@agentdeck/bridge 0.2.2 → 0.2.3

package/dist/apme/store.d.ts

@@ -6,12 +6,21 @@
  * Callers should check `store.enabled` before assuming persistence.
  */
 import type { ApmeRunRow, ApmeStepRow, ApmeArtifactRow, ApmeEvalRowDb, ApmeRubricRow, ApmeVibeRow, ApmeScorecardRow, ApmeTaskRow } from './types.js';
+import type { ApmeSampleEventRow, ApmeSampleScorecardRow, SessionSample, TrajectoryEvent } from '@agentdeck/shared';
 export declare class ApmeStore {
     private db;
     enabled: boolean;
     readonly dbPath: string;
+    /** Populated when init() returns false. Surfaced in daemon startup logs so
+     *  silent APME outages (the failure mode that left ~/.agentdeck/apme.sqlite
+     *  stale for 11 days in 2026-04 user data) become diagnosable. */
+    lastInitError: string | null;
     constructor(dbPath?: string);
-    /** Attempt to open the DB. Returns false if better-sqlite3 is unavailable. */
+    /** Attempt to open the DB. Returns false on failure; check `lastInitError`
+     *  for the reason. The two common failure modes are:
+     *    1. better-sqlite3 native binding missing (CI / setups without build tools)
+     *    2. DB file unreadable / DDL fails (disk full, permissions, WAL lock from a
+     *       crashed prior process). */
     init(): Promise<boolean>;
     close(): void;
     /** Add columns that may be missing from databases created before this version. */
@@ -43,6 +52,12 @@ export declare class ApmeStore {
     updateTask(id: string, patch: Partial<ApmeTaskRow>): void;
     getTask(id: string): ApmeTaskRow | null;
     listTasksForRun(runId: string): ApmeTaskRow[];
+    /** All tasks across runs, newest first. Used by `agentdeck apme export --by task`
+     *  to dump a flat dataset of meaningful task units (one row per closed task). */
+    listAllTasks(opts?: {
+        limit?: number;
+        closedOnly?: boolean;
+    }): ApmeTaskRow[];
     listTurnsForTask(taskId: string): Array<Record<string, unknown>>;
     /** Ended tasks (boundary hit) that haven't been judged yet — backfill candidates. */
     listTasksNeedingSummary(limit?: number): Array<{
@@ -86,6 +101,23 @@ export declare class ApmeStore {
     /** Orphaned runs: started long ago, never closed, no turns.
      *  Typically from session bridges that crashed without cleanup. */
     listOrphanedRuns(staleSec?: number): string[];
+    /** Append one typed trajectory event. INSERT OR IGNORE on the UNIQUE
+     *  (task_id, dedup_key) index makes storage-time dedup atomic — duplicates
+     *  never persist. Returns true if a row was actually inserted. */
+    insertSampleEvent(row: ApmeSampleEventRow): boolean;
+    /** Update a previously-inserted event (e.g. a tool pending→resolved) by id. */
+    updateSampleEvent(id: number, fields: Partial<ApmeSampleEventRow>): void;
+    /** Find a tool event still pending for (task, turn, toolName), to resolve it. */
+    findPendingToolEvent(taskId: string, turnIndex: number, toolName: string): ApmeSampleEventRow | null;
+    listSampleEventRows(taskId: string): ApmeSampleEventRow[];
+    listSampleEvents(taskId: string): TrajectoryEvent[];
+    /** Next monotonic seq within a task. */
+    nextSampleSeq(taskId: string): number;
+    /** Assemble the full SessionSample (header + cost + typed trajectory). */
+    getSample(taskId: string): SessionSample | null;
+    /** Recompute the task's cost aggregate by summing its ModelEvents. */
+    recomputeSampleCost(taskId: string): void;
+    sampleScorecard(): ApmeSampleScorecardRow[];
     scorecard(): ApmeScorecardRow[];
     categoryScorecard(): Array<{
         taskCategory: string;

package/dist/apme/store.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"store.d.ts","sourceRoot":"","sources":["../../src/apme/store.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAaH,OAAO,KAAK,EACV,UAAU,EACV,WAAW,EACX,eAAe,EACf,aAAa,EACb,aAAa,EACb,WAAW,EACX,gBAAgB,EAChB,WAAW,EACZ,MAAM,YAAY,CAAC;AAoWpB,qBAAa,SAAS;IACpB,OAAO,CAAC,EAAE,CAA+B;IAClC,OAAO,UAAS;IACvB,SAAgB,MAAM,EAAE,MAAM,CAAC;gBAEnB,MAAM,CAAC,EAAE,MAAM;IAK3B,8EAA8E;IACxE,IAAI,IAAI,OAAO,CAAC,OAAO,CAAC;IA0B9B,KAAK,IAAI,IAAI;IAMb,kFAAkF;IAClF,OAAO,CAAC,aAAa;IAiCrB,OAAO,CAAC,iBAAiB;IAqCzB,SAAS,CAAC,GAAG,EAAE,UAAU,GAAG,IAAI;IA4BhC,SAAS,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,UAAU,CAAC,GAAG,IAAI;IAoCvD,oFAAoF;IACpF,SAAS,CAAC,EAAE,EAAE,MAAM,GAAG,IAAI;IAW3B,MAAM,CAAC,EAAE,EAAE,MAAM,GAAG,UAAU,GAAG,IAAI;IAMrC,QAAQ,CAAC,IAAI,GAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,UAAU,EAAE;IAgB3F,UAAU,CAAC,IAAI,EAAE;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI;IAOxJ,UAAU,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI;IAkB7D,OAAO,CAAC,EAAE,EAAE,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI;IAKnD,SAAS,CAAC,KAAK,EAAE,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAOxD,UAAU,CAAC,GAAG,EAAE,WAAW,GAAG,IAAI;IAWlC,UAAU,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,WAAW,CAAC,GAAG,IAAI;IA0BzD,OAAO,CAAC,EAAE,EAAE,MAAM,GAAG,WAAW,GAAG,IAAI;IAMvC,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,EAAE;IAQ7C,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAOhE,qFAAqF;IACrF,uBAAuB,CAAC,KAAK,GAAE,MAAW,GAAG,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,CAAC;IAY9G,iBAAiB,CAAC,GAAG,EAAE,aAAa,GAAG;QAAE,MAAM,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI;IAShE,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,aAAa,EAAE;IAkBjD,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,aAAa,EAAE;IAkBjD,iBAAiB,CAAC,GAAG,EAAE,aAAa,GAAG;QAAE,MAAM,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI;IAWhE,UAAU,CAAC,GAAG,EAAE,WAAW,GAAG,IAAI;IAOlC,SAAS,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,EAAE;IAevC,cAAc,CAAC,GAAG,EAAE,eAAe,GAAG,IAAI;IAS1C,UAAU,CAAC,GAAG,EAAE,aAAa,GAAG,IAAI;IAYpC,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,aAAa,EAAE;IAoB/C,gBAAgB,CAAC,OAAO,GAAE,MAAkB,GAAG,aAAa,GAAG,IAAI;IAQnE,YAAY,CAAC,GAAG,EAAE,IAAI,CAAC,aAAa,EAAE,SAAS,CAAC,GAAG,MAAM;IAYzD,UAAU,CAAC,GAAG,EAAE,WAAW,GAAG,IAAI;IAOlC,sEAAsE;IACtE,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,GAAG,IAAI;IAenD,2FAA2F;IAC3F,mBAAmB,CAAC,KAAK,GAAE,MAAW,GAAG;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,EAAE;IAYrF,2FAA2F;IAC3F,oBAAoB,CAAC,KAAK,GAAE,MAAU,GAAG;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,EAAE;IAYrF,6EAA6E;IAC7E,uBAAuB,CAAC,KAAK,GAAE,MAAW,GAAG,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC;IAYjF;uEACmE;IACnE,gBAAgB,CAAC,QAAQ,GAAE,MAAa,GAAG,MAAM,EAAE;IAgBnD,SAAS,IAAI,gBAAgB,EAAE;IAc/B,iBAAiB,IAAI,KAAK,CAAC;QAAE,YAAY,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,CAAC;CAYtK"}
+{"version":3,"file":"store.d.ts","sourceRoot":"","sources":["../../src/apme/store.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAaH,OAAO,KAAK,EACV,UAAU,EACV,WAAW,EACX,eAAe,EACf,aAAa,EACb,aAAa,EACb,WAAW,EACX,gBAAgB,EAChB,WAAW,EACZ,MAAM,YAAY,CAAC;AACpB,OAAO,KAAK,EACV,kBAAkB,EAClB,sBAAsB,EACtB,aAAa,EAEb,eAAe,EAChB,MAAM,mBAAmB,CAAC;AA+a3B,qBAAa,SAAS;IACpB,OAAO,CAAC,EAAE,CAA+B;IAClC,OAAO,UAAS;IACvB,SAAgB,MAAM,EAAE,MAAM,CAAC;IAC/B;;sEAEkE;IAC3D,aAAa,EAAE,MAAM,GAAG,IAAI,CAAQ;gBAE/B,MAAM,CAAC,EAAE,MAAM;IAK3B;;;;uCAImC;IAC7B,IAAI,IAAI,OAAO,CAAC,OAAO,CAAC;IA6B9B,KAAK,IAAI,IAAI;IAMb,kFAAkF;IAClF,OAAO,CAAC,aAAa;IAkDrB,OAAO,CAAC,iBAAiB;IAqCzB,SAAS,CAAC,GAAG,EAAE,UAAU,GAAG,IAAI;IA4BhC,SAAS,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,UAAU,CAAC,GAAG,IAAI;IAoCvD,oFAAoF;IACpF,SAAS,CAAC,EAAE,EAAE,MAAM,GAAG,IAAI;IAW3B,MAAM,CAAC,EAAE,EAAE,MAAM,GAAG,UAAU,GAAG,IAAI;IAMrC,QAAQ,CAAC,IAAI,GAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,OAAO,CAAC,EAAE,MAAM,CAAA;KAAO,GAAG,UAAU,EAAE;IAgB3F,UAAU,CAAC,IAAI,EAAE;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,MAAM,CAAC,EAAE,MAAM,CAAC;QAAC,SAAS,EAAE,MAAM,CAAC;QAAC,SAAS,CAAC,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI;IAOxJ,UAAU,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI;IAkB7D,OAAO,CAAC,EAAE,EAAE,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,GAAG,IAAI;IAKnD,SAAS,CAAC,KAAK,EAAE,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAOxD,UAAU,CAAC,GAAG,EAAE,WAAW,GAAG,IAAI;IAWlC,UAAU,CAAC,EAAE,EAAE,MAAM,EAAE,KAAK,EAAE,OAAO,CAAC,WAAW,CAAC,GAAG,IAAI;IAgCzD,OAAO,CAAC,EAAE,EAAE,MAAM,GAAG,WAAW,GAAG,IAAI;IAMvC,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,EAAE;IAQ7C;qFACiF;IACjF,YAAY,CAAC,IAAI,GAAE;QAAE,KAAK,CAAC,EAAE,MAAM,CAAC;QAAC,UAAU,CAAC,EAAE,OAAO,CAAA;KAAO,GAAG,WAAW,EAAE;IAUhF,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,KAAK,CAAC,MAAM,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;IAOhE,qFAAqF;IACrF,uBAAuB,CAAC,KAAK,GAAE,MAAW,GAAG,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAC;QAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,CAAC;IAY9G,iBAAiB,CAAC,GAAG,EAAE,aAAa,GAAG;QAAE,MAAM,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI;IAShE,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,aAAa,EAAE;IAkBjD,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,aAAa,EAAE;IAkBjD,iBAAiB,CAAC,GAAG,EAAE,aAAa,GAAG;QAAE,MAAM,EAAE,MAAM,CAAA;KAAE,GAAG,IAAI;IAWhE,UAAU,CAAC,GAAG,EAAE,WAAW,GAAG,IAAI;IAOlC,SAAS,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,EAAE;IAevC,cAAc,CAAC,GAAG,EAAE,eAAe,GAAG,IAAI;IAS1C,UAAU,CAAC,GAAG,EAAE,aAAa,GAAG,IAAI;IAYpC,eAAe,CAAC,KAAK,EAAE,MAAM,GAAG,aAAa,EAAE;IAoB/C,gBAAgB,CAAC,OAAO,GAAE,MAAkB,GAAG,aAAa,GAAG,IAAI;IAQnE,YAAY,CAAC,GAAG,EAAE,IAAI,CAAC,aAAa,EAAE,SAAS,CAAC,GAAG,MAAM;IAYzD,UAAU,CAAC,GAAG,EAAE,WAAW,GAAG,IAAI;IAOlC,sEAAsE;IACtE,gBAAgB,CAAC,KAAK,EAAE,MAAM,GAAG,WAAW,GAAG,IAAI;IAenD,2FAA2F;IAC3F,mBAAmB,CAAC,KAAK,GAAE,MAAW,GAAG;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,EAAE;IAarF,2FAA2F;IAC3F,oBAAoB,CAAC,KAAK,GAAE,MAAU,GAAG;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,WAAW,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,EAAE;IAYrF,6EAA6E;IAC7E,uBAAuB,CAAC,KAAK,GAAE,MAAW,GAAG,KAAK,CAAC;QAAE,EAAE,EAAE,MAAM,CAAC;QAAC,KAAK,EAAE,MAAM,CAAA;KAAE,CAAC;IAYjF;uEACmE;IACnE,gBAAgB,CAAC,QAAQ,GAAE,MAAa,GAAG,MAAM,EAAE;IAgBnD;;sEAEkE;IAClE,iBAAiB,CAAC,GAAG,EAAE,kBAAkB,GAAG,OAAO;IAiBnD,+EAA+E;IAC/E,iBAAiB,CAAC,EAAE,EAAE,MAAM,EAAE,MAAM,EAAE,OAAO,CAAC,kBAAkB,CAAC,GAAG,IAAI;IAiBxE,iFAAiF;IACjF,oBAAoB,CAAC,MAAM,EAAE,MAAM,EAAE,SAAS,EAAE,MAAM,EAAE,QAAQ,EAAE,MAAM,GAAG,kBAAkB,GAAG,IAAI;IAWpG,mBAAmB,CAAC,MAAM,EAAE,MAAM,GAAG,kBAAkB,EAAE;IAQzD,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,eAAe,EAAE;IAInD,wCAAwC;IACxC,aAAa,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM;IAMrC,0EAA0E;IAC1E,SAAS,CAAC,MAAM,EAAE,MAAM,GAAG,aAAa,GAAG,IAAI;IAmC/C,sEAAsE;IACtE,mBAAmB,CAAC,MAAM,EAAE,MAAM,GAAG,IAAI;IAYzC,eAAe,IAAI,sBAAsB,EAAE;IAiB3C,SAAS,IAAI,gBAAgB,EAAE;IAc/B,iBAAiB,IAAI,KAAK,CAAC;QAAE,YAAY,EAAE,MAAM,CAAC;QAAC,OAAO,EAAE,MAAM,CAAC;QAAC,IAAI,EAAE,MAAM,CAAC;QAAC,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,YAAY,EAAE,MAAM,GAAG,IAAI,CAAC;QAAC,SAAS,EAAE,MAAM,GAAG,IAAI,CAAA;KAAE,CAAC;CAYtK"}

package/dist/apme/store.js

@@ -88,11 +88,43 @@ CREATE TABLE IF NOT EXISTS tasks (
   outcome          TEXT,
   composite_score  REAL,
   task_category    TEXT,
-  notes_json       TEXT
+  notes_json       TEXT,
+  model_id         TEXT,
+  model_config     TEXT,
+  input_tokens     INTEGER,
+  output_tokens    INTEGER,
+  cost_usd         REAL,
+  latency_ms       INTEGER
 );
 
 CREATE INDEX IF NOT EXISTS idx_tasks_run ON tasks(run_id);
 
+-- Typed trajectory events — the SessionSample.events projection. The single
+-- source of truth that BOTH the timeline and APME eval derive from. Storage-
+-- time dedup via the UNIQUE index + INSERT OR IGNORE (no race-sensitive window).
+CREATE TABLE IF NOT EXISTS sample_events (
+  id            INTEGER PRIMARY KEY AUTOINCREMENT,
+  task_id       TEXT NOT NULL REFERENCES tasks(id) ON DELETE CASCADE,
+  run_id        TEXT NOT NULL REFERENCES runs(id) ON DELETE CASCADE,
+  turn_index    INTEGER,
+  seq           INTEGER NOT NULL,
+  ts            INTEGER NOT NULL,
+  kind          TEXT NOT NULL,
+  model         TEXT,
+  input_tokens  INTEGER,
+  output_tokens INTEGER,
+  cost_usd      REAL,
+  latency_ms    INTEGER,
+  tool_name     TEXT,
+  tool_status   TEXT,
+  tool_error    TEXT,
+  payload       TEXT,
+  dedup_key     TEXT
+);
+
+CREATE INDEX IF NOT EXISTS idx_sevents_task ON sample_events(task_id, seq);
+CREATE UNIQUE INDEX IF NOT EXISTS idx_sevents_dedup ON sample_events(task_id, dedup_key);
+
 CREATE TABLE IF NOT EXISTS artifacts (
   id        INTEGER PRIMARY KEY AUTOINCREMENT,
   run_id    TEXT NOT NULL REFERENCES runs(id) ON DELETE CASCADE,
@@ -179,6 +211,28 @@ FROM runs r
 LEFT JOIN v_run_metrics m ON m.run_id = r.id
 WHERE r.task_category IS NOT NULL AND r.task_category != 'unknown'
 GROUP BY r.task_category, r.model_id;
+
+-- Sample-granularity scorecard: quality vs cost per (agent, model, category).
+-- The recommender + Pareto frontier read this. Uses the task's own model_id /
+-- cost (the sample header), falling back to the run's model when unset.
+CREATE VIEW IF NOT EXISTS v_sample_scorecard AS
+SELECT
+  r.agent_type AS agent_type,
+  COALESCE(t.model_id, r.model_id, 'unknown') AS model_id,
+  t.task_category AS task_category,
+  COUNT(*) AS samples,
+  AVG(t.composite_score) AS avg_quality,
+  SUM(t.cost_usd) AS total_cost,
+  AVG(t.latency_ms) AS avg_latency_ms,
+  CASE
+    WHEN AVG(t.composite_score) > 0
+    THEN SUM(t.cost_usd) / AVG(t.composite_score)
+    ELSE NULL
+  END AS cost_per_quality
+FROM tasks t
+JOIN runs r ON r.id = t.run_id
+WHERE t.ended_at IS NOT NULL AND t.composite_score IS NOT NULL
+GROUP BY r.agent_type, COALESCE(t.model_id, r.model_id, 'unknown'), t.task_category;
 `;
 // ─── Default rubric v1 (seeded on first boot) ──────────────────────────────────
 const DEFAULT_RUBRIC_V1 = {
@@ -323,29 +377,50 @@ Return strict JSON: {"correctness":N,"safety":N,"completeness":N,"overall":N,"re
     },
     task_rollup: {
         purpose: 'task_rollup',
-        prompt: `You are evaluating a multi-turn AI agent task that the agent itself has just signaled as complete
-(for example: every TodoWrite item was marked \`completed\`, or the user issued \`/clear\` to reset context).
+        prompt: `You are evaluating a multi-turn AI agent task that has just ended.
+The boundary signal that closed the task tells you HOW it ended:
+  - todo_complete : the agent itself marked every TodoWrite item as completed (self-declared done)
+  - clear         : the user typed /clear to reset context (often: user gave up or moved on)
+  - session_end   : the agent process exited (could be done, could be interrupted)
+  - manual        : a human marked the boundary explicitly
 
-You will be given the full sequence of turns (user prompt → agent response) that make up this task.
-Produce a concise rollup: a single-sentence summary plus axis scores.
+You receive: the task's category (coding/planning/research/…), the number of turns,
+the boundary signal, and the full Turn 0..N transcript (user prompt → agent response).
 
-Score each axis as a float in [0,1] where 0=failed and 1=excellent.
+Your job is a one-sentence rollup PLUS axis scores in [0,1].
 
-Axes:
-- completion: Did the agent actually finish what the task was about? A high score means the task was
-  delivered end-to-end; low if large pieces were skipped or declared done without evidence.
-- coherence: Did the turns build on each other toward a single goal? Penalize incoherent jumps,
-  redundant re-planning, or lost context between turns.
-- efficiency: Were the turns appropriately scoped? Penalize unnecessary back-and-forth, repeated
-  tool calls with the same inputs, or churn. Reward focused, decisive progress.
-- overall: Holistic judgment. Weight completion most heavily — an efficient, coherent task that
-  never actually finishes is worse than a slightly messier task that delivered.
-
-Summary guidance: one sentence, ≤ 140 characters, past tense, describing what the task accomplished.
-Start with a verb (e.g. "Added", "Fixed", "Investigated", "Refactored"). No hedging.
-
-Return strict JSON exactly:
-{"summary":"<one sentence>","completion":N,"coherence":N,"efficiency":N,"overall":N,"reasoning":"...","done":["item1"],"missed":["item1"]}.`,
+Identify FIRST: what was the user actually trying to accomplish? Read Turn 0's prompt and any
+later prompts that pivot or refine. The task's success is measured against THAT goal — not
+against how busy the turns look.
+
+Axes (each in [0,1], 0=failed, 1=excellent):
+- completion: Did the agent actually deliver against the user's identified goal? High = goal
+  reached with evidence in the final turns. Low = goal half-done, abandoned, or only declared
+  done (e.g. "I've completed all the items" with nothing visible). For boundary=clear, completion
+  is usually low — the user reset before satisfaction.
+- coherence: Did the turns build on each other toward the goal? Penalize incoherent jumps,
+  redundant re-planning, lost context, or the agent forgetting earlier decisions.
+- efficiency: Were the turns appropriately scoped? Penalize repeated tool calls with the same
+  inputs, long discovery loops the agent could have shortcut, or churn. Reward focused progress.
+- overall: Holistic judgment. Weight completion most heavily — an efficient coherent task
+  that never finishes is worse than a slightly messier task that delivered.
+
+Summary guidance: one sentence, ≤ 280 characters, past tense, describing what the task ACCOMPLISHED
+(not what the agent attempted). Start with a verb: "Added", "Fixed", "Investigated", "Refactored",
+"Failed to". Be specific about the artefact when possible. No hedging, no "the agent…" preamble.
+
+reasoning: 1-3 sentences explaining the key evidence behind the overall score. Cite turn numbers.
+done: list the concrete deliverables visible in the turns (≤5 short items).
+missed: list what the user asked for but the agent did NOT deliver (≤5 items, empty array if none).
+
+Return strict JSON exactly, no prose before or after:
+{"summary":"<one sentence>","completion":N,"coherence":N,"efficiency":N,"overall":N,"reasoning":"...","done":["…"],"missed":["…"]}
+
+Examples of well-calibrated overall scores:
+  0.9 — User asked to add a feature; final turns show the feature implemented + test passing.
+  0.6 — User asked for a feature; agent built most of it but left a TODO they self-declared "done".
+  0.3 — User asked a question; agent rambled across 5 turns without ever answering.
+  0.1 — User asked to fix a bug; agent introduced two more bugs and called /clear.`,
         weights: JSON.stringify({ completion: 0.5, coherence: 0.25, efficiency: 0.25 }),
         notes: 'task-unit rollup (TodoWrite all-completed / /clear / session_end)',
     },
@@ -354,19 +429,28 @@ export class ApmeStore {
     db = null;
     enabled = false;
     dbPath;
+    /** Populated when init() returns false. Surfaced in daemon startup logs so
+     *  silent APME outages (the failure mode that left ~/.agentdeck/apme.sqlite
+     *  stale for 11 days in 2026-04 user data) become diagnosable. */
+    lastInitError = null;
     constructor(dbPath) {
         const dataDir = process.env.AGENTDECK_DATA_DIR || join(homedir(), '.agentdeck');
         this.dbPath = dbPath ?? join(dataDir, 'apme.sqlite');
     }
-    /** Attempt to open the DB. Returns false if better-sqlite3 is unavailable. */
+    /** Attempt to open the DB. Returns false on failure; check `lastInitError`
+     *  for the reason. The two common failure modes are:
+     *    1. better-sqlite3 native binding missing (CI / setups without build tools)
+     *    2. DB file unreadable / DDL fails (disk full, permissions, WAL lock from a
+     *       crashed prior process). */
     async init() {
         try {
             let Ctor = null;
             try {
                 Ctor = require('better-sqlite3');
             }
-            catch {
-                debug('APME', 'better-sqlite3 not installed — APME store disabled');
+            catch (err) {
+                this.lastInitError = `better-sqlite3 native binding unavailable (${String(err).slice(0, 200)}). Run \`pnpm install\` in bridge/.`;
+                debug('APME', this.lastInitError);
                 return false;
             }
             const dir = dirname(this.dbPath);
@@ -379,11 +463,13 @@ export class ApmeStore {
             this.migrateSchema();
             this.seedDefaultRubric();
             this.enabled = true;
+            this.lastInitError = null;
             debug('APME', `store ready at ${this.dbPath}`);
             return true;
         }
         catch (err) {
-            debug('APME', `store init failed: ${String(err)}`);
+            this.lastInitError = `store init failed at ${this.dbPath}: ${String(err).slice(0, 300)}`;
+            debug('APME', this.lastInitError);
             return false;
         }
     }
@@ -439,6 +525,26 @@ export class ApmeStore {
             }
             catch { /* ignore */ }
         }
+        // Tasks sample-header columns (model identity + cost) — added for the
+        // SessionSample rebuild. Older DBs get them via ALTER.
+        const taskCols = this.db.prepare("PRAGMA table_info(tasks)").all().map(c => c.name);
+        for (const [col, sql] of [
+            ['model_id', 'ALTER TABLE tasks ADD COLUMN model_id TEXT'],
+            ['model_config', 'ALTER TABLE tasks ADD COLUMN model_config TEXT'],
+            ['input_tokens', 'ALTER TABLE tasks ADD COLUMN input_tokens INTEGER'],
+            ['output_tokens', 'ALTER TABLE tasks ADD COLUMN output_tokens INTEGER'],
+            ['cost_usd', 'ALTER TABLE tasks ADD COLUMN cost_usd REAL'],
+            ['latency_ms', 'ALTER TABLE tasks ADD COLUMN latency_ms INTEGER'],
+        ]) {
+            if (!taskCols.includes(col)) {
+                try {
+                    this.db.exec(sql);
+                }
+                catch { /* ignore */ }
+            }
+        }
+        // sample_events table + indexes are created via CREATE TABLE IF NOT EXISTS
+        // in DDL; nothing to ALTER. The v_sample_scorecard view likewise.
     }
     seedDefaultRubric() {
         if (!this.db)
@@ -605,6 +711,12 @@ export class ApmeStore {
             taskCategory: 'task_category',
             notesJson: 'notes_json',
             boundarySignal: 'boundary_signal',
+            modelId: 'model_id',
+            modelConfig: 'model_config',
+            inputTokens: 'input_tokens',
+            outputTokens: 'output_tokens',
+            costUsd: 'cost_usd',
+            latencyMs: 'latency_ms',
         };
         const sets = [];
         const vals = [];
@@ -632,6 +744,18 @@ export class ApmeStore {
         const rows = this.db.prepare('SELECT * FROM tasks WHERE run_id = ? ORDER BY task_index ASC').all(runId);
         return rows.map(rowToTask);
     }
+    /** All tasks across runs, newest first. Used by `agentdeck apme export --by task`
+     *  to dump a flat dataset of meaningful task units (one row per closed task). */
+    listAllTasks(opts = {}) {
+        if (!this.db)
+            return [];
+        const limit = opts.limit ?? 100;
+        const sql = opts.closedOnly
+            ? `SELECT * FROM tasks WHERE ended_at IS NOT NULL ORDER BY started_at DESC LIMIT ?`
+            : `SELECT * FROM tasks ORDER BY started_at DESC LIMIT ?`;
+        const rows = this.db.prepare(sql).all(limit);
+        return rows.map(rowToTask);
+    }
     listTurnsForTask(taskId) {
         if (!this.db)
             return [];
@@ -782,6 +906,7 @@ export class ApmeStore {
             return [];
         const rows = this.db.prepare(`SELECT r.id, r.project_path FROM runs r
        WHERE r.ended_at IS NOT NULL
+         AND (r.task_category IS NULL OR r.task_category != '_empty')
          AND NOT EXISTS (SELECT 1 FROM evals e WHERE e.run_id = r.id)
        ORDER BY r.ended_at DESC
        LIMIT ?`).all(limit);
@@ -823,6 +948,136 @@ export class ApmeStore {
        LIMIT 20`).all(cutoff);
         return rows.map((r) => r.id);
     }
+    // ─── Sample events (typed trajectory) ────────────────────────────────────────
+    /** Append one typed trajectory event. INSERT OR IGNORE on the UNIQUE
+     *  (task_id, dedup_key) index makes storage-time dedup atomic — duplicates
+     *  never persist. Returns true if a row was actually inserted. */
+    insertSampleEvent(row) {
+        if (!this.db)
+            return false;
+        const res = this.db.prepare(`INSERT OR IGNORE INTO sample_events
+        (task_id, run_id, turn_index, seq, ts, kind, model, input_tokens, output_tokens,
+         cost_usd, latency_ms, tool_name, tool_status, tool_error, payload, dedup_key)
+       VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)`).run(row.taskId, row.runId, row.turnIndex ?? null, row.seq, row.ts, row.kind, row.model ?? null, row.inputTokens ?? null, row.outputTokens ?? null, row.costUsd ?? null, row.latencyMs ?? null, row.toolName ?? null, row.toolStatus ?? null, row.toolError ?? null, row.payload ?? null, row.dedupKey ?? null);
+        return res.changes > 0;
+    }
+    /** Update a previously-inserted event (e.g. a tool pending→resolved) by id. */
+    updateSampleEvent(id, fields) {
+        if (!this.db)
+            return;
+        const map = {
+            toolStatus: 'tool_status', toolError: 'tool_error', payload: 'payload',
+            costUsd: 'cost_usd', latencyMs: 'latency_ms', model: 'model',
+            inputTokens: 'input_tokens', outputTokens: 'output_tokens', ts: 'ts',
+        };
+        const sets = [];
+        const vals = [];
+        for (const [k, v] of Object.entries(fields)) {
+            const col = map[k];
+            if (!col || v === undefined)
+                continue;
+            sets.push(`${col} = ?`);
+            vals.push(v);
+        }
+        if (sets.length === 0)
+            return;
+        vals.push(id);
+        this.db.prepare(`UPDATE sample_events SET ${sets.join(', ')} WHERE id = ?`).run(...vals);
+    }
+    /** Find a tool event still pending for (task, turn, toolName), to resolve it. */
+    findPendingToolEvent(taskId, turnIndex, toolName) {
+        if (!this.db)
+            return null;
+        const row = this.db.prepare(`SELECT * FROM sample_events
+       WHERE task_id = ? AND turn_index = ? AND kind = 'tool' AND tool_name = ?
+         AND (tool_status IS NULL OR tool_status = 'pending')
+       ORDER BY seq DESC LIMIT 1`).get(taskId, turnIndex, toolName);
+        return row ? rowToSampleEvent(row) : null;
+    }
+    listSampleEventRows(taskId) {
+        if (!this.db)
+            return [];
+        const rows = this.db.prepare('SELECT * FROM sample_events WHERE task_id = ? ORDER BY seq ASC').all(taskId);
+        return rows.map(rowToSampleEvent);
+    }
+    listSampleEvents(taskId) {
+        return this.listSampleEventRows(taskId).map(sampleEventRowToTrajectory).filter((e) => e !== null);
+    }
+    /** Next monotonic seq within a task. */
+    nextSampleSeq(taskId) {
+        if (!this.db)
+            return 0;
+        const row = this.db.prepare('SELECT COALESCE(MAX(seq),-1)+1 AS s FROM sample_events WHERE task_id = ?').get(taskId);
+        return row.s;
+    }
+    /** Assemble the full SessionSample (header + cost + typed trajectory). */
+    getSample(taskId) {
+        if (!this.db)
+            return null;
+        const task = this.getTask(taskId);
+        if (!task)
+            return null;
+        const run = this.getRun(task.runId);
+        const events = this.listSampleEvents(taskId);
+        let modelConfig = null;
+        if (task.modelConfig) {
+            try {
+                modelConfig = JSON.parse(task.modelConfig);
+            }
+            catch { /* ignore */ }
+        }
+        const modelId = task.modelId ?? run?.modelId ?? modelConfig?.modelId ?? 'unknown';
+        return {
+            id: task.id,
+            runId: task.runId,
+            sessionId: run?.sessionId ?? '',
+            agentType: (run?.agentType ?? 'claude-code'),
+            index: task.taskIndex,
+            boundarySignal: task.boundarySignal,
+            startedAt: task.startedAt,
+            endedAt: task.endedAt ?? null,
+            model: modelConfig ?? { modelId },
+            projectName: run?.projectName ?? null,
+            projectPath: run?.projectPath ?? null,
+            events,
+            cost: {
+                inputTokens: task.inputTokens ?? 0,
+                outputTokens: task.outputTokens ?? 0,
+                costUsd: task.costUsd ?? 0,
+                latencyMs: task.latencyMs ?? 0,
+            },
+            summary: task.summary ?? null,
+            outcome: task.outcome ?? null,
+            compositeScore: task.compositeScore ?? null,
+            taskCategory: task.taskCategory ?? null,
+        };
+    }
+    /** Recompute the task's cost aggregate by summing its ModelEvents. */
+    recomputeSampleCost(taskId) {
+        if (!this.db)
+            return;
+        const row = this.db.prepare(`SELECT COALESCE(SUM(input_tokens),0) AS it, COALESCE(SUM(output_tokens),0) AS ot,
+              COALESCE(SUM(cost_usd),0) AS cu, COALESCE(SUM(latency_ms),0) AS lm
+       FROM sample_events WHERE task_id = ? AND kind = 'model'`).get(taskId);
+        this.updateTask(taskId, {
+            inputTokens: row.it, outputTokens: row.ot, costUsd: row.cu, latencyMs: row.lm,
+        });
+    }
+    sampleScorecard() {
+        if (!this.db)
+            return [];
+        const rows = this.db.prepare('SELECT * FROM v_sample_scorecard').all();
+        return rows.map((r) => ({
+            agentType: r.agent_type,
+            modelId: r.model_id,
+            taskCategory: r.task_category ?? null,
+            samples: r.samples,
+            avgQuality: r.avg_quality ?? null,
+            totalCost: r.total_cost ?? null,
+            avgLatencyMs: r.avg_latency_ms ?? null,
+            costPerQuality: r.cost_per_quality ?? null,
+        }));
+    }
     // ─── Scorecard ───────────────────────────────────────────────────────────────
     scorecard() {
         if (!this.db)
@@ -895,8 +1150,62 @@ function rowToTask(r) {
         compositeScore: r.composite_score ?? null,
         taskCategory: r.task_category ?? null,
         notesJson: r.notes_json ?? null,
+        modelId: r.model_id ?? null,
+        modelConfig: r.model_config ?? null,
+        inputTokens: r.input_tokens ?? null,
+        outputTokens: r.output_tokens ?? null,
+        costUsd: r.cost_usd ?? null,
+        latencyMs: r.latency_ms ?? null,
     };
 }
+function rowToSampleEvent(r) {
+    return {
+        id: r.id,
+        taskId: r.task_id,
+        runId: r.run_id,
+        turnIndex: r.turn_index ?? null,
+        seq: r.seq,
+        ts: r.ts,
+        kind: r.kind,
+        model: r.model ?? null,
+        inputTokens: r.input_tokens ?? null,
+        outputTokens: r.output_tokens ?? null,
+        costUsd: r.cost_usd ?? null,
+        latencyMs: r.latency_ms ?? null,
+        toolName: r.tool_name ?? null,
+        toolStatus: r.tool_status ?? null,
+        toolError: r.tool_error ?? null,
+        payload: r.payload ?? null,
+        dedupKey: r.dedup_key ?? null,
+    };
+}
+/** Decode a stored sample_events row back into a typed TrajectoryEvent. */
+function sampleEventRowToTrajectory(r) {
+    const base = { ts: r.ts, turnIndex: r.turnIndex ?? 0 };
+    let p = {};
+    if (r.payload) {
+        try {
+            p = JSON.parse(r.payload);
+        }
+        catch { /* ignore */ }
+    }
+    switch (r.kind) {
+        case 'user_message':
+            return { ...base, kind: 'user_message', text: p.text ?? '' };
+        case 'assistant_message':
+            return { ...base, kind: 'assistant_message', text: p.text ?? '', responseKind: (p.responseKind ?? 'text') };
+        case 'model':
+            return { ...base, kind: 'model', model: r.model ?? 'unknown', inputTokens: r.inputTokens ?? 0, outputTokens: r.outputTokens ?? 0, costUsd: r.costUsd ?? 0, latencyMs: r.latencyMs ?? 0 };
+        case 'tool':
+            return { ...base, kind: 'tool', name: r.toolName ?? 'tool', input: p.input, output: p.output, error: r.toolError ?? null, status: r.toolStatus ?? undefined };
+        case 'state':
+            return { ...base, kind: 'state', from: p.from ?? null, to: p.to ?? 'unknown' };
+        case 'info':
+            return { ...base, kind: 'info', label: p.label ?? 'info', detail: p.detail ?? null };
+        default:
+            return null;
+    }
+}
 function rowToRubric(r) {
     return {
         version: r.version,