- fix(byteplus): report token usage for streamed responses — AI-assisted PR (written with Claude). I understand what the change does. What Problem This Solves Fixes an issue where users on the byteplus and byteplus-plan providers get no token usage for...
- fix(release): authorize staged ClawHub publication — What Problem This Solves The protected OpenClaw release publisher could stage ClawHub packages without a parent-owned authorization artifact bound to the exact candidate, tooling, child attempt, and...
- feat(nodes): expose plugin duplex channels — What Problem This Solves Gateway-loaded plugins previously could not open a generic, provider-neutral, bidirectional binary channel to a declared command on a paired node. Existing node invocation...
- Dreaming deep phase: minUniqueQueries gate bypassed by day-diversity counting — Summary The deep phase promotion gate minUniqueQueries does not require genuinely distinct recall queries. In the source code (short-term-promotion-.js, lines ~1198-1200 and ~1438-1439), the gate...
- fix(memory): keep queued dreaming work within subagent limits — Closes #95746 What Problem This Solves Fixes an issue where Memory Core dreaming work could bypass agents.defaults.subagents.maxConcurrent by using per-session lanes. Routing that work through the...
- read tool emits omission note for non-vision models but still sends the full image payload (#127009 source fix) — Summary The read tool base64-encodes the full original image into the LLM context even when the active model declares no image input capability (input: ["text"]). The text note for a non-vision model...
- Document practical requirements for QMD memory backend on CPU-only VPS deployments
- fix(read): drop image payload for non-vision models — What Problem This Solves Closes #127009. The read tool base64-encodes the full original image into the LLM context even when the active model declares no image input (input: ["text"]). For non-vision...
- fix(scripts): typecheck hangs forever when tsgo wedges instead of failing — Typechecks can hang forever when tsgo wedges, leaving local terminals and CI lanes blocked while an orphaned compiler holds gigabytes of memory. The wrapper prepared a managed invocation but bypassed...
- fix: Tool Search directory/tools wrap native read/exec in tool_call — Closes #126460 What Problem This Solves Fixes an issue where Tool Search directory/tools modes caused openai-completions models to call meta tool_call for ordinary read / exec work instead of the...
- fix: announce Worktrees operation failures — What Problem This Solves Fixes an issue where operators using the Worktrees page could see a failed list, create, remove, restore, or cleanup operation, but screen readers received no announcement...
- fix: Android Chat mic tap records a voice note when on-device dictation is unavailable — Related: #125322 Related: #125273 What Problem This Solves Fixes an issue where users tapping the Android Chat composer microphone would see “On-device speech recognition is unavailable.” (localized,...
- fix: HTTP chat binds one unbounded session from a constant OpenAI user — Closes #126461 Related client fix: https://github.com/leveragedrobot/raycast-moltbot/pull/7 What Problem This Solves Fixes an issue where HTTP Chat Completions clients that send a constant OpenAI...
- Bug: DeepSeek V4 Flash incomplete turn (payloads=0, tools=2, replaySafe=no, stopReason=stop) in 2026.5.27/28 — Summary DeepSeek V4 Flash via OpenRouter produces incomplete turns (payloads=0, tools=2, replaySafe=no, stopReason=stop) in OpenClaw 2026.5.27 and 2026.5.28. The model worked without issues in...
- fix(talk): activate configured realtime voice provider — Closes #126594 What Problem This Solves Fixes an issue where users configuring Realtime Talk through talk.realtime would fail to start a bundled realtime voice provider unless they also configured an...
- fix: HTTP chat still sends a full system prompt when tools.profile is minimal — Closes #126459 What Problem This Solves Fixes an issue where HTTP /v1/chat/completions always ran the full agent system prompt, so a chat agent with tools.profile: "minimal" still paid tens of...
- fix(plugins): preserve channel adapters in narrow tool scopes — Problem External plugin tools execute inside a registry containing only their owning plugin. Channel adapter lookup selected that entire scoped registry, so a narrow registry without Telegram hid the...
- fix(gateway): recover embedded channel runs in chat startup projection — Related to #121756 What Problem This Solves When a run started from a chat app (e.g. Telegram) is still active, opening that session in Control UI shows only an empty working timer — the page cannot...
- [Feature]: [Feature Request] Plugin contract should support thinkingOverride to enable dynamic thinking level (low/medium/high/xhigh) per round — Summary Problem Currently, OpenClaw's plugin contract does not support thinkingOverride. As a result, it is impossible to dynamically adjust the thinking intensity level (e.g., low / medium / high /...
- Expose structured progress metrics in hook events for plugin-based live progress cards — Background We have integrated openclaw-progress-plugin in Feishu DM and validated stable live card updates (send/update flow works). Current UX gap: progress is mostly textual (进行中/执行中) because...
- improve(ui): put session row actions on the title line — What Problem This Solves Two-line sidebar rows centre the pin and menu buttons across the whole row, so they float between the title and the line below. Hover then reserves their width from the...
- fix(cli): reject --agent on models aliases and scan instead of ignoring it — Closes #126597. What Problem This Solves openclaw models aliases {list,add,remove} and openclaw models scan accept --agent from the parent models command and do nothing with it — no scoping, and no...
- Reasoning stream — Summary Is it possible to have openclaw write and overwrite lines like openai and grok do to indicate the thinking process? Normally, messages appear like "searching google, scanning code, installing...
- fix(cron): surface Code Mode tool failures in cron run history — Related: #120385 What Problem This Solves Operators debugging a failed Code Mode cron run see only a generic ⚠️ 🛠️ Exec failed in run history, even though OpenClaw already captured the specific...
- OpenClaw repo stats for 2026-08-21 — 386983 stars, 81284 forks, 3116 contributors
- OpenClaw sponsors (94 total) — 94 sponsors supporting OpenClaw
- fix(ci): prevent moved tooling from reaching package publish (#3499) — fix(ci): revalidate trusted tooling before publish fix(ci): bind tooling trust to publisher run
- fix: await publish webhook follow-up (#3496)
- fix(api): stream bulk skill exports (#3498) — fix(api): stream bulk skill exports fix(api): tolerate stale bulk export storage test(cli): make GitHub publish e2e hermetic fix(api): harden streamed export boundaries docs(api): define streamed...
- 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 a security nightmare dressed up as a daydream
- OpenClaw surpasses React to become the most-starred software project on GitHub
- Show HN: Klaus – OpenClaw on a VM, batteries included
- Beyond AI Chat: Building Real Infrastructure with OpenClaw + Kubernetes — 💻 This week I started exploring OpenClaw in a hands-on lab by integrating it with Minikube, Kubernetes, and Helm. Rather than using AI as…Continue reading on Medium »
- I Didn’t Hire an AI Team. I Configured One. — How I run 13 AI agents out of a Discord server on one Linux box — and how you can copy the idea.Continue reading on Medium »
- Inqora Engineering — This talks about the engineering done while building Inqora and some experiments. It covers the following topics:Continue reading on Medium »
- Donely’s BYO Model Problem — Do not pay for Donely until one free-tier agent proves that the dashboard saves more operator time than the hosting premium costs. The…Continue reading on Medium »