OpenClaw v2026.6.5-beta.3: Anthropic extended thinking recovers after restarts, recovery notes

v2026.6.5-beta.3 is easiest to understand through failure. The official record describes a provider, model, voice, or context change; the sections below spell out what to save, what to interrupt, and what must remain available afterward.

This candidate is still pre-release. Treat provider, channel, and migration behavior as unconfirmed until your own bounded test passes. Treat the version as one input to the decision. Local credentials, enabled plugins, channels, and host support still need their own evidence.

Failure and recovery path

01 / Safer channel output

Source signal. QQBot strips model reasoning and thinking scaffolding before native delivery, so users see the final answer rather than raw internal markup.

Safer channel output lives at the handoff between the Gateway and a conversation. Send one harmless message, reconnect once, and compare the account, thread, sender, and final transcript.

02 / MCP results no longer poison sessions

Source signal. resource_link, audio, malformed images, and future non-text blocks are normalized before provider conversion, avoiding Anthropic 400s and broken follow-up history.

A catalog label for MCP results no longer poison sessions is only the first signal. Confirm the account, endpoint, context limit, and tool policy that the running Gateway actually selected.

03 / Anthropic extended thinking recovers after restarts

Source signal. prompt-cache expiry and early-signature failures wait for a real message start, so they reach the existing retry path rather than ending a turn.

The operator question around Anthropic extended thinking recovers after restarts is where the reply lands after a restart. Record the conversation identifiers before testing and compare them after the channel is healthy again.

04 / Parallel web search is bundled

Source signal. API-key discovery, guarded endpoint handling, cache-safe session IDs, onboarding, and documentation make it a first-class search provider.

The safe reading of Parallel web search is bundled is operational: prove the model can answer the intended task, then force a bounded failure and verify that no side effect is repeated.

The limits of the recovery promise

Better Matrix and Vertex reliability

Matrix voice notes and threaded conversations retain their context, while Google Vertex ADC models regain static catalog and runtime resolution.

Use Better Matrix and Vertex reliability as a routing test: capture the destination before the send, interrupt the transport, and verify that retry logic did not duplicate or redirect the response.

Search/providers

add the Parallel bundled web-search plugin, registration contracts, onboarding integration, and guarded api.parallel.ai/v1/search support.

Check Search/providers with a known prompt and a disposable account. Save the selected route, reasoning setting, latency, and refusal or fallback result so another operator can reproduce it.

Matrix/channels

add voice-message preflight and thread-aware read/reply behavior.

The operator question around Matrix/channels is where the reply lands after a restart. Record the conversation identifiers before testing and compare them after the channel is healthy again.

Run the smallest useful rehearsal

1. Write down the baseline for Safer channel output, including the installed version, host, provider, and workspace.
2. Confirm what can be restored, then take a verified backup before applying the candidate package.
3. Exercise the affected channel with synthetic text and note the thread, sender, reaction, or attachment id that comes back.
4. Restart or reconnect the Gateway, then check for one intended reply rather than a duplicate, missing, or misrouted message.
5. Keep the experiment bounded; an unexpected result belongs in the release record before any wider rollout.

openclaw --version
openclaw gateway status
openclaw models

Leave a usable maintenance note

Save the command, output, source link, and environment used for this check. Future operators should be able to tell whether a difference belongs to the release or to local provider, channel, or platform state.

Fix details to carry forward

Agents. sessions_send now honors an explicit sessionKey when stale label metadata is also present, and denied session-id sends no longer echo the resolved canonical session key. Fixes #64699; refs #74009 and #41199. This is the edge case to include in a regression check, especially when the release is pre-release.

Channel content boundaries. QQBot now strips reasoning/thinking tags before sending, preserving final answers while hiding internal model narration from users. Treat the fix as a boundary condition and retain the source wording beside your local result.

Keep the full v2026.6.5-beta.3 notes with your test result. Read the official release index for neighboring versions and the model provider reference for configuration limits.