- compaction: absolute token thresholds break when switching models with different context windows — Problem Four compaction parameters use absolute token counts, not relative percentages of the context window: Meanwhile, contextThreshold (0.75) and maxHistoryShare (0.8) correctl
- fix(slack): keep command tests inside plugin boundary — What Problem This Solves Resolves a problem where the Slack slash-command test exercised core reply-dispatch internals instead of the public plugin SDK contract. That made the release-validation test...
- fix(memory): bisect embedding batches on provider item-count 400s — Closes #125640 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 Fixes an issue where...
- feat(agent): allow explicit one-shot tools — Related: #126255 What Problem This Solves openclaw agent exec uses a restrictive one-shot tool profile, but callers cannot add one explicitly trusted tool without writing persistent configuration....
- fix(plugins): pretag pack checks stop hanging on stalled commands — What Problem This Solves\n\nFixes an issue where the fast plugin pretag check could return after timing out a stalled build or pack command while bash/npm descendant processes remained alive and...
- fix(release): accept frozen QA completion evidence — What Problem This Solves Resolves a release-validation problem where a successful frozen QA runtime-pair run could be rejected after the legacy producer omitted its terminal run.status, forcing an...
- fix(release): preserve validation state across reruns — What Problem This Solves Fixes an issue where rerunning a failed Full Release Validation parent job could lose attempt-scoped artifacts, preventing the parent from adopting its already-dispatched...
- Feature: Surface cron job name / session label in /status and statusline — Problem When viewing /status in the TUI, cron job sessions and their runs display as raw UUIDs: The session store already has a label field (e.g., "Cron: claude-keepalive") and cron jobs have a name...
- fix(control-ui): retain webchat context for config restarts — Summary Carry { channel: "webchat", to: sessionKey } on Control UI config.apply and config.patch requests. This preserves the restart wake for the originating webchat session while preventing a stale...
- feat(browser): add policy-bound Browser Use CLI 3.0 — Related: #19289 What Problem This Solves OpenClaw agents need a supported way to use Browser Use CLI 3.0 against an orchestrator-owned Cloud browser without giving an external executable independent...
- Codex dynamic tool progress duplicates a single call in Telegram — Bug type Behavior bug Beta release blocker No Summary A single Codex dynamic tool invocation can appear twice in the same Telegram progress draft: once as an id-less textual result summary and once...
- fix(codex): keep Computer Use working after desktop updates — Related: #127024 What Problem This Solves Fixes an issue where Codex-native Computer Use could keep using an app-server loaded from the previous ChatGPT or Codex desktop build after the desktop app...
- [Bug]: Model picker only applies changes to new sessions; Ollama Cloud requests sign-in despite configured API key — Bug type Behavior bug (incorrect output/state without crash) Beta release blocker No Summary I am seeing two related problems in the model picker in the OpenClaw Control UI: 1. Changing the model...
- refactor(plugins): unify synchronous message hooks — What Problem This Solves The synchronous message-hook wrappers duplicated sequencing and failure policy. That left tool_result_persist and before_message_write vulnerable to drifting apart on event...
- fix(onboard): emit JSON on option-validation failures — What Problem This Solves openclaw onboard --json promises a machine-readable result, but only some terminal outcomes produce one. The success path writes a JSON object, and the top-level CLI error...
- No plugin-facing hook can force agent-run termination -- before_tool_call denies calls, nothing can stop a turn — Summary No plugin-facing hook can force an agent run/turn to stop. A plugin can deny individual tool calls (before_tool_call), but has no way to end the session once it has correctly detected that...
- improve(gateway): avoid targetless session cache allocations — What Problem This Solves Ordinary authenticated Gateway RPCs without a session target, such as status, allocated two empty session lookup maps during authorization even though they never performed a...
- macOS: surface realtime Talk settings — What Problem This Solves macOS operators need a direct handoff from Voice & Talk settings to the existing Control UI Talk configuration. The Control UI must also preserve a configured realtime...
- fix(mantis): publish proof comments on auto-triggered runs — What Problem This Solves Auto-triggered Mantis proof runs (label/clawsweeper_label and other non-comment request sources) end with no PR comment at all. The durable evidence publisher runs with...
- [Bug]: Android Realtime Talk suppresses the microphone for all of assistant playback, so it cannot be interrupted — Bug type Behavior bug (incorrect output/state without crash) Beta release blocker No Summary On Android, Realtime Talk discards captured microphone audio for the entire duration of assistant...
- improve(gateway): avoid copying single-chunk request bodies — What Problem This Solves Reduces CPU and allocation overhead when Gateway HTTP endpoints read request bodies delivered as a single chunk. The previous completion path copied that chunk into a new...
- chore(deps): bump the actions group across 1 directory with 11 updates — [//]: (dependabot-start) ⚠️ Dependabot is rebasing this PR ⚠️ Rebasing might not happen immediately, so don't worry if this takes some time. Note: if you make any changes to this PR yourself, they...
- fix(onboard): reject --gateway-bind custom without a valid customBindHost — What Problem This Solves Non-interactive onboarding accepts --gateway-bind custom without a gateway.customBindHost, writes bind: "custom" to the config, and reports success. The Gateway then refuses...
- fix(talk): share and bound Apple relay lifecycle — What Problem This Solves The iOS-owned realtime Talk relay could not be reused by macOS and had unbounded or ambiguous audio lifecycle failure paths. Why This Change Was Made The canonical Apple...
- fix(exec): prevent zsh literal expansion failures — Closes #126521 What Problem This Solves Fixes an issue where users running exec commands through zsh would have the entire command chain abort when model-generated shell text contained an...
- macOS: add realtime Gateway-relay Talk support — Related: https://github.com/openclaw/openclaw/issues/71195 and https://github.com/openclaw/openclaw/issues/116201 What Problem This Solves macOS Talk lacked a reusable, bounded Gateway realtime relay...
- perf(sqlite): skip schema write lock on current state — What Problem This Solves Every fresh writable process open entered state.schema.ensure with BEGIN IMMEDIATE, even when the shared-state database was already canonical for the running OpenClaw...
- fix: stop limited access status from shifting page layouts — What Problem This Solves Fixes an issue where users connected with limited device access would see the global access status consume a full row inside routed pages, making Automations and other narrow...
- OpenClaw repo stats for 2026-08-22 — 387080 stars, 81307 forks, 3120 contributors
- OpenClaw sponsors (94 total) — 94 sponsors supporting OpenClaw
- fix(ci): bind package publish to release parent (#3500) — Bind scoped upload and publish credentials to one server-derived authorization transaction, retain fail-closed v2 validation through staged finalization, and preserve idempotent same-storage retries...
- fix: resolve skill evals for projected skill pages — Resolve SkillEvaluator lookup for projected skill detail rows by falling back to NVIDIA source repo/path and returning null when no unique skill can be resolved.
- docs: update Git & GitHub skill count to 160
- ykarout/Qwen3.5-9b-Opus-Openclaw-Distilled-GGUF — text-generation
- sulpikar2/Qwen3.5-4B-Claude-4.6-Opus-Reasoning-openclaw
- 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...
- StepFun 3.5 Flash is #1 cost-effective model for OpenClaw tasks (300 battles)
- OpenClaw’s memory is unreliable, and you don’t know when it will break
- Setting Up an OpenClaw Agent o… — Originally published at norvik.tech Introduction Explore the setup of an OpenClaw agent...
- I thought OpenClaw would just install and instead I got dropped into provider hell — A practical DEV post on why OpenClaw installs often fail at the provider layer, not the package layer. The piece argues for a boring but reliable setu
- Grok Bot Is Staff. The Computer Is the Product. — What it is, how I use it, and who it is actually for.Continue reading on Medium »
- The Assistant Needs a Computer. Who Runs It? — OpenClaw, Hermes, and Grok Bot are the same category. The step up is not a smarter model.Continue reading on Medium »