OpenClaw 2026.4.29: Commitments, People Wiki, and NVIDIA Support

OpenClaw v2026.4.29 shipped today — a release dense enough that "highlights" section alone covers five distinct subsystems. Here is what matters and what you need to watch out for before upgrading.

Inferred Commitments — Agents That Follow Through

The headline feature is the new commitments system (#74189). When enabled, OpenClaw can now extract implicit follow-up promises from conversations — phrases like "I'll check in on that tomorrow" or "remind me Friday" — and turn them into scheduled heartbeat deliveries.

The config surface is minimal by design:

commitments:
  enabled: true
  maxPerDay: 5

Extraction happens in a hidden batched pass, scoped per agent and per channel. Heartbeat-interval due-time clamping prevents commitments from echoing back immediately. It is opt-in for now but signals where the project is heading: agents that don't just respond, but actually follow through.

Memory Grows Into a People-Aware Wiki

Memory gets a substantial upgrade this release. OpenClaw now maintains an agent-facing people wiki with:

• Canonical aliases and person cards
• Relationship graphs between contacts
• Privacy and provenance reports showing where each piece of information came from
• Evidence-kind drilldown (who said what, when, in which conversation)
• Per-conversation allowedChatIds / deniedChatIds filters for targeted recall

For teams running shared OpenClaw deployments, the provenance reports address a real concern: knowing why the agent thinks it knows something about a person. The partial recall on timeout improvement (#73219) is also worth noting — if the memory sub-agent times out, you now get a bounded summary of whatever was recovered rather than losing all context silently.

NVIDIA Provider Now Available

The NVIDIA provider lands in this release (#71204), joining the existing roster of OpenAI, Anthropic, Bedrock, and others. Setup includes API-key onboarding, a static model catalog, and literal model-ref picker support so you can reference NVIDIA-hosted models with their provider prefix (e.g., nvidia/llama-3.1-nemotron-70b-instruct).

Bedrock also gets Opus 4.7 thinking parity this release, and Codex/OpenAI-compatible replay and streaming behavior sees several safety fixes.

⚠️ Security Breaking Change: Tool Profiles Now Enforced

This is the one you need to read before upgrading.

Previously, adding tools.exec or tools.fs sections to your config would silently widen access even under restrictive profiles like messaging or minimal. That behavior is gone.

Configured tool sections (tools.exec, tools.fs) no longer implicitly widen restrictive profiles (messaging, minimal). Users who need those tools under a restricted profile must add explicit alsoAllow entries.

A startup warning will identify affected configs, so the migration path is clear — but deployments relying on the old implicit behavior will see tools disabled on first boot. Check your configs. Fixes #47487, thanks to @amknight.

Channel Fixes Across the Board

This release clears a long backlog of channel-specific bugs:

• Slack: Block Kit limits no longer cause silent failures; bot relay allowlisting now requires explicit channel owner presence (#59284)
• Telegram: Proxy, webhook, polling, and send resilience all improved
• Discord: Startup and rate-limit handling fixed; group turns no longer complete silently when the message tool is unavailable (#74868)
• WhatsApp: Delivery and liveness improvements
• Signal: Group allowlists now match inbound group IDs, not just sender IDs (#53308)
• Yuanbao: Tencent's Yuanbao channel is now officially documented in the channel listing

OpenGrep Security Scanning

A new OpenGrep rulepack (#69483) ships with PR and full-scan GitHub Actions workflows that validate first-party code changes and upload SARIF results to GitHub Code Scanning. For contributors, this means security-relevant code changes now get static analysis as part of CI.

Docker and Operations

Two Docker-specific improvements are worth calling out:

• OPENCLAW_SKIP_ONBOARDING (#55518) lets automated Docker installs skip interactive onboarding while still applying gateway defaults
• Gateway startup now exits with sysexits 78 for port conflicts, which maps cleanly to RestartPreventExitStatus=78 in systemd — so restart loops on occupied ports stop instead of spinning (#75115)

Internationalization Push

New locale support lands for Persian (fa), Dutch (nl), Vietnamese (vi), Italian (it), Arabic (ar), and Thai (th) — with matching docs glossaries for fa, nl, vi, and zh-TW. The Control UI language picker now stays aligned with the docs translation pipeline across all surfaces.

Upgrade Notes

The tools profile change (see above) is the only breaking change confirmed in this release. The commitments and people-wiki features are both opt-in and require explicit config to activate. Everything else is additive or a fix.

Full changelog on GitHub.