akm-cli 0.9.0-beta.5 → 0.9.0-beta.51

package/dist/commands/improve/loop-stages.js

@@ -0,0 +1,1083 @@
+// This Source Code Form is subject to the terms of the Mozilla Public
+// License, v. 2.0. If a copy of the MPL was not distributed with this
+// file, You can obtain one at https://mozilla.org/MPL/2.0/.
+import path from "node:path";
+import { parseAssetRef } from "../../core/asset/asset-ref.js";
+import { daysToMs } from "../../core/common.js";
+import { loadConfig } from "../../core/config/config.js";
+import { UsageError } from "../../core/errors.js";
+import { appendEvent } from "../../core/events.js";
+import { openLogsDatabase, purgeOldTaskLogs } from "../../core/logs-db.js";
+import { getDbPath } from "../../core/paths.js";
+import { purgeOldEvents, purgeOldImproveRuns, withStateDb } from "../../core/state-db.js";
+import { info, warn } from "../../core/warn.js";
+import { closeDatabase, openIndexDatabase } from "../../indexer/db/db.js";
+import { runGraphExtractionPass } from "../../indexer/graph/graph-extraction.js";
+import { withIndexWriterLease } from "../../indexer/index-writer-lock.js";
+import { collectPendingMemories, runMemoryInferencePass, } from "../../indexer/passes/memory-inference.js";
+import { runStalenessDetectionPass } from "../../indexer/passes/staleness-detect.js";
+import { getWritableStashDirs, resolveSourceEntries } from "../../indexer/search/search-source.js";
+import { resolveImproveProcessRunnerFromProfile } from "../../integrations/agent/runner.js";
+import { isProcessEnabled } from "../../llm/feature-gate.js";
+import { withLlmStage } from "../../llm/usage-telemetry.js";
+import { createProposal, expireStaleProposals, isProposalSkipped, listProposals, purgeOrphanProposals, } from "../proposal/validators/proposals.js";
+import { checkDeadUrls } from "../url-checker.js";
+import { deriveLessonRef } from "./distill.js";
+import { deriveKnowledgeRef } from "./distill-promotion-policy.js";
+// Eligibility / candidate-selection predicates live in ./eligibility.
+import { findAssetFilePath, isDistillCandidateRef } from "./eligibility.js";
+import { writeEvalCase } from "./eval-cases.js";
+import { makeGateConfig, runAutoAcceptGate } from "./improve-auto-accept.js";
+import { resolveProcessEnabled, shouldSkipRef } from "./improve-profiles.js";
+// The pre-loop preparation pipeline lives in ./preparation.
+import { maybeAutoTuneThreshold } from "./preparation.js";
+import { akmProcedural } from "./procedural.js";
+import { akmRecombine } from "./recombine.js";
+import { recordNoOp, resetConsecutiveNoOps } from "./salience.js";
+// ── improve loop / post-loop / maintenance stages ───────────────────
+// The cycle stages run by akmImprove, extracted from improve.ts.
+export async function runImproveLoopStage(args) {
+    const { scope, options, primaryStashDir, reflectFn, distillFn, loopRefs, actions, signalBearingSet, distillCooledRefs, distillOnlyRefs, recentErrors, rejectedProposalsByRef, utilityMap, startMs, budgetMs, eventsCtx, improveProfile, } = args;
+    // O-1 (#364): compute remaining budget at call time so each sub-call
+    // receives only its fair share of the wall-clock budget.
+    const remainingBudgetMs = () => Math.max(0, budgetMs - (Date.now() - startMs));
+    const RECENT_ERRORS_CAP = 3;
+    // requirePlannedRefs guard: when the distill profile sets this flag, skip
+    // distill for distill-only refs if the reflect phase produced no planned refs.
+    // Prevents the distill loop from generating hundreds of distill-skipped events
+    // on quiet passes (all refs on reflect cooldown, no new signal to distill).
+    const requirePlannedRefs = improveProfile?.processes?.distill?.requirePlannedRefs === true;
+    const _distillOnlyRefNames = new Set(distillOnlyRefs.map((r) => r.ref));
+    const hasReflectEligibleRefs = loopRefs.some((r) => !_distillOnlyRefNames.has(r.ref));
+    const skipDistillDueToRequirePlannedRefs = requirePlannedRefs && !hasReflectEligibleRefs;
+    // R-2 / #389: Self-Consistency multi-sample voting helpers.
+    // Wang et al. arXiv:2203.11171 — N=3 samples beat single-shot on reasoning tasks.
+    const SC_THRESHOLD = options.selfConsistencyThreshold ?? 0.7;
+    const SC_N = Math.min(Math.max(2, options.selfConsistencyN ?? 3), 5);
+    /**
+     * Compute Jaccard token overlap between two strings.
+     * Tokenizes by whitespace; returns 0 when both are empty.
+     */
+    function jaccardSimilarity(a, b) {
+        const tokensA = new Set(a.split(/\s+/).filter(Boolean));
+        const tokensB = new Set(b.split(/\s+/).filter(Boolean));
+        if (tokensA.size === 0 && tokensB.size === 0)
+            return 1;
+        let intersection = 0;
+        for (const t of tokensA) {
+            if (tokensB.has(t))
+                intersection++;
+        }
+        const union = tokensA.size + tokensB.size - intersection;
+        return union > 0 ? intersection / union : 0;
+    }
+    /**
+     * Given N reflect results, return the one with the highest average Jaccard
+     * similarity to all other successful results (majority-vote winner).
+     * Falls back to the first successful result when N < 2.
+     */
+    function pickMajorityVote(results) {
+        const successful = results.filter((r) => r.ok);
+        if (successful.length === 0)
+            return (results[0] ?? {
+                schemaVersion: 1,
+                ok: false,
+                reason: "non_zero_exit",
+                error: "all samples failed",
+                exitCode: null,
+            });
+        if (successful.length === 1)
+            return successful[0];
+        let bestIdx = 0;
+        let bestScore = -1;
+        for (let i = 0; i < successful.length; i++) {
+            let totalSim = 0;
+            for (let j = 0; j < successful.length; j++) {
+                if (i === j)
+                    continue;
+                totalSim += jaccardSimilarity(successful[i].proposal.payload.content ?? "", successful[j].proposal.payload.content ?? "");
+            }
+            const avgSim = totalSim / (successful.length - 1);
+            if (avgSim > bestScore) {
+                bestScore = avgSim;
+                bestIdx = i;
+            }
+        }
+        return successful[bestIdx] ?? successful[0];
+    }
+    // O-5 / #378: helper to push per-originator errors into the rolling window.
+    function pushRecentError(originator, msg) {
+        if (!recentErrors[originator])
+            recentErrors[originator] = [];
+        recentErrors[originator].push(msg);
+        if (recentErrors[originator].length > RECENT_ERRORS_CAP)
+            recentErrors[originator].shift();
+    }
+    // Build a Set for O(1) membership test — these refs skip the reflect call (Bug D2).
+    const distillOnlyRefSet = new Set(distillOnlyRefs.map((r) => r.ref));
+    let completedCount = 0;
+    let reflectsWithErrorContext = 0;
+    const memoryRefsForInference = new Set();
+    // Pre-load all pending proposals once instead of querying per asset in the loop.
+    const dedupeStashDirForProposals = primaryStashDir ?? options.stashDir;
+    const pendingProposalRefSet = new Set(dedupeStashDirForProposals
+        ? listProposals(dedupeStashDirForProposals, { status: "pending" }).map((p) => p.ref)
+        : []);
+    let gateAutoAcceptedCount = 0;
+    let gateAutoAcceptFailedCount = 0;
+    const reflectGateCfg = makeGateConfig("reflect", {
+        globalThreshold: options.autoAccept,
+        dryRun: options.dryRun ?? false,
+        stashDir: primaryStashDir,
+        config: options.config ?? loadConfig(),
+        eventsCtx,
+        stateDbPath: eventsCtx?.dbPath,
+        // candidateCount drives the exploration budget. loopRefs is the per-phase
+        // set for reflect/distill; pass it so exploration budget is proportional.
+        candidateCount: loopRefs.length,
+    });
+    const distillGateCfg = makeGateConfig("distill", {
+        globalThreshold: options.autoAccept,
+        dryRun: options.dryRun ?? false,
+        stashDir: primaryStashDir,
+        config: options.config ?? loadConfig(),
+        eventsCtx,
+        stateDbPath: eventsCtx?.dbPath,
+        candidateCount: loopRefs.length,
+    });
+    for (const planned of loopRefs) {
+        if (Date.now() - startMs >= budgetMs) {
+            const remaining = loopRefs.length - completedCount;
+            info(`[improve] budget exhausted after ${Math.round((Date.now() - startMs) / 60000)}min — ${remaining} assets skipped`);
+            appendEvent({
+                eventType: "improve_skipped",
+                ref: planned.ref,
+                metadata: {
+                    reason: "budget_exhausted",
+                    remaining,
+                },
+            }, eventsCtx);
+            // B11: Emit improve_skipped for all remaining assets that will not be processed.
+            for (const remainingRef of loopRefs.slice(completedCount + 1)) {
+                appendEvent({
+                    eventType: "improve_skipped",
+                    ref: remainingRef.ref,
+                    metadata: { reason: "budget_exhausted_batch", remaining: loopRefs.length - completedCount - 1 },
+                }, eventsCtx);
+            }
+            actions.push({
+                ref: planned.ref,
+                mode: "error",
+                result: { ok: false, error: "timeout: improve wall-clock budget exhausted" },
+            });
+            break;
+        }
+        try {
+            // Bug D2: distillOnlyRefs skip the reflect call but still run the distill path.
+            // Bug D1: in-loop distill-cooldown check removed — distill-cooled candidates
+            //         have their synthetic actions emitted in runImprovePreparationStage.
+            const isDistillOnly = distillOnlyRefSet.has(planned.ref);
+            const parsedPlannedRef = parseAssetRef(planned.ref);
+            // B6: derived memories are machine-generated; skip reflect to avoid noisy proposals.
+            // shouldDistillMemoryRef already returns false for .derived refs, so the distill
+            // path is also a no-op for them — we just avoid unnecessary agent spawns.
+            // D2: distillOnlyRefs also skip the reflect call (reflect-cooled, distill path only).
+            if (!isDistillOnly && !planned.ref.endsWith(".derived")) {
+                // Type guard: skip reflect for unsupported types (script, env, task, etc.)
+                // and raw wiki directories, driven by the active improve profile.
+                const reflectSkip = shouldSkipRef(planned.ref, "reflect", improveProfile);
+                if (reflectSkip.skip) {
+                    actions.push({
+                        ref: planned.ref,
+                        mode: "reflect-skipped",
+                        result: { ok: true, reason: reflectSkip.reason },
+                    });
+                }
+                else {
+                    // O-5 / #378: only inject reflect-originator errors into the reflect call.
+                    // Cross-task errors (e.g. schema-repair) must NOT contaminate reflect prompts.
+                    const reflectErrors = recentErrors.reflect ?? [];
+                    if (reflectErrors.length > 0)
+                        reflectsWithErrorContext++;
+                    // O-1 (#364): pass remaining budget as timeoutMs so the agent spawn is
+                    // bounded by the wall-clock deadline rather than the default per-profile timeout.
+                    const reflectBudgetMs = remainingBudgetMs();
+                    // Wire profile.processes.reflect.{mode, profile, timeoutMs} into the reflect
+                    // dispatch when present. Falls back to akmReflect's own config-based resolution
+                    // (profiles.improve.<name>.processes.reflect → defaults.llm) when the profile
+                    // does not specify.
+                    const reflectProfileRunner = resolveImproveProcessRunnerFromProfile(improveProfile.processes?.reflect, options.config ?? loadConfig());
+                    const reflectCallArgs = {
+                        ref: planned.ref,
+                        task: options.task,
+                        ...(options.stashDir ? { stashDir: options.stashDir } : {}),
+                        ...(reflectErrors.length > 0 ? { avoidPatterns: [...reflectErrors] } : {}),
+                        agentProcess: options.agentProcess ?? "reflect",
+                        eventSource: "improve",
+                        // #639 — resolve the low-value filter from the ACTIVE improve profile
+                        // (default off when unset), so the running profile decides instead of
+                        // a hardcoded profiles.improve.default path.
+                        lowValueFilter: improveProfile.processes?.reflect?.lowValueFilter?.enabled === true,
+                        ...(reflectBudgetMs > 0 ? { timeoutMs: reflectBudgetMs } : {}),
+                        ...(reflectProfileRunner ? { runner: reflectProfileRunner } : {}),
+                        // Attribution: carry the eligibility lane so reflect stamps it on
+                        // the reflect_invoked event and the persisted proposal.
+                        ...(planned.eligibilitySource ? { eligibilitySource: planned.eligibilitySource } : {}),
+                    };
+                    // R-2 / #389: Self-consistency multi-sample voting for high-utility refs.
+                    // Self-Consistency arXiv:2203.11171 — N=3 samples beat single-shot quality.
+                    const refUtility = utilityMap.get(planned.ref) ?? 0;
+                    const useConsistency = refUtility >= SC_THRESHOLD && SC_N >= 2;
+                    let reflectResult;
+                    if (useConsistency) {
+                        const samples = [];
+                        for (let s = 0; s < SC_N; s++) {
+                            if (remainingBudgetMs() <= 0)
+                                break;
+                            // draftMode: skip DB write so each sample doesn't create a proposal.
+                            samples.push(await withLlmStage("reflect", () => reflectFn({ ...reflectCallArgs, draftMode: true })));
+                        }
+                        const winner = pickMajorityVote(samples.length > 0
+                            ? samples
+                            : [await withLlmStage("reflect", () => reflectFn({ ...reflectCallArgs, draftMode: true }))]);
+                        // Persist only the majority-vote winner as a single real proposal.
+                        if (winner.ok && primaryStashDir) {
+                            const persistResult = createProposal(primaryStashDir, {
+                                ref: winner.proposal.ref,
+                                source: "reflect",
+                                sourceRun: `reflect-sc-${Date.now()}`,
+                                payload: winner.proposal.payload,
+                                // Attribution: the self-consistency path persists the winner here
+                                // (draftMode skips reflect's own createProposal), so stamp the lane.
+                                ...(planned.eligibilitySource ? { eligibilitySource: planned.eligibilitySource } : {}),
+                            });
+                            reflectResult = isProposalSkipped(persistResult)
+                                ? {
+                                    schemaVersion: 1,
+                                    ok: false,
+                                    reason: "cooldown",
+                                    error: `SC proposal skipped: ${persistResult.message}`,
+                                    ref: winner.ref,
+                                    exitCode: null,
+                                }
+                                : { ...winner, proposal: persistResult };
+                        }
+                        else {
+                            reflectResult = winner;
+                        }
+                    }
+                    else {
+                        reflectResult = await withLlmStage("reflect", () => reflectFn(reflectCallArgs));
+                    }
+                    const isCooldown = !reflectResult.ok && reflectResult.reason === "cooldown";
+                    // Content-policy guard hits (reflect size-rail rejections) are NOT
+                    // LLM faults — the agent responded fine, the downstream guard
+                    // blocked the output. Route them to a distinct `reflect-guard-rejected`
+                    // mode so health metrics can split deterministic guard hits out of
+                    // true LLM failures. See
+                    // `/tmp/akm-health-investigations/metrics-taxonomy-review.md` §1a.
+                    const isGuardReject = !reflectResult.ok && reflectResult.reason === "content_policy_reject";
+                    // Type-guard rejection (reflect refused a script/env/task ref) is
+                    // also NOT an LLM failure — the LLM is never invoked. Route to the
+                    // existing `reflect-skipped` bucket so it does not inflate the
+                    // failure-rate numerator. ~9% of `reflect-failed` events in the
+                    // user's stack were this case; see review §1a row "Reflect refused
+                    // asset type".
+                    const isTypeRefused = !reflectResult.ok && reflectResult.reason === "unsupported_type";
+                    // Noise-gate suppression (#580): the candidate edit was an empty
+                    // diff or a cosmetic-only reformat of the current asset. Like
+                    // `unsupported_type`, this is a deterministic skip — not an LLM
+                    // fault — so it routes to the `reflect-skipped` bucket and stays
+                    // out of recentErrors/avoidPatterns.
+                    const isNoChange = !reflectResult.ok && reflectResult.reason === "no_change";
+                    actions.push({
+                        ref: planned.ref,
+                        mode: reflectResult.ok
+                            ? "reflect"
+                            : isCooldown
+                                ? "reflect-cooldown"
+                                : isGuardReject
+                                    ? "reflect-guard-rejected"
+                                    : isTypeRefused || isNoChange
+                                        ? "reflect-skipped"
+                                        : "reflect-failed",
+                        result: reflectResult,
+                    });
+                    // Cooldown skips, guard rejects, type-refused skips, and noise-gate
+                    // skips are not failures — do not pollute recentErrors with them
+                    // (those get injected as `avoidPatterns` into the next reflect
+                    // prompt). Guard rejects ARE worth showing the LLM as a learn-signal
+                    // so the next iteration sees "your last expansion was too large";
+                    // type-refused and no-change are deterministic and add no learning
+                    // signal.
+                    if (!reflectResult.ok && !isCooldown && !isTypeRefused && !isNoChange) {
+                        const errMsg = reflectResult.error ?? reflectResult.reason ?? "unknown reflect error";
+                        pushRecentError("reflect", errMsg);
+                    }
+                    // improve_reflect_outcome — per-asset metric for tuning the reflect path.
+                    appendEvent({
+                        eventType: "improve_reflect_outcome",
+                        ref: planned.ref,
+                        metadata: {
+                            ok: reflectResult.ok,
+                            durationMs: reflectResult.ok ? reflectResult.durationMs : undefined,
+                            agentProfile: reflectResult.ok ? reflectResult.agentProfile : undefined,
+                            reason: reflectResult.ok ? undefined : reflectResult.reason,
+                        },
+                    }, eventsCtx);
+                    // Plasticity counter (plan §WS-1 step 8): record no-ops so the
+                    // WS-1 selection comparator (effectiveScore, ~line 3073) can dampen
+                    // repeatedly-silent assets during consolidation-selection.
+                    // A no_change reflect means the LLM was invoked but found nothing to
+                    // improve — the asset is stable. Track it. A successful reflect means
+                    // the asset changed; reset the counter so the dampener lifts.
+                    if (isNoChange && eventsCtx?.db) {
+                        try {
+                            recordNoOp(eventsCtx.db, planned.ref);
+                        }
+                        catch {
+                            // best-effort: plasticity counter failure never blocks the run
+                        }
+                    }
+                    else if (reflectResult.ok && eventsCtx?.db) {
+                        try {
+                            resetConsecutiveNoOps(eventsCtx.db, planned.ref);
+                        }
+                        catch {
+                            // best-effort
+                        }
+                    }
+                    if (reflectResult.ok) {
+                        const reflectGr = await runAutoAcceptGate([{ proposalId: reflectResult.proposal.id, confidence: reflectResult.proposal.confidence }], reflectGateCfg);
+                        gateAutoAcceptedCount += reflectGr.promoted.length;
+                        gateAutoAcceptFailedCount += reflectGr.failed.length;
+                    }
+                } // end else (reflect type/profile check)
+            }
+            else if (!isDistillOnly && planned.ref.endsWith(".derived")) {
+                // B6: .derived refs skip reflect; record synthetic skip action.
+                actions.push({
+                    ref: planned.ref,
+                    mode: "distill-skipped",
+                    result: { ok: true, reason: "derived-memory-reflect-skipped" },
+                });
+                appendEvent({
+                    eventType: "improve_skipped",
+                    ref: planned.ref,
+                    metadata: { reason: "derived_memory_reflect_skipped" },
+                }, eventsCtx);
+            }
+            // isDistillOnly refs: no reflect action emitted — proceed directly to distill path below.
+            const hasRecentFeedbackSignal = signalBearingSet.has(planned.ref);
+            const explicitRefScope = scope.mode === "ref";
+            // Profile gate: apply the full type-filter / raw-wiki / disabled rules to
+            // distill so callers who configure `profile.processes.distill.allowedTypes`
+            // or land on raw-wiki refs get a recorded skip action instead of silently
+            // proceeding.
+            const distillSkip = shouldSkipRef(planned.ref, "distill", improveProfile);
+            if (distillSkip.skip) {
+                actions.push({
+                    ref: planned.ref,
+                    mode: "distill-skipped",
+                    result: { ok: true, reason: distillSkip.reason },
+                });
+                completedCount++;
+                info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
+                continue;
+            }
+            // requirePlannedRefs guard: skip distill for distill-only refs when no
+            // reflect-eligible refs were planned this run, preventing mass skip events.
+            if (skipDistillDueToRequirePlannedRefs && isDistillOnly) {
+                actions.push({
+                    ref: planned.ref,
+                    mode: "distill-skipped",
+                    result: { ok: true, reason: "require_planned_refs" },
+                });
+                completedCount++;
+                info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
+                continue;
+            }
+            // See `isDistillCandidateRef` — excludes `lesson:*` (and anything else in
+            // DISTILL_REFUSED_INPUT_TYPES) so distill never gets queued for an input
+            // it will refuse.
+            const shouldAttemptDistill = isDistillCandidateRef(planned.ref, options.stashDir);
+            const skipMemoryDistillForWeakSignal = !isDistillOnly && parsedPlannedRef.type === "memory" && !hasRecentFeedbackSignal && !explicitRefScope;
+            // distillCooledRefs guard: pre-filter emitted synthetic actions for distill-candidate
+            // refs; non-candidate refs in the set are blocked here.
+            // O-2 (#365): bypass the distill cooldown when the user explicitly targeted
+            // this ref via --scope — their intent overrides unattended-run policies.
+            if (shouldAttemptDistill &&
+                !skipMemoryDistillForWeakSignal &&
+                (!distillCooledRefs.has(planned.ref) || explicitRefScope)) {
+                // TODO(refactor): single call site needs both lesson+knowledge refs for proposal dedup. If a third target ref type is added, extract deriveAllTargetRefs(inputRef): string[].
+                const lessonRef = deriveLessonRef(planned.ref);
+                const knowledgeRef = deriveKnowledgeRef(planned.ref);
+                const dedupeStashDir = primaryStashDir ?? options.stashDir;
+                if (dedupeStashDir) {
+                    // B2: check both lesson ref and knowledge ref since auto-promoted memories
+                    // create knowledge: proposals, not lesson: proposals.
+                    const hasExistingPending = pendingProposalRefSet.has(lessonRef) || pendingProposalRefSet.has(knowledgeRef);
+                    if (hasExistingPending) {
+                        actions.push({
+                            ref: planned.ref,
+                            mode: "distill-skipped",
+                            result: { ok: true, reason: "pending proposal exists" },
+                        });
+                        appendEvent({
+                            eventType: "improve_skipped",
+                            ref: planned.ref,
+                            metadata: { reason: "pending_proposal_exists" },
+                        }, eventsCtx);
+                        completedCount++;
+                        info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
+                        continue;
+                    }
+                    // D-2 (#370): reject-aware cooldown for distill. When the reviewer
+                    // recently rejected a distilled lesson or knowledge proposal for this
+                    // asset, skip re-distillation for a 1-day grace window. Prevents the
+                    // same rejected proposal from being regenerated immediately. The
+                    // window is fixed (the 0.8.0 redesign moved per-ref cooldowns to
+                    // signal-delta gates and dropped --distill-cooldown-days; a short
+                    // reject grace is preserved here so a fresh rejection isn't
+                    // overridden by the same run).
+                    // References: ExpeL arXiv:2308.10144, STaR arXiv:2203.14465.
+                    const DISTILL_REJECT_COOLDOWN_MS = daysToMs(1);
+                    const recentlyRejectedLesson = !explicitRefScope && // O-2: bypass when --scope <ref> is explicit
+                        (rejectedProposalsByRef.has(lessonRef) || rejectedProposalsByRef.has(knowledgeRef));
+                    if (recentlyRejectedLesson) {
+                        const rejectedEntry = rejectedProposalsByRef.get(lessonRef) ?? rejectedProposalsByRef.get(knowledgeRef);
+                        const rejectedAgeMs = rejectedEntry ? Date.now() - new Date(rejectedEntry.ts).getTime() : 0;
+                        if (rejectedAgeMs < DISTILL_REJECT_COOLDOWN_MS) {
+                            actions.push({
+                                ref: planned.ref,
+                                mode: "distill-skipped",
+                                result: { ok: true, reason: "distill reject grace window" },
+                            });
+                            appendEvent({
+                                eventType: "improve_skipped",
+                                ref: planned.ref,
+                                metadata: {
+                                    reason: "distill_reject_grace_window",
+                                },
+                            }, eventsCtx);
+                            completedCount++;
+                            info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
+                            continue;
+                        }
+                    }
+                }
+                const distillResult = await withLlmStage("distill", () => distillFn({
+                    ref: planned.ref,
+                    ...(parsedPlannedRef.type === "memory" ? { proposalKind: "auto" } : {}),
+                    ...(options.stashDir ? { stashDir: options.stashDir } : {}),
+                    // Attribution: carry the eligibility lane so distill stamps it on the
+                    // distill_invoked event and the persisted proposal.
+                    ...(planned.eligibilitySource ? { eligibilitySource: planned.eligibilitySource } : {}),
+                }));
+                actions.push({ ref: planned.ref, mode: "distill", result: distillResult });
+                if (distillResult.outcome === "queued" && distillResult.proposal) {
+                    const distillGr = await runAutoAcceptGate([{ proposalId: distillResult.proposal.id, confidence: distillResult.proposal.confidence }], distillGateCfg);
+                    gateAutoAcceptedCount += distillGr.promoted.length;
+                    gateAutoAcceptFailedCount += distillGr.failed.length;
+                }
+                if (parsedPlannedRef.type === "memory") {
+                    const promotedToKnowledge = distillResult.outcome === "queued" && distillResult.proposalKind === "knowledge";
+                    if (!promotedToKnowledge)
+                        memoryRefsForInference.add(planned.ref);
+                }
+                // Plasticity counter (plan §WS-1 step 8) for the distill path.
+                // quality_rejected: the LLM ran but produced output that didn't pass the
+                // quality gate — the asset is not yielding useful distill output.
+                // queued: a proposal was produced; reset the no-op counter.
+                if (eventsCtx?.db) {
+                    try {
+                        if (distillResult.outcome === "quality_rejected" || distillResult.outcome === "skipped") {
+                            recordNoOp(eventsCtx.db, planned.ref);
+                        }
+                        else if (distillResult.outcome === "queued") {
+                            resetConsecutiveNoOps(eventsCtx.db, planned.ref);
+                        }
+                    }
+                    catch {
+                        // best-effort: plasticity counter failure never blocks the run
+                    }
+                }
+                if (distillResult.outcome === "quality_rejected" && primaryStashDir) {
+                    const slug = planned.ref
+                        .replace(/[^a-z0-9]/gi, "-")
+                        .toLowerCase()
+                        .slice(0, 60);
+                    writeEvalCase(primaryStashDir, {
+                        ref: planned.ref,
+                        failureReason: distillResult.reason ?? "quality gate rejected",
+                        assetType: parseAssetRef(planned.ref).type ?? "unknown",
+                        rejectedAt: Date.now(),
+                        source: "distill_quality_rejected",
+                        slug: `${slug}-${Date.now()}`,
+                    });
+                }
+                // D6: use pre-loaded map instead of per-iteration DB query
+                const rejectedProposalEvent = rejectedProposalsByRef.get(planned.ref);
+                if (rejectedProposalEvent && primaryStashDir) {
+                    const slug = planned.ref
+                        .replace(/[^a-z0-9]/gi, "-")
+                        .toLowerCase()
+                        .slice(0, 60);
+                    writeEvalCase(primaryStashDir, {
+                        ref: planned.ref,
+                        failureReason: rejectedProposalEvent.metadata?.reason ?? "proposal rejected",
+                        assetType: parseAssetRef(planned.ref).type ?? "unknown",
+                        rejectedAt: new Date(rejectedProposalEvent.ts).getTime(),
+                        source: "proposal_rejected",
+                        slug: `${slug}-rejected`,
+                    });
+                }
+            }
+            else if (skipMemoryDistillForWeakSignal) {
+                actions.push({
+                    ref: planned.ref,
+                    mode: "distill-skipped",
+                    result: { ok: true, reason: "memory requires recent feedback signal" },
+                });
+                appendEvent({
+                    eventType: "improve_skipped",
+                    ref: planned.ref,
+                    metadata: { reason: "memory_distill_requires_feedback" },
+                }, eventsCtx);
+            }
+        }
+        catch (err) {
+            // B7: UsageError thrown by akmDistill on validation_failed should be recorded
+            // as mode:"distill" with outcome:"validation_failed", NOT as a generic error.
+            // The distill_invoked event was already emitted inside akmDistill before the throw.
+            if (err instanceof UsageError) {
+                actions.push({
+                    ref: planned.ref,
+                    mode: "distill",
+                    result: { ok: false, outcome: "validation_failed", error: err.message },
+                });
+            }
+            else {
+                actions.push({
+                    ref: planned.ref,
+                    mode: "error",
+                    result: { ok: false, error: err instanceof Error ? err.message : String(err) },
+                });
+            }
+        }
+        completedCount++;
+        info(`[improve] ${completedCount}/${loopRefs.length} ${planned.ref}`);
+    }
+    // WS-4: Per-phase threshold auto-tune — runs AFTER the loop so the gate
+    // has processed all candidates for this run. Persists each phase's tuned
+    // threshold to state.db for the NEXT run's makeGateConfig to read.
+    // Best-effort: a tune failure must never fail the improve run.
+    const stateDbPathForTune = eventsCtx?.dbPath;
+    if (options.autoAccept !== undefined && stateDbPathForTune) {
+        const phaseGateCfgMap = {
+            reflect: reflectGateCfg,
+            distill: distillGateCfg,
+        };
+        for (const phase of ["reflect", "distill"]) {
+            const phaseCfg = phaseGateCfgMap[phase];
+            try {
+                maybeAutoTuneThreshold(phaseCfg.phaseThreshold ?? options.autoAccept, options.config ?? loadConfig(), stateDbPathForTune, undefined, phase);
+            }
+            catch (err) {
+                warn(`[improve] calibration auto-tune (${phase}) skipped: ${err instanceof Error ? err.message : String(err)}`);
+            }
+        }
+    }
+    return { reflectsWithErrorContext, memoryRefsForInference, gateAutoAcceptedCount, gateAutoAcceptFailedCount };
+}
+export async function runImprovePostLoopStage(args) {
+    const { scope, options, primaryStashDir, actionableRefs, appliedCleanup, cleanupWarnings, memoryRefsForInference, reindexFn, eventsCtx, budgetSignal, improveProfile, consolidationRan, } = args;
+    const allWarnings = [...cleanupWarnings, ...(appliedCleanup?.warnings ?? [])];
+    info("[improve] post-loop maintenance starting");
+    const maintenanceResult = await runImproveMaintenancePasses({
+        options,
+        primaryStashDir,
+        actionableRefs,
+        memoryRefsForInference,
+        allWarnings,
+        reindexFn,
+        consolidationRan,
+        // O-1 (#364): forward the budget signal to memory inference + graph extraction.
+        budgetSignal,
+        eventsCtx,
+        improveProfile,
+    });
+    let deadUrls;
+    if (scope.mode === "all" && primaryStashDir && actionableRefs.length > 0) {
+        try {
+            const knowledgeEntries = actionableRefs
+                .filter((r) => {
+                try {
+                    return parseAssetRef(r.ref).type === "knowledge";
+                }
+                catch {
+                    return false;
+                }
+            })
+                .slice(0, 10)
+                .map((r) => ({ ref: r.ref, body: "" }));
+            if (knowledgeEntries.length > 0) {
+                info(`[improve] checking URLs in ${knowledgeEntries.length} knowledge refs`);
+                deadUrls = await checkDeadUrls(primaryStashDir, knowledgeEntries);
+                info(`[improve] URL check complete (${deadUrls.length} dead/timeout URLs)`);
+            }
+        }
+        catch {
+            // best-effort
+        }
+    }
+    // #609 — recombine / synthesize pass. Whole-corpus cross-episodic
+    // generalization. Runs in the post-loop stage under consolidate.lock (it
+    // reads the consolidated corpus and writes proposals). Opt-in: gated on the
+    // `recombine` process being enabled, whole-stash / type scope (never `ref`),
+    // and not a dry run. Mirrors the proactiveMaintenance opt-in wiring.
+    let recombination;
+    if (primaryStashDir &&
+        improveProfile &&
+        resolveProcessEnabled("recombine", improveProfile) &&
+        scope.mode !== "ref" &&
+        !options.dryRun) {
+        const recombineFn = options.recombineFn ?? akmRecombine;
+        try {
+            recombination = await recombineFn({
+                stashDir: primaryStashDir,
+                config: options.config ?? loadConfig(),
+                ...(options.runId ? { sourceRun: options.runId } : {}),
+                ...(budgetSignal ? { signal: budgetSignal } : {}),
+                ...(options.autoAccept !== undefined ? { autoAccept: options.autoAccept } : {}),
+                eligibilitySource: "recombine",
+                ...(eventsCtx ? { ctx: eventsCtx } : {}),
+                minClusterSize: improveProfile.processes?.recombine?.minClusterSize,
+                maxClustersPerRun: improveProfile.processes?.recombine?.maxClustersPerRun,
+                relatednessSource: improveProfile.processes?.recombine?.relatednessSource,
+                confirmThreshold: improveProfile.processes?.recombine?.confirmThreshold,
+                // #632 — clustering-tuning knobs. UNSET = pre-#632 behaviour.
+                maxClusterSize: improveProfile.processes?.recombine?.maxClusterSize,
+                excludeTags: improveProfile.processes?.recombine?.excludeTags,
+                excludeEntities: improveProfile.processes?.recombine?.excludeEntities,
+            });
+        }
+        catch (e) {
+            allWarnings.push(`recombine: ${String(e)}`);
+        }
+    }
+    // #615 — procedural-compilation pass. Detects recurring successful ordered
+    // action sequences and compiles them into workflow proposals. Opt-in: gated
+    // on the `procedural` process being enabled, whole-stash / type scope (never
+    // `ref`), and not a dry run. Mirrors the recombine opt-in wiring.
+    let proceduralCompilation;
+    if (primaryStashDir &&
+        improveProfile &&
+        resolveProcessEnabled("procedural", improveProfile) &&
+        scope.mode !== "ref" &&
+        !options.dryRun) {
+        const proceduralFn = options.proceduralFn ?? akmProcedural;
+        try {
+            proceduralCompilation = await proceduralFn({
+                stashDir: primaryStashDir,
+                config: options.config ?? loadConfig(),
+                ...(options.runId ? { sourceRun: options.runId } : {}),
+                ...(budgetSignal ? { signal: budgetSignal } : {}),
+                ...(options.autoAccept !== undefined ? { autoAccept: options.autoAccept } : {}),
+                eligibilitySource: "procedural",
+                ...(eventsCtx ? { ctx: eventsCtx } : {}),
+                minRecurrence: improveProfile.processes?.procedural?.minRecurrence,
+                maxProposalsPerRun: improveProfile.processes?.procedural?.maxProposalsPerRun,
+            });
+        }
+        catch (e) {
+            allWarnings.push(`procedural: ${String(e)}`);
+        }
+    }
+    return {
+        allWarnings,
+        deadUrls,
+        ...(recombination ? { recombination } : {}),
+        ...(proceduralCompilation ? { proceduralCompilation } : {}),
+        ...(maintenanceResult.memoryInference ? { memoryInference: maintenanceResult.memoryInference } : {}),
+        ...(maintenanceResult.graphExtraction ? { graphExtraction: maintenanceResult.graphExtraction } : {}),
+        ...(maintenanceResult.stalenessDetection ? { stalenessDetection: maintenanceResult.stalenessDetection } : {}),
+        ...(maintenanceResult.actions && maintenanceResult.actions.length > 0
+            ? { maintenanceActions: maintenanceResult.actions }
+            : {}),
+        memoryInferenceDurationMs: maintenanceResult.memoryInferenceDurationMs,
+        graphExtractionDurationMs: maintenanceResult.graphExtractionDurationMs,
+        orphansPurged: maintenanceResult.orphansPurged,
+        proposalsExpired: maintenanceResult.proposalsExpired,
+        // Consolidation's auto-accept gate counts now accrue in the preparation
+        // stage (#551); post-loop no longer runs an auto-accept gate of its own.
+        gateAutoAcceptedCount: 0,
+        gateAutoAcceptFailedCount: 0,
+    };
+}
+// TODO(refactor): mutates the passed-in `allWarnings` array as a hidden side channel. Return warnings in ImproveMaintenanceResult and merge in caller — invasive signature change deferred to next refactor pass.
+// Exported for tests (#584/#585 DB-locking regression coverage); production
+// callers reach it only through akmImprove → runImprovePostLoopStage.
+// TODO(refactor): mutates the passed-in `allWarnings` array as a hidden side channel. Return warnings in ImproveMaintenanceResult and merge in caller — invasive signature change deferred to next refactor pass.
+// Exported for tests (#584/#585 DB-locking regression coverage); production
+// callers reach it only through akmImprove → runImprovePostLoopStage.
+export async function runImproveMaintenancePasses(args) {
+    const { options, primaryStashDir, memoryRefsForInference, allWarnings, reindexFn, consolidationRan, budgetSignal, eventsCtx, improveProfile, } = args;
+    if (!primaryStashDir)
+        return { memoryInferenceDurationMs: 0, graphExtractionDurationMs: 0 };
+    const config = options.config ?? loadConfig();
+    const sources = resolveSourceEntries(options.stashDir, config);
+    const memoryInferenceFn = options.memoryInferenceFn ?? runMemoryInferencePass;
+    const graphExtractionFn = options.graphExtractionFn ?? runGraphExtractionPass;
+    const stalenessDetectionFn = options.stalenessDetectionFn ?? runStalenessDetectionPass;
+    let db;
+    let memoryInference;
+    let graphExtraction;
+    let stalenessDetection;
+    let reindexedAfterInference = false;
+    const actions = [];
+    let memoryInferenceDurationMs = 0;
+    let graphExtractionDurationMs = 0;
+    let orphansPurged = 0;
+    let proposalsExpired = 0;
+    const openIndexDb = () => openIndexDatabase(getDbPath(), config.embedding?.dimension ? { embeddingDim: config.embedding.dimension } : undefined);
+    // #584: reindexFn opens its own write handle on the same index.db WAL file.
+    // Holding our handle across that call produced SQLITE_BUSY / "database is
+    // locked" failures in production, so the handle is closed BEFORE every
+    // reindex and reopened after — the fresh handle also sees the post-reindex
+    // state that graph extraction and staleness detection below rely on. The
+    // reopen runs in `finally` so a failed reindex still leaves a usable handle.
+    const reindexWithIndexDbReleased = async (stashDir) => {
+        if (db) {
+            closeDatabase(db);
+            db = undefined;
+        }
+        try {
+            await reindexFn({ stashDir });
+        }
+        finally {
+            db = openIndexDb();
+        }
+    };
+    await withIndexWriterLease({ purpose: "improve-maintenance", signal: budgetSignal }, async () => {
+        try {
+            db = openIndexDb();
+            // Memory inference candidate-discovery (post-Item 9 fix from
+            // memory:akm-improve-critical-review-2026-05-20). Previously this pass
+            // was gated on memoryRefsForInference.size > 0 AND passed those refs as a
+            // candidateRefs filter. But memoryRefsForInference is populated from refs
+            // distilled THIS RUN — by the time that happens, those parents are
+            // already split (`inferenceProcessed: true`) and `isPendingMemory` excludes
+            // them. The genuinely-pending parents in the stash never entered the
+            // filter. Result: 0/0/0 for 25 consecutive runs.
+            //
+            // Fix: always run the pass when the feature is enabled; let the pass's
+            // own `collectPendingMemories` + `isPendingMemory` predicate find
+            // candidates from the filesystem-of-truth. The this-run set is still
+            // logged as a hint but no longer used as a filter.
+            const memoryInferenceDisabledByProfile = improveProfile?.processes?.memoryInference?.enabled === false;
+            const minPendingCount = improveProfile?.processes?.memoryInference?.minPendingCount;
+            const pendingBelowMinCount = (() => {
+                if (!primaryStashDir || minPendingCount === undefined || minPendingCount <= 0)
+                    return false;
+                const pending = collectPendingMemories(primaryStashDir).length;
+                if (pending < minPendingCount) {
+                    info(`[improve] memory inference skipped (${pending} pending < minPendingCount ${minPendingCount})`);
+                    return true;
+                }
+                return false;
+            })();
+            if (memoryInferenceDisabledByProfile) {
+                info("[improve] memory inference skipped (disabled by improve profile)");
+            }
+            else if (pendingBelowMinCount) {
+                // skipped — message already emitted above
+            }
+            else {
+                const hintRefs = memoryRefsForInference.size;
+                info(hintRefs > 0
+                    ? `[improve] memory inference starting (${hintRefs} hint refs touched this run; pass discovers all pending)`
+                    : "[improve] memory inference starting (discovering pending parents)");
+                const inferenceStart = Date.now();
+                try {
+                    // O-1 (#364): pass budget signal so a hung inference call is cancelled.
+                    memoryInference = await withLlmStage("memory-inference", () => memoryInferenceFn({
+                        config,
+                        sources,
+                        signal: budgetSignal,
+                        db,
+                        reEnrich: false,
+                        onProgress: (event) => {
+                            const current = event.currentRef ? ` ${event.currentRef}` : "";
+                            info(`[improve] memory inference ${event.processed}/${event.total}${current} (written ${event.writtenFacts}, skipped ${event.skippedNoFacts})`);
+                        },
+                    }));
+                    memoryInferenceDurationMs = Date.now() - inferenceStart;
+                    actions.push({ ref: "memory:_inference", mode: "memory-inference", result: memoryInference });
+                    info(`[improve] memory inference complete (${memoryInference.writtenFacts} facts written from ${memoryInference.splitParents} parents)`);
+                }
+                catch (err) {
+                    memoryInferenceDurationMs = Date.now() - inferenceStart;
+                    allWarnings.push(`memory inference failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+            }
+            if (memoryInference && (memoryInference.splitParents > 0 || memoryInference.writtenFacts > 0)) {
+                info("[improve] reindexing after memory inference writes");
+                try {
+                    await reindexWithIndexDbReleased(primaryStashDir);
+                    reindexedAfterInference = true;
+                    info("[improve] reindex after memory inference complete");
+                }
+                catch (err) {
+                    allWarnings.push(`reindex after memory inference failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+            }
+            const graphEnabled = isProcessEnabled("index", "graph_extraction", config);
+            const graphExtractionDisabledByProfile = improveProfile?.processes?.graphExtraction?.enabled === false;
+            const graphExtractionFullScan = improveProfile?.processes?.graphExtraction?.fullScan === true;
+            // #624 P2: optional incremental high-signal-first cap. Unset = process all
+            // eligible (byte-identical to today; no ranking/slice).
+            const graphExtractionTopN = improveProfile?.processes?.graphExtraction?.topN;
+            // Build the set of refs actually touched this run.
+            const touchedRefs = new Set();
+            for (const r of args.actionableRefs)
+                touchedRefs.add(r.ref);
+            for (const r of memoryRefsForInference)
+                touchedRefs.add(r);
+            // INVARIANT: graph extraction normally runs only on files touched by
+            // actionable refs (candidatePaths). Full-corpus scans are opt-in via
+            // profile.processes.graphExtraction.fullScan = true (used by the
+            // `graph-refresh` built-in profile and its weekly scheduled task).
+            // The empty-Set fallback is intentional when no refs were touched —
+            // the extractor's filter rejects every file and returns empty, keeping
+            // the pass invoked so the action is recorded and tests stay exercised.
+            if (graphExtractionDisabledByProfile) {
+                info("[improve] graph extraction skipped (disabled by improve profile)");
+            }
+            else if (sources.length > 0 && graphEnabled) {
+                info(`[improve] graph extraction starting${graphExtractionFullScan ? " (full-corpus scan)" : ""}`);
+                const extractionStart = Date.now();
+                try {
+                    // D9: if consolidation ran but memory inference did not reindex, force a reindex
+                    // so graph extraction sees current DB state after consolidation writes.
+                    if (consolidationRan && !reindexedAfterInference) {
+                        info("[improve] reindexing after consolidation (graph extraction needs current state)");
+                        try {
+                            await reindexWithIndexDbReleased(primaryStashDir);
+                            reindexedAfterInference = true;
+                            info("[improve] reindex after consolidation complete");
+                        }
+                        catch (err) {
+                            allWarnings.push(`reindex after consolidation failed: ${err instanceof Error ? err.message : String(err)}`);
+                        }
+                    }
+                    // #584: no close/reopen needed here — reindexWithIndexDbReleased
+                    // already swapped in a fresh post-reindex handle.
+                    // Resolve touched refs to absolute file paths. Skipped for fullScan
+                    // (candidatePaths stays undefined → extractor processes all files).
+                    let candidatePaths;
+                    if (!graphExtractionFullScan) {
+                        candidatePaths = new Set();
+                        if (primaryStashDir && touchedRefs.size > 0) {
+                            const writableDirSet = new Set(getWritableStashDirs(primaryStashDir).map((d) => path.resolve(d)));
+                            const resolved = await Promise.all([...touchedRefs].map((ref) => findAssetFilePath(ref, primaryStashDir, writableDirSet).catch(() => null)));
+                            for (const p of resolved) {
+                                if (typeof p === "string" && p.length > 0)
+                                    candidatePaths.add(p);
+                            }
+                        }
+                    }
+                    const progressHandler = (event) => {
+                        const current = event.currentPath ? ` ${path.basename(event.currentPath)}` : "";
+                        info(`[improve] graph extraction ${event.processed}/${event.total}${current} (extracted ${event.extracted}, entities ${event.totalEntities}, relations ${event.totalRelations})`);
+                    };
+                    // O-1 (#364): pass budget signal so a hung graph extraction call is cancelled.
+                    graphExtraction = await withLlmStage("graph-extraction", () => graphExtractionFn({
+                        config,
+                        sources,
+                        signal: budgetSignal,
+                        db,
+                        reEnrich: false,
+                        onProgress: progressHandler,
+                        options: { candidatePaths, ...(graphExtractionTopN != null ? { topN: graphExtractionTopN } : {}) },
+                    }));
+                    graphExtractionDurationMs = Date.now() - extractionStart;
+                    actions.push({ ref: "graph:_artifact", mode: "graph-extraction", result: graphExtraction });
+                    info(`[improve] graph extraction complete (${graphExtraction.quality.extractedFiles} files, ${graphExtraction.quality.entityCount} entities, ${graphExtraction.quality.relationCount} relations)`);
+                }
+                catch (err) {
+                    graphExtractionDurationMs = Date.now() - extractionStart;
+                    allWarnings.push(`graph extraction failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+            }
+            else if (sources.length > 0 && !graphEnabled) {
+                info("[improve] graph extraction skipped (features.index.graph_extraction is disabled)");
+            }
+            // Orphan proposal purge — reject pending reflect proposals whose target
+            // asset no longer exists on disk. Runs after graph extraction so newly
+            // promoted assets from accept flows during this run are already present.
+            if (primaryStashDir) {
+                try {
+                    const purgeResult = purgeOrphanProposals(primaryStashDir, sources.map((s) => s.path));
+                    orphansPurged = purgeResult.rejected;
+                    if (purgeResult.rejected > 0) {
+                        info(`[improve] orphan purge: ${purgeResult.rejected}/${purgeResult.checked} orphaned proposals rejected (${purgeResult.durationMs}ms)`);
+                    }
+                    appendEvent({
+                        eventType: "proposal_orphan_purge",
+                        ref: "proposals:_orphan-purge",
+                        metadata: {
+                            checked: purgeResult.checked,
+                            rejected: purgeResult.rejected,
+                            durationMs: purgeResult.durationMs,
+                            byType: purgeResult.byType,
+                            orphans: purgeResult.orphans.map((o) => o.ref),
+                        },
+                    }, eventsCtx);
+                }
+                catch (err) {
+                    allWarnings.push(`orphan purge failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+                // Phase 6B (Advantage D6b): expire pending proposals that have aged past
+                // the retention window. Runs AFTER orphan purge so we never double-archive
+                // a proposal that orphan-purge already moved. `expireStaleProposals` emits
+                // its own per-proposal `proposal_expired` events; we additionally emit a
+                // single roll-up event here for parity with the orphan-purge surface.
+                try {
+                    const expireResult = expireStaleProposals(primaryStashDir, config);
+                    proposalsExpired = expireResult.expired;
+                    if (expireResult.expired > 0) {
+                        info(`[improve] expiration: ${expireResult.expired}/${expireResult.checked} pending proposals expired ` +
+                            `(retention=${expireResult.retentionDays}d, ${expireResult.durationMs}ms)`);
+                    }
+                    appendEvent({
+                        eventType: "proposal_expiration_pass",
+                        ref: "proposals:_expiration",
+                        metadata: {
+                            checked: expireResult.checked,
+                            expired: expireResult.expired,
+                            durationMs: expireResult.durationMs,
+                            retentionDays: expireResult.retentionDays,
+                            expiredProposals: expireResult.expiredProposals,
+                        },
+                    }, eventsCtx);
+                }
+                catch (err) {
+                    allWarnings.push(`proposal expiration failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+            }
+            // Fix #2 (observability 0.8.0): trim the events table in state.db so it
+            // doesn't grow unbounded. `akm health` writes a `health_probe` row on every
+            // invocation, and every command surface emits at least one event besides —
+            // without this trim, state.db is a permanent append-only log. Config key
+            // `improve.eventRetentionDays` (default 90, set 0 to disable) controls the
+            // window. The purge runs against state.db (a different SQLite file from
+            // the index `db` above).
+            {
+                const retentionDays = typeof config.improve?.eventRetentionDays === "number" ? config.improve.eventRetentionDays : 90;
+                if (retentionDays > 0) {
+                    // #585: reuse the long-lived eventsCtx.db connection when akmImprove
+                    // opened one — opening a second state.db write connection while
+                    // eventsDb is still live made two simultaneous writers contend on the
+                    // same WAL file ("database is locked"). Only the eventsCtx.dbPath
+                    // fallback path (state.db failed to open up-front) opens — and then
+                    // owns and closes — its own handle. C2 still holds: the fallback uses
+                    // the boundary-pinned path, never a live `process.env` re-read.
+                    try {
+                        withStateDb((stateDb) => {
+                            const purgedCount = purgeOldEvents(stateDb, retentionDays);
+                            if (purgedCount > 0) {
+                                info(`[improve] events purge: ${purgedCount} event(s) older than ${retentionDays}d removed from state.db`);
+                            }
+                            appendEvent({
+                                eventType: "events_purged",
+                                ref: "events:_purge",
+                                metadata: { purgedCount, retentionDays },
+                            }, eventsCtx);
+                            // improve_runs uses the same retention window as events — both are
+                            // observability/audit data, both grow append-only, both have a
+                            // dedicated purge helper. Mirroring the events purge here means a
+                            // single retention knob (improve.eventRetentionDays) governs both.
+                            const improveRunsPurged = purgeOldImproveRuns(stateDb, retentionDays);
+                            if (improveRunsPurged > 0) {
+                                info(`[improve] improve_runs purge: ${improveRunsPurged} run(s) older than ${retentionDays}d removed from state.db`);
+                            }
+                            appendEvent({
+                                eventType: "improve_runs_purged",
+                                ref: "improve_runs:_purge",
+                                metadata: { purgedCount: improveRunsPurged, retentionDays },
+                            }, eventsCtx);
+                        }, { path: eventsCtx?.dbPath, borrowed: eventsCtx?.db });
+                    }
+                    catch (err) {
+                        allWarnings.push(`events purge failed: ${err instanceof Error ? err.message : String(err)}`);
+                    }
+                    // task_logs in logs.db (#579) shares the same retention window as
+                    // events/improve_runs — all three are observability data governed by
+                    // the single improve.eventRetentionDays knob. Separate try/finally
+                    // because logs.db is a different file: a locked/missing logs.db must
+                    // not block the state.db purges above.
+                    let logsDb;
+                    try {
+                        logsDb = openLogsDatabase();
+                        const taskLogsPurged = purgeOldTaskLogs(logsDb, retentionDays);
+                        if (taskLogsPurged > 0) {
+                            info(`[improve] task_logs purge: ${taskLogsPurged} log line(s) older than ${retentionDays}d removed from logs.db`);
+                        }
+                        appendEvent({
+                            eventType: "task_logs_purged",
+                            ref: "task_logs:_purge",
+                            metadata: { purgedCount: taskLogsPurged, retentionDays },
+                        }, eventsCtx);
+                    }
+                    catch (err) {
+                        allWarnings.push(`task_logs purge failed: ${err instanceof Error ? err.message : String(err)}`);
+                    }
+                    finally {
+                        if (logsDb) {
+                            try {
+                                logsDb.close();
+                            }
+                            catch {
+                                // best-effort
+                            }
+                        }
+                    }
+                }
+            }
+            // Phase 4A (staleness detection). Activates the `deprecated` belief-state
+            // machinery shipped in Phase 1A. Default OFF — gated by
+            // `features.index.staleness_detection.enabled`. Runs after orphan purge
+            // and before the URL check (which lives in the outer caller).
+            if (sources.length > 0) {
+                try {
+                    stalenessDetection = await withLlmStage("staleness-detection", () => stalenessDetectionFn({ config, sources, signal: budgetSignal, db }));
+                    if (stalenessDetection.considered > 0) {
+                        info(`[improve] staleness detection complete (considered ${stalenessDetection.considered}, ` +
+                            `deprecated ${stalenessDetection.deprecated}, confirmed ${stalenessDetection.confirmed}, ` +
+                            `skipped ${stalenessDetection.skipped}, ${stalenessDetection.durationMs}ms)`);
+                    }
+                    for (const w of stalenessDetection.warnings)
+                        allWarnings.push(`[improve] staleness detection: ${w}`);
+                }
+                catch (err) {
+                    allWarnings.push(`staleness detection failed: ${err instanceof Error ? err.message : String(err)}`);
+                }
+            }
+        }
+        finally {
+            if (db)
+                closeDatabase(db);
+        }
+    });
+    return {
+        ...(memoryInference ? { memoryInference } : {}),
+        ...(graphExtraction ? { graphExtraction } : {}),
+        ...(stalenessDetection ? { stalenessDetection } : {}),
+        ...(actions.length > 0 ? { actions } : {}),
+        memoryInferenceDurationMs,
+        graphExtractionDurationMs,
+        orphansPurged,
+        proposalsExpired,
+    };
+}