Schema Quarantine · GLM-5.2 · Kimi K2.7 · Telegram Rich Text
June 16, 2026 · Stable Release · 143 PRs · 66 Issues
OpenAI/Anthropic tool schemas are now quarantined when unreadable or post-hook. Sustained security pass across quarantine unreadable OpenAI/Anthropic tool schemas series — 8+ Anthropic PRs, 10+ OpenAI PRs.
GLM-5.2 (Zhipu AI), Kimi K2.7 Code (Moonshot), Claude Haiku 4.5 catalog entries. Plus OpenRouter/Vertex provider-prefix normalization, OAuth image routing through Codex, LM Studio binary-thinking support.
Structured rich text: tables, lists, expandable blockquotes, preserved intentional line breaks, prompt-preserving CLI backend delivery, safer rich-media boundaries. WhatsApp now honors configured ACP bindings.
Anthropic thinking-signature recovery on replay + OpenAI reasoning-signature recovery. Storeless OpenAI Responses replay compatibility preserved. Eliminates a major class of replay bugs.
Full provider support added. Chinese frontier-class model. #92796
Coding-specialized variant. Long-context reasoning. #92554
Anthropic's fast/cheap tier — now in static catalog. #90116
Provider-qualified model IDs normalized across OpenRouter and Google Vertex paths. #92627, #91218
Tables, lists, expandable blockquotes, line breaks, prompt-preserving CLI delivery, native draft migration retired, safer rich-media boundaries.
Honors configured ACP bindings. Login fix: no longer reports success before auth durably persisted.
Outbound message_sent hooks. Warns on name-keyed channel map. Same-channel final reply mirroring.
60s timeout + 4,096-token reasoning-model budget for auto-thread titles. Default-enabled (no plugins.entries.discord.enabled required).
A 15+ PR series hardening the agent runtime against malformed tool schemas from upstream providers.
quarantine unreadable tool schemas — #92921guard OpenAI transport tool descriptors — #89016guard OpenAI tool schema conversion — #89378harden OpenAI strict schema inspection — #89543guard OpenAI strict tool diagnostics — #90200skip unreadable OpenAI completion tool schemas — #90283skip unreadable OpenAI responses tool schemas — #90286, #90397guard post-hook tool payloads — #92928guard responses tool payload names — #89703quarantine invalid direct tool schemas — #92896quarantine invalid projected tools — #89418guard Anthropic tool descriptors — #89221guard Anthropic tool schema conversion — #90228skip unreadable Anthropic tool schemas — #89622, #90278guard Anthropic provider tool descriptors — #89229quarantine unreadable Anthropic payload tools — #92908Embedded runner: freshly streamed thinking signatures intermittently invalid on replay. Recovery wrapper now fires for genericized error text. History thinking-block stripping when thinking is disabled.
Recover invalid reasoning signatures in OpenAI Responses API. Storeless replay compatibility preserved. gpt-5.5 tool reasoning effort omitted where unsupported.
Replays of long agent sessions no longer fail when thinking-block signatures degrade across provider round-trips. Reliability for long-horizon agents materially improved.
/btw CLI CommandNow supported in CLI-backed sessions. CLI usage-error exits classified as usage failures, not successful runs.
Trusted subagent thinking override clamped through provider/model fallback. Yielded media completions preserved.
HTTP session/model override requires admin privileges. session_kill endpoint protected. Subagent pauses on abort signal.
Unknown OpenAI agent selectors explicitly rejected. wrapToolWithBeforeToolCallHook made idempotent to prevent double execution.
Oversized OpenAI embedding batches split before 431s. Nested embedding fetch failures now surfaced.
memory_search stays available in transient QMD mode. Skipped short-term recall hits explained.
WAL avoided on NFS state volumes. Full reindex preserves rollback/cache recovery.
Stuck-session recovery no longer resets warning backoff. Reset archive fallback for missing async transcripts.
Survives streaming — previously lost when long responses came in.
Can collapse and start collapsed. Cleaner default state.
Remains interactive above desktop workbench.
Stale foreground gateways force reconnect. TUI parent stdin stays paused after exit.
Previously, Parallel Free, DuckDuckGo, Ollama, and Codex Hosted Search were automatically chosen when no API-backed provider was configured. This silently degraded result quality for many users. They are now explicit opt-ins — defaults now require a configured API-backed provider.
/models and models.list catalog loading now bounded — fixes perf regression from v2026.6.1npm install -g @openclaw/cli@latest
openclaw update stable
openclaw doctor
web_search.provider in config