akm-cli 0.9.0-beta.34 → 0.9.0-beta.36

package/CHANGELOG.md

@@ -6,6 +6,69 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [0.9.0-beta.36] — 2026-06-22
+
+### Added
+
+- **Stash standards + wiki schemas are surfaced to authoring agents at write
+  time.** When an agent edits a page under `wikis/<name>/`, that wiki's
+  `schema.md` body is injected into the prompt; when it creates/edits a non-wiki
+  asset, the bodies of `category: convention`/`meta` `fact` assets are injected.
+  Two mutually-exclusive features selected by target type, sharing one
+  `standardsContext` prompt seam. Wired into reflect, propose, and every
+  improve authoring pass (distill, consolidate, recombine, procedural, extract,
+  schema-repair). (#642)
+- **Unified, validator-sourced authoring-rules seam.** A new
+  `authoringRulesForType(type)` injects the hard authoring rules (no
+  pseudo-frontmatter in body, exactly two `---` fences, description/`when_to_use`
+  length + shape) into every authoring prompt. The numeric bounds live in one
+  module that the validators import, so the prompt can no longer drift from what
+  the gate enforces. (#645)
+
+### Fixed
+
+- **High-salience reflect lane now reflects each asset at most once.** The
+  `#608` admission gate lacked the cooldown its sibling high-retrieval gate has,
+  so zero-feedback assets were re-selected on every run (auto-accept emits
+  `promoted`, not `feedback`), burning LLM calls and churning assets. (#643)
+- **Stuck validation-failing proposals no longer dead-end.** The triage drain no
+  longer overwrites an `auto-rejected` gate stamp with a misleading
+  `auto-accepted` (the failure stays truthful and visible). A bounded,
+  content-preserving auto-repair (strip pseudo-frontmatter / stray `---`, repair
+  truncated descriptions) runs at the promote boundary and re-validates — fixable
+  proposals promote; genuinely unrepairable ones stay `pending` for manual
+  review, with nothing fabricated and validation never bypassed. (#645)
+- Corrected a prompt/validator drift where the distill system prompt asked for an
+  80–200 char description while the gate enforced 20–400. (#645)
+
+## [0.9.0-beta.35] — 2026-06-21
+
+### Fixed
+
+- **Default extract discovery window is now "since the last run" (floored at 48h),
+  not a fixed 24h.** An intermittently-online host that was off for longer than
+  the old 24h window could permanently miss sessions that ended during the gap.
+  Discovery now looks back to the last recorded extract run for the harness, never
+  less than 48h. Widening is free of redundant LLM cost — the content-hash ledger
+  skips unchanged sessions with zero LLM calls. An explicit `--since`/`defaultSince`
+  still wins.
+- **Per-session lock prevents concurrent double-extraction.** A session-end hook
+  firing `extract --session-id` while the periodic `akm improve` extract pass runs
+  discovery could both LLM-process the SAME session (duplicate spend + near-dup
+  proposals). A per-(harness, session) advisory lock (co-located with state.db,
+  PID + age staleness recovery) now makes the second run skip without any LLM call.
+- **`minNewSessions` is read from the ACTIVE improve profile, not always `default`.**
+  A non-default profile (e.g. `frequent`) setting `minNewSessions` was silently
+  ignored because the gate (and its candidate-count discovery window) read
+  `profiles.improve.default`. They now read the resolved active profile, matching
+  how `extract.enabled` already resolves.
+
+### Docs
+
+- Documented that `processes.extract.indexSessions` (default on) makes a second
+  LLM call per processed session (the session summary); set it to `false` to halve
+  per-session extract cost. Unchanged/skipped sessions still cost zero.
+
 ## [0.9.0-beta.34] — 2026-06-21
 
 ### Fixed

package/dist/assets/prompts/distill-lesson-system.md

@@ -15,7 +15,7 @@ when_to_use: <one complete sentence describing the concrete trigger condition>
 <lesson body — plain markdown, 1–3 short paragraphs of practical guidance>
 
 ## description field (MANDATORY)
-- A single complete sentence in present tense, 80-200 chars, NO markdown.
+- A single complete sentence in present tense, 20–400 chars, NO markdown.
 - Self-contained: a reviewer must understand the lesson from this field alone.
 - DO NOT start with "When ", "If ", or a connector word — that belongs in when_to_use.
 - DO NOT copy a section heading ("Key takeaways", "For example", "Key pitfalls").

package/dist/assets/prompts/extract-session.md

@@ -28,7 +28,7 @@ Things NOT to extract:
 - Started: {{STARTED_AT}}
 - Ended: {{ENDED_AT}}
 - Project hint: {{PROJECT_HINT}}
-
+{{STANDARDS}}
 ## Filtered session transcript
 
 The transcript below has already had read-only `akm` meta-ops and platform boilerplate stripped. Only content that might carry signal remains.

package/dist/commands/improve/consolidate.js

@@ -15,6 +15,7 @@ import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
 import { ConfigError } from "../../core/errors.js";
 // Note: appendEvent import removed (WS-3a: archive TTL machinery retired)
 import { parseEmbeddedJsonResponse } from "../../core/parse.js";
+import { resolveStashStandards } from "../../core/standards/resolve-stash-standards.js";
 import { detectTruncatedDescription } from "../../core/text-truncation.js";
 import { hasHotCaptureMode, hasSupersededStatus, MERGE_ABSOLUTE_FLOOR_CHARS, MERGE_SHRINK_RATIO_MIN, validateProposalFrontmatter, } from "../proposal/validators/proposal-quality-validators.js";
 import { createProposal, isProposalSkipped, listProposals } from "../proposal/validators/proposals.js";
@@ -377,7 +378,7 @@ async function clusterMemoriesBySimilarity(memories, config, stateDb) {
  * is precomputed once per run by `loadPendingConsolidateProposalHashes`
  * so the cost stays O(memories) inside the chunk loop.
  */
-export function buildChunkPrompt(sourceName, memories, chunkIndex, totalChunks, bodyTruncation, pendingProposalBodyHashes = new Set()) {
+export function buildChunkPrompt(sourceName, memories, chunkIndex, totalChunks, bodyTruncation, pendingProposalBodyHashes = new Set(), standardsContext = "") {
     const start = memories[0] ? `memory:${memories[0].name}` : "";
     const end = memories[memories.length - 1] ? `memory:${memories[memories.length - 1].name}` : "";
     const annotationsByIndex = [];
@@ -405,6 +406,11 @@ export function buildChunkPrompt(sourceName, memories, chunkIndex, totalChunks,
         `Chunk ${chunkIndex + 1} of ${totalChunks}, memories ${start}–${end}:`,
         "",
     ];
+    if (standardsContext.trim()) {
+        lines.push("Standards to follow (the rulebook for this target):");
+        lines.push(standardsContext.trim());
+        lines.push("");
+    }
     // Top-of-prompt protection block for hot refs. Neutral phrasing — avoid
     // op-words like "promote", "merge", "contradict" so the model doesn't
     // accidentally treat the warning as a hint to use that op elsewhere
@@ -1246,6 +1252,10 @@ async function akmConsolidateInner(opts, config, stashDir, startMs, sourceRun, w
     }
     warn(`[consolidate] ${memories.length} memories / ${chunks.length} chunk(s) / chunk_size=${chunkSize}` +
         ` / pending-proposal hashes: ${pendingProposalBodyHashes.size}`);
+    // Consolidate output merges memories (non-wiki) → stash authoring standards.
+    // Resolved ONCE per run and passed to each chunk prompt (facts not re-read
+    // per chunk).
+    const standardsContext = resolveStashStandards(stashDir);
     const chunkOpsArrays = [];
     // Structured skip-reason histogram (2026-05-26): every deterministic
     // post-LLM op rejection site below also calls `pushSkipReason` so the
@@ -1363,7 +1373,7 @@ async function akmConsolidateInner(opts, config, stashDir, startMs, sourceRun, w
             continue;
         }
         warn(`[consolidate] chunk ${chunkIdx + 1}/${chunks.length} (${chunk.length} memories) …`);
-        const userPrompt = buildChunkPrompt(sourceName, chunk, chunkIdx, chunks.length, bodyTruncation, pendingProposalBodyHashes);
+        const userPrompt = buildChunkPrompt(sourceName, chunk, chunkIdx, chunks.length, bodyTruncation, pendingProposalBodyHashes, standardsContext);
         let raw = await tryLlmFeature("memory_consolidation", config, async () => {
             if (!llmConfig)
                 return { ok: false, error: "No LLM configured for consolidation" };

package/dist/commands/improve/distill.js

@@ -58,12 +58,14 @@ import { parseAssetRef } from "../../core/asset/asset-ref.js";
 import { assembleAssetFromString } from "../../core/asset/asset-serialize.js";
 import { parseFrontmatter, writeSalienceToFrontmatter } from "../../core/asset/frontmatter.js";
 import { stripMarkdownFences } from "../../core/asset/markdown.js";
+import { authoringRulesForType } from "../../core/authoring-rules.js";
 import { resolveStashDir, timestampForFilename } from "../../core/common.js";
 import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
 import { ConfigError, UsageError } from "../../core/errors.js";
 import { appendEvent, readEvents } from "../../core/events.js";
 import { lintLessonContent } from "../../core/lesson-lint.js";
 import { getDbPath } from "../../core/paths.js";
+import { resolveStashStandards } from "../../core/standards/resolve-stash-standards.js";
 import { openStateDatabase } from "../../core/state-db.js";
 import { warnVerbose } from "../../core/warn.js";
 import { closeDatabase, getAllEntries, openDatabase } from "../../indexer/db/db.js";
@@ -307,6 +309,18 @@ export function buildDistillPrompt(input) {
     const lines = [];
     lines.push(`Asset ref: ${input.inputRef}`);
     lines.push("");
+    if (input.standardsContext?.trim()) {
+        lines.push("Standards to follow (the rulebook for this target):");
+        lines.push(input.standardsContext.trim());
+        lines.push("");
+    }
+    {
+        const authoringRules = authoringRulesForType(input.proposalKind ?? "lesson");
+        if (authoringRules) {
+            lines.push(authoringRules);
+            lines.push("");
+        }
+    }
     lines.push("Asset content:");
     if (input.assetContent) {
         const body = input.assetContent.trim().slice(0, 3000);
@@ -939,12 +953,16 @@ export async function akmDistill(options) {
             // Fail open — CLS is supplemental, never required.
         }
     }
+    // Distill output is a lesson/knowledge (non-wiki) → stash authoring
+    // standards. Resolved once for this single call.
+    const standardsContext = resolveStashStandards(stash);
     const baseUserPrompt = buildDistillPrompt({
         inputRef,
         assetContent,
         feedback,
         proposalKind: effectiveProposalKind,
         ...(rejectedForRef.length > 0 ? { rejectedProposals: rejectedForRef } : {}),
+        ...(standardsContext.trim() ? { standardsContext } : {}),
     });
     const userPrompt = clsContext ? `${baseUserPrompt}${clsContext}` : baseUserPrompt;
     const messages = [

package/dist/commands/improve/extract-prompt.js

@@ -149,6 +149,11 @@ export function buildExtractPrompt(input) {
     const ref = input.data.ref;
     const startedAt = ref.startedAt ? new Date(ref.startedAt).toISOString() : "unknown";
     const endedAt = ref.endedAt ? new Date(ref.endedAt).toISOString() : "unknown";
+    // Optional standards block — rendered to the lead-in + body when present,
+    // or an empty string (no section) when absent. Gated on non-empty.
+    const standards = input.standardsContext?.trim()
+        ? `\n## Standards to follow (the rulebook for this target)\n\n${input.standardsContext.trim()}\n`
+        : "";
     return promptTemplate
         .replace("{{HARNESS}}", ref.harness)
         .replace("{{TITLE}}", ref.title ?? "(no title)")
@@ -156,6 +161,7 @@ export function buildExtractPrompt(input) {
         .replace("{{ENDED_AT}}", endedAt)
         .replace("{{PROJECT_HINT}}", ref.projectHint ?? "(no project hint)")
         .replace("{{ALREADY_PRESERVED}}", formatAlreadyPreserved(input.inlineRefs))
+        .replace("{{STANDARDS}}", standards)
         .replace("{{TRANSCRIPT}}", formatTranscript(input.events));
 }
 /**

package/dist/commands/improve/extract.js

@@ -23,12 +23,16 @@
  *     descriptionQualityValidator passes — same pattern as the
  *     consolidate-writer fix.
  */
+import fs from "node:fs";
+import path from "node:path";
 import { assembleAsset } from "../../core/asset/asset-serialize.js";
 import { resolveStashDir, timestampForFilename } from "../../core/common.js";
 import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
 import { ConfigError, UsageError } from "../../core/errors.js";
 import { appendEvent } from "../../core/events.js";
-import { getExtractedSessionsMap, openStateDatabase, shouldSkipAlreadyExtractedSession, upsertExtractedSession, } from "../../core/state-db.js";
+import { probeLock, releaseLock, tryAcquireLockSync } from "../../core/file-lock.js";
+import { resolveStashStandards } from "../../core/standards/resolve-stash-standards.js";
+import { getExtractedSessionsMap, getLastExtractRunAt, getStateDbPath, openStateDatabase, shouldSkipAlreadyExtractedSession, upsertExtractedSession, } from "../../core/state-db.js";
 import { repairTruncatedDescription } from "../../core/text-truncation.js";
 import { warn } from "../../core/warn.js";
 import { resolveImproveProcessRunnerFromProfile, runnerIsLlm } from "../../integrations/agent/runner.js";
@@ -63,6 +67,88 @@ const DEFAULT_MIN_CONTENT_CHARS = 10;
  * processed by subsequent runs, so coverage is preserved — just spread out.
  */
 const DEFAULT_MAX_SESSIONS_PER_RUN = 25;
+/**
+ * Floor for the default discovery window (48h). When no explicit `--since` /
+ * `defaultSince` is configured, discovery looks back to the LAST recorded
+ * extract run for the harness (so an intermittently-online host that was off for
+ * days still rediscovers sessions that ended during the gap), but never LESS
+ * than this — looking back less than the prior window could drop a session that
+ * a previous run deferred via `maxSessionsPerRun`. Widening is free of redundant
+ * LLM cost: the content-hash ledger skips unchanged sessions with zero LLM calls.
+ */
+const DEFAULT_SINCE_FLOOR_MS = 48 * 60 * 60 * 1000;
+/**
+ * Staleness window for the per-session extract lock. A single session's
+ * processing is bounded by the per-session LLM timeout (default 60s) plus the
+ * session-summary call, so a lock older than this must belong to a crashed
+ * holder and is safe to reclaim.
+ */
+const EXTRACT_SESSION_LOCK_STALE_MS = 5 * 60 * 1000;
+/**
+ * Resolve the discovery `sinceMs` cutoff when no explicit `since`/`defaultSince`
+ * is set: the later of (last recorded extract run for this harness) and
+ * (now − 48h). See {@link DEFAULT_SINCE_FLOOR_MS}. Best-effort — any state.db
+ * error falls back to the 48h floor.
+ */
+function resolveDefaultSinceMs(harnessName, now, opts) {
+    const floor = now - DEFAULT_SINCE_FLOOR_MS;
+    if (opts.skipTracking)
+        return floor;
+    let db = opts.stateDb;
+    let opened = false;
+    try {
+        if (!db) {
+            db = openStateDatabase(opts.stateDbPath);
+            opened = true;
+        }
+        const lastRun = getLastExtractRunAt(db, harnessName);
+        return lastRun != null ? Math.min(lastRun, floor) : floor;
+    }
+    catch {
+        return floor;
+    }
+    finally {
+        if (opened && db) {
+            try {
+                db.close();
+            }
+            catch {
+                // best-effort close
+            }
+        }
+    }
+}
+/** Filesystem-safe per-session lock path, co-located with the state.db. */
+function getExtractSessionLockPath(harness, sessionId, stateDbPath) {
+    const safe = `${harness}-${sessionId}`.replace(/[^A-Za-z0-9._-]/g, "_");
+    return path.join(path.dirname(stateDbPath), "extract-locks", `extract-${safe}.lock`);
+}
+/**
+ * Try to claim the per-session extract lock so a concurrent extract (e.g. a
+ * session-end hook firing `--session-id` while the hourly improve pass runs
+ * discovery) cannot double-process the SAME session — duplicate LLM spend and
+ * near-duplicate proposals. Reclaims a stale lock (dead holder PID or age past
+ * {@link EXTRACT_SESSION_LOCK_STALE_MS}). Returns false when another LIVE run
+ * holds it — the caller then skips the session without any LLM call. Best-effort:
+ * any filesystem error resolves to `true` (proceed) so locking never blocks
+ * extraction outright.
+ */
+function acquireExtractSessionLock(lockPath) {
+    try {
+        fs.mkdirSync(path.dirname(lockPath), { recursive: true });
+        if (tryAcquireLockSync(lockPath, String(process.pid)))
+            return true;
+        const probe = probeLock(lockPath, { staleAfterMs: EXTRACT_SESSION_LOCK_STALE_MS });
+        if (probe.state === "held")
+            return false;
+        // absent (released between attempt + probe) or stale → reclaim and retry once.
+        releaseLock(lockPath);
+        return tryAcquireLockSync(lockPath, String(process.pid));
+    }
+    catch {
+        return true;
+    }
+}
 // ── Helpers ──────────────────────────────────────────────────────────────────
 /**
  * Parse a since-string into an absolute ms-epoch cutoff. Accepts:
@@ -180,7 +266,11 @@ triage, sessionIndexing, schemaSimilarityCtx,
 // prior row + bypass flags are threaded in from the caller. Skipping here still
 // costs ZERO LLM calls (the expensive resource #602 protects); only the cheap
 // file read is incurred.
-prior, force) {
+prior, force, 
+// Stash authoring standards (convention/meta fact bodies) for non-wiki
+// output. Resolved ONCE per run by the caller and threaded in so facts are
+// not re-read per session. Empty string when none exist.
+standardsContext) {
     const warnings = [];
     let data;
     try {
@@ -275,7 +365,12 @@ prior, force) {
             };
         }
     }
-    const prompt = buildExtractPrompt({ data, events: filtered.events, inlineRefs: data.inlineRefs });
+    const prompt = buildExtractPrompt({
+        data,
+        events: filtered.events,
+        inlineRefs: data.inlineRefs,
+        ...(standardsContext.trim() ? { standardsContext } : {}),
+    });
     // #561 — ADDITIVE session indexing. Generate + write the session asset
     // (`sessions/<harness>/<id>.md`). FAIL-OPEN: any failure only records a
     // warning; it NEVER changes the proposal/skip outcome of extract. Returns the
@@ -625,7 +720,16 @@ export async function akmExtract(options) {
         candidates = [target];
     }
     else {
-        const sinceMs = parseSinceArg(effectiveSince);
+        // No explicit `--since`/`defaultSince` → default to "since the last run"
+        // (floored at 48h) so an intermittently-online host doesn't lose sessions
+        // that ended while it was off. See {@link resolveDefaultSinceMs}.
+        const sinceMs = effectiveSince
+            ? parseSinceArg(effectiveSince)
+            : resolveDefaultSinceMs(harness.name, startMs, {
+                ...(options.stateDb ? { stateDb: options.stateDb } : {}),
+                ...(options.stateDbPath ? { stateDbPath: options.stateDbPath } : {}),
+                ...(options.skipTracking ? { skipTracking: options.skipTracking } : {}),
+            });
         candidates = harness.listSessions({
             sinceMs,
             ...(options.location ? { location: options.location } : {}),
@@ -677,6 +781,10 @@ export async function akmExtract(options) {
             embedFn: options.schemaSimilarityEmbedFn,
         };
     }
+    // Stash authoring standards (convention/meta fact bodies) for non-wiki
+    // extract output. Resolved ONCE per run and threaded into each session's
+    // prompt so facts are not re-read per session.
+    const extractStandardsContext = resolveStashStandards(stashDir);
     for (const summary of candidates) {
         // #602 — the already-extracted skip moved INTO processSession (the content
         // hash needs the session body, only available after readSession). The prior
@@ -690,8 +798,33 @@ export async function akmExtract(options) {
             topLevelWarnings.push(`Reached maxSessionsPerRun=${maxSessionsPerRun}; ${candidates.length - processedCount - skippedCount} session(s) deferred to a later run.`);
             break;
         }
+        // Q5 — per-session lock so two concurrent extracts (e.g. a session-end hook
+        // firing `--session-id` while the hourly improve discovery pass runs) can't
+        // both LLM-process the SAME session. The holder records the outcome; a
+        // second run skips without any LLM call. Engaged only for real cross-process
+        // runs (those that open their own state.db): dry-run is read-only, an
+        // injected `stateDb` handle is an in-process/test scenario with no cross-
+        // process race, and skip-tracking-off opts out entirely.
+        let sessionLockPath;
+        if (trackingEnabled && !dryRun && !options.stateDb) {
+            sessionLockPath = getExtractSessionLockPath(harness.name, summary.sessionId, options.stateDbPath ?? getStateDbPath());
+            if (!acquireExtractSessionLock(sessionLockPath)) {
+                sessions.push({
+                    sessionId: summary.sessionId,
+                    harness: harness.name,
+                    candidateCount: 0,
+                    proposalIds: [],
+                    preFilter: { inputCount: 0, outputCount: 0, truncatedCount: 0 },
+                    warnings: ["concurrent extract holds this session's lock — skipped (handled by the other run)"],
+                    skipped: true,
+                    skipReason: "locked_concurrent",
+                });
+                skippedCount += 1;
+                continue;
+            }
+        }
         try {
-            const result = await processSession(harness, summary, stashDir, config, llmConfig, chat, options.ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, minContentChars, triage, sessionIndexing, schemaSimilarityCtx, prior, options.force === true);
+            const result = await processSession(harness, summary, stashDir, config, llmConfig, chat, options.ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, minContentChars, triage, sessionIndexing, schemaSimilarityCtx, prior, options.force === true, extractStandardsContext);
             sessions.push(result);
             // #626 — triage aggregation. A session reached the triage gate only when it
             // was NOT already preempted by an earlier skip (read_failed / too_short /
@@ -779,6 +912,10 @@ export async function akmExtract(options) {
             });
             skippedCount += 1;
         }
+        finally {
+            if (sessionLockPath)
+                releaseLock(sessionLockPath);
+        }
     }
     // Close the state.db connection we opened. Callers that injected stateDb
     // via the test seam own its lifecycle.
@@ -838,13 +975,21 @@ export async function akmExtract(options) {
 export function countNewExtractCandidates(config, options = {}) {
     const extractProcess = config.profiles?.improve?.default?.processes?.extract;
     const effectiveSince = options.since ?? extractProcess?.defaultSince;
-    const sinceMs = parseSinceArg(effectiveSince);
+    // Mirror akmExtract: when no explicit window is set, default per-harness to
+    // "since the last run" (floored at 48h) instead of a fixed 24h. Keeps this
+    // gate's discovery window identical to what akmExtract will actually scan.
+    const explicitSinceMs = effectiveSince ? parseSinceArg(effectiveSince) : undefined;
     const harnesses = (options.harnesses ?? getAvailableHarnesses()).filter((h) => h.isAvailable());
     let stateDb = options.stateDb;
     let openedStateDb = false;
     let total = 0;
     try {
         for (const harness of harnesses) {
+            const sinceMs = explicitSinceMs ??
+                resolveDefaultSinceMs(harness.name, Date.now(), {
+                    ...(options.stateDb ? { stateDb: options.stateDb } : {}),
+                    ...(options.stateDbPath ? { stateDbPath: options.stateDbPath } : {}),
+                });
             const candidates = harness.listSessions({ sinceMs });
             if (candidates.length === 0)
                 continue;

package/dist/commands/improve/improve.js

@@ -1887,7 +1887,10 @@ async function runImprovePreparationStage(args) {
     // memory mtimes, so a skipped extract never flags work for the NEXT run's
     // consolidation mtime-gate (the downstream trigger #554 asks us to suppress).
     const EXTRACT_DEFAULT_MIN_NEW_SESSIONS = 0;
-    const configuredMinNewSessions = extractConfig.profiles?.improve?.default?.processes?.extract?.minNewSessions;
+    // Read from the ACTIVE resolved profile (not always `default`), matching how
+    // `extract.enabled` resolves — otherwise a non-default profile (e.g.
+    // `frequent`) setting `minNewSessions` was silently ignored.
+    const configuredMinNewSessions = improveProfile.processes?.extract?.minNewSessions;
     const minNewSessions = typeof configuredMinNewSessions === "number" ? configuredMinNewSessions : EXTRACT_DEFAULT_MIN_NEW_SESSIONS;
     // #593/#594: the ACTIVE resolved improve profile is the single source of
     // truth for whether extract runs. (Previously this also ANDed in the legacy
@@ -1904,6 +1907,11 @@ async function runImprovePreparationStage(args) {
             const countFn = options.extractCandidateCountFn ?? countNewExtractCandidates;
             const newCandidateCount = countFn(extractConfig, {
                 ...(options.extractHarnesses ? { harnesses: options.extractHarnesses } : {}),
+                // Use the ACTIVE profile's discovery window so the gate counts over the
+                // same window akmExtract will scan (not always `default`).
+                ...(improveProfile.processes?.extract?.defaultSince
+                    ? { since: improveProfile.processes.extract.defaultSince }
+                    : {}),
                 // C2: pin the candidate-count state.db open to the boundary-resolved path.
                 ...(eventsCtx?.dbPath ? { stateDbPath: eventsCtx.dbPath } : {}),
             });
@@ -2455,6 +2463,14 @@ async function runImprovePreparationStage(args) {
     // Cap: at most 10% of the effective run limit so the lane cannot crowd out
     // reactive feedback. Requires state.db to have an asset_salience row — refs
     // without a row (pre-#608 assets still on the type-weight stub) are skipped.
+    //
+    // Cooldown: a ref qualifies at most once — when no prior reflect proposal
+    // exists for it (`!lastReflectProposalTs.has`). Without this guard the lane
+    // re-selects the same high-salience refs on EVERY run (auto-accept emits a
+    // `promoted` event, not `feedback`, so the ref never leaves
+    // noFeedbackCandidates), burning LLM calls and churning the asset. This
+    // mirrors the P0-A high-retrieval gate's `!lastReflectProposalTs.has(r.ref)`
+    // guard above so all "rescue" lanes share the same once-per-asset semantics.
     const highSalienceRefs = [];
     const salienceCfg = (options.config ?? loadConfig()).improve?.salience;
     const salienceThreshold = salienceCfg?.salienceThreshold ?? 0.75;
@@ -2470,7 +2486,7 @@ async function runImprovePreparationStage(args) {
                 if (highSalienceRefs.length >= highSalienceCap)
                     break;
                 const row = getAssetSalience(dbForHighSalience, r.ref);
-                if (row && row.encoding_salience >= salienceThreshold) {
+                if (row && row.encoding_salience >= salienceThreshold && !lastReflectProposalTs.has(r.ref)) {
                     highSalienceRefs.push(r);
                 }
             }

package/dist/commands/improve/procedural.js

@@ -27,6 +27,7 @@ import { resolveStashDir } from "../../core/common.js";
 import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
 import { appendEvent } from "../../core/events.js";
 import { parseEmbeddedJsonResponse } from "../../core/parse.js";
+import { resolveStashStandards } from "../../core/standards/resolve-stash-standards.js";
 import { warn } from "../../core/warn.js";
 import { closeDatabase, getAllEntries, openExistingDatabase } from "../../indexer/db/db.js";
 import { resolveImproveProcessRunnerFromProfile, runnerIsLlm } from "../../integrations/agent/runner.js";
@@ -144,12 +145,17 @@ export function deriveProceduralWorkflowRef(cluster) {
 }
 // ── Prompt + parse ──────────────────────────────────────────────────────────────
 /** Assemble the per-sequence user prompt fed to the procedural LLM. */
-export function buildProceduralPrompt(cluster) {
+export function buildProceduralPrompt(cluster, standardsContext = "") {
     const lines = [
         `A recurring successful action sequence observed across ${cluster.members.length} sessions.`,
         "",
-        "Ordered actions (turn EACH into exactly one step, in this order):",
     ];
+    if (standardsContext.trim()) {
+        lines.push("Standards to follow (the rulebook for this target):");
+        lines.push(standardsContext.trim());
+        lines.push("");
+    }
+    lines.push("Ordered actions (turn EACH into exactly one step, in this order):");
     cluster.normalized.forEach((step, i) => {
         lines.push(`${i + 1}. ${step}`);
     });
@@ -312,6 +318,9 @@ export async function akmProcedural(opts) {
         warnings.push("procedural: no LLM configured — skipping");
         return finish({ sequencesScanned, clustersFormed: 0 });
     }
+    // Procedural output is a workflow (non-wiki) → stash authoring standards.
+    // Resolved ONCE per run and passed to each sequence prompt.
+    const standardsContext = resolveStashStandards(stashDir);
     for (const cluster of clusters) {
         if (opts.signal?.aborted) {
             warnings.push("aborted-mid-run");
@@ -319,7 +328,7 @@ export async function akmProcedural(opts) {
         }
         clustersFormed += 1;
         const workflowRef = deriveProceduralWorkflowRef(cluster);
-        const prompt = buildProceduralPrompt(cluster);
+        const prompt = buildProceduralPrompt(cluster, standardsContext);
         const raw = await llmFn(prompt);
         const doc = parseProceduralWorkflow(raw, cluster.normalized.length);
         if (!doc) {

package/dist/commands/improve/recombine.js

@@ -45,6 +45,7 @@ import { resolveStashDir } from "../../core/common.js";
 import { getDefaultLlmConfig, loadConfig } from "../../core/config/config.js";
 import { appendEvent } from "../../core/events.js";
 import { parseEmbeddedJsonResponse } from "../../core/parse.js";
+import { resolveStashStandards } from "../../core/standards/resolve-stash-standards.js";
 import { decayUnseenRecombineHypotheses, findMatchingRecombineHypothesis, getRecombineHypothesis, getStateDbPath, markRecombineHypothesisPromoted, openStateDatabase, recordRecombineInduction, } from "../../core/state-db.js";
 import { warn } from "../../core/warn.js";
 import { closeDatabase, getAllEntries, getEntitiesByEntryIds, openExistingDatabase, } from "../../indexer/db/db.js";
@@ -228,12 +229,17 @@ function readBody(entry) {
     }
 }
 /** Assemble the per-cluster user prompt fed to the recombine LLM. */
-export function buildClusterPrompt(cluster) {
+export function buildClusterPrompt(cluster, standardsContext = "") {
     const lines = [
         `Shared signal: ${cluster.signature}`,
         `Cluster of ${cluster.members.length} related memories:`,
         "",
     ];
+    if (standardsContext.trim()) {
+        lines.push("Standards to follow (the rulebook for this target):");
+        lines.push(standardsContext.trim());
+        lines.push("");
+    }
     for (const m of cluster.members) {
         lines.push(`[memory:${m.entry.name}]`);
         if (m.entry.description)
@@ -399,6 +405,9 @@ export async function akmRecombine(opts) {
     // Refs re-induced (defensible generalization passed the quality gate) THIS
     // run — everything else is decayed after the loop.
     const seenThisRun = new Set();
+    // Recombine output is knowledge/lesson (non-wiki) → stash authoring
+    // standards. Resolved ONCE per run and passed to each cluster prompt.
+    const standardsContext = resolveStashStandards(stashDir);
     try {
         for (const cluster of clusters) {
             if (opts.signal?.aborted) {
@@ -406,7 +415,7 @@ export async function akmRecombine(opts) {
                 break;
             }
             clustersFormed += 1;
-            const prompt = buildClusterPrompt(cluster);
+            const prompt = buildClusterPrompt(cluster, standardsContext);
             const raw = await llmFn(prompt);
             const generalization = parseGeneralization(raw);
             if (!generalization) {

package/dist/commands/improve/reflect.js

@@ -34,6 +34,7 @@ import { loadConfig } from "../../core/config/config.js";
 import { ConfigError, UsageError } from "../../core/errors.js";
 import { appendEvent, readEvents } from "../../core/events.js";
 import { lintLessonContent } from "../../core/lesson-lint.js";
+import { resolveStandardsContext } from "../../core/standards/resolve-standards-context.js";
 import { lookup } from "../../indexer/indexer.js";
 import { runAgent, } from "../../integrations/agent/index.js";
 import { resolveProcessAgentProfile } from "../../integrations/agent/config.js";
@@ -745,6 +746,9 @@ export async function akmReflect(options = {}) {
     // Reflexion-style verbal-RL: inject rejected proposals so the agent avoids
     // reproducing proposals that have already been reviewed and refused.
     const rejectedProposals = readRejectedProposals(stash, options.ref);
+    // Standards "rulebook" for this target — wiki schema (wiki page) or stash
+    // convention/meta facts (non-wiki asset); empty when neither fires.
+    const standardsContext = resolveStandardsContext(options.ref, stash);
     // 5. Spawn the agent — with optional Self-Refine loop (R-1 / #372).
     //
     // maxRefineIters controls how many agent invocations are made:
@@ -813,6 +817,7 @@ export async function akmReflect(options = {}) {
                 ...(schemaHints.length > 0 ? { schemaHints } : {}),
                 ...(relatedLessons.length > 0 ? { relatedLessons } : {}),
                 ...(options.task ? { task: options.task } : {}),
+                ...(standardsContext.trim() ? { standardsContext } : {}),
                 ...(options.avoidPatterns && options.avoidPatterns.length > 0 ? { avoidPatterns: options.avoidPatterns } : {}),
                 ...(rejectedProposals.length > 0 ? { rejectedProposals } : {}),
                 // R-1: inject prior draft as self-critique target on iterations > 0

package/dist/commands/proposal/drain.js

@@ -370,6 +370,13 @@ export async function drainProposals(opts, promoteFn = akmProposalAccept, reject
     // — these are re-stamped `no-judge-configured` when no runner resolves them.
     const needsJudge = new Set();
     for (const proposal of pending) {
+        // Do NOT reclassify a proposal that was already conclusively stamped
+        // `auto-rejected` by a prior gate run (e.g. the improve confidence gate).
+        // Overwriting an authoritative rejection with `auto-accepted` would corrupt
+        // the audit trail and silently promote content the gate explicitly rejected.
+        // Such proposals remain pending for manual review (or TTL expiry).
+        if (proposal.gateDecision?.outcome === "auto-rejected")
+            continue;
         const decision = classifyProposal(proposal, opts.policy, opts.maxDiffLines);
         if (decision === null)
             continue;

package/dist/commands/proposal/propose.js

@@ -18,6 +18,7 @@ import { TYPE_DIRS } from "../../core/asset/asset-spec.js";
 import { resolveStashDir } from "../../core/common.js";
 import { ConfigError, UsageError } from "../../core/errors.js";
 import { appendEvent } from "../../core/events.js";
+import { resolveStandardsContext } from "../../core/standards/resolve-standards-context.js";
 import { runAgent, } from "../../integrations/agent/index.js";
 import { resolveProcessAgentProfile } from "../../integrations/agent/config.js";
 import { buildProposePrompt, parseAgentProposalPayload } from "../../integrations/agent/prompts.js";
@@ -93,10 +94,14 @@ export async function akmPropose(options) {
     // directly using its file tools rather than returning JSON via stdout.
     const draftFilePath = import("node:os").then((os) => import("node:path").then((path) => path.join(os.tmpdir(), `akm-propose-${options.type}-${options.name.replace(/[^a-z0-9_-]/gi, "_")}-${Date.now()}.md`)));
     const resolvedDraftPath = await draftFilePath;
+    // Standards "rulebook" for this target — wiki schema (wiki page) or stash
+    // convention/meta facts (non-wiki asset); empty when neither fires.
+    const standardsContext = resolveStandardsContext(`${options.type}:${options.name}`, stash);
     const prompt = buildProposePrompt({
         type: options.type,
         name: options.name,
         task: options.task,
+        ...(standardsContext.trim() ? { standardsContext } : {}),
         draftFilePath: resolvedDraftPath,
     });
     // 4. Spawn the agent.