* feat(bedrock): add Bedrock Mantle (OpenAI-compatible) provider
New amazon-bedrock-mantle extension that provides auto-discovery and
authentication for Amazon Bedrock Mantle endpoints.
Mantle (bedrock-mantle.<region>.api.aws) is Amazon Bedrock's OpenAI-
compatible API surface, separate from the existing bedrock-runtime
(ConverseStream) endpoint. It has its own model catalog including
models not available via ConverseStream (e.g. openai.gpt-oss-120b,
mistral.devstral-2-123b).
Extension structure:
- discovery.ts: Model discovery via GET /v1/models (OpenAI format),
bearer token resolution, implicit provider configuration
- register.sync.runtime.ts: Provider registration with catalog,
error classification (rate limits, context overflow)
- openclaw.plugin.json: Plugin manifest, enabledByDefault
Auth support:
- Long-lived Bedrock API key (AWS_BEARER_TOKEN_BEDROCK env var)
created from the AWS Console → used directly as Bearer token
- Pre-generated SigV4-derived tokens (via aws-bedrock-token-generator)
set in AWS_BEARER_TOKEN_BEDROCK → works transparently
Provider config (auto-resolved when AWS_BEARER_TOKEN_BEDROCK is set):
api: "openai-completions"
baseUrl: "https://bedrock-mantle.<region>.api.aws/v1"
auth: "api-key" (bearer token)
Available in 12 regions: us-east-1, us-east-2, us-west-2,
ap-northeast-1, ap-south-1, ap-southeast-3, eu-central-1,
eu-west-1, eu-west-2, eu-south-1, eu-north-1, sa-east-1
Tests: 15 passing (13 discovery + 2 plugin registration)
* chore(bedrock): clarify mantle bearer auth scope
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* feat(bedrock): add inference profile discovery and region injection
Inference profiles (cross-region and application) work with ConverseStream
but require the SDK client region to match the profile region. Without
this, users get "The provided model identifier is invalid" errors when
using cross-region profiles like us.anthropic.claude-sonnet-4-6.
Changes:
1. Inference profile discovery (discovery.ts):
- Call ListInferenceProfiles alongside ListFoundationModels (parallel)
- Inference profiles INHERIT capabilities from their underlying
foundation model (modalities, reasoning, context window, cost)
- resolveBaseModelId() maps profile → foundation model:
"us.anthropic.claude-sonnet-4-6" → "anthropic.claude-sonnet-4-6"
Application ARNs → extract model ID from models[].modelArn
- Graceful degradation if IAM lacks bedrock:ListInferenceProfiles
- Provider filter applies to profiles via underlying model ARNs
2. Region injection (register.sync.runtime.ts):
- Extract region from provider baseUrl or bedrockDiscovery.region
- Pass through to pi-ai options.region in wrapStreamFn
- Ensures SDK client connects to correct regional endpoint
3. Inference profile model detection (anthropic-family-cache-semantics.ts):
- isAnthropicBedrockModel() now recognizes application inference
profile ARNs (arn:aws:bedrock:...:application-inference-profile/*)
4. Tests (discovery.test.ts):
- New: inference profile inheritance test (4 models: 1 foundation +
3 profiles, verifies capability inheritance, inactive filtering)
- New: graceful AccessDeniedException handling test
- Updated: all existing tests for dual-API discovery pattern
Fixes#55642
* fix(bedrock): preserve inference profile model lookup
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
resolveTrackedFacadePluginId triggers config loading (plugin auto-enable,
channel discovery) which can re-enter loadBundledPluginPublicSurfaceModuleSync
for the same module. Because the sentinel was still empty at that point,
re-entrant callers saw undefined exports (e.g. shouldNormalizeGoogleProviderConfig).
Move Object.assign(sentinel, loaded) before the plugin ID resolution so any
re-entrant lookup through the cached sentinel finds the real exports.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The cron tool schema used type arrays (['string','null']), the 'not'
keyword, and 'const' — all unsupported by the OpenAPI 3.0 subset that
Gemini-backed providers (e.g. GitHub Copilot) enforce. This caused
HTTP 400 for every request when cron was enabled.
Replace type arrays with scalar types, remove not/const from
CronFailureAlertSchema, and add 'not' to the Gemini unsupported
keywords list as defense-in-depth.
Fixes#61206
* fix(acpx): store agent session ID when session/load fails
When an ACP agent (e.g. Gemini CLI) rejects the acpx-generated session
ID via session/load and falls back to session/new, the agent-returned
session ID was previously discarded. This caused identity stuck at
pending forever, multi-turn failures, lost completion events, and
persistent reconcile warnings.
- Parse ACP protocol stream in runTurn() to capture agent session IDs
- Flip resolveRuntimeResumeSessionId() to prefer agentSessionId
- Add createIdentityFromHandleEvent() for handle-sourced identity
- Layer handle event identity before status in reconcile
- Add regression tests for load fallback and restart resume
Closes#52182
* ACPX: prefer decoded session ids
* ACPX: refresh runtime handle state from status
---------
Co-authored-by: Wesley <imwyvern@users.noreply.github.com>
* fix(infra): windows-task-restart fallback to startup entry when schtasks task is unregistered
* fix code style problem
* use /min for startup fallback and assert schtasks pre-check in test
* fix: windows restart fallback when scheduled task is unregistered (#58943) (thanks @imechZhangLY)
---------
Co-authored-by: Luyao Zhang <zhangluyao@microsoft.com>
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* plugins: include resolved workspaceDir in provider hook cache keys
resolveProviderPluginsForHooks, resolveProviderPluginsForCatalogHooks, and
resolveProviderRuntimePlugin used the raw params.workspaceDir for cache keys
and plugin-id discovery while resolvePluginProviders already fell back to
the active registry workspace. Resolve workspaceDir once at the top of each
function so cache keys, candidate filtering, and loading all use the same
workspace root.
* fix(plugins): inherit runtime workspace for snapshot loads
* test(gateway): stub runtime registry seam
* fix(plugins): restore workspace fallback after rebase
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(bedrock): stop injecting fake apiKey marker for aws-sdk auth when no env vars exist
When the Bedrock provider uses auth: "aws-sdk" and no AWS environment
variables are set (EC2 instance roles, ECS task roles, etc.),
resolveAwsSdkApiKeyVarName() fell back to "AWS_PROFILE" unconditionally.
This string was injected as apiKey in the provider config during
normalisation, which poisoned the downstream auth resolver — it treated
the marker as a literal key and failed with "No API key found".
The fix:
- resolveAwsSdkApiKeyVarName() now returns undefined (not "AWS_PROFILE")
when no AWS env vars are present
- resolveBedrockConfigApiKey() (extension) gets the same fix
- resolveMissingProviderApiKey() guards both the providerApiKeyResolver
and direct aws-sdk branches: if the resolver returns nothing, the
provider config is returned unchanged (no apiKey injected)
- The aws-sdk credential chain then resolves credentials at request time
via IMDS/ECS task role/etc. as intended
When AWS env vars ARE present (AWS_ACCESS_KEY_ID, AWS_PROFILE,
AWS_BEARER_TOKEN_BEDROCK), the marker is still injected correctly.
Closes#49891Closes#50699Fixes#54274
* test(bedrock): update resolveBedrockConfigApiKey test for undefined return on empty env
The test previously expected "AWS_PROFILE" when no env vars are set.
Now expects undefined (matching the fix), and adds a separate assertion
that AWS_PROFILE is returned when the env var is actually present.
* fix(bedrock): lock aws-sdk env marker behavior
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(google): restore proxy-safe image generation (#59873)
* fix(ssrf): preserve transport policy without pinned dns
* fix(ssrf): use undici fetch for dispatcher requests
* fix(ssrf): type dispatcher fetch path
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(telegram): enable voice-note transcription in DMs
The preflight transcription condition only triggered for group chats
(isGroup && requireMention), so voice notes sent in direct messages
were never transcribed -- they arrived as raw <media:audio> placeholders.
This regression was introduced when the Telegram channel was moved from
src/telegram/ to extensions/telegram/, losing the fix from c15385fc94.
Widen the condition to fire whenever there is audio and no accompanying
text, regardless of chat type. Group-specific guards (requireMention,
disableAudioPreflight, senderAllowedForAudioPreflight) still apply
only in group contexts.
* fix: restore Telegram DM voice-note transcription (#61008) (thanks @manueltarouca)
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix: ensure bypassPermissions on custom CLI backend args
When users override cliBackends.claude-cli.args (e.g. to add --verbose
or change --output-format), the override array replaces the default
entirely. The normalization step only re-added --permission-mode
bypassPermissions when the legacy --dangerously-skip-permissions flag
was present — if neither flag existed, it did nothing.
This causes cron and heartbeat runs to silently fail with "exec denied:
Cron runs cannot wait for interactive exec approval" because the CLI
subprocess launches in interactive permission mode.
Fix: always inject --permission-mode bypassPermissions when no explicit
permission-mode flag is found in the resolved args, regardless of
whether the legacy flag was present.
* test(anthropic): add claude-cli permission normalization coverage
* fix(test-utils): include video generation providers
* fix: preserve claude-cli bypassPermissions on custom args (#61114) (thanks @cathrynlavery)
---------
Co-authored-by: Shadow <hi@shadowing.dev>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(telegram): trim menu descriptions before dropping commands
* fix: note Telegram command menu trimming (#61129) (thanks @neeravmakwana)
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* UI: improve mobile chat layout
* change .chat-group-messages min-width: from 604 to 602
* UI: fix chat-group-messages overflow in split-view and mobile layouts
* UI: revert chat.css import order in styles.css and components.css
* UI: simplify mobile chat layout overrides in grouped.css
* ui: move .chat and .chat-thread styles to chat/layout.css
* fix: document mobile chat layout improvements
* fix: improve narrow mobile chat width
---------
Co-authored-by: Altay <altay@uinaf.dev>
Add PR limits section explaining:
- 10 open PRs per author cap
- r: too-many-prs label auto-close mechanism
- How to get exception via #clawtributors Discord
Fixes: #38283
* fix(agents): prefer completion wake over polling
* fix(changelog): note completion wake guidance
* fix(agents): qualify quiet exec completion wake
* fix(agents): qualify disabled exec completion wake
* fix(agents): split process polling from control actions
* fix(cron): suppress NO_REPLY sentinel in direct delivery path
* fix: set deliveryAttempted on filtered NO_REPLY to prevent timer fallback
* fix: mark silent NO_REPLY direct deliveries as delivered
* fix(cron): unify silent direct delivery handling
* fix: suppress NO_REPLY direct cron leaks (#45737) (thanks @openperf)
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix(agents): handle LiveSessionModelSwitchError in subagent execution
Add retry loop for cross-provider model switches in the subagent
command path, mirroring the existing logic in agent-runner-execution.ts.
- Wrap runWithModelFallback in a while(true) loop inside agentCommandInternal
- Catch LiveSessionModelSwitchError and update provider, model,
fallbackProvider, fallbackModel, providerForAuthProfileValidation,
sessionEntry.authProfileOverride, and storedModelOverride before retrying
- Guard storedModelOverride update: only set when the model genuinely
changed (compared before mutation) or a session override already existed
- Reset lifecycleEnded flag so the retried iteration can emit lifecycle events
- Add comprehensive tests covering retry success, error propagation,
lifecycle reset, auth-profile forwarding, and fallback override state
Fixes#57998
* fix(agents): include provider change in storedModelOverride guard
* fix(agents): validate allowlist and clear stale compaction count on live model switch
* fix(agents): remove broken allowlist guard on live model switch
* fix(agents): address security review — bound retry loop, validate allowlist, redact error in lifecycle events
* fix(agents): restore error observability in lifecycle events using err.message
* fix(agents): sanitize log inputs and shallow-copy sessionEntry on live model switch
* fix(agents): enforce allowlist on empty set and sanitize error message
* fix: handle subagent live model switches (#58178) (thanks @openperf)
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix(agents): classify generic provider errors for failover
Anthropic returns bare 'An unknown error occurred' during API instability
and OpenRouter wraps upstream failures as 'Provider returned error'. Neither
message was recognized by the failover classifier, so the error surfaced
directly to users instead of triggering the configured fallback chain.
Add both patterns to the serverError classifier so they are classified as
transient server errors (timeout) and trigger model failover.
Closes#49706Closes#45834
* fix(agents): scope unknown-error failover by provider
* docs(changelog): note provider-scoped unknown-error failover
---------
Co-authored-by: Aaron Zhu <aaron@Aarons-MacBook-Air.local>
Co-authored-by: Altay <altay@uinaf.dev>
* fix(cli): route skills list output to stdout when --json is active
runSkillsAction used defaultRuntime.log() which goes through console.log.
The --json preAction hook calls routeLogsToStderr(), redirecting console.log
to stderr. Switch to defaultRuntime.writeStdout() which writes directly to
process.stdout, consistent with how other --json commands (e.g. skills search)
already emit their output.
Fixes#57599
* test(cli): add skills JSON stdout regression coverage
* test(cli): refine skills CLI stream coverage
* fix(cli): add changelog entry for skills JSON stdout fix
---------
Co-authored-by: Aftabbs <aftabbs.wwe@gmail.com>
* fix(google-gemini-cli-auth): fix Gemini CLI OAuth failures on Windows
Two issues prevented Gemini CLI OAuth from working on Windows:
1. resolveGeminiCliDirs: the first candidate `dirname(dirname(resolvedPath))`
can resolve to an unrelated ancestor directory (e.g. the nvm root
`C:\Users\<user>\AppData\Local\nvm`) when gemini is installed via nvm.
The subsequent `findFile` recursive search (depth 10) then picks up an
`oauth2.js` from a completely different package (e.g.
`discord-api-types/payloads/v10/oauth2.js`), which naturally does not
contain Google OAuth credentials, causing silent extraction failure.
Fix: validate candidate directories before including them — only keep
candidates that contain a `package.json` or a `node_modules/@google/
gemini-cli-core` subdirectory.
2. resolvePlatform: returns "WINDOWS" on win32, but Google's loadCodeAssist
API rejects it as an invalid Platform enum value (400 INVALID_ARGUMENT),
just like it rejects "LINUX".
Fix: use "PLATFORM_UNSPECIFIED" for all non-macOS platforms.
* test(google-gemini-cli-auth): keep oauth regressions portable
* chore(changelog): add google gemini cli auth fix note
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
The YAML parser's outer loop was exiting the tasks block when it
encountered 'interval:' or 'prompt:' lines, causing only the first
task to be parsed. Added isTaskField check to skip those lines.
Fixes: #3034790131
- Fix: Pass startedAt into resolveHeartbeatRunPrompt
- Fix: Return proper object instead of null for no-tasks-due
- Fix: Add early return when prompt is null
- Fix: Persist timestamps on successful exits
- Fix YAML parsing to capture interval:/prompt: before breaking
- Record task timestamps AFTER successful execution (not before)
- Initialize task state on first run (handle undefined session)
- Skip API call when no tasks due (return null)
- Use startedAt consistently for due-task filtering
Fixes: #3030568439, #3033833124, #3030570872, #3030568408, #3030570872, #3035434022, #3035434368
The heartbeat task batching feature uses heartbeatTaskState to track
last run times for periodic tasks, but this property was missing
from the SessionEntry type, causing TypeScript compilation errors.
- Add parseHeartbeatTasks() to parse YAML-like task definitions
- Add isTaskDue() to check if task interval has elapsed
- Add heartbeatTaskState to session store for tracking last run times
- Modify resolveHeartbeatRunPrompt to build batched prompts for due tasks
- Update task last run times after successful heartbeat execution
Implements openclaw#29570
* fix(cron): prevent agent default model from overriding cron payload model (#58065)
When a cron job specifies a model override via the Advanced settings,
runWithModelFallback could silently fall back to the agent's configured
primary model. This happened because fallbacksOverride was undefined
when neither payload.fallbacks nor per-agent fallbacks were configured,
causing resolveFallbackCandidates to append the agent primary as a
last-resort candidate. A transient failure on the cron-selected model
(rate limit, model-not-found, etc.) would then succeed on the agent
default, making it appear as if the override was ignored entirely.
Fix: when the cron payload carries an explicit model override, ensure
fallbacksOverride is always a defined array (empty when no fallbacks
are configured) so the agent primary is never silently appended.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test: use stricter toEqual([]) assertion for fallbacksOverride
Replace toBeDefined() + toBeInstanceOf(Array) with toEqual([])
to catch regressions where the array unexpectedly gains entries.
Addresses review feedback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* fix: preserve cron override fallback semantics (#58294)
---------
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
When the gateway client reconnects using a stored device token, it was
defaulting to ["operator.admin"] scopes instead of preserving the
previously authorized scopes from the stored token. This caused the
operator device token to be regenerated without operator.read scope,
breaking status/probe/health commands.
This fix:
1. Loads the stored scopes along with the stored token in selectConnectAuth
2. Uses the stored scopes when reconnecting with a valid device token
3. Falls back to explicitly requested scopes or default admin-only scope
when no stored scopes exist
Fixes#46000
- Remove redundant name === 'MiniMax-M*' condition (already matched by startsWith)
- Use !== undefined guard instead of falsy check in deriveWindowLabelFromTimestamps
- Pass chatRemains directly to deriveWindowLabel when available
- Remove JSDoc comment style to match codebase conventions
- Pick the chat model entry (MiniMax-M*) from model_remains instead of using the first BFS candidate, which could be a speech/video/image model with total_count=0.
- Derive window label from start_time/end_time timestamps when window_hours/window_minutes fields are absent; fixes the hardcoded 5h default for 4h windows.
- Include model name in plan label so users can distinguish free-tier coding-plan quota from paid API balance.
Closes#52335
* fix(google-cli): restore gemini json reporting
* fix(google-cli): fall back to stats when usage is empty
* fix(changelog): note gemini cli cache reporting
MiniMax M2.7 returns reasoning_content in OpenAI-style delta chunks
({delta: {content: "", reasoning_content: "..."}}) when thinking is
active, rather than native Anthropic thinking block SSE events. Pi-ai's
Anthropic provider does not handle this format, causing the model's
internal reasoning to appear as visible chat output.
Add createMinimaxThinkingDisabledWrapper that injects
thinking: {type: "disabled"} into the outgoing payload for any MiniMax
anthropic-messages request where thinking is not already explicitly
configured, preventing the provider from generating reasoning_content
deltas during streaming.
Fixes#55739
The VLM image analysis fetch had no timeout, causing sessions to hang
indefinitely when the MiniMax API is slow or unresponsive. Other
vision/model API calls in the codebase already use timeouts. Adds
AbortSignal.timeout(60_000) consistent with image upload workloads.
Fixes#54139
When cron tasks or subagents use browser automation, the browser
processes were not cleaned up after the task completed. This caused
orphaned Chrome processes (PPID=1) to accumulate over time.
Root cause: closeTrackedBrowserTabsForSessions was only called during
session-reset/session-delete (via ensureSessionRuntimeCleanup), but
isolated cron runs and subagent completions never triggered these paths.
Fix: Add browser tab cleanup in two places:
1. server-cron.ts: wrap runCronIsolatedAgentTurn in try/finally to
ensure browser tabs are cleaned up after every cron run.
2. subagent-registry-lifecycle.ts: call closeTrackedBrowserTabsForSessions
when a subagent run completes, before the announce cleanup flow.
Both cleanup calls are best-effort (caught errors) so they never mask
the actual task result or break the completion flow.
Fixes#60104
When `browser stop` is called for an `attachOnly` or remote CDP
profile, `profileState.running` is null (no process was launched), so
`stopRunningBrowser()` returned early without closing the Playwright
CDP connection. This left emulation overrides (prefers-color-scheme,
viewport, etc.) permanently applied until a full gateway restart.
Now call `closePlaywrightBrowserConnectionForProfile()` before
returning for attachOnly and remote CDP profiles, matching the cleanup
behavior already present in `resetProfile()`. Regular profiles that
were never started still return `{ stopped: false }`.
Fixes#60095
Add native MiniMax Search integration via their Coding Plan search API
(POST /v1/coding_plan/search). This brings MiniMax in line with Brave,
Kimi, Grok, Gemini, and other providers that already have bundled web
search support.
- Implement WebSearchProviderPlugin with caching, credential resolution,
and trusted endpoint wrapping
- Support both global (api.minimax.io) and CN (api.minimaxi.com)
endpoints, inferred from explicit region config, model provider base
URL, or minimax-portal OAuth base URL
- Prefer MINIMAX_CODE_PLAN_KEY over MINIMAX_API_KEY in credential
fallback, matching existing repo precedence
- Accept SecretRef objects for webSearch.apiKey (type: [string, object])
- Register in bundled registry, provider-id compat map, and fast-path
plugin id list with full alignment test coverage
- Add unit tests for endpoint/region resolution and edge cases
Closes#47927
Related #11399
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
When the liveModelSwitchPending flag is set but the current model already
matches the persisted selection (e.g. the switch was applied as an override
and the current attempt is already using the new model), the flag is now
consumed eagerly via a fire-and-forget clearLiveModelSwitchPending() call.
Without this, the stale flag could persist across fallback iterations and
later cause a spurious LiveSessionModelSwitchError when the model rotates
to a fallback candidate that differs from the persisted selection.
Also expands JSDoc on shouldSwitchToLiveModel to document the stale-flag
clearing and deferral semantics.
Replace the ambiguous comparison-based approach (hasDifferentLiveSessionModelSelection
+ in-memory map EMBEDDED_RUN_MODEL_SWITCH_REQUESTS) with a persisted
`liveModelSwitchPending` flag on SessionEntry.
The root cause: the in-memory map was never populated in production because
requestLiveSessionModelSwitch() was removed in commit 622b91d04e and replaced
with refreshQueuedFollowupSession(). This left the comparison-based detection
as the only path, which could not distinguish user-initiated model switches
(via /model command) from system-initiated fallback rotations.
The fix:
- Add `liveModelSwitchPending?: boolean` to SessionEntry (persisted)
- Set the flag to true ONLY when /model command applies a model override
- New `shouldSwitchToLiveModel()` checks the flag + model mismatch together
- New `clearLiveModelSwitchPending()` resets the flag after consumption
- Replace throw-site logic in run.ts to use the new flag-based functions
- Remove orphaned resolveCurrentLiveSelection helper
Only the /model command sets this flag, so system-initiated fallback rotations
are never mistaken for user-initiated model switches. This restores the
live-switch-during-active-run feature that was accidentally broken.
Fixes#57857, #57760, #58137
When auth choice explicitly sets a preferred provider (e.g., volcengine-api-key or byteplus-api-key), the model picker should always filter by that provider. Previously, it relied on providerIds.includes(preferredProvider), which could be false if the catalog hadn't loaded that provider's models yet due to a race condition between auth choice setup and catalog loading.
This ensures that selecting a provider via auth choice consistently filters the model list to only that provider's models, rather than showing all providers.
MiniMax's usage_percent / usagePercent fields report the *remaining* quota
as a percentage, not the consumed quota. When count fields (prompt_limit /
prompt_remain) are also present, fromCounts already computed the correct
usedPercent and the inverted value was silently ignored. But when only
usage_percent is returned (no count fields), the code treated it as a
used-percent and passed it through unchanged, causing the menu bar to show
"2% left" instead of "98% left".
Move usage_percent and usagePercent from PERCENT_KEYS to a new
REMAINING_PERCENT_KEYS array. deriveUsedPercent now inverts remaining-percent
values to obtain usedPercent, matching the behaviour already validated by the
existing "prefers count-based usage when percent looks inverted" test. Count-
based fromCounts still takes priority over both key groups.
Fixes#60193
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
When using custom providers like LM Studio, Ollama, or DashScope,
token counts in /status show as 0 because the agent meta store
does not always have usage data populated for these providers.
Fix: set includeTranscriptUsage: true in both /status command and
the session_status tool. This enables the existing fallback path
that reads usage from the session transcript JSONL file when the
meta store has zero/missing token counts.
The merge logic already guards against overwriting valid data:
- totalTokens: only updated when zero or transcript value is larger
- inputTokens/outputTokens: only filled when zero/missing
- model/contextTokens: only filled when missing
Fixes#54995
When sessions report an already-qualified model id (e.g. ollama/qwen3:30b),
resolveServerChatModelValue was re-qualifying it using modelProvider,
producing incorrect values like openai-codex/qwen3:30b.
Preserve already-qualified model refs as-is before applying provider prefix.
Adds test coverage for qualified model preservation.
Fixes#49839
The Kimi Coding plugin registers with provider ID `kimi` and default
model ID `kimi-code`, making the correct model ref `kimi/kimi-code`.
The docs incorrectly showed `kimi-coding/k2p5` as the provider/model
ref. This is confusing because `kimi-coding` is only a plugin alias,
not the actual provider ID used in config.
Updated all references in:
- docs/concepts/model-providers.md
- docs/providers/moonshot.md
- docs/zh-CN/concepts/model-providers.md
- docs/zh-CN/providers/moonshot.md
Switch DEFAULT_MINIMAX_TTS_BASE_URL from api.minimaxi.com (CN) to
api.minimax.io (global) so international API keys work out of the box.
Add vol and pitch to resolveTalkOverrides for parity with resolveTalkConfig.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add MiniMax as a fourth TTS provider alongside OpenAI, ElevenLabs, and
Microsoft. Registers a SpeechProviderPlugin in the existing minimax
extension with config resolution, directive parsing, and Talk Mode
support. Hex-encoded audio response from the T2A v2 API is decoded to
MP3.
Closes#52720
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Remove DEFAULT_KIMI_MODEL (moonshot-v1-128k) and align resolveKimiModel
fallback to DEFAULT_KIMI_SEARCH_MODEL (kimi-k2.5). The legacy model
does not support the $web_search builtin_function tool, so env-var-only
users without a configured model would hit the original bug.
Config and Plugin SDK drift detection now compares SHA-256 hashes instead
of full JSON content. The .sha256 files (6 lines total) are tracked in git;
the full JSON baselines are gitignored and generated locally for inspection.
Same CI guarantee, zero repo churn on schema changes.
CLI connections with valid shared auth (token/password) now bypass device
pairing, fixing the chicken-and-egg problem where Docker CLI commands fail
with 'pairing required' (1008) despite sharing the gateway's network
namespace and auth token.
The existing shouldSkipBackendSelfPairing only matched gateway-client/backend
mode. CLI connections use cli/cli mode and were excluded. Additionally,
isLocalDirectRequest produces false negatives in Docker (host networking,
network_mode sharing) even when remoteAddress is 127.0.0.1, so CLI connections
with valid shared auth skip the locality check entirely — the token is the
trust anchor.
Closes#55067
Related: #12210, #23471, #30740
* fix(matrix): migrate room allow aliases to enabled
* test(matrix): keep migration coverage on the channel seam
* chore(config): refresh baselines after matrix alias cleanup
* memory-core: add dreaming promotion flow with weighted thresholds
* docs(memory): mark dreaming as experimental
* memory-core: address dreaming promotion review feedback
* memory-core: harden short-term promotion concurrency
* acpx: make abort-process test timer-independent
* memory-core: simplify dreaming config with mode presets
* memory-core: add /dreaming command and tighten recall tracking
* ui: add Dreams tab with sleeping lobster animation
Adds a new Dreams tab to the gateway UI under the Agent group.
The tab is gated behind the memory-core dreaming config — it only
appears in the sidebar when dreaming.mode is not 'off'.
Features:
- Sleeping vector lobster with breathing animation
- Floating Z's, twinkling starfield, moon glow
- Rotating dream phrase bubble (17 whimsical phrases)
- Memory stats bar (short-term, long-term, promoted)
- Active/idle visual states
- 14 unit tests
* plugins: fix --json stdout pollution from hook runner log
The hook runner initialization message was using log.info() which
writes to stdout via console.log, breaking JSON.parse() in the
Docker smoke test for 'openclaw plugins list --json'. Downgrade to
log.debug() so it only appears when debugging is enabled.
* ui: keep Dreams tab visible when dreaming is off
* tests: fix contracts and stabilize extension shards
* memory-core: harden dreaming recall persistence and locking
* fix: stabilize dreaming PR gates (#60569) (thanks @vignesh07)
* test: fix rebase drift in telegram and plugin guards
* fix(daemon): preserve Windows Task Scheduler settings on reinstall and exit early on failed restart
* fix(daemon): add test coverage for Create/Change paths, fix early exit grace period
* fix(daemon): fix startup-fallback tests for new isRegisteredScheduledTask call
* fix(daemon): report early restart failure accurately
* fix: preserve Windows scheduled task restart/install behavior (#59335) (thanks @tmimmanuel)
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* fix: add enabledByDefault to groq and deepgram media plugin manifests
The groq and deepgram plugin manifests were missing the
enabledByDefault: true flag. Without this flag, both plugins are
treated as bundled-but-disabled-by-default, so resolveRuntimePluginRegistry
loads without them. When buildProviderRegistry later needs to resolve
audio providers, the active registry is used first (short-circuits
the compat path in resolvePluginCapabilityProviders), leaving groq
and deepgram absent from the registry.
This caused 'Media provider not available: groq' errors when users
configured tools.media.audio.models with groq or deepgram, even
with GROQ_API_KEY / DEEPGRAM_API_KEY set correctly.
The fix mirrors the pattern used by other audio/media-only providers
such as mistral, which already has enabledByDefault: true.
Fixes#59875
* fix: enable groq and deepgram bundled media providers by default (#59982) (thanks @yxjsxy)
---------
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
- Skip sensitive fields with a note directing users to openclaw config set
or the Web UI (WizardPrompter has no masked input)
- Clear number fields to undefined when input is empty instead of storing 0
- Allow clearing array fields to undefined via empty input
pruneProcessedHistoryImages was stripping image blocks from every
already-answered user turn on each run. Turn N sends image bytes → provider
caches the prefix. Turn N+1 replaces image with text marker → bytes diverge
at that message → cache miss from there onward.
Now only prune images older than 3 assistant turns. Recent history stays
byte-identical so the cached prefix survives, while legacy sessions with
persisted image payloads still get cleaned up.
Wire uiHints from plugin manifests into the TUI wizard so sandbox/tool
plugins get interactive config prompts during openclaw onboard (manual
flow) and openclaw configure --section plugins.
- Add setup.plugin-config.ts: discovers plugins with non-advanced uiHints,
generates type-aware prompts (enum→select, boolean→confirm, array→csv,
string/number→text) from jsonSchema + uiHints metadata.
- Onboard: new step after Skills, before Hooks (skipped in QuickStart).
Only shows plugins with unconfigured fields.
- Configure: new 'plugins' section in the section menu. Shows all
configurable plugins with configured/total field counts.
Closes#60030
* fix(cache): compact newest tool results first to preserve prompt cache prefix
compactExistingToolResultsInPlace iterated front-to-back, replacing the
oldest tool results with placeholders when context exceeded 75%. This
rewrote messages[k] for small k, invalidating the provider prompt cache
from that point onward on every subsequent turn.
Reverse the loop to compact newest-first. The cached prefix stays intact;
the tradeoff is the model loses recent tool output instead of old, which
is acceptable since this guard only fires as an emergency measure past
the 75% threshold.
* fix(cache): compact newest tool results first to preserve prompt cache prefix (#58036) Thanks @bcherny
---------
Co-authored-by: George Zhang <georgezhangtj97@gmail.com>
auth_permanent errors (e.g. API_KEY_INVALID) can be caused by transient
provider outages rather than genuinely revoked credentials. Previously
these used the same 5h-24h billing backoff, which left providers disabled
long after the upstream issue resolved.
Introduce separate authPermanentBackoffMinutes (default: 10) and
authPermanentMaxMinutes (default: 60) config options so auth_permanent
failures recover in minutes rather than hours.
Fixes#56838
Preserve the !author.bot || sender.isPluralKit guard when short-circuiting
wasMentioned on mentionedEveryone, so bot relay messages don't spuriously
trigger mention-gate logic. Add test coverage for the wasMentioned path.
Forward mediaReadFile and mediaAccess through the sendMessageDiscord shortcut
in sendDiscordComponentMessage, so local-file media works correctly when
falling back to classic Discord messages. Add test coverage.
* fix(telegram): enable HTML formatting for model switch messages
The model switch confirmation message was displaying raw Markdown
(**text**) instead of bold formatting because parse_mode was not set.
Changes:
- Add optional extra parameter to editMessageWithButtons for parse_mode
- Change format from Markdown ** to HTML <b> tags
- Pass parse_mode: 'HTML' when editing model switch message
Fixes the issue where model names appeared as **provider/model**
instead of bold text in Telegram.
* fix(telegram): escape HTML entities in model switch confirmation
Add defensive `escapeHtml` helper to sanitize `selection.provider`
and `selection.model` before interpolating them into the HTML
callback message. This prevents potential API rejection (HTTP 400)
if future provider or model names contain `<`, `>`, or `&`.
Addresses review feedback on unescaped HTML interpolation.
* test(telegram): cover HTML model switch confirmation
---------
Co-authored-by: Frank Yang <frank.ekn@gmail.com>
The implementation fix propagates the hydrated cfg to reactMessageDiscord
and removeReactionDiscord. Update test assertions to expect the cfg
property in the options argument using expect.objectContaining to handle
the dynamic session store path.
When extracting `reactMessageDiscord`, it defaulted to reading the raw config (which contains `SecretRef`s) if a hydrated `cfg` was omitted. We now pass the pre-resolved `cfg` context into the reaction options so the plugin SDK resolves the token via memory rather than the raw file.
resolveDiscordThreadStarter only checked content and embeds, returning
null for forwarded messages where the text lives in message_snapshots.
Add a local resolveStarterForwardedText helper that extracts text
directly from the message_snapshots array on the REST response object.
This avoids fragile type casts and keeps the change self-contained
within threading.ts.
Fixes#60129
Fixes#60158
MCP tools with parameter-free schemas may return truly empty objects
`{}` without a `type` field. The existing normalization handled
`{ type: "object" }` → `{ type: "object", properties: {} }` but
missed the truly empty case.
OpenAI gpt-5.4 rejects tool schemas without `type: "object"` and
`properties`, causing HTTP 400 errors:
```
Invalid schema for function 'flux-mcp__get_flux_instance':
In context=(), object schema missing properties.
```
This change catches empty schemas (no type, no properties, no unions)
before the final pass-through and converts them to the required format.
Added test case for parameter-free MCP tool schemas.
* feat(gateway): add skills.search and skills.detail RPC methods
Expose ClawHub search and detail capabilities through the Gateway protocol,
enabling desktop/web clients to browse and inspect skills from the registry.
New RPCs:
- skills.search: search ClawHub skills by query with optional limit
- skills.detail: fetch full detail for a single skill by slug
Both methods delegate to existing agent-layer functions
(searchSkillsFromClawHub, fetchSkillDetailFromClawHub) which wrap
the ClawHub HTTP client. No new external dependencies.
Signed-off-by: samzong <samzong.lu@gmail.com>
* feat(skills): add ClawHub skill search and detail in Control UI
Add skills.search and skills.detail Gateway RPC methods with typed
protocol schemas, AJV validators, and handler implementations. Wire
the new RPCs into the Control UI Skills panel with a debounced search
input, results list, detail dialog, and one-click install from ClawHub.
Gateway:
- SkillsSearchParams/ResultSchema and SkillsDetailParams/ResultSchema
- Handler calls searchClawHubSkills and fetchClawHubSkillDetail directly
- Remove zero-logic fetchSkillDetailFromClawHub wrapper
- 9 handler tests including boundary validation
Control UI:
- searchClawHub, loadClawHubDetail, installFromClawHub controllers
- 300ms debounced search input to avoid 429 rate limits
- Dedicated install busy state (clawhubInstallSlug) with success/error feedback
- Install buttons disabled during install with progress text
- Detail dialog with owner, version, changelog, platform metadata
Part of #43301
Signed-off-by: samzong <samzong.lu@gmail.com>
* fix(skills): guard search and detail responses against stale writes
Signed-off-by: samzong <samzong.lu@gmail.com>
* fix(skills): reset loading flags on query clear and detail close
Signed-off-by: samzong <samzong.lu@gmail.com>
* fix(gateway): register skills.search/detail in read scope and method list
Add skills.search and skills.detail to the operator READ scope group
and the server methods list. Without this, unclassified methods default
to operator.admin, blocking read-only operator sessions.
Also guard the detail loading reset in the finally block by the active
slug to prevent a transient flash when rapidly switching skills.
Signed-off-by: samzong <samzong.lu@gmail.com>
* fix(skills): guard search loading reset by active query
Signed-off-by: samzong <samzong.lu@gmail.com>
* test: cover ClawHub skills UI flow
* fix: clear stale ClawHub search results
---------
Signed-off-by: samzong <samzong.lu@gmail.com>
Co-authored-by: Frank Yang <frank.ekn@gmail.com>
resolveAgentConfig().subagents.allowAgents reads only the per-agent
entry, never falling back to agents.defaults.subagents.allowAgents.
Other subagent defaults like runTimeoutSeconds correctly read from
cfg.agents.defaults.subagents — allowAgents was missed.
Root cause: subagent-spawn.ts:463 and agents-list-tool.ts:49 both
use resolveAgentConfig() which returns only per-agent config without
defaults merging. The same pattern is already established at
subagent-spawn.ts:403 for runTimeoutSeconds.
Fix: add cfg.agents.defaults.subagents.allowAgents as fallback when
per-agent entry doesn't specify allowAgents. Both call sites fixed.
Closes#59938
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Signed-off-by: HCL <chenglunhu@gmail.com>
@@ -30,6 +30,7 @@ Use this skill for Parallels guest workflows and smoke interpretation. Do not lo
- Flow: fresh snapshot -> install npm package baseline -> smoke -> install current main tgz on the same guest -> smoke again.
- Same-guest update verification should set the default model explicitly to `openai/gpt-5.4` before the agent turn and use a fresh explicit `--session-id` so old session model state does not leak into the check.
- The aggregate npm-update wrapper must resolve the Linux VM with the same Ubuntu fallback policy as `parallels-linux-smoke.sh` before both fresh and update lanes. Treat any Ubuntu guest with major version `>= 24` as acceptable when the exact default VM is missing, preferring the closest version match. On Peter's current host today, missing `Ubuntu 24.04.3 ARM64` should fall back to `Ubuntu 25.10`.
- On macOS same-guest update checks, restart the gateway after the npm upgrade before `gateway status` / `agent`; launchd can otherwise report a loaded service while the old process has exited and the fresh process is not RPC-ready yet.
- On Windows same-guest update checks, restart the gateway after the npm upgrade before `gateway status` / `agent`; in-place global npm updates can otherwise leave stale hashed `dist/*` module imports alive in the running service.
- For Windows same-guest update checks, prefer the done-file/log-drain PowerShell runner pattern over one long-lived `prlctl exec ... powershell -EncodedCommand ...` transport. The guest can finish successfully while the outer `prlctl exec` still hangs.
- The Windows same-guest update helper should write stage markers to its log before long steps like tgz download and `npm install -g` so the outer progress monitor does not sit on `waiting for first log line` during healthy but quiet installs.
@@ -46,6 +47,7 @@ Use this skill for Parallels guest workflows and smoke interpretation. Do not lo
- Default to the snapshot closest to `macOS 26.3.1 latest`.
- On Peter's Tahoe VM, `fresh-latest-march-2026` can hang in `prlctl snapshot-switch`; if restore times out there, rerun with `--snapshot-hint 'macOS 26.3.1 latest'` before blaming auth or the harness.
-`parallels-macos-smoke.sh` now retries `snapshot-switch` once after force-stopping a stuck running/suspended guest. If Tahoe still times out after that recovery path, then treat it as a real Parallels/host issue and rerun manually.
- The macOS smoke should include a dashboard load phase after gateway health: resolve the tokenized URL with `openclaw dashboard --no-open`, verify the served HTML contains the Control UI title/root shell, then open Safari and require an established localhost TCP connection from Safari to the gateway port.
- If a packaged install regresses with `500` on `/`, `/healthz`, or `__openclaw/control-ui-config.json` after `fresh.install-main` or `upgrade.install-main`, suspect bundled plugin runtime deps resolving from the package root `node_modules` rather than `dist/extensions/*/node_modules`. Repro quickly with a real `npm pack`/global install lane before blaming dashboard auth or Safari.
-`prlctl exec` is fine for deterministic repo commands, but use the guest Terminal or `prlctl enter` when installer parity or shell-sensitive behavior matters.
@@ -64,6 +66,7 @@ Use this skill for Parallels guest workflows and smoke interpretation. Do not lo
- Use PowerShell only as the transport with `-ExecutionPolicy Bypass`, then call the `.cmd` shims from inside it.
- Multi-word `openclaw agent --message ...` checks should call `& $openclaw ...` inside PowerShell, not `Start-Process ... -ArgumentList` against `openclaw.cmd`, or Commander can see split argv and throw `too many arguments for 'agent'`.
- Windows installer/tgz phases now retry once after guest-ready recheck; keep new Windows smoke steps idempotent so a transport-flake retry is safe.
- If a Windows retry sees the VM become `suspended` or `stopped`, resume/start it before the next `prlctl exec`; otherwise the second attempt just repeats the same `rc=255`.
- Windows global `npm install -g` phases can stay quiet for a minute or more even when healthy; inspect the phase log before calling it hung, and only treat it as a regression once the retry wrapper or timeout trips.
- Fresh Windows ref-mode onboard should use the same background PowerShell runner plus done-file/log-drain pattern as the npm-update helper, including startup materialization checks, host-side timeouts on short poll `prlctl exec` calls, and retry-on-poll-failure behavior for transient transport flakes.
- Fresh Windows ref-mode agent verification should set `OPENAI_API_KEY` in the PowerShell environment before invoking `openclaw.cmd agent`, for the same pairing-required fallback reason as macOS.
@@ -82,6 +85,7 @@ Use this skill for Parallels guest workflows and smoke interpretation. Do not lo
- Fresh `main` tgz smoke still needs the latest-release installer first because the snapshot has no Node or npm before bootstrap.
- This snapshot does not have a usable `systemd --user` session; managed daemon install is unsupported.
- The Linux smoke now falls back to a manual `setsid openclaw gateway run --bind loopback --port 18789 --force` launch with `HOME=/root` and the provider secret exported, then verifies `gateway status --deep --require-rpc` when available.
- The Linux manual gateway launch should wait for `gateway status --deep --require-rpc` inside the `gateway-start` phase; otherwise the first status probe can race the background bind and fail a healthy lane.
- If Linux gateway bring-up fails, inspect `/tmp/openclaw-parallels-linux-gateway.log` in the guest phase logs first; the common failure mode is a missing provider secret in the launched gateway environment.
"Closing this PR because it looks dirty (too many unrelated or unexpected changes). This usually happens when a branch picks up unrelated commits or a merge went sideways. Please recreate the PR from a clean branch.";
if (pullRequest) {
// `bad-barnacle` exempts PRs that Barnacle incorrectly marked dirty.
if (labelSet.has(dirtyLabel) && !labelSet.has(badBarnacleLabel)) {
if (labelSet.has(badBarnacleLabel)) {
core.info(`Skipping PR auto-response checks for #${pullRequest.number} because ${badBarnacleLabel} is present.`);
- Rule: protocol changes are contract changes. Prefer additive evolution; incompatible changes require explicit versioning, docs, and client/codegen follow-through.
- Config contract boundary:
- Canonical public config lives in exported config types, zod/schema surfaces, schema help/labels, generated config metadata, config baselines, and any user-facing gateway/config payloads. Keep those surfaces aligned.
- When a legacy config key is retired from the public contract, remove it from every public config surface above. Keep backward compatibility only through raw-config migration/doctor seams unless explicit product policy says otherwise.
- Do not reintroduce removed legacy aliases into public types/schema/help/baselines “for convenience”. If old configs still need to load, handle that in `legacy.migrations.*`, config ingest, or `openclaw doctor --fix`.
-`hooks.internal.entries` is the canonical public hook config model. `hooks.internal.handlers` is compatibility-only input and must not be re-exposed in public schema/help/baseline surfaces.
- Bundled plugin contract boundary:
- Public docs: `docs/plugins/architecture.md`, `docs/plugins/manifest.md`, `docs/plugins/sdk-overview.md`
- README (GitHub): keep absolute docs URLs (`https://docs.openclaw.ai/...`) so links work on GitHub.
- Docs content must be generic: no personal device names/hostnames/paths; use placeholders like `user@gateway-host` and “gateway host”.
## Docs i18n (zh-CN)
## Docs i18n (generated publish locales)
-`docs/zh-CN/**` is generated; do not edit unless the user explicitly asks.
- Pipeline: update English docs → adjust glossary (`docs/.i18n/glossary.zh-CN.json`) → run `scripts/docs-i18n` → apply targeted fixes only if instructed.
-Generated publish output lives in the sibling `openclaw-docs` repo; do not add or edit `docs/zh-CN/**`, `docs/ja-JP/**`, `docs/es/**`, `docs/pt-BR/**`, `docs/ko/**`, `docs/de/**`, `docs/fr/**`, or `docs/ar/**` here.
- Pipeline: update English docs here → adjust glossary (`docs/.i18n/glossary.zh-CN.json`, `docs/.i18n/glossary.ja-JP.json`, `docs/.i18n/glossary.es.json`, `docs/.i18n/glossary.pt-BR.json`, `docs/.i18n/glossary.ko.json`, `docs/.i18n/glossary.de.json`, `docs/.i18n/glossary.fr.json`, `docs/.i18n/glossary.ar.json`) → let the publish-repo sync + `scripts/docs-i18n` run in `openclaw-docs` → apply targeted fixes only if instructed.
- Before rerunning `scripts/docs-i18n`, add glossary entries for any new technical terms, page titles, or short nav labels that must stay in English or use a fixed translation (for example `Doctor` or `Polls`).
-`pnpm docs:check-i18n-glossary` enforces glossary coverage for changed English doc titles and short internal doc labels before translation reruns.
- Translation memory: locale TM files such as`docs/.i18n/zh-CN.tm.jsonl`, `docs/.i18n/ja-JP.tm.jsonl`, `docs/.i18n/es.tm.jsonl`, `docs/.i18n/pt-BR.tm.jsonl`, `docs/.i18n/ko.tm.jsonl`, `docs/.i18n/de.tm.jsonl`, `docs/.i18n/fr.tm.jsonl`, and `docs/.i18n/ar.tm.jsonl` (generated in the publish repo).
- See `docs/.i18n/README.md`.
- The pipeline can be slow/inefficient; if it’s dragging, ping @jospalmbier on Discord instead of hacking around it.
@@ -125,10 +130,10 @@
- Formatting gate: the pre-commit hook runs `pnpm format` before `pnpm check`. If you want a formatting-only preflight locally, run `pnpm format` explicitly.
- If you need a fast commit loop, `FAST_COMMIT=1 git commit ...` skips the hook’s repo-wide `pnpm format` and `pnpm check`; use that only when you are deliberately covering the touched surface some other way.
- Generated baseline artifacts live together under `docs/.generated/`.
- Generated baseline drift detection uses SHA-256 hash files under `docs/.generated/` (`.sha256` files tracked in git; full JSON baselines are gitignored, generated locally for inspection).
- If you change config schema/help or the public Plugin SDK surface, update the matching baseline artifact and keep the two drift-check flows adjacent in scripts/workflows/docs guidance rather than inventing a third pattern.
- If you change config schema/help or the public Plugin SDK surface, run the matching gen command and commit the updated `.sha256` hash file. Keep the two drift-check flows adjacent in scripts/workflows/docs guidance rather than inventing a third pattern.
- For narrowly scoped changes, prefer narrowly scoped tests that directly validate the touched behavior. If no meaningful scoped test exists, say so explicitly and use the next most direct validation available.
- Verification modes for work on `main`:
- Default mode: `main` is relatively stable. Count pre-commit hook coverage when it already verified the current tree, avoid rerunning the exact same checks just for ceremony, and prefer keeping CI/main green before landing.
@@ -140,6 +145,14 @@
- For narrowly scoped changes, if unrelated failures already exist on latest `origin/main`, state that clearly, report the scoped tests you ran, and ask before broadening scope into unrelated fixes or landing despite those failures.
- Do not use scoped tests as permission to ignore plausibly related failures.
## Prompt Cache Stability
- Treat prompt-cache stability as correctness/perf-critical, not cosmetic.
- Any code that assembles model or tool payloads from maps, sets, registries, plugin lists, MCP catalogs, filesystem reads, or network results must make ordering deterministic before building the request.
- Do not rewrite older transcript/history bytes on every turn unless you intentionally want to invalidate the cached prefix. Legacy cleanup, pruning, normalization, and migration logic should preserve recent prompt bytes when possible.
- If truncation or compaction is required, prefer mutating newest or tail content first so the cached prefix stays byte-identical for as long as possible.
- For cache-sensitive changes, require a regression test that proves turn-to-turn prefix stability or deterministic request assembly; helper-local tests alone are not enough.
- Test performance guardrail: inside an extension package, prefer a thin local seam (`./api.ts`, `./runtime-api.ts`, or a narrower local `*.runtime-api.ts`) over direct `openclaw/plugin-sdk/*` imports for internal production code. Keep local seams curated and lightweight; only reach for direct `plugin-sdk/*` imports when you are crossing a real package boundary or when no suitable local seam exists yet.
- Test performance guardrail: keep expensive runtime fallback work such as snapshotting, migration, installs, or bootstrap behind dedicated `*.runtime.ts` boundaries so tests can mock the seam instead of accidentally invoking real work.
- Test performance guardrail: for import-only/runtime-wrapper tests, keep the wrapper lazy. Do not eagerly load heavy verification/bootstrap/runtime modules at module top level if the exported function can import them on demand.
- Test performance guardrail: prefer explicit mock factories over `importOriginal()` for broad modules. Reserve `importOriginal()` for narrow modules where partial-real behavior is genuinely needed.
- Test performance guardrail: do not partial-mock broad `openclaw/plugin-sdk/*` barrels in hot tests. Add a plugin-local `*.runtime.ts` seam and mock that seam instead.
- Test performance guardrail: when production code already accepts `deps`, callbacks, or runtime injection, use that seam in tests before adding module-level mocks.
- Test performance guardrail: prefer narrow public SDK subpaths such as `models-provider-runtime`, `skill-commands-runtime`, and `reply-dispatch-runtime` over older broad helper barrels when both expose the needed helper.
- Test performance guardrail: treat import-dominated test time as a boundary bug. Refactor the import surface before adding more cases to the slow file.
- Agents MUST NOT modify baseline, inventory, ignore, snapshot, or expected-failure files to silence failing checks without explicit approval in this chat.
- For targeted/local debugging, keep using the wrapper: `pnpm test -- <path-or-filter> [vitest args...]` (for example `pnpm test -- src/commands/onboard-search.test.ts -t "shows registered plugin providers"`); do not default to raw `pnpm vitest run ...` because it bypasses wrapper config/profile/pool routing.
- For targeted/local debugging, use the native root-project entrypoint: `pnpm test <path-or-filter> [vitest args...]` (for example `pnpm test src/commands/onboard-search.test.ts -t "shows registered plugin providers"`); do not default to raw `pnpm vitest run ...` because it bypasses the repo's default config/profile/pool routing.
- Do not set test workers above 16; tried already.
- Keep Vitest on `forks` only. Do not introduce or reintroduce any non-`forks` Vitest pool or alternate execution mode in configs, wrapper scripts, or default test commands without explicit approval in this chat. This includes `threads`, `vmThreads`, `vmForks`, and any future/nonstandard pool variant.
- If local Vitest runs cause memory pressure, the wrapper now derives budgets from host capabilities (CPU, memory band, current load). For a conservative explicit override during land/gate runs, use `OPENCLAW_TEST_PROFILE=serial OPENCLAW_TEST_SERIAL_GATEWAY=1 pnpm test`.
- Vitest now defaults to native root-project `threads`, with hard `forks` exceptions for `gateway`, `agents`, and `commands`. Keep new pool changes explicit and justified; use `OPENCLAW_VITEST_POOL=forks` for full local fork debugging.
- If local Vitest runs cause memory pressure, the default worker budget now derives from host capabilities (CPU, memory band, current load). For a conservative explicit override during land/gate runs, use `OPENCLAW_VITEST_MAX_WORKERS=1 pnpm test`.
- Live tests (real keys): `OPENCLAW_LIVE_TEST=1 pnpm test:live` (OpenClaw-only) or `LIVE=1 pnpm test:live` (includes provider live tests). Docker: `pnpm test:docker:live-models`, `pnpm test:docker:live-gateway`. Onboarding Docker E2E: `pnpm test:docker:onboard`.
- Full kit + what’s covered: `docs/help/testing.md`.
@@ -247,6 +265,8 @@
- "Bump version everywhere" means all version locations above **except**`appcast.xml` (only touch appcast when cutting a new macOS Sparkle release).
- **Restart apps:** “restart iOS/Android apps” means rebuild (recompile/install) and relaunch, not just kill/launch.
- **Device checks:** before testing, verify connected real devices (iOS/Android) before reaching for simulators/emulators.
- Mobile pairing: `ws://` (cleartext) is allowed for private LAN addresses (RFC 1918, link-local, mDNS `.local`) and loopback. Private LAN hosts typically lack PKI-backed identity, so requiring TLS there adds complexity without meaningful security gain. `wss://` is required for Tailscale and public endpoints.
- Security report scope: reports that treat cleartext `ws://` mobile pairing over private LAN as a vulnerability are out of scope unless they demonstrate a trust-boundary bypass beyond passive network observation on the same LAN.
- iOS Team ID lookup: `security find-identity -p codesigning -v` → use Apple Development (…) TEAMID. Fallback: `defaults read com.apple.dt.Xcode IDEProvisioningTeamIdentifiers`.
- A2UI bundle hash: `src/canvas-host/a2ui/.bundle.hash` is auto-generated; ignore unexpected changes, and only regenerate via `pnpm canvas:a2ui:bundle` (or `scripts/bundle-a2ui.sh`) when needed. Commit the hash as a separate commit.
- Release signing/notary credentials are managed outside the repo; maintainers keep that setup in the private [maintainer release docs](https://github.com/openclaw/maintainers/tree/main/release).
- Config: remove legacy public config aliases such as `talk.voiceId` / `talk.apiKey`, `agents.*.sandbox.perSession`, `browser.ssrfPolicy.allowPrivateNetwork`, `hooks.internal.handlers`, and channel/group/room `allow` toggles in favor of the canonical public paths and `enabled`, while keeping load-time compatibility and `openclaw doctor --fix` migration support for existing configs. (#60726) Thanks @vincentkoc.
### Changes
- Agents/cache: diagnostics: add prompt-cache break diagnostics, trace live cache scenarios through embedded runner paths, and show cache reuse explicitly in `openclaw status --verbose`. Thanks @vincentkoc.
- Agents/cache: stabilize cache-relevant system prompt fingerprints by normalizing equivalent structured prompt whitespace, line endings, hook-added system context, and runtime capability ordering so semantically unchanged prompts reuse KV/cache more reliably. Thanks @vincentkoc.
- Agents/tool prompts: remove the duplicate in-band tool inventory from agent system prompts so tool-calling models rely on the structured tool definitions as the single source of truth, improving prompt stability and reducing stale tool guidance.
- Agents/Claude CLI: expose OpenClaw tools to background Claude CLI runs through a loopback MCP bridge that reuses gateway tool policy, honors session/account/channel scoping, and only advertises the bridge when the local runtime is actually live. (#35676) Thanks @mylukin.
- Agents/Claude CLI: switch bundled Claude CLI runs to stdin + `stream-json` partial-message streaming so prompts stop riding argv, long replies show live progress, and final session/usage metadata still land cleanly.
- Docs/IRC: replace public IRC hostname examples with `irc.example.com` and recommend private servers for bot coordination while listing common public networks for intentional use.
- Channels/context visibility: add configurable `contextVisibility` per channel (`all`, `allowlist`, `allowlist_quote`) so supplemental quote, thread, and fetched history context can be filtered by sender allowlists instead of always passing through as received.
- Config/schema: enrich the exported `openclaw config schema` JSON Schema with field titles and descriptions so editors, agents, and other schema consumers receive the same config help metadata. (#60067) Thanks @solavrc.
- Control UI/skills: add ClawHub search, detail, and install flows directly in the Skills panel. (#60134) Thanks @samzong.
- Matrix/exec approvals: add Matrix-native exec approval prompts with account-scoped approvers, channel-or-DM delivery, and room-thread aware resolution handling. (#58635) Thanks @gumadeiras.
- Memory/dreaming (experimental): add weighted short-term recall promotion, managed dreaming modes (`off|core|rem|deep`), a `/dreaming` command, Dreams UI, multilingual conceptual tagging, and doctor/status repair support so durable memory promotion can run in the background with less manual setup. (#60569, #60697) Thanks @vignesh07.
- MiniMax/TTS: add a bundled MiniMax speech provider backed by the T2A v2 API so speech synthesis can run through MiniMax-native voices and auth. (#55921) Thanks @duncanita.
- Plugins/install: add `openclaw plugins install --force` to overwrite existing plugin and hook-pack install targets without using the dangerous-code override flag. (#60544) Thanks @gumadeiras.
- Plugins/onboarding: add plugin config TUI prompts to onboard and configure wizards so more plugin setup can stay in the guided flow. (#60590) Thanks @odysseus0.
- Prompt caching: keep prompt prefixes more reusable across transport fallback, deterministic MCP tool ordering, compaction, and embedded image history so follow-up turns hit cache more reliably. (#58036, #58037, #58038, #59054, #60603, #60691) Thanks @bcherny.
- Providers/Amazon Bedrock: discover regional and global inference profiles, inherit their backing model capabilities, and inject the Bedrock request region automatically so cross-region Claude profiles work without manual provider overrides. (#61299) Thanks @wirjo.
- Providers/Amazon Bedrock Mantle: add a bundled OpenAI-compatible Mantle provider with bearer-token discovery, automatic OSS model catalog loading, and Bedrock Mantle region detection for hosted GPT-OSS, Qwen, Kimi, GLM, and similar routes. (#61296) Thanks @wirjo.
- Providers/Anthropic: remove setup-token from new onboarding and auth-command setup paths, keep existing configured legacy token profiles runnable, and steer new Anthropic setup to Claude CLI or API keys.
- Providers/Fireworks: add a bundled Fireworks AI provider plugin with `FIREWORKS_API_KEY` onboarding, Fire Pass Kimi defaults, and dynamic Fireworks model-id support.
- Providers/Ollama: add a bundled Ollama Web Search provider for key-free `web_search` via your configured Ollama host and `ollama signin`. (#59318) Thanks @BruceMacD.
- Providers/OpenAI Codex: add forward-compat `openai-codex/gpt-5.4-mini` synthesis across provider runtime, model catalog, and model listing so Codex mini works before bundled Pi catalog updates land.
- Providers/request overrides: add shared model and media request transport overrides across OpenAI-, Anthropic-, Google-, and compatible provider paths, including headers, auth, proxy, and TLS controls. (#60200)
- Providers/StepFun: add the bundled StepFun provider plugin with standard and Step Plan endpoints, China/global onboarding choices, `step-3.5-flash` on both catalogs, and `step-3.5-flash-2603` currently exposed on Step Plan. (#60032) Thanks @hengm3467.
- Tools/web_search: add a bundled MiniMax Search provider backed by the Coding Plan search API, with region reuse from `MINIMAX_API_HOST` and plugin-owned credential config. (#54648) Thanks @fengmk2.
### Fixes
-Telegram/reactions: preserve `reactionNotifications: "own"` across gateway restarts by persisting sent-message ownership state instead of treating cold cache as a permissive fallback. (#59207) Thanks @samzong.
-Gateway/startup: detect PID recycling in gateway lock files on Windows and macOS, and add startup progress so stale lock conflicts no longer block healthy restarts. (#59843) Thanks @TonyDerek-dot.
-MS Teams/DM media: download inline images in 1:1 chats via Graph API so Teams DM image attachments stop failing to load. (#52212) Thanks @Ted-developer.
-MS Teams/threading: preserve channel reply threading in proactive fallback so replies stay in the original thread instead of dropping into the channel root. (#55198) Thanks @hyojin.
-Telegram/media: preserve `<media:...>` placeholders and `file_id` in captioned messages when Bot API downloads fail, so agents still receive media context. (#59948) Thanks @v1p0r.
-Telegram/media: keep inbound image attachments readable on upgraded installs where legacy state roots still differ from the managed config-dir media cache. (#59971) Thanks @neeravmakwana.
-Telegram/local Bot API: thread `channels.telegram.apiRoot` through buffered reply-media and album downloads so self-hosted Bot API file paths stop falling back to `api.telegram.org` and 404ing. (#59544) Thanks @SARAMALI15792.
-Telegram/replies: preserve explicit topic targets when `replyTo` is present while still inheriting the current topic for same-chat replies without an explicit topic. (#59634) Thanks @dashhuang.
-Telegram/native commands: clean up metadata-driven progress placeholders when replies fall back, edits fail, or local exec approval prompts are suppressed. (#59300) Thanks @jalehman.
-Media/request overrides: resolve shared and capability-filtered media request SecretRefs correctly and expose media transport override fields to schema-driven config consumers. (#59848) Thanks @vincentkoc.
-Matrix: allow secret-storage recreation during automatic repair bootstrap so clients that lose their recovery key can recover and persist new cross-signing keys. (#59846) Thanks @al3mart.
-Matrix/crypto persistence: capture and write the IndexedDB snapshot while holding the snapshot file lock so concurrent gateway and CLI persists cannot overwrite newer crypto state. (#59851) Thanks @al3mart.
-Ollama/auth: prefer real cloud auth over local marker during model auth resolution so cloud-backed Ollama auth does not get shadowed by stale local-only markers.
-Plugins/Kimi Coding: parse tagged Kimi tool-call text into structured tool calls on the provider stream path so tools execute instead of echoing raw markup. (#60051) Thanks @obviyus.
-Channels/passive hooks: emit passive message hooks for mention-skipped Telegram and Signal group messages when `ingest` is enabled, including wildcard/default fallback and per-group override handling. (#60018) Thanks @obviyus.
-Plugins/manifest registry: stop warning when an explicit manifest `id` intentionally differs from the discovery hint. (#59185) Thanks @samzong.
-WhatsApp/streaming: honor `channels.whatsapp.blockStreaming` again for inbound auto-replies so progressive block replies can be enabled explicitly instead of being forced to final-only delivery. Thanks @mcaxtr.
-Plugins/runtime: reuse compatible active registries for `web_search` and `web_fetch` provider snapshot resolution so repeated runtime reads do not re-import the same bundled plugin set on each agent message. Related #48380.
-Infra/tailscale: ignore `OPENCLAW_TEST_TAILSCALE_BINARY` outside explicit test environments and block it from workspace `.env`, so test-only binary overrides cannot be injected through trusted repository state. (#58468) Thanks @eleqtrizit.
- Agents/tool policy: preserve restrictive plugin-only allowlists instead of silently widening access to core tools, and keep allowlist warnings aligned with the enforced policy. (#58476) Thanks @eleqtrizit.
-Hooks/session_end: preserve deterministic reason metadata for custom reset aliases and overlapping idle-plus-daily rollovers so plugins can rely on lifecycle reason reporting. (#59715) Thanks @jalehman.
-Tools/image generation: stop inferring unsupported resolution overrides for OpenAI reference-image edits when no explicit `size` or `resolution` is provided, so default edit flows no longer fail before the provider request is sent.
- Agents/sessions: release embedded runner session locks even when teardown cleanup throws, so timed-out or failed cleanup paths no longer leave sessions wedged until the stale-lock watchdog recovers them. (#59194) Thanks @samzong.
-ACP/agents: inherit the target agent workspace for cross-agent ACP spawns and fall back safely when the inherited workspace no longer exists. (#58438) Thanks @zssggle-rgb.
-ACPX/Windows: preserve backslashes and absolute `.exe` paths in Claude CLI parsing, and fail fast on wrapper-script targets with guidance to use `cmd.exe /c`, `powershell.exe -File`, or `node <script>`. (#60689) Thanks @steipete.
-Agents/Anthropic: preserve native `toolu_*` replay ids on direct Anthropic and Anthropic Vertex paths so cache-sensitive history stops rewriting known-valid Anthropic tool-use ids. (#52612)
-Agents/cache: preserve the full 3-turn prompt-cache image window across tool loops, keep colliding bundled MCP tool definitions deterministic, and reapply Anthropic Vertex cache shaping after payload hook replacements so KV/cache reuse stays stable. Thanks @vincentkoc.
-Agents/Claude CLI/images: reuse stable hydrated image file paths and preserve shared media extensions like HEIC when passing image refs to local CLI runs, so Claude CLI image prompts stop thrashing KV cache prefixes and oddball image formats do not fall back to `.bin`. Thanks @vincentkoc.
-Agents/Claude CLI/security: clear inherited Claude Code config-root and plugin-root env overrides like `CLAUDE_CONFIG_DIR` and `CLAUDE_CODE_PLUGIN_*`, so OpenClaw-launched Claude CLI runs cannot be silently pointed at an alternate Claude config/plugin tree with different hooks, plugins, or auth context. Thanks @vincentkoc.
-Agents/Claude CLI/security: clear inherited Claude Code provider-routing and managed-auth env overrides, and mark OpenClaw-launched Claude CLI runs as host-managed, so Claude CLI backdoor sessions cannot be silently redirected to proxy, Bedrock, Vertex, Foundry, or parent-managed token contexts. Thanks @vincentkoc.
-Agents/Claude CLI/security: force host-managed Claude CLI backdoor runs to `--setting-sources user`, even under custom backend arg overrides, so repo-local `.claude` project/local settings, hooks, and plugin discovery do not silently execute inside non-interactive OpenClaw sessions. Thanks @vincentkoc.
-Agents/Claude CLI: keep non-interactive `--permission-mode bypassPermissions` when custom `cliBackends.claude-cli.args` override defaults, including fallback resolution before the runtime plugin registry is active, so cron and heartbeat Claude CLI runs do not regress to interactive approval mode. (#61114) Thanks @cathrynlavery and @thewilloftheshadow.
-Agents/Claude CLI: persist explicit `openclaw agent --session-id` runs under a stable session key so follow-ups can reuse the stored CLI binding and resume the same underlying Claude session.
-Agents/Claude CLI: treat malformed bare `--permission-mode` backend overrides as missing and fail safe back to `bypassPermissions`, so custom `cliBackends.claude-cli.args` security config cannot accidentally consume the next flag as a bogus permission mode. Thanks @vincentkoc.
-Agents/CLI backends: invalidate stored CLI session reuse when local CLI login state or the selected auth profile credential changes, so relogin and token rotation stop resuming stale sessions.
-Agents/compaction: keep assistant tool calls and displaced tool results in the same compaction chunk so strict summarization providers stop rejecting orphaned tool pairs. (#58849) Thanks @openperf.
-Agents/errors: surface an explicit disk-full message when local session or transcript writes fail with `ENOSPC`/`disk full`, so those runs stop degrading into opaque `NO_REPLY`-style failures. Thanks @vincentkoc.
-Agents/exec approvals: let `exec-approvals.json` agent security override stricter gateway tool defaults so approved subagents can use `security: "full"` without falling back to allowlist enforcement again. (#60310) Thanks @lml2468.
-Agents/exec: restore `host=node` routing for node-pinned and `host=auto` sessions, while still blocking sandboxed `auto` sessions from jumping to gateway. (#60788) Thanks @openperf.
-Agents/failover: scope Anthropic `An unknown error occurred` failover matching by provider so generic internal unknown-error text no longer triggers retryable timeout fallback. (#59325) Thanks @aaron-he-zhu.
-Agents/GPT: add explicit work-item lifecycle events for embedded runs, use them to surface real progress more reliably, and stop counting tool-started turns as planning-only retries.
-Agents/Kimi tool-call repair: preserve tool arguments that were already present on streamed tool calls when later malformed deltas fail reevaluation, while still dropping stale repair-only state before `toolcall_end`.
- Agents/OpenAI: mark Claude-compatible file tool schemas as `additionalProperties: false` so direct OpenAI GPT-5 routes stop rejecting the `read` tool with invalid strict-schema errors.
-Agents/output delivery: suppress `phase:"commentary"` assistant text at the embedded subscribe boundary so internal planning text cannot leak into user-visible replies or Telegram partials. (#61282) Thanks @mbelinky.
-Agents/runtime: make default subagent allowlists, inherited skills/workspaces, and duplicate session-id resolution behave more predictably, and include value-shape hints in missing-parameter tool errors. (#59944, #59992, #59858, #55317) Thanks @hclsys, @gumadeiras, @joelnishanth, and @priyansh19.
- Agents/scheduling: steer background-now work toward automatic completion wake and treat `process` polling as on-demand inspection or intervention instead of default completion handling. (#60877) Thanks @vincentkoc.
- Agents/skills: skip `.git` and `node_modules` when mirroring skills into sandbox workspaces so read-only sandboxes do not copy repo history or dependency trees. (#61090) Thanks @joelnishanth.
- Agents/subagents: honor allowlist validation, auth-profile handoff, and session override state when a subagent retries after `LiveSessionModelSwitchError`. (#58178) Thanks @openperf.
- Amazon Bedrock/aws-sdk auth: stop injecting the fake `AWS_PROFILE` apiKey marker when no AWS auth env vars exist, so instance-role and other default-chain setups keep working without poisoning provider config. (#61194) Thanks @wirjo.
- Android/canvas security: require exact normalized A2UI URL matches before forwarding canvas bridge actions, rejecting query mismatches and descendant paths while still allowing fragment-only A2UI navigation.
- Android/Talk Mode: cancel in-flight `talk.speak` playback when speech is explicitly stopped, so stale replies stop starting after barge-in or manual stop. (#61164) Thanks @obviyus.
- Android/Talk Mode: restore spoken assistant replies on node-scoped sessions by keeping reply routing synced to the resolved node session key and pausing mic capture during reply playback. (#60306) Thanks @MKV21.
- Android/Talk Mode: restore voice replies on gateway-backed talk mode sessions by updating embedded runner transport overrides to the current agent transport API. (#61214) Thanks @obviyus.
- Anthropic CLI onboarding: rewrite migrated fallback model refs during non-interactive Claude CLI setup too, so onboarding and scripted setup no longer keep stale `anthropic/*` fallbacks after switching the primary model to `claude-cli/*`. Thanks @vincentkoc.
- Auth/failover: persist selected fallback overrides before retrying, shorten `auth_permanent` lockouts, and refresh websocket/shared-auth sessions only when real auth changes occur so retries and secret rotations behave predictably. (#60404, #60323, #60387) Thanks @extrasmall0 and @mappel-nv.
- CLI/Commander: preserve Commander-computed exit codes for argument and help-error paths, and cover the user-argv parse mode in the regression tests so invalid CLI invocations no longer report success when exits are intercepted. (#60923) Thanks @Linux2010.
- CLI/skills JSON: route `skills list --json`, `skills info --json`, and `skills check --json` output to stdout instead of stderr so machine-readable consumers receive JSON on the expected stream again. (#60914; fixes #57599; landed from contributor PR #57611 by @Aftabbs) Thanks @Aftabbs.
- Config/All Settings: keep the raw config view intact when sensitive fields are blank instead of corrupting or dropping the rendered snapshot. (#28214) Thanks @solodmd.
- Control UI/avatar: honor `ui.assistant.avatar` when serving `/avatar/:agentId` so Appearance UI avatar paths stop falling back to initials placeholders. (#60778) Thanks @hannasdev.
- Control UI/cron: highlight the Cron refresh button while refresh is in flight so the page's loading state stays visible even when prior data remains on screen. (#60394) Thanks @coder-zhuzm.
- Control UI/Overview: prevent gateway access token/password visibility toggle buttons from overlapping their inputs at narrow widths. (#56924) Thanks @bbddbb1.
- Control UI: keep Stop visible during tool-only execution, preserve pending-send busy state, and clear stale ClawHub search results as soon as the query changes. (#54528, #59800, #60267) Thanks @chziyue and @frankekn.
- Cron: replay interrupted recurring jobs on the first gateway restart instead of waiting for a second restart. (#60583) Thanks @joelnishanth.
- Cron: send failure notifications through the job's primary delivery channel using the same session context as successful delivery when no explicit `failureDestination` is configured. (#60622) Thanks @artwalker.
- Device pairing/security: keep non-operator device scope checks bound to the requested role prefix so bootstrap verification cannot redeem `operator.*` scopes through `node` auth. (#57258) Thanks @jlapenna.
- Device pairing: reject rotating device tokens into roles that were never approved during pairing, and keep reconnect role checks bounded to the paired device's approved role set. (#60462) Thanks @eleqtrizit.
- Discord: keep REST, webhook, and monitor traffic on the configured proxy, preserve component-only media sends, honor `@everyone` and `@here` mention gates, keep ACK reactions on the active account, and split voice connect/playback timeouts so auto-join is more reliable. (#57465, #60361, #60345) Thanks @geekhuashan.
- Doctor/config: compare normalized `talk` configs by deep structural equality instead of key-order-sensitive serialization so `openclaw doctor --fix` stops repeatedly reporting/applying no-op `talk.provider/providers` normalization. (#59911) Thanks @ejames-dev.
- Exec approvals/node host: forward prepared `system.run` approval plans on the async node invoke path so mutable script operands keep their approval-time binding and drift revalidation instead of dropping back to unbound execution.
- Exec approvals: reuse durable exact-command `allow-always` approvals in allowlist mode so identical reruns stop prompting, and tighten Windows interpreter/path approval handling so wrapper and malformed-path cases fail closed more consistently. (#59880, #59780, #58040, #59182) Thanks @luoyanglang, @SnowSky1, and @pgondhi987.
- Exec/heartbeat: use the canonical `exec-event` wake reason for `notifyOnExit` so background exec completions still trigger follow-up turns when `HEARTBEAT.md` is empty or comments-only. (#41479) Thanks @rstar327.
- Feishu/reasoning: only expose streamed reasoning previews when the session is explicitly `reasoning:stream`, so hidden reasoning traces do not surface on normal streaming sessions. Thanks @vincentkoc.
- Gateway/auth: serialize async shared-secret auth attempts per client so concurrent Tailscale-capable failures cannot overrun the intended auth rate-limit budget. Thanks @Telecaster2147.
- Gateway/channels: pin the initial startup channel registry before later plugin-registry churn so configured channels stay visible and `channels.status` stops falling back to empty `channelOrder` / `channels` payloads after runtime plugin loads.
- Gateway/device auth: reuse cached device-token scopes only for cached-token reconnects, while keeping explicit `deviceToken` scope requests and empty-cache fallbacks intact so reconnects preserve `operator.read` without breaking explicit auth flows. (#46032) Thanks @caicongyang.
- Gateway/device pairing: require non-admin paired-device sessions to manage only their own device for token rotate/revoke and paired-device removal, blocking cross-device token theft inside pairing-scoped sessions. (#50627) Thanks @coygeek.
- Gateway/macOS: let launchd `KeepAlive` own in-process gateway restarts again, adding a short supervised-exit delay so rapid restarts avoid launchd crash-loop unloads while `openclaw gateway restart` still reports real LaunchAgent errors synchronously.
- Gateway/macOS: re-bootstrap the LaunchAgent if `launchctl kickstart -k` unloads it during restart so failed restarts do not leave the gateway unmanaged until manual repair.
- Gateway/macOS: recover installed-but-unloaded LaunchAgents during `openclaw gateway start` and `restart`, while still preferring live unmanaged gateways during restart recovery. (#43766) Thanks @HenryC-3.
- Gateway/plugin routes: keep gateway-auth plugin runtime routes on write-only fallback scopes unless a trusted-proxy caller explicitly declares narrower `x-openclaw-scopes`, so plugin HTTP handlers no longer mint admin-level runtime scopes on missing or untrusted HTTP scope headers. (#59815) Thanks @pgondhi987.
- Gateway/startup: default `gateway.mode` to `local` when unset, detect PID recycling in gateway lock files on Windows and macOS, and show startup progress so healthy restarts stop getting blocked by stale locks. (#54801, #60085, #59843) Thanks @BradGroux and @TonyDerek-dot.
- Gateway/Windows scheduled tasks: preserve Task Scheduler settings on reinstall, fail loudly when `/Run` does not start, and report fast failed restarts accurately instead of pretending they timed out after 60 seconds. (#59335) Thanks @tmimmanuel.
- Google Gemini CLI auth: detect bundled npm installs by scanning packaged bundle files for the Gemini OAuth client config, so `npm install -g @google/gemini-cli` layouts work again. (#60486) Thanks @wzfmini01.
- Google Gemini CLI auth: detect personal OAuth mode from local Gemini settings and skip Code Assist project discovery for those logins, so personal Google accounts stop failing with `loadCodeAssist 400 Bad Request`. (#49226) Thanks @bobworrall.
- Google Gemini CLI auth: improve OAuth credential discovery across Windows nvm and Homebrew libexec installs, and align Code Assist metadata so Gemini login stops failing on packaged CLI layouts. (#40729) Thanks @hughcube.
- Google Gemini CLI models: add forward-compat support for stable `gemini-2.5-*` model ids by letting the bundled CLI provider clone them from Google templates, so `gemini-2.5-flash-lite` and related configured models stop showing up as missing. (#35274) Thanks @mySebbe.
- Google image generation: disable pinned DNS for Gemini image requests and honor explicit `pinDns` overrides in shared provider HTTP helpers so proxy-backed image generation works again. (#59873) Thanks @luoyanglang.
- Heartbeat: skip wake delivery when the target session lane is already busy so the pending event is retried instead of getting drained too early. (#40526) Thanks @lucky7323.
- Live model switching: only treat explicit user-driven model changes as pending live switches, so fallback rotation, heartbeat overrides, and compaction no longer trip `LiveSessionModelSwitchError` before making an API call. (#60266) Thanks @kiranvk-2011.
- Matrix: recover more reliably when secret storage or recovery keys are missing by recreating secret storage during repair and backup reset, hold crypto snapshot locks during persistence, and surface explicit too-large attachment markers. (#59846, #59851, #60599, #60289) Thanks @al3mart, @emonty, and @efe-arv.
- Mattermost/config schema: accept `groups.*.requireMention` again so existing Mattermost configs no longer fail strict validation after upgrade. (#58271) Thanks @MoerAI.
- Media understanding: auto-register image-capable config providers for vision routing, so custom GLM-style provider ids with image models stop failing with “no media-understanding provider registered”. (#51418) Thanks @xydt-610.
- MiniMax/pricing: keep bundled MiniMax highspeed pricing distinct in provider catalogs and preserve the lower M2.5 cache-read pricing when onboarding older MiniMax models. (#54214) Thanks @octo-patch.
- MiniMax: advertise image input on bundled `MiniMax-M2.7` and `MiniMax-M2.7-highspeed` model definitions so image-capable flows can route through the M2.7 family correctly. (#54843) Thanks @MerlinMiao88888888.
- Mobile pairing/Android: tighten secure endpoint handling so Tailscale and public remote setup reject cleartext endpoints, private LAN pairing still works, merged-role approvals mint both node and operator device tokens, and bootstrap tokens survive node auto-pair until operator approval finishes. (#60128, #60208, #60221) Thanks @obviyus.
- Mobile pairing/bootstrap: keep QR bootstrap handoff tokens bounded to the mobile-safe contract so node handoff stays unscoped and operator handoff drops mixed `node.*`, `operator.admin`, and `operator.pairing` scopes.
- Mobile pairing/security: fail closed for internal `/pair` setup-code issuance, cleanup, and approval paths when gateway pairing scopes are missing, and keep approval-time requested-scope enforcement on the internal command path. (#55996) Thanks @coygeek.
- Model picker/providers: treat bundled BytePlus and Volcengine plan aliases as their native providers during setup, and expose their bundled standard/coding catalogs before auth so setup can suggest the right models. (#58819) Thanks @Luckymingxuan.
- Models/Anthropic CLI auth: replace migrated `agents.defaults.models` allowlists when `openclaw models auth login --provider anthropic --method cli --set-default` switches to `claude-cli/*`, so stale `anthropic/*` entries do not linger beside the migrated Claude CLI defaults. Thanks @vincentkoc.
- Models/MiniMax: honor `MINIMAX_API_HOST` for implicit bundled MiniMax provider catalogs so China-hosted API-key setups pick `api.minimaxi.com/anthropic` without manual provider config. (#34524) Thanks @caiqinghua.
- MS Teams: download inline DM images via Graph API and preserve channel reply threading in proactive fallback. (#52212, #55198) Thanks @Ted-developer and @hyojin.
- MS Teams: replace the deprecated Teams SDK HttpPlugin stub with `httpServerAdapter` so recurring gateway deprecation warnings stop firing and the Express 5 compatibility workaround stays on the supported SDK path. (#60939) Thanks @coolramukaka-sys.
- Outbound/sanitizer: strip leaked `<tool_call>`, `<function_calls>`, and model special tokens from shared user-visible assistant text, including truncated tool-call streams, so internal scaffolding no longer bleeds into replies across surfaces. (#60619) Thanks @oliviareid-svg.
- Plugin SDK/context engines: export the missing context-engine result and subagent lifecycle types from `openclaw/plugin-sdk` so context engine plugins can type `ContextEngine` implementations without local workarounds. (#61251) Thanks @DaevMithran.
- Plugin SDK/facades: back-fill bundled plugin facade sentinels before plugin-id tracking re-enters config loading, so CLI/provider startup no longer crashes with `shouldNormalizeGoogleProviderConfig is not a function` or other empty-facade reads during bundled plugin re-entry. Thanks @adam91holt.
- Plugins/auth-choice: apply provider-owned auth config patches without recursively preserving replaced default-model maps, so Anthropic Claude CLI and similar migrations can intentionally swap model allowlists during onboarding and setup instead of accumulating stale entries. Thanks @vincentkoc.
- Plugins/cache: inherit the active gateway workspace for provider, web-search, and web-fetch snapshot loads when callers omit `workspaceDir`, so compatible plugin registries and snapshot caches stop missing on gateway-owned runtime paths. (#61138) Thanks @jzakirov.
- Plugins/facades: back-fill facade sentinels before tracked-plugin resolution re-enters config loading, so facade exports stay defined during circular provider normalization. (#61180) Thanks @adam91holt.
- Providers/OpenAI GPT: treat short approval turns like `ok do it` and `go ahead` as immediate action turns, and trim overly memo-like GPT-5 chat confirmations so OpenAI replies stay shorter and more conversational by default.
- Plugins/install: preserve unsafe override flags across linked plugin and hook-pack probes so local `--link` installs honor the documented override behavior. (#60624) Thanks @JerrettDavis.
- Plugins/Kimi Coding: parse tagged tool calls and keep Anthropic-native tool payloads so Kimi coding endpoints execute tools instead of echoing raw markup. (#60051, #60391) Thanks @obviyus and @Eric-Guo.
- Plugins/marketplace: block remote marketplace symlink escapes without breaking ordinary local marketplace install paths. (#60556) Thanks @eleqtrizit.
- Plugins/media understanding: enable bundled Groq and Deepgram providers by default so configured transcription models work without extra plugin activation config. (#59982) Thanks @yxjsxy.
- Plugins/onboarding: write dotted plugin uiHint paths like Brave `webSearch.mode` as nested plugin config so `llm-context` setup stops failing validation. (#61159) Thanks @obviyus.
- Plugins/OpenAI: enable `gpt-image-1` reference-image edits through `/images/edits` multipart uploads, and stop inferring unsupported resolution overrides when no explicit `size` or `resolution` is provided.
- Plugins/OpenAI: tune the OpenAI prompt overlay for live-chat cadence so GPT replies stay shorter, more human, and less wall-of-text by default.
- Prompt caching: order stable workspace project-context files before `HEARTBEAT.md` and keep `HEARTBEAT.md` below the system-prompt cache boundary so heartbeat churn does not invalidate the stable project-context prefix. (#58979) Thanks @yozu and @vincentkoc.
- Prompt caching: route Codex Responses and Anthropic Vertex through boundary-aware cache shaping, and report the actual outbound system prompt in cache traces so cache reuse and misses line up with what providers really receive. Thanks @vincentkoc.
- Providers/Anthropic Vertex: honor `cacheRetention: "long"` with the real 1-hour prompt-cache TTL on Vertex AI endpoints, and default `anthropic-vertex` cache retention like direct Anthropic. (#60888) Thanks @affsantos.
- Providers/Anthropic: keep `claude-cli/*` auth on live Claude CLI credentials at runtime, avoid persisting stale bearer-token profiles, and suppress macOS Keychain prompts during non-interactive Claude CLI setup. (#61234) Thanks @darkamenosa.
- Providers/compat: stop forcing OpenAI-only defaults on proxy and custom OpenAI-compatible routes, preserve native vendor-specific reasoning/tool/streaming behavior across Anthropic-compatible, Moonshot, Mistral, ModelStudio, OpenRouter, xAI, and Z.ai endpoints, and route GitHub Copilot Claude models through Anthropic Messages instead of OpenAI Responses.
- Providers/GitHub Copilot: send IDE identity headers on runtime model requests and GitHub token exchange so IDE-authenticated Copilot runs stop failing with missing `Editor-Version`. (#60641) Thanks @VACInc and @vincentkoc.
- Providers/Google: add model-level `cacheRetention` support for direct Gemini system prompts by creating, reusing, and refreshing `cachedContents` automatically on Google AI Studio runs. (#51372) Thanks @rafaelmariano-glitch.
- Providers/Model Studio: preserve native streaming usage reporting for DashScope-compatible endpoints even when they are configured under a generic provider key, so streamed token totals stop sticking at zero. (#52395) Thanks @IVY-AI-gif.
- Providers/OpenAI Codex: split native `contextWindow` from runtime `contextTokens`, keep the default effective cap at `272000`, and expose a per-model `contextTokens` override on `models.providers.*.models[]`.
- Providers/OpenAI-compatible WS: compute fallback token totals from normalized usage when providers omit or zero `total_tokens`, so DashScope-compatible sessions stop storing zero totals after alias normalization. (#54940) Thanks @lyfuci.
- Providers/OpenAI: make GPT-5 and Codex runs act sooner with lower-verbosity defaults, visible progress during tool work, and a one-shot retry when a turn only narrates the plan instead of taking action.
- Providers/OpenAI: preserve native `reasoning.effort: "none"` and strict tool schemas on direct OpenAI-family endpoints, keep compat routes on compat shaping, fix Responses WebSocket warm-up behavior, keep stable session and turn metadata, and fall back more gracefully after early WebSocket failures.
- Providers/OpenAI: support GPT-5.4 assistant `phase` metadata across OpenAI-family Responses replay and the Gateway `/v1/responses` compatibility layer, including `commentary` tool preambles and `final_answer` replies.
- Providers/OpenRouter failover: classify `403 "Key limit exceeded"` spending-limit responses as billing so model fallback continues instead of stopping on generic auth. (#59892) Thanks @rockcent.
- Providers/Z.AI: preserve explicitly registered `glm-5-*` variants like `glm-5-turbo` instead of intercepting them with the generic GLM-5 forward-compat shim. (#48185) Thanks @haoyu-haoyu.
- Security: preserve restrictive plugin-only tool allowlists, require owner access for `/allowlist add` and `/allowlist remove`, fail closed when `before_tool_call` hooks crash, block browser SSRF redirect bypasses earlier, and keep non-interactive auth-choice inference scoped to bundled and already-trusted plugins. (#58476, #59836, #59822, #58771, #59120) Thanks @eleqtrizit and @pgondhi987.
- Slack: route live DM replies back to the concrete inbound DM channel while keeping persisted routing metadata user-scoped, so normal assistant replies stop disappearing when pairing and system messages still arrive. (#59030) Thanks @afurm.
- Status/cache: restore `cacheRead` and `cacheWrite` in transcript fallback so `/status` keeps showing cache hit percentages when session logs are the only complete usage source. (#59247) Thanks @stuartsy.
- Status/usage: let `/status` and `session_status` fall back to transcript token totals when the session meta store stayed at zero, so LM Studio, Ollama, DashScope, and similar OpenAI-compatible providers stop showing `Context: 0/...`. (#55041) Thanks @jjjojoj.
- Synology Chat/security: default low-level HTTPS helper TLS verification to on so helper/API defaults match the shipped safe account default, and only explicit `allowInsecureSsl: true` opts out.
- Synology Chat/security: route webhook token comparison through the shared constant-time secret helper for consistency with other bundled plugins.
- Tasks/maintenance: reconcile stale cron and chat-backed CLI task rows against live cron-job and agent-run ownership instead of treating any persisted session key as proof that the task is still running. (#60310) Thanks @lml2468.
- Telegram/local Bot API: honor `channels.telegram.apiRoot` for buffered media downloads, add `channels.telegram.network.dangerouslyAllowPrivateNetwork` for trusted fake-IP setups, and require `channels.telegram.trustedLocalFileRoots` before reading absolute Bot API `file_path` values. (#59544, #60705) Thanks @SARAMALI15792 and @obviyus.
- Telegram/native command menu: trim long menu descriptions before dropping commands so sub-100 command sets can still fit Telegram's payload budget and keep more `/` entries visible. (#61129) Thanks @neeravmakwana.
- Telegram/reasoning: only create a Telegram reasoning preview lane when the session is explicitly `reasoning:stream`, so hidden `<think>` traces from streamed replies stop surfacing as chat previews on normal sessions. Thanks @vincentkoc.
- Telegram: fix current-model checks in the model picker, HTML-format non-default `/model` confirmations, explicit topic replies, persisted reaction ownership across restarts, caption-media placeholder and `file_id` preservation on download failure, and upgraded-install inbound image reads. (#60384, #60042, #59634, #59207, #59948, #59971) Thanks @sfuminya, @GitZhangChi, @dashhuang, @samzong, @v1p0r, and @neeravmakwana.
- Telegram: restore DM voice-note preflight transcription so direct-message audio stops arriving as raw `<media:audio>` placeholders. (#61008) Thanks @manueltarouca.
- Tools/web_search (Kimi): when `tools.web.search.kimi.baseUrl` is unset, inherit native Moonshot chat `baseUrl` (`.ai` / `.cn`) so China console keys authenticate on the same host as chat. Fixes #44851. (#56769) Thanks @tonga54.
- Update/npm: prefer the npm binary that owns the installed global OpenClaw prefix so mixed Homebrew-plus-nvm setups update the right install. (#60153) Thanks @jayeshp19.
- Usage/MiniMax: invert remaining-style `usage_percent` fields when MiniMax reports only remaining percentage data, so usage bars stop showing nearly-full remaining quota as nearly-exhausted usage. (#60254) Thanks @jwchmodx.
- Voice-call/OpenAI: pass full plugin config into realtime transcription provider resolution so streaming calls can discover the bundled OpenAI realtime transcription provider again. Fixes #60936. Thanks @sliekens and @vincentkoc.
- WhatsApp: restore `channels.whatsapp.blockStreaming` and reset watchdog timeouts after reconnect so quiet chats stop falling into reconnect loops. (#60007, #60069) Thanks @MonkeyLeeT and @mcaxtr.
- Windows/restart: fall back to the installed Startup-entry launcher when the scheduled task was never registered, so `/restart` can relaunch the gateway on Windows setups where `schtasks` install fell back during onboarding. (#58943) Thanks @imechZhangLY.
- Sandbox/security: block credential-path binds even when sandbox home paths resolve through canonical aliases, so agent containers cannot mount user secret stores through alternate home-directory paths. (#59157) Thanks @eleqtrizit.
- Gateway/Windows scheduled tasks: preserve Task Scheduler settings on reinstall, fail loud when Scheduled Task `/Run` does not start, and report fast failed restarts with the actual elapsed time instead of a fake 60s timeout. (#59335) Thanks @tmimmanuel.
- Control UI/model picker: preserve already-qualified `provider/model` refs from the server so models whose ids already contain slashes stop being double-prefixed and remapped to the wrong provider. (#49874) Thanks @ShionEria.
- Models/selection: resolve bare model ids in session model switches against the configured allowlist before falling back to the current session provider, so Control UI model picks stop drifting into `google/k2p5` and similar wrong-provider refs. (#51580) Thanks @honwee.
## 2026.4.1-beta.1
- Providers/transport policy: centralize request auth, proxy, TLS, and header shaping across shared HTTP, stream, and websocket paths, block insecure TLS/runtime transport overrides, and keep proxy-hop TLS separate from target mTLS settings. (#59682) Thanks @vincentkoc.
- Providers/OpenRouter: gate documented OpenRouter attribution to native OpenRouter endpoints or the default route so custom proxy base URLs do not inherit OpenRouter request headers.
- Providers/Copilot: classify native GitHub Copilot API hosts in the shared provider endpoint resolver and harden token-derived proxy endpoint parsing so Copilot base URL routing stays centralized and fails closed on malformed hints. (#59644) Thanks @vincentkoc.
- Providers/streaming headers: centralize default and attribution header merging across OpenAI websocket, embedded-runner, and proxy stream paths so provider-specific headers stay consistent and caller overrides only win where intended. (#59542) Thanks @vincentkoc.
- Providers/media HTTP: centralize base URL normalization, default auth/header injection, and explicit header override handling across shared OpenAI-compatible audio, Deepgram audio, Gemini media/image, and Moonshot video request paths. (#59469) Thanks @vincentkoc.
@@ -90,6 +226,7 @@ Docs: https://docs.openclaw.ai
- Image generation/providers: stop inferring private-network access from configured OpenAI, MiniMax, and fal image base URLs, and cap shared HTTP error-body reads so hostile or misconfigured endpoints fail closed without relaxing SSRF policy or buffering unbounded error payloads. Thanks @vincentkoc.
- Browser/host inspection: keep static Chrome inspection helpers out of the activated browser runtime so `openclaw doctor browser` and related checks do not eagerly load the bundled browser plugin. (#59471) Thanks @vincentkoc.
- Browser/CDP: normalize trailing-dot localhost absolute-form hosts before loopback checks so remote CDP websocket URLs like `ws://localhost.:...` rewrite back to the configured remote host. (#59236) Thanks @mappel-nv.
- Browser/attach-only profiles: disconnect cached Playwright CDP sessions when stopping attach-only or remote CDP profiles, while still reporting never-started local managed profiles as not stopped. (#60097) Thanks @pedh.
- Agents/output sanitization: strip namespaced `antml:thinking` blocks from user-visible text so Anthropic-style internal monologue tags do not leak into replies. (#59550) Thanks @obviyus.
- Kimi Coding/tools: normalize Anthropic tool payloads into the OpenAI-compatible function shape Kimi Coding expects so tool calls stop losing required arguments. (#59440) Thanks @obviyus.
- Image tool/paths: resolve relative local media paths against the agent `workspaceDir` instead of `process.cwd()` so inputs like `inbox/receipt.png` pass the local-path allowlist reliably. (#57222) Thanks Priyansh Gupta.
- Telegram/exec approvals: fall back to the origin session key for async approval followups and keep resume-failure status delivery sanitized so Telegram followups still land without leaking raw exec metadata. (#59351) Thanks @seonang.
- Node-host/exec approvals: bind `pnpm dlx` invocations through the approval planner's mutable-script path so the effective runtime command is resolved for approval instead of being left unbound. (#58374)
- Exec/node hosts: stop forwarding the gateway workspace cwd to remote node exec when no workdir was explicitly requested, so cross-platform node approvals fall back to the node default cwd instead of failing with `SYSTEM_RUN_DENIED`. (#58977) Thanks @Starhappysh.
- TUI/chat: keep pending local sends visible and reconciled across history reloads, make busy/error recovery clearer through fallback and terminal-error paths, and reclaim transcript width for long links and paths. (#59800) Thanks @vincentkoc.
- Exec approvals/channels: decouple initiating-surface approval availability from native delivery enablement so Telegram, Slack, and Discord still expose approvals when approvers exist and native target routing is configured separately. (#59776) Thanks @joelnishanth.
-Plugins/runtime: reuse compatible active registries for `web_search` and `web_fetch` provider snapshot resolution so repeated runtime reads do not re-import the same bundled plugin set on each agent message. Related #48380.
-Plugins/OpenAI: enable reference-image edits for `gpt-image-1` by routing edit calls to `/images/edits` with multipart image uploads, and update image-generation capability/docs metadata accordingly.
-Android/assistant: keep queued App Actions prompts pending when auto-send enqueue is rejected, so transient chat-health drops do not silently lose the assistant request. Thanks @obviyus.
-Agents/tools: include value-shape hints in missing-parameter tool errors so dropped, empty-string, and wrong-type write payloads are easier to diagnose from logs. (#55317) Thanks @priyansh19.
-Plugins/Google: separate OAuth CSRF state from PKCE code verifier during Gemini browser sign-in so state validation and token exchange use independent values. (#59116) Thanks @eleqtrizit.
-Plugins/browser: block SSRF redirect bypass by installing a real-time Playwright route handler before `page.goto()` so navigation to private/internal IPs is intercepted and aborted mid-redirect instead of checked post-hoc. (#58771) Thanks @pgondhi987.
-Gateway/connect: omit admin-scoped config and auth metadata from lower-privilege `hello-ok` snapshots while preserving those fields for admin reconnects. (#58469) Thanks @eleqtrizit.
-iOS/canvas: restrict A2UI bridge trust to the bundled scaffold and exact capability-backed remote canvas URLs, so generic `canvas.navigate` and `canvas.present` loads no longer gain action-dispatch authority. (#58471) Thanks @eleqtrizit.
- Android/gateway: require TLS for non-loopback remote gateway endpoints while still allowing local loopback and emulator cleartext setup flows. (#58475) Thanks @eleqtrizit.
-Zalo/webhook: scope replay-dedupe cache key to path and account using `JSON.stringify` so multi-account deployments do not silently drop events due to cross-account cache poisoning. (#59387) Thanks @pgondhi987.
- Exec/Windows: hide transient console windows for `runExec` and `runCommandWithTimeout` child-process launches, matching other Windows exec paths and stopping visible shell flashes during tool runs. (#59466) Thanks @lawrence3699.
- Exec/Windows: reject malformed drive-less rooted executable paths like `:\Users\...` so approval and allowlist candidate resolution no longer treat them as cwd-relative commands. (#58040) Thanks @SnowSky1.
- Exec/preflight: fail closed on complex interpreter invocations that would otherwise skip script-content validation, and correctly inspect quoted script paths before host execution. Thanks @pgondhi987.
- Exec/Windows: include Windows-compatible env override keys like `ProgramFiles(x86)` in system-run approval binding so changed approved values are rejected instead of silently passing unbound. (#59182) Thanks @pgondhi987.
- ACP/Windows spawn: fail closed on unresolved `.cmd` and `.bat` OpenClaw wrappers unless a caller explicitly opts into shell fallback, so Windows ACP launches do not silently drop into shell-mediated execution when wrapper unwrapping fails. (#58436) Thanks @eleqtrizit.
- Exec/Windows: prefer strict-inline-eval denial over generic allowlist prompts for interpreter carriers, while keeping persisted Windows allow-always approvals argv-bound. (#59780) Thanks @luoyanglang.
-Agents/logging: keep orphaned-user transcript repair warnings focused on interactive runs, and downgrade background-trigger repairs (`heartbeat`, `cron`, `memory`, `overflow`) to debug logs to reduce false-alarm gateway noise.
-Gateway/node pairing: require `operator.pairing` for node approvals end-to-end, while still requiring `operator.write` or `operator.admin` when the pending node commands need those higher scopes. (#60461) Thanks @eleqtrizit.
-Providers/OpenRouter: gate Anthropic prompt-cache `cache_control` markers to native/default OpenRouter routes and preserve them for native OpenRouter hosts behind custom provider ids. Thanks @vincentkoc.
-Browser/CDP: validate both initial and discovered CDP websocket endpoints before connect so strict SSRF policy blocks cross-host pivots and direct websocket targets. (#60469) Thanks @eleqtrizit.
-Browser/profiles: reject remote browser profile `cdpUrl` values that violate strict SSRF policy before saving config, with clearer validation errors for blocked endpoints. (#60477) Thanks @eleqtrizit.
-Browser/screenshots: stop sending `fromSurface: false` on CDP screenshots so managed Chrome 146+ browsers can capture images again. (#60682) Thanks @mvanhorn.
-Mattermost/slash commands: harden native slash-command callback token validation to use constant-time secret comparison, matching the existing interaction-token path.
-Control UI/mobile chat: reduce narrow-screen overflow by shrinking the chat pane minimum width, removing extra mobile padding, widening message groups, and hiding avatars on very small screens. (#60220) Thanks @macdao.
- Android/Talk Mode: route spoken replies through `talk.speak`, keep compressed playback cleanup deterministic, and fall back to local TTS for legacy gateways that omit Talk error reasons. (#60954) Thanks @obviyus.
-Android/Talk Mode: keep reply-speaker routing and teardown behavior aligned with the new remote playback path. (#60954) Thanks @MKV21.
- Telegram/local Bot API: preserve media MIME types for absolute-path downloads so local audio files still trigger transcription and other MIME-based handling. (#54603) Thanks @jzakirov
- Channels/WhatsApp: pass inbound message timestamp to model context so the AI can see when WhatsApp messages were sent. (#58590) Thanks @Maninae
- QQBot/voice: lazy-load `silk-wasm` in `audio-convert.ts` so qqbot still starts when the optional voice dependency is missing, while voice encode/decode degrades gracefully instead of crashing at module load time. (#58829) Thanks @WideLee.
- WhatsApp/groups: fix bot waking up on self-number quoted replies in groups with `selfChatMode` enabled. (#60148) Thanks @lurebat
- Device pairing: require `operator.pairing` or `operator.admin` for internal `/pair` setup-code, QR, and cleanup commands so lower-privilege gateway callers cannot mint or revoke pairing bootstrap material. (#60491) Thanks @eleqtrizit.
- Agents/failover: unify structured and raw provider error classification so provider-specific `400`/`422` payloads no longer get forced into generic format failures before retry, billing, or compaction logic can inspect them. (#58856) Thanks @aaron-he-zhu.
- Auth profiles/store: coerce misplaced SecretRef objects out of plaintext `key` and `token` fields during store load so agents without ACP runtime stop crashing on `.trim()` after upgrade. (#58923) Thanks @openperf.
- ACPX/runtime: repair `queue owner unavailable` session recovery by replacing dead named sessions and resuming the backend session when ACPX exposes a stable session id, so the first ACP prompt no longer inherits a dead handle. (#58669) Thanks @neeravmakwana
- ACPX/runtime: retry dead-session queue-owner repair without `--resume-session` when the reported ACPX session id is stale, so recovery still creates a fresh named session instead of failing session init. Thanks @obviyus.
- Tools/web_search (Kimi): replay native Moonshot `$web_search` arguments verbatim, disable thinking for `kimi-k2.5`, and add Moonshot region/model setup prompts so bundled Kimi web search works again. (#59356) Thanks @Innocent-children.
- Plugins/Matrix: encrypt E2EE image thumbnails with `thumbnail_file` while keeping unencrypted-room previews on `thumbnail_url`, so encrypted Matrix image events keep thumbnail metadata without leaking plaintext previews. (#54711) thanks @frischeDaten.
- Telegram/forum topics: keep native `/new` and `/reset` routed to the active topic by preserving the topic target on forum-thread command context. (#35963)
- Status/port diagnostics: treat single-process dual-stack loopback gateway listeners as healthy in `openclaw status --all`, suppressing false "port already in use" conflict warnings. (#53398) Thanks @DanWebb1949.
- CLI/Docker: treat loopback private-host CLI gateway connects as local for silent pairing auto-approval, while keeping remote backend and public-host CLI connects behind pairing. (#55113) Thanks @sar618.
- Feishu/topic threads: fetch full thread context, including prior bot replies, when starting a topic-thread session so follow-up turns in Feishu topics keep the right conversation state. (#45254) Thanks @Coobiw.
- Feishu/media: keep native image, file, audio, and video/media handling aligned across outbound sends, inbound downloads, thread replies, directory/action aliases, and capability docs so unsupported areas are explicit instead of implied. (#47968) Thanks @Takhoffman.
- Feishu/webhooks: harden signed webhook verification to use constant-time signature comparison and keep malformed short signatures fail-closed in webhook E2E coverage.
- Ollama/tool replay: deserialize stringified tool-call arguments on native and OpenAI-compatible Ollama paths while preserving unsafe integer ids across round-trips. (#52253) Thanks @Adam-Researchh.
- WhatsApp/reconnect: restore the append recency filter in the extension inbox monitor and handle protobuf `Long` timestamps correctly, so fresh post-reconnect append messages are processed while stale history sync stays suppressed. (#42588) Thanks @MonkeyLeeT.
- WhatsApp/login: wait for pending creds writes before reopening after Baileys `515` pairing restarts in both QR login and `channels login` flows, and keep the restart coverage pinned to the real wrapped error shape plus per-account creds queues. (#27910) Thanks @asyncjason.
- Telegram/message send: forward `--force-document` through the `sendPayload` path as well as `sendMedia`, so Telegram payload sends with `channelData` keep uploading images as documents instead of silently falling back to compressed photo sends. (#47119) Thanks @thepagent.
- Telegram/message chunking: preserve spaces, paragraph separators, and word boundaries when HTML overflow rechunking splits formatted replies. (#47274) Thanks @obviyus.
- Z.AI/onboarding: detect a working default model even for explicit `zai-coding-*` endpoint choices, so Coding Plan setup can keep the selected endpoint while defaulting to `glm-5` when available or `glm-4.7` as fallback. (#45969) Thanks @obviyus.
4.**Test/CI-only PRs for known `main` failures** → Don't open a PR. The Maintainer team is already tracking those failures, and PRs that only tweak tests or CI to chase them will be closed unless they are required to validate a new fix.
Text("Paste setup code or enter host/port manually.",style=onboardingCaption1Style,color=onboardingTextSecondary)
Text("Paste setup code or enter host/port manually. Private LAN ws:// is supported; Tailscale/public hosts need wss://.",style=onboardingCaption1Style,color=onboardingTextSecondary)
This folder stores **generated** and **config** files for documentation translations.
This folder stores translation config for the source docs repo.
## Files
Generated locale trees and live translation memory now live in the publish repo:
-`glossary.<lang>.json` — preferred term mappings (used in prompt guidance).
-`<lang>.tm.jsonl` — translation memory (cache) keyed by workflow + model + text hash.
-repo: `openclaw/docs`
-local checkout: `~/Projects/openclaw-docs`
## Source of truth
- English docs are authored in `openclaw/openclaw`.
- The source docs tree lives under `docs/`.
- The source repo no longer keeps committed generated locale trees such as `docs/zh-CN/**`, `docs/ja-JP/**`, `docs/es/**`, `docs/pt-BR/**`, `docs/ko/**`, `docs/de/**`, `docs/fr/**`, or `docs/ar/**`.
## End-to-end flow
1. Edit English docs in `openclaw/openclaw`.
2. Push to `main`.
3.`openclaw/openclaw/.github/workflows/docs-sync-publish.yml` mirrors the docs tree into `openclaw/docs`.
4. The sync script rewrites the publish `docs/docs.json` so the generated locale picker blocks exist there even though they are no longer committed in the source repo.
5.`openclaw/docs/.github/workflows/translate-zh-cn.yml` refreshes `docs/zh-CN/**` once a day, on demand, and after source-repo release dispatches.
6.`openclaw/docs/.github/workflows/translate-ja-jp.yml` does the same for `docs/ja-JP/**`.
7.`openclaw/docs/.github/workflows/translate-es.yml`, `translate-pt-br.yml`, `translate-ko.yml`, `translate-de.yml`, `translate-fr.yml`, and `translate-ar.yml` do the same for `docs/es/**`, `docs/pt-BR/**`, `docs/ko/**`, `docs/de/**`, `docs/fr/**`, and `docs/ar/**`.
## Why the split exists
- Keep generated locale output out of the main product repo.
- Keep Mintlify on a single published docs tree.
- Preserve the built-in language switcher by letting the publish repo own generated locale trees.
## Files in this folder
-`glossary.<lang>.json` — preferred term mappings used as prompt guidance.
-`ar-navigation.json`, `de-navigation.json`, `es-navigation.json`, `fr-navigation.json`, `ja-navigation.json`, `ko-navigation.json`, `pt-BR-navigation.json`, `zh-Hans-navigation.json` — Mintlify locale picker blocks reinserted into the publish repo during sync.
-`<lang>.tm.jsonl` — translation memory keyed by workflow + model + text hash.
In this repo, generated locale TM files such as `docs/.i18n/zh-CN.tm.jsonl`, `docs/.i18n/ja-JP.tm.jsonl`, `docs/.i18n/es.tm.jsonl`, `docs/.i18n/pt-BR.tm.jsonl`, `docs/.i18n/ko.tm.jsonl`, `docs/.i18n/de.tm.jsonl`, `docs/.i18n/fr.tm.jsonl`, and `docs/.i18n/ar.tm.jsonl` are intentionally no longer committed.
## Glossary format
@@ -14,9 +44,7 @@ This folder stores **generated** and **config** files for documentation translat
```json
{
"source":"troubleshooting",
"target":"故障排除",
"ignore_case":true,
"whole_word":false
"target":"故障排除"
}
```
@@ -25,7 +53,19 @@ Fields:
-`source`: English (or source) phrase to prefer.
-`target`: preferred translation output.
## Notes
## Translation mechanics
-Glossary entries are passed to the model as **prompt guidance** (no deterministic rewrites).
-The translation memory is updated by `scripts/docs-i18n`.
-`scripts/docs-i18n` still owns translation generation.
-Doc mode writes `x-i18n.source_hash` into each translated page.
- Each publish workflow precomputes a pending file list by comparing the current English source hash to the stored locale `x-i18n.source_hash`.
- If the pending count is `0`, the expensive translation step is skipped entirely.
- If there are pending files, the workflow translates only those files.
- The publish workflow retries transient model-format failures, but unchanged files stay skipped because the same hash check runs on each retry.
- The source repo also dispatches zh-CN, ja-JP, es, pt-BR, ko, de, fr, and ar refreshes after published GitHub releases so release docs can catch up without waiting for the daily cron.
## Operational notes
- Sync metadata is written to `.openclaw-sync/source.json` in the publish repo.
- Jobs persist at `~/.openclaw/cron/jobs.json` so restarts do not lose schedules.
- All cron executions create [background task](/automation/tasks) records.
- One-shot jobs (`--at`) auto-delete after success by default.
- Isolated cron runs best-effort close tracked browser tabs/processes for their `cron:<jobId>` session when the run completes, so detached browser automation does not leave orphaned processes behind.
- Isolated cron runs also guard against stale acknowledgement replies. If the
first result is just an interim status update (`on it`, `pulling everything
together`, and similar hints) and no descendant subagent run is still
responsible for the final answer, OpenClaw re-prompts once for the actual
result before delivery.
Task reconciliation for cron is runtime-owned: an active cron task stays live while the
cron runtime still tracks that job as running, even if an old child session row still exists.
Once the runtime stops owning the job and the 5-minute grace window expires, maintenance can
mark the task `lost`.
## Schedule types
@@ -60,6 +71,12 @@ Recurring top-of-hour expressions are automatically staggered by up to 5 minutes
**Main session** jobs enqueue a system event and optionally wake the heartbeat (`--wake now` or `--wake next-heartbeat`). **Isolated** jobs run a dedicated agent turn with a fresh session. **Custom sessions** (`session:xxx`) persist context across runs, enabling workflows like daily standups that build on previous summaries.
For isolated jobs, runtime teardown now includes best-effort browser cleanup for that cron session. Cleanup failures are ignored so the actual cron result still wins.
When isolated cron runs orchestrate subagents, delivery also prefers the final
descendant output over stale parent interim text. If descendants are still
running, OpenClaw suppresses that partial parent update instead of announcing it.
### Payload options for isolated jobs
-`--message`: prompt text (required for isolated)
@@ -67,6 +84,29 @@ Recurring top-of-hour expressions are automatically staggered by up to 5 minutes
-`--tools exec,read`: restrict which tools the job can use
`--model` uses the selected allowed model for that job. If the requested model
is not allowed, cron logs a warning and falls back to the job's agent/default
model selection instead. Configured fallback chains still apply, but a plain
model override with no explicit per-job fallback list no longer appends the
agent primary as a hidden extra retry target.
Model-selection precedence for isolated jobs is:
1. Gmail hook model override (when the run came from Gmail and that override is allowed)
2. Per-job payload `model`
3. Stored cron session model override
4. Agent/default model selection
Fast mode follows the resolved live selection too. If the selected model config
has `params.fastMode`, isolated cron uses that by default. A stored session
`fastMode` override still wins over config in either direction.
If an isolated run hits a live model-switch handoff, cron retries with the
switched provider/model and persists that live selection before retrying. When
the switch also carries a new auth profile, cron persists that auth profile
override too. Retries are bounded: after the initial attempt plus 2 switch
retries, cron aborts instead of looping forever.
## Delivery and output
| Mode | What happens |
@@ -77,6 +117,22 @@ Recurring top-of-hour expressions are automatically staggered by up to 5 minutes
Use `--announce --channel telegram --to "-1001234567890"` for channel delivery. For Telegram forum topics, use `-1001234567890:topic:123`. Slack/Discord/Mattermost targets should use explicit prefixes (`channel:<id>`, `user:<id>`).
For cron-owned isolated jobs, the runner owns the final delivery path. The
agent is prompted to return a plain-text summary, and that summary is then sent
through `announce`, `webhook`, or kept internal for `none`. `--no-deliver`
does not hand delivery back to the agent; it keeps the run internal.
If the original task explicitly says to message some external recipient, the
agent should note who/where that message should go in its output instead of
trying to send it directly.
Failure notifications follow a separate destination path:
-`cron.failureDestination` sets a global default for failure notifications.
-`job.delivery.failureDestination` overrides that per job.
- If neither is set and the job already delivers via `announce`, failure notifications now fall back to that primary announce target.
-`delivery.failureDestination` is only supported on `sessionTarget="isolated"` jobs unless the primary delivery mode is `webhook`.
Heartbeat is a periodic main-session turn (default every 30 minutes). It batches multiple checks (inbox, calendar, notifications) in one agent turn with full session context. Heartbeat turns do not create task records. Use `HEARTBEAT.md`to define what the agent checks.
Heartbeat is a periodic main-session turn (default every 30 minutes). It batches multiple checks (inbox, calendar, notifications) in one agent turn with full session context. Heartbeat turns do not create task records. Use `HEARTBEAT.md`for a small checklist, or a `tasks:` block when you want due-only periodic checks inside heartbeat itself. Empty heartbeat files skip as `empty-heartbeat-file`; due-only task mode skips as `no-tasks-due`.
See [Heartbeat](/gateway/heartbeat).
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.