synabun 2.0.0-beta → 2.0.5

package/neural-interface/lib/assistant-dispatch.js

@@ -29,6 +29,7 @@ import { TASK_CLASS_META, classPlaybook, collectedMedia, describeTarget, designF
 import { MAX_MEDIA_BYTES, MAX_RUN_MEDIA_BYTES, copyRunMedia, mediaType, mediaUrl, servableRunMedia } from './assistant-media.js';
 import { loadPlaybook } from './assistant-playbooks.js';
 import { loadStyleGuideForRun } from './style-guide/store.js';
+import { claimCheckWaived, normalizeTestDensity, testingApplies } from './assistant-test-density.js';
 import { normalizeEffort } from './effort-levels.js';
 import { createTurnEvidence, recordProviderEvent, evidenceSnapshot, claimEvidenceLine } from './assistant-evidence.js';
 import { codexUsageCostUsd } from './assistant-budget.js';
@@ -374,6 +375,9 @@ export function createAssistantDispatcher({
   // The project's Style Guide for a run's prompt (lib/style-guide/store.js): (cwd, taskClass) =>
   // { summary, designPath, tokenFiles, proposals } | null. Reads <dataDir>/style-guides.
   loadStyleGuide = (cwd, taskClass) => loadStyleGuideForRun(cwd, taskClass, { dataDir }),
+  // The session's Test density (assistant-test-density.js): (assistantSessionId) => level | null.
+  // Read when a prompt is composed; null or an unknown value is the default level.
+  testDensity = () => null,
 } = {}) {
   if (typeof getRuntime !== 'function') throw new Error('createAssistantDispatcher requires getRuntime()');
   if (!dataDir) throw new Error('createAssistantDispatcher requires dataDir');
@@ -1262,7 +1266,9 @@ export function createAssistantDispatcher({
     const ask = {
       status: noBlock && !entry.outputSchema,
       cause: noBlock || parsed.status === 'blocked',
-      claim: (noBlock || parsed.status === 'done') && evidence.hasEvidence,
+      // At Test density None a "Not tested" result is what the brief asked for, not a claim to check.
+      claim: (noBlock || parsed.status === 'done') && evidence.hasEvidence
+        && !claimCheckWaived({ level: entry.testDensity || null, taskClass: entry.taskClass || null, text, evidence }),
     };
     if (!ask.status && !ask.cause && !ask.claim) return null;
     const judged = await bounded(judge.workerOutcome({
@@ -1349,6 +1355,7 @@ export function createAssistantDispatcher({
       entry, ctx: null, ring: [], ringChars: 0, followUps: [], waitResolve: null, idleTimer: null,
       pendingResolvers: new Map(), dead: false, lastEventAt: now(), stalledFlagged: false, turnCounter: 0, abortReason: null,
       evidence: null, // the current turn's commands / tool results / edited files (assistant-evidence.js)
+      completed: false, // its turn loop ended on a completion (agent_complete, a closed finished tab, a removal)
     };
     live.set(entry.runId, liveState);
     const extraEnv = {};
@@ -1403,7 +1410,7 @@ export function createAssistantDispatcher({
         source: 'assistant',
         focus: entry.focus === true,
         title: entry.title,
-        buildPrompt: () => buildTaskPrompt(playbook || styleGuide ? { ...state, ...(playbook ? { playbook } : {}), ...(styleGuide ? { styleGuide } : {}) } : state),
+        buildPrompt: () => buildTaskPrompt({ ...state, ...(playbook ? { playbook } : {}), ...(styleGuide ? { styleGuide } : {}), testDensity: briefTestDensity(entry) }),
         wrapAdapter: (adapter, ctx) => createTaskAdapter(adapter, ctx, liveState, state),
       });
       entry.runtime = { status: descriptor.status };
@@ -1428,6 +1435,27 @@ export function createAssistantDispatcher({
   function styleGuideFor(entry) {
     try { return loadStyleGuide(entry.cwd, entry.taskClass || null) || null; } catch { return null; }
   }
+  /** The session's Test density right now (the user's setting; the default when it cannot be read). */
+  function testDensityFor(entry) {
+    try { return normalizeTestDensity(testDensity(entry.assistantSessionId || null)); } catch { return normalizeTestDensity(null); }
+  }
+  /**
+   * The level a run's brief carries, read when the prompt is composed and kept on the entry: the
+   * claim check and a later follow-up go by what the worker was told. null for a class with no
+   * TESTING block.
+   */
+  function briefTestDensity(entry) {
+    entry.testDensity = testingApplies(entry.taskClass || null) ? testDensityFor(entry) : null;
+    return entry.testDensity;
+  }
+  /** A follow-up turn names the level only when the user changed it since the run was last told. */
+  function changedTestDensity(entry) {
+    if (!testingApplies(entry.taskClass || null)) return null;
+    const level = testDensityFor(entry);
+    if (level === (entry.testDensity || null)) return null;
+    entry.testDensity = level;
+    return level;
+  }
   /**
    * The rules a run of the entry's class follows (assistant-playbooks.js: the user's override
    * at <dataDir>/playbooks/<name>.md, else the shipped copy), or null. The run's notes say
@@ -1674,7 +1702,7 @@ export function createAssistantDispatcher({
           emit(entry, 'idle');
           const next = await waitForFollowUp();
           if (next.kind === 'abort') { entry.completionReason = entry.completionReason || liveState.abortReason || 'aborted'; return { providerSessionId: inner.identity?.()?.providerSessionId }; }
-          if (next.kind === 'complete') { entry.completionReason = 'complete'; return { providerSessionId: inner.identity?.()?.providerSessionId }; }
+          if (next.kind === 'complete') { entry.completionReason = next.reason || 'complete'; liveState.completed = true; return { providerSessionId: inner.identity?.()?.providerSessionId }; }
           if (next.kind === 'timeout') { entry.completionReason = 'idle_timeout'; return { providerSessionId: inner.identity?.()?.providerSessionId }; }
           if (liveState.turnCounter >= entry.maxTurns) {
             entry.completionReason = 'max_turns';
@@ -1688,7 +1716,7 @@ export function createAssistantDispatcher({
             return budgetStop(entry, blocked);
           }
           try {
-            const followPrompt = buildFollowUpPrompt(next.text, state, { turn: liveState.turnCounter + 1, maxTurns: entry.maxTurns, origin: next.origin });
+            const followPrompt = buildFollowUpPrompt(next.text, state, { turn: liveState.turnCounter + 1, maxTurns: entry.maxTurns, origin: next.origin, testDensity: changedTestDensity(entry) });
             await runInnerTurn(followPrompt, meta, { origin: next.origin, label: clip(next.text, 80), followUp: next.text, taskId: next.taskId || entry.taskId || null });
           } catch (error) {
             return handleTurnError(error, liveState, entry, ctx);
@@ -2404,12 +2432,30 @@ export function createAssistantDispatcher({
     emit(entry, 'follow_up_queued', { queued: liveState.followUps.length });
     return { ok: true, queued: true, position: liveState.followUps.length, run: view(entry) };
   }
-  function complete(runId) {
+  function complete(runId, { reason = 'complete' } = {}) {
     const entry = requireEntry(runId);
     const liveState = live.get(runId);
     if (TERMINAL_STATES.has(entry.state)) return { ok: true, run: view(entry) };
-    if (liveState?.waitResolve) { liveState.waitResolve({ kind: 'complete' }); return { ok: true, run: view(entry) }; }
-    return stop(runId, 'complete');
+    if (liveState?.waitResolve) { liveState.waitResolve({ kind: 'complete', reason }); return { ok: true, run: view(entry) }; }
+    return stop(runId, reason);
+  }
+  /**
+   * The user closed the sidepanel tab of a worker that reported its task done
+   * and waits warm for a follow-up (its X, or Back to Assistant): the run is
+   * completed the way agent_complete completes it, with the reason
+   * `tab_closed`. Its result stays as the worker reported it, and the brain is
+   * handed no stop to react to. Any other run is refused (RUN_NOT_FINISHED):
+   * the tab then stops it as its X always did. A run that already ended is
+   * left as it is.
+   */
+  function completeClosedTab(runId) {
+    const entry = requireEntry(runId);
+    if (TERMINAL_STATES.has(entry.state)) return { ok: true, completed: false, alreadyTerminal: true, run: view(entry) };
+    if (!finishedIdle(entry, { outcomes: ['done'] }) || !live.get(runId)?.waitResolve) {
+      const outcome = entry.lastResult?.status || null;
+      throw new DispatchError('RUN_NOT_FINISHED', `Run ${runId} is ${entry.state}${outcome ? ` (last result: ${outcome})` : ''}: closing its tab completes only a worker that reported its task done and waits for a follow-up.`, { status: 409, state: entry.state, outcome });
+    }
+    return { ...complete(runId, { reason: 'tab_closed' }), completed: true };
   }
   async function stop(runId, reason = 'assistant') {
     const entry = requireEntry(runId);
@@ -2442,13 +2488,16 @@ export function createAssistantDispatcher({
   // ── runtime feed → registry ────────────────────────────────────────────────
   function finalize(entry, runtimeRun) {
     if (TERMINAL_STATES.has(entry.state)) return;
-    const status = runtimeRun?.status || 'completed';
+    const liveState = live.get(entry.runId);
+    let status = runtimeRun?.status || 'completed';
+    // A completed worker whose loop file was swept meanwhile (a stale-file cleanup) reaches the
+    // runtime's next check as "stopped": its turn loop had already ended on the completion.
+    if (status === 'stopped' && liveState?.completed && /^state_(removed|inactive)$/.test(String(runtimeRun?.stoppedReason || ''))) status = 'completed';
     entry.state = TERMINAL_STATES.has(status) ? status : 'completed';
     entry.turnState = 'terminal';
     entry.finishedAt = iso(now());
     entry.completionReason = entry.completionReason || runtimeRun?.stoppedReason || (status === 'completed' ? 'complete' : status);
     if (runtimeRun?.error) entry.error = runtimeRun.error;
-    const liveState = live.get(entry.runId);
     try { finishUsage(entry, liveState || null); } catch (error) { log(entry.runId, 'assistant:meter-error', error?.message || String(error)); }
     if (liveState) {
       liveState.dead = true;
@@ -2654,9 +2703,9 @@ export function createAssistantDispatcher({
    * failed run whose cause is still being judged is not finished yet.
    */
   function removable(entry) { return TERMINAL_STATES.has(entry.state) && !entry.judging; }
-  function finishedIdle(entry) {
+  function finishedIdle(entry, { outcomes = ['done', 'blocked'] } = {}) {
     return entry.state === 'idle' && !entry.judging
-      && ['done', 'blocked'].includes(entry.lastResult?.status)
+      && outcomes.includes(entry.lastResult?.status)
       && !(live.get(entry.runId)?.followUps?.length);
   }
   async function finishIdleForRemoval(entry) {
@@ -2700,7 +2749,7 @@ export function createAssistantDispatcher({
       && (!assistantSessionId || entry.assistantSessionId === String(assistantSessionId))
       && (!workflowId || entry.workflowId === String(workflowId))
       && (!wanted || wanted.has(entry.runId)));
-    await Promise.all(scoped.filter(finishedIdle).map((entry) => finishIdleForRemoval(entry).catch(() => {})));
+    await Promise.all(scoped.filter((entry) => finishedIdle(entry)).map((entry) => finishIdleForRemoval(entry).catch(() => {})));
     const targets = scoped.filter(removable);
     return { ok: true, removed: markRemoved(targets), skipped: scoped.filter((entry) => !removable(entry)).map((entry) => entry.runId) };
   }
@@ -2731,7 +2780,7 @@ export function createAssistantDispatcher({
   stalledTimer.unref?.();
 
   return {
-    dispatch, sendTurn, complete, stop, killAll, get, peek, list, totals, wait, transcript, pendingPermissions,
+    dispatch, sendTurn, complete, completeClosedTab, stop, killAll, get, peek, list, totals, wait, transcript, pendingPermissions,
     respondPermission, subscribe, shutdown, limits,
     sessionBudget, checkSession, refreshLimits,
     resolveRoute, declineRoute, heldRouteRefusal, escalate,

package/neural-interface/lib/assistant-envelope.js

@@ -158,13 +158,16 @@ export function codexControlResponse(pending, response = {}) {
   // Codex matches a reply by id value and type: 0 and "0" are different requests to it.
   const base = { requestId: native.rpcId ?? pending.request_id, responseToken: randomUUID(), threadId: native.threadId || null, turnId: native.turnId || null, toolCallId: native.toolCallId || null };
   if (method === 'item/commandExecution/requestApproval' || method === 'item/fileChange/requestApproval') {
-    const available = Array.isArray(native.availableDecisions) && native.availableDecisions.length
-      ? native.availableDecisions.map((d) => (typeof d === 'string' ? d : Object.keys(d)[0]))
+    const available = Array.isArray(native.availableDecisions)
+      ? native.availableDecisions
       : ['accept', 'acceptForSession', 'decline', 'cancel'];
-    let decision;
-    if (response.decision && available.includes(response.decision)) decision = response.decision;
-    else if (allow) decision = response.always && available.includes('acceptForSession') ? 'acceptForSession' : 'accept';
-    else decision = response.cancel && available.includes('cancel') ? 'cancel' : 'decline';
+    // The UI puts policy amendment objects in result.decision and only their
+    // label in decision. Preserve the wire value for strict server validation.
+    let decision = response.result?.decision ?? response.decision;
+    if (decision == null) {
+      if (allow) decision = response.always && available.includes('acceptForSession') ? 'acceptForSession' : 'accept';
+      else decision = response.cancel && available.includes('cancel') ? 'cancel' : 'decline';
+    }
     return { ...base, result: { decision } };
   }
   if (method === 'item/permissions/requestApproval') {