Review on #98102: the durable-retry classifier treated every MediaFetchError
`fetch_failed` as retryable, but `fetch_failed` also covers permanent failures
(SSRF/guard denials, local Bot API path/read errors). On spooled replay those
would requeue forever with the user-facing warning suppressed.
Move the policy to its owner: add `isDurablyRetryableMediaFetchError` to
`src/media/fetch.ts`, defined as `shouldRetryMediaFetch` (the canonical in-loop
transient policy) plus shutdown/abort `fetch_failed`. A restart-window abort is
the primary inbound-media loss vector and is recoverable on replay, whereas
in-loop retry mid-shutdown is futile; permanent `fetch_failed`, other 4xx, and
size limits stay non-retryable so they cannot loop in the spool.
Telegram reuses the shared helper instead of a local classifier
(`isAbortError`/`isTransientNetworkError` live in core; the classification must
not drift). Adds permanent-`fetch_failed` regression coverage at the unit and
behavioral level. Media-group partial delivery (#55216) unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A single-message inbound media fetch failure was handled best-effort: warn
and let the update complete, so bot-core acked the spooled update. During a
restart window the warning send also failed and the document was silently
lost with no durable retry.
Classify transient media fetch failures (network/abort fetch_failed and
408/429/5xx HTTP) and record them as failed-retryable, so spooled-replay
updates are kept and re-driven by the ingress spool with backoff instead of
acked. Live updates still ack and warn; permanent failures (size limit, 4xx)
stay best-effort. The retry notice is suppressed on spooled replay to avoid
repeating it on each replay.
The durable-retry classifier deliberately diverges from media/fetch.ts
shouldRetryMediaFetch on shutdown aborts: in-loop retry of a shutdown abort
is futile, but durable re-spool must retry it -- that abort is the primary
restart-window loss vector. Media-group partial delivery (#55216) is left
unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(imessage): shed emoji anywhere in poll-vote echo match
iMessage native poll options carry a trailing emoji ("Lobster 🦞 ") while the
agent echoes its vote with a leading one ("🦞 Lobster.") under message_tool_only
reply mode. normalizePollEchoText stripped only a leading emoji prefix, so
"lobster 🦞" never matched "lobster" and the redundant text leaked past the
poll_vote_echo guard (shipped in #98421).
Shed emoji anywhere across every class: pictographic, regional-indicator flags
(🇺🇸), subdivision-flag tag chars, ZWJ, skin-tone, and keycap sequences cleared
as a unit (so "1️⃣" -> "" not "1", while a plain "1"/"#"/"*" survives). Internal
punctuation stays so C#/C++/Node.js remain distinct; emoji-only labels normalize
to empty and the guard's existing empty-option check fails open. Re-exports
normalizePollEchoText with a focused unit test across every emoji class plus a
message-tool integration test for the emoji-suffixed-option case.
* fix(imessage): make poll-echo emoji regex lint-safe
ClawSweeper/CI check-lint flagged no-misleading-character-class: combining
marks (VS16, U+20E3), ZWJ, and skin-tone modifiers are misleading inside a
single character class. Rewrite the emoji strip as an alternation using
property escapes (\p{Regional_Indicator}, \p{Emoji_Modifier}) plus standalone
atoms so the same normalization behavior passes lint. Behavior unchanged;
poll-echo unit + integration tests still green.
* fix(imessage): preserve poll option emoji identity
---------
Co-authored-by: Omar Shahine <10343873+omarshahine@users.noreply.github.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Adds an `on-exit` cron schedule kind: a job fires once when a watched command/process
exits, via gateway ProcessSupervisor exit watchers. Covers CLI (`--on-exit`/`--on-exit-cwd`),
tool/protocol schema, RPC list-filter, Control UI + macOS read-only display, SQLite
round-trip, and origin-aware wake routing. Restart-safe one-shot (persists completion
before firing); platform-aware shell; bounded watched-command execution.
Squashed from 22 iterative commits for a clean rebase onto current main.
The built-in default footer (added in #92657) only defined a `discord`
surface. Telegram and every other surface fall through to `output.default`,
which was authored for a wide terminal: space-separated badges, a leading
space-less join onto the message body, and ↕️ token + 🗄 cache% segments.
On a narrow Telegram screen this wrapped at arbitrary spaces (often mid-meter
or between 📚 and its bar) and welded onto the end of the message text.
Make `default` read well on a phone:
- Leading \n so the footer sits on its own line below the body.
- Tight badges (no separators between model/flags/reasoning/fast).
- NBSP before "|" and 💰; 📚 glued hard to the meter so it can't wrap there.
- Single regular space after "|" as the only sanctioned wrap point.
- Drop ↕️ tokens and 🗄 cache% from the default — noise for most users
(still available to anyone via a custom messages.usageTemplate).
Discord surface trimmed to match for consistency (keeps its -# dim header).
ClawSweeper review flagged that the new suppressReply test replaced the
existing direct-message binding regression test instead of being added
beside it. Restore the DM binding test to maintain coverage for the
plugin command dispatch surface.
Adds to the type so plugin
commands that handle their own transport delivery (e.g. via Telegram Bot API
directly with retry logic, IPv4 forcing, etc.) can signal the channel adapter
to skip the fallback reply.
When a plugin command handler returns , the Telegram
native command dispatcher now:
1. Cleans up any progress placeholder
2. Returns early without sending the "No response generated. Please try again." fallback
Includes detailed JSDoc for the new flag explaining its use
for plugin commands that deliver their own responses via channel-native APIs.
Fixes#80756
* fix(embedded-agent-runner): pump async streamFn through pumpStreamWithRecovery for mid-stream error recovery
When wrapEmbeddedAgentStreamFn returns an async function (e.g. when
authStorage or resolvedApiKey is present), the stream is a Promise that
resolves to AssistantMessageEventStreamLike. The old Promise branch in
wrapAnthropicStreamWithRecovery only handled Promise rejections via
.catch(), missing {type:"error"} events that arrive after the Promise
resolves — such as Anthropic thinking-signature replay rejections.
Now the Promise branch:
- Awaits the resolved stream and runs pumpStreamWithRecovery on it,
so mid-stream error events trigger thinking-signature recovery.
- Handles Promise rejection via .then(onFulfilled, onRejected) with
the same retryStreamWithoutThinking path used by the sync branch.
- Returns a unified AssistantMessageEventStream (outer) identical to
the non-Promise branch pattern.
Tests updated to match the new return type (outer stream + .result()
instead of a bare Promise).
Fixes#95429
* fix(anthropic): remove orphaned wrapRetryStreamWithRecoveryNotification
Function is dead code — the Promise-handling was inlined into
pumpStreamWithRecovery, leaving only a recursive self-call
with no external entry point.
* fix(embedded-agent-runner): add Promise-resolved-stream regression test
* fix: use createTestStreamErrorMessage for type-correct stream error in Promise-resolved test
* test(agents): cover async thinking stream recovery
---------
Co-authored-by: lzyyzznl <lzyyzznl@users.noreply.github.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
isCodexToolResultError fail-closes any dynamic-tool result whose
details.status is absent from its non-error allowlist. get_goal returns
details.status "found" or "missing" (a successful read of the thread
goal, or its absence), neither of which was in the allowlist, so every
get_goal call was classified as an error: reported to codex as
success: false and persisted on the transcript with isError: true.
Sibling #96856 added the write-side goal statuses (created/updated) and
the accepted spawn status but missed the read-side get_goal statuses.
This adds found/missing alongside them. Genuinely failed statuses stay
fail-closed.
* fix(memory-wiki): disambiguate the reserved index page stem for synthesis and ingest
A create_synthesis or source ingest whose title slugifies to index was
written to the auto-generated <dir>/index.md. That file is the only
compiler-owned page filename: compile writes it per page group, and
compile/query/status exclude it from page scanning. So the stored fact
was both unretrievable and clobbered by the generated directory-index on
the next compile.
The collision only happens at the direct page-stem boundary: synthesis
(syntheses/<slug>.md) and ingest (sources/<slug>.md) build a bare
<dir>/<slug>.md path. The bridge, unsafe-local, and OKF callers compose
slugifyWikiSegment output into prefixed and hashed identities that can
never equal a bare index.md, and OKF already guards its own write path
via OKF_RESERVED_FILENAMES.
Add a dedicated slugifyWikiPageStem helper that escapes the reserved
index stem and route only the two direct page writers through it,
leaving slugifyWikiSegment output stable for the composed callers so
their persisted page ids and paths do not churn on upgrade. Only index
is reserved: the compiler never owns <dir>/log.md in these directories
(the wiki log lives at .openclaw-wiki/log.jsonl), so a Log title keeps
its existing direct path. Additive and deterministic, no migration.
* fix(memory-wiki): preserve reserved-page ids
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(whatsapp): thread authDir through command authorization and owner bypass for LID JID resolution
WhatsApp group commands (/new, /stop) can be ignored when Baileys reports the sender as a LID JID (@lid) instead of a phone JID (@s.whatsapp.net). The resolveWhatsAppCommandAuthorized() and isOwnerSender() functions called getSelfIdentity/getSenderIdentity without passing authDir, so the LID-to-phone reverse mapping could not happen.
Fix: thread account.authDir through both command authorization and group owner-bypass identity resolution paths so that LID JIDs are properly resolved to phone E.164 identities before owner/allowlist checks.
* fix(whatsapp): replace deprecated top-level fields with admission overrides in LID JID test
* fix(feishu): send blocks as independent messages when blockStreaming is enabled
* fix(feishu): preserve mention targets on first independently sent block
* fix(feishu): route independent block sends through chunked text delivery
Route the independent block-send path through sendChunkedTextReply
instead of calling sendMessageFeishu directly, so the configured
Feishu textChunkLimit is honored for completed blocks. Preserve
mention targets only on the first emitted block chunk and keep
sentBlockText based on the original block text for duplicate-final
suppression.
Add regression test with a small textChunkLimit that verifies a
long block is split into chunked messages and mentions only appear
on the first chunk.
* fix(feishu): preserve block delivery target
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Claude Code >= 2.1.156 tightened its PermissionResult schema so an
approved can_use_tool control_response allow branch must carry an
updatedInput record. The Claude live-session bridge answered with the
deprecated { behavior: "allow" } shape and no updatedInput, so the CLI
rejected every approval-gated native tool call (Bash, WebFetch, and
AskUserQuestion among them) with a ZodError before the tool ran.
Echo the tool input back unchanged as updatedInput on the allow branch,
matching the shape the Claude Code 2.1 binary expects. The deny branch
and all other behavior are unchanged.
Fixes#95171.
hasProviderAuthForTool/hasAuthForProvider now thread cfg and workspaceDir into resolveEnvApiKey, so provider plugins that authenticate via environment variables and are only discoverable with config are detected during image/video model auto-selection instead of being dropped as unconfigured.
