OpenClaw merged a Mattermost reliability fix this morning for threaded conversations that survive Gateway restarts or session resets. PR #150644, "fix(mattermost): restore thread context after restart or reset," closes issue #93204 and continues earlier work from PR #114523.
The user-facing problem was simple and frustrating: a Mattermost bot could reply without earlier thread context when local history was empty after a restart or reset.
What Changed
OpenClaw now performs bounded server-side history recovery through the existing Mattermost client before preparing the inbound turn. The recovery is scoped through the sender-visibility policy, current stored session ID, and lifecycle revision.
That binding matters. An old request cannot seed a reset session or overwrite retry state, and concurrent live messages are preserved. Waiting turns get separately filtered context instead of a shared blob of recovered history.
The recovery budget is intentionally limited:
- At most one server page
- Up to 200 posts
- A five-second deadline
- Fixed retry budget and cooldown
- No pairing-request creation
- No broader sender access
Flat DMs are unchanged, and no configuration or migration is required.
Why It Matters
Threaded collaboration depends on context. If a Mattermost channel or group thread has already established a fact, a bot that forgets that fact after a Gateway restart can produce confusing or wrong replies.
This fix narrows that failure mode without turning history recovery into a broad backfill system. The PR explicitly says this is server history recovery, not a stored-data migration. The added recovery owner keeps monitor-lifetime Map and Set markers and does not alter schemas, retention, session serialization, or normal reply dispatch.
That is the right shape for a channel integration fix: recover enough context to answer correctly, but avoid changing durable data models just to cover a transient local-history gap.
Validation
The focused test set is broad: 161 tests cover the post handler, HTTP client, SQLite session reader, sender authorization, history windows, reset races, coalescing, deadlines, and retry ownership.
The strongest evidence is the real-server acceptance test. The PR used disposable Mattermost Team Edition 10.11.0 with PostgreSQL, a built plugin running inside an isolated OpenClaw Gateway, and memory disabled. The provider request included a previously posted thread fact, and the server accepted and delivered the bot reply.
The same check passed immediately after an administrator session reset that kept the session ID but changed the lifecycle revision. Additional cold threaded-DM delivery was also verified.
The final diff is Mattermost-only after upstream Slack test fixes landed on main. Independent source review was clean through P2, and subsequent lint-only cleanup preserved the reviewed production behavior.
For teams using Mattermost as an OpenClaw channel, this should make resumed threads feel less brittle after operational interruptions.
Source: OpenClaw PR #150644.