akm-cli 0.9.0-beta.1 → 0.9.0-beta.10

package/dist/commands/read/curate.js

@@ -54,6 +54,9 @@ function logCurateEvent(query, result) {
     try {
         const db = openExistingDatabase();
         try {
+            // Summary row (entry_ref = NULL): preserves the query → itemRefs audit
+            // trail. Retrieval counting ignores NULL-ref rows, so this row is purely
+            // informational.
             insertUsageEvent(db, {
                 event_type: "curate",
                 query,
@@ -63,6 +66,20 @@ function logCurateEvent(query, result) {
                 }),
                 source: "user",
             });
+            // Per-item rows with entry_ref populated so curation registers as a real
+            // retrieval signal in getRetrievalCounts (which counts 'curate' events).
+            // Only stash items expose a canonical asset ref; registry hits
+            // (`registry:<id>`) have no asset ref and are skipped here.
+            for (const item of result.items) {
+                if (!("ref" in item) || typeof item.ref !== "string")
+                    continue;
+                insertUsageEvent(db, {
+                    event_type: "curate",
+                    query,
+                    entry_ref: item.ref,
+                    source: "user",
+                });
+            }
         }
         finally {
             closeDatabase(db);

package/dist/commands/remember.js

@@ -149,6 +149,10 @@ export async function runLlmEnrich(body) {
         return { tags: [] };
     }
     const { chatCompletion, parseEmbeddedJsonResponse: parseJsonResponse } = await import("../llm/client.js");
+    // #576: attribute this entry point's LLM call to the `remember` stage. The
+    // wrapper is ambient — if a usage sink is active it tags the record; if not,
+    // it is a no-op.
+    const { withLlmStage } = await import("../llm/usage-telemetry.js");
     const prompt = `You are a memory tagger for a developer knowledge base.
 Given the memory text below, return ONLY a JSON object with these fields:
 - "tags": array of 1-5 short lowercase keyword tags
@@ -164,10 +168,10 @@ Return ONLY the JSON object, no prose, no markdown fences.`;
         const result = await (async () => {
             try {
                 return await Promise.race([
-                    chatCompletion(llmConfig, [
+                    withLlmStage("remember", () => chatCompletion(llmConfig, [
                         { role: "system", content: "Return only valid JSON. No prose." },
                         { role: "user", content: prompt },
-                    ], { maxTokens: 256, temperature: 0.1 }),
+                    ], { maxTokens: 256, temperature: 0.1 })),
                     new Promise((_, reject) => {
                         timeoutHandle = setTimeout(() => reject(new Error("LLM enrichment timed out")), LLM_ENRICH_TIMEOUT_MS);
                     }),

package/dist/commands/sources/stash-cli.js

@@ -71,6 +71,11 @@ export const indexCommand = defineCommand({
             description: "When combined with --clean, report stale entries without deleting them.",
             default: false,
         },
+        background: {
+            type: "boolean",
+            description: "Run as a background process (suppresses interactive output, manages PID file).",
+            default: false,
+        },
     },
     async run({ args }) {
         await runWithJsonErrors(async () => {
@@ -80,6 +85,7 @@ export const indexCommand = defineCommand({
             if (getHyphenatedBoolean(args, "re-enrich") || parseFlagValue(process.argv, "--re-enrich") !== undefined) {
                 throw new UsageError("`akm index --re-enrich` has been removed. Re-enrichment of index-time LLM passes is not exposed in this slice.");
             }
+            const isBackground = args.background === true;
             const outputMode = getOutputMode();
             const controller = new AbortController();
             const abort = () => controller.abort(new Error("index interrupted"));
@@ -88,7 +94,7 @@ export const indexCommand = defineCommand({
             const indexLogFile = path.join(getCacheDir(), "logs", "index", `${new Date().toISOString().replace(/[:.]/g, "-")}.log`);
             setLogFile(indexLogFile);
             const verbose = isVerbose();
-            const spin = !verbose && outputMode.format === "text" ? p.spinner() : null;
+            const spin = !verbose && !isBackground && outputMode.format === "text" ? p.spinner() : null;
             if (spin) {
                 spin.start(`Building search index${args.full ? " (full rebuild)" : ""}...`);
             }
@@ -114,7 +120,9 @@ export const indexCommand = defineCommand({
                 if (spin) {
                     spin.stop(`Indexed ${result.totalEntries} assets.`);
                 }
-                output("index", result);
+                if (!isBackground) {
+                    output("index", result);
+                }
             }
             catch (error) {
                 if (spin) {

package/dist/commands/tasks/tasks.js

@@ -218,7 +218,13 @@ export async function akmTasksSetEnabled(id, enabled) {
     fs.writeFileSync(filePath, updated, "utf8");
     const sched = selectBackend();
     try {
-        await sched.setEnabled(normalised, enabled);
+        // Reinstall from the (just-updated) definition rather than only toggling
+        // the comment. A plain toggle leaves a stale schedule in place if the
+        // .yml's `schedule:` changed while the task was disabled — re-enabling
+        // would silently keep the old cron line. install() renders the block with
+        // both the current schedule and the new enabled state, and is idempotent.
+        const task = parseTaskDocument({ yaml: updated, filePath, id: normalised });
+        await sched.install(task);
     }
     catch (err) {
         // Roll the file back so the YAML source-of-truth and the OS
@@ -254,9 +260,12 @@ export async function akmTasksHistory(input) {
  * Reconcile the on-disk task files with the OS scheduler.
  *   • install missing tasks (after validating them — invalid files are
  *     skipped with a per-task reason rather than aborting the whole sync)
+ *   • reinstall tasks whose schedule or enabled state changed in the .yml
+ *     (drift detected by comparing the backend's installed signature against
+ *     the signature the current definition would produce)
  *   • remove orphan scheduler entries that no longer have a backing file
  */
-export async function akmTasksSync() {
+export async function akmTasksSync(deps = {}) {
     const stashDir = resolveStashDir();
     const typeRoot = path.join(stashDir, "tasks");
     if (fs.existsSync(typeRoot))
@@ -267,10 +276,13 @@ export async function akmTasksSync() {
             .filter((f) => f.endsWith(".yml"))
             .map((f) => f.slice(0, -4))
         : [];
-    const sched = selectBackend();
+    const sched = deps.backend ?? selectBackend();
     const backend = backendNameForPlatform();
-    const present = new Set((await sched.list()).map((t) => t.id));
+    // Map id → installed signature so sync can detect schedule/enabled drift on
+    // tasks that already exist in the scheduler, not just presence/absence.
+    const present = new Map((await sched.list()).map((t) => [t.id, t.signature]));
     const installed = [];
+    const updated = [];
     const unchanged = [];
     const skipped = [];
     for (const id of fileIds) {
@@ -290,22 +302,34 @@ export async function akmTasksSync() {
             skipped.push({ id, reason: err instanceof Error ? err.message : String(err) });
             continue;
         }
-        if (present.has(id)) {
+        if (!present.has(id)) {
+            await sched.install(task);
+            installed.push(id);
+            continue;
+        }
+        // Already installed — reconcile against the current definition. Compare the
+        // installed signature to what this task would render to; reinstall on drift.
+        // When the backend can't produce a signature (no expectedSignature, or it
+        // didn't record one), reinstall unconditionally — install() is idempotent,
+        // so the cost is one crontab write and correctness is guaranteed.
+        const installedSig = present.get(id);
+        const expectedSig = sched.expectedSignature?.(task);
+        if (installedSig !== undefined && expectedSig !== undefined && installedSig === expectedSig) {
             unchanged.push(id);
         }
         else {
             await sched.install(task);
-            installed.push(id);
+            updated.push(id);
         }
     }
     const removed = [];
-    for (const installedId of present) {
+    for (const installedId of present.keys()) {
         if (!fileIds.includes(installedId)) {
             await sched.uninstall(installedId);
             removed.push(installedId);
         }
     }
-    return { installed, removed, unchanged, skipped, backend: sched.name };
+    return { installed, updated, removed, unchanged, skipped, backend: sched.name };
 }
 export async function akmTasksDoctor() {
     const warnings = [];

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

@@ -138,7 +138,35 @@ export const ImproveProcessConfigSchema = z
     // Extract process config (only meaningful for extract process)
     defaultSince: z.string().min(1).optional(),
     maxTotalChars: positiveInt.optional(),
+    // Extract process: minimum raw session size (pre-filter inputCount) below
+    // which the extract LLM call is skipped (#595/#596). 0 disables the gate.
+    // Absent = default 10 (skip only truly empty sessions). Only meaningful
+    // on the `extract` process.
+    minContentChars: z.number().int().min(0).optional(),
     maxChunkSize: z.number().int().min(1).max(50).optional(),
+    // Consolidate process: narrow candidate pool to memories modified within
+    // this duration window plus their graph neighbours. Only meaningful on
+    // the `consolidate` process. Absent = full-pool sweep.
+    incrementalSince: z.string().optional(),
+    // Consolidate process: hard cap on memories processed per pass.
+    // Reflect/distill: max refs processed (same as profile-level `limit`).
+    limit: positiveInt.optional(),
+    // Consolidate process: graph neighbours per changed memory during
+    // incremental consolidation. Default 5. Only meaningful with incrementalSince.
+    neighborsPerChanged: z.number().int().min(1).optional(),
+    // Distill process: skip distill entirely when reflect produced zero planned refs.
+    requirePlannedRefs: z.boolean().optional(),
+    // proactiveMaintenance process (Layer 2): staleness gate + rotation cooldown
+    // in days (default 30). Only meaningful on `proactiveMaintenance`.
+    dueDays: z.number().int().min(0).optional(),
+    // proactiveMaintenance process: top-N bound per run (default 25). Alias for
+    // `limit`; `maxPerRun` wins when both are set.
+    maxPerRun: positiveInt.optional(),
+    // proactiveMaintenance process: optional per-type importance overrides,
+    // merged over the built-in defaults. Only meaningful on `proactiveMaintenance`.
+    importanceWeights: z.record(z.string().min(1), z.number()).optional(),
+    // MemoryInference process: minimum pending memory count to run the pass.
+    minPendingCount: z.number().int().min(0).optional(),
     // Extract process: minimum number of new (unseen, in-window) candidate
     // sessions below which the extract pass skips entirely (emits an
     // `improve_skipped` event with `reason: "below_min_new_sessions"`). 0
@@ -176,6 +204,7 @@ const ImproveProfileProcessesSchema = z
     graphExtraction: ImproveProcessConfigSchema.optional(),
     validation: ImproveProcessConfigSchema.optional(),
     triage: ImproveProcessConfigSchema.optional(),
+    proactiveMaintenance: ImproveProcessConfigSchema.optional(),
 })
     .passthrough()
     .superRefine((val, ctx) => {
@@ -199,6 +228,7 @@ const ImproveProfileProcessesSchema = z
         "validation",
         "extract",
         "triage",
+        "proactiveMaintenance",
     ]);
     for (const k of Object.keys(raw)) {
         if (!allowed.has(k)) {

package/dist/core/logs-db.js

@@ -0,0 +1,304 @@
+// 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/.
+/**
+ * logs.db — Dedicated SQLite database for task/run log lines (#579).
+ *
+ * Replaces grep-the-flat-file consumption of `<cacheDir>/tasks/logs/<id>/<ts>.log`
+ * with structured, indexed rows: `{ts, task_id, run_id, stream, level, line}`.
+ * The strategic direction (stop scattering data across files/folders) means
+ * every NEW log consumer queries this database; the per-run text file written
+ * by the task runner is retained only as a transitional tail for humans —
+ * see docs/technical/logs-audit.md for the full producer audit.
+ *
+ * ## Why a separate database from state.db
+ *
+ * Log lines are high-volume, append-only, and freely purgeable; state.db rows
+ * (events, proposals, task_history) are durable records. Separating them keeps
+ * state.db small and lets log retention be aggressive without touching durable
+ * state. Cross-db queries (e.g. "failed task_history row → its log lines") use
+ * SQLite ATTACH — see {@link attachStateDatabase}.
+ *
+ * ## run_id
+ *
+ * state.db's `task_history` identifies a run by the unique pair
+ * `(task_id, started_at)` (see migration 002 in state-db.ts). logs.db encodes
+ * that pair as a single string — {@link buildTaskRunId} — so log rows can be
+ * joined back to their history row:
+ *
+ *   l.run_id = th.task_id || '@' || th.started_at
+ *
+ * ## Schema evolution
+ *
+ * Same migration-safety contract as state.db: append-only `MIGRATIONS` applied
+ * through the shared runner in src/storage/engines/sqlite-migrations.ts.
+ *
+ * @module logs-db
+ */
+import fs from "node:fs";
+import path from "node:path";
+import { openDatabase } from "../storage/database.js";
+import { runMigrations as runSqliteMigrations } from "../storage/engines/sqlite-migrations.js";
+import { getDataDir } from "./paths.js";
+import { getStateDbPath } from "./state-db.js";
+// ── Path helper ──────────────────────────────────────────────────────────────
+/**
+ * Default path: `<dataDir>/logs.db` — alongside state.db so cooperating
+ * processes sharing a data root automatically share the same logs database
+ * (same `AKM_DATA_DIR` / XDG env-isolation as {@link getStateDbPath}).
+ */
+export function getLogsDbPath() {
+    return path.join(getDataDir(), "logs.db");
+}
+// ── Database open ────────────────────────────────────────────────────────────
+/**
+ * Open (and initialise / migrate) the logs database.
+ *
+ * @param dbPath - Override the database file path (tests pass a tmpdir path).
+ *
+ * PRAGMA rationale:
+ *
+ *   journal_mode = WAL
+ *     Readers never block writers and vice-versa; crashes are safe (the WAL is
+ *     replayed on next open). Required because the task runner writes log rows
+ *     while `akm health` may be reading them.
+ *
+ *   busy_timeout = 30000
+ *     Log writes happen at the end of scheduled task runs, which can pile up
+ *     (cron fan-out). 30 s of retry absorbs a slow concurrent writer instead of
+ *     surfacing SQLITE_BUSY and dropping log lines.
+ */
+export function openLogsDatabase(dbPath) {
+    const resolvedPath = dbPath ?? getLogsDbPath();
+    const dir = path.dirname(resolvedPath);
+    if (!fs.existsSync(dir)) {
+        fs.mkdirSync(dir, { recursive: true });
+    }
+    const db = openDatabase(resolvedPath);
+    // PRAGMAs must run before any DDL or DML.
+    db.exec("PRAGMA journal_mode = WAL");
+    db.exec("PRAGMA busy_timeout = 30000");
+    runMigrations(db);
+    return db;
+}
+// ── Migrations ───────────────────────────────────────────────────────────────
+/**
+ * All migrations in application order. APPEND only — never insert in the
+ * middle or reorder. Same contract as state.db's MIGRATIONS array.
+ */
+const MIGRATIONS = [
+    // ── Migration 001 — task_logs ───────────────────────────────────────────────
+    //
+    // One row per log line emitted by a task run.
+    //
+    // Indexed (query) columns:
+    //   ts       TEXT — ISO-8601 UTC; range queries ("logs in the last hour").
+    //   task_id  TEXT — task identifier; per-task log views.
+    //   run_id   TEXT — buildTaskRunId(task_id, started_at); per-run log views
+    //                   and the join key back to state.db task_history.
+    //
+    // Non-indexed columns:
+    //   stream   TEXT — 'stdout' | 'stderr'; which pipe the line came from.
+    //   level    TEXT — 'info' | 'warn' | 'error'; runner-assigned severity
+    //                   ('info' for captured stdout, 'error' for stderr and
+    //                   failure diagnostics).
+    //   line     TEXT — the log line itself (no trailing newline).
+    //
+    // ADD COLUMN extension points (future migrations):
+    //   ALTER TABLE task_logs ADD COLUMN seq INTEGER DEFAULT NULL;
+    //   ALTER TABLE task_logs ADD COLUMN source TEXT DEFAULT NULL;
+    //
+    // TTL: rows where ts < NOW() - retention can be deleted by purgeOldTaskLogs().
+    // No automatic deletion occurs here.
+    {
+        id: "001-task-logs",
+        up: `
+      CREATE TABLE IF NOT EXISTS task_logs (
+        id      INTEGER PRIMARY KEY AUTOINCREMENT,
+        ts      TEXT    NOT NULL,
+        task_id TEXT    NOT NULL,
+        run_id  TEXT    NOT NULL,
+        stream  TEXT    NOT NULL DEFAULT 'stdout',
+        level   TEXT    NOT NULL DEFAULT 'info',
+        line    TEXT    NOT NULL
+      );
+
+      -- Query patterns:
+      --   SELECT … WHERE ts >= ? AND ts <= ?   → idx_task_logs_ts (purge, windows)
+      --   SELECT … WHERE task_id = ?           → idx_task_logs_task_id
+      --   SELECT … WHERE run_id = ?            → idx_task_logs_run_id (per-run tail)
+      CREATE INDEX IF NOT EXISTS idx_task_logs_ts      ON task_logs(ts);
+      CREATE INDEX IF NOT EXISTS idx_task_logs_task_id ON task_logs(task_id);
+      CREATE INDEX IF NOT EXISTS idx_task_logs_run_id  ON task_logs(run_id);
+    `,
+    },
+];
+/**
+ * Apply every pending migration. Called automatically by
+ * {@link openLogsDatabase}; exported for the same test seams state-db exposes.
+ */
+export function runMigrations(db) {
+    runSqliteMigrations(db, MIGRATIONS);
+}
+// ── run_id ───────────────────────────────────────────────────────────────────
+/**
+ * Encode a task run's identity — the unique `(task_id, started_at)` pair from
+ * state.db `task_history` — as a single run_id string.
+ *
+ * The format MUST stay in sync with the SQL expression
+ * `task_id || '@' || started_at` used by {@link queryFailedRunLogLines}.
+ */
+export function buildTaskRunId(taskId, startedAtIso) {
+    return `${taskId}@${startedAtIso}`;
+}
+/**
+ * Insert a batch of log lines for one task run in a single transaction.
+ * Returns the number of rows inserted. Lines are stored in array order
+ * (ascending rowid), so reading back `ORDER BY id` reproduces emission order.
+ *
+ * Errors propagate — the task runner wraps this in its own best-effort
+ * handling (mirroring `appendHistory`) so an unwritable logs.db never fails
+ * a task run.
+ */
+export function insertTaskLogLines(db, input) {
+    if (input.lines.length === 0)
+        return 0;
+    const stmt = db.prepare(`INSERT INTO task_logs (ts, task_id, run_id, stream, level, line)
+     VALUES (?, ?, ?, ?, ?, ?)`);
+    db.transaction(() => {
+        for (const entry of input.lines) {
+            stmt.run(input.ts, input.taskId, input.runId, entry.stream ?? "stdout", entry.level ?? "info", entry.line);
+        }
+    })();
+    return input.lines.length;
+}
+/**
+ * Read log lines matching the filter, in emission order (ascending id).
+ *
+ * Connection-lifetime rule (WS5): `.all()` materializes a plain array before
+ * returning.
+ */
+export function queryTaskLogs(db, options = {}) {
+    const conditions = [];
+    const params = [];
+    if (options.taskId) {
+        conditions.push("task_id = ?");
+        params.push(options.taskId);
+    }
+    if (options.runId) {
+        conditions.push("run_id = ?");
+        params.push(options.runId);
+    }
+    if (options.stream) {
+        conditions.push("stream = ?");
+        params.push(options.stream);
+    }
+    if (options.since) {
+        conditions.push("ts >= ?");
+        params.push(options.since);
+    }
+    if (options.until) {
+        conditions.push("ts < ?");
+        params.push(options.until);
+    }
+    const where = conditions.length > 0 ? `WHERE ${conditions.join(" AND ")}` : "";
+    const limit = options.limit !== undefined && options.limit >= 0 ? ` LIMIT ${Math.floor(options.limit)}` : "";
+    return db
+        .prepare(`SELECT id, ts, task_id, run_id, stream, level, line FROM task_logs ${where} ORDER BY id ASC${limit}`)
+        .all(...params);
+}
+/**
+ * Bulk membership check: which of `runIds` have at least one log row?
+ * Used by `akm health` to compute the log-backing rate from the database
+ * instead of `fs.existsSync` over scattered files. Chunked to stay under
+ * SQLite's bound-parameter ceiling.
+ */
+export function getLoggedRunIds(db, runIds) {
+    const out = new Set();
+    if (runIds.length === 0)
+        return out;
+    const CHUNK = 500;
+    for (let i = 0; i < runIds.length; i += CHUNK) {
+        const chunk = runIds.slice(i, i + CHUNK);
+        const placeholders = chunk.map(() => "?").join(",");
+        const rows = db
+            .prepare(`SELECT DISTINCT run_id FROM task_logs WHERE run_id IN (${placeholders})`)
+            .all(...chunk);
+        for (const row of rows)
+            out.add(row.run_id);
+    }
+    return out;
+}
+// ── Cross-db: ATTACH state.db ────────────────────────────────────────────────
+/**
+ * ATTACH state.db to an open logs.db handle under the schema name `state`,
+ * enabling cross-db joins like task_history × task_logs.
+ *
+ * The state.db file must already exist (callers always open state.db first in
+ * practice); attaching a non-existent path would silently create an empty,
+ * unmigrated database file, so this throws instead.
+ */
+export function attachStateDatabase(db, stateDbPath) {
+    const resolved = stateDbPath ?? getStateDbPath();
+    if (!fs.existsSync(resolved)) {
+        throw new Error(`Cannot ATTACH state.db: file does not exist at ${resolved}`);
+    }
+    // prepare().run() rather than db.run(): both drivers support parameterised
+    // ATTACH through a prepared statement, and no other call site uses db.run().
+    db.prepare("ATTACH DATABASE ? AS state").run(resolved);
+}
+/**
+ * Convenience: open logs.db with state.db attached as `state`. The returned
+ * handle supports cross-db queries such as {@link queryFailedRunLogLines}.
+ * Close it like any other handle (DETACH is implicit on close).
+ */
+export function openLogsDatabaseWithState(logsDbPath, stateDbPath) {
+    const db = openLogsDatabase(logsDbPath);
+    try {
+        attachStateDatabase(db, stateDbPath);
+    }
+    catch (err) {
+        db.close();
+        throw err;
+    }
+    return db;
+}
+/**
+ * Cross-db join: every log line belonging to a FAILED task_history run whose
+ * `started_at` is `>= since` (all failed runs when omitted). Requires a handle
+ * opened via {@link openLogsDatabaseWithState}.
+ *
+ * The join key is the run_id encoding documented on {@link buildTaskRunId}:
+ * `task_logs.run_id = task_history.task_id || '@' || task_history.started_at`.
+ */
+export function queryFailedRunLogLines(db, options = {}) {
+    const conditions = ["th.status = 'failed'"];
+    const params = [];
+    if (options.since) {
+        conditions.push("th.started_at >= ?");
+        params.push(options.since);
+    }
+    const limit = options.limit !== undefined && options.limit >= 0 ? ` LIMIT ${Math.floor(options.limit)}` : "";
+    return db
+        .prepare(`SELECT th.task_id, l.run_id, th.started_at, th.status, l.ts, l.stream, l.level, l.line
+       FROM state.task_history th
+       JOIN task_logs l ON l.run_id = th.task_id || '@' || th.started_at
+       WHERE ${conditions.join(" AND ")}
+       ORDER BY th.started_at DESC, l.id ASC${limit}`)
+        .all(...params);
+}
+// ── Retention ────────────────────────────────────────────────────────────────
+/**
+ * Delete task_logs rows older than `retentionDays` (default: 90). Mirrors
+ * `purgeOldEvents` / `purgeOldImproveRuns` in state-db.ts — same default, same
+ * return shape (rows deleted), same disabled-when-non-positive semantics.
+ * Wired into the improve maintenance pass alongside the state.db purges.
+ */
+export function purgeOldTaskLogs(db, retentionDays = 90) {
+    if (!Number.isFinite(retentionDays) || retentionDays <= 0)
+        return 0;
+    const cutoff = new Date(Date.now() - retentionDays * 86_400_000).toISOString();
+    const result = db.prepare("DELETE FROM task_logs WHERE ts < ?").run(cutoff);
+    const changes = result.changes ?? 0;
+    return typeof changes === "bigint" ? Number(changes) : changes;
+}

package/dist/core/paths.js

@@ -215,6 +215,9 @@ export function getDataDir(env = process.env, platform = process.platform) {
 export function getDbPath() {
     return path.join(getDataDir(), "index.db");
 }
+export function getIndexWriterLockPath() {
+    return path.join(getDataDir(), "index.db.write.lock");
+}
 export function getWorkflowDbPath() {
     return path.join(getDataDir(), "workflow.db");
 }