OpenClaw· 2026-10-12TODAY@openclaw ↗ OpenClaw v2026.10.5-beta.2: 91-PR hotfix patch — update clears stale handoff leases, completed receipts, and legacy v2026.9.8 recovery records without stopping; Doctor preserves installed-plugin configuration and data when path safety checks block them; tool-approval hook hardening (ordinary hooks cannot rewrite trusted inline-approved arguments after approval); inter-session messages keep their safety envelope through replay; Control UI paired-device chat turns survive heartbeat/tab-reconnect noise; Codex runtime admission rejects official builds below v2026.9.7; updates skip rehashing unchanged runtime files (59 contributors, commit 467e83d)
v2026.10.5-beta.2, published October 12, 2026 (commit 467e83d), npm integrity: sha512-9OgCXW1ZUmMdBdtojEIp7P9qRh9hTHhTARNQfTLPTtKQViuNenvcjgJh6JaJsZx9b20J0FL4qXA4q1dkw795eg==, covering 91 pull requests since v2026.10.5-beta.1, with 59 contributors (including @bluet and @abacha). ① Updates and recovery — `openclaw update` clears stale handoff leases, completed receipts, and legacy v2026.9.8 recovery records; no longer stops on `update-recovery-pending`; handles Windows short-name paths; no longer reports accepted handoffs or successful updates as failures. ② Doctor — preserves installed plugins' configuration and data when path safety checks block them; leaves malformed UTF-8 root memory files untouched; captures Canvas and Memory Core files in recovery copies before migrations. ③ Tool approval hooks — ordinary hooks cannot rewrite trusted inline-approved arguments; plugins that transform arguments must request approval for their final arguments (see docs/plugins/hooks/tool-policy.md). ④ Subagent provenance — inter-session messages keep their [Inter-session message] safety envelope through replay, so child sessions do not treat parent instructions as direct user instructions. ⑤ Control UI — chat turns from paired devices survive routine pairing activity such as node heartbeats and tab reconnects; real revocations still stop them. ⑥ Codex runtime — plugin admission rejects official Codex runtimes below v2026.9.7 with update guidance; incompatible plugins stay inspectable without being imported. ⑦ Performance — updates skip rehashing unchanged runtime files, reducing I/O. ⑧ Other fixes — Tool Search and Code Mode network results taint the turn; node command approvals replay correctly with Windows quoted arguments; restart recovery for mid-turn main sessions; daily memory-note flushes reject copied content; prompt-prefix byte stability for whitespace-ending first messages; redundant host compaction skipped; Telegram polling reconnects; API-key sign-in discovers models immediately; OpenRouter cache-TTL pruning for Grok and Gemini 2.5+ models; MCP computer control no longer returns COMPUTER_HOST_BUSY; cron validation errors reported correctly; session store reservation no longer overflows; ChatGPT sign-in sessions keep Ultrafast on resumed turns.
OpenClaw· 2026-10-12TODAY@openclaw ↗ OpenClaw v2026.10.2: 85-PR stable release — first upgrade from v2026.10.1 now completes cleanly (fixes Windows, cloned-filesystem, private-Node paths); Doctor no longer stalls on residual legacy imports, Gateway locks, slow state inspection, memory reindex locks, or removed channel account allowlists; OpenAI and Anthropic API-key routes default to provider-native compaction; credential leaks across chunk boundaries fixed (large tool output no longer stalls or overflows stack, several times faster); Anthropic prompt cache stays byte-identical across follow-ups, replays, restarts, and subagent completions; fixes for Telegram, Claude CLI, subagents, OpenAI/Codex, cron/hooks, Control UI, and macOS (65 contributors, commit b8952d4)
v2026.10.2, published October 12, 2026 (commit b8952d4), npm integrity: sha512-FuLLF+WI12ypwQ0r+pcKDcmB+kXus5N4QClaKEnaAn6GN4I1AitkOhwv0GFl2rVKpvuUBDQWIfglHNy6Trv/EQ==, merges 85 pull requests from 65 contributors (including @jlapenna and @hawkit). ① Update reliability — the first upgrade from stable v2026.10.1 now completes cleanly; successful in-place updates no longer end in errors; Windows, cloned-filesystem, and private-Node update paths are fixed; live npm dist-tag drift no longer fails validated release candidates. ② Doctor and recovery — Doctor no longer stops on residual legacy imports, stale Gateway locks, slow state inspection, memory reindex locks, or allowlists for removed channel accounts; suppressed channels are reported as warnings rather than failures. ③ Provider-native compaction — OpenAI and Anthropic API-key routes now default to provider-native compaction; compaction recovers when the primary auth profile is cooling down. ④ Secret redaction — credentials split across chunk boundaries in long tool output are now masked; redaction of large text no longer stalls or overflows the stack and runs several times faster. ⑤ Prompt-cache stability — Anthropic checkpoints, enriched prompts, tool results, and inter-session provenance stay byte-identical across follow-ups, replays, restarts, and subagent completions. ⑥ Gateway responsiveness — Control UI reconnects, model changes with large plugins, dashboard title work, and catalog loading no longer freeze the Gateway or hang new sessions. ⑦ Key bug fixes — Telegram: replies lost while an agent starts, streamed replies vanishing when the replacement message never lands, local owner rejection treated as ambiguous delivery; Claude CLI: runtime prompts shown as user messages, routed CLI prompts on owner's side of history, lost replies after long streamed turns, post-turn compaction failing without an Anthropic API key; subagents and ACP: truncated answers, superseded child answers delivered by late receipts, requester turns stopping when completion delivery expires; agent runs: runs aborted as stalled after async tools or plugin approval, lost context-engine advancement after steering, tool loop refilling from the wrong context window; OpenAI and Codex: turns failing on regex-lookaround tool schemas, ChatGPT WebSocket expiry, 16 MiB SSE limits, strict chat templates, Code Mode repeated tool-call IDs, sessions stopping after thread unloads, and client leaks after model discovery timeouts; sessions: every turn failing after a truncated image, agent disowning earlier image answers, memory_search snake_case arguments rejected through tool_call; cron and hooks: scheduled runs losing Gateway plugin registry, manual runs losing tool access, runs rejected after catalog publication, preflight compaction skipped for completion turns, Max thinking dropped on cron and hook turns; Control UI: hold-to-dictate stopping on release, typed /stop letting queued turns restart the agent, repeated browser replies, pending Home notices piling up; updates: failures from absolute avatar paths, refused finalize steps leaving Gateway stopped, SQLite worker launch error hidden behind TypeError; workers and browser: capacity leaks after forced descendant cleanup, browser tab eviction at managed tab cap; platform: Feishu replaying obsolete streaming previews, Bedrock custom provider using wrong region, egress proxy CA rejected by strict TLS clients, quoted Windows shell arguments; macOS: Swift compiler crash in sidebar picker, Rosetta segfaults, Intel worker startup, managed llama.cpp setup on macOS below 13.3.
Hermes· 2026-10-11@NousResearch ↗ Hermes v0.21.7: 204-PR patch — source installs follow latest stable tag instead of main (Desktop adds update channel switch); signed thinking blocks persist across hermes chat --resume and gateway turns; Bedrock Converse reasoning models allowed on resumed sessions; new ctx.runAction/ctx.pet.say plugin API; skills installed mid-session reach the model without a prompt rebuild; Krea 2 Medium Turbo becomes default image lane for paid subscribers with FAL fallback; 7 new community plugins: memory-shield, command-ledger, Memoria, Model Router, wyoming-voice-bridge, codex-account-switch, redline (commit 71d34ba)
v0.21.7, published October 12, 2026 at 05:53 UTC (release body notes October 11, 2026, commit 71d34bad), merges 204 pull requests and 740 non-merge commits since v0.21.6, closing 69 issues and touching 4,156 files (+86,775/−26,980); full contributor credits are deferred to v0.22.0. ① Update channels — source installs now update to the latest tagged vX.Y.Z release instead of every commit on main; installs newer than the latest release stay where they are; to follow every commit run `hermes update --set-channel main`; forks and mirrors still follow main; Store, MSIX, and bundled desktop installs are unaffected; the desktop app adds an update channel switch under Settings ▸ About ▸ Updates, and the update dialog shows the current channel. ② Reasoning persistence — signed thinking blocks replay after `hermes chat --resume` and on the next gateway turn; Bedrock Converse reasoning models are no longer rejected on resumed sessions. ③ Update reliability — installs on main no longer fail on the release record fetch that some regional firewalls answer with 403; the stable check uses a GitHub token (GITHUB_TOKEN or gh auth login) when one is set; GitHub rate limits are now reported as rate limits. ④ Plugin API — new ctx.runAction, ctx.pet.say, and core-made Codex requests so plugins never hold the token. ⑤ Mid-session skills — skills installed during a session reach the model without a prompt rebuild. ⑥ Image lane — Krea 2 Medium Turbo is the default for new paid subscribers; FAL fallback when managed Krea refuses a request. ⑦ Other — faster desktop updates through cached compilation and reuse of unchanged inputs; pinned installs check out only the pin; compaction marks emitted to Relay; 7 new community plugins: memory-shield, command-ledger, Memoria, Model Router, wyoming-voice-bridge, codex-account-switch, redline.
OpenClaw· 2026-10-11@openclaw ↗ OpenClaw v2026.10.5-beta.1: X (Twitter) channel plugin, voice-call per-call briefs/call reports/live steering/voicemail detection, ClawHub built into Gateway for plugin and skill browsing and search, model picker with recommended models first (newer Claude/ChatGPT/Ollama without waiting for a release), safer updates (Windows elevation refusal, Gateway restart on interrupted activation, diagnostics overhaul), restart-safe conversations resuming after Gateway restarts, Gateway multi-threaded reads (~80 PRs), concurrent worktree creation with background GC; ~1,971 PRs, 200 contributors (commit f283c3c)
v2026.10.5-beta.1, published October 11, 2026 at 13:08 UTC (commit f283c3c), npm integrity: sha512-t9A6CO0HO2gKret1smbimDnxtaupU9cWqtdUQhqKpXtgnsDtorFZnbStUxqpE34x9CDOZ7CqQNCN2bUD07O8ug==, approximately 1,971 unique PRs, 200 contributors (including shanselman and vincentkoc). ① New channel — X (Twitter) channel plugin for listening to tweet mentions. ② Voice calls — per-call briefs, call reports, live steering, callback support, and voicemail detection. ③ ClawHub — plugin and skill browsing and search built directly into the Gateway. ④ Model picker — recommended models listed first; newer Claude, ChatGPT, and Ollama models surface without waiting for a release. ⑤ Safer updates — Windows refuses non-elevated updates immediately with an elevation hint (#166695); Gateway restarts and records the interruption when activation is signaled mid-run (#164849 and others); diagnostics name each candidate, preflight result, and resolved runtime path; Doctor migration fixes (archive migration committed per batch, legacy session entry state migration, SQLite backup preservation); install fixes (npm network environment settings preserved, Windows Bun packaging unable to find npm). ⑥ Restart-safe conversations — interrupted turns and deferred tasks resume from a checkpoint after a Gateway restart. ⑦ Gateway performance — chat sends acknowledged before skill preparation; session and history reads moved to worker threads; reconnect and restart speeds improved; shared reads and broadcast encoding optimized (~80 PRs); worker and CLI overhead reduced with cron and session scans off main thread (~40 PRs); update and Git skip unnecessary fetches (~30 PRs); OAuth refresh and auth saves moved to workers (~17 PRs). ⑧ Worktrees and Git — new session worktrees created concurrently; background GC with capacity limits and backoff; leaked branches recovered; missing commit-graph tips and stale pack indexes healed. ⑨ Workboard — automation, home, and Board sessions kept off Sessions boards; run and PR facts accurate; dispatch refusals explained correctly. ⑩ Skill Workshop — approval-first setups no longer write skills without review while keeping automatic learning for approval-policy configs. ⑪ Plugins — fixes for Windows Gateway startup stalls on plugin file capture (#168313), mixed-case plugin ID resolution (#164683), Gateway freezing on model changes with large plugins (#161267), and Copilot tool calls failing after the first turn (#167621).
OpenClaw· 2026-10-10@openclaw ↗ OpenClaw v2026.10.1: MCP Apps run beside chat with attachment and workspace-file read/write access, external backups to disk/Cloudflare R2 with schedules and restores, conversation snooze on Web/iOS/Android/Mac sidebar, Workboard for organizing sessions by run status and pins, Bun-only install on macOS/Linux without Node (Bun 1.4+), agent installed-skill search in embedded/Code Mode/Codex sessions, Code Mode working values persist across cells and restarts, YouTube cards and local audio/video inline up to 4 GiB, Mac sidebar grouping/multi-select/search/plugin conversations; model lists and prices update online without restart; upgrade note: pre-July 2026 Workboard data needs doctor --fix on 2026.9.7 first (3,055 commits, 2,425 PRs, 402 contributors, commit e178c39)
v2026.10.1, published October 10, 2026 at 23:54 UTC (commit e178c39), merges 2,425 pull requests and 3,055 commits from 402 contributors. npm integrity hash: sha512-C072/lxy83KP1nMCnJbipr4O+EtIR9XXbjzQFM+ol7VE5ci2MkKk5XTxfSG70jxpEVaGFeXcEmObHxR/ZHLdHw==. ① New features — MCP Apps: compatible apps run beside chat, take selected content as attachments, and can view or edit workspace files with approval controls; file editors need to check for stale saves to avoid overwrites. External backups: full backups can go to an external drive, mounted storage, or Cloudflare R2, with opt-in schedules and retention; restores land in a fresh staging directory. Conversation snooze: Web, iOS, Android, and the experimental Mac sidebar can hide eligible conversations until a chosen time. Workboard: session boards organize conversations by run status, placement rules, and pins (note: Workboard data from before July 2026 needs `doctor --fix` on 2026.9.7 first; rolling back shows Sessions boards as empty Cards boards). Bun-only install: macOS and Linux can run openclaw without Node given Bun 1.4 or newer. Agent installed-skill discovery: embedded agents, Code Mode, and Codex sessions can search installed skills and read a selected skill's full instructions; search covers only part of very long instruction files. Code Mode working values: small JSON results persist across cells, replies, and restarts. Bundled desktop runtimes: fresh Mac and Linux setups can use bundled runtimes, and the Mac app can keep its Gateway running after the app closes. Model discovery: model lists and prices update without a restart; OpenAI API-key users can select chat models their account offers beyond the bundled catalog. Media enhancements: assistant replies can show YouTube cards; allowed local audio and video play inline up to 4 GiB; HTML previews can load nearby assets within limits. Mac sidebar gains grouping, owners, multi-select, search, plugin conversations, and translations. Resumable worker installs: interrupted cloud-worker downloads resume and are verified before use. Smaller npm package: native binaries for unrelated platforms are excluded. Memory plugins: compatible plugins serve recall and save context before compaction. Offline availability: cached conversations and drafts stay usable offline; saved drafts and queued messages are easier to recover. ② Security notes — cached conversations and drafts can be read by anyone with access to that browser profile; overwritten or deleted Code Mode values remain in conversation history and exports; on multi-user Gateways an agent with the exact path can read another session's media in the shared store; pairing credentials may gain a longer lifetime after upgrading, and closing enrollment alone does not revoke existing approvals; old Teams and Nextcloud Talk webhook listeners remain reachable on all interfaces until migrated. ③ Upgrade notes — replacing standalone webhook plugins requires core 2026.9.9 or later; Nextcloud Talk users with the retired private-network setting need the 2026.9.5 intermediate upgrade first; older updaters may reject the smaller npm package and need a manual reinstall.
Claude Code· 2026-10-09@AnthropicAI ↗ Claude Code v2.1.296: MCP tool-description and server-instruction limit raised to 4,096 chars; gateway `code` policy key for Claude Desktop Code tab; subagent `autoCompactWindow`; `CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL` and `CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS` env vars; Read `allow_large`; Sonnet 5.5 cache-read repriced to $0.10/Mtok (was $0.20); ← now stops background turns instead of finishing silently; 60+ fixes including managed PreToolUse hooks, Haiku 5.5 token counts, non-UTF-8 file edits
v2.1.296, published October 9, 2026. ① Changes — MCP tool-description and server-instruction default limit raised from 2,048 to 4,096 chars; pressing ← to background a session now stops an in-progress turn or ! command instead of letting it finish silently; [VSCode] Claude in Chrome asks before browser actions every session, consistent with the terminal; [Claude Tag] Slack replies show the footer with session link, model and cost only on the thread's newest reply. ② Additions — a `code` key in `managed.policies[]` carries the same settings as `cli` into Claude Desktop's Code tab; `autoCompactWindow` in subagent frontmatter and `--agents` lets a subagent compact before the main window would; `CLAUDE_CODE_WORKFLOW_SUBAGENT_MODEL` runs every workflow agent on one model; `CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS` sets the maximum backoff for 529 retries; `/plugin` notes hooks dropped because another same-named plugin is enabled; Read tool gains `allow_large` to read a large text file in one call when context allows; [Cloud sessions] status filter (multi-select) added to Sessions and Runners lists in admin; [Claude Tag] members with Claude Tag Admin may create, edit and delete workspace and channel memory files on the Activity page. ③ Improvements — fullscreen mode underlines hovered links; Ctrl+O transcript toggle is faster by caching syntax-highlighted blocks; `--debug` names unrecognized frontmatter fields in custom agent files; auto mode shows a dim "Not run" row for tool calls with no usable answer; command hooks log command, plugin, outcome and duration in `--debug`. ④ Updates — Sonnet 5.5 cache-read repriced from $0.20 to $0.10/Mtok; dataviz skill color update. ⑤ Fixes — managed PreToolUse hooks continuing after deny with `"continue": false`; PostToolUse `updatedMCPToolOutput` not applying in some sessions; headless sessions starting disabled .mcp.json servers; a gateway with `"gateway"` in allowedProviders locking out user-configured laptops; 2.1.295 regression ignoring saved gateway sign-in without forceLoginGatewayUrl; `--teleport` opening empty session on unreadable history; token-count failures for Haiku 5.5 and adaptive-thinking-only models; resumed subagents wrongly told a tool call was rejected; hook output tampered when it contained plugin-hint-like text; secret redaction missing values after key-with-no-value; stray escape in MATE Terminal after copy; notifications queued unseen while /diff panel open; Esc during UserPromptSubmit hook ending headless sessions; Edit/NotebookEdit corrupting non-UTF-8 files (now refused); prompt running twice silently after ←; Bash BASH_ARGV0 assignment auto-approved; plus 60+ more fixes across workflow scripts, plugins, cloud sessions, Claude Tag, and Code Review.
Codex· 2026-10-09@OpenAI ↗ Codex CLI rust-v0.162.1: fix TUI crash when async questions span multiple lines; fix startup failures from background-server feature-setting mismatches with CLI defaults (commit 092d3ac)
rust-v0.162.1, published October 9, 2026 (commit 092d3ac). This release contains two bug fixes: ① fixed a TUI crash when asynchronous questions span multiple lines, preserving line breaks and complete hyperlink destinations (PR #51866); ② fixed startup failures caused by mismatches between a running background server's feature settings and CLI defaults — compatibility checks now apply only to explicit command-line feature overrides (PR #52648).
Claude Code· 2026-10-08@AnthropicAI ↗ Claude Code v2.1.295: onFailure:"block" for command and HTTP hooks, OSC 7501 Program Status Protocol for terminal working-state display, $.ui.notify for mods raises native notifications, timeouts.upstream_ttfb_ms for gateway upstreams, forceLoginMethod:"gateway" in user settings, MCP tool-search description limit raised to 16,384 chars; fixes [1m] model failures on gateway/Bedrock/Vertex/Foundry refusal, terminal freezes on deep nesting/quotes and gigabytes of memory, syntax highlighting freezes on very long lines, VS Code file-send row, and 50+ more (commit 602df92)
v2.1.295, published October 8, 2026 at 19:48 UTC (commit 602df92). ① New features — `onFailure: "block"` for command and HTTP hooks: a hook that fails to start, times out, or exits unexpectedly blocks its action; Program Status Protocol (OSC 7501) lets terminals show whether Claude Code is working, waiting, or done; `$.ui.notify` for mods raises a native notification and names the sending channel; mod `Button` now accepts children (strings and `Text`); `timeouts.upstream_ttfb_ms` for the gateway's cloud upstreams (Bedrock, Vertex, Foundry) limits how long a stream may take to start, then fails over or returns 502; `forceLoginMethod: "gateway"` and `forceLoginGatewayUrl` in user settings force `/login` to open the gateway on machines without managed settings; `CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS` caps how long unattended retry mode waits on 429/529; optional `models` list on each gateway upstream restricts which models are sent; MCP tool-search description limit raised from 2,048 to 16,384 characters; subagents preload at most 32 skills, each once. ② Key bug fixes — fixed every request failing on a `[1m]` context model when a gateway, Bedrock, Vertex, or Foundry refuses the context-1m beta (Claude Code now resends without it); fixed terminal freezing and consuming gigabytes of memory when replies nested thousands of levels of quotes or lists; fixed syntax highlighting freezing on very long lines, long blank-line runs, or unclosed strings; fixed `claude -p` dropping earlier turn output when background work started another turn; fixed remote MCP servers staying disconnected after a 15-second outage or reconnecting in a tight loop, backing off up to 30 s; fixed Bash calls failing when the model passes `command_description`; fixed text with tabs or bidirectional control characters losing its end or drawing over nearby rows; VS Code: added a chat row for files Claude sends, click to open in editor; fixed cloud-session messages right after unarchiving getting no reply; fixed `/model` claiming a Max effort pick was saved as default (Max is session-only). ③ Other changes — claude.ai connectors negotiate MCP protocol 2026-07-28 by default; Grep accepts `-l`, `-c`, `-r` flags; `claude agents` stops when its background service stops; Artifact pages with no background now show on white.
