OpenClaw’s June beta fixes 5 release risks
5 release updates in OpenClaw 2026.6.5-beta.2 tighten AI output, MCP handling, provider recovery, and release validation.
OpenClaw 2026.6.5-beta.2 tightens AI output handling, provider recovery, and release safety.
This release note covers 5 clusters of changes in OpenClaw 2026.6.5-beta.2, a beta published on 07 Jun with a full release CI report and npm tarball integrity check.
| Item | What changed | Why it matters |
|---|---|---|
| Output boundaries | QQBot strips reasoning/thinking scaffolding before delivery | Users see final answers, not internal model narration |
| MCP/tool handling | Non-text/image blocks are coerced at materialize time | Prevents Anthropic 400s and poisoned session history |
| Provider recovery | Anthropic, Vertex, and cooldown flows recover more cleanly | Fewer broken sessions after cache expiry or restarts |
| Platform support | Parallel search, Matrix voice/thread flow, and auth storage updates | Broader feature coverage across channels and installs |
| Release safety | Monthly patch numbering, migration fixes, and CI tightening | Cleaner validation and safer publish paths |
1. Output boundary fixes
Get the latest AI news in your inbox
Weekly picks of model releases, tools, and deep dives — no spam, unsubscribe anytime.
No spam. Unsubscribe at any time.
The most visible user-facing change is in QQBot: OpenClaw now strips model reasoning and thinking scaffolding before sending replies. That keeps raw chain-of-thought style text out of channel messages while preserving the final answer.
This matters because the release also closes related boundary bugs in Feishu merged content and ClickClack reply handling. The common theme is simple: what the model produces internally should not leak into what users read.
- QQBot removes thinking tags before native delivery.
- Feishu streaming cards keep merged content intact.
- ClickClack reply tools respect
toolsAllow.
2. MCP tool-result cleanup
OpenClaw now coerces richer MCP tool-result blocks at the materialize boundary, including resource links, audio, malformed image blocks, and future non-text/image content. That reduces provider errors when a tool returns something more complex than plain text.
The practical payoff is fewer Anthropic 400 responses and less corrupted session history after a tool call. If you build on MCP, this is the sort of fix that prevents one odd tool payload from breaking the rest of the conversation.
resource_link -> text-safe materialization
resource -> text-safe materialization
audio -> handled before provider conversion
malformed image -> normalized or rejected cleanly3. Provider recovery and model resolution
Several fixes focus on keeping provider sessions alive when state changes underneath them. Anthropic extended-thinking sessions now recover after prompt-cache expiry or Gateway restarts, and Google Vertex ADC users get static catalog rows plus runtime model resolution again.
There are also guardrails around cooldown and auth resolution. Single-provider cooldown recovery is more reliable, unknown model auth fails closed, and model listing avoids resolving auth too early.
- Anthropic stream start waits for
message_start. - Vertex ADC catalog rows and runtime resolution work again.
- Cooldown recovery re-probes the primary provider.
- Unknown model auth is treated as unknown, not assumed valid.
4. New channel and integration coverage
This beta expands the number of places OpenClaw can operate cleanly. Parallel is now bundled as a web search provider with API-key discovery and onboarding support, while Matrix gets voice-note preflight and thread-aware read and reply behavior.
Auth and plugin install state also got sturdier. Auth profiles now live in SQLite, official npm plugin install records keep trusted pins, and ClawHub skills can install from GitHub repositories through the resolved install API.
- Parallel search has bundled provider support and docs.
- Matrix handles voice notes before mention gating.
- Matrix thread reads and replies survive pagination.
- ClawHub installs pinned GitHub-backed skills.
5. Release and CI safety checks
The release process itself changed in a meaningful way: OpenClaw switched to YYYY.M.PATCH monthly patch numbering, with pre-transition tags kept compatible and June 2026 pinned at 2026.6.5 after the beta. That makes versioning easier to read and compare across the train.
CI and validation also got tighter. The release note points to a full release CI report, npm preflight, and full release validation run, plus test-state isolation work that reduces noisy failures. If you maintain a release pipeline, this is the part that lowers the odds of shipping hidden drift.
release-evidence.md
npm package: [email protected]
tarball: openclaw-2026.6.5-beta.2.tgz
integrity: sha512-6Bmx2rlReO1MOEi9ehuhsVO59keQ1xQQ8/PvwbNq0CatJkNwyCX/MBlhWrscByYAmBl6JCOSr+cLejoDk1sDBA==How to decide
If you care most about user-visible chat quality, start with the output-boundary fixes. If your pain is provider instability or tool-call errors, the MCP and recovery items will matter more. If you run releases, the versioning and CI section is the one to read first.
For teams using OpenClaw across multiple channels, the best signal is breadth: this beta is less about one headline feature than about removing the failure modes that make AI systems feel brittle.
// Related Articles