Memory promotion audit surfaces (MEMORY.md annotations, dreaming ranked/applied logs, memory promote listing, and promote-explain text/JSON) now show the aggregate signal count the promotion engine actually gates on, instead of recall-only numbers that misread daily/grounded promotions as zero-signal. The candidate carries a required signalCount owned by totalSignalCountForEntry (inline recomputation removed), and the promotion-annotation contamination regex accepts both legacy and signals= annotation generations.
Fixes#87588
Thanks to @bladin for the contribution.
Co-authored-by: heichl_xydigit <1740879+bladin@users.noreply.github.com>
Co-authored-by: Ayaan Zaidi <hi@obviy.us>
* [AI] fix(memory): use truncateUtf16Safe for dreaming snippet truncation
Replace .slice(0, N) with truncateUtf16Safe() at 5 call sites in
dreaming-phases.ts so complex emoji and surrogate pairs near the
truncation boundary are not split into lone surrogates.
truncateUtf16Safe is the standard SDK helper, already used in
session-cost-usage, cron, exec-approval, and node-host modules
for the same purpose.
* [AI] fix(memory): use SDK facade import for truncateUtf16Safe, add surrogate-proof test
Replace direct @openclaw/normalization-core/utf16-slice import with
openclaw/plugin-sdk/memory-core-host-engine-foundation SDK facade,
matching manager-search.ts in the same extension.
Add surrogate-proof.test.ts that runs a full dreaming ingestion sweep
with emoji at the 280-char boundary and verifies zero lone surrogates.
* [AI] fix(memory-proof): use relative imports in surrogate-proof test
* [AI] fix(memory-proof): place emoji at post-prefix 280-char boundary, assert session corpus file
* [AI] fix(memory-proof): avoid unnecessary String callback in oxlint
* [AI] fix(memory-proof): add curly braces per eslint curly rule
* [AI] feat(proof): add standalone dreaming surrogate truncation proof script
* [AI] chore(proof): remove unused SQLite-dependent proof script
* [AI] fix(proof): correct emoji boundary placement for all 5 call sites
* [AI] fix(proof): add curly braces per eslint curly rule in proof script
* [AI] fix(proof): session emoji at actual corpus 280-char boundary
* test: streamline memory dreaming UTF-16 coverage
---------
Co-authored-by: hailory <hailory@xydigit.com>
Co-authored-by: Peter Steinberger <steipete@gmail.com>
* fix(memory-core): clamp widen-fallback kNN k to sqlite-vec 4096 limit
The widen fallback in searchVector re-runs the kNN with k = vectorCount
(the full vector table size), unclamped. sqlite-vec hard-caps `k` at 4096,
so once a memory index exceeds 4096 chunks any query that triggers the
fallback fails with "k value in knn query too large".
Clamp the widen re-query to MAX_VECTOR_KNN_K (4096). 4096 oversampled
candidates is far more than enough to feed top-N hybrid ranking, so there
is no meaningful recall loss.
* test(memory-core): cover widen-fallback kNN clamp to sqlite-vec 4096 ceiling
Add two regression tests for the searchVector widen-fallback clamp:
- asserts a >4096-vector widen re-query caps k at 4096 (out-of-range k
would raise sqlite-vec "k out of range" and kill the whole search)
- asserts a sub-ceiling widen (30) passes through untouched, proving the
clamp is a ceiling and not a fixed value
Uses a prepare-mock on the KNN/COUNT statements rather than materializing
4097+ real vectors plus the native extension; the assertion is purely on
the k bind handed to the widen re-query.
* fix(memory-core): preserve filtered recall beyond KNN cap
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(memory-core): clear daily-ingestion sqlite namespace on dreaming repair
repairDreamingArtifacts() cleared the dreaming-session-ingestion-files and
dreaming-session-ingestion-seen namespaces but not the migrated
dreaming-daily-ingestion namespace. After #92020 taught auditDreamingArtifacts()
to treat all three ingestion namespaces as ingestion state, the repair path
became asymmetric: the memory status --fix re-audit still reported
sessionIngestionExists=true from the surviving daily rows, and the daily
ingestion bookkeeping leaked past repair so daily memory files were not
re-ingested on the next sweep.
Clear DREAMING_DAILY_INGESTION_NAMESPACE alongside the session files/seen
namespaces so repair fully resets dreaming ingestion state, matching the audit.
* chore: retrigger CI for real behavior proof check
* fix(memory-core): keep daily ingestion outside session repair
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
Co-authored-by: Alix-007 <li.long15@xydigit.com>
---------
Co-authored-by: Peter Steinberger <steipete@gmail.com>
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* [AI] fix(memory): add batch completed and batch failed logs for embedding ops
The embedBatchWithRetry function logged 'batch start' but never logged
'batch completed' or 'batch failed' after the embedding batch call,
leaving operators with no post-request feedback. When batches hang or
time out, only 'batch start' appears in logs with no diagnostic signal.
Add 'batch completed' log after runEmbeddingOperationWithTimeout success
and 'batch failed' log in the catch handler. This is an observability
improvement, not a functional fix for the underlying hang (#93312).
Related to #93312
* [AI] fix(memory): use formatErrorMessage for embedding batch error log
Replace String(err) with formatErrorMessage(err) in the batch failed
catch handler to redact sensitive provider error text (e.g. API keys
and tokens embedded in error messages) before logging.
Related to #94732
---------
Co-authored-by: Vincent Koc <vincentkoc@ieee.org>
* fix(memory-core): guard qmd mcporter JSON.parse against non-JSON stdout
runQmdSearchViaMcporter parsed mcporter subprocess stdout with JSON.parse
outside the runMcporter try/catch (qmd-manager.ts:2722). A non-JSON stdout
(daemon warning bleeding onto stdout, output truncated by maxOutputChars, CLI
killed early, or flag mismatch) threw a raw SyntaxError that propagated
uncaught out of runQmdSearchViaMcporter, surfacing in agent logs as a
context-free SyntaxError with no hint of the actual mcporter failure.
Wrap JSON.parse in try/catch and throw a typed Error carrying a stdout snippet
(matching the guard pattern already used in parseListedCollections in this
file, and the recent matrix #97973 / sms #97999 / signal #98073 /
telegram-ingress #98372 JSON.parse guard series).
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(memory-core): preserve cause in qmd mcporter JSON.parse guard
Add { cause: err } to the re-thrown Error to satisfy the preserve-caught-error
lint rule; the original SyntaxError is now chained, improving diagnosability.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(memory-core): redact raw stdout from qmd mcporter error (security-boundary)
ClawSweeper flagged that the previous error message exposed raw mcporter
stdout (first 200 chars) before session visibility filtering, which could
leak sensitive content. Drop the stdout preview from the thrown message;
keep the original SyntaxError as `cause` for diagnostics so the parse-failure
reason is still reachable without surfacing unfiltered subprocess output.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(memory-core): keep qmd mcporter error message generic (no raw stdout leak)
The SyntaxError thrown by JSON.parse embeds a snippet of the raw input in its
message (e.g. Unexpected token '<', then the raw bytes). Including that
SyntaxError message in the thrown Error would surface unfiltered mcporter
stdout before session visibility filtering. Drop the parse-error message from
the thrown Error; keep the original SyntaxError as cause for developer
diagnostics only.
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(memory-core): keep qmd mcporter cause generic (no raw stdout leak via formatErrorMessage)
formatErrorMessage walks the .cause chain into the user-visible path.
Keeping the JSON.parse SyntaxError on .cause leaked its embedded raw
stdout snippet through formatErrorMessage even with a generic message.
Give the cause a generic message too; the raw snippet no longer reaches
the user.
Co-Authored-By: Claude <noreply@anthropic.com>
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix(cron): reject sub-millisecond durations
* fix(skill-workshop): preserve proposal terminal newline
Preserve proposal_content exactly at the agent tool boundary and make
renderProposalMarkdown defensively emit a terminal newline.
Add focused regressions for the tool write path and markdown renderer.
* fix(skill-workshop): reject blank raw proposal content
* fix: treat empty-string optional integer tool params as unset
Optional positive-integer tool params (e.g. Telegram replyTo/threadId)
threw ToolInputError when a tool-calling model populated them with an
empty-string or whitespace-only default. Those defaults carry no value,
so readPositiveIntegerParam/readNonNegativeIntegerParam now treat a
blank string as unset (undefined) instead of throwing, while still
rejecting genuinely invalid present values (0, "42.5", "-3"). This
prevents silent message-delivery failures when models emit empty
routing-param defaults. Adds unit tests covering blank vs invalid.
* fix(gateway): log start session persistence failures
The gateway session-lifecycle "start" event persistence
(persistGatewaySessionLifecycleEvent, which surfaces write failures via
requireWriteSuccess) was fired as void ...catch(() => undefined), swallowing
the rejection with zero logging. A failed start-marker write silently dropped
the run's start record from restart-recovery accounting, with no
operator-visible trace.
The sibling terminal-phase catch already logs this since #97839; the start
path was the unfixed sibling. Mirror that fix: log the swallowed start-phase
persistence failure with the same redacted message shape via formatForLog,
keeping the fire-and-forget semantics unchanged. Adds a focused regression
test asserting the log fires on start-persist rejection.
* fix(memory): report close-time pending work failures
* fix(shared): return "" from sliceUtf16Safe when end <= start, matching native .slice
sliceUtf16Safe silently swapped reversed bounds (to < from) instead of
returning "" like String.prototype.slice, creating a subtle footgun for
callers with dynamic start/end pairs.
Caller scan across src/, extensions/, and packages/ confirmed no production
code relies on the old swap behavior — all callers use (text, 0, N) or
(text, -N) forms only.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
* fix(plugins): require plugin manifest in npm verifier
* fix(android): propagate CancellationException in CameraHandler catch blocks
Catch (err: Throwable) swallows kotlinx.coroutines.CancellationException,
breaking structured concurrency when the coroutine scope is cancelled
during camera operations (handleList/handleSnap/handleClip).
Add CancellationException rethrow before each Throwable catch to match
the existing pattern used in GatewaySession and TalkModeManager.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* fix(android): propagate CancellationException past GatewaySession invoke boundary
* chore: sync native i18n inventory after gateway session line shift
* fix(agents): prevent native hook relay bridge race condition on renew and registration
Remove synchronous bridge record write after server.listen() that races
before the TCP server binds, and guard renew handler with server.listening
check to prevent stale relay registrations.
Closes#98650
* fix: harden small reliability fixes
Co-authored-by: cxbAsDev <cxbAsDev@users.noreply.github.com>
* docs(agents): explain buffered LSP spawn failures
* docs(agents): clarify LSP spawn timeout invariant
---------
Co-authored-by: qingminlong <qing.minlong@xydigit.com>
Co-authored-by: anyech <anyech@gmail.com>
Co-authored-by: snotty <snotty@users.noreply.github.com>
Co-authored-by: masatohoshino <g515hoshino@gmail.com>
Co-authored-by: lin-hongkuan <lin-hongkuan@users.noreply.github.com>
Co-authored-by: simon-w <weng.qimeng@xydigit.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
Co-authored-by: 宇宙熊Yzx <53250620+849261680@users.noreply.github.com>
Co-authored-by: xialonglee <li.xialong@xydigit.com>
Co-authored-by: nankingjing <1079826437@qq.com>
Co-authored-by: cxbAsDev <cxbAsDev@users.noreply.github.com>
The lineRangeOverlapsDreamingFence guard tracked insideFence state from
the marker lines but did not flag ranges that included the marker lines
themselves. A relocated promotion range that ended on a start marker,
began on an end marker, or covered only a marker line passed the guard,
and the snippet built from those raw lines carried the marker text into
MEMORY.md.
Treat marker lines as inside-fence content for range overlap purposes
and add unit + integration coverage. The integration test exercises the
real applyShortTermPromotions path: pre-patch a recall whose stored
snippet is the start-marker text produces a 'Promoted From Short-Term
Memory' entry containing the literal marker; post-patch the same input
yields applied=0 and no MEMORY.md write. Refs #80613.
* fix(memory): detect unindexed session transcripts in status mode (fixes#97814)
The status-purpose MemoryIndexManager init skips ensureSessionStartupCatchup(),
so openclaw memory status reports dirty=false while unindexed session files
exist on disk. This is a false-clean state: the operator sees no backlog,
but memory search silently misses unindexed transcripts.
Fix: run markSessionStartupCatchupDirtyFiles() in status mode. It checks
for on-disk session files without corresponding memory_index_sources rows
and sets sessionsDirty=true if found, without triggering a full sync.
The full catchup+sync runs on the next non-transient manager cycle (CLI
sync or normal runtime init).
Co-Authored-By: Claude <noreply@anthropic.com>
* fix(memory): await status dirty detection before status() reads the flag (fixes#97814)
The status-purpose MemoryIndexManager init skips ensureSessionStartupCatchup(),
so openclaw memory status reports dirty=false while unindexed session files
exist on disk. This is a false-clean state: the operator sees no backlog,
but memory search silently misses unindexed transcripts.
Fix: move status-mode markSessionStartupCatchupDirtyFiles() from the
constructor (fire-and-forget via void) into the create() factory method
where it is awaited. This guarantees sessionsDirty is set before any
caller reads manager.status(). The detection does NOT trigger a sync,
so status remains lightweight.
Review: verified manually — unit tests 21/21 pass, compilation 0 errors.
* test(memory): add regression test for status-purpose dirty detection (fixes#97814)
Adds a regression test that exercises the actual purpose:'status' manager
flow: create a session file without index row, create status manager, verify
status().dirty is true. This addresses the ClawSweeper P1 finding requesting
a test that exercises the real status path, not only the protected helper.
Also fixes the timing issue: move dirty detection from constructor (void)
to create() factory (await), guaranteeing sessionsDirty is set before any
caller reads manager.status().
---------
Co-authored-by: Claude <noreply@anthropic.com>
* fix(memory): gate dreaming config changes
* fix(memory): document dreaming privilege gate
Explain that persistent dreaming toggles require channel owner status or Gateway admin scope, while status and help remain read-only.
vsearch/search are documented as vector/lexical-only modes, but
runQmdSearchViaMcporter omitted `rerank` from the query tool callArgs,
so QMD's query tool (which defaults rerank:true) ran the LLM reranker
anyway. Pass rerank:false for those modes; full "query" mode keeps it.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
