feat(i18n): German language support + fix fire-and-forget race in agent_end

[Banger455]

Summary

Fixes #393 — Memory store reports STORE-OK but recall in a new session returns RECALLED: NONE for German-speaking users.

Root cause

Three separate bugs compound into a silent data-loss scenario:

1. Missing German trigger patterns — MEMORY_TRIGGERS, AUTO_CAPTURE_EXPLICIT_REMEMBER_RE, and FORCE_RETRIEVE_PATTERNS cover English, Czech, and Chinese but have no German entries. Phrases like "Merke dir: X" fall through shouldCapture() → nothing is stored.
2. Fire-and-forget void backgroundRun in the agent_end hook — works in gateway mode (persistent process) but races against process.exit() in --local / embedded CLI mode. The capture Promise gets killed before it flushes to LanceDB.

Changes

Commit 1 — feat(i18n): add German language triggers

Array	File	Added patterns
MEMORY_TRIGGERS	index.ts	5 regexes: "merke dir", "vergiss nicht", "ich bevorzuge", "wir haben entschieden", "ab jetzt/sofort/in Zukunft", personal info
AUTO_CAPTURE_EXPLICIT_REMEMBER_RE	index.ts	German + English explicit remember commands
FORCE_RETRIEVE_PATTERNS	src/adaptive-retrieval.ts	2 regexes: "erinnerst du dich", "weißt du noch", "gestern", "habe ich erwähnt", etc.

Commit 2 — fix(auto-capture): await backgroundRun with timeout

• Hook becomes async; void backgroundRun replaced by await Promise.race([backgroundRun, 15s timeout])
• AgentEndAutoCaptureHook type widened to Promise<void> | void
• Session-lock / channel-delivery safety (Issue autoCapture + smartExtraction agent_end hook causes Telegram messages to be silently dropped #260): OpenClaw core calls runAgentEnd() fire-and-forget with .catch(), so the new await does not block session locks or downstream deliveries in gateway mode. The timeout is a safety net for hung API calls only.

Note on extractMinMessages

The default value of extractMinMessages (currently 4) means smart extraction never fires for CLI one-shot commands (only 1 eligible text). We worked around this with a config override (extractMinMessages: 1), but the default might be worth revisiting upstream — leaving that decision to the maintainers.

Verification (on live OpenClaw instance)

Test	Result
Gateway log shows auto-capture	✅ regex fallback found 1 capturable text(s) → auto-captured 1 memories
openclaw memory-pro search "Emerald Bear"	✅ Found with 75% score
Recall in fresh session: "Welche Testtokens kennst du?"	✅ All 6 test tokens correctly recalled
Memory stats: 0 garbage	✅ 32 clean memories

[rwmjhb]

Thanks for the PR @Banger455 — both the German i18n support and the agent_end race fix address real user pain points (issue #260). A few things need attention before this can merge.

Must Fix

Build failure — test/plugin-manifest-regression.mjs fails because the test expects api.hooks["command:new"] === undefined, but appendSelfImprovementNote already registers on command:new upstream. This is likely a pre-existing main/test mismatch rather than something your diff introduced, but the branch still needs to be green. A rebase onto latest main and re-verification should clarify this.

German regex issues (MEDIUM)

Three related problems with the new German trigger patterns:

1. False positives from substring matching — /immer|niemals|wichtig/i at index.ts:1266 matches inside compound words: Zimmermann, Schwimmerin, Flimmern, Wichtigkeit. Consider using word boundary assertions (\b) or a more targeted pattern.

2. Ungrouped alternation in fact triggers — /mein\s+\w+\s+ist|heißt|wohne|arbeite/i at index.ts:1265 causes heißt, wohne, and arbeite to match anywhere, not just after mein. Sentences like Er heißt Peter or Die Stadt heißt Berlin will spuriously trigger capture. Wrap the alternation: /mein\s+\w+\s+ist|mein\s+\w+\s+heißt|ich\s+wohne|ich\s+arbeite/i or similar.

3. Inconsistency between explicit-remember detection and capture triggers — isExplicitRememberCommand at index.ts:741 uses merke?\s+dir while shouldCapture at index.ts:1262 uses merk es dir. Result: "merk es dir" bypasses the explicit path but hits the capture path; "vergiss das nicht" does the opposite. These should be aligned.

Nice to have

• Timer cleanup (index.ts:2895-2904): The setTimeout in the Promise.race keeps Node alive up to 15s after backgroundRun resolves. Consider clearTimeout in the .then() handler. Minor, not blocking.
• Missing tests: The new German regex patterns and async hook timeout have no direct test coverage. Given the regex issues above, adding a few positive/negative test cases would catch future regressions.

[Banger455]

Thanks for the thorough review @rwmjhb! All points addressed in e773153:

Regex fixes (MUST FIX) ✅

1. Word boundaries — /immer|niemals|wichtig/ now uses \b(immer|niemals|wichtig)\b to prevent false positives on Zimmermann, Schwimmerin, Flimmern, Wichtigkeit, etc.

2. Ungrouped alternation — rewritten to mein\s+\w+\s+(?:ist|heißt)|ich\s+(?:wohne|arbeite)\b so heißt/wohne/arbeite require proper context. "Die Stadt heißt Berlin" no longer triggers.

3. Explicit-remember consistency — Both AUTO_CAPTURE_EXPLICIT_REMEMBER_RE and MEMORY_TRIGGERS now cover the same German variants: merk dir, merke dir, merk es dir, vergiss nicht, vergiss das nicht, nicht vergessen, erinnere dich, erinner dich.

Timer cleanup (NICE TO HAVE) ✅

setTimeout handle is now stored and clearTimeout'd in a finally block so the 15s safety timer doesn't keep Node alive after backgroundRun resolves.

Test coverage (NICE TO HAVE) ✅

Two new test files, 58 test cases total:

• test/german-i18n-triggers.test.mjs (54 cases)

  • shouldCapture positive: 24 cases (remember commands, preferences, decisions, personal info, emphasis)
  • shouldCapture false-positive prevention: 9 cases (compound words, missing context prefix, length bounds)
  • shouldSkipRetrieval force-retrieve: 15 cases (German retrieval triggers + short-command skip verification)
  • Explicit remember consistency: 6 cases (all German variants padded above length threshold)

• test/agent-end-async-capture.test.mjs (4 cases)

  • Hook returns Promise (async signature verification)
  • Error swallowing (unreachable endpoint doesn't throw)
  • __lastRun stores a Promise for downstream consumers
  • Early return for empty events (no 15s timer leak)

Build failure note

The plugin-manifest-regression.mjs L155 failure (command:new === undefined) is a pre-existing upstream mismatch: 26abb04 changed selfImprovement.enabled to default true, which registers on command:new, but the test wasn't updated. This is unrelated to our diff — happy to fix it in a separate PR if helpful.

Rebase

Branch rebased onto latest master (7fe2ae0).

[AliceLJY]

Thanks for the well-structured PR! Two things to confirm before merging:

1. async hook safety: Making agentEndAutoCaptureHook async is the right direction for --local embedded mode, but could you confirm the OpenClaw core runAgentEnd() call path handles unhandled rejections from async hooks correctly? If core has any special handling for sync-only hooks, the async conversion could silently swallow errors in edge cases.

2. German immer trigger precision: /\b(immer|niemals|wichtig)\b/i — immer is a very common German adverb ("always", "still") and could produce false-positive captures on ordinary sentences like "Das ist immer so". Would you consider scoping it to explicit memory-intent contexts like /\b(immer\s+(?:wenn|daran|denken)|niemals|wichtig)\b/i, or adding it to a lower-confidence tier? The merke dir / vergiss nicht patterns look solid.

Happy to approve once these two points are addressed!