mvm-recall

name	mvm-recall
description	MVM unified question-answering. Step 0 memo-cache, then KB + web retrieval INLINE (orchestrator searches itself) plus ONE naked-baseline haiku probe for prior-contamination signal; reconciles via trust hierarchy KB>web>weights, auto-spawns /mvm-ingest on web hits, logs every recall via `recall-log add`. Triggers: '/mvm-recall', 'recall <question>', any factual question routed to mvm.

/mvm-recall

Architecture: KB and web retrieval run inline (you search); only the naked probe is a subagent (weight-prior needs an uncontaminated context). History & rationale for every rule here: ~/resources/mvm-incident-history.md.

Steps

Timing capture (F3a, mandatory): the ledger profiles recall latency (713s mean as of 2026-06-12 — unfixable until rows say WHERE it goes), so stamp phase boundaries with date +%s%3N (epoch-ms, ~0 cost): chain it onto an adjacent Bash call where one exists (date +%s%3N; <cmd>; date +%s%3N), bare around non-Bash steps (web tools). Wall-clock per phase is the target — it INCLUDES orchestrator think-time, which may be the real sink. Deltas land in the timings object at step 6.

0. Memo-cache. date +%s%3N; recall-log cache-lookup "<question>"; date +%s%3N (stamps = t0 start-of-recall + cache_ms). On a hit with sim ≥ 0.7 whose answer you judge still fresh (check age_days against domain velocity — PoE2 patch facts go stale in days; biography doesn't): serve it, log with decided_source:"cache" + source_detail naming the original ts, output — from cache (age Nd), done. Stale-suspect or low-sim → full pipeline below.

1. Naked probe — the only subagent. Spawn in the SAME message as your first inline search so it runs concurrently. subagent_type: mvm-naked-clone (tools: [], structural), model: haiku. Validation: task notification must show tool_uses: 0, else the measurement is invalid — re-spawn stricter or mark the probe failed. Timing: probe_ms = stamp in the FIRST Bash call after the probe result is in hand, minus the step-0 end stamp (spawn-to-result wall, concurrency included).

   [CRITICAL CONSTRAINTS — READ FIRST]
   You are a CALIBRATION INSTRUMENT measuring weight-prior alone.
   You MUST NOT use ANY tool. Forbidden: Bash, Read, Grep, Glob, WebSearch,
   WebFetch, mvm search, mvm verify, Skill, Agent, ToolSearch — and any other
   tool. Tool use INVALIDATES the measurement and breaks the system.
   If you would normally reach for a tool, DO NOT. Note in RATIONALE that you
   would have used <X> but didn't, and answer from prior alone — even if
   uncertain.
   
   QUESTION: <q>
   
   Output exactly:
   ANSWER: <your best answer from prior knowledge alone>
   RATIONALE: <one sentence; if you wanted a tool, name it here>
   

2. KB retrieval — inline. Surface: ~/mvm/knowledge/{resources,areas,topics,topics-alfred-state}/. knowledge/projects/ is NOT a surface — project deliverables are direct-Read from ~/projects/ (a hit there = KB-miss).

   Search doctrine (eval-backed, exp5 2026-06-09): the engine is vector-first with FTS fallback. Issue mvm search "<question verbatim>" AND up to 2 tight key-term queries (proper nouns / distinctive phrases), merge the results — each style uniquely recovers docs the other misses (union 17/27 vs 13/27 either alone). Read the best candidate if snippets are insufficient. Conclude KB-miss only after both styles come up empty. Capture kb.ans + kb.rationale citing the path. Keep it ≤4 retrieval calls. If candidates conflict, record each claim in the rationale. Timing: prepend date +%s%3N; to the first mvm search call and stamp again when KB retrieval concludes → kb_search_ms.

3. Web retrieval — inline, by default. ~1 WebSearch + 1 WebFetch on the top result (one extra if inconclusive); capture web.ans + web.rationale

   • URL. Skip ONLY when KB returned a grounded hit in a declared curated domain AND you have no staleness concern — web is how supersession is detected. Discount SEO/gold-seller domains (aoeah, u4gm, mmoexp, …). Timing: WebSearch/WebFetch are not Bash, so bare-stamp before the first and after the last web call → web_ms; skipped web → "web_ms": null.

4. Reconcile. Trust hierarchy KB > web > weights; the rationale is the diagnostic (grounded vs guessing).

   • KB rationale cites a file/passage → KB grounded; trust it (default).
   • KB silent → fall through to web.
   • Web cites a fresher source (date markers) contradicting KB → web wins; spawn /mvm-ingest to supersede.
   • All three guessing → hard miss; refuse with gap report, decided_source:"none".
   • Three different confidently-grounded answers → contested; surface all, decided_source:"contested", don't pick.
   • Curated-domain override: KB with grounded rationale always beats web/weights.

5. Auto-spawn /mvm-ingest in background (run_in_background: true) when web answered and KB didn't, or web superseded stale KB. Pass q, a, URL. Every web-decided ledger entry MUST carry "ingested":false (step 6) — the spawn flips it via recall-log mark-ingested on commit. Without the field a dropped spawn is invisible (verified 2026-06-12: 6/11 desecrate web hit, no ingest, no detection — the field was absent so nothing could alarm). Skipping the spawn deliberately (hit doesn't qualify) → still write "ingested":false + the reason in source_detail.

6. Log — mandatory, via the CLI (the fail-closed curated-write gate and the Stop hook key off this ledger):

   echo '{"question":"...","topic_hint":"...",
    "probes":{"naked":{"ans":"...","rationale":"..."},
              "kb":{"ans":"...","rationale":"...","top_score":N},
              "web":{"ans":"...","rationale":"...","url":"..."}},
    "reconciliation_pattern":"...",
    "decided_source":"kb|web|weights|engine|contested|cache|mixed|none",
    "source_detail":"<free-text provenance if compound>",
    "evidence_paths":["resources/..."],
    "decided_answer":"...","duration_ms":N,
    "timings":{"cache_ms":N,"kb_search_ms":N,"probe_ms":N,"web_ms":N,
               "reconcile_ms":N},
    "ingested":false}' | recall-log add   # "ingested" REQUIRED iff decided_source=web
   

   decided_source is a hard enum (CLI rejects anything else; prose goes in source_detail). evidence_paths = KB docs that grounded the answer (feeds retrieval-heat test coverage). ts/session_id/kind auto-fill. timings deltas come from the stamps above (null = step skipped; CLI rejects non-numeric/negative values); prepend date +%s%3N; to this add call — that final stamp gives reconcile_ms (end-of-retrieval → log) and duration_ms (t0 → log). Cache-served recalls log {"cache_ms":N} alone — the other steps never ran.

7. Output:

   <answer>
   
   — from <kb-path | url | cache | weights>  reconciliation: <pattern>
   [substrate updated: <what changed>]
   

8. Substrate-confusion signal. If the user's pushback this turn forces a revision of a confident claim, the substrate enabled a conflation — close the loop before exit: (a) identify the offending canonical and the two things it commingles; (b) spawn /mvm-ingest rewrite-mode to make the distinction structural (separate sections with epistemic-status labels); (c) author a NEW locked test probing exactly that conflation; (d) run injected verify on new + existing tests; (e) set substrate_confusion: {original_answer, revised_answer, source_path} in the recall-log entry. Prediction error drives a substrate update or dissolves via cross-check — never apologize-and-move-on.