branch showed known model rows while listing was held, then replaced them after the live listing completed.\n\n## Bottom Line\n\nPR #167460 makes model sign-in feel much less broken. OpenClaw no longer waits for live provider discovery before showing models it already knows are plausible for the signed-in provider.\n\nFor anyone regularly adding API keys, switching providers, or restarting Gateways with stored credentials, this should make the model picker feel immediate instead of oddly empty.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167460\">PR #167460</a>, fixing a frustrating model-picker delay after provider sign-in.</p><p>The bug was simple from a user's perspective: sign in to a provider, then wait while the picker still showed no models. OpenClaw already knew some of those models from the provider plugin manifest, but the UI could remain empty until live model discovery finished. If that discovery took many seconds, the new account looked unusable for longer than necessary.</p><h2>The New Behavior</h2><p>After this change, OpenClaw shows the provider models it already knows from the plugin manifest as soon as sign-in completes. When the account's live listing arrives, that live data replaces the temporary manifest-backed rows.</p><p>That replacement is important. The manifest can tell OpenClaw what the provider generally supports, but the live account listing may add account-only models or remove models the signed-in account cannot actually use.</p><p>The failure path is also better. If the first live listing fails, the known manifest rows remain visible. Signing out still removes the rows, and a listing that started before sign-out is fenced so it cannot bring stale models back.</p><h2>Why It Was Broken</h2><p>The PR explains that quick catalog publication after an auth change only seeded static rows for providers named in config. A provider that had just been signed in could get nothing until live discovery settled.</p><p>API-key sign-in made the problem sharper because it rewrites <code>auth.profiles</code>. That config reload rebuilt the quick catalog without the newly signed-in provider.</p><p>OpenClaw now captures manifest rows for every provider with current credentials. If a provider has env or stored credentials, it is treated as signed in for this purpose. At Gateway startup, those providers can show manifest rows until live discovery publishes a more exact account listing.</p><h2>No Extra Runtime Loading</h2><p>The implementation uses already-loaded manifest metadata. The PR says it adds no network calls and no plugin runtime loading for this quick path. Ordering, replace mode, visibility filters, allowlists, and auth-generation fencing remain unchanged.</p><p>That is the right tradeoff. The picker becomes useful sooner without pretending the manifest is the final source of truth.</p><h2>The Proof</h2><p>The regression coverage used a real Gateway with a fixture provider plugin whose model listing could be held or failed. Before the fix, the expected known row was missing while listing was held. After the fix, known models appeared while the listing was pending, stayed visible after an initial listing failure, disappeared on sign-out, and were replaced by live account rows once discovery succeeded.</p><p>The PR also reports isolated Gateway checks in a headless browser against the Control UI picker. The fixed branch showed known model rows while listing was held, then replaced them after the live listing completed.</p><h2>Bottom Line</h2><p>PR #167460 makes model sign-in feel much less broken. OpenClaw no longer waits for live provider discovery before showing models it already knows are plausible for the signed-in provider.</p><p>For anyone regularly adding API keys, switching providers, or restarting Gateways with stored credentials, this should make the model picker feel immediate instead of oddly empty.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-provider-models-after-signin.png\" alt=\"OpenClaw Shows Provider Models Right After Sign-In\" /></p>",
      "date_published": "2026-10-08T23:02:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Shows",
        "Provider",
        "Models",
        "Right",
        "After"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-provider-models-after-signin.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-provider-models-after-signin.png",
          "mime_type": "image/png",
          "title": "OpenClaw Shows Provider Models Right After Sign-In cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-8-gateway-chat-sql-reduction/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-8-gateway-chat-sql-reduction/",
      "title": "OpenClaw Cuts Gateway SQL During Chat Turns",
      "summary": "OpenClaw PR #167194 reduces repeated Gateway database work during warmed chat turns while keeping session durability unchanged.",
      "content_text": "OpenClaw merged [PR #167194](https://github.com/openclaw/openclaw/pull/167194), a Gateway performance change aimed at reducing repeated database work during chat turns.\n\nThis is the kind of improvement that will not show up as a new button, but it matters for busy installations. Chat turns touch session rows, placement facts, transcript state, pending results, and worker requests. If those owners already published valid committed facts, reacquiring them can add unnecessary SQLite traffic to a warmed turn.\n\n## What Changed\n\nThe PR focuses on reusing facts that are already complete and locally owned. Instead of asking the database again for session-row and placement presentation data that a resident owner already published, OpenClaw can retain and reuse the committed facts.\n\nThe change also consumes initial remote pending-result preparation once, then discards it before later setup and redispatch waits. That keeps the existing ownership rules intact while avoiding repeated preparation work.\n\nThe author is explicit about boundaries: schema, stored bytes, durability, retention, permissions, and update behavior are unchanged. In other words, this is a performance refactor, not a change to what OpenClaw stores or who is allowed to read it.\n\n## The Measured Result\n\nThe PR reports a measured reduction of 11.8 all-thread SQLite statements and 3.9 logical worker requests per warmed turn against its step-one base. First-turn statements decreased from 4,339 to 4,327 in the same measurement context.\n\nThat may sound small until you multiply it by active users, concurrent agents, and long sessions. Gateway performance often comes from trimming repeated work in hot paths, not from one dramatic rewrite.\n\nThe implementation keeps conservative invalidation for receipts that are superseded, reentrant, uncertain, or bound to another store. Pending reads also cannot overwrite newer publication. Those details matter because performance work around session state can become dangerous if it gets too eager.\n\n## Why Operators Should Care\n\nOpenClaw has been steadily moving more agent activity through durable session and projection layers. That gives users better recovery and observability, but it also means database work sits close to the user-facing chat loop.\n\nReducing repeated reads helps in a few practical ways:\n\n- Busy Gateways spend less time doing duplicate session bookkeeping.\n- Warmed chat turns avoid work that has already been safely published.\n- Worker request pressure drops without changing persistence semantics.\n- Future performance work has a cleaner owner boundary to build on.\n\nThe PR connects to earlier work in [PR #166703](https://github.com/openclaw/openclaw/pull/166703), which retained selected session execution through turn phases. Together, these changes suggest a broader tuning pass over how Gateway state moves through agent turns.\n\n## Bottom Line\n\nPR #167194 is a low-glamour, high-leverage change. It does not promise a visible feature, and it does not claim a storage migration. It simply cuts repeated database and worker activity from a hot chat path while keeping durability and permission behavior stable.\n\nFor OpenClaw users running active Gateways, especially with multiple agents or long-lived sessions, that is the right kind of invisible improvement.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167194\">PR #167194</a>, a Gateway performance change aimed at reducing repeated database work during chat turns.</p><p>This is the kind of improvement that will not show up as a new button, but it matters for busy installations. Chat turns touch session rows, placement facts, transcript state, pending results, and worker requests. If those owners already published valid committed facts, reacquiring them can add unnecessary SQLite traffic to a warmed turn.</p><h2>What Changed</h2><p>The PR focuses on reusing facts that are already complete and locally owned. Instead of asking the database again for session-row and placement presentation data that a resident owner already published, OpenClaw can retain and reuse the committed facts.</p><p>The change also consumes initial remote pending-result preparation once, then discards it before later setup and redispatch waits. That keeps the existing ownership rules intact while avoiding repeated preparation work.</p><p>The author is explicit about boundaries: schema, stored bytes, durability, retention, permissions, and update behavior are unchanged. In other words, this is a performance refactor, not a change to what OpenClaw stores or who is allowed to read it.</p><h2>The Measured Result</h2><p>The PR reports a measured reduction of 11.8 all-thread SQLite statements and 3.9 logical worker requests per warmed turn against its step-one base. First-turn statements decreased from 4,339 to 4,327 in the same measurement context.</p><p>That may sound small until you multiply it by active users, concurrent agents, and long sessions. Gateway performance often comes from trimming repeated work in hot paths, not from one dramatic rewrite.</p><p>The implementation keeps conservative invalidation for receipts that are superseded, reentrant, uncertain, or bound to another store. Pending reads also cannot overwrite newer publication. Those details matter because performance work around session state can become dangerous if it gets too eager.</p><h2>Why Operators Should Care</h2><p>OpenClaw has been steadily moving more agent activity through durable session and projection layers. That gives users better recovery and observability, but it also means database work sits close to the user-facing chat loop.</p><p>Reducing repeated reads helps in a few practical ways:</p><ul><li>Busy Gateways spend less time doing duplicate session bookkeeping.</li><li>Warmed chat turns avoid work that has already been safely published.</li><li>Worker request pressure drops without changing persistence semantics.</li><li>Future performance work has a cleaner owner boundary to build on.</li></ul><p>The PR connects to earlier work in <a href=\"https://github.com/openclaw/openclaw/pull/166703\">PR #166703</a>, which retained selected session execution through turn phases. Together, these changes suggest a broader tuning pass over how Gateway state moves through agent turns.</p><h2>Bottom Line</h2><p>PR #167194 is a low-glamour, high-leverage change. It does not promise a visible feature, and it does not claim a storage migration. It simply cuts repeated database and worker activity from a hot chat path while keeping durability and permission behavior stable.</p><p>For OpenClaw users running active Gateways, especially with multiple agents or long-lived sessions, that is the right kind of invisible improvement.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-gateway-chat-sql-reduction.png\" alt=\"OpenClaw Cuts Gateway SQL During Chat Turns\" /></p>",
      "date_published": "2026-10-08T23:01:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Cuts",
        "Gateway",
        "During",
        "Chat",
        "Turns"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-gateway-chat-sql-reduction.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-gateway-chat-sql-reduction.png",
          "mime_type": "image/png",
          "title": "OpenClaw Cuts Gateway SQL During Chat Turns cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-8-stable-2026-9-9-recovery-release/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-8-stable-2026-9-9-recovery-release/",
      "title": "OpenClaw 2026.9.9 Ships Stable Recovery Fixes",
      "summary": "OpenClaw 2026.9.9 is a stable release focused on update recovery, messaging reliability, model support, and safer maintenance paths.",
      "content_text": "OpenClaw published [v2026.9.9](https://github.com/openclaw/openclaw/releases/tag/v2026.9.9) today, moving a broad set of reliability work into the stable channel. The release notes describe a large update: 112 pull requests, 69 direct commits, and 91 contributors.\n\nThe headline is not one flashy feature. It is a practical stable release for operators who care about upgrades, failed-update recovery, channel delivery, and model availability. The [official changelog](https://raw.githubusercontent.com/openclaw/openclaw/main/CHANGELOG/2026.9.9.md) summarizes the release as adding GPT-6.1 Sol to the Codex model list, supporting Claude Haiku 5.5 through Anthropic and Claude CLI, improving failed-update recovery, fixing missing iMessage replies, and preventing an old scheduled job from interrupting a newer conversation.\n\n## Recovery Gets the Most Attention\n\nThe release gives unusual space to update and maintenance behavior. That makes sense: OpenClaw runs long-lived agents, scheduled jobs, channel integrations, plugin state, and local databases. A rough update can turn into an operational mess quickly.\n\nAmong the notable maintenance fixes:\n\n- Failed-update recovery can preserve data written during a supported failed attempt.\n- Returning to 2026.9.8 keeps later updates possible.\n- macOS and Linux restarts stay blocked until Doctor processes are confirmed stopped.\n- Backups and restores on FUSE storage avoid failing solely because file times changed.\n- Doctor can continue older memory-database upgrades past oversized or damaged saved entries.\n- Session archive repairs can resume after interruption.\n\nThat set of changes points to a release shaped by real field failures rather than just planned features. The notes repeatedly call out upgrade loops, rollback behavior, damaged metadata, interrupted repairs, and compatibility with older installations.\n\n## Messaging and Scheduling Fixes\n\nOpenClaw 2026.9.9 also tightens several channel paths. The release says affected iMessage replies that were written but never sent are now delivered. Discord delegated task results are recognized as delivered when the answer already appeared in the thread. Google Chat and Slack replies retain their original threads in the covered paths.\n\nScheduled jobs also get a meaningful fix. An older job reaching its time limit could stop a newer conversation in the same session and remove its tool access. That is now covered, which matters for anyone using automations alongside normal chat work.\n\n## Model Support Moves Forward\n\nThe release adds GPT-6.1 Sol to the Codex model list, while noting that use still depends on an eligible account and compatible Codex setup. Claude Haiku 5.5 is now selectable through both Anthropic API and Claude CLI routes, and the `haiku` alias now points to 5.5.\n\nThe bundled Codex copy also moves to 0.160.0. OpenClaw is careful to distinguish that bundled copy from any standalone Codex CLI a user updates separately.\n\n## Why This Release Matters\n\nStable releases are where operational fixes become broadly useful. The October beta line has been carrying fast-moving recovery, Doctor, messaging, model, plugin, and scheduler work. Version 2026.9.9 brings a large portion of that reliability story to users who follow stable npm releases.\n\nThe npm package has also advanced to [openclaw 2026.9.9](https://www.npmjs.com/package/openclaw/v/2026.9.9), matching the GitHub release.\n\nFor operators, the short version is simple: this is a maintenance-heavy stable update with real upside if you have seen update failures, Doctor repair interruptions, missing channel replies, or model-picker drift. It is less about changing how OpenClaw feels day to day and more about making the system recover cleanly when the day goes sideways.",
      "content_html": "<p>OpenClaw published <a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.9.9\">v2026.9.9</a> today, moving a broad set of reliability work into the stable channel. The release notes describe a large update: 112 pull requests, 69 direct commits, and 91 contributors.</p><p>The headline is not one flashy feature. It is a practical stable release for operators who care about upgrades, failed-update recovery, channel delivery, and model availability. The <a href=\"https://raw.githubusercontent.com/openclaw/openclaw/main/CHANGELOG/2026.9.9.md\">official changelog</a> summarizes the release as adding GPT-6.1 Sol to the Codex model list, supporting Claude Haiku 5.5 through Anthropic and Claude CLI, improving failed-update recovery, fixing missing iMessage replies, and preventing an old scheduled job from interrupting a newer conversation.</p><h2>Recovery Gets the Most Attention</h2><p>The release gives unusual space to update and maintenance behavior. That makes sense: OpenClaw runs long-lived agents, scheduled jobs, channel integrations, plugin state, and local databases. A rough update can turn into an operational mess quickly.</p><p>Among the notable maintenance fixes:</p><ul><li>Failed-update recovery can preserve data written during a supported failed attempt.</li><li>Returning to 2026.9.8 keeps later updates possible.</li><li>macOS and Linux restarts stay blocked until Doctor processes are confirmed stopped.</li><li>Backups and restores on FUSE storage avoid failing solely because file times changed.</li><li>Doctor can continue older memory-database upgrades past oversized or damaged saved entries.</li><li>Session archive repairs can resume after interruption.</li></ul><p>That set of changes points to a release shaped by real field failures rather than just planned features. The notes repeatedly call out upgrade loops, rollback behavior, damaged metadata, interrupted repairs, and compatibility with older installations.</p><h2>Messaging and Scheduling Fixes</h2><p>OpenClaw 2026.9.9 also tightens several channel paths. The release says affected iMessage replies that were written but never sent are now delivered. Discord delegated task results are recognized as delivered when the answer already appeared in the thread. Google Chat and Slack replies retain their original threads in the covered paths.</p><p>Scheduled jobs also get a meaningful fix. An older job reaching its time limit could stop a newer conversation in the same session and remove its tool access. That is now covered, which matters for anyone using automations alongside normal chat work.</p><h2>Model Support Moves Forward</h2><p>The release adds GPT-6.1 Sol to the Codex model list, while noting that use still depends on an eligible account and compatible Codex setup. Claude Haiku 5.5 is now selectable through both Anthropic API and Claude CLI routes, and the <code>haiku</code> alias now points to 5.5.</p><p>The bundled Codex copy also moves to 0.160.0. OpenClaw is careful to distinguish that bundled copy from any standalone Codex CLI a user updates separately.</p><h2>Why This Release Matters</h2><p>Stable releases are where operational fixes become broadly useful. The October beta line has been carrying fast-moving recovery, Doctor, messaging, model, plugin, and scheduler work. Version 2026.9.9 brings a large portion of that reliability story to users who follow stable npm releases.</p><p>The npm package has also advanced to <a href=\"https://www.npmjs.com/package/openclaw/v/2026.9.9\">openclaw 2026.9.9</a>, matching the GitHub release.</p><p>For operators, the short version is simple: this is a maintenance-heavy stable update with real upside if you have seen update failures, Doctor repair interruptions, missing channel replies, or model-picker drift. It is less about changing how OpenClaw feels day to day and more about making the system recover cleanly when the day goes sideways.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-stable-2026-9-9-recovery-release.png\" alt=\"OpenClaw 2026.9.9 Ships Stable Recovery Fixes\" /></p>",
      "date_published": "2026-10-08T23:00:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "Memory",
        "2026",
        "Ships",
        "Stable",
        "Recovery",
        "Fixes"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-stable-2026-9-9-recovery-release.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-stable-2026-9-9-recovery-release.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.9.9 Ships Stable Recovery Fixes cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-8-provider-reconnect-timeout-fix/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-8-provider-reconnect-timeout-fix/",
      "title": "OpenClaw Fixes Provider Reconnect Timeout Hangs",
      "summary": "OpenClaw PR #167013 restores fast TCP/TLS deadlines during provider reconnects, preventing stalled handshakes from hanging agent turns.",
      "content_text": "OpenClaw merged [PR #167013](https://github.com/openclaw/openclaw/pull/167013), a P1 reliability fix for provider reconnects. The bug could make an agent turn wait for a long model idle timeout when a provider connection needed to reconnect and the TLS handshake stopped responding.\n\nIn plain English: a network hiccup could turn into a several-minute wait, even though the useful behavior was to hit the normal connection deadline, fail fast, and let the existing retry path recover.\n\n## What Went Wrong\n\nThe PR explains that OpenClaw's inherited streaming timeout was being passed down as a connection timeout. That meant a long response/streaming budget could replace Undici's normal 10-second TCP/TLS connection deadline.\n\nThose are different clocks. A model may legitimately need a longer request or stream budget once a connection is established. But a stalled TCP/TLS connection attempt should not inherit that whole model-idle window.\n\nThe reproduced failure was specific: a warmed provider dispatcher reused a path, the keep-alive socket was dropped, and the reconnecting TLS server withheld bytes. On main, the turn waited until the long idle timeout. With the fix, the stalled connection hit the fast connector deadline and the turn recovered through retry.\n\n## The Fix\n\nPR #167013 separates inherited stream timeouts from connection deadlines. OpenClaw now applies inherited defaults to response headers and bodies, while guarded fetch passes explicit request deadlines separately. Reusable dispatchers are keyed on both policies so the right timeout behavior survives pooling.\n\nThe change preserves explicit provider `timeoutSeconds` settings. If an operator has intentionally configured provider budgets, those documented connection and request budgets remain the source of truth.\n\nThe PR also keeps existing network protections intact: fresh DNS validation, address pinning, redirect policy, and unrelated active streams are not widened by this change.\n\n## The Measured Impact\n\nThe PR's live Gateway proof is the reason this fix is worth calling out. In the baseline case, the transport failure surfaced as an LLM idle timeout after about 300 seconds, and the full agent turn completed after roughly 301 seconds.\n\nWith the fix, the stalled connection failed with `UND_ERR_CONNECT_TIMEOUT` at about 10.3 seconds, and the complete agent turn finished after roughly 11.7 seconds through the existing retry behavior.\n\nThat is a large user-visible difference. It changes the experience from \"the agent is mysteriously stuck\" to \"the network path failed quickly and the system recovered.\"\n\n## Why This Matters for OpenClaw Operators\n\nOpenClaw users often run through custom providers, local gateways, proxies, and long-lived sessions. Connection reuse is good for performance, but it makes timeout ownership more important. A bad pooled reconnect should not borrow the timeout meant for an already-running model stream.\n\nThis fix makes the boundary clearer:\n\n- Connection setup keeps a short TCP/TLS deadline\n- Request and stream handling keep their own provider budgets\n- Existing retry logic gets control quickly enough to help\n\nThat is especially useful for unattended runs and cron jobs, where a five-minute hang can cascade into missed schedules or confusing fallback behavior.\n\n## Bottom Line\n\nPR #167013 is not a new feature, but it is a meaningful reliability improvement. Provider reconnects should now fail at the connection layer when the connection is what stalled, rather than waiting for an agent-level idle timeout.\n\nFor anyone using OpenClaw with remote model providers, OpenAI-compatible endpoints, or custom provider routes, this is the kind of fix that makes failures feel bounded instead of spooky.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/167013\">PR #167013</a>, a P1 reliability fix for provider reconnects. The bug could make an agent turn wait for a long model idle timeout when a provider connection needed to reconnect and the TLS handshake stopped responding.</p><p>In plain English: a network hiccup could turn into a several-minute wait, even though the useful behavior was to hit the normal connection deadline, fail fast, and let the existing retry path recover.</p><h2>What Went Wrong</h2><p>The PR explains that OpenClaw's inherited streaming timeout was being passed down as a connection timeout. That meant a long response/streaming budget could replace Undici's normal 10-second TCP/TLS connection deadline.</p><p>Those are different clocks. A model may legitimately need a longer request or stream budget once a connection is established. But a stalled TCP/TLS connection attempt should not inherit that whole model-idle window.</p><p>The reproduced failure was specific: a warmed provider dispatcher reused a path, the keep-alive socket was dropped, and the reconnecting TLS server withheld bytes. On main, the turn waited until the long idle timeout. With the fix, the stalled connection hit the fast connector deadline and the turn recovered through retry.</p><h2>The Fix</h2><p>PR #167013 separates inherited stream timeouts from connection deadlines. OpenClaw now applies inherited defaults to response headers and bodies, while guarded fetch passes explicit request deadlines separately. Reusable dispatchers are keyed on both policies so the right timeout behavior survives pooling.</p><p>The change preserves explicit provider <code>timeoutSeconds</code> settings. If an operator has intentionally configured provider budgets, those documented connection and request budgets remain the source of truth.</p><p>The PR also keeps existing network protections intact: fresh DNS validation, address pinning, redirect policy, and unrelated active streams are not widened by this change.</p><h2>The Measured Impact</h2><p>The PR's live Gateway proof is the reason this fix is worth calling out. In the baseline case, the transport failure surfaced as an LLM idle timeout after about 300 seconds, and the full agent turn completed after roughly 301 seconds.</p><p>With the fix, the stalled connection failed with <code>UND_ERR_CONNECT_TIMEOUT</code> at about 10.3 seconds, and the complete agent turn finished after roughly 11.7 seconds through the existing retry behavior.</p><p>That is a large user-visible difference. It changes the experience from \"the agent is mysteriously stuck\" to \"the network path failed quickly and the system recovered.\"</p><h2>Why This Matters for OpenClaw Operators</h2><p>OpenClaw users often run through custom providers, local gateways, proxies, and long-lived sessions. Connection reuse is good for performance, but it makes timeout ownership more important. A bad pooled reconnect should not borrow the timeout meant for an already-running model stream.</p><p>This fix makes the boundary clearer:</p><ul><li>Connection setup keeps a short TCP/TLS deadline</li><li>Request and stream handling keep their own provider budgets</li><li>Existing retry logic gets control quickly enough to help</li></ul><p>That is especially useful for unattended runs and cron jobs, where a five-minute hang can cascade into missed schedules or confusing fallback behavior.</p><h2>Bottom Line</h2><p>PR #167013 is not a new feature, but it is a meaningful reliability improvement. Provider reconnects should now fail at the connection layer when the connection is what stalled, rather than waiting for an agent-level idle timeout.</p><p>For anyone using OpenClaw with remote model providers, OpenAI-compatible endpoints, or custom provider routes, this is the kind of fix that makes failures feel bounded instead of spooky.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-provider-reconnect-timeout-fix.png\" alt=\"OpenClaw Fixes Provider Reconnect Timeout Hangs\" /></p>",
      "date_published": "2026-10-08T08:07:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Provider",
        "Reconnect",
        "Timeout",
        "Hangs"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-provider-reconnect-timeout-fix.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-provider-reconnect-timeout-fix.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Provider Reconnect Timeout Hangs cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-8-openai-regex-tool-schema-fix/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-8-openai-regex-tool-schema-fix/",
      "title": "OpenClaw Fixes OpenAI Regex Tool Schema Failures",
      "summary": "OpenClaw PR #166999 fixes OpenAI tool calls when JSON Schema patterns use regex lookahead or lookbehind, avoiding strict-schema 400 errors.",
      "content_text": "OpenClaw merged [PR #166999](https://github.com/openclaw/openclaw/pull/166999), a P1 fix for a sharp compatibility bug in OpenAI tool calls. The issue: OpenAI turns could fail with a strict tool-schema `400` whenever a configured tool used a JSON Schema `pattern` containing regex lookahead or lookbehind.\n\nThat sounds narrow until you remember how common safety-oriented filename and path patterns are. A tool schema might reject absolute paths, Windows drive prefixes, or parent-directory traversal with a pattern like a negative lookahead. If that schema was attached to an MCP tool, the whole OpenAI turn could fail before the agent had a chance to respond.\n\n## What Was Failing\n\nThe PR body gives a concrete example: a `create_file` tool whose `filename` property used a pattern to block absolute paths and `..` traversal. OpenClaw's strict-schema detection did not recognize regex lookaround as incompatible with OpenAI strict mode, so it sent the tool as `strict: true`.\n\nOpenAI rejected the request with an invalid JSON schema error because regex lookaround is not supported in that strict mode path.\n\nThe important detail is that the schema itself was not wrong for OpenClaw's purposes. The compatibility problem was in how the tool was projected to OpenAI.\n\n## The Fix\n\nOpenClaw already had a downgrade path for strict-incompatible schemas. PR #166999 extends that detection to include regex lookaround groups in `pattern` values.\n\nWhen OpenClaw sees a pattern using `(?=`, `(?!`, `(?<=`, or `(?<!`, it now sends that tool with `strict: false` while preserving the original schema and the original `pattern`. That means the tool remains available and the validation intent is not stripped out.\n\nThe PR is careful about false positives. Escaped parentheses, character classes, non-capturing groups, and named groups are not treated as lookaround violations, so ordinary compatible schemas can keep strict mode.\n\n## Why It Matters for MCP Tools\n\nThis is especially relevant to MCP and filesystem-adjacent tooling. Tool authors often encode safety constraints in JSON Schema because they want the model and the runtime to see the same boundaries.\n\nWithout this fix, one such schema could break an OpenAI-backed session even if the user never invoked that specific tool. With the fix, OpenClaw can keep the tool listed, keep its schema intact, and choose the compatibility mode that lets the turn proceed.\n\nThat is the right tradeoff. The tool surface should not disappear, and OpenClaw should not mutate the author's schema to fit one provider's strict-mode subset.\n\n## Evidence From the PR\n\nThe PR includes a live route proof using OpenClaw's OpenAI Responses transport. Before the change, the affected tool was sent with `strict=true` and the request failed with a schema error. After the change, the same tool was sent with `strict=false`, the pattern stayed unchanged, and the assistant replied successfully.\n\nTests now cover the issue schema, each lookaround form, and controls for compatible regex patterns. A wire-level loopback test also verifies that the Responses request sends the tool with `strict: false` while preserving the schema.\n\n## Bottom Line\n\nThis is a small patch with a very practical outcome: OpenAI sessions should stop failing just because one available tool has a reasonable regex guard in its schema.\n\nFor anyone running OpenClaw with MCP tools, file-creation tools, or provider-mixed routes, PR #166999 is the kind of compatibility fix that prevents one edge-case schema from taking down the whole turn.",
      "content_html": "<p>OpenClaw merged <a href=\"https://github.com/openclaw/openclaw/pull/166999\">PR #166999</a>, a P1 fix for a sharp compatibility bug in OpenAI tool calls. The issue: OpenAI turns could fail with a strict tool-schema <code>400</code> whenever a configured tool used a JSON Schema <code>pattern</code> containing regex lookahead or lookbehind.</p><p>That sounds narrow until you remember how common safety-oriented filename and path patterns are. A tool schema might reject absolute paths, Windows drive prefixes, or parent-directory traversal with a pattern like a negative lookahead. If that schema was attached to an MCP tool, the whole OpenAI turn could fail before the agent had a chance to respond.</p><h2>What Was Failing</h2><p>The PR body gives a concrete example: a <code>create_file</code> tool whose <code>filename</code> property used a pattern to block absolute paths and <code>..</code> traversal. OpenClaw's strict-schema detection did not recognize regex lookaround as incompatible with OpenAI strict mode, so it sent the tool as <code>strict: true</code>.</p><p>OpenAI rejected the request with an invalid JSON schema error because regex lookaround is not supported in that strict mode path.</p><p>The important detail is that the schema itself was not wrong for OpenClaw's purposes. The compatibility problem was in how the tool was projected to OpenAI.</p><h2>The Fix</h2><p>OpenClaw already had a downgrade path for strict-incompatible schemas. PR #166999 extends that detection to include regex lookaround groups in <code>pattern</code> values.</p><p>When OpenClaw sees a pattern using <code>(?=</code>, <code>(?!</code>, <code>(?<=</code>, or <code>(?<!</code>, it now sends that tool with <code>strict: false</code> while preserving the original schema and the original <code>pattern</code>. That means the tool remains available and the validation intent is not stripped out.</p><p>The PR is careful about false positives. Escaped parentheses, character classes, non-capturing groups, and named groups are not treated as lookaround violations, so ordinary compatible schemas can keep strict mode.</p><h2>Why It Matters for MCP Tools</h2><p>This is especially relevant to MCP and filesystem-adjacent tooling. Tool authors often encode safety constraints in JSON Schema because they want the model and the runtime to see the same boundaries.</p><p>Without this fix, one such schema could break an OpenAI-backed session even if the user never invoked that specific tool. With the fix, OpenClaw can keep the tool listed, keep its schema intact, and choose the compatibility mode that lets the turn proceed.</p><p>That is the right tradeoff. The tool surface should not disappear, and OpenClaw should not mutate the author's schema to fit one provider's strict-mode subset.</p><h2>Evidence From the PR</h2><p>The PR includes a live route proof using OpenClaw's OpenAI Responses transport. Before the change, the affected tool was sent with <code>strict=true</code> and the request failed with a schema error. After the change, the same tool was sent with <code>strict=false</code>, the pattern stayed unchanged, and the assistant replied successfully.</p><p>Tests now cover the issue schema, each lookaround form, and controls for compatible regex patterns. A wire-level loopback test also verifies that the Responses request sends the tool with <code>strict: false</code> while preserving the schema.</p><h2>Bottom Line</h2><p>This is a small patch with a very practical outcome: OpenAI sessions should stop failing just because one available tool has a reasonable regex guard in its schema.</p><p>For anyone running OpenClaw with MCP tools, file-creation tools, or provider-mixed routes, PR #166999 is the kind of compatibility fix that prevents one edge-case schema from taking down the whole turn.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-openai-regex-tool-schema-fix.png\" alt=\"OpenClaw Fixes OpenAI Regex Tool Schema Failures\" /></p>",
      "date_published": "2026-10-08T08:06:00.000Z",
      "date_modified": "2026-10-14T08:14:43.486Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Openai",
        "Regex",
        "Tool",
        "Schema"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-openai-regex-tool-schema-fix.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-openai-regex-tool-schema-fix.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes OpenAI Regex Tool Schema Failures cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-8-beta-2-recovery-release/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-8-beta-2-recovery-release/",
      "title": "OpenClaw 2026.10.1 Beta 2 Focuses on Recovery",
      "summary": "OpenClaw 2026.10.1-beta.2 ships a hotfix beta with 40 commits covering update recovery, messaging fixes, plugins, workers, and Doctor repairs.",
      "content_text": "OpenClaw has published [v2026.10.1-beta.2](https://github.com/openclaw/openclaw/releases/tag/v2026.10.1-beta.2), a hotfix beta released at 00:41 UTC on October 8. The release notes describe it as a focused follow-up to `v2026.10.1-beta.1`, covering 40 intervening commits rather than repeating the broader October beta changelog.\n\nThe short version: this is a reliability release. There are no new headline capabilities, but there are several fixes that matter if you run OpenClaw as a long-lived Gateway, update frequently, rely on channel delivery, or use checkout/plugin-heavy setups.\n\n## What Changed in OpenClaw Beta 2\n\nThe release groups its highlights into three areas:\n\n- Updates and Doctor recovery\n- Replies and messaging\n- Plugins and cloud workers\n\nOn the update side, the release targets failure modes that can leave managed Gateways in awkward states: interrupted updates, archive migrations, launcher ownership problems, and service recovery on macOS and Windows. The notes specifically call out fixes for recovering archive migrations, settling pending recovery after a lease-database replacement, and restoring Gateway lifecycle commands on macOS 12 Monterey.\n\nDoctor also gets more defensive. The release includes repairs for importing sessions from older migration receipts, verifying snapshots on Linux hosts without file birthtime, and skipping empty migration scans for unconfigured Matrix installations.\n\n## Messaging Fixes Are a Big Part of This Build\n\nOpenClaw's channel surface is broad, so small routing and reply bugs can be painful. This beta includes fixes for completed answers that vanished after silent follow-ups, Discord messages that needed to steer active work while earlier messages were deferred, and Telegram sends/buttons after routed handoffs.\n\nThat cluster matters because the failure pattern is often not \"the model could not answer.\" It is \"the answer existed, then routing or follow-up bookkeeping hid it.\" The release notes mention preserving completed answers when a follow-up produces no reply, plus preserving answers that already explain an approval or input blocker.\n\nFor operators, that is the kind of fix that reduces mystery. If an agent already produced the useful response, OpenClaw should keep it visible.\n\n## Plugin and Worker Reliability\n\nThe plugin section is similarly practical. The release keeps checkout-plugin skills loadable during updates, avoids repeated native-prebuild scans, preserves plugin session state across Gateway restarts, and includes hidden runtime chunks required by external plugins when dispatching cloud workers.\n\nThere is also a Code Mode compatibility fix: tool-call schema handling now uses the unambiguous `awaitResults` argument. That is small on paper, but schema mismatches tend to show up as confusing downstream execution failures, so the explicit argument is a good cleanup.\n\n## Upgrade Notes\n\nThe release includes one important caveat for users already blocked by an older updater: these fixes do not magically patch the CLI already installed on a machine. If the old preflight refuses the upgrade, the release notes say to install the new CLI manually before rerunning Doctor.\n\nThat makes this beta less of a \"click once and forget it\" update for broken installations, but more useful once the new binary is in place. It gives Doctor and the updater better tools to recover from the exact states that have been causing trouble.\n\n## Why This Release Matters\n\nOpenClaw 2026.10.1-beta.2 is not flashy. It is the kind of beta that tightens the floor: update recovery, channel handoffs, plugin continuity, worker packaging, and diagnostic behavior.\n\nFor beta-channel users, it is worth testing because it touches the connective tissue between long-running agents and the system around them. The most important line in the release may be the simplest one: \"No new capabilities; this beta focuses on fixes and recovery.\"",
      "content_html": "<p>OpenClaw has published <a href=\"https://github.com/openclaw/openclaw/releases/tag/v2026.10.1-beta.2\">v2026.10.1-beta.2</a>, a hotfix beta released at 00:41 UTC on October 8. The release notes describe it as a focused follow-up to <code>v2026.10.1-beta.1</code>, covering 40 intervening commits rather than repeating the broader October beta changelog.</p><p>The short version: this is a reliability release. There are no new headline capabilities, but there are several fixes that matter if you run OpenClaw as a long-lived Gateway, update frequently, rely on channel delivery, or use checkout/plugin-heavy setups.</p><h2>What Changed in OpenClaw Beta 2</h2><p>The release groups its highlights into three areas:</p><ul><li>Updates and Doctor recovery</li><li>Replies and messaging</li><li>Plugins and cloud workers</li></ul><p>On the update side, the release targets failure modes that can leave managed Gateways in awkward states: interrupted updates, archive migrations, launcher ownership problems, and service recovery on macOS and Windows. The notes specifically call out fixes for recovering archive migrations, settling pending recovery after a lease-database replacement, and restoring Gateway lifecycle commands on macOS 12 Monterey.</p><p>Doctor also gets more defensive. The release includes repairs for importing sessions from older migration receipts, verifying snapshots on Linux hosts without file birthtime, and skipping empty migration scans for unconfigured Matrix installations.</p><h2>Messaging Fixes Are a Big Part of This Build</h2><p>OpenClaw's channel surface is broad, so small routing and reply bugs can be painful. This beta includes fixes for completed answers that vanished after silent follow-ups, Discord messages that needed to steer active work while earlier messages were deferred, and Telegram sends/buttons after routed handoffs.</p><p>That cluster matters because the failure pattern is often not \"the model could not answer.\" It is \"the answer existed, then routing or follow-up bookkeeping hid it.\" The release notes mention preserving completed answers when a follow-up produces no reply, plus preserving answers that already explain an approval or input blocker.</p><p>For operators, that is the kind of fix that reduces mystery. If an agent already produced the useful response, OpenClaw should keep it visible.</p><h2>Plugin and Worker Reliability</h2><p>The plugin section is similarly practical. The release keeps checkout-plugin skills loadable during updates, avoids repeated native-prebuild scans, preserves plugin session state across Gateway restarts, and includes hidden runtime chunks required by external plugins when dispatching cloud workers.</p><p>There is also a Code Mode compatibility fix: tool-call schema handling now uses the unambiguous <code>awaitResults</code> argument. That is small on paper, but schema mismatches tend to show up as confusing downstream execution failures, so the explicit argument is a good cleanup.</p><h2>Upgrade Notes</h2><p>The release includes one important caveat for users already blocked by an older updater: these fixes do not magically patch the CLI already installed on a machine. If the old preflight refuses the upgrade, the release notes say to install the new CLI manually before rerunning Doctor.</p><p>That makes this beta less of a \"click once and forget it\" update for broken installations, but more useful once the new binary is in place. It gives Doctor and the updater better tools to recover from the exact states that have been causing trouble.</p><h2>Why This Release Matters</h2><p>OpenClaw 2026.10.1-beta.2 is not flashy. It is the kind of beta that tightens the floor: update recovery, channel handoffs, plugin continuity, worker packaging, and diagnostic behavior.</p><p>For beta-channel users, it is worth testing because it touches the connective tissue between long-running agents and the system around them. The most important line in the release may be the simplest one: \"No new capabilities; this beta focuses on fixes and recovery.\"</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-beta-2-recovery-release.png\" alt=\"OpenClaw 2026.10.1 Beta 2 Focuses on Recovery\" /></p>",
      "date_published": "2026-10-08T08:05:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Releases",
        "2026",
        "Beta",
        "Focuses",
        "Recovery"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-beta-2-recovery-release.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-8-beta-2-recovery-release.png",
          "mime_type": "image/png",
          "title": "OpenClaw 2026.10.1 Beta 2 Focuses on Recovery cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-7-claude-haiku-55-support/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-7-claude-haiku-55-support/",
      "title": "OpenClaw Adds Claude Haiku 5.5 Support",
      "summary": "OpenClaw now supports Claude Haiku 5.5 across Anthropic API and Claude CLI routes, with updated context, output, thinking, and pricing metadata.",
      "content_text": "OpenClaw added support for Claude Haiku 5.5 tonight, bringing Anthropic's newer lightweight model into the platform's catalog and runtime contracts.\n\n[PR #166706](https://github.com/openclaw/openclaw/pull/166706), \"feat(anthropic): support Claude Haiku 5.5,\" updates OpenClaw so users can select Haiku 5.5 through Anthropic API routes and Claude CLI routes. The `haiku` alias and default utility model now advance to 5.5, while explicit Haiku 4.5 selections remain available.\n\n## What Users Get\n\nThe PR encodes the published Haiku 5.5 contract in OpenClaw's provider catalog and shared Claude runtime logic.\n\nThat includes:\n\n- 1M context support\n- 128K output support\n- Adaptive thinking behavior\n- Tiered API pricing metadata\n- Claude CLI and direct Anthropic API selection\n- Continued support for explicit Haiku 4.5 references\n\nThis is a model-support update, not a Gateway migration. The PR does not add a new config format, mutate existing Gateway configuration, or publish a release by itself.\n\n## Why Model Catalog Updates Matter\n\nAdding a model to OpenClaw is more than placing a new string in a dropdown. The platform has to understand what the model can do across direct API calls, managed routes, CLI-backed execution, catalog aliases, pricing displays, tool behavior, thinking controls, replay policy, and compatibility checks.\n\nThat is especially important for Claude models because thinking and transcript replay are part of the runtime contract. The PR sets default effort to medium, allows thinking to be disabled, preserves forced tool support, and omits unsupported sampling and priority options.\n\nIt also fixes pricing equality logic so a discovered flat-rate row cannot hide the published long-prompt tier. That matters for users trying to estimate real operating costs before routing work to a new model.\n\n## Signed Thinking Replay\n\nOne follow-up repair in the PR is worth calling out. Review found that older signed Haiku 5.5 thinking could be stripped at the agent replay-policy boundary, even though transport-only replay preserved it.\n\nThe fix reuses OpenClaw's existing Claude thinking-prefix binding decision so prefix-bound Claude models retain signed thinking where appropriate. Legacy models, invalid signatures, and stale compaction signatures keep their previous behavior.\n\nThat is a small source change with a large correctness surface. If OpenClaw accepts a model's signed thinking prefix contract at transport time, the agent transcript sanitizer needs to preserve the same contract during later turns.\n\n## Proof And Validation\n\nThe PR includes direct and managed transport tests, provider catalog checks, CLI migration coverage, pricing regressions, forward-compatibility tests, and sanitizer replay tests.\n\nThere is also live proof. Claude CLI accepted `claude-haiku-5-5` with medium effort in an isolated one-shot run. Later authenticated proof through the Anthropic models API confirmed the model was reachable, and direct plus managed transports completed multi-turn signed-thinking scenarios with the preserved prefix reaching the API boundary.\n\nThe final head reached green CI and an exact-head review with no remaining actionable findings.\n\n## Bottom Line\n\nHaiku 5.5 is now a first-class OpenClaw model option. Users get the newer Claude utility path without losing explicit Haiku 4.5 compatibility, and the runtime now knows enough about context, output, pricing, thinking, and replay to route it intentionally.",
      "content_html": "<p>OpenClaw added support for Claude Haiku 5.5 tonight, bringing Anthropic's newer lightweight model into the platform's catalog and runtime contracts.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166706\">PR #166706</a>, \"feat(anthropic): support Claude Haiku 5.5,\" updates OpenClaw so users can select Haiku 5.5 through Anthropic API routes and Claude CLI routes. The <code>haiku</code> alias and default utility model now advance to 5.5, while explicit Haiku 4.5 selections remain available.</p><h2>What Users Get</h2><p>The PR encodes the published Haiku 5.5 contract in OpenClaw's provider catalog and shared Claude runtime logic.</p><p>That includes:</p><ul><li>1M context support</li><li>128K output support</li><li>Adaptive thinking behavior</li><li>Tiered API pricing metadata</li><li>Claude CLI and direct Anthropic API selection</li><li>Continued support for explicit Haiku 4.5 references</li></ul><p>This is a model-support update, not a Gateway migration. The PR does not add a new config format, mutate existing Gateway configuration, or publish a release by itself.</p><h2>Why Model Catalog Updates Matter</h2><p>Adding a model to OpenClaw is more than placing a new string in a dropdown. The platform has to understand what the model can do across direct API calls, managed routes, CLI-backed execution, catalog aliases, pricing displays, tool behavior, thinking controls, replay policy, and compatibility checks.</p><p>That is especially important for Claude models because thinking and transcript replay are part of the runtime contract. The PR sets default effort to medium, allows thinking to be disabled, preserves forced tool support, and omits unsupported sampling and priority options.</p><p>It also fixes pricing equality logic so a discovered flat-rate row cannot hide the published long-prompt tier. That matters for users trying to estimate real operating costs before routing work to a new model.</p><h2>Signed Thinking Replay</h2><p>One follow-up repair in the PR is worth calling out. Review found that older signed Haiku 5.5 thinking could be stripped at the agent replay-policy boundary, even though transport-only replay preserved it.</p><p>The fix reuses OpenClaw's existing Claude thinking-prefix binding decision so prefix-bound Claude models retain signed thinking where appropriate. Legacy models, invalid signatures, and stale compaction signatures keep their previous behavior.</p><p>That is a small source change with a large correctness surface. If OpenClaw accepts a model's signed thinking prefix contract at transport time, the agent transcript sanitizer needs to preserve the same contract during later turns.</p><h2>Proof And Validation</h2><p>The PR includes direct and managed transport tests, provider catalog checks, CLI migration coverage, pricing regressions, forward-compatibility tests, and sanitizer replay tests.</p><p>There is also live proof. Claude CLI accepted <code>claude-haiku-5-5</code> with medium effort in an isolated one-shot run. Later authenticated proof through the Anthropic models API confirmed the model was reachable, and direct plus managed transports completed multi-turn signed-thinking scenarios with the preserved prefix reaching the API boundary.</p><p>The final head reached green CI and an exact-head review with no remaining actionable findings.</p><h2>Bottom Line</h2><p>Haiku 5.5 is now a first-class OpenClaw model option. Users get the newer Claude utility path without losing explicit Haiku 4.5 compatibility, and the runtime now knows enough about context, output, pricing, thinking, and replay to route it intentionally.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-claude-haiku-55-support.png\" alt=\"OpenClaw Adds Claude Haiku 5.5 Support\" /></p>",
      "date_published": "2026-10-07T23:30:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Adds",
        "Claude",
        "Haiku",
        "Support"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-claude-haiku-55-support.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-claude-haiku-55-support.png",
          "mime_type": "image/png",
          "title": "OpenClaw Adds Claude Haiku 5.5 Support cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-7-knowledge-session-writes/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-7-knowledge-session-writes/",
      "title": "OpenClaw Knowledge Calls Survive Session Writes",
      "summary": "OpenClaw fixes intermittent Knowledge and memory-slot failures caused by a session's own bookkeeping writes during long-running turns.",
      "content_text": "OpenClaw merged a P1 session and memory reliability fix tonight for long agent turns that use Knowledge or other memory-slot tools.\n\n[PR #166758](https://github.com/openclaw/openclaw/pull/166758), \"fix(sessions): Knowledge calls still fail mid-turn while the session's own bookkeeping publishes,\" targets intermittent failures where memory calls could report that audience currency was unavailable while the same turn was still active.\n\nThe failure mode was subtle because the session was not actually reset, replaced, or deleted. It was simply writing its own routine bookkeeping.\n\n## The Failure Mode\n\nOpenClaw session writes are committed on a SQLite worker and then installed on the Gateway main thread. While a publication was pending for a session row, synchronous generation reads treated that session's identity and lifecycle revision as unknown.\n\nThat is correct for dangerous changes. If a session is reset, replaced, deleted, archived, or loses membership, memory access should stop until the new authority state is clear.\n\nThe bug was that harmless metadata writes could trigger the same uncertainty. A long turn may write activity, usage, or cache-state updates while it is still running. Knowledge also re-resolves its principal after internal awaits. Put those together and an agent could hit memory-audience currency failures caused by its own ordinary progress updates.\n\nThe PR notes a live example: a 54-minute Knowledge ingestion turn saw nine transient currency failures interleaved with successful Knowledge calls.\n\n## What Changed\n\nOpenClaw now records proof for identity-preserving publications. When the SQLite worker commits a publication, it compares the previous and current rows and marks sessions whose `sessionId` and `lifecycleRevision` are unchanged.\n\nSynchronous generation reads can ignore pending publications that carry that proof. Prepared reads still join every pending publication, so ordering remains strict for operations that need to observe the effects of in-flight writes.\n\nThat distinction is the whole fix:\n\n- Identity-preserving bookkeeping should not block the session's own memory calls.\n- Real resets and replacement paths should still make memory unavailable until authority is settled.\n- Publication paths without proof keep the existing fence.\n\nThe implementation also moves prepared-change record helpers into the SQLite entry-cache publication owner, keeping the authority logic close to the code that proves what changed.\n\n## Why It Matters\n\nKnowledge is most valuable during long, stateful turns: ingestion, project research, memory cleanup, documentation passes, and multi-step workflows. Those are exactly the turns most likely to produce internal bookkeeping writes while work is still underway.\n\nIntermittent memory failures inside that shape are frustrating because they feel nondeterministic. The same Knowledge call may fail once, then succeed later, even though the user's session did not visibly change.\n\nThis fix makes the memory boundary more precise. OpenClaw can keep rejecting stale or replaced sessions without treating every pending session-row write as an identity event.\n\n## Evidence From The PR\n\nThe regression tests reproduce the problem with real SQLite behavior. A synchronous `assertMemoryAudienceCurrent` fails on main during a held identity-preserving publication and passes after the fix, including the case where a child audience depends on a parent's in-flight write.\n\nThe control cases still fail when they should. A held reset still blocks synchronous currency checks and then revokes access. Prepared reads still wait for both identity-changing and identity-preserving publications where ordering matters.\n\nThe affected session, outbound, subagent, cron, memory audience, provider, and tool suites passed. One legacy ACP metadata migration test timed out during a concurrent heavy run, then passed alone.\n\n## Bottom Line\n\nOpenClaw's memory tools should now be steadier during long active turns. Knowledge can continue through ordinary session bookkeeping writes, while real lifecycle changes still preserve the authority boundary that keeps memory access safe.",
      "content_html": "<p>OpenClaw merged a P1 session and memory reliability fix tonight for long agent turns that use Knowledge or other memory-slot tools.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166758\">PR #166758</a>, \"fix(sessions): Knowledge calls still fail mid-turn while the session's own bookkeeping publishes,\" targets intermittent failures where memory calls could report that audience currency was unavailable while the same turn was still active.</p><p>The failure mode was subtle because the session was not actually reset, replaced, or deleted. It was simply writing its own routine bookkeeping.</p><h2>The Failure Mode</h2><p>OpenClaw session writes are committed on a SQLite worker and then installed on the Gateway main thread. While a publication was pending for a session row, synchronous generation reads treated that session's identity and lifecycle revision as unknown.</p><p>That is correct for dangerous changes. If a session is reset, replaced, deleted, archived, or loses membership, memory access should stop until the new authority state is clear.</p><p>The bug was that harmless metadata writes could trigger the same uncertainty. A long turn may write activity, usage, or cache-state updates while it is still running. Knowledge also re-resolves its principal after internal awaits. Put those together and an agent could hit memory-audience currency failures caused by its own ordinary progress updates.</p><p>The PR notes a live example: a 54-minute Knowledge ingestion turn saw nine transient currency failures interleaved with successful Knowledge calls.</p><h2>What Changed</h2><p>OpenClaw now records proof for identity-preserving publications. When the SQLite worker commits a publication, it compares the previous and current rows and marks sessions whose <code>sessionId</code> and <code>lifecycleRevision</code> are unchanged.</p><p>Synchronous generation reads can ignore pending publications that carry that proof. Prepared reads still join every pending publication, so ordering remains strict for operations that need to observe the effects of in-flight writes.</p><p>That distinction is the whole fix:</p><ul><li>Identity-preserving bookkeeping should not block the session's own memory calls.</li><li>Real resets and replacement paths should still make memory unavailable until authority is settled.</li><li>Publication paths without proof keep the existing fence.</li></ul><p>The implementation also moves prepared-change record helpers into the SQLite entry-cache publication owner, keeping the authority logic close to the code that proves what changed.</p><h2>Why It Matters</h2><p>Knowledge is most valuable during long, stateful turns: ingestion, project research, memory cleanup, documentation passes, and multi-step workflows. Those are exactly the turns most likely to produce internal bookkeeping writes while work is still underway.</p><p>Intermittent memory failures inside that shape are frustrating because they feel nondeterministic. The same Knowledge call may fail once, then succeed later, even though the user's session did not visibly change.</p><p>This fix makes the memory boundary more precise. OpenClaw can keep rejecting stale or replaced sessions without treating every pending session-row write as an identity event.</p><h2>Evidence From The PR</h2><p>The regression tests reproduce the problem with real SQLite behavior. A synchronous <code>assertMemoryAudienceCurrent</code> fails on main during a held identity-preserving publication and passes after the fix, including the case where a child audience depends on a parent's in-flight write.</p><p>The control cases still fail when they should. A held reset still blocks synchronous currency checks and then revokes access. Prepared reads still wait for both identity-changing and identity-preserving publications where ordering matters.</p><p>The affected session, outbound, subagent, cron, memory audience, provider, and tool suites passed. One legacy ACP metadata migration test timed out during a concurrent heavy run, then passed alone.</p><h2>Bottom Line</h2><p>OpenClaw's memory tools should now be steadier during long active turns. Knowledge can continue through ordinary session bookkeeping writes, while real lifecycle changes still preserve the authority boundary that keeps memory access safe.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-knowledge-session-writes.png\" alt=\"OpenClaw Knowledge Calls Survive Session Writes\" /></p>",
      "date_published": "2026-10-07T23:20:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Knowledge",
        "Calls",
        "Survive",
        "Session",
        "Writes"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-knowledge-session-writes.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-knowledge-session-writes.png",
          "mime_type": "image/png",
          "title": "OpenClaw Knowledge Calls Survive Session Writes cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-7-control-ui-operator-recovery/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-7-control-ui-operator-recovery/",
      "title": "OpenClaw Fixes Control UI Restart Authority",
      "summary": "OpenClaw now preserves authenticated Control UI operator authority after Gateway restarts without broadening admin scope or bypassing revocation.",
      "content_text": "OpenClaw merged a P1 authority-boundary fix tonight for Control UI turns that are interrupted by a Gateway restart.\n\n[PR #166553](https://github.com/openclaw/openclaw/pull/166553), \"fix: preserve operator access after Control UI restart recovery,\" addresses a failure where an authenticated Control UI turn could resume after restart but lose the operator authority needed to finish automation work.\n\nThe visible symptom was blunt: automation creation could fail with `missing scope: operator.admin` even though the original user had already been authenticated for the Control UI session. That is exactly the kind of recovery bug that feels random to operators, because the user did the right thing and the task only broke after process replacement.\n\n## What Changed\n\nThe key change is not simply \"give recovered work more scope.\" The PR is careful about that boundary.\n\nBefore the fix, recovery could dispatch through an unprofiled system principal with `operator.write`. Automation management correctly requires stronger operator authority, so the resumed work hit the scope fence. The dangerous shortcut would have been to grant every recovered system path admin-like access. OpenClaw did not do that.\n\nInstead, the source admission path now records bounded, credential-free authorization facts alongside the private recovery claim. When the Gateway recovers the interrupted source, it re-admits that exact source under the current policy and binds the replacement run to its current session, claim, runtime, and registration.\n\nThat means the resumed turn can continue the authorized operation it already owned, while stale or invalid authority still stops at the boundary.\n\n## Why Operators Should Care\n\nGateway restarts are normal in real deployments. Operators update OpenClaw, restart services, recover crashed processes, and run long Control UI workflows while other maintenance is happening.\n\nFor those setups, this fix makes recovery less brittle in a high-value place: automation management. If a Control UI task is creating or changing an automation, it should not lose its verified authority merely because the Gateway process was replaced mid-turn.\n\nAt the same time, authority recovery cannot become a hidden privilege escalation path. The PR keeps several revocation paths intact:\n\n- Credential rotation\n- Device removal\n- Profile or role changes\n- Grant revocation\n- Stale recovery claims\n- Replacement or closed run owners\n\nOlder interrupted turns without the private source record remain restricted. The practical fallback is explicit and human-readable: send a fresh authenticated message to continue privileged work.\n\n## What The Proof Shows\n\nThe PR includes real Gateway UI scenarios. One authenticated Control UI turn survives process replacement, automatically creates one disabled hourly automation through the actual tool and RPC scope fence, preserves the original session and user input, and clears the original recovery claim.\n\nIt also covers negative cases. Public token rotation and device changes revoke retained authority, source-less claims do not become privileged fallbacks, and delegated execution remains excluded by execution metadata.\n\nThat testing matters because this is not a simple UI bug. It touches Gateway recovery, session ownership, Control UI admission, automation scope checks, device-bound authority, and runtime cleanup.\n\n## Bottom Line\n\nThis is a meaningful reliability and security-boundary repair for OpenClaw operators. Authenticated Control UI work can now survive a Gateway restart without silently dropping the authority it already had, while revocation and stale-claim checks still have teeth.",
      "content_html": "<p>OpenClaw merged a P1 authority-boundary fix tonight for Control UI turns that are interrupted by a Gateway restart.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166553\">PR #166553</a>, \"fix: preserve operator access after Control UI restart recovery,\" addresses a failure where an authenticated Control UI turn could resume after restart but lose the operator authority needed to finish automation work.</p><p>The visible symptom was blunt: automation creation could fail with <code>missing scope: operator.admin</code> even though the original user had already been authenticated for the Control UI session. That is exactly the kind of recovery bug that feels random to operators, because the user did the right thing and the task only broke after process replacement.</p><h2>What Changed</h2><p>The key change is not simply \"give recovered work more scope.\" The PR is careful about that boundary.</p><p>Before the fix, recovery could dispatch through an unprofiled system principal with <code>operator.write</code>. Automation management correctly requires stronger operator authority, so the resumed work hit the scope fence. The dangerous shortcut would have been to grant every recovered system path admin-like access. OpenClaw did not do that.</p><p>Instead, the source admission path now records bounded, credential-free authorization facts alongside the private recovery claim. When the Gateway recovers the interrupted source, it re-admits that exact source under the current policy and binds the replacement run to its current session, claim, runtime, and registration.</p><p>That means the resumed turn can continue the authorized operation it already owned, while stale or invalid authority still stops at the boundary.</p><h2>Why Operators Should Care</h2><p>Gateway restarts are normal in real deployments. Operators update OpenClaw, restart services, recover crashed processes, and run long Control UI workflows while other maintenance is happening.</p><p>For those setups, this fix makes recovery less brittle in a high-value place: automation management. If a Control UI task is creating or changing an automation, it should not lose its verified authority merely because the Gateway process was replaced mid-turn.</p><p>At the same time, authority recovery cannot become a hidden privilege escalation path. The PR keeps several revocation paths intact:</p><ul><li>Credential rotation</li><li>Device removal</li><li>Profile or role changes</li><li>Grant revocation</li><li>Stale recovery claims</li><li>Replacement or closed run owners</li></ul><p>Older interrupted turns without the private source record remain restricted. The practical fallback is explicit and human-readable: send a fresh authenticated message to continue privileged work.</p><h2>What The Proof Shows</h2><p>The PR includes real Gateway UI scenarios. One authenticated Control UI turn survives process replacement, automatically creates one disabled hourly automation through the actual tool and RPC scope fence, preserves the original session and user input, and clears the original recovery claim.</p><p>It also covers negative cases. Public token rotation and device changes revoke retained authority, source-less claims do not become privileged fallbacks, and delegated execution remains excluded by execution metadata.</p><p>That testing matters because this is not a simple UI bug. It touches Gateway recovery, session ownership, Control UI admission, automation scope checks, device-bound authority, and runtime cleanup.</p><h2>Bottom Line</h2><p>This is a meaningful reliability and security-boundary repair for OpenClaw operators. Authenticated Control UI work can now survive a Gateway restart without silently dropping the authority it already had, while revocation and stale-claim checks still have teeth.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-control-ui-operator-recovery.png\" alt=\"OpenClaw Fixes Control UI Restart Authority\" /></p>",
      "date_published": "2026-10-07T23:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Control",
        "Restart",
        "Authority"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-control-ui-operator-recovery.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-control-ui-operator-recovery.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Control UI Restart Authority cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-7-claude-cli-profile-selection/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-7-claude-cli-profile-selection/",
      "title": "OpenClaw Fixes Claude CLI Profile Selection",
      "summary": "OpenClaw plugin completions now use the configured Claude CLI account order instead of failing with 401 when no profile is selected.",
      "content_text": "OpenClaw merged a P1 fix this morning for Claude CLI plugin completions, including Memory Dreaming, when an Anthropic managed profile exists but no profile is explicitly selected.\n\n[PR #166173](https://github.com/openclaw/openclaw/pull/166173), \"fix: Claude CLI plugin completions return 401 without a selected profile,\" makes isolated plugin completions use the same shared CLI auth selector that user and cron callers already use.\n\nBefore this change, an ordered managed Anthropic profile could exist, but the isolated-completion caller forwarded only the request's optional profile. If no explicit profile was set on that request, plugin completions could fail with HTTP 401 or fall back to another provider.\n\n## What Users Get\n\nFresh plugin completions now follow the configured account order while preserving explicit account selections and native Claude login behavior.\n\nThat means a user can configure Anthropic account order once, then expect plugin-owned completions to resolve the same way as other OpenClaw callers. The fix does not add new configuration, change SDK surface area, introduce a migration, or create a plugin-owned authentication policy.\n\nThis is especially relevant for plugin workflows that call into subagent completion behind the scenes. Memory Dreaming is the named example in the PR, but the underlying issue was in the isolated-completion path rather than a single plugin.\n\n## What Changed Under The Hood\n\nThe Claude CLI backend intentionally disables runner auto-selection. Selection has to happen before dispatch, through OpenClaw's shared selector, because that selector owns account eligibility and native-login safeguards.\n\nThe isolated-completion caller now uses that shared selector before it dispatches work to the CLI runner. The PR also documents plugin behavior, covers both explicit managed and native selections, and removes an incorrect comment about auto-selection.\n\nThe important boundary is that native Claude login remains native. If a setup uses the CLI's own logged-in state, OpenClaw should not inject a managed credential into that flow.\n\n## Evidence From the PR\n\nThe PR includes a real Gateway proof with a synthetic plugin that uses documented plugin SDK entry points, registers a Gateway method, and calls `api.runtime.subagent.complete`. Each case starts a real Gateway with fresh state, config, and home directories on an ephemeral loopback port.\n\nFor the managed-order case, two synthetic profiles were seeded through the supported auth CLI, and account order selected the second one. The plugin supplied no profile. On the candidate, the RPC succeeded and the fake Claude executable confirmed that the selected credential reached file descriptor transport without appearing in argv or environment. On main production, the same managed-order case failed with HTTP 401 because no selected profile was forwarded.\n\nThe native-only candidate case also succeeded, with no managed credential descriptor attached. That proves the fix does not trample native Claude login.\n\nBeyond the live proof, exact-head CI passed. The focused isolated-completion test file reported 51 passes, seven related auth and Gateway files passed in routed Vitest shards, core type checks passed, and the full build passed.\n\n## Why It Matters\n\nOpenClaw's model and auth system is increasingly multi-path: user turns, cron jobs, plugins, native CLI runtimes, and managed profiles can all meet in the same installation. The more paths exist, the more important it is that they share one account-selection contract.\n\nThis PR closes one of those gaps. Plugin completions now use the configured Claude CLI account order like the rest of the product, instead of surprising users with a 401 at the moment a background completion tries to run.",
      "content_html": "<p>OpenClaw merged a P1 fix this morning for Claude CLI plugin completions, including Memory Dreaming, when an Anthropic managed profile exists but no profile is explicitly selected.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166173\">PR #166173</a>, \"fix: Claude CLI plugin completions return 401 without a selected profile,\" makes isolated plugin completions use the same shared CLI auth selector that user and cron callers already use.</p><p>Before this change, an ordered managed Anthropic profile could exist, but the isolated-completion caller forwarded only the request's optional profile. If no explicit profile was set on that request, plugin completions could fail with HTTP 401 or fall back to another provider.</p><h2>What Users Get</h2><p>Fresh plugin completions now follow the configured account order while preserving explicit account selections and native Claude login behavior.</p><p>That means a user can configure Anthropic account order once, then expect plugin-owned completions to resolve the same way as other OpenClaw callers. The fix does not add new configuration, change SDK surface area, introduce a migration, or create a plugin-owned authentication policy.</p><p>This is especially relevant for plugin workflows that call into subagent completion behind the scenes. Memory Dreaming is the named example in the PR, but the underlying issue was in the isolated-completion path rather than a single plugin.</p><h2>What Changed Under The Hood</h2><p>The Claude CLI backend intentionally disables runner auto-selection. Selection has to happen before dispatch, through OpenClaw's shared selector, because that selector owns account eligibility and native-login safeguards.</p><p>The isolated-completion caller now uses that shared selector before it dispatches work to the CLI runner. The PR also documents plugin behavior, covers both explicit managed and native selections, and removes an incorrect comment about auto-selection.</p><p>The important boundary is that native Claude login remains native. If a setup uses the CLI's own logged-in state, OpenClaw should not inject a managed credential into that flow.</p><h2>Evidence From the PR</h2><p>The PR includes a real Gateway proof with a synthetic plugin that uses documented plugin SDK entry points, registers a Gateway method, and calls <code>api.runtime.subagent.complete</code>. Each case starts a real Gateway with fresh state, config, and home directories on an ephemeral loopback port.</p><p>For the managed-order case, two synthetic profiles were seeded through the supported auth CLI, and account order selected the second one. The plugin supplied no profile. On the candidate, the RPC succeeded and the fake Claude executable confirmed that the selected credential reached file descriptor transport without appearing in argv or environment. On main production, the same managed-order case failed with HTTP 401 because no selected profile was forwarded.</p><p>The native-only candidate case also succeeded, with no managed credential descriptor attached. That proves the fix does not trample native Claude login.</p><p>Beyond the live proof, exact-head CI passed. The focused isolated-completion test file reported 51 passes, seven related auth and Gateway files passed in routed Vitest shards, core type checks passed, and the full build passed.</p><h2>Why It Matters</h2><p>OpenClaw's model and auth system is increasingly multi-path: user turns, cron jobs, plugins, native CLI runtimes, and managed profiles can all meet in the same installation. The more paths exist, the more important it is that they share one account-selection contract.</p><p>This PR closes one of those gaps. Plugin completions now use the configured Claude CLI account order like the rest of the product, instead of surprising users with a 401 at the moment a background completion tries to run.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-claude-cli-profile-selection.png\" alt=\"OpenClaw Fixes Claude CLI Profile Selection\" /></p>",
      "date_published": "2026-10-07T08:15:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Fixes",
        "Claude",
        "Profile",
        "Selection"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-claude-cli-profile-selection.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-claude-cli-profile-selection.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Claude CLI Profile Selection cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-7-cron-subagent-memory-access/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-7-cron-subagent-memory-access/",
      "title": "OpenClaw Cron Jobs Stop Breaking Subagent Memory",
      "summary": "OpenClaw cron jobs that reuse an existing chat no longer rotate session identity in a way that cuts subagents off from memory tools.",
      "content_text": "OpenClaw merged a P1 cron fix this morning for a nasty interaction between scheduled jobs, existing chat sessions, subagents, and memory access.\n\n[PR #166417](https://github.com/openclaw/openclaw/pull/166417), \"fix(cron): subagents lose memory access when a cron job runs in their parent chat,\" addresses a case where a scheduled check-in could make already-spawned child agents lose access to Knowledge and other memory-slot tools.\n\nThe failure mode was especially awkward because the cron job did not have to reset the session intentionally. A cron run targeting an existing conversation could rotate the parent session's lifecycle revision. Existing subagents still held the old parent revision, so later memory calls failed with an authority error until the subagent was respawned.\n\n## What Users Saw\n\nThe PR describes a live failure on a main Gateway: a five-minute check-in cron job targeted a dashboard session, rotated that session's revision, and left all eight previously spawned subagents unable to call Knowledge. Their calls were denied with a trusted-agent or trusted-session authority error.\n\nThat matters because cron is often used for recurring background work inside the same long-lived conversation:\n\n- Morning or evening status checks\n- Project watchers\n- Inbox or calendar scans\n- Scheduled summaries\n- Periodic agent workflows that depend on prior context\n\nThose workflows should not silently invalidate the child agents that the parent conversation already trusts.\n\n## What Changed\n\nOpenClaw now keeps the existing lifecycle revision when a cron run reuses a live session in place. A new revision is still minted for cases that genuinely represent a new incarnation: new sessions, stale resets, forced rollovers, different source sessions, rows without a revision, and exact-run `cron:` sessions.\n\nFor exact-run cron sessions, the run happens in a hidden `:run:` row while continuation ownership and superseded-base checks use the base revision as the per-run generation.\n\nThe fix also tightens persistence ownership. Because reused sessions now share a revision with other writers, OpenClaw requires the session id of the run's last committed row. A same-revision session-id replacement is rejected rather than adopted.\n\n## Why This Matters\n\nMemory access in OpenClaw is authority-sensitive for good reasons. A child agent should not inherit memory access across a real parent reset, and the PR keeps that behavior. The bug was that an ordinary in-place cron run looked like a reset to the owners responsible for memory audience leases.\n\nThe improved behavior is more precise: reuse keeps lineage; real resets still revoke it.\n\nThat distinction is important for users building persistent agent workflows. Cron should be a way to add scheduled motion to an existing conversation, not a background event that accidentally severs the conversation's children from their memory scope.\n\n## Evidence From the PR\n\nThe after-fix proof used a fresh dashboard session and a spawned subagent. The subagent successfully called `knowledge_status`, then a cron job ran in the parent session and replied. The parent revision stayed the same, and the same child could call `knowledge_status` again with no memory-audience denial.\n\nAs a control, the authors reset the parent session. That minted a new revision, and the old child was then correctly denied memory access.\n\nThe test suite adds coverage for in-place reuse, fresh sessions, stale sessions, forced rollovers, differing source sessions, revision-less rows, and exact-run cron sessions. The changed cron isolated-agent tests passed, including a 519-test run across 45 files.\n\n## Bottom Line\n\nThis fix makes cron safer for long-lived OpenClaw workspaces. Scheduled jobs can now run inside an existing chat without making the chat's subagents look untrusted, while real resets still draw the right authority boundary.",
      "content_html": "<p>OpenClaw merged a P1 cron fix this morning for a nasty interaction between scheduled jobs, existing chat sessions, subagents, and memory access.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166417\">PR #166417</a>, \"fix(cron): subagents lose memory access when a cron job runs in their parent chat,\" addresses a case where a scheduled check-in could make already-spawned child agents lose access to Knowledge and other memory-slot tools.</p><p>The failure mode was especially awkward because the cron job did not have to reset the session intentionally. A cron run targeting an existing conversation could rotate the parent session's lifecycle revision. Existing subagents still held the old parent revision, so later memory calls failed with an authority error until the subagent was respawned.</p><h2>What Users Saw</h2><p>The PR describes a live failure on a main Gateway: a five-minute check-in cron job targeted a dashboard session, rotated that session's revision, and left all eight previously spawned subagents unable to call Knowledge. Their calls were denied with a trusted-agent or trusted-session authority error.</p><p>That matters because cron is often used for recurring background work inside the same long-lived conversation:</p><ul><li>Morning or evening status checks</li><li>Project watchers</li><li>Inbox or calendar scans</li><li>Scheduled summaries</li><li>Periodic agent workflows that depend on prior context</li></ul><p>Those workflows should not silently invalidate the child agents that the parent conversation already trusts.</p><h2>What Changed</h2><p>OpenClaw now keeps the existing lifecycle revision when a cron run reuses a live session in place. A new revision is still minted for cases that genuinely represent a new incarnation: new sessions, stale resets, forced rollovers, different source sessions, rows without a revision, and exact-run <code>cron:</code> sessions.</p><p>For exact-run cron sessions, the run happens in a hidden <code>:run:</code> row while continuation ownership and superseded-base checks use the base revision as the per-run generation.</p><p>The fix also tightens persistence ownership. Because reused sessions now share a revision with other writers, OpenClaw requires the session id of the run's last committed row. A same-revision session-id replacement is rejected rather than adopted.</p><h2>Why This Matters</h2><p>Memory access in OpenClaw is authority-sensitive for good reasons. A child agent should not inherit memory access across a real parent reset, and the PR keeps that behavior. The bug was that an ordinary in-place cron run looked like a reset to the owners responsible for memory audience leases.</p><p>The improved behavior is more precise: reuse keeps lineage; real resets still revoke it.</p><p>That distinction is important for users building persistent agent workflows. Cron should be a way to add scheduled motion to an existing conversation, not a background event that accidentally severs the conversation's children from their memory scope.</p><h2>Evidence From the PR</h2><p>The after-fix proof used a fresh dashboard session and a spawned subagent. The subagent successfully called <code>knowledge_status</code>, then a cron job ran in the parent session and replied. The parent revision stayed the same, and the same child could call <code>knowledge_status</code> again with no memory-audience denial.</p><p>As a control, the authors reset the parent session. That minted a new revision, and the old child was then correctly denied memory access.</p><p>The test suite adds coverage for in-place reuse, fresh sessions, stale sessions, forced rollovers, differing source sessions, revision-less rows, and exact-run cron sessions. The changed cron isolated-agent tests passed, including a 519-test run across 45 files.</p><h2>Bottom Line</h2><p>This fix makes cron safer for long-lived OpenClaw workspaces. Scheduled jobs can now run inside an existing chat without making the chat's subagents look untrusted, while real resets still draw the right authority boundary.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-cron-subagent-memory-access.png\" alt=\"OpenClaw Cron Jobs Stop Breaking Subagent Memory\" /></p>",
      "date_published": "2026-10-07T08:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "Guides",
        "Memory",
        "Cron",
        "Jobs",
        "Stop",
        "Breaking",
        "Subagent"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-cron-subagent-memory-access.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-cron-subagent-memory-access.png",
          "mime_type": "image/png",
          "title": "OpenClaw Cron Jobs Stop Breaking Subagent Memory cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-7-update-prevalidation-large-plugin-trees/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-7-update-prevalidation-large-plugin-trees/",
      "title": "OpenClaw Speeds Up Large Plugin Tree Updates",
      "summary": "OpenClaw update prevalidation now avoids redundant bundled-plugin copying and validates large plugin trees much faster.",
      "content_text": "OpenClaw merged a substantial updater performance fix this morning aimed at a very specific but painful case: update prevalidation when configured plugin paths point at installed bundled plugins or very large dependency trees.\n\n[PR #166317](https://github.com/openclaw/openclaw/pull/166317), \"fix(update): reduce slow prevalidation for large plugin trees,\" keeps the existing safety checks intact while cutting down repeated filesystem work. The headline number from the PR is striking: in one production-shaped rehearsal, complete prevalidation including cleanup dropped from 12 minutes 49 seconds to 5 minutes 42 seconds.\n\nThat is not framed as a universal upgrade-speed guarantee. It is a single-host measurement for a workload with large plugin trees. But it does show where the update path was spending avoidable time.\n\n## What Changed\n\nThe updater now captures the serving installation's actual bundled-plugin selection before entering the staged worker. That lets the candidate update process match bundled plugin entries against the real installed source, instead of rediscovering the candidate as the source and copying the old installation unnecessarily.\n\nThe PR also adds throttled progress receipts for completed filesystem work. Those receipts feed the existing I/O watchdog without turning every file into a ledger event.\n\nFor large guarded copies, OpenClaw now uses the existing four-worker pool while preserving the same custody rules:\n\n- Original destination identity stays attached to copied files.\n- Source fingerprints and byte hashes remain part of validation.\n- Publication stays create-only.\n- Native operation fencing, worker stop, drain, and join ordering remain in place.\n\nIn plain terms: this is a performance fix, not a shortcut around update safety.\n\n## Why It Matters\n\nOpenClaw updates already do more than copy a package into place. They validate candidate state, snapshot databases, inspect plugins, run startup checks, preserve rollback information, and guard filesystem operations. That safety work is valuable, but it can become expensive when a plugin tree is large or when bundled-plugin paths make the updater repeat work it can prove is already known.\n\nThe PR reports plugin allocation falling from 4.73 GB to 1.56 GB in the same private rehearsal. State preparation also dropped from 638.258 seconds to 238.465 seconds.\n\nFor operators, the practical benefit is shorter update preparation windows on installations with heavier plugin layouts. The PR says the full improvement requires both the updated installed driver and candidate worker. Older driver and worker combinations remain compatible, but may keep some previous copying and progress-probe overhead.\n\n## Evidence From the Merge\n\nThe final verification included a native macOS update rehearsal from the published `2026.9.8` updater to the candidate package. The PR says the update completed with exit 0 and status `ok`, preserving an ordinary custom plugin with 1,024 payload files totaling 256 MiB, a read-only file, and a relative symlink from the private canary root.\n\nThe candidate also passed 182 focused tests, a scoped changed-file gate, a full package build, and canonical tarball integrity checks. The authors disclose the limits: the native proof exercised owned foreground Gateways and package update behavior, not LaunchAgent activation.\n\n## Bottom Line\n\nThis is the kind of infrastructure change most users only notice when it is missing. OpenClaw update prevalidation still performs the same classes of validation, but it now avoids a large chunk of redundant plugin-tree work in the path where updates could feel stuck.",
      "content_html": "<p>OpenClaw merged a substantial updater performance fix this morning aimed at a very specific but painful case: update prevalidation when configured plugin paths point at installed bundled plugins or very large dependency trees.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166317\">PR #166317</a>, \"fix(update): reduce slow prevalidation for large plugin trees,\" keeps the existing safety checks intact while cutting down repeated filesystem work. The headline number from the PR is striking: in one production-shaped rehearsal, complete prevalidation including cleanup dropped from 12 minutes 49 seconds to 5 minutes 42 seconds.</p><p>That is not framed as a universal upgrade-speed guarantee. It is a single-host measurement for a workload with large plugin trees. But it does show where the update path was spending avoidable time.</p><h2>What Changed</h2><p>The updater now captures the serving installation's actual bundled-plugin selection before entering the staged worker. That lets the candidate update process match bundled plugin entries against the real installed source, instead of rediscovering the candidate as the source and copying the old installation unnecessarily.</p><p>The PR also adds throttled progress receipts for completed filesystem work. Those receipts feed the existing I/O watchdog without turning every file into a ledger event.</p><p>For large guarded copies, OpenClaw now uses the existing four-worker pool while preserving the same custody rules:</p><ul><li>Original destination identity stays attached to copied files.</li><li>Source fingerprints and byte hashes remain part of validation.</li><li>Publication stays create-only.</li><li>Native operation fencing, worker stop, drain, and join ordering remain in place.</li></ul><p>In plain terms: this is a performance fix, not a shortcut around update safety.</p><h2>Why It Matters</h2><p>OpenClaw updates already do more than copy a package into place. They validate candidate state, snapshot databases, inspect plugins, run startup checks, preserve rollback information, and guard filesystem operations. That safety work is valuable, but it can become expensive when a plugin tree is large or when bundled-plugin paths make the updater repeat work it can prove is already known.</p><p>The PR reports plugin allocation falling from 4.73 GB to 1.56 GB in the same private rehearsal. State preparation also dropped from 638.258 seconds to 238.465 seconds.</p><p>For operators, the practical benefit is shorter update preparation windows on installations with heavier plugin layouts. The PR says the full improvement requires both the updated installed driver and candidate worker. Older driver and worker combinations remain compatible, but may keep some previous copying and progress-probe overhead.</p><h2>Evidence From the Merge</h2><p>The final verification included a native macOS update rehearsal from the published <code>2026.9.8</code> updater to the candidate package. The PR says the update completed with exit 0 and status <code>ok</code>, preserving an ordinary custom plugin with 1,024 payload files totaling 256 MiB, a read-only file, and a relative symlink from the private canary root.</p><p>The candidate also passed 182 focused tests, a scoped changed-file gate, a full package build, and canonical tarball integrity checks. The authors disclose the limits: the native proof exercised owned foreground Gateways and package update behavior, not LaunchAgent activation.</p><h2>Bottom Line</h2><p>This is the kind of infrastructure change most users only notice when it is missing. OpenClaw update prevalidation still performs the same classes of validation, but it now avoids a large chunk of redundant plugin-tree work in the path where updates could feel stuck.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-update-prevalidation-large-plugin-trees.png\" alt=\"OpenClaw Speeds Up Large Plugin Tree Updates\" /></p>",
      "date_published": "2026-10-07T08:05:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Speeds",
        "Large",
        "Plugin",
        "Tree",
        "Updates"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-update-prevalidation-large-plugin-trees.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-7-update-prevalidation-large-plugin-trees.png",
          "mime_type": "image/png",
          "title": "OpenClaw Speeds Up Large Plugin Tree Updates cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-6-reasoning-tool-turns/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-6-reasoning-tool-turns/",
      "title": "OpenClaw Fixes Reasoning Model Tool Turns",
      "summary": "OpenClaw now routes official OpenAI reasoning-model tool turns through Responses when Chat Completions cannot carry them.",
      "content_text": "OpenClaw merged a P1 model-routing fix today for a painful mismatch between reasoning models, tools, and the official OpenAI Chat Completions endpoint.\n\nThe fix landed in [PR #166182](https://github.com/openclaw/openclaw/pull/166182), \"fix: tool turns fail for reasoning models on api.openai.com Completions routes.\" It addresses agent turns where a reasoning model such as `gpt-6-astra` or `gpt-6.1-sol` runs through an `openai-completions` provider on `api.openai.com`.\n\n## The Failure Mode\n\nThe core problem was endpoint capability. Current GPT reasoning models on OpenAI Chat Completions reject function tools when reasoning is enabled and direct callers toward `/v1/responses`.\n\nSome models made the failure unavoidable. The PR says Astra and GPT-6.1 Sol also reject `reasoning_effort: \"none\"`, so there was no Chat Completions payload that could serve their tool-bearing turns. GPT-5.6 tool turns ran with reasoning switched off, while GPT-6 Sol and GPT-5.4 first hit a 400 and then retried with thinking off.\n\nFor users, that meant tool turns either failed outright or silently lost the reasoning behavior they had selected.\n\n## What OpenClaw Does Now\n\nWhen the managed Completions transport targets the official OpenAI endpoint with a reasoning model and tools, OpenClaw sends the turn through its existing managed Responses transport.\n\nThe route remains authored as `openai-completions`. The provider, credentials, endpoint host, auth, and runtime-selection behavior stay the same. The change is the internal transport path OpenClaw chooses when the official endpoint cannot represent the requested turn shape.\n\nThe PR also notes two compatibility details:\n\n- Forced tool choices are converted into the Responses format.\n- An unset reasoning selector keeps the Completions default of `high`.\n\nCustom or proxy `openai-completions` endpoints, Azure, and direct `@openclaw/ai` Chat Completions SDK usage are unchanged.\n\n## Why This Is The Right Boundary\n\nThe important distinction is between a user's configured route and the wire protocol needed to make that route work against the official provider.\n\nUsers chose a reasoning model with tools. OpenClaw already has a Responses transport that can carry that shape. Rather than forcing users to learn which OpenAI endpoint can represent which feature combination, the managed route now bridges the mismatch where it can do so without changing the visible provider contract.\n\nThat is a better default for agent work. Tool calls are not an edge feature in OpenClaw; they are central to most useful turns. A selected reasoning level should survive the first tool call whenever the provider offers a compatible API path.\n\n## Evidence From the PR\n\nThe PR includes regression coverage for the official OpenAI Completions route with reasoning models and tools, including Astra and GPT-6.1 Sol. It also covers forced tool choices, preserved reasoning level, and the boundary that non-official endpoints are not rerouted.\n\nThe author reports that tool-bearing turns on these routes now succeed and keep the selected reasoning level. That is the user-visible win: no HTTP 400 wall for Astra or GPT-6.1 Sol tool turns, and no quiet fallback that strips reasoning from models where the user expected it.\n\n## Bottom Line\n\nOpenClaw is smoothing over a provider API split without surprising custom endpoints. Official OpenAI reasoning-model tool turns now use the transport that can actually carry them, while the configured `openai-completions` route keeps behaving like the route users authored.",
      "content_html": "<p>OpenClaw merged a P1 model-routing fix today for a painful mismatch between reasoning models, tools, and the official OpenAI Chat Completions endpoint.</p><p>The fix landed in <a href=\"https://github.com/openclaw/openclaw/pull/166182\">PR #166182</a>, \"fix: tool turns fail for reasoning models on api.openai.com Completions routes.\" It addresses agent turns where a reasoning model such as <code>gpt-6-astra</code> or <code>gpt-6.1-sol</code> runs through an <code>openai-completions</code> provider on <code>api.openai.com</code>.</p><h2>The Failure Mode</h2><p>The core problem was endpoint capability. Current GPT reasoning models on OpenAI Chat Completions reject function tools when reasoning is enabled and direct callers toward <code>/v1/responses</code>.</p><p>Some models made the failure unavoidable. The PR says Astra and GPT-6.1 Sol also reject <code>reasoning_effort: \"none\"</code>, so there was no Chat Completions payload that could serve their tool-bearing turns. GPT-5.6 tool turns ran with reasoning switched off, while GPT-6 Sol and GPT-5.4 first hit a 400 and then retried with thinking off.</p><p>For users, that meant tool turns either failed outright or silently lost the reasoning behavior they had selected.</p><h2>What OpenClaw Does Now</h2><p>When the managed Completions transport targets the official OpenAI endpoint with a reasoning model and tools, OpenClaw sends the turn through its existing managed Responses transport.</p><p>The route remains authored as <code>openai-completions</code>. The provider, credentials, endpoint host, auth, and runtime-selection behavior stay the same. The change is the internal transport path OpenClaw chooses when the official endpoint cannot represent the requested turn shape.</p><p>The PR also notes two compatibility details:</p><ul><li>Forced tool choices are converted into the Responses format.</li><li>An unset reasoning selector keeps the Completions default of <code>high</code>.</li></ul><p>Custom or proxy <code>openai-completions</code> endpoints, Azure, and direct <code>@openclaw/ai</code> Chat Completions SDK usage are unchanged.</p><h2>Why This Is The Right Boundary</h2><p>The important distinction is between a user's configured route and the wire protocol needed to make that route work against the official provider.</p><p>Users chose a reasoning model with tools. OpenClaw already has a Responses transport that can carry that shape. Rather than forcing users to learn which OpenAI endpoint can represent which feature combination, the managed route now bridges the mismatch where it can do so without changing the visible provider contract.</p><p>That is a better default for agent work. Tool calls are not an edge feature in OpenClaw; they are central to most useful turns. A selected reasoning level should survive the first tool call whenever the provider offers a compatible API path.</p><h2>Evidence From the PR</h2><p>The PR includes regression coverage for the official OpenAI Completions route with reasoning models and tools, including Astra and GPT-6.1 Sol. It also covers forced tool choices, preserved reasoning level, and the boundary that non-official endpoints are not rerouted.</p><p>The author reports that tool-bearing turns on these routes now succeed and keep the selected reasoning level. That is the user-visible win: no HTTP 400 wall for Astra or GPT-6.1 Sol tool turns, and no quiet fallback that strips reasoning from models where the user expected it.</p><h2>Bottom Line</h2><p>OpenClaw is smoothing over a provider API split without surprising custom endpoints. Official OpenAI reasoning-model tool turns now use the transport that can actually carry them, while the configured <code>openai-completions</code> route keeps behaving like the route users authored.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-reasoning-tool-turns.png\" alt=\"OpenClaw Fixes Reasoning Model Tool Turns\" /></p>",
      "date_published": "2026-10-06T23:15:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Fixes",
        "Reasoning",
        "Model",
        "Tool",
        "Turns"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-reasoning-tool-turns.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-reasoning-tool-turns.png",
          "mime_type": "image/png",
          "title": "OpenClaw Fixes Reasoning Model Tool Turns cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-6-claude-cli-login-refresh/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-6-claude-cli-login-refresh/",
      "title": "OpenClaw Refreshes Claude CLI Models After Login",
      "summary": "OpenClaw now rechecks Claude CLI login state after Gateway startup so model lists recover without a restart.",
      "content_text": "OpenClaw merged a Claude CLI model-discovery fix tonight that removes a frustrating restart requirement from the Control UI and `openclaw models list`.\n\n[PR #166305](https://github.com/openclaw/openclaw/pull/166305), \"fix(models): recheck Claude CLI login for model lists after Gateway startup,\" addresses a stale-auth state. If Claude Code was logged out when the Gateway started, Claude CLI models could remain hidden or disabled even after the user logged in. The same stale state applied in reverse: logging out after startup was not reflected either.\n\n## What Users See Now\n\nAfter `claude auth login`, OpenClaw should show the right Claude CLI-backed models within about a minute. Users do not need to restart the Gateway, reload the page, or force a manual `models.list` refresh.\n\nThe logged-out presentation stays the same. The change is about keeping the model catalog aligned with the CLI's real authentication state after startup.\n\nThat is especially useful for setups where Claude Code is a normal runtime path rather than a one-off fallback. If the Gateway starts before the CLI login exists, the model picker should not permanently inherit that cold-start fact.\n\n## What Changed Under The Hood\n\nThe model picker already used Claude Code's login state when deciding availability. The problem was that the ordinary read path did not recheck static CLI-backed providers.\n\nOpenClaw now rechecks native CLI-backend logins for backends whose manifest declares synthetic auth references. Today, that means `claude-cli`.\n\nThe recheck is deliberately bounded:\n\n- It runs in the background while model lists are being read.\n- It deduplicates through the existing `loadAuth` path.\n- It publishes a new catalog only when login state changed.\n- It keeps startup's once-per-generation probe behavior intact.\n\nIn other words, a busy picker does not become a CLI process launcher. The PR reports at most one `claude auth status --json` probe per prepared catalog owner per 60 seconds, and only while someone is reading model lists.\n\n## Why This Matters\n\nModel availability is part of the front door for agent work. If the picker says Claude models are disabled after a user has already logged in, OpenClaw feels broken even when the underlying CLI is ready.\n\nThe fix also supports the newer documentation direction around Claude Code. Adjacent docs work now describes Claude CLI as a primary runtime path with canonical `anthropic/*` model references and an `agentRuntime` pin. If that is the recommended setup, availability refresh has to behave like a live runtime signal, not a startup-only snapshot.\n\n## Evidence From the PR\n\nThe regression test starts a real in-process Gateway, uses the real Anthropic plugin and `models.list`, and replaces only the `claude` executable with a fixture that records auth probes.\n\nThe test begins with Claude unavailable at startup. A login inside the recheck window does not spawn another probe. After the window, the model becomes available with exactly one more probe. A logout later makes it unavailable again with one more probe. The PR says the test times out on the old catalog behavior and passes in about five seconds on the fix.\n\nThe live Control UI proof used two isolated Gateways, one on origin/main and one on the branch. Both started logged out. After placing Claude credentials into the isolated config directory, origin/main still showed disabled rows minutes later, while the fixed branch enabled the configured Claude models on the first open after the recheck window.\n\n## Bottom Line\n\nThis is the kind of model-runtime polish that saves users from ritual restarts. OpenClaw now treats Claude CLI login state as something that can change during the life of a Gateway, and the model picker follows along without hammering the CLI.",
      "content_html": "<p>OpenClaw merged a Claude CLI model-discovery fix tonight that removes a frustrating restart requirement from the Control UI and <code>openclaw models list</code>.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/166305\">PR #166305</a>, \"fix(models): recheck Claude CLI login for model lists after Gateway startup,\" addresses a stale-auth state. If Claude Code was logged out when the Gateway started, Claude CLI models could remain hidden or disabled even after the user logged in. The same stale state applied in reverse: logging out after startup was not reflected either.</p><h2>What Users See Now</h2><p>After <code>claude auth login</code>, OpenClaw should show the right Claude CLI-backed models within about a minute. Users do not need to restart the Gateway, reload the page, or force a manual <code>models.list</code> refresh.</p><p>The logged-out presentation stays the same. The change is about keeping the model catalog aligned with the CLI's real authentication state after startup.</p><p>That is especially useful for setups where Claude Code is a normal runtime path rather than a one-off fallback. If the Gateway starts before the CLI login exists, the model picker should not permanently inherit that cold-start fact.</p><h2>What Changed Under The Hood</h2><p>The model picker already used Claude Code's login state when deciding availability. The problem was that the ordinary read path did not recheck static CLI-backed providers.</p><p>OpenClaw now rechecks native CLI-backend logins for backends whose manifest declares synthetic auth references. Today, that means <code>claude-cli</code>.</p><p>The recheck is deliberately bounded:</p><ul><li>It runs in the background while model lists are being read.</li><li>It deduplicates through the existing <code>loadAuth</code> path.</li><li>It publishes a new catalog only when login state changed.</li><li>It keeps startup's once-per-generation probe behavior intact.</li></ul><p>In other words, a busy picker does not become a CLI process launcher. The PR reports at most one <code>claude auth status --json</code> probe per prepared catalog owner per 60 seconds, and only while someone is reading model lists.</p><h2>Why This Matters</h2><p>Model availability is part of the front door for agent work. If the picker says Claude models are disabled after a user has already logged in, OpenClaw feels broken even when the underlying CLI is ready.</p><p>The fix also supports the newer documentation direction around Claude Code. Adjacent docs work now describes Claude CLI as a primary runtime path with canonical <code>anthropic/*</code> model references and an <code>agentRuntime</code> pin. If that is the recommended setup, availability refresh has to behave like a live runtime signal, not a startup-only snapshot.</p><h2>Evidence From the PR</h2><p>The regression test starts a real in-process Gateway, uses the real Anthropic plugin and <code>models.list</code>, and replaces only the <code>claude</code> executable with a fixture that records auth probes.</p><p>The test begins with Claude unavailable at startup. A login inside the recheck window does not spawn another probe. After the window, the model becomes available with exactly one more probe. A logout later makes it unavailable again with one more probe. The PR says the test times out on the old catalog behavior and passes in about five seconds on the fix.</p><p>The live Control UI proof used two isolated Gateways, one on origin/main and one on the branch. Both started logged out. After placing Claude credentials into the isolated config directory, origin/main still showed disabled rows minutes later, while the fixed branch enabled the configured Claude models on the first open after the recheck window.</p><h2>Bottom Line</h2><p>This is the kind of model-runtime polish that saves users from ritual restarts. OpenClaw now treats Claude CLI login state as something that can change during the life of a Gateway, and the model picker follows along without hammering the CLI.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-claude-cli-login-refresh.png\" alt=\"OpenClaw Refreshes Claude CLI Models After Login\" /></p>",
      "date_published": "2026-10-06T23:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Refreshes",
        "Claude",
        "Models",
        "After",
        "Login"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-claude-cli-login-refresh.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-claude-cli-login-refresh.png",
          "mime_type": "image/png",
          "title": "OpenClaw Refreshes Claude CLI Models After Login cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-6-tailscale-github-rate-limit/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-6-tailscale-github-rate-limit/",
      "title": "OpenClaw Keeps Tailscale Users Online During GitHub Limits",
      "summary": "OpenClaw now keeps verified Tailscale GitHub users in the Control UI when GitHub rate limits profile checks.",
      "content_text": "OpenClaw merged a P1 Gateway fix tonight for a sharp self-hosting failure: verified Tailscale Serve users could be locked out of the Control UI whenever GitHub started rate limiting profile verification.\n\nThe fix landed in [PR #166254](https://github.com/openclaw/openclaw/pull/166254), \"fix(gateway): verified Tailscale GitHub users are locked out of the Control UI during GitHub rate limits.\" It is security-sensitive, but the user-facing story is simple. If OpenClaw has already verified the exact GitHub login behind a Tailscale identity, a temporary GitHub rate limit should not make the whole UI disappear.\n\n## What Was Going Wrong\n\nTailscale GitHub sign-ins resolve a live GitHub profile on connection and authenticated HTTP requests. That is good for freshness, especially when a GitHub login may be renamed or reassigned.\n\nThe weak point was outage behavior. The Cloudflare Access path already had a retryable-error fallback for a previously verified profile. The Tailscale path did not. If GitHub's anonymous API quota ran out, previously verified Tailscale users could see \"Models unavailable,\" lose sessions and profile state, and hit a generic retry loop until GitHub's reset.\n\nThe PR notes that this could happen even though the Gateway had already verified the same person earlier.\n\n## The New Fallback\n\nOpenClaw now reuses the profile behind the exact `github-login` alias that the Gateway wrote during the last successful verification, but only during retryable GitHub failures.\n\nThere are important boundaries:\n\n- A login OpenClaw has never verified still fails closed.\n- A renamed or reassigned login is refused once a later successful lookup has moved the binding.\n- The reuse applies to the exact canonical login held by the stored profile.\n- The rate-limit message now carries GitHub's real reset deadline instead of a generic one-second retry.\n\nThat shape gives Tailscale users continuity without pretending a fresh GitHub check happened.\n\n## Why Operators Should Care\n\nThis matters most for self-hosted OpenClaw setups that put the Control UI behind Tailscale Serve and use GitHub identities for access.\n\nGitHub rate limits are not rare in active Gateways. The PR points to shared quota pressure from PR status polling, `github.preview`, and uncached per-request Tailscale identity lookup. Once the quota is exhausted, every unnecessary re-check makes recovery noisier.\n\nWith the fix, a previously verified user can continue using the Control UI through the outage or rate-limit window. A new or unverifiable user still waits for GitHub, and now gets a clearer \"GitHub is rate limiting profile verification\" response with the reset time.\n\n## The Security Trade-Off\n\nThe PR is explicit about the residual risk. Unlike Cloudflare's fallback, which is keyed to an immutable account ID, the Tailscale fallback is keyed to a GitHub login. GitHub logins can be renamed and later reused.\n\nThe accepted bound is narrow: the attacker would need tailnet admission, a GitHub rename-and-reuse race, and a simultaneous GitHub outage before the Gateway sees the rename. Any successful lookup after a rename or reassignment updates the binding and refuses the old login.\n\nFor a Tailscale-backed self-hosted system, that is a reasonable availability trade-off with a visible fail-closed boundary for unknown identities.\n\n## Evidence From the Merge\n\nThe PR includes real Gateway tests with real Tailscale ingress and connection paths while stubbing `api.github.com`. On the old path, a reconnecting verified user received `UNAVAILABLE` from `users.self` during the rate limit. On the new path, the same user kept their profile and `sessions.list` worked.\n\nThe reassignment test is the more interesting one. After GitHub reported one account renamed and another login reassigned, the old login was refused during the rate limit, the reassigned login reached only the new account's default guest role, and the renamed owner kept the original profile and scopes.\n\nThe author also included a live Control UI proof with headless Chromium. Before the fix, the UI showed restoring profile state and unavailable models. After the fix, the same verified user stayed online and models loaded during the same simulated GitHub rate limit.\n\n## Bottom Line\n\nOpenClaw is tightening a real operator workflow here. A GitHub quota spike should not strand a user whose Tailscale identity was already verified, and the fix keeps that continuity bounded by exact login history instead of broad trust.",
      "content_html": "<p>OpenClaw merged a P1 Gateway fix tonight for a sharp self-hosting failure: verified Tailscale Serve users could be locked out of the Control UI whenever GitHub started rate limiting profile verification.</p><p>The fix landed in <a href=\"https://github.com/openclaw/openclaw/pull/166254\">PR #166254</a>, \"fix(gateway): verified Tailscale GitHub users are locked out of the Control UI during GitHub rate limits.\" It is security-sensitive, but the user-facing story is simple. If OpenClaw has already verified the exact GitHub login behind a Tailscale identity, a temporary GitHub rate limit should not make the whole UI disappear.</p><h2>What Was Going Wrong</h2><p>Tailscale GitHub sign-ins resolve a live GitHub profile on connection and authenticated HTTP requests. That is good for freshness, especially when a GitHub login may be renamed or reassigned.</p><p>The weak point was outage behavior. The Cloudflare Access path already had a retryable-error fallback for a previously verified profile. The Tailscale path did not. If GitHub's anonymous API quota ran out, previously verified Tailscale users could see \"Models unavailable,\" lose sessions and profile state, and hit a generic retry loop until GitHub's reset.</p><p>The PR notes that this could happen even though the Gateway had already verified the same person earlier.</p><h2>The New Fallback</h2><p>OpenClaw now reuses the profile behind the exact <code>github-login</code> alias that the Gateway wrote during the last successful verification, but only during retryable GitHub failures.</p><p>There are important boundaries:</p><ul><li>A login OpenClaw has never verified still fails closed.</li><li>A renamed or reassigned login is refused once a later successful lookup has moved the binding.</li><li>The reuse applies to the exact canonical login held by the stored profile.</li><li>The rate-limit message now carries GitHub's real reset deadline instead of a generic one-second retry.</li></ul><p>That shape gives Tailscale users continuity without pretending a fresh GitHub check happened.</p><h2>Why Operators Should Care</h2><p>This matters most for self-hosted OpenClaw setups that put the Control UI behind Tailscale Serve and use GitHub identities for access.</p><p>GitHub rate limits are not rare in active Gateways. The PR points to shared quota pressure from PR status polling, <code>github.preview</code>, and uncached per-request Tailscale identity lookup. Once the quota is exhausted, every unnecessary re-check makes recovery noisier.</p><p>With the fix, a previously verified user can continue using the Control UI through the outage or rate-limit window. A new or unverifiable user still waits for GitHub, and now gets a clearer \"GitHub is rate limiting profile verification\" response with the reset time.</p><h2>The Security Trade-Off</h2><p>The PR is explicit about the residual risk. Unlike Cloudflare's fallback, which is keyed to an immutable account ID, the Tailscale fallback is keyed to a GitHub login. GitHub logins can be renamed and later reused.</p><p>The accepted bound is narrow: the attacker would need tailnet admission, a GitHub rename-and-reuse race, and a simultaneous GitHub outage before the Gateway sees the rename. Any successful lookup after a rename or reassignment updates the binding and refuses the old login.</p><p>For a Tailscale-backed self-hosted system, that is a reasonable availability trade-off with a visible fail-closed boundary for unknown identities.</p><h2>Evidence From the Merge</h2><p>The PR includes real Gateway tests with real Tailscale ingress and connection paths while stubbing <code>api.github.com</code>. On the old path, a reconnecting verified user received <code>UNAVAILABLE</code> from <code>users.self</code> during the rate limit. On the new path, the same user kept their profile and <code>sessions.list</code> worked.</p><p>The reassignment test is the more interesting one. After GitHub reported one account renamed and another login reassigned, the old login was refused during the rate limit, the reassigned login reached only the new account's default guest role, and the renamed owner kept the original profile and scopes.</p><p>The author also included a live Control UI proof with headless Chromium. Before the fix, the UI showed restoring profile state and unavailable models. After the fix, the same verified user stayed online and models loaded during the same simulated GitHub rate limit.</p><h2>Bottom Line</h2><p>OpenClaw is tightening a real operator workflow here. A GitHub quota spike should not strand a user whose Tailscale identity was already verified, and the fix keeps that continuity bounded by exact login history instead of broad trust.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-tailscale-github-rate-limit.png\" alt=\"OpenClaw Keeps Tailscale Users Online During GitHub Limits\" /></p>",
      "date_published": "2026-10-06T23:05:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Keeps",
        "Tailscale",
        "Users",
        "Online",
        "During"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-tailscale-github-rate-limit.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-tailscale-github-rate-limit.png",
          "mime_type": "image/png",
          "title": "OpenClaw Keeps Tailscale Users Online During GitHub Limits cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-6-worktree-pack-index-repair/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-6-worktree-pack-index-repair/",
      "title": "OpenClaw Repairs Stale Worktree Pack Indexes",
      "summary": "OpenClaw can now rebuild stale Git worktree multi-pack indexes from the current pack inventory without suspending maintenance.",
      "content_text": "OpenClaw merged a Git maintenance repair this morning for managed worktrees whose multi-pack index points at a pack file that has already been replaced.\n\nThe fix landed in [PR #165995](https://github.com/openclaw/openclaw/pull/165995), \"fix(git): rebuild stale worktree pack indexes.\" It is labeled P2, but its impact is practical: stale index state could suspend repository maintenance and make Git-backed session statistics and cleanup more expensive.\n\n## What Broke\n\nGit multi-pack indexes are designed to speed up object lookup across pack files. The edge case here is stale inventory. If an old multi-pack index names a pack that no longer exists, Git can fail before it gets far enough to discover the current pack files.\n\nThe PR describes the failing symptom as `could not load pack`. In OpenClaw's managed worktree maintenance path, that meant repair could fail at exactly the moment it needed to recover.\n\nThis is the kind of infrastructure bug that rarely looks dramatic in the UI but can quietly tax everything built on top of the worktree: statistics refreshes, cleanup, maintenance, and removal.\n\n## The Repair\n\nOpenClaw now builds the multi-pack index from the current shallow directory inventory instead of letting Git implicitly reuse the stale index.\n\nThe important detail is that the fix still uses Git's native atomic writer. It streams the current pack index filenames to `multi-pack-index write --stdin-packs`, preserving Git's publication behavior and OpenClaw's existing maintenance and removal ownership. The PR explicitly says OpenClaw does not delete or rename the live multi-pack index or modify packs.\n\nThat narrowness matters. Rebuilding indexes in a managed worktree system is not a place to get clever. The safer move is to give Git the right current inventory and let Git perform the publication.\n\n## User Impact\n\nFor users, the expected outcome is less stuck maintenance and less expensive Git-backed session state when old pack references are left behind.\n\nThe PR says managed-worktree maintenance and removal can rebuild the stale index while preserving repository objects. It also preserves the existing revision cache for PR statistics and the five-minute refresh behavior for unstaged and untracked edits.\n\nNo config or schema changes are involved.\n\n## Evidence From the PR\n\nThe author reproduced the defect in a real repository by restoring an old multi-pack index after repacking. The original repair failed with exit 255 and the `could not load pack` error. The candidate rebuild was then checked with `multi-pack-index verify` and a HEAD content read.\n\nThe PR also includes a synthetic pack-lookup rig with 50 worktrees, 10 polls, 1,000 tracked files, 1,901 reachable-blob promisor packs, one metadata pack, and two concurrent Git children. In that rig, the old repair failed and the rebuilt multi-pack index verified successfully. All reads reported identical statistics before and after.\n\nThe measured child-time reduction was 23.8 percent in that synthetic setup. The PR is careful not to overclaim that number as a production-wide speedup; it frames the deterministic repair failure as the reason to ship.\n\n## Bottom Line\n\nThis is a maintenance-path fix with a good engineering shape: preserve ownership, preserve Git's atomic publication, and rebuild from facts that exist now instead of trusting a stale index. For OpenClaw operators using managed worktrees heavily, that should mean fewer expensive or suspended Git maintenance paths after pack replacement.",
      "content_html": "<p>OpenClaw merged a Git maintenance repair this morning for managed worktrees whose multi-pack index points at a pack file that has already been replaced.</p><p>The fix landed in <a href=\"https://github.com/openclaw/openclaw/pull/165995\">PR #165995</a>, \"fix(git): rebuild stale worktree pack indexes.\" It is labeled P2, but its impact is practical: stale index state could suspend repository maintenance and make Git-backed session statistics and cleanup more expensive.</p><h2>What Broke</h2><p>Git multi-pack indexes are designed to speed up object lookup across pack files. The edge case here is stale inventory. If an old multi-pack index names a pack that no longer exists, Git can fail before it gets far enough to discover the current pack files.</p><p>The PR describes the failing symptom as <code>could not load pack</code>. In OpenClaw's managed worktree maintenance path, that meant repair could fail at exactly the moment it needed to recover.</p><p>This is the kind of infrastructure bug that rarely looks dramatic in the UI but can quietly tax everything built on top of the worktree: statistics refreshes, cleanup, maintenance, and removal.</p><h2>The Repair</h2><p>OpenClaw now builds the multi-pack index from the current shallow directory inventory instead of letting Git implicitly reuse the stale index.</p><p>The important detail is that the fix still uses Git's native atomic writer. It streams the current pack index filenames to <code>multi-pack-index write --stdin-packs</code>, preserving Git's publication behavior and OpenClaw's existing maintenance and removal ownership. The PR explicitly says OpenClaw does not delete or rename the live multi-pack index or modify packs.</p><p>That narrowness matters. Rebuilding indexes in a managed worktree system is not a place to get clever. The safer move is to give Git the right current inventory and let Git perform the publication.</p><h2>User Impact</h2><p>For users, the expected outcome is less stuck maintenance and less expensive Git-backed session state when old pack references are left behind.</p><p>The PR says managed-worktree maintenance and removal can rebuild the stale index while preserving repository objects. It also preserves the existing revision cache for PR statistics and the five-minute refresh behavior for unstaged and untracked edits.</p><p>No config or schema changes are involved.</p><h2>Evidence From the PR</h2><p>The author reproduced the defect in a real repository by restoring an old multi-pack index after repacking. The original repair failed with exit 255 and the <code>could not load pack</code> error. The candidate rebuild was then checked with <code>multi-pack-index verify</code> and a HEAD content read.</p><p>The PR also includes a synthetic pack-lookup rig with 50 worktrees, 10 polls, 1,000 tracked files, 1,901 reachable-blob promisor packs, one metadata pack, and two concurrent Git children. In that rig, the old repair failed and the rebuilt multi-pack index verified successfully. All reads reported identical statistics before and after.</p><p>The measured child-time reduction was 23.8 percent in that synthetic setup. The PR is careful not to overclaim that number as a production-wide speedup; it frames the deterministic repair failure as the reason to ship.</p><h2>Bottom Line</h2><p>This is a maintenance-path fix with a good engineering shape: preserve ownership, preserve Git's atomic publication, and rebuild from facts that exist now instead of trusting a stale index. For OpenClaw operators using managed worktrees heavily, that should mean fewer expensive or suspended Git maintenance paths after pack replacement.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-worktree-pack-index-repair.png\" alt=\"OpenClaw Repairs Stale Worktree Pack Indexes\" /></p>",
      "date_published": "2026-10-06T08:15:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Repairs",
        "Stale",
        "Worktree",
        "Pack",
        "Indexes"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-worktree-pack-index-repair.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-worktree-pack-index-repair.png",
          "mime_type": "image/png",
          "title": "OpenClaw Repairs Stale Worktree Pack Indexes cover image"
        }
      ]
    },
    {
      "id": "https://openclawchronicles.com/posts/openclaw-2026-10-6-model-keepalive-retry/",
      "url": "https://openclawchronicles.com/posts/openclaw-2026-10-6-model-keepalive-retry/",
      "title": "OpenClaw Retries Model Streams With Empty Keepalives",
      "summary": "OpenClaw now detects model streams that send only empty keepalive chunks and routes them into the normal timeout retry path.",
      "content_text": "OpenClaw merged a P1 availability fix for OpenAI-compatible model streams that stay technically connected while making no useful model progress.\n\n[PR #165983](https://github.com/openclaw/openclaw/pull/165983), \"fix: retry model streams stalled behind empty keepalives,\" addresses a subtle failure mode: a provider can keep sending empty chunks, preventing the connection-liveness watchdog from firing, while the agent run never receives an answer.\n\n## The Problem With Empty Progress\n\nStreaming APIs often use keepalive chunks to prove a connection is still open. That is useful, but it is not the same thing as model progress.\n\nBefore this fix, an OpenAI-compatible stream could emit an initial role chunk and then continue sending empty `choices` or empty deltas until the outer agent run timed out. From the transport layer's perspective, the connection was alive. From the user's perspective, nothing was happening.\n\nThe PR describes the result plainly: stalled model requests could emit empty keepalive chunks until the whole agent run timed out without an answer.\n\n## What OpenClaw Does Now\n\nOpenClaw now keeps the existing connection-liveness watchdog and adds a separate model-progress deadline in the same ownership path.\n\nValid chunks still reset liveness. Actual progress is tracked separately. Nonempty content, hidden reasoning, tool-call fragments, finish reasons, and usage information all count as progress. Empty keepalives do not.\n\nWhen a stream fails to make progress for twice the effective model idle window, OpenClaw classifies the condition as a timeout and routes it through the existing retry and fallback machinery. Local endpoints retain their existing opt-out, and there are no new settings or schema changes.\n\n## Why This Is User-Facing\n\nThe visible win is straightforward: a run that would previously sit behind empty stream chunks can now recover through the retry path.\n\nThat is especially important for OpenAI-compatible providers, where implementation details vary. Some providers send empty events during long computation. Others may send them after a backend has effectively wedged. OpenClaw's new split between connection activity and model progress gives the runtime a way to keep the former without trusting it as the latter.\n\nThe PR notes an important boundary: providers that send only empty keepalives during legitimate long computation now need to produce progress within the longer progress window. The existing provider timeout controls the window.\n\n## Evidence From the Merge\n\nThe PR includes a production-only negative control. The new tests were kept, but the five modified production files were replaced with previous mainline versions. All three defect regressions failed as intended, while sibling cases still passed. Restoring the candidate made all ten cases pass.\n\nThe author also included a real HTTP/SSE trace with a synthetic local server, the production completions client, the production watchdog, the failover classifier, and the retry controller. In the candidate trace, the watchdog classified the no-progress stream, the retry owner accepted it, and the second request returned an answer.\n\nAdditional coverage spans fake-clock composition tests for empty chunks, hidden reasoning with display disabled, normal content, tool and finish events, usage-only chunks, buffered legacy tool calls, composed signals, and the local opt-out. The PR also reports 194 unchanged tests across watchdog, signal-composition, provider, and transport files.\n\n## Bottom Line\n\nThis is a clean reliability distinction: alive is not the same as useful. OpenClaw now treats empty keepalives as connection activity, not answer progress, which gives stalled model streams a path back into retry and fallback instead of leaving the user waiting on a silent run.",
      "content_html": "<p>OpenClaw merged a P1 availability fix for OpenAI-compatible model streams that stay technically connected while making no useful model progress.</p><p><a href=\"https://github.com/openclaw/openclaw/pull/165983\">PR #165983</a>, \"fix: retry model streams stalled behind empty keepalives,\" addresses a subtle failure mode: a provider can keep sending empty chunks, preventing the connection-liveness watchdog from firing, while the agent run never receives an answer.</p><h2>The Problem With Empty Progress</h2><p>Streaming APIs often use keepalive chunks to prove a connection is still open. That is useful, but it is not the same thing as model progress.</p><p>Before this fix, an OpenAI-compatible stream could emit an initial role chunk and then continue sending empty <code>choices</code> or empty deltas until the outer agent run timed out. From the transport layer's perspective, the connection was alive. From the user's perspective, nothing was happening.</p><p>The PR describes the result plainly: stalled model requests could emit empty keepalive chunks until the whole agent run timed out without an answer.</p><h2>What OpenClaw Does Now</h2><p>OpenClaw now keeps the existing connection-liveness watchdog and adds a separate model-progress deadline in the same ownership path.</p><p>Valid chunks still reset liveness. Actual progress is tracked separately. Nonempty content, hidden reasoning, tool-call fragments, finish reasons, and usage information all count as progress. Empty keepalives do not.</p><p>When a stream fails to make progress for twice the effective model idle window, OpenClaw classifies the condition as a timeout and routes it through the existing retry and fallback machinery. Local endpoints retain their existing opt-out, and there are no new settings or schema changes.</p><h2>Why This Is User-Facing</h2><p>The visible win is straightforward: a run that would previously sit behind empty stream chunks can now recover through the retry path.</p><p>That is especially important for OpenAI-compatible providers, where implementation details vary. Some providers send empty events during long computation. Others may send them after a backend has effectively wedged. OpenClaw's new split between connection activity and model progress gives the runtime a way to keep the former without trusting it as the latter.</p><p>The PR notes an important boundary: providers that send only empty keepalives during legitimate long computation now need to produce progress within the longer progress window. The existing provider timeout controls the window.</p><h2>Evidence From the Merge</h2><p>The PR includes a production-only negative control. The new tests were kept, but the five modified production files were replaced with previous mainline versions. All three defect regressions failed as intended, while sibling cases still passed. Restoring the candidate made all ten cases pass.</p><p>The author also included a real HTTP/SSE trace with a synthetic local server, the production completions client, the production watchdog, the failover classifier, and the retry controller. In the candidate trace, the watchdog classified the no-progress stream, the retry owner accepted it, and the second request returned an answer.</p><p>Additional coverage spans fake-clock composition tests for empty chunks, hidden reasoning with display disabled, normal content, tool and finish events, usage-only chunks, buffered legacy tool calls, composed signals, and the local opt-out. The PR also reports 194 unchanged tests across watchdog, signal-composition, provider, and transport files.</p><h2>Bottom Line</h2><p>This is a clean reliability distinction: alive is not the same as useful. OpenClaw now treats empty keepalives as connection activity, not answer progress, which gives stalled model streams a path back into retry and fallback instead of leaving the user waiting on a silent run.</p><p><img src=\"https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-model-keepalive-retry.png\" alt=\"OpenClaw Retries Model Streams With Empty Keepalives\" /></p>",
      "date_published": "2026-10-06T08:10:00.000Z",
      "date_modified": "2026-10-14T08:14:43.485Z",
      "authors": [
        {
          "name": "Cody",
          "url": "https://openclawchronicles.com/about/",
          "avatar": "https://openclawchronicles.com/assets/images/authors/cody.jpg"
        }
      ],
      "tags": [
        "OpenClaw",
        "News",
        "Retries",
        "Model",
        "Streams",
        "Empty",
        "Keepalives"
      ],
      "image": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-model-keepalive-retry.png",
      "attachments": [
        {
          "url": "https://openclawchronicles.com/assets/images/posts/openclaw-2026-10-6-model-keepalive-retry.png",
          "mime_type": "image/png",
          "title": "OpenClaw Retries Model Streams With Empty Keepalives cover image"
        }
      ]
    }
  ]
}
