Agent Breaking Changes_

← OpenClaw

OpenClaw openclaw 2026.10.5

2026-10-18 · official notes

Summary

• Updates finish and recover on their own: openclaw update completes across devices, filesystems, pnpm 12, Windows elevation, and macOS login shells, settles stale handoff leases and package-activation
• Restart-safe conversations: interrupted turns, first messages during worktree preparation, subagent work, and deferred tasks resume after a Gateway restart, and user input and committed session data s
• Faster, lighter Gateway: session, transcript, history, and metadata reads move onto worker threads with far less SQLite work per turn, startup maintenance runs after readiness, and Gateway startup and

Full notes

## 2026.10.5

### Highlights

- **Updates finish and recover on their own:** `openclaw update` completes across devices, filesystems, pnpm 12, Windows elevation, and macOS login shells, settles stale handoff leases and package-activation journals instead of stranding the install, keeps busy or unhealthy Gateways from blocking the update, and `--no-restart` package updates now report success honestly. (#164769, #165863, #166695, #167302, #169410, #170046, #172196, #172640, #173358, #173409, #174649, #171015) Thanks @shanselman, @patrickschiller, @RomneyDa, @SergeiAug, @giulianoaccorsi, @mwbroxson, and @Bamorem.
- **Restart-safe conversations:** interrupted turns, first messages during worktree preparation, subagent work, and deferred tasks resume after a Gateway restart, and user input and committed session data survive restarts, concurrent writes, and failed publications. (#165333, #166042, #165184, #168650, #169144, #170161, #170223, #165502, #167570, #166010, #167662) Thanks @DonnieFi, @jjjhenriksen, and @RomneyDa.
- **Faster, lighter Gateway:** session, transcript, history, and metadata reads move onto worker threads with far less SQLite work per turn, startup maintenance runs after readiness, and Gateway startup and memory search workers use less memory, back in the 2026.10.4 range. (#164744, #165470, #165614, #166587, #167194, #164779, #165028, #166041, #167278, #165993, #174718, #172392) Thanks @obviyus, @RomneyDa, @DonnieFi, and @jjjhenriksen.
- **Reliable turns and subagents:** turns recover from stalled streams, reused tool-call IDs, expired ChatGPT websockets, and malformed OpenAI or Google streams, incomplete turns end with a visible notice, and subagent results keep returning through restarts. (#165983, #167013, #165212, #167972, #166846, #168400, #173404, #173612) Thanks @cestercian, @Irish-Joseph, @rmyers64, @sszyh2026, and @RomneyDa.
- **X channel:** answer X (Twitter) mentions through a new channel plugin with streaming, guest mode for repository questions, and enforced API cost limits. (#165138, #165337, #165354, #165351)
- **Voice calls:** voice calls gain per-call briefs, call reports, live steering, callbacks, and voicemail detection. (#165716) Thanks @mbelinky and @jacobtomlinson.
- **Smarter model pickers and new models:** pickers lead with each provider's recommended models, show newly listed Claude, ChatGPT, and Ollama models without waiting for an OpenClaw release, and Google adds Gemini 4 Argon and Gemini 3.8 Flash. (#167370, #166737, #167658, #167387, #167463, #171852, #172339) Thanks @obviyus.
- **ClawHub in the Gateway:** browse and search ClawHub plugins and skills from the Gateway, and plugins automatically pick the newest official release your OpenClaw supports. (#163459, #167485, #166588) Thanks @jacobtomlinson and @charlesBerg-ux.

### Changes

- **Tool approval hooks:** ordinary hooks can no longer rewrite trusted-inline-approved tool arguments without another approval, and transforming plugins must request separate approval for their final arguments or the call is blocked; see `docs/plugins/hooks/tool-policy.md` for migration. (#142700) Thanks @giodl73-repo and @giodl73.
- **Sessions:** allow directed sends to a session without granting transcript access. (#164444)
- **Worker-local inference:** add a native worker inference runtime with Gateway dispatch, client placement, required worker policy and destination checks before local inference, and clear setup-failure explanations. (#158902, #163645, #163646, #163647, #166611, #166613, #165960) Thanks @RomneyDa.
- **Device joins:** the CLI and Control UI join command now installs a background node service that can host agent sessions (`--service --session-host`), with a command-only service alternative, and already-paired devices get a working command to enable session hosting. (#168623) Thanks @obviyus.
- **Public threads:** open public threads at their regular links. (#166049)
- **Subagent progress:** count running subagents on the working line, name launched subagents in the operations group, and continue a resumed answer in the block that handed off. (#165477, #165725, #165542) Thanks @shakkernerd.
- **X channel:** add the X (Twitter) mentions channel plugin with Activity API streaming and backfill, repository-question guest mode with hidden same-agent helpers, GitHub-verified repository writers, and daily and monthly X API cost limits. (#165138, #165337, #165376, #165354, #165543, #168071, #165351) Thanks @obviyus.
- **Automation repair:** let a failing automation's owner conversation repair it before alerting, give a second repair turn when the error changes, and post a clear warning when a repair turn times out. (#161007, #162702, #162772) Thanks @obviyus.
- **Channel progress:** show useful Telegram and Discord progress by default, add per-tool glyphs to Discord progress drafts, restore per-tool emoji on Telegram tool rows, and hide routine Slack reaction rows. (#166861, #168032, #167392, #166880) Thanks @obviyus, @keshavbotagent, and @notmnp.
- **Skills:** keep learned skills focused on repeatable procedures rather than facts or codebase knowledge, and keep longer legacy Workshop skills manageable. (#167404, #167639) Thanks @obviyus.
- **Memory:** explain why trigger recall is skipped and why embedding responses are rejected, and report managed Ollama and LM Studio embedding readiness. (#164409, #168035, #156483) Thanks @Ayushdevo, @wangmiao0668000666, and @obviyus.
- **Models and catalogs:** show each provider's recommended models first in pickers, publish curated per-provider recommendations in catalog v2, lead NVIDIA recommendations with its featured models, and apply downloaded catalogs without a Gateway restart. (#167370, #166737, #167989, #158000) Thanks @obviyus.
- **Google:** add Gemini 4 Argon and Gemini 3.8 Flash, and make Gemini 4 Argon the default Google model. (#171852, #172339)
- **Anthropic and Claude CLI:** show newly listed Claude models without waiting for an OpenClaw release, list each Claude CLI model once, and keep logged-out Claude CLI models listed with a login reason. (#167658, #166311, #166338, #167942) Thanks @obviyus.
- **OpenAI:** support custom realtime voice endpoints, show account chat models missing from the manifest, surface new ChatGPT-login models without a release, and stop offering ChatGPT-only or bundled-Codex-rejected models that would fail. (#165632, #162563, #167387, #167633, #167975) Thanks @jacobtomlinson, @obviyus, and @RomneyDa.
- **Codex runtime compatibility:** plugin admission rejects official Codex runtimes below 2026.9.7 with update and repair guidance, and incompatible plugins stay inspectable without being imported. (#168930, #168748, #168749, #168977) Thanks @ezimerman and @RomneyDa.
- **Context limits:** status shows corrected context limits after model discovery, and llama.cpp keeps live router model limits and tools on refresh. (#165991, #168309) Thanks @snotty, @saju01, and @RomneyDa.
- **Tool rows and subagent progress:** show Tool Search calls in native tool rows with shared tool icons across clients, show the subagent wait in place without handoff markers, and guide users after HTTP 400 prompt-size rejections. (#165095, #165378, #159712) Thanks @shakkernerd, @sunlit-deng, and @RomneyDa.
- **Voice calls:** add per-call briefs, call reports, live steering, callbacks, and voicemail detection to voice calls. (#165716) Thanks @mbelinky and @jacobtomlinson.
- **Control UI subagents:** show when a chat is waiting on subagents, open the Subagents panel from the chat's subagent count and names, and surface child failures and attention in the parent chat. (#164190, #166142, #166224, #166312, #166347) Thanks @shakkernerd and @vincentkoc.
- **Control UI customization:** let themes customize Control UI branding, use agent avatars or unlocked Lobsterdex icons for browser tabs, choose favicon shapes, and add seven Lobsterdex character variants. (#168027, #165201, #166325, #167809, #159815) Thanks @vyctorbrzezowski, @jacobtomlinson, and @hannesrudolph.
- **Terminal:** add configurable terminal fonts with built-in Nerd Font icons. (#166180) Thanks @jacobtomlinson.
- **Control UI navigation:** copy file paths from chat tooltips, pin agents directly from the switcher, make agent menus easier to scan, let users hide a session's branch PR row, and simplify optional question cards. (#164431, #165640, #165641, #166366, #168040, #150587) Thanks @jjjhenriksen, @vyctorbrzezowski, and @RomneyDa.
- **Skill Workshop:** show each skill with its changes, files, and history in one view. (#167440) Thanks @obviyus.
- **iOS:** connect the Gateway through native Cloudflare Access with app-owned browser and profile admission. (#147244, #147238) Thanks @vincentkoc, @Solvely-Colin, @Patrick-Erichsen, and @RomneyDa.
- **ClawHub:** browse and search ClawHub plugins and skills through the Gateway. (#163459) Thanks @jacobtomlinson.
- **Plugins:** let plugin tools deliver their finished reply without a second model turn, allow every loaded plugin its own queue modes, and select plugin-owned self-hosted executors. (#164558, #165965, #161609) Thanks @mbelinky and @sjf-oa.
- **Plugins:** pick the newest official plugin release the running OpenClaw supports and keep a working plugin when its matching release is unavailable. (#167485, #166588) Thanks @charlesBerg-ux.
- **Workboard:** pin, unpin, and delete boards from the sidebar, keep card notes selectable with a roomier notes editor, and keep column edits on rename. (#165020, #166625, #167348) Thanks @rogeriochaves, @llmsommelier, and @obviyus.
- **GitHub:** support asynchronous PR merges and compact the Publish PR card. (#163579, #165219) Thanks @vincentkoc.
- **Session performance:** serve per-turn session projections, entry reads, authority checks, compaction reads, and first-turn headers from worker threads to keep the Gateway main thread responsive. (#164779, #165028, #165205, #165406, #165401, #165581, #165819, #165027, #165367, #164747, #164989, #165073)
- **Session list and search performance:** stop Control UI list refresh bursts, bound list materialization, scope and prefilter transcript search, speed up cold suggestion lists, and retain previews across metadata edits. (#162280, #164764, #165080, #165089, #165570, #165150, #165834, #165096, #165932, #166635, #165887, #167909) Thanks @RomneyDa.
- **Turn database performance:** reduce SQLite statements, worker requests, and redundant reads per chat turn, batch snapshot and metadata queries, and reuse admitted schema and reader facts. (#164490, #166041, #166053, #166054, #166074, #166102, #166541, #166503, #166519, #166698, #167278, #167686, #167685, #167690, #167932, #167046, #166966, #166925, #167291, #166703, #167029, #165360, #165193)
- **State database performance:** speed up cold agent database admission and crash recovery, reuse integrity checks on clean restarts, stop reopening executors when switching agents, and release leases promptly during shutdown. (#164846, #165070, #165186, #165293, #165680, #166112, #165993, #167001, #168008, #164788, #164962, #165085, #165910, #166822) Thanks @RomneyDa.
- **Maintenance and history performance:** lighten idle and manual maintenance, lifecycle cleanup, archive, reset, and trajectory retention work, and speed history reads with nested tool activity. (#164851, #164755, #164938, #165081, #165082, #165254, #165090, #165074, #165274, #165678, #167268, #167070, #167549, #167643, #168013, #167449, #165221, #165181, #165093, #165086, #165111, #165377, #165495, #165909, #166069, #166631, #166778, #166804, #167434, #167915, #167959) Thanks @Patrick-Erichsen, @obviyus, and @DonnieFi.
- **Agent performance:** reduce retained heap, per-turn prompt allocations, transcript reads, and main-thread SQLite work during runs, preserve prompt cache across instruction refreshes and image cleanup, and reuse verified history images across turns. (#167454, #167922, #167682, #167547, #167698, #167613, #167355, #165142, #165404, #167282, #164685, #165012, #165410, #165370, #165552) Thanks @RomneyDa and @DonnieFi.
- **Model catalog performance:** speed up multi-provider catalog refreshes, avoid discovery on cold account reads and unrequested plugin catalogs, retain account catalogs during refresh, and reduce catalog worker heap. (#167101, #167334, #167359, #167009, #167008, #166007, #166434, #165911, #165708, #165547, #165034, #165491, #166602) Thanks @vincentkoc, @obviyus, and @xialonglee.
- **Subagent and ACP performance:** reduce startup registry restore latency, move bulk registry work off the Gateway thread, batch registry cohorts, and avoid full-session scans when resuming ACP sessions. (#165942, #165684, #165098, #167684, #167010, #167372, #165550, #167086) Thanks @RomneyDa.
- **Update performance:** updates stop rehashing unchanged runtime files while retaining the running updater and publishing the candidate, reducing update I/O. (#168958)

### Fixes

- **Gateway memory:** Gateway startup memory is back in the 2026.10.4 range: on a cold start, memory indexes open on the first search or turn that needs them instead of all at once after readiness, catalog, SQLite broker, and transcript worker heaps are capped, and memory search workers used by `openclaw agent` and memory recall retain less memory. (#174718, #172392) Thanks @RomneyDa, @xialonglee, @DonnieFi, @jjjhenriksen, and @jacobtomlinson.
- **Session data safety:** preserve user input and committed facts through restart recovery and failed publication, keep accepted locked appends settled before releasing the writer, and fence cross-store worker commits. (#165502, #167570, #165894, #168018, #168029)
- **Security boundaries:** fail closed when the process census cannot prove a temp root is unused, reject incompatible worker prompt contexts before launch, keep browser pairing credentials out of opener process arguments, expand `file:` SQLite targets before the exec live-state guard checks them, require approval before setup diagnostics repair state, run plugin `before_tool_call` hooks under `openclaw agent exec`, and reject oversized local workspace document saves. (#165079, #166268, #170289, #170253, #171960, #159333, #171938) Thanks @Dhotei, @vincentkoc, @obviyus, @eleqtrizit, @joseairosa, and @RomneyDa.
- **Tool Search provenance:** network results reached through Tool Search and Code Mode now taint the turn the same way direct network tool calls do. (#169560) Thanks @RomneyDa and @farisoftfry.
- **Command owners:** Discord and Slack honor `discord:user:<id>` and `slack:user:<id>` command owners for voice access and exec approvals, so owner-derived approvers are no longer dropped. (#172385, #172424) Thanks @yetval and @RomneyDa.
- **Session authority and privacy:** check caller authority before restoring cold history for compaction, keep authority current after committed session changes, reject malformed hook session keys, and hide pre-session edits to branch-added files from uncommitted diffs. (#166506, #167875, #166406, #164671) Thanks @yetval, @obviyus, and @RomneyDa.
- **Session cleanup:** remove leftover GitHub publication receipts and managed worktrees on cleanup and Incognito reset, and publish conversation, plugin-state, sandbox, worktree, and GitHub authority receipts. (#163833, #165185, #167731, #168025) Thanks @jayzhou2309 and @obviyus.
- **Turn reliability:** stop unrelated session writes from failing Codex and chat turns, keep Knowledge calls working while a session publishes its bookkeeping, and stop group turns from failing silently when history cannot load. (#165196, #166010, #166758, #135253, #166059) Thanks @jalehman, @izshaw, @DonnieFi, and @RomneyDa.
- **Restart recovery:** settle sessions left running after restart, avoid duplicate recovery after ordinary interruptions, keep recovery reconciliation on the host, and preserve database admission, custody, and hook retries for post-restart and first-session turns. (#165174, #154175, #166395, #166867, #166609, #167629, #165859, #165738, #166891) Thanks @Grynn and @RomneyDa.
- **Restart resume:** resume Control UI turns interrupted by a Gateway restart, keep a new session's first message when the Gateway restarts during worktree preparation, resume tasks after deferred startup recovery, resume turns interrupted while a session was still preparing, return interrupted subagents to their parent with their history (including quiet children of parents created by the `agent` RPC), resume automation and node-hosted turns instead of failing with an active turn claim, keep streamed text from a killed native worker in history, keep symlinked state directories working after restart, and settle restart intents and update reports in workers. (#165333, #166042, #165184, #168650, #169144, #170162, #170161, #170223, #168361, #167572, #165069, #165670, #175278) Thanks @obviyus, @RomneyDa, @DonnieFi, and @jjjhenriksen.
- **Main session restarts:** main sessions that were mid-turn during a Gateway restart resume on the next boot, and interrupted completion replies are delivered instead of lost. (#168651, #167814, #167766) Thanks @goutamadwant, @marknilsn, @DonnieFi, and @RomneyDa.
- **Agent and session creation:** fix cold shared-store and fresh-store creation failures and agent database recovery after orphan session window repair. (#167257, #167515, #167366, #166860) Thanks @RomneyDa.
- **Windows and path aliases:** canonicalize the shared state database path and compare link-stable identities so Windows and POSIX path aliases share one store and channel lifecycle. (#166755, #166062, #166787) Thanks @taihartman and @RomneyDa.
- **SQLite deadlocks and stalls:** prevent writer-reader admission and cleanup deadlocks, keep committed reads responsive during writer stalls, stop writer-held transcript reads waiting on readers, and keep foreign changes visible during refresh. (#167662, #167446, #167901, #167687, #166331, #165898, #167332) Thanks @sasan1200, @vincentkoc, and @DonnieFi.
- **State and updates:** stop leftover update-check caches and completed recovery history from blocking updates, and give one actionable answer for state written by a newer release while keeping status reads read-only. (#165461, #167358, #166963) Thanks @RomneyDa.
- **Startup and integrity checks:** return pending admission during startup index repair, retain integrity proof through startup and idle retirement, and stop healthy starts logging SQLite lock-wait warnings. (#165323, #166972, #166412, #167754, #166253, #168310) Thanks @RomneyDa.
- **Session search and history:** search transcripts for configured ACP owners, stop false indexing notices, keep search parallel with accurate indexing, preserve dot-prefixed checkpoint references and CLI history after interrupted progress replies, keep canonical transcript owners distinct, and refresh empty previews after cold history is restored. (#164897, #165879, #167174, #167163, #164920, #166658, #164512, #168397) Thanks @sunlit-deng, @obviyus, @techdou, @shakkernerd, @VACInc, and @RomneyDa.
- **Session display:** stop Markdown link markup leaking into previews, show correct context windows for plugin-discovered models, log failed wakes with their target, keep owning agents in lifecycle commands and trajectory export, and decode UTF-16 BOM text files. (#166653, #159731, #166458, #166415, #165777) Thanks @jayzhou2309, @obviyus, @rudycelekli, @Yigtwxx, @yetval, and @RomneyDa.
- **Startup ordering:** stop agents with small transcript rebuild backlogs from waiting behind larger agents at startup. (#165207)
- **Data integrity:** reject malformed UTF-8 backup manifests before verification, preserve malformed UTF-8 and trailing-space connect targets, and keep accents in named legacy text attachments. (#167389, #167399, #167116, #165405) Thanks @Alix-007, @obviyus, and @Chessing234.
- **Turn lifecycle:** publish one terminal event per stopped or runtime-lost run, settle terminal runs before session mutations, keep queued message run ids and replies aligned, prevent