- chore: migrate tooling and source installs to pnpm 12 — What Problem This Solves Migrates OpenClaw source development, CI, installers, and package tooling from pnpm 11.22.0 to stable pnpm 12.0.0. A pin-only upgrade would break lockfile consumers,...
- fix: Responses sessions compact before reaching context limit — Summary High Level TLDR Fixes two failures in the same long-session compaction pipeline: 1. OpenAI Responses usage could be discarded as a context boundary, making a session look over budget and...
- fix: Git update can report success while Web UI serves old build — Related: #92241 What Problem This Solves Fixes an issue where users running openclaw update --channel dev could see the update report success while the Web UI continued serving the previous build....
- feat(auto-reply): clean empty staged inbound media directories (#104358) — This is a replacement PR for #117184, which was incorrectly closed due to temporary branch pollution. What Problem This Solves When inbound media is staged in the host workspace and the turn is...
- feat(comfy): support custom headers and per-request seed randomization — What Problem This Solves The ComfyUI provider plugin's local mode only ever sent Content-Type on requests, so a local ComfyUI instance reachable behind HTTP auth (Basic auth, a bearer token, etc.)...
- fix(openshell): preserve nested host symlinks during mirror sync — Closes #131111 What Problem This Solves Fixes an issue where OpenShell mirror users lose existing host symlinks inside subdirectories after an ordinary sandbox command, even when that command changes...
- [Feature]: ADR: replay-safe session Goal protocol — Executive summary 1. Goal start currently writes the human turn and Goal state through separate paths. 2. A retry can therefore create partial, duplicate, or conflicting state. 3. The proposal adds...
- [Bug]: Aborting parked Code Mode cells exhausts shared execution slots — Bug type Behavior bug (incorrect output/state without crash) Beta release blocker No Summary Aborting an agent run after Code Mode has returned a parked cell leaves its snapshot occupying the...
- test(whatsapp): fix order-dependent pairing-prompt flake with event-driven inbound drain — What Problem This Solves Fixes an order-dependent flake where extensions/whatsapp/src/monitor-inbox.policy.test.ts > delivery coordinator blocks unauthorized senders outside allowFrom fails in full...
- feat(slack): deliver TTS voice notes via captioned media upload — What Problem This Solves Issue #130838 requests voice-message (tts.voice) delivery for the Slack channel, matching the capability already shipped on the Feishu and Telegram channels. Today the Slack...
- [codex] Repair Browser Session Credential Steward MVP — What Problem This Solves The Browser / Session / Credential Steward MVP needed reliable boundary enforcement across browser-node routing, session targeting, approval provenance, and credential...
- fix(config): channel schema ownership follows the declared preferOver replacement — Related: #123209 What Problem This Solves Resolves a problem where openclaw config validate rejects channel config keys that the plugin actually serving the channel accepts. When one plugin declares...
- fix(code-mode): resume work after safe reconciliation — Related: #128028 Additional instructions MUST: Keep Allow edits from maintainers enabled for this PR so maintainers can help update the branch when needed. Summary High Level TLDR Code Mode already...
- fix(cli): resolve alias targets from the write snapshot — Closes #127618 Additional instructions MUST: Keep Allow edits from maintainers enabled for this PR so maintainers can help update the branch when needed. What Problem This Solves openclaw models...
- [Bug]: image tool resolves sandbox upload file_ reference as workspace-relative path → ENOENT — Summary When a file (e.g. an uploaded image) is referenced in a sandbox agent session by its file_ handle, the image tool treats the bare file_ string as a workspace-relative path, prepends the agent...
- feat(ci): add admin-only Crabbox PR gate fallback — What Problem This Solves Provides a bounded fallback when hosted PR infrastructure cannot produce openclaw/ci-gate, without changing the default gate path or allowing broad admin bypass. Why This...
- perf(codex): narrow the direct message tool schema and teach deferred-tool batching — Closes #125314 What Problem This Solves When Codex admits message as a direct source-reply tool, OpenClaw exposes the full rich message-manager schema on the hot path. Most calls are text replies to...
- fix(agents): prevent early final responses in sequential subagent runs — Closes #129455 What Problem This Solves Sequential native subagent workflows could finalize after the first child wave even though the resumed requester had intentionally yielded to start the next...
- agents: apply_patch reports success while rewriting bytes outside the hunk (indentation, UTF-8 BOM, mixed CRLF, EOF newline) — Summary apply_patch writes the model's own bytes verbatim for every + line in an update hunk, so a successful patch silently rewrites file bytes the hunk never asked to change. Four user-visible...
- feat(ui): review install policy warnings — New behavior An authenticated administrator can review an install-policy warning in the Control UI and deliberately continue that plugin install. - plugins.install accepts the optional literal...
- fix(device-pair): verify retained subscribers before archiving the migration source — > REVIEW REQUIRED — persistent-state retention semantics. Prepared by the auto-QA campaign; intentionally left as a draft for maintainer review, not autonomous landing. What Problem This Solves...
- improve(installer): reduce Windows Git clone downloads — What Problem This Solves Fixes an issue where Windows users installing OpenClaw from Git download every repository blob even though the installer can use a blobless checkout, creating unnecessary...
- fix(gateway): pin runtime generation at admission — What Problem This Solves Gateway agent turns accepted on prepared runtime generation A could perform a later execution-time lookup after config, plugin, or auth publication advanced to generation B....
- refactor(gateway): consolidate node request validation — What Problem This Solves This maintainer-requested cleanup removes duplicated request-validation plumbing from node, push, and environment Gateway handlers. It is an internal refactor, not a change...
- fix(cron): surface command prompt capability failures — Summary - fail a retained agentTurn automation before model execution when its structured Command to run: block cannot use any shell/process tool - persist a structured cron-preflight error that...
- OpenClaw repo stats for 2026-08-27 — 387820 stars, 81416 forks, 3168 contributors
- OpenClaw sponsors (93 total) — 93 sponsors supporting OpenClaw
- fix: accept sourced repo for skill eval claims — Resolve GitHub source repo through githubSkillSources during SkillEvaluator claim stale checks, and let the temporary doca prioritizer reset its stale-version row.
- openclaw v2026.7.1-2 on npm — Multi-channel AI gateway with extensible messaging integrations
- OpenClaw — Deploy your own personal AI assistant with OpenClaw on a DigitalOcean Droplet® server, a powerful open-source platform that runs entirely on your infrastructure. OpenClaw connects to the messaging...
- OpenClaw is changing my life
- Nanobot: Ultra-Lightweight Alternative to OpenClaw
- StepFun 3.5 Flash is #1 cost-effective model for OpenClaw tasks (300 battles)
- How to Host OpenClaw for Multiple Clients in Production — The first OpenClaw deployment is usually straightforward. You provision a machine, configure one...
- I thought running OpenClaw from my phone was the goal, then I realized the phone should only say yes or no — A practical argument for using your phone as an approval layer for OpenClaw and agent workflows, not as the full control plane. Covers architecture, o
- You Send OpenClaw One Message. What Happens Next? A Deep Dive Into Its Architecture — Understanding the Gateway, Agents, Sessions, Models, Tools, and Skills Behind Every OpenClaw TaskContinue reading on Medium »
- One in Four of Those Skills Was a Trap — A skill called “What Would Elon Do?” was functionally malwareContinue reading on Medium »
- The Hidden Cost of Long-Running Agents — After six months of running OpenClaw across a 12-person team, we found that stale, re-fetchable tool output — not conversation — was often…Continue reading on onthelook »