akm-cli 0.9.0-beta.10 → 0.9.0-beta.11

package/CHANGELOG.md

@@ -6,6 +6,29 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/).
 
 ## [Unreleased]
 
+## [0.9.0-beta.11] - 2026-06-15
+
+### Added
+
+- **`extract.maxSessionsPerRun`** (default 25) — caps the NEW sessions the
+  extract pass LLM-processes in a single run so a backlog (e.g. after downtime)
+  can't push one run past its scheduled-task timeout. Overflow sessions stay
+  unseen and are picked up by later runs, so coverage is preserved. `0` disables.
+
+### Fixed
+
+- **Auto-accept validation failures are no longer a blind leak.** When a
+  confidence-passing proposal fails promotion validation, the gate now captures
+  the reason (the `validateProposal` finding kind, e.g. `validation:description-quality`),
+  records it on the proposal (`akm proposal show` explains the rejection), logs
+  it, and exposes `failedByReason` on the gate result — so the ~5% leak is
+  diagnosable instead of silently warned-and-dropped.
+- **Inflated skip-reason aggregates in `akm health`.** `no_new_signal` /
+  `profile_filtered_all_passes` are per-run snapshots of a stable set; the
+  window aggregator summed their per-run counts (≈2.7M / 3M). It now uses the
+  most recent run's count for these aggregated-snapshot reasons while still
+  summing genuine per-occurrence skips.
+
 ## [0.9.0-beta.10] - 2026-06-15
 
 ### Added

package/dist/commands/health.js

@@ -833,19 +833,33 @@ function computeWallTimeStats(durationsMs, byPhase) {
     };
 }
 function buildImproveSkipSummary(events) {
-    const skipReasons = {};
-    let skipped = 0;
+    // Two kinds of skip events:
+    //  - Per-occurrence (no `count`): one event per skipped ref → SUM is correct.
+    //  - Aggregated snapshot (carries `count`): a single per-run event whose count
+    //    is the number of refs that hit a STABLE, whole-stash condition that run
+    //    (`no_new_signal`, `profile_filtered_all_passes`). Each run re-counts the
+    //    same stable set, so summing across the window re-counts it N times (the
+    //    2.7M / 3M inflation). For these we keep the MOST RECENT run's count — the
+    //    current snapshot — matching how memorySummary/profileFilteredRefs are
+    //    handled. Events arrive in chronological (offset) order, so the last
+    //    count-bearing event per reason is the latest run's value.
+    const summed = {};
+    const latestSnapshot = {};
     for (const event of events) {
         const reason = typeof event.metadata?.reason === "string" && event.metadata.reason.trim() ? event.metadata.reason : "unknown";
-        // Aggregated skip events (e.g. `no_new_signal`, `profile_filtered_all_passes`)
-        // carry a `count` of the refs they represent in a single row instead of one
-        // event per ref. Honor that count so the skip histogram reflects the true
-        // number of skipped refs; per-ref events without a count contribute 1.
         const rawCount = event.metadata?.count;
-        const count = typeof rawCount === "number" && Number.isFinite(rawCount) && rawCount > 0 ? rawCount : 1;
+        if (typeof rawCount === "number" && Number.isFinite(rawCount) && rawCount > 0) {
+            latestSnapshot[reason] = rawCount; // overwrite → keeps the latest run's snapshot
+        }
+        else {
+            summed[reason] = (summed[reason] ?? 0) + 1;
+        }
+    }
+    const skipReasons = { ...summed };
+    for (const [reason, count] of Object.entries(latestSnapshot)) {
         skipReasons[reason] = (skipReasons[reason] ?? 0) + count;
-        skipped += count;
     }
+    const skipped = Object.values(skipReasons).reduce((a, b) => a + b, 0);
     return { skipped, skipReasons };
 }
 function probeStateDbRoundTrip(stateDbPath) {

package/dist/commands/improve/extract.js

@@ -51,6 +51,14 @@ const DEFAULT_MIN_SESSION_DURATION_MINUTES = 5;
  * (0 chars, journal files) are safe to skip.
  */
 const DEFAULT_MIN_CONTENT_CHARS = 10;
+/**
+ * Default cap on NEW sessions the extract pass will LLM-process in a single run
+ * (`processes.extract.maxSessionsPerRun` overrides; `0` disables). Bounds per-run
+ * wall time + token spend so a backlog of accumulated sessions can't run a single
+ * pass past its scheduled-task timeout. Overflow sessions stay unseen and are
+ * processed by subsequent runs, so coverage is preserved — just spread out.
+ */
+const DEFAULT_MAX_SESSIONS_PER_RUN = 25;
 // ── Helpers ──────────────────────────────────────────────────────────────────
 /**
  * Parse a since-string into an absolute ms-epoch cutoff. Accepts:
@@ -383,6 +391,12 @@ export async function akmExtract(options) {
     // #595/#596 — minimum raw session size; sessions below it skip the LLM call
     // entirely. Set `processes.extract.minContentChars: 0` to disable the gate.
     const minContentChars = typeof extractProcess?.minContentChars === "number" ? extractProcess.minContentChars : DEFAULT_MIN_CONTENT_CHARS;
+    // Cap on NEW sessions LLM-processed per run; 0 disables. Absent = default.
+    // Bounds per-run wall time / LLM cost so a backlog can't push a run past its
+    // task timeout — the overflow stays unseen and is picked up by later runs.
+    const maxSessionsPerRun = typeof extractProcess?.maxSessionsPerRun === "number"
+        ? extractProcess.maxSessionsPerRun
+        : DEFAULT_MAX_SESSIONS_PER_RUN;
     // Default discovery window — process config can override the built-in 24h.
     const effectiveSince = options.since ?? extractProcess?.defaultSince;
     // #561 — resolve session-indexing config. Default ON: we only reach this code
@@ -526,6 +540,13 @@ export async function akmExtract(options) {
             skippedCount += 1;
             continue;
         }
+        // Per-run cap on LLM-processed sessions (skip-tracked seen sessions above
+        // don't count). Single-session / --force modes bypass the cap (explicit
+        // intent). Overflow sessions are left unseen for the next run.
+        if (!options.sessionId && !options.force && maxSessionsPerRun > 0 && processedCount >= maxSessionsPerRun) {
+            topLevelWarnings.push(`Reached maxSessionsPerRun=${maxSessionsPerRun}; ${candidates.length - processedCount - skippedCount} session(s) deferred to a later run.`);
+            break;
+        }
         try {
             const result = await processSession(harness, summary, stashDir, config, llmConfig, chat, options.ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, minContentChars, sessionIndexing);
             sessions.push(result);

package/dist/commands/improve/improve-auto-accept.js

@@ -5,6 +5,23 @@ import { loadConfig } from "../../core/config/config.js";
 import { appendEvent } from "../../core/events.js";
 import { info, warn } from "../../core/warn.js";
 import { promoteProposal, recordGateDecision } from "../proposal/validators/proposals.js";
+/**
+ * Derive a stable, low-cardinality reason bucket from an auto-accept promotion
+ * error. `promoteProposal` throws a `validateProposal` report formatted as
+ * `[kind] message` lines; we extract the first finding kind. Non-validation
+ * throws collapse to `promote-error`.
+ */
+function classifyPromoteFailure(err) {
+    const message = err instanceof Error ? err.message : String(err);
+    const finding = /\[([a-z][a-z0-9-]*)\]/i.exec(message);
+    if (finding)
+        return `validation:${finding[1]}`;
+    if (/not pending/i.test(message))
+        return "not-pending";
+    if (/unknown asset type/i.test(message))
+        return "unknown-type";
+    return "promote-error";
+}
 // ---------------------------------------------------------------------------
 // Gate implementation
 // ---------------------------------------------------------------------------
@@ -18,7 +35,7 @@ import { promoteProposal, recordGateDecision } from "../proposal/validators/prop
  * @param promoteFn   Injectable override for `promoteProposal` (test seam).
  */
 export async function runAutoAcceptGate(candidates, cfg, promoteFn = promoteProposal) {
-    const result = { promoted: [], skipped: [], failed: [] };
+    const result = { promoted: [], skipped: [], failed: [], failedByReason: {} };
     // --- Guard: gate is disabled or context is incomplete ---
     if (cfg.dryRun || cfg.globalThreshold === undefined || !cfg.stashDir) {
         result.skipped = candidates.map((c) => c.proposalId);
@@ -84,8 +101,19 @@ export async function runAutoAcceptGate(candidates, cfg, promoteFn = promoteProp
             result.promoted.push(proposalId);
         }
         catch (err) {
-            warn(`[improve] ${cfg.phase} auto-accept failed for ${proposalId}: ${err instanceof Error ? err.message : String(err)}`);
+            const reason = classifyPromoteFailure(err);
+            warn(`[improve] ${cfg.phase} auto-accept failed for ${proposalId} (${reason}): ${err instanceof Error ? err.message : String(err)}`);
             result.failed.push(proposalId);
+            result.failedByReason[reason] = (result.failedByReason[reason] ?? 0) + 1;
+            // Record WHY on the proposal so `akm proposal show` explains the rejection
+            // and the leak is no longer blind. Best-effort.
+            stamp(proposalId, {
+                outcome: "auto-rejected",
+                reason,
+                confidence,
+                thresholds: { autoAccept: effectiveThreshold },
+                gate: gateLabel,
+            });
         }
     }
     return result;

package/dist/core/config/config-schema.js

@@ -173,6 +173,9 @@ export const ImproveProcessConfigSchema = z
     // disables the guard. Only meaningful on the `extract` process. Default 0
     // (disabled) so existing behaviour is preserved; only opted-in profiles set it.
     minNewSessions: z.number().int().min(0).optional(),
+    // Extract process: cap on NEW sessions processed (LLM-called) per run; the
+    // rest roll to the next run (still unseen). 0 disables. Absent = default 25.
+    maxSessionsPerRun: z.number().int().min(0).optional(),
     // #561 — index agent sessions as a searchable `session` asset (extract
     // process). Absent = on-when-an-LLM-is-available (fail-open when offline).
     indexSessions: z.boolean().optional(),

package/dist/scripts/migrate-storage.js

@@ -15543,6 +15543,7 @@ var init_config_schema = __esm(() => {
     importanceWeights: exports_external.record(exports_external.string().min(1), exports_external.number()).optional(),
     minPendingCount: exports_external.number().int().min(0).optional(),
     minNewSessions: exports_external.number().int().min(0).optional(),
+    maxSessionsPerRun: exports_external.number().int().min(0).optional(),
     indexSessions: exports_external.boolean().optional(),
     minSessionDuration: exports_external.number().min(0).optional(),
     applyMode: exports_external.enum(["queue", "promote"]).optional(),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "akm-cli",
-  "version": "0.9.0-beta.10",
+  "version": "0.9.0-beta.11",
   "type": "module",
   "description": "akm (Agent Knowledge Management) — A package manager for AI agent skills, commands, tools, and knowledge. Works with Claude Code, OpenCode, Cursor, and any AI coding assistant.",
   "keywords": [