The gateway sends agent stderr to /dev/null, so F1's `log.info("tier1-retrieval
...")` was unobservable — couldn't confirm it fired. Add recordTier1Diag() which
best-effort appends each outcome (query, hits, injected, latency, reason,
sessionKey, agentId) to ~/.openclaw/logs/tier1-retrieval.jsonl, independent of
subsystem log routing. Never throws. attempt.ts calls it alongside the log line.

Verify a fresh-thread session with: tail ~/.openclaw/logs/tier1-retrieval.jsonl

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…al root cause)

Agents running via provider=claude-cli use the CLI runner, NOT the embedded/API
runner (attempt.ts) where F1 was first wired — so it never fired on real turns.
Move the Tier-1 retrieval into the CLI runner's context assembly
(cli-runner/prepare.ts), gated on a populated cold-start contextFiles,
prepending the slice ahead of MEMORY.md.

Also:
- cleanTier1Query: extract the real user message from the Slack relay preamble
  (the "## Message (user-controlled text...)" boundary) — was embedding the whole
  ~8KB wrapper, which is junk + made bge-m3/Ollama throw.
- Tier1Diag.err: capture the caught search-error text in the trace.
- recordTier1Gate (temp instrumentation): log the eligibility breakdown per run.

Verified live: clean query, Ollama bge-m3 embed ~156ms, search runs.
Open: recall QUALITY (bge-m3 cosine too compressed to discriminate) -> follow-up
reranker+embedder lane.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

…il-open)

After the cosine candidate set in buildTier1RetrievalContextFile, an optional LLM
judge (qwen2.5:7b-instruct on the carbon-white RTX 5090) scores each candidate for
"does this actually answer THIS question?", re-sorts by judge score, and drops below
query.reranker.minScore. Catches topically-adjacent hard negatives the embedder alone
can't separate (the 0.40-vs-0.38 problem). FAIL-OPEN: any judge error/timeout/empty
returns the cosine order unchanged - the rerank refines, never drops, retrieval.

- zod-schema.agent-runtime.ts: query.reranker {enabled,baseUrl,apiKey,model,timeoutMs,minScore,topK}
- memory-search.ts: resolve query.reranker into ResolvedMemorySearchConfig
- tier1-retrieval.ts: rerankTier1Candidates() + observability (kind:"rerank" diag line)

Verified live on the mini1 brain: cold-start turn fired the judge (before:2 -> after:1),
reason:"ok", both embedder+judge serving on the 5090.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>

Harness-agnostic entrypoint for tier1 retrieval-at-start: `openclaw memory
tier1 <query>` builds the same RETRIEVED-CONTEXT-TIER1.md slice the runners
inject at cold session start, so any external harness (or an operator) can
fetch it without going through a runner.

- Registration + runtime follow the existing `memory search` pattern
  (positional query or --query, --agent, --json, --out, --session-key,
  --max-results, --max-bytes).
- Honors the `query.tier1.enabled` flag gate exactly as the runners do:
  disabled -> `{ injected: false, reason: "disabled" }`, never an error.
  Fail-open contract: every retrieval miss (disabled / no-signal / timeout /
  error / below-threshold) exits 0; only a missing query is a usage error.
- Exposes buildTier1RetrievalContextFile through the established
  memory-core-host-runtime-core plugin-sdk surface.
- tier1 module: small additive caller overrides for maxResults/maxBytes,
  hard-capped at 8 hits / 8192 bytes.
- Brings the ported tier1/reranker module up to main's gates: adds the
  missing `query.reranker` member on the MemorySearchConfig input type,
  exports rerankTier1Candidates, and fixes sort/boolean-compare lint.

Co-Authored-By: Claude Fable 5 <[email protected]>

- rerankTier1Candidates: re-orders by judge score within topK, drops
  below-minScore candidates, limits to topK on success, and fails open on
  judge HTTP error, timeout (abort), non-JSON responses, an emptied set,
  and missing baseUrl/model or single-candidate input.
- buildTier1RetrievalContextFile: caller maxResults/maxBytes overrides are
  honored and clamped to the hard caps; config-enabled reranker re-orders
  the injected slice.
- `memory tier1` CLI: usage error on missing query (exit 1), disabled gate
  emits JSON without opening a manager, enabled path injects with
  memory-scoped search + per-command session key, --session-key override,
  fail-open exit 0 when the search throws, --out writes only when injected,
  and the default mode prints the body to stdout.

recordTier1Diag is intentionally not unit-tested: TIER1_DIAG_LOG resolves
os.homedir() at module load, and mocking the homedir globally is worse than
the missing coverage.

Co-Authored-By: Claude Fable 5 <[email protected]>

Co-Authored-By: Claude Fable 5 <[email protected]>