akm-cli 0.9.0-beta.2 → 0.9.0-beta.3

package/dist/commands/health.js

@@ -4,11 +4,13 @@
 import fs from "node:fs";
 import { ConfigError, UsageError } from "../core/errors.js";
 import { appendEvent, readEvents } from "../core/events.js";
+import { buildTaskRunId, getLoggedRunIds, openLogsDatabase } from "../core/logs-db.js";
 import { getStateDbPathInDataDir } from "../core/paths.js";
 import { listExistingTableNames, openStateDatabase, queryCompletedTaskIntervals, queryImproveRuns, queryTaskHistory, } from "../core/state-db.js";
 import { parseSinceToIso } from "../core/time.js";
 import { readSemanticStatus } from "../indexer/search/semantic-status.js";
 import { getExecutionLogCandidates } from "../integrations/session-logs/index.js";
+import { LLM_USAGE_EVENT } from "../llm/usage-persist.js";
 import { HEALTH_CHECKS } from "./health/checks.js";
 const DEFAULT_SINCE_MS = 24 * 60 * 60 * 1000;
 const IMPROVE_COMPLETED_EVENT = "improve_completed";
@@ -882,14 +884,84 @@ function computeDeltas(first, last) {
     }
     return out;
 }
-function buildWindowMetrics(db, stateDbPath, since, until, now = () => Date.now()) {
+/**
+ * Partition task_history rows into "should have a log" (non-null log_path) and
+ * "log is actually backed". A run counts as backed when logs.db holds rows for
+ * its run_id (#579 — the DB is the primary record); rows written before logs.db
+ * existed fall back to the transitional on-disk file check. `logsDb` may be
+ * undefined when logs.db could not be opened — then only the file check runs.
+ */
+function partitionLogBackedRows(taskRows, logsDb) {
+    const withLogs = taskRows.filter((row) => row.log_path !== null);
+    const loggedRunIds = logsDb
+        ? getLoggedRunIds(logsDb, withLogs.map((row) => buildTaskRunId(row.task_id, row.started_at)))
+        : new Set();
+    const backed = withLogs.filter((row) => loggedRunIds.has(buildTaskRunId(row.task_id, row.started_at)) ||
+        (row.log_path !== null && fs.existsSync(row.log_path)));
+    return { withLogs, backed };
+}
+/** Stage key used for `llm_usage` events recorded outside any stage scope. */
+const UNATTRIBUTED_STAGE = "unattributed";
+function emptyLlmUsageStageAggregate() {
+    return {
+        calls: 0,
+        totalDurationMs: 0,
+        promptTokens: 0,
+        completionTokens: 0,
+        totalTokens: 0,
+        reasoningTokens: 0,
+    };
+}
+function emptyLlmUsageAggregate() {
+    return { ...emptyLlmUsageStageAggregate(), byStage: {} };
+}
+/**
+ * Aggregate `llm_usage` events (#576) into a window total plus a per-stage
+ * breakdown of call count, wall-time, and token usage. Token fields absent from
+ * a best-effort record contribute 0. Calls with no `stage` land under
+ * {@link UNATTRIBUTED_STAGE}.
+ */
+function summarizeLlmUsage(events) {
+    const aggregate = emptyLlmUsageAggregate();
+    for (const event of events) {
+        const meta = event.metadata ?? {};
+        const stageKey = typeof meta.stage === "string" && meta.stage ? meta.stage : UNATTRIBUTED_STAGE;
+        let stage = aggregate.byStage[stageKey];
+        if (!stage) {
+            stage = emptyLlmUsageStageAggregate();
+            aggregate.byStage[stageKey] = stage;
+        }
+        const durationMs = toFiniteNumber(meta.durationMs);
+        const promptTokens = toFiniteNumber(meta.promptTokens);
+        const completionTokens = toFiniteNumber(meta.completionTokens);
+        const totalTokens = toFiniteNumber(meta.totalTokens);
+        const reasoningTokens = toFiniteNumber(meta.reasoningTokens);
+        for (const target of [aggregate, stage]) {
+            target.calls += 1;
+            target.totalDurationMs += durationMs;
+            target.promptTokens += promptTokens;
+            target.completionTokens += completionTokens;
+            target.totalTokens += totalTokens;
+            target.reasoningTokens += reasoningTokens;
+        }
+    }
+    return aggregate;
+}
+function readLlmUsageAggregate(stateDbPath, since, until) {
+    const events = readEvents({ since, type: LLM_USAGE_EVENT }, { dbPath: stateDbPath }).events.filter((event) => {
+        if (until === undefined)
+            return true;
+        return new Date(event.ts ?? since).getTime() < new Date(until).getTime();
+    });
+    return summarizeLlmUsage(events);
+}
+function buildWindowMetrics(db, stateDbPath, since, until, now = () => Date.now(), logsDb) {
     const taskRows = queryTaskHistory(db, { since }).filter((row) => {
         const startMs = new Date(row.started_at).getTime();
         const untilMs = new Date(until).getTime();
         return !Number.isFinite(untilMs) || startMs < untilMs;
     });
-    const taskRowsWithLogs = taskRows.filter((row) => row.log_path !== null);
-    const existingLogRows = taskRowsWithLogs.filter((row) => row.log_path && fs.existsSync(row.log_path));
+    const { withLogs: taskRowsWithLogs, backed: existingLogRows } = partitionLogBackedRows(taskRows, logsDb);
     const failedTaskRows = taskRows.filter((row) => row.status === "failed");
     const activeRows = taskRows.filter((row) => row.status === "active");
     const stuckActiveRuns = activeRows.filter((row) => now() - new Date(row.started_at).getTime() > ACTIVE_RUN_WARN_MS).length;
@@ -923,6 +995,7 @@ function buildWindowMetrics(db, stateDbPath, since, until, now = () => Date.now(
         stuckActiveRuns,
         logBackingRate: roundRate(logBackingRate),
         probeRoundTripMs: null,
+        llmUsage: readLlmUsageAggregate(stateDbPath, since, until),
     };
     return { improve: improveSummary, metrics, runs: runCount };
 }
@@ -961,6 +1034,16 @@ export function akmHealth(options = {}) {
     catch (error) {
         throw new ConfigError(`Unable to open state.db: ${error instanceof Error ? error.message : String(error)}`, "INVALID_CONFIG_FILE");
     }
+    // logs.db backs the log-backing metric (#579). Best-effort: when it cannot
+    // be opened, partitionLogBackedRows falls back to the on-disk file check, so
+    // health never hard-fails on a missing/locked logs database.
+    let logsDb;
+    try {
+        logsDb = openLogsDatabase(options.logsDbPath);
+    }
+    catch {
+        logsDb = undefined;
+    }
     try {
         const tables = listExistingTableNames(db, ["events", "task_history", "proposals", "schema_migrations"]);
         const tableNames = tables.map((row) => row.name).sort();
@@ -968,8 +1051,7 @@ export function akmHealth(options = {}) {
         const missingTables = requiredTables.filter((name) => !tableNames.includes(name));
         const probe = probeStateDbRoundTrip(stateDbPath);
         const taskRows = queryTaskHistory(db, { since });
-        const taskRowsWithLogs = taskRows.filter((row) => row.log_path !== null);
-        const existingLogRows = taskRowsWithLogs.filter((row) => row.log_path && fs.existsSync(row.log_path));
+        const { withLogs: taskRowsWithLogs, backed: existingLogRows } = partitionLogBackedRows(taskRows, logsDb);
         const failedTaskRows = taskRows.filter((row) => row.status === "failed");
         const activeRows = taskRows.filter((row) => row.status === "active");
         const stuckActiveRuns = activeRows.filter((row) => now() - new Date(row.started_at).getTime() > ACTIVE_RUN_WARN_MS).length;
@@ -1041,6 +1123,7 @@ export function akmHealth(options = {}) {
             stuckActiveRuns,
             logBackingRate: roundRate(logBackingRate),
             probeRoundTripMs: probe.durationMs,
+            llmUsage: readLlmUsageAggregate(stateDbPath, since),
         };
         const hardFailure = hardChecks.some((check) => check.status === "fail");
         const deterministicWarnings = [...hardChecks, ...advisories].some((check) => check.status === "warn" && check.kind === "deterministic");
@@ -1062,7 +1145,7 @@ export function akmHealth(options = {}) {
             windowResults = windowSpecs.map((spec) => {
                 const winSince = parseHealthSince(spec.since);
                 const winUntil = spec.until ? parseHealthSince(spec.until) : new Date(now()).toISOString();
-                const bundle = buildWindowMetrics(db, stateDbPath, winSince, winUntil, now);
+                const bundle = buildWindowMetrics(db, stateDbPath, winSince, winUntil, now, logsDb);
                 return {
                     name: spec.name,
                     since: winSince,
@@ -1112,6 +1195,14 @@ export function akmHealth(options = {}) {
     }
     finally {
         db.close();
+        if (logsDb) {
+            try {
+                logsDb.close();
+            }
+            catch {
+                // best-effort
+            }
+        }
     }
 }
 // ── Markdown renderers ───────────────────────────────────────────────────────

package/dist/commands/improve/extract.js

@@ -42,6 +42,14 @@ import { buildExtractPrompt, EXTRACT_JSON_SCHEMA, parseExtractPayload } from "./
 import { buildSessionSummaryPrompt, parseSessionSummary, SESSION_SUMMARY_JSON_SCHEMA, sessionMeetsDurationGate, writeSessionAsset, } from "./session-asset.js";
 /** Default minimum session duration (minutes) for session indexing (#561). */
 const DEFAULT_MIN_SESSION_DURATION_MINUTES = 5;
+/**
+ * Default minimum raw session size (chars) below which the extract LLM call is
+ * skipped (#595/#596). Deliberately tiny: analysis of 218 candidate-producing
+ * sessions showed sessions of 22–368 raw chars regularly yield 1–5 candidates,
+ * so size is not a reliable proxy for value — only truly empty sessions
+ * (0 chars, journal files) are safe to skip.
+ */
+const DEFAULT_MIN_CONTENT_CHARS = 10;
 // ── Helpers ──────────────────────────────────────────────────────────────────
 /**
  * Parse a since-string into an absolute ms-epoch cutoff. Accepts:
@@ -115,7 +123,7 @@ function buildCandidateProposal(candidate, sourceRef) {
  * proposal validation failure) the session result records a warning and
  * keeps going — one session's bad luck never aborts a multi-session run.
  */
-async function processSession(harness, sessionRef, stashDir, config, llmConfig, chat, ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, sessionIndexing) {
+async function processSession(harness, sessionRef, stashDir, config, llmConfig, chat, ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, minContentChars, sessionIndexing) {
     const warnings = [];
     let data;
     try {
@@ -136,6 +144,31 @@ async function processSession(harness, sessionRef, stashDir, config, llmConfig,
     const filtered = preFilterSession(data, {
         ...(typeof maxTotalChars === "number" ? { maxTotalChars } : {}),
     });
+    // #595/#596 — minContentChars gate: skip the LLM call for sessions whose RAW
+    // size is below threshold. Measured on the raw event text BEFORE the noise
+    // pre-filter, NOT on post-filter output — the pre-filter strips boilerplate
+    // so aggressively that even signal-bearing sessions can have tiny output
+    // (#596: gating post-filter filtered out 100% of sessions). Note: the 0.8.x
+    // fix gated on `filtered.stats.inputCount`, which is an EVENT count, not a
+    // char count — this port measures actual raw chars so the threshold matches
+    // the config key's documented unit.
+    const rawContentChars = data.events.reduce((sum, event) => sum + event.text.length, 0);
+    if (minContentChars > 0 && rawContentChars < minContentChars) {
+        return {
+            sessionId: sessionRef.sessionId,
+            harness: harness.name,
+            candidateCount: 0,
+            proposalIds: [],
+            preFilter: {
+                inputCount: filtered.stats.inputCount,
+                outputCount: filtered.stats.outputCount,
+                truncatedCount: filtered.stats.truncatedCount,
+            },
+            warnings: [],
+            skipped: true,
+            skipReason: "too_short",
+        };
+    }
     const prompt = buildExtractPrompt({ data, events: filtered.events, inlineRefs: data.inlineRefs });
     // #561 — ADDITIVE session indexing. Generate + write the session asset
     // (`sessions/<harness>/<id>.md`). FAIL-OPEN: any failure only records a
@@ -339,6 +372,9 @@ export async function akmExtract(options) {
         60_000;
     // Pre-filter budget — process config can raise it for large-context models.
     const maxTotalChars = typeof extractProcess?.maxTotalChars === "number" ? extractProcess.maxTotalChars : undefined;
+    // #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;
     // 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
@@ -483,7 +519,7 @@ export async function akmExtract(options) {
             continue;
         }
         try {
-            const result = await processSession(harness, summary, stashDir, config, llmConfig, chat, options.ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, sessionIndexing);
+            const result = await processSession(harness, summary, stashDir, config, llmConfig, chat, options.ctx, sourceRun, dryRun, timeoutMs, maxTotalChars, minContentChars, sessionIndexing);
             sessions.push(result);
             if (result.skipped)
                 skippedCount += 1;

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

@@ -4,7 +4,7 @@
 import { loadConfig } from "../../core/config/config.js";
 import { appendEvent } from "../../core/events.js";
 import { info, warn } from "../../core/warn.js";
-import { promoteProposal } from "../proposal/validators/proposals.js";
+import { promoteProposal, recordGateDecision } from "../proposal/validators/proposals.js";
 // ---------------------------------------------------------------------------
 // Gate implementation
 // ---------------------------------------------------------------------------
@@ -26,14 +26,40 @@ export async function runAutoAcceptGate(candidates, cfg, promoteFn = promoteProp
     }
     const effectiveThreshold = Math.max(cfg.globalThreshold, cfg.minimumThreshold ?? 0) / 100;
     const resolvedConfig = typeof cfg.config === "function" ? cfg.config() : cfg.config;
+    const gateLabel = `improve:${cfg.phase}`;
+    // #577: stamp the gate's verdict onto each proposal so `akm proposal show`
+    // can explain why a proposal is pending (e.g. "deferred: below-threshold,
+    // 0.72 < 0.90"). Best-effort — a recording failure must never abort the gate.
+    const stamp = (proposalId, decision) => {
+        try {
+            recordGateDecision(cfg.stashDir, proposalId, decision);
+        }
+        catch (err) {
+            warn(`[improve] ${cfg.phase} failed to record gate decision for ${proposalId}: ${err instanceof Error ? err.message : String(err)}`);
+        }
+    };
     for (const candidate of candidates) {
         const { proposalId, confidence } = candidate;
         if (confidence === undefined || confidence < effectiveThreshold) {
+            stamp(proposalId, {
+                outcome: "deferred",
+                reason: confidence === undefined ? "no-confidence" : "below-threshold",
+                ...(confidence !== undefined ? { confidence } : {}),
+                thresholds: { autoAccept: effectiveThreshold },
+                gate: gateLabel,
+            });
             result.skipped.push(proposalId);
             continue;
         }
         try {
             const promotion = await promoteFn(cfg.stashDir, resolvedConfig, proposalId, {}, undefined);
+            stamp(promotion.proposal.id, {
+                outcome: "auto-accepted",
+                reason: "above-threshold",
+                confidence,
+                thresholds: { autoAccept: effectiveThreshold },
+                gate: gateLabel,
+            });
             appendEvent({
                 eventType: "promoted",
                 ref: promotion.ref,