@agentdeck/bridge 1.0.24 → 1.2.0

package/dist/apme/store.js

@@ -15,13 +15,143 @@ import { debug } from '../logger.js';
 // workspace symlinks, regardless of the process CWD (vitest runs from the
 // repo root, where the symlink doesn't exist).
 const require = createRequire(import.meta.url);
+import { TASK_ATTENTION_WINDOW_MS, TASK_ATTENTION_RED_SCORE } from '@agentdeck/shared';
 // ─── Schema ────────────────────────────────────────────────────────────────────
+const SCORECARD_DDL = `
+-- Pre-aggregate metrics at the level they actually describe. New rows prefer
+-- turn attribution; runs with no turn identity remain available as a legacy
+-- fallback instead of being silently dropped after the schema migration.
+CREATE VIEW IF NOT EXISTS v_run_metrics AS
+SELECT
+  run_id,
+  MAX(CASE WHEN metric='overall' AND layer='llm_judge' THEN score END) AS overall,
+  MAX(CASE WHEN metric='tests_pass' AND layer='deterministic' THEN score END) AS tests_pass
+FROM evals
+GROUP BY run_id;
+
+CREATE VIEW IF NOT EXISTS v_model_scorecard AS
+WITH turn_eval AS (
+  SELECT
+    turn_id,
+    MAX(CASE WHEN metric='overall' THEN score END) AS overall,
+    MAX(CASE WHEN metric='tests_pass' THEN score END) AS tests_pass
+  FROM evals
+  WHERE turn_id IS NOT NULL
+  GROUP BY turn_id
+), turn_cost AS (
+  SELECT
+    turn_id,
+    CASE WHEN MIN(COALESCE(cost_known, 0))=1 THEN SUM(cost_usd) ELSE NULL END AS cost_usd,
+    MIN(COALESCE(cost_known, 0)) AS cost_known
+  FROM sample_events
+  WHERE turn_id IS NOT NULL AND kind='model'
+  GROUP BY turn_id
+), attributed_units AS (
+  SELECT
+    r.agent_type AS agent_type,
+    t.run_id AS run_id,
+    t.id AS turn_id,
+    t.model_id AS model_id,
+    COALESCE(t.provider, r.provider) AS provider,
+    COALESCE(
+      t.composite_score,
+      e.overall
+    ) AS overall,
+    e.tests_pass AS tests_pass,
+    c.cost_usd AS cost_usd,
+    COALESCE(c.cost_known, 0) AS cost_known
+  FROM turns t
+  JOIN runs r ON r.id=t.run_id
+  LEFT JOIN turn_eval e ON e.turn_id=t.id
+  LEFT JOIN turn_cost c ON c.turn_id=t.id
+  WHERE t.model_id IS NOT NULL AND t.model_id != ''
+), legacy_units AS (
+  SELECT
+    r.agent_type AS agent_type,
+    r.id AS run_id,
+    NULL AS turn_id,
+    COALESCE(r.model_id, 'unknown') AS model_id,
+    r.provider AS provider,
+    m.overall AS overall,
+    m.tests_pass AS tests_pass,
+    CASE WHEN r.cost_known=1 THEN r.cost_usd ELSE NULL END AS cost_usd,
+    COALESCE(r.cost_known, 0) AS cost_known
+  FROM runs r
+  LEFT JOIN v_run_metrics m ON m.run_id=r.id
+  WHERE NOT EXISTS (
+    SELECT 1 FROM turns t
+    WHERE t.run_id=r.id AND t.model_id IS NOT NULL AND t.model_id != ''
+  )
+), units AS (
+  SELECT * FROM attributed_units
+  UNION ALL
+  SELECT * FROM legacy_units
+)
+SELECT
+  agent_type,
+  model_id,
+  provider,
+  COUNT(DISTINCT run_id) AS runs,
+  AVG(overall) AS avg_overall,
+  AVG(tests_pass) AS avg_tests_pass,
+  CASE WHEN MIN(cost_known)=1 THEN SUM(cost_usd) ELSE NULL END AS total_cost,
+  MIN(cost_known) AS cost_known,
+  CASE WHEN MIN(cost_known)=1 AND AVG(overall)>0 THEN SUM(cost_usd)/AVG(overall) ELSE NULL END AS cost_per_quality
+FROM units
+GROUP BY agent_type, model_id, provider;
+
+CREATE VIEW IF NOT EXISTS v_category_scorecard AS
+WITH task_metrics AS (
+  SELECT
+    task_id,
+    MAX(CASE WHEN metric='tests_pass' AND layer='deterministic' THEN score END) AS tests_pass
+  FROM evals
+  WHERE task_id IS NOT NULL
+  GROUP BY task_id
+)
+SELECT
+  t.task_category AS task_category,
+  COALESCE(t.model_id, r.model_id, 'unknown') AS model_id,
+  COALESCE(t.provider, r.provider) AS provider,
+  COUNT(DISTINCT t.run_id) AS runs,
+  AVG(t.composite_score) AS avg_overall,
+  AVG(m.tests_pass) AS avg_tests_pass,
+  CASE WHEN MIN(COALESCE(t.cost_known, 0))=1 THEN SUM(t.cost_usd) ELSE NULL END AS total_cost,
+  MIN(COALESCE(t.cost_known, 0)) AS cost_known
+FROM tasks t
+JOIN runs r ON r.id=t.run_id
+LEFT JOIN task_metrics m ON m.task_id=t.id
+WHERE t.task_category IS NOT NULL AND t.task_category != 'unknown'
+GROUP BY t.task_category, COALESCE(t.model_id, r.model_id, 'unknown'), COALESCE(t.provider, r.provider);
+
+-- Sample-granularity scorecard: quality vs cost per (agent, provider, model,
+-- category). A task that used more than one model is stored as 'mixed' rather
+-- than attributed to whichever assistant message happened to arrive last.
+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,
+  COALESCE(t.provider, r.provider) AS provider,
+  t.task_category AS task_category,
+  COUNT(*) AS samples,
+  AVG(t.composite_score) AS avg_quality,
+  CASE WHEN MIN(COALESCE(t.cost_known, 0))=1 THEN SUM(t.cost_usd) ELSE NULL END AS total_cost,
+  MIN(COALESCE(t.cost_known, 0)) AS cost_known,
+  AVG(t.latency_ms) AS avg_latency_ms,
+  CASE WHEN MIN(COALESCE(t.cost_known, 0))=1 AND 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'), COALESCE(t.provider, r.provider), t.task_category;
+`;
 const DDL = `
 CREATE TABLE IF NOT EXISTS runs (
   id            TEXT PRIMARY KEY,
   session_id    TEXT NOT NULL,
   agent_type    TEXT NOT NULL,
   model_id      TEXT,
+  provider      TEXT,
   project_name  TEXT,
   project_path  TEXT,
   task_prompt   TEXT,
@@ -30,6 +160,7 @@ CREATE TABLE IF NOT EXISTS runs (
   input_tokens  INTEGER,
   output_tokens INTEGER,
   cost_usd      REAL,
+  cost_known    INTEGER NOT NULL DEFAULT 0,
   exit_code     INTEGER,
   git_before    TEXT,
   git_after     TEXT,
@@ -57,6 +188,8 @@ CREATE TABLE IF NOT EXISTS turns (
   run_id      TEXT NOT NULL REFERENCES runs(id) ON DELETE CASCADE,
   task_id     TEXT,
   turn_index  INTEGER NOT NULL,
+  model_id    TEXT,
+  provider    TEXT,
   prompt      TEXT,
   response    TEXT,
   started_at  INTEGER NOT NULL,
@@ -99,10 +232,12 @@ CREATE TABLE IF NOT EXISTS tasks (
   task_category    TEXT,
   notes_json       TEXT,
   model_id         TEXT,
+  provider         TEXT,
   model_config     TEXT,
   input_tokens     INTEGER,
   output_tokens    INTEGER,
   cost_usd         REAL,
+  cost_known       INTEGER NOT NULL DEFAULT 0,
   latency_ms       INTEGER
 );
 
@@ -123,6 +258,7 @@ CREATE TABLE IF NOT EXISTS sample_events (
   input_tokens  INTEGER,
   output_tokens INTEGER,
   cost_usd      REAL,
+  cost_known    INTEGER NOT NULL DEFAULT 0,
   latency_ms    INTEGER,
   tool_name     TEXT,
   tool_status   TEXT,
@@ -181,67 +317,7 @@ CREATE INDEX IF NOT EXISTS idx_runs_started ON runs(started_at);
 CREATE INDEX IF NOT EXISTS idx_evals_run ON evals(run_id);
 CREATE INDEX IF NOT EXISTS idx_steps_run ON steps(run_id);
 
--- Pre-aggregate per-run eval metrics to avoid inflating cost_usd when
--- multiple eval rows exist per run (e.g. 3 deterministic + 5 judge axes).
-CREATE VIEW IF NOT EXISTS v_run_metrics AS
-SELECT
-  run_id,
-  MAX(CASE WHEN metric='overall' AND layer='llm_judge' THEN score END) AS overall,
-  MAX(CASE WHEN metric='tests_pass' AND layer='deterministic' THEN score END) AS tests_pass
-FROM evals
-GROUP BY run_id;
-
-CREATE VIEW IF NOT EXISTS v_model_scorecard AS
-SELECT
-  r.agent_type AS agent_type,
-  COALESCE(r.model_id, 'unknown') AS model_id,
-  COUNT(*) AS runs,
-  AVG(m.overall) AS avg_overall,
-  AVG(m.tests_pass) AS avg_tests_pass,
-  SUM(r.cost_usd) AS total_cost,
-  CASE
-    WHEN AVG(m.overall) > 0
-    THEN SUM(r.cost_usd) / AVG(m.overall)
-    ELSE NULL
-  END AS cost_per_quality
-FROM runs r
-LEFT JOIN v_run_metrics m ON m.run_id = r.id
-GROUP BY r.agent_type, r.model_id;
-
-CREATE VIEW IF NOT EXISTS v_category_scorecard AS
-SELECT
-  r.task_category AS task_category,
-  COALESCE(r.model_id, 'unknown') AS model_id,
-  COUNT(*) AS runs,
-  AVG(m.overall) AS avg_overall,
-  AVG(m.tests_pass) AS avg_tests_pass,
-  SUM(r.cost_usd) AS total_cost
-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;
+${SCORECARD_DDL}
 `;
 // ─── Default rubric v1 (seeded on first boot) ──────────────────────────────────
 const DEFAULT_RUBRIC_V1 = {
@@ -455,6 +531,81 @@ Examples of well-calibrated overall scores:
         notes: 'task-unit rollup (TodoWrite all-completed / /clear / session_end)',
     },
 };
+/** The latest overall judge score for a task, else its composite. Inlined into
+ *  several `TASK_VIEW_SQL` buckets so "judged" means one thing everywhere. */
+const TASK_SCORE_SQL = `COALESCE((SELECT e.score FROM evals e WHERE e.task_id = t.id AND e.metric = 'overall'
+             ORDER BY e.created_at DESC LIMIT 1), t.composite_score)`;
+/** The recency window and red-score band live in shared/src/eval-schema.ts
+ *  (TASK_ATTENTION_WINDOW_MS / TASK_ATTENTION_RED_SCORE): both daemons build
+ *  the same attention SQL, so the numbers are cross-daemon contracts, not
+ *  store-local tuning knobs. Re-exported for existing importers. */
+export { TASK_ATTENTION_WINDOW_MS } from '@agentdeck/shared';
+/** The attention bucket, anchored at a caller-supplied cutoff (epoch ms) so
+ *  the filter, the row flag and the badge are computed against the SAME
+ *  instant within one request. Needs eyes = recent AND (reaper-closed, a
+ *  closed task with an unarchived reply the judge can only partly see, or a
+ *  judged score below the dashboard's red band). `cutoffMs` is always a
+ *  number we computed — never user input — so inlining it is safe. */
+function taskAttentionSql(cutoffMs) {
+    // IFNULL is load-bearing: a NULL score makes `score < 0.4` NULL, NULL OR
+    // false is NULL, and SQLite sorts NULL LAST under DESC — so without it,
+    // recent unjudged rows sank BELOW week-old rows (which get a determinate 0
+    // from the cutoff test) on the attention-first ordering.
+    return `IFNULL((t.started_at >= ${Math.floor(cutoffMs)}
+      AND (t.boundary_signal = 'orphaned'
+        OR (t.ended_at IS NOT NULL AND EXISTS (
+              SELECT 1 FROM turns tu WHERE tu.task_id = t.id AND tu.response IS NULL))
+        OR ${TASK_SCORE_SQL} < ${TASK_ATTENTION_RED_SCORE})), 0)`;
+}
+/** The NARROWING filters `listTaskPage` and `taskViewCounts` share — one
+ *  builder so a filtered board's badges and its rows read the same WHERE. */
+function buildTaskFilterWhere(opts) {
+    const where = [];
+    const args = [];
+    // `_empty` runs are bookkeeping shells, never work the user did.
+    where.push("COALESCE(r.task_category, '') != '_empty'");
+    if (opts.agentType) {
+        where.push('r.agent_type = ?');
+        args.push(opts.agentType);
+    }
+    if (opts.sessionId) {
+        where.push('r.session_id = ?');
+        args.push(opts.sessionId);
+    }
+    if (opts.projectName) {
+        where.push('r.project_name = ?');
+        args.push(opts.projectName);
+    }
+    if (opts.category) {
+        where.push('COALESCE(t.task_category, r.task_category) = ?');
+        args.push(opts.category);
+    }
+    if (opts.outcome) {
+        where.push('t.outcome = ?');
+        args.push(opts.outcome);
+    }
+    if (opts.q) {
+        where.push('(t.summary LIKE ? OR r.task_prompt LIKE ?)');
+        const like = `%${opts.q}%`;
+        args.push(like, like);
+    }
+    return { where, args };
+}
+/** ONE SQL definition per Work-board lifecycle bucket — the filter, the row's
+ *  `attention` flag and the tab badges all read these same expressions, so a
+ *  badge count can never disagree with the rows its tab lists.
+ *
+ *  attention  — see `taskAttentionSql` (recency-windowed, hence a function).
+ *  inprogress — still accumulating turns.
+ *  judged     — carries an overall judge score or a composite.
+ *  reported   — closed but never scored (no judge configured / skipped).
+ *  orphaned   — reaper-closed; the un-windowed ledger of segmentation debt. */
+const TASK_VIEW_SQL = {
+    inprogress: 't.ended_at IS NULL',
+    judged: `${TASK_SCORE_SQL} IS NOT NULL`,
+    reported: `(t.ended_at IS NOT NULL AND ${TASK_SCORE_SQL} IS NULL)`,
+    orphaned: `t.boundary_signal = 'orphaned'`,
+};
 export class ApmeStore {
     db = null;
     enabled = false;
@@ -517,6 +668,8 @@ export class ApmeStore {
             return;
         const cols = this.db.prepare("PRAGMA table_info(runs)").all().map(c => c.name);
         const migrations = [
+            ['provider', 'ALTER TABLE runs ADD COLUMN provider TEXT'],
+            ['cost_known', 'ALTER TABLE runs ADD COLUMN cost_known INTEGER NOT NULL DEFAULT 0'],
             ['task_signals', 'ALTER TABLE runs ADD COLUMN task_signals TEXT'],
             ['task_category', 'ALTER TABLE runs ADD COLUMN task_category TEXT'],
             ['task_category_source', "ALTER TABLE runs ADD COLUMN task_category_source TEXT DEFAULT 'auto'"],
@@ -560,6 +713,18 @@ export class ApmeStore {
             }
             catch { /* ignore */ }
         }
+        if (!turnCols.includes('model_id')) {
+            try {
+                this.db.exec('ALTER TABLE turns ADD COLUMN model_id TEXT');
+            }
+            catch { /* ignore */ }
+        }
+        if (!turnCols.includes('provider')) {
+            try {
+                this.db.exec('ALTER TABLE turns ADD COLUMN provider TEXT');
+            }
+            catch { /* ignore */ }
+        }
         const evalCols = this.db.prepare("PRAGMA table_info(evals)").all().map(c => c.name);
         if (!evalCols.includes('task_id')) {
             try {
@@ -572,10 +737,12 @@ export class ApmeStore {
         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'],
+            ['provider', 'ALTER TABLE tasks ADD COLUMN provider 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'],
+            ['cost_known', 'ALTER TABLE tasks ADD COLUMN cost_known INTEGER NOT NULL DEFAULT 0'],
             ['latency_ms', 'ALTER TABLE tasks ADD COLUMN latency_ms INTEGER'],
         ]) {
             if (!taskCols.includes(col)) {
@@ -597,6 +764,12 @@ export class ApmeStore {
         //     pointer to the run it continues, so one conversation shows up as N
         //     disconnected components. One live session here had 127 such runs.
         const sevCols = this.db.prepare("PRAGMA table_info(sample_events)").all().map(c => c.name);
+        if (!sevCols.includes('cost_known')) {
+            try {
+                this.db.exec('ALTER TABLE sample_events ADD COLUMN cost_known INTEGER NOT NULL DEFAULT 0');
+            }
+            catch { /* ignore */ }
+        }
         if (!sevCols.includes('turn_id')) {
             try {
                 this.db.exec('ALTER TABLE sample_events ADD COLUMN turn_id TEXT');
@@ -626,6 +799,21 @@ export class ApmeStore {
             }
             catch { /* ignore */ }
         }
+        // A positive historical value proves that a price existed. Historical
+        // zeroes remain unknown because they could be either a free local model or
+        // the old UNKNOWN_PRICE fallback — the ambiguity this column removes.
+        try {
+            this.db.exec('UPDATE runs SET cost_known=1 WHERE cost_usd > 0');
+        }
+        catch { /* ignore */ }
+        try {
+            this.db.exec('UPDATE tasks SET cost_known=1 WHERE cost_usd > 0');
+        }
+        catch { /* ignore */ }
+        try {
+            this.db.exec('UPDATE sample_events SET cost_known=1 WHERE cost_usd > 0');
+        }
+        catch { /* ignore */ }
         // ── Covering indexes for the per-run/per-task rollups ──
         // better-sqlite3 is synchronous on a single connection, so ANY slow query
         // here stalls the daemon's whole HTTP path — this is a latency budget, not
@@ -647,6 +835,16 @@ export class ApmeStore {
             }
             catch { /* ignore */ }
         }
+        // CREATE VIEW IF NOT EXISTS preserves stale executable SQL forever. Rebuild
+        // after the ALTERs above so both fresh and upgraded databases group by the
+        // turn/provider identity this version actually writes.
+        for (const view of ['v_sample_scorecard', 'v_category_scorecard', 'v_model_scorecard', 'v_run_metrics']) {
+            try {
+                this.db.exec(`DROP VIEW IF EXISTS ${view}`);
+            }
+            catch { /* ignore */ }
+        }
+        this.db.exec(SCORECARD_DDL);
     }
     seedDefaultRubric() {
         if (!this.db)
@@ -688,11 +886,12 @@ export class ApmeStore {
     insertRun(row) {
         if (!this.db)
             return;
+        const costKnown = row.costKnown === true || (row.costUsd != null && row.costUsd > 0);
         this.db.prepare(`INSERT INTO runs
-        (id, session_id, agent_type, model_id, project_name, project_path, task_prompt,
-         started_at, ended_at, input_tokens, output_tokens, cost_usd, exit_code,
+        (id, session_id, agent_type, model_id, provider, project_name, project_path, task_prompt,
+         started_at, ended_at, input_tokens, output_tokens, cost_usd, cost_known, exit_code,
          git_before, git_after, hw_profile)
-       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`).run(row.id, row.sessionId, row.agentType, row.modelId ?? null, row.projectName ?? null, row.projectPath ?? null, row.taskPrompt ?? null, row.startedAt, row.endedAt ?? null, row.inputTokens ?? null, row.outputTokens ?? null, row.costUsd ?? null, row.exitCode ?? null, row.gitBefore ?? null, row.gitAfter ?? null, row.hwProfile ?? null);
+       VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)`).run(row.id, row.sessionId, row.agentType, row.modelId ?? null, row.provider ?? null, row.projectName ?? null, row.projectPath ?? null, row.taskPrompt ?? null, row.startedAt, row.endedAt ?? null, row.inputTokens ?? null, row.outputTokens ?? null, row.costUsd ?? null, costKnown ? 1 : 0, row.exitCode ?? null, row.gitBefore ?? null, row.gitAfter ?? null, row.hwProfile ?? null);
     }
     updateRun(id, patch) {
         if (!this.db)
@@ -701,6 +900,7 @@ export class ApmeStore {
         const values = [];
         const map = {
             modelId: 'model_id',
+            provider: 'provider',
             projectName: 'project_name',
             projectPath: 'project_path',
             taskPrompt: 'task_prompt',
@@ -708,6 +908,7 @@ export class ApmeStore {
             inputTokens: 'input_tokens',
             outputTokens: 'output_tokens',
             costUsd: 'cost_usd',
+            costKnown: 'cost_known',
             exitCode: 'exit_code',
             gitBefore: 'git_before',
             gitAfter: 'git_after',
@@ -726,7 +927,11 @@ export class ApmeStore {
             if (!col || v === undefined)
                 continue;
             fields.push(`${col} = ?`);
-            values.push(v);
+            values.push(k === 'costKnown' ? (v ? 1 : 0) : v);
+        }
+        if (patch.costUsd != null && patch.costUsd > 0 && patch.costKnown === undefined) {
+            fields.push('cost_known = ?');
+            values.push(1);
         }
         if (fields.length === 0)
             return;
@@ -784,6 +989,7 @@ export class ApmeStore {
             outcome: 'outcome', compositeScore: 'composite_score', efficiencyJson: 'efficiency_json',
             prompt: 'prompt', response: 'response', taskId: 'task_id',
             endSource: 'end_source',
+            modelId: 'model_id', provider: 'provider',
         };
         const sets = [];
         const vals = [];
@@ -886,10 +1092,12 @@ export class ApmeStore {
             notesJson: 'notes_json',
             boundarySignal: 'boundary_signal',
             modelId: 'model_id',
+            provider: 'provider',
             modelConfig: 'model_config',
             inputTokens: 'input_tokens',
             outputTokens: 'output_tokens',
             costUsd: 'cost_usd',
+            costKnown: 'cost_known',
             latencyMs: 'latency_ms',
         };
         const sets = [];
@@ -899,7 +1107,11 @@ export class ApmeStore {
             if (!col || v === undefined)
                 continue;
             sets.push(`${col} = ?`);
-            vals.push(v);
+            vals.push(k === 'costKnown' ? (v ? 1 : 0) : v);
+        }
+        if (patch.costUsd != null && patch.costUsd > 0 && patch.costKnown === undefined) {
+            sets.push('cost_known = ?');
+            vals.push(1);
         }
         if (sets.length === 0)
             return;
@@ -945,36 +1157,18 @@ export class ApmeStore {
             return { total: 0, tasks: [] };
         const limit = Math.min(Math.max(opts.limit ?? 50, 1), 500);
         const offset = Math.max(opts.offset ?? 0, 0);
-        const where = [];
-        const args = [];
-        // `_empty` runs are bookkeeping shells, never work the user did.
-        where.push("COALESCE(r.task_category, '') != '_empty'");
-        if (opts.agentType) {
-            where.push('r.agent_type = ?');
-            args.push(opts.agentType);
-        }
-        if (opts.projectName) {
-            where.push('r.project_name = ?');
-            args.push(opts.projectName);
-        }
-        if (opts.category) {
-            where.push('COALESCE(t.task_category, r.task_category) = ?');
-            args.push(opts.category);
-        }
-        if (opts.outcome) {
-            where.push('t.outcome = ?');
-            args.push(opts.outcome);
-        }
+        const { where, args } = buildTaskFilterWhere(opts);
+        const attentionSql = taskAttentionSql(Date.now() - TASK_ATTENTION_WINDOW_MS);
+        if (opts.view)
+            where.push(opts.view === 'attention' ? attentionSql : TASK_VIEW_SQL[opts.view]);
         if (opts.state === 'closed')
             where.push('t.ended_at IS NOT NULL');
         if (opts.state === 'open')
             where.push('t.ended_at IS NULL');
-        if (opts.q) {
-            where.push('(t.summary LIKE ? OR r.task_prompt LIKE ?)');
-            const like = `%${opts.q}%`;
-            args.push(like, like);
-        }
         const whereSql = `WHERE ${where.join(' AND ')}`;
+        const orderSql = opts.order === 'attention'
+            ? 'ORDER BY attention DESC, t.started_at DESC'
+            : 'ORDER BY t.started_at DESC';
         const total = this.db.prepare(`SELECT COUNT(*) AS n FROM tasks t JOIN runs r ON r.id = t.run_id ${whereSql}`).get(...args).n;
         const rows = this.db.prepare(`SELECT t.*,
               r.session_id, r.agent_type, r.model_id AS run_model_id, r.project_name,
@@ -983,15 +1177,78 @@ export class ApmeStore {
               (SELECT COUNT(*) FROM turns tu WHERE tu.task_id = t.id AND tu.response IS NOT NULL) AS answered_turns,
               (SELECT COUNT(*) FROM sample_events se WHERE se.task_id = t.id) AS event_count,
               (SELECT COUNT(*) FROM sample_events se WHERE se.task_id = t.id AND se.kind = 'tool') AS tool_count,
+              (SELECT COALESCE(SUM(COALESCE(tu.files_modified, 0) + COALESCE(tu.files_created, 0)), 0)
+                 FROM turns tu WHERE tu.task_id = t.id) AS files_touched,
               (SELECT tu.prompt FROM turns tu WHERE tu.task_id = t.id ORDER BY tu.turn_index ASC LIMIT 1) AS first_prompt,
               (SELECT e.score FROM evals e WHERE e.task_id = t.id AND e.metric = 'overall' ORDER BY e.created_at DESC LIMIT 1) AS overall_score,
-              (SELECT COUNT(*) FROM evals e WHERE e.task_id = t.id) AS eval_count
+              (SELECT COUNT(*) FROM evals e WHERE e.task_id = t.id) AS eval_count,
+              (${attentionSql}) AS attention
        FROM tasks t JOIN runs r ON r.id = t.run_id
        ${whereSql}
-       ORDER BY t.started_at DESC
+       ${orderSql}
        LIMIT ? OFFSET ?`).all(...args, limit, offset);
         return { total, tasks: rows.map(rowToTaskListRow) };
     }
+    /** 10 s TTL cache for `taskViewCounts`, keyed by the filter tuple. The Work
+     *  board polls every 15 s and refetches on every interaction; the badges
+     *  are a full-table aggregate with correlated eval subqueries, so paying it
+     *  at most once per TTL bounds the cost of leaving the dashboard open. */
+    viewCountsCache = new Map();
+    /** Per-lifecycle-bucket totals for the Work board's tab badges. Takes the
+     *  same NARROWING filters as `listTaskPage` (agent/project/category/
+     *  outcome/q) so a filtered board's badges count what its tabs would list.
+     *  Buckets and filters read the same SQL definitions; the TTL means a badge
+     *  may lag its rows by up to 10 s after a task moves buckets — a bounded
+     *  staleness, not a different definition. */
+    taskViewCounts(filters = {}) {
+        const empty = { all: 0, attention: 0, inprogress: 0, judged: 0, reported: 0, orphaned: 0 };
+        if (!this.db)
+            return { ...empty };
+        const key = JSON.stringify([filters.agentType, filters.sessionId, filters.projectName, filters.category, filters.outcome, filters.q]);
+        const cached = this.viewCountsCache.get(key);
+        if (cached && Date.now() - cached.at < 10_000)
+            return cached.value;
+        const attentionSql = taskAttentionSql(Date.now() - TASK_ATTENTION_WINDOW_MS);
+        const { where, args } = buildTaskFilterWhere(filters);
+        const row = this.db.prepare(`SELECT COUNT(*) AS all_n,
+              SUM(CASE WHEN ${attentionSql} THEN 1 ELSE 0 END) AS attention_n,
+              SUM(CASE WHEN ${TASK_VIEW_SQL.inprogress} THEN 1 ELSE 0 END) AS inprogress_n,
+              SUM(CASE WHEN ${TASK_VIEW_SQL.judged} THEN 1 ELSE 0 END) AS judged_n,
+              SUM(CASE WHEN ${TASK_VIEW_SQL.reported} THEN 1 ELSE 0 END) AS reported_n,
+              SUM(CASE WHEN ${TASK_VIEW_SQL.orphaned} THEN 1 ELSE 0 END) AS orphaned_n
+       FROM tasks t JOIN runs r ON r.id = t.run_id
+       WHERE ${where.join(' AND ')}`).get(...args);
+        const value = {
+            all: row.all_n ?? 0,
+            attention: row.attention_n ?? 0,
+            inprogress: row.inprogress_n ?? 0,
+            judged: row.judged_n ?? 0,
+            reported: row.reported_n ?? 0,
+            orphaned: row.orphaned_n ?? 0,
+        };
+        if (this.viewCountsCache.size > 20)
+            this.viewCountsCache.clear();
+        this.viewCountsCache.set(key, { at: Date.now(), value });
+        return value;
+    }
+    /** Per-tool call counts for a page of tasks, in one grouped query. Feeds the
+     *  shared `foldActionCounts` projection — the store returns raw counts, the
+     *  HTTP layer folds. */
+    toolCountsForTasks(taskIds) {
+        const out = new Map();
+        if (!this.db || taskIds.length === 0)
+            return out;
+        const placeholders = taskIds.map(() => '?').join(',');
+        const rows = this.db.prepare(`SELECT task_id, tool_name, COUNT(*) AS n FROM sample_events
+       WHERE kind = 'tool' AND tool_name IS NOT NULL AND task_id IN (${placeholders})
+       GROUP BY task_id, tool_name`).all(...taskIds);
+        for (const r of rows) {
+            const list = out.get(r.task_id) ?? [];
+            list.push({ name: r.tool_name, count: r.n });
+            out.set(r.task_id, list);
+        }
+        return out;
+    }
     /** Distinct values behind the task list's filters, so the UI offers what the
      *  data actually contains rather than a hardcoded menu. */
     taskFacets() {
@@ -1280,10 +1537,14 @@ export class ApmeStore {
     insertSampleEvent(row) {
         if (!this.db)
             return false;
+        // A positive monetary amount is self-proving provenance even for legacy
+        // callers that predate costKnown. Zero remains ambiguous and must only be
+        // marked known by a pricing table / known-local producer.
+        const costKnown = row.costKnown === true || (row.costUsd != null && row.costUsd > 0);
         const res = this.db.prepare(`INSERT OR IGNORE INTO sample_events
         (task_id, run_id, turn_index, turn_id, 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.turnId ?? 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);
+         cost_usd, cost_known, latency_ms, tool_name, tool_status, tool_error, payload, dedup_key)
+       VALUES (?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?,?)`).run(row.taskId, row.runId, row.turnIndex ?? null, row.turnId ?? null, row.seq, row.ts, row.kind, row.model ?? null, row.inputTokens ?? null, row.outputTokens ?? null, row.costUsd ?? null, costKnown ? 1 : 0, 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. */
@@ -1292,7 +1553,7 @@ export class ApmeStore {
             return;
         const map = {
             toolStatus: 'tool_status', toolError: 'tool_error', payload: 'payload',
-            costUsd: 'cost_usd', latencyMs: 'latency_ms', model: 'model',
+            costUsd: 'cost_usd', costKnown: 'cost_known', latencyMs: 'latency_ms', model: 'model',
             inputTokens: 'input_tokens', outputTokens: 'output_tokens', ts: 'ts',
         };
         const sets = [];
@@ -1302,7 +1563,11 @@ export class ApmeStore {
             if (!col || v === undefined)
                 continue;
             sets.push(`${col} = ?`);
-            vals.push(v);
+            vals.push(k === 'costKnown' ? (v ? 1 : 0) : v);
+        }
+        if (fields.costUsd != null && fields.costUsd > 0 && fields.costKnown === undefined) {
+            sets.push('cost_known = ?');
+            vals.push(1);
         }
         if (sets.length === 0)
             return;
@@ -1319,6 +1584,17 @@ export class ApmeStore {
        ORDER BY seq DESC LIMIT 1`).get(taskId, turnIndex, toolName);
         return row ? rowToSampleEvent(row) : null;
     }
+    /** One assistant-message slot per turn. Response producers overlap, so text
+     *  hashes cannot be the identity: a truncated projection and the full final
+     *  are two payload versions of the same logical event. */
+    findAssistantMessageEvent(taskId, turnId) {
+        if (!this.db)
+            return null;
+        const row = this.db.prepare(`SELECT * FROM sample_events
+       WHERE task_id = ? AND turn_id = ? AND kind = 'assistant_message'
+       ORDER BY seq ASC LIMIT 1`).get(taskId, turnId);
+        return row ? rowToSampleEvent(row) : null;
+    }
     listSampleEventRows(taskId) {
         if (!this.db)
             return [];
@@ -1369,6 +1645,7 @@ export class ApmeStore {
                 inputTokens: task.inputTokens ?? 0,
                 outputTokens: task.outputTokens ?? 0,
                 costUsd: task.costUsd ?? 0,
+                costKnown: task.costKnown ?? false,
                 latencyMs: task.latencyMs ?? 0,
             },
             summary: task.summary ?? null,
@@ -1382,10 +1659,15 @@ export class ApmeStore {
         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
+              CASE WHEN COUNT(*) > 0 AND MIN(COALESCE(cost_known,0))=1
+                   THEN SUM(cost_usd) ELSE NULL END AS cu,
+              CASE WHEN COUNT(*) > 0 AND MIN(COALESCE(cost_known,0))=1
+                   THEN 1 ELSE 0 END AS ck,
+              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,
+            inputTokens: row.it, outputTokens: row.ot, costUsd: row.cu,
+            costKnown: row.ck === 1, latencyMs: row.lm,
         });
     }
     sampleScorecard() {
@@ -1395,10 +1677,12 @@ export class ApmeStore {
         return rows.map((r) => ({
             agentType: r.agent_type,
             modelId: r.model_id,
+            provider: r.provider ?? null,
             taskCategory: r.task_category ?? null,
             samples: r.samples,
             avgQuality: r.avg_quality ?? null,
             totalCost: r.total_cost ?? null,
+            costKnown: r.cost_known === 1,
             avgLatencyMs: r.avg_latency_ms ?? null,
             costPerQuality: r.cost_per_quality ?? null,
         }));
@@ -1411,10 +1695,12 @@ export class ApmeStore {
         return rows.map((r) => ({
             agentType: r.agent_type,
             modelId: r.model_id,
+            provider: r.provider ?? null,
             runs: r.runs,
             avgOverall: r.avg_overall ?? null,
             avgTestsPass: r.avg_tests_pass ?? null,
             totalCost: r.total_cost ?? null,
+            costKnown: r.cost_known === 1,
             costPerQuality: r.cost_per_quality ?? null,
         }));
     }
@@ -1425,10 +1711,12 @@ export class ApmeStore {
         return rows.map((r) => ({
             taskCategory: r.task_category,
             modelId: r.model_id,
+            provider: r.provider ?? null,
             runs: r.runs,
             avgOverall: r.avg_overall ?? null,
             avgTestsPass: r.avg_tests_pass ?? null,
             totalCost: r.total_cost ?? null,
+            costKnown: r.cost_known === 1,
         }));
     }
 }
@@ -1439,6 +1727,7 @@ function rowToRun(r) {
         sessionId: r.session_id,
         agentType: r.agent_type,
         modelId: r.model_id ?? null,
+        provider: r.provider ?? null,
         projectName: r.project_name ?? null,
         projectPath: r.project_path ?? null,
         taskPrompt: r.task_prompt ?? null,
@@ -1447,6 +1736,7 @@ function rowToRun(r) {
         inputTokens: r.input_tokens ?? null,
         outputTokens: r.output_tokens ?? null,
         costUsd: r.cost_usd ?? null,
+        costKnown: r.cost_known === 1,
         exitCode: r.exit_code ?? null,
         gitBefore: r.git_before ?? null,
         gitAfter: r.git_after ?? null,
@@ -1477,10 +1767,12 @@ function rowToTask(r) {
         taskCategory: r.task_category ?? null,
         notesJson: r.notes_json ?? null,
         modelId: r.model_id ?? null,
+        provider: r.provider ?? null,
         modelConfig: r.model_config ?? null,
         inputTokens: r.input_tokens ?? null,
         outputTokens: r.output_tokens ?? null,
         costUsd: r.cost_usd ?? null,
+        costKnown: r.cost_known === 1,
         latencyMs: r.latency_ms ?? null,
     };
 }
@@ -1497,12 +1789,17 @@ function rowToTaskListRow(r) {
         projectPath: r.project_path ?? null,
         parentRunId: r.parent_run_id ?? null,
         firstPrompt: r.first_prompt ?? r.run_prompt ?? null,
+        // The task's OWN first turn prompt, with NO run-prompt fallback: title
+        // derivation must never name task 2+ of a split run after task 0's intent.
+        ownFirstPrompt: r.first_prompt ?? null,
         turnCount: r.turn_count ?? 0,
         answeredTurns: r.answered_turns ?? 0,
         eventCount: r.event_count ?? 0,
         toolCount: r.tool_count ?? 0,
+        filesTouched: r.files_touched ?? 0,
         evalCount: r.eval_count ?? 0,
         overallScore: r.overall_score ?? r.composite_score ?? null,
+        attention: r.attention === 1,
     };
 }
 function rowToSampleEvent(r) {
@@ -1519,6 +1816,7 @@ function rowToSampleEvent(r) {
         inputTokens: r.input_tokens ?? null,
         outputTokens: r.output_tokens ?? null,
         costUsd: r.cost_usd ?? null,
+        costKnown: r.cost_known === 1,
         latencyMs: r.latency_ms ?? null,
         toolName: r.tool_name ?? null,
         toolStatus: r.tool_status ?? null,