OpenClaw v2026.6.5-beta.2: MCP results no longer poison sessions, operator field notes

The source record for v2026.6.5-beta.2 names a provider, model, voice, or context change. This is an operator's reading of that record, with the exact source facts kept next to a bounded experiment rather than a generic feature list.

Pre-release status matters here: stage the tag beside the stable package and do not infer support for every account or platform. A release tag describes the project change. Your Gateway's provider, channel, platform, and account support should be recorded separately.

A small field test

Safer channel output

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

Check it with. The operator question around Safer channel output is where the reply lands after a restart. Record the conversation identifiers before testing and compare them after the channel is healthy again.

MCP results no longer poison sessions

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

Check it with. Check MCP results no longer poison sessions 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.

Anthropic extended thinking recovers after restarts

What changed. 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.

Check it with. For Anthropic extended thinking recovers after restarts, delivery is only proven when identity survives an interruption. Keep a disposable channel, send one request, and check for a single reply in the intended thread.

Parallel web search is bundled

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

Check it with. A catalog label for Parallel web search is bundled is only the first signal. Confirm the account, endpoint, context limit, and tool policy that the running Gateway actually selected.

The details that affect a runbook

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.

For Better Matrix and Vertex reliability, delivery is only proven when identity survives an interruption. Keep a disposable channel, send one request, and check for a single reply in the intended thread.

Search/providers

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

A catalog label for Search/providers is only the first signal. Confirm the account, endpoint, context limit, and tool policy that the running Gateway actually selected.

Matrix/channels

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

Use Matrix/channels 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.

Operator checklist before rollout

1. Before touching persistent state, capture the version, host, provider, and workspace for Safer channel output.
2. Read the linked release record and verify a backup whenever the change can alter configuration or session data.
3. Interrupt one delivery after admission, reconnect the transport, and compare the final reply with the original conversation.
4. Force a short transport interruption and verify that retry or recovery preserves the original thread and sender.
5. Close the test with the evidence attached, and do not widen the rollout while a behavior remains unexplained.

openclaw --version
openclaw gateway status
openclaw models

Keep the source trail attached

Close the loop with an explicit next action: keep the candidate isolated, proceed with the staged upgrade, or stop and investigate. Do not treat a green install as proof that the runtime path is correct.

Regression cases from the source

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. Treat the fix as a boundary condition and retain the source wording beside your local result.

Channel content boundaries. QQBot now strips reasoning/thinking tags before sending, preserving final answers while hiding internal model narration from users. Keep the reproduction and the post-fix result together in the acceptance record.

This page is a practical reading of the v2026.6.5-beta.2 source record; the release index and model provider reference remain the references to use for migrations and support boundaries.