@piagent/platform 1.49.0 → 1.50.0

package/packages/piagent-core/managed/command-tools.mjs

@@ -1,5 +1,6 @@
 import { spawn } from 'node:child_process';
 import { beforeCheckOrPush } from './check-questions.mjs';
+import { approvalRefusal } from './approval-refusal.mjs';
 import { asksOnlyHandoff, broadAccess, effectivePermission, networkConfirmation } from './permission.mjs';
 import { handoffReason } from '../runtime/policy/handoff-policy.mjs';
 import { engineAlive, engineLauncher } from './container-engine.mjs';
@@ -38,7 +39,7 @@ const KINDS = {
   },
   host: {
     name: 'run_on_host', label: 'Run on this machine', scope: 'host-command-once', provider: 'host', riskClass: 'high',
-    description: 'Run one shell command on the member\'s machine, outside the sandbox, as the member, after they approve that exact command (only Bypass runs it without asking, and even there a push, PR, publish, deploy or a command that reaches credentials is asked). Use it when the sandbox blocks what the task needs and no other tool covers it: install a runtime or tool the project needs (brew install openjdk@21, brew install node@22, sdk install java…), git pull with the member\'s sign-in, git push only when the member asked for it, a private registry, a file outside the project. It has the member\'s files, Keychain and sign-ins, so ask only for what the task needs and say why in reason. No input can be typed: commands that need the member\'s password (sudo, a .pkg installer) or an interactive sign-in are refused; ask the member with ask_user for those.',
+    description: 'Run one shell command on the member\'s machine, outside the sandbox, as the member, in their login shell (zsh on macOS: an unquoted $var is not split into words, unlike bash), after they approve that exact command (only Bypass runs it without asking, and even there a push, PR, publish, deploy or a command that reaches credentials is asked). Use it when the sandbox blocks what the task needs and no other tool covers it: install a runtime or tool the project needs (brew install openjdk@21, brew install node@22, sdk install java…), git pull with the member\'s sign-in, git push only when the member asked for it, a private registry, a file outside the project. It has the member\'s files, Keychain and sign-ins, so ask only for what the task needs and say why in reason. No input can be typed: commands that need the member\'s password (sudo, a .pkg installer) or an interactive sign-in are refused; ask the member with ask_user for those.',
     allow: 'Run this exact command once on this machine, outside the sandbox, as you: it can read and change your files and use your sign-ins.',
     title: 'Chạy lệnh trên máy (ngoài sandbox)', allowText: 'chạy đúng lệnh này một lần trên máy bạn, ngoài sandbox, bằng quyền của bạn: lệnh đọc/sửa được file và dùng được tài khoản đã đăng nhập trên máy.',
     // Full access does not cover it: the member approves each command (with
@@ -78,6 +79,7 @@ export function stillAsks(self, reason) {
 // member may allow it once, allow the same exact action for the rest of the
 // conversation (never a handoff or a waiver of checks: its scope names it), deny it, or deny it and
 // stop the turn so the agent does not try again.
+const HANDOFF_APPROVAL_MS = 30 * 60_000;
 const CHOICES = { once: 'Đồng ý một lần', session: 'Đồng ý cho cả cuộc trò chuyện (đúng lệnh này)', deny: 'Từ chối', stop: 'Từ chối và dừng lượt này' };
 export async function approve(self, id, signal, ctx, action, { title, body }) {
   const grantable = !/handoff|waiver/.test(String(action.requestedScope ?? '')), key = `${action.toolName}\u0000${action.commandPreview ?? JSON.stringify(action.rawAction ?? {})}`;
@@ -91,9 +93,11 @@ export async function approve(self, id, signal, ctx, action, { title, body }) {
       const picked = await ctx.ui.select(`${title}\n\n${body}`, choices.map(choice => CHOICES[choice]));
       return choices.find(choice => CHOICES[choice] === picked) ?? 'deny';
     } } : {}),
+    // A push waits longer than other cards: the member may be away while a suite ran.
+    ...(/handoff/.test(String(action.requestedScope ?? '')) ? { ttlMs: HANDOFF_APPROVAL_MS } : {}),
     unavailableFallback: 'terminal-confirm', recheck: () => !signal?.aborted && Boolean(self.grant) });
   if (!decision.allowed && decision.choice === 'stop') { try { ctx?.abort?.(); } catch { /* the denial below still stands */ } throw Error('managed-operation-denied-stop: bạn đã từ chối và dừng lượt này; không thử lại, chờ member.'); }
-  if (!decision.allowed || !decision.consume() || signal?.aborted) throw Error('managed-operation-denied');
+  if (!decision.allowed || !decision.consume() || signal?.aborted) throw approvalRefusal(decision);
   if (decision.choice === 'session' && grantable) (self.sessionApprovals ??= new Set()).add(key);
 }
 
@@ -126,7 +130,7 @@ function commandTool(self, kind) {
       } finally {
         // What the command installed (a JDK, Node) is there for the next one.
         if (kind === 'host') self.boundary.refreshEnvironment();
-        await self.claimChanges(before); await self.run?.afterTool('bash', args, ok, () => self.digest()); await self.refreshReview();
+        await self.claimChanges(before); await self.run?.afterTool('bash', args, ok, () => self.digest(), self); await self.refreshReview();
       }
     } };
 }

package/packages/piagent-core/managed/fast-mode.mjs

@@ -1,10 +1,12 @@
 // Fast mode of a company conversation. The member turns it on per
 // conversation, and only where the team's Harness in Studio allows it
 // (harness.configuration.fast_mode); a conversation starts with it off. While
-// on, the main agent's calls ask a Codex model for the priority service tier:
-// the same model and thinking level, answered faster, and charged to the
-// key's quota at a higher rate (Studio decides the rate and refuses the tier
-// for a run whose Harness does not allow it). Subagents run as usual.
+// on, every call of the conversation, the main agent's and its subagents', asks
+// for the priority service tier where its model serves it: the same model and
+// thinking level, answered faster, and charged to the key's quota at a higher
+// rate (Studio decides the rate and refuses the tier for a run whose Harness
+// does not allow it). A model that does not serve it runs at standard speed,
+// so a Claude main agent with Codex subagents still gets it on the subagents.
 const ENTRY = 'agent-watch-fast';
 export const NOTICE = 'agent-watch-fast-notice';
 export const FAST_SERVICE_TIER = 'priority';
@@ -27,7 +29,7 @@ export function fastSetter(owner, manager) {
   return enabled => {
     if (enabled && !fastAllowed(owner.manifest)) throw Error('managed-fast-unavailable');
     if (Boolean(enabled) === Boolean(owner.fast)) return false;
-    owner.fast = Boolean(enabled); manager.appendCustomEntry(ENTRY, { enabled: owner.fast, allowed: fastAllowed(owner.manifest), at: new Date().toISOString() });
+    owner.fast = Boolean(enabled); owner.fastDeclined?.clear(); manager.appendCustomEntry(ENTRY, { enabled: owner.fast, allowed: fastAllowed(owner.manifest), at: new Date().toISOString() });
     return true;
   };
 }
@@ -44,21 +46,22 @@ export const fastServed = route => route?.provider === 'codex' && CODEX_FAST.has
 // Studio's refusal of the tier (the Harness or the model list changed), or
 // the provider's of a request carrying it: asked again once without it.
 export const fastRefusal = code => code === 'policy_denied' || code === 'upstream_request_rejected';
-// A fast request was refused before the model did any work: the conversation
-// goes back to standard speed and says so, instead of failing the turn.
-export function declineFast(owner, code) {
-  owner.fast = false;
+// A fast request was refused before the model did any work: that model goes
+// back to standard speed and the others keep fast mode. Only the main agent's
+// model says so; turning Fast off and on asks again.
+export function declineFast(owner, code, role = 'main', route = null) {
+  (owner.fastDeclined ??= new Set()).add(route?.native ?? '');
+  if (role !== 'main') return;
   try {
-    recordFast(owner, owner.session.sessionManager);
     void owner.session.sendCustomMessage?.({ customType: NOTICE, display: true, details: { code },
-      content: `Fast mode was refused for this request [${code}]: this conversation continues at standard speed. Turn Fast on again to retry.` }, { triggerTurn: false })?.catch?.(() => {});
+      content: `Fast mode was refused for this model [${code}]: its calls continue at standard speed. Turn Fast off and on again to retry.` }, { triggerTurn: false })?.catch?.(() => {});
   } catch { /* the request goes on at standard speed either way */ }
 }
-// The payload hook of one request: the main agent's call on a model that
-// serves fast mode carries the tier while fast mode is on and still allowed
-// (the Harness may have turned it off since the member turned it on).
-export function fastPayload(owner, role, route, onPayload) {
-  if (role !== 'main' || !owner.fast || !fastAllowed(owner.manifest) || !fastServed(route)) return onPayload;
+// The payload hook of one request: a call on a model that serves fast mode
+// carries the tier while fast mode is on and still allowed (the Harness may
+// have turned it off since the member turned it on).
+export function fastPayload(owner, route, onPayload) {
+  if (!owner.fast || !fastAllowed(owner.manifest) || !fastServed(route) || owner.fastDeclined?.has(route.native)) return onPayload;
   return async (payload, model) => {
     const next = (onPayload ? await onPayload(payload, model) : undefined) ?? payload;
     return next && typeof next === 'object' && !Array.isArray(next) ? { ...next, service_tier: FAST_SERVICE_TIER } : next;

package/packages/piagent-core/managed/helper-roles.mjs

@@ -87,7 +87,7 @@ export function countedCheck(managed, tool) {
     try { const result = await tool.execute(...args); ok = !result?.isError; return result; }
     finally {
       const after = counted && before !== null ? await folder() : null;
-      if (after !== null && after === before) await managed.run.afterTool('bash', args[1], ok, () => managed.digest());
+      if (after !== null && after === before) await managed.run.afterTool('bash', args[1], ok, () => managed.digest(), managed);
     }
   } };
 }

package/packages/piagent-core/managed/helper-run.mjs

@@ -4,6 +4,7 @@
 // refusal, starts once more: Studio then gives the new helper the next model
 // of its role that serves (an auto role's list).
 import { fenced, HELPER_UNTRUSTED_RULE } from './untrusted.mjs';
+import { relevantChecks } from './check-evidence.mjs';
 import { jobTitle } from './call-purpose.mjs';
 import { ManagedToolBoundary } from './tool-boundary.mjs';
 import { managedResourceLoader } from './resource-loader.mjs';
@@ -144,7 +145,7 @@ async function answerUnder(managed, key, task, signal, harness, snapshot, parts,
     (managed.jobTitles ??= new WeakMap()).set(runtime, jobTitle(task, title));
     const setup = HELPER_SETUP[role], memory = helperMemoryTools(managed, () => grant, renew);
     boundary = new ManagedToolBoundary({ cwd: managed.cwd, sdkRoot: managed.sdk, protectedRoots: managed.protectedRoots, readOnly: true, commands: setup.commands });
-    const checks = role === 'verify' ? `\n\nRepository checks: ${managed.checks.commands.length ? managed.checks.commands.map(c => '`' + c + '`').join(', ') : 'none declared or detected; choose the tests, type check or build that cover the claims'}.${managed.run?.fullAsked || managed.run?.scopeChoice === 'full' ? '' : ' Run only the tests that cover the claims: the member decides about a whole-suite run.'}` : '';
+    const checks = role === 'verify' ? `\n\nRepository checks: ${managed.checks.commands.length ? relevantChecks(managed, managed.checks.commands).map(c => '`' + c + '`').join(', ') : 'none declared or detected; choose the tests, type check or build that cover the claims'}.${managed.run?.fullAsked || managed.run?.scopeChoice === 'full' ? '' : ' Run only the tests that cover the claims: the member decides about a whole-suite run.'}` : '';
     // One helper session answers the brief (with one part of the patch for a
     // review in parts); its reply and tokens. A helper runs at its role's
     // level from Studio (fixed by the Harness, or its main agent's).
@@ -201,7 +202,7 @@ async function answerUnder(managed, key, task, signal, harness, snapshot, parts,
     for (const session of sessions) session.dispose(); await boundary?.dispose();
     await settleHelper(managed, renew.lease, report); // with the role's current token, before it closes
     await closeRole(managed, key);
-    for (const who of watched) helperState(managed, who, null);
+    for (const who of watched) { helperState(managed, who, null); saveHelperSteps(managed, who); }
   }
 }
 
@@ -216,12 +217,29 @@ export function helperState(managed, key, state) {
 // latest tool step, told to listeners as it changes (company-terminal.mjs).
 function watchHelper(managed, key, session, task, title) {
   const job = String(title || task.split('\n').find(line => line.trim()) || '').trim().slice(0, 120);
+  const logs = managed.helperLogs ??= new Map();
+  const log = logs.get(key) ?? { job, startedAt: new Date().toISOString(), total: 0, steps: [] }; logs.set(key, log);
   helperState(managed, key, { job, step: 'đọc brief' });
   session.subscribe?.(event => {
     if (event?.type !== 'tool_execution_start') return;
-    helperState(managed, key, { job, step: stepText(event.toolName, event.args) });
+    const args = stepArgs(event.args);
+    log.total += 1; log.steps.push({ tool: String(event.toolName), args, at: new Date().toISOString() }); log.steps.splice(0, log.steps.length - HELPER_STEPS_KEPT);
+    helperState(managed, key, { job, step: stepText(event.toolName, event.args), tool: String(event.toolName), args });
   });
 }
+// A helper's session lives in memory: what it called is kept here so the
+// dashboard's Activity can still show it once the helper has ended.
+const HELPER_STEPS_KEPT = 200;
+function saveHelperSteps(managed, key) {
+  const log = managed.helperLogs?.get(key); if (!log) return;
+  managed.helperLogs.delete(key);
+  managed.session?.sessionManager?.appendCustomEntry?.('agent-watch-helper-steps', { agent: key, ...log, endedAt: new Date().toISOString() });
+}
+function stepArgs(args = {}) {
+  const out = {};
+  for (const name of ['path', 'pattern', 'command', 'query', 'url']) if (typeof args?.[name] === 'string') out[name] = args[name].slice(0, 300);
+  return out;
+}
 const STEP = { read: 'đọc', grep: 'tìm', find: 'tìm file', ls: 'xem thư mục', bash: 'chạy', web_search: 'tìm trên web', web_fetch: 'đọc trang', memory_search: 'tra team memory', memory_ask: 'hỏi team memory' };
 export function stepText(tool, args = {}) {
   const what = args?.path ?? args?.pattern ?? args?.command ?? args?.query ?? args?.url ?? '';

package/packages/piagent-core/managed/ledger.mjs

@@ -120,7 +120,7 @@ export function rejudgeBrief(item) {
 // One step of the completion gate (true: the gate goes round again). A
 // rejected premise objection goes back to its helper, which withdraws it or
 // keeps it for the member; then the items the main agent never decided are
-// put to it once, with no new review.
+// put to it once each, with no new review: a later review's findings are new items.
 export async function settleLedger(managed, run, { note, send, signal }) {
   const rejected = toRejudge(run)[0];
   if (rejected) {
@@ -133,9 +133,9 @@ export async function settleLedger(managed, run, { note, send, signal }) {
     await note(`Harness: the ${rejected.role} subagent keeps its objection; the member decides.\n${itemText(rejected)}`, { phase: 'dispute', role: rejected.role, id: rejected.id });
     return true;
   }
-  const pending = undecided(run);
-  if (!pending.length || run.decideAsked) return false;
-  run.decideAsked = true;
+  const asked = run.decidedAsked ??= new Set(), pending = undecided(run).filter(item => !asked.has(item.id));
+  if (!pending.length) return false;
+  for (const item of pending) asked.add(item.id);
   await send(decideText(pending), { phase: 'answer', ids: pending.map(i => i.id), issues: pending.slice(0, 10).map(i => ({ kind: i.kind, detail: i.detail })) });
   return true;
 }
@@ -151,7 +151,7 @@ export function ledgerWords(run) {
   const items = ledgerOf(run), count = s => items.filter(i => i.status === s).length, words = [];
   if (!items.length && !run.minorFindings) return null;
   for (const [status, label] of [['accepted', 'accepted'], ['rejected', 'rejected'], ['withdrawn', 'withdrawn by the helper'], ['cleared', 'cleared by a later check'], ['disputed', 'for the member'], ['open', 'left without a decision'], ['noted', 'noted']])
-    if (count(status)) words.push(`${count(status)} ${label}`);
+    if (count(status)) words.push(`${count(status)} ${label}${status === 'open' ? ` (${items.filter(i => i.status === 'open').map(i => i.id).join(', ')})` : ''}`);
   if (run.minorFindings) words.push(`${run.minorFindings} minor finding(s) listed`);
   return `objections: ${words.join(', ')}`;
 }

package/packages/piagent-core/managed/memory-lessons.mjs

@@ -55,7 +55,7 @@ export class CheckLessons {
  // shape it becomes a failure lesson for the team memory.
  #toolError(call,event,output) {
   if(!COMMAND_TOOLS.has(call.tool)||call.check||this.lessons.length>=MAX_LESSONS)return;
-  if(/managed[-_](?:tool[-_]cancelled|operation[-_]denied|helper[-_]cancelled)|Command aborted|Command timed out/i.test(output))return;
+  if(/managed[-_](?:tool[-_]cancelled|operation[-_](?:denied|expired|cancelled)|helper[-_]cancelled)|Command aborted|Command timed out/i.test(output))return;
   const line=output.split('\n').find(l=>ERROR_LINE.test(l));
   if(!line)return;
   const command=commandShape(call.command);
@@ -85,7 +85,7 @@ export class CheckLessons {
   const output=textOf(event.result), c=call.check;
   const exitFailed=event.isError || (event.result?.details?.exitCode !== undefined && event.result.details.exitCode!==0);
   if(exitFailed || (c.piped && failedOutput(output))) {
-   if(/managed[-_](?:tool[-_]cancelled|operation[-_]denied|helper[-_]cancelled)|Command aborted|Command timed out/i.test(output))return;
+   if(/managed[-_](?:tool[-_]cancelled|operation[-_](?:denied|expired|cancelled)|helper[-_]cancelled)|Command aborted|Command timed out/i.test(output))return;
    this.failed++;this.last='failed';
    const diagnostic=output.split('\n').find(l=>/error|fail|exception|not found|EADDRINUSE/i.test(l)) ?? 'Check exited unsuccessfully';
    if(this.failures.size>=20)this.failures.delete(this.failures.keys().next().value);

package/packages/piagent-core/managed/repository-operation.mjs

@@ -1,4 +1,5 @@
 import { broadAccess } from './permission.mjs';
+import { approvalRefusal } from './approval-refusal.mjs';
 import fs from 'node:fs';
 import os from 'node:os';
 import path from 'node:path';
@@ -121,7 +122,7 @@ export function fetchOriginTool(self) {
           reason:`Fetch origin branches from ${plan.repository}`,riskClass:'low',allowConsequence:'Download origin branches once; keep the current branch and working files unchanged.',denyConsequence:'No network request or credential read.'},
         terminalConfirm:()=>ctx?.ui?.confirm?.('Tải nhánh từ origin',`Tải các nhánh của ${plan.repository} về ${plan.cwd} một lần.\n\nĐồng ý: chỉ tải nhánh; nhánh hiện tại và file đang làm không đổi.\nTừ chối: không gửi request mạng, không đọc thông tin đăng nhập.`)??Promise.resolve(false),
         unavailableFallback:'terminal-confirm',recheck:()=>!signal?.aborted&&Boolean(self.grant)});
-      if(!decision.allowed||!decision.consume()||signal?.aborted)throw Error('managed-operation-denied');
+      if(!decision.allowed||!decision.consume()||signal?.aborted)throw approvalRefusal(decision);
       return executeRepositoryFetch(self.boundary,plan,signal);
     }};
 }

package/packages/piagent-core/managed/request-stream.mjs

@@ -92,7 +92,7 @@ export function wrapRoleStreams(owner, runtime, role, { provider, verifyGrant })
             // same company account and prompt cache; a helper names its role.
             const conversation = role === 'main' ? conversationSession(owner) : null;
             // Fast mode, where the member turned it on and the model serves it (fast-mode.mjs).
-            const onPayload = fastPayload(owner, role, route, options?.onPayload), fast = onPayload !== options?.onPayload;
+            const onPayload = fastPayload(owner, route, options?.onPayload), fast = onPayload !== options?.onPayload;
             // Studio's id for this request, from its answer's headers: a
             // stream that ends here without an answer is ended there too.
             let studioRequest = null;
@@ -118,7 +118,7 @@ export function wrapRoleStreams(owner, runtime, role, { provider, verifyGrant })
               out.end(told(result)); return;
             }
             // Refused with the fast tier: asked again at standard speed.
-            if (fast && fastRefusal(failureCode(refusal.errorMessage))) { letGo(); declineFast(owner, failureCode(refusal.errorMessage)); continue; }
+            if (fast && fastRefusal(failureCode(refusal.errorMessage))) { letGo(); declineFast(owner, failureCode(refusal.errorMessage), role, route); continue; }
             refused = String(refusal.errorMessage ?? '');
             // A lapsed or replaced grant: the next ask renews it, not this share.
             const stale = share; letGo();

package/packages/piagent-core/managed/session.mjs

@@ -482,10 +482,10 @@ export class ManagedSession {
   publishHelpers() {
     // One of each enabled helper role can run at a time.
     const maximum = helperRoles(this.manifest).length;
-    this.session.sessionManager.appendCustomEntry('agent-watch-helpers', {active:this.helpers.size,maximum});
+    this.session.sessionManager.appendCustomEntry('agent-watch-helpers', {active:this.helpers.size,maximum,roles:[...this.helpers.keys()]});
     // A helper runs while the main agent is quiet: listeners (the WebUI's
     // "1/4 running") are told, since no agent event would say so.
-    this.notify({ type: 'managed_helpers', active: this.helpers.size, maximum });
+    this.notify({ type: 'managed_helpers', active: this.helpers.size, maximum, roles: [...this.helpers.keys()] });
   }
   notify(event) {
     for (const listener of [...this.listeners]) { try { listener(event); } catch { /* a listener's failure is its own */ } }

package/packages/piagent-core/managed/tool-boundary.mjs

@@ -112,7 +112,7 @@ ${proxyPort && !network ? `(allow mach-lookup (global-name "com.apple.trustd.age
 (allow network-outbound (literal ${literal(docker)}))` : ''}` : loopback ? '(allow network-outbound (remote ip "localhost:*"))' : proxyPort ? `(allow network-outbound (remote ip "localhost:${Number(proxyPort)}"))` : ''}
 ${denied.map(root => `(deny file-read* file-write* (require-all (subpath ${literal(root)})${skills.filter(dir => inside(dir, root) && dir !== root).map(dir => ` (require-not (subpath ${literal(dir)}))`).join('')}))`).join('\n')}
 (deny file-read* file-read-metadata file-write* (require-all (require-any (regex #"(^|/)(\\.env([./]|$)|credentials([./]|$)|\\.npmrc$|\\.netrc$)") (regex #"/\\.[^/]+/auth\\.json$") ${projectRoots.map(root => `(literal ${literal(path.join(root, 'auth.json'))})`).join(' ')})
-  (require-not (regex #"(^|/)\\.env\\.(example|sample|template|dist|defaults)$"))))
+  (require-not (regex #"(^|/)\\.env\\.(example|sample|template|dist|defaults)$"))))  (allow file-read* file-read-metadata file-write* (require-all (subpath ${literal(temporary)}) (regex #"/\\.env([.][^/]*)?$")))
 `;
 }
 // Toolchains from the pinned node/git and Homebrew, compilers and SDK; the

package/packages/piagent-core/managed/tool-failures.mjs

@@ -15,7 +15,7 @@ const INPUT_CHARS = 500, ERROR_HEAD = 1000, ERROR_TAIL = 2000;
 // One request carries at most BATCH failures; KEEP wait for the next one.
 const BATCH = 20, KEEP = 100, ATTEMPTS = 3;
 // The member stopped the tool or declined to approve it: not a failure.
-const NOT_FAILURES = new Set(['managed_tool_cancelled', 'managed_helper_cancelled', 'managed_operation_denied']);
+const NOT_FAILURES = new Set(['managed_tool_cancelled', 'managed_helper_cancelled', 'managed_operation_denied', 'managed_operation_expired', 'managed_operation_cancelled']);
 
 const text = result => typeof result === 'string' ? result
   : Array.isArray(result?.content) ? result.content.filter(part => part?.type === 'text').map(part => part.text ?? '').join('\n') : '';

package/packages/piagent-core/managed/workflow.mjs

@@ -4,7 +4,8 @@ import { TsDiagnostics } from './ts-diagnostics.mjs';
 import { checkpointsOf } from './checkpoints.mjs';
 import { declaredChecks, detectedChecks, repositoryChecks, isCheckCommand, isReadOnlyCommand, nonCodePath, onlyNonCode } from './repository-checks.mjs';
 export { declaredChecks, detectedChecks, repositoryChecks, isCheckCommand, isReadOnlyCommand, nonCodePath, onlyNonCode };
-import { askCheckScope, asksFullChecks, beforeCheckOrPush, checkScope } from './check-questions.mjs';
+import { askCheckScope, asksFullChecks, beforeCheckOrPush, checkScope, suitePart } from './check-questions.mjs';
+import { recordCheck, relevantChecks } from './check-evidence.mjs';
 import { LANE_LIMITS, patchStats, changeLane, lanePolicy } from './risk-lanes.mjs';
 import { briefCounts, decisionsText, ledgerGuard, ledgerOf, ledgerWords, reviewFocus, saveLedger, setStatus, settleLedger } from './ledger.mjs';
 
@@ -81,7 +82,7 @@ export function workflowPrompt(policy, checks) {
     const L = LANE_LIMITS;
     lines.push(`- Size: the process follows the size of your change. A small change (at most ${L.tinyFiles} files and ${L.tinyLines} changed lines, no migration, auth, security, CI, dependency or AGENTS.md file) ends after its checks, without a review: do not ask for one. A risky change (a complex request, one of those files, more than ${L.riskyFiles} files or ${L.riskyLines} lines) must pass the checks${policy.review !== 'off' ? ' and the harness review' : ''} even where they are only suggested.`);
   }
-  return lines.length ? `\n\nProcess for code changes (company Harness):\n${lines.join('\n')}\n${process.platform === 'darwin' ? 'End-to-end and UI checks run in bash with the project\'s Playwright (@playwright/test; add it with npm when the project lacks it) and the browsers installed on this Mac, against a server the test or start_process starts; while a local proxy listens they need run_with_network. When Playwright\'s browsers are missing, ask with run_on_host to run `npx playwright install chromium`.' : 'End-to-end tests that start a local server and a browser run through run_with_network, which the user approves.'} If a check cannot run because the machine lacks something (a runtime such as a JDK, Docker, a database or another service, the network), do not stop there: request the approval that makes it run (run_with_network, run_with_docker, start_docker, or run_on_host to install what is missing) or ask the member with ask_user; never install or start anything without that approval.\nA command stops after 10 minutes and anything left running in the background stops when the turn ends: run the checks that cover your change first and a long suite in parts, not in the background.${process.platform === 'darwin' ? ' A dev server, watcher or local service you need across turns runs with start_process (read it with process_output, stop it with stop_process).' : ''}` : '';
+  return lines.length ? `\n\nProcess for code changes (company Harness):\n${lines.join('\n')}\n${process.platform === 'darwin' ? 'A bash command stops after 10 minutes unless you pass timeout in seconds, at most 1800: a long suite needs that, not another tool. End-to-end and UI checks run in bash with the project\'s Playwright (@playwright/test; add it with npm when the project lacks it) and the browsers installed on this Mac, against a server the test or start_process starts; while a local proxy listens they need run_with_network. When Playwright\'s browsers are missing, ask with run_on_host to run `npx playwright install chromium`.' : 'End-to-end tests that start a local server and a browser run through run_with_network, which the user approves.'} If a check cannot run because the machine lacks something (a runtime such as a JDK, Docker, a database or another service, the network), do not stop there: request the approval that makes it run (run_with_network, run_with_docker, start_docker, or run_on_host to install what is missing) or ask the member with ask_user; never install or start anything without that approval.\nA command stops after 10 minutes and anything left running in the background stops when the turn ends: run the checks that cover your change first and a long suite in parts, not in the background.${process.platform === 'darwin' ? ' A dev server, watcher or local service you need across turns runs with start_process (read it with process_output, stop it with stop_process).' : ''}` : '';
 }
 
 // Every helper checks the brief it was given: it may object with evidence, it
@@ -173,7 +174,7 @@ export class RunProcess {
       // asked for, and where the harness last sent the agent back (to read
       // its answer and see whether the code moved).
       ledger: [], disputes: [], verifyFails: [], claimsAsked: false, claimsAt: null, claimsRechecked: false, claimsDone: false, mainReviews: 0, minorFindings: 0,
-      sentAt: null, sentDigest: null, rejudged: false, reviewSettled: false, decideAsked: false,
+      sentAt: null, sentDigest: null, rejudged: false, reviewSettled: false,
       // How much of the suite the member wants run (check-questions.mjs).
       scopeChoice: null, scopeWords: null, fullAsked: asksFullChecks(request), pushAllowedOn: undefined,
       // The patch's risk lane (workflow.lanes) and the main agent's own web lookups.
@@ -197,15 +198,17 @@ export class RunProcess {
   changedWithoutPlan() {
     this.unplannedChange = true; this.planSkipped = true;
   }
-  async afterTool(name, args, ok, digestOf) {
+  // `owner`: the conversation the command ran in, for what passed in which repository.
+  async afterTool(name, args, ok, digestOf, owner = null) {
     if (['write', 'edit'].includes(name) && ok) this.edited = true;
     if (name === 'bash' && isCheckCommand(args?.command, this.checks.commands)) {
       const scope = checkScope(args.command, this.checks.commands), digest = await digestOf(), last = this.lastCheck;
       this.checksRun += 1; if (!ok) this.checksFailed += 1;
       // A lint after a partial test run does not make it the whole suite.
       this.lastCheck = { ok, digest, mutations: this.mutations, scope: scope ?? (last?.scope === 'partial' && last.digest === digest ? 'partial' : 'other') };
-      // The member's "run it all" covers one run; after a fix they are asked again.
-      if (scope === 'suite' && this.scopeChoice === 'full') this.scopeChoice = null;
+      await recordCheck(this, owner, args.command, scope, ok, suitePart(args.command));
+      // The member's "run it all" covers one run in each repository a push needs; after a failure they are asked again.
+      if (scope === 'suite' && this.scopeChoice === 'full' && (!ok || !this.fullFor?.size)) { this.scopeChoice = null; this.fullFor = null; }
     // A command may change files, so it counts. Without a git repository to
     // compare the code against, one that only reads does not: counting every
     // command reported code changes in turns that only read and checked.
@@ -298,7 +301,7 @@ export async function completionGate(managed, run, signal) {
       }
       if (policy.verify === 'require' && (!run.verified || run.scopeWords) && !run.docsOnly && !verifyGaveUp && run.waivedDigest !== digest) {
         run.fixLoops += 1;
-        const list = run.checks.commands.length ? run.checks.commands.map(c => '`' + c + '`').join(', ') : 'the tests, type check or build that cover the change';
+        const list = run.checks.commands.length ? relevantChecks(managed, run.checks.commands).map(c => '`' + c + '`').join(', ') : 'the tests, type check or build that cover the change';
         const words = run.scopeWords; run.scopeWords = null;
         const ask = words ? `The member answered about testing: "${String(words).slice(0, 500)}". Follow that and fix what fails.`
           : run.scopeChoice === 'partial' ? 'Run only the tests that cover your change (one file, a filter, one package), not the whole suite, and fix what fails.'
@@ -314,6 +317,9 @@ export async function completionGate(managed, run, signal) {
         const focus = run.sentAt !== null ? await reviewFocus(managed) : null, previous = (review?.findings ?? []).filter(f => f.severity === 'blocking').map((f, i) => `${i + 1}. ${findingText(f)}`).join('\n');
         const brief = `Review the current patch against the member's request.\n\nChecks: ${run.partOnlyOn(digest) ? 'only part of the test suite passed on this code' : run.verified ? 'a repository check passed on this code' : 'no check has passed on this code'}.${answer ? `\n\nThe main agent's answer to your previous blocking findings (the code changed after it):\n${answer}` : ''}${focus ? `\n\nSince your last review only these files changed: ${focus.files.join(', ')}. Judge whether each previous blocking finding is fixed (repeat it when it is not) and review the changes in these files; the rest of the patch was reviewed and has not changed.${previous ? `\nYour previous blocking findings:\n${previous}` : ''}` : ''}`;
         managed.reviewFocus = focus;
+        // The answer so far is not the result yet: the dashboard marks it so.
+        // An entry, not a message: the model has nothing to do with it.
+        session.sessionManager.appendCustomEntry('agent-watch-review-start', {});
         try { await managed.delegate({ role: 'review', task: brief }, signal, { harness: true }); }
         catch (error) {
           run.reviewUnavailable = !stopped(); run.reviewUnavailableReason = reviewUnavailableReason(error?.message); interrupted = stopped();
@@ -443,7 +449,7 @@ export function processEditTools(session, tools) {
         }
         return result;
       }
-      finally { await session.run?.afterTool(tool.name, args[1], ok, () => session.digest()); await session.refreshReview(); }
+      finally { await session.run?.afterTool(tool.name, args[1], ok, () => session.digest(), session); await session.refreshReview(); }
     } };
   });
 }

package/packages/piagent-core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@piagent/core",
-  "version": "1.49.0",
+  "version": "1.50.0",
   "type": "module",
   "private": true,
   "license": "MIT",

package/packages/piagent-core/runtime/inspection/approval-broker.ts

@@ -8,7 +8,7 @@ const PUBLIC_REF = /^[A-Za-z0-9][A-Za-z0-9._:@~-]{0,159}$/;
 const REVISION = /^[A-Za-z0-9][A-Za-z0-9._:~-]{0,159}$/;
 const DIGEST = /^sha256:[a-f0-9]{64}$/;
 const TIMESTAMP = /^\d{4}-\d{2}-\d{2}T\d{2}:\d{2}:\d{2}\.\d{3}Z$/;
-const MAX_PENDING = 32, MAX_RECENT = 64, DEFAULT_TTL_MS = 5 * 60_000;
+const MAX_PENDING = 32, MAX_RECENT = 64, DEFAULT_TTL_MS = 5 * 60_000, MAX_TTL_MS = 30 * 60_000;
 
 type Identity = { projectRef: string; runtimeInstanceId: string; sessionRef: string; taskId: string; taskRunId: string;
   agentOperationId: string; toolCallId: string };
@@ -161,7 +161,7 @@ export class PiApprovalBroker {
     if (!binding || !authority || authority.taskState !== "active" || input.expectedTask && (authority.identity.taskId !== input.expectedTask.taskId || authority.identity.taskRunId !== input.expectedTask.taskRunId)
       || !REF.test(authority.identity.agentOperationId)
       || !toolCallId || binding.pending.size >= MAX_PENDING) return this.#unavailable(input);
-    const now = Date.now(), requestedAt = new Date(now).toISOString(), expiresAt = new Date(now + Math.min(Math.max(input.ttlMs ?? DEFAULT_TTL_MS, 5_000), DEFAULT_TTL_MS)).toISOString();
+    const now = Date.now(), requestedAt = new Date(now).toISOString(), expiresAt = new Date(now + Math.min(Math.max(input.ttlMs ?? DEFAULT_TTL_MS, 5_000), MAX_TTL_MS)).toISOString();
     const approvalRef = ref("approval", binding.secret, [toolCallId, now, randomUUID()]);
     const decisionToken = randomBytes(32).toString("base64url");
     const nextRevision = this.#nextRevision(binding, "requested", approvalRef);

package/packages/piagent-webui/dist/client/assets/ActivityPanel-CmjxPmNT.js

@@ -0,0 +1 @@
+import{n as e,r as t}from"./rolldown-runtime-hePW80VL.js";import{An as n,Rn as r}from"./vendor-DNZbaWhT.js";import{i,n as a}from"./ui-preferences-OfS30-ZV.js";import{B as o,K as s,x as c}from"./api-BQS4Dtrr.js";import{c as l}from"./view-model-CWNiS4JN.js";var u=t(r(),1);function d(e,t=`vi`){let n=e.state===`running`?`running`:e.state===`failed`?`failed`:`passed`,r=n===`running`?t===`vi`?`đang chạy`:`running`:n===`failed`?t===`vi`?`thất bại`:`failed`:t===`vi`?`hoàn tất`:`completed`,i=e.startedAt??e.finishedAt??``;return{activityRef:`live.${e.toolCallRef}`,kind:[`bash`,`shell`,`exec`].includes(e.toolLabel)?`command`:`tool`,state:n,label:`${e.agent?`${e.agent} · `:``}${e.toolLabel} ${r}`,preview:e.detail??e.fileLabel??(t===`vi`?`Cập nhật trực tiếp từ Gateway`:`Live update from the Gateway`),toolCallId:e.toolCallRef,toolName:e.toolLabel,commandDigest:null,logRef:null,exitCode:null,exitCodeExact:!1,startedAt:i,finishedAt:n===`running`?null:e.finishedAt??i}}function f(e,t=`vi`){let n=e.state===`failed`?t===`vi`?`Lượt chạy gặp lỗi`:`Operation failed`:e.state===`blocked`?t===`vi`?`Lượt chạy bị chặn`:`Operation blocked`:e.state===`aborted`?t===`vi`?`Lượt chạy đã dừng`:`Operation stopped`:t===`vi`?`Chưa xác định kết quả lượt chạy`:`Operation outcome unknown`;return{activityRef:e.activityRef,kind:`system`,state:e.state,label:n,preview:e.reasonCode,toolCallId:null,toolName:null,commandDigest:null,logRef:null,exitCode:null,exitCodeExact:!1,startedAt:e.settledAt,finishedAt:e.settledAt}}function p(e,t,n,r=`vi`,i){let a=[...e,...t],o=new Set(a.map(e=>e.activityRef)),s=n.filter(e=>!o.has(e.activityRef)).map(e=>f(e,r));if(i===void 0)return{rows:[...e,...s,...t],terminalCount:s.length,runningCount:e.length};let c=i.filter(e=>e.state===`running`).map(e=>d(e,r));return{rows:[...c,...s,...t],terminalCount:s.length,runningCount:c.length}}function m(e,t=`vi`){return e.state===`passed`&&/ recovered$/i.test(e.label)?t===`vi`?`Đã khôi phục`:`Recovered`:e.state===`passed`&&/ warning$/i.test(e.label)?t===`vi`?`Đã xử lý cảnh báo`:`Handled warning`:e.state===`passed`&&e.exitCodeExact?`Pass · exit ${e.exitCode??0}`:e.state===`failed`&&e.exitCodeExact?`Fail · exit ${e.exitCode??`?`}`:e.state===`blocked`?`Blocked`:e.state===`aborted`?t===`vi`?`Đã dừng`:`Stopped`:e.state===`unknown`?t===`vi`?`Chưa xác định`:`Unknown`:e.state===`failed`?t===`vi`?`Thất bại`:`Failed`:e.state===`running`?t===`vi`?`Đang chạy`:`Running`:e.state.replaceAll(`-`,` `)}function h(e,t=`vi`){let n=Date.parse(e.startedAt),r=e.finishedAt?Date.parse(e.finishedAt):Date.now();if(!Number.isFinite(n)||!Number.isFinite(r)||r<n)return t===`vi`?`Thời gian chưa xác định`:`Unknown duration`;let i=Math.round((r-n)/1e3);return i<60?`${i}s`:`${Math.floor(i/60)}m ${i%60}s`}function g(e,t=[]){let n=new Map([[`main`,{agent:`main`,steps:[],total:0}]]),r=e=>{let t=n.get(e);if(t)return t;let r={agent:e,steps:[],total:0};return n.set(e,r),r};for(let t of e){if(t.helperRun){let e=r(t.helperRun.agent);e.total+=t.helperRun.total,e.steps.push(...t.helperRun.steps.map((e,n)=>({key:`${t.messageRef}:${n}`,toolName:e.toolName,target:e.target,state:`completed`,at:e.at})));continue}if(t.role!==`assistant`)continue;let e=r(`main`);for(let n of t.toolCalls)e.total+=1,e.steps.push({key:`${t.messageRef}:${n.toolCallRef}`,toolName:n.toolName,target:n.summary?.target??n.summary?.detail??null,state:n.state===`failed`||n.result?.isError?`failed`:`completed`,at:t.recordedAt})}for(let e of t){if(!e.agent||e.state===`running`)continue;let t=r(e.agent);t.total+=1,t.steps.push({key:e.toolCallRef,toolName:e.toolLabel,target:e.detail??null,state:e.state,at:e.startedAt??``})}for(let e of n.values())e.steps.sort((e,t)=>t.at.localeCompare(e.at));return[...n.values()].filter(e=>e.steps.length>0)}var _=n();function v(e,t){return e===`main`?a(t,`Main agent`,`Main agent`):`Subagent ${e}`}function y(e,t){let n=new Date(e);return Number.isFinite(n.getTime())?n.toLocaleTimeString(t===`vi`?`vi-VN`:`en-US`,{hour:`2-digit`,minute:`2-digit`,second:`2-digit`}):``}var b=5e3;function x({sessionRef:e,running:t,live:n,locale:r,empty:i}){let[o,c]=(0,u.useState)([]);(0,u.useEffect)(()=>{let n=new AbortController,r=()=>void s(e,null,200,n.signal).then(e=>c(e.items)).catch(()=>void 0);r();let i=t?window.setInterval(r,b):void 0;return()=>{n.abort(),i&&window.clearInterval(i)}},[e,t]);let d=(0,u.useMemo)(()=>g(o,n),[o,n]);return d.length?(0,_.jsx)(`div`,{className:`agent-activity-log`,children:d.map(e=>(0,_.jsxs)(`details`,{open:e.agent===`main`||t,className:`agent-activity-group`,children:[(0,_.jsxs)(`summary`,{children:[(0,_.jsx)(`strong`,{children:v(e.agent,r)}),` `,(0,_.jsxs)(`span`,{children:[e.total,` `,a(r,`lần gọi`,`calls`)]}),e.total>e.steps.length&&(0,_.jsxs)(`span`,{children:[` · `,a(r,`hiện ${e.steps.length} gần nhất`,`showing the latest ${e.steps.length}`)]})]}),e.steps.map(e=>(0,_.jsxs)(`article`,{className:`activity-row`,children:[(0,_.jsx)(`span`,{className:`activity-state tone-${l(e.state)}`,"aria-label":e.state}),(0,_.jsxs)(`div`,{className:`activity-copy`,children:[(0,_.jsx)(`div`,{children:(0,_.jsx)(`strong`,{children:e.toolName})}),e.target&&(0,_.jsx)(`p`,{children:e.target})]}),(0,_.jsx)(`div`,{className:`activity-result`,children:(0,_.jsx)(`span`,{children:y(e.at,r)})})]},e.key))]},e.agent))}):i?(0,_.jsx)(`div`,{className:`activity-empty`,children:i}):null}var S=e({ActivityPanel:()=>w});function C({activity:e,preview:t,onPreview:n,locale:r}){return(0,_.jsxs)(`article`,{className:`activity-row`,children:[(0,_.jsx)(`span`,{className:`activity-state tone-${l(e.state)}`,"aria-label":e.state}),(0,_.jsxs)(`div`,{className:`activity-copy`,children:[(0,_.jsxs)(`div`,{children:[(0,_.jsx)(`strong`,{children:e.label}),(0,_.jsx)(`span`,{children:e.kind})]}),(0,_.jsx)(`p`,{children:e.preview||a(r,`Không có mô tả ngắn`,`No short description`)}),t&&(0,_.jsxs)(`div`,{className:`log-preview ${t.state===`error`?`error-state`:``}`,children:[t.state===`loading`?a(r,`Đang tải log ngắn…`,`Loading log preview…`):t.state===`error`?a(r,`Log preview không còn sẵn sàng.`,`Log preview is no longer available.`):t.value?.preview??`${a(r,`Không có nội dung log`,`No log content`)} · ${t.value?.reasonCode??`unavailable`}`,t.value?.truncated&&(0,_.jsx)(`small`,{children:a(r,`Preview đã được rút gọn`,`Preview truncated`)})]})]}),(0,_.jsxs)(`div`,{className:`activity-result`,children:[(0,_.jsx)(`strong`,{children:m(e,r)}),(0,_.jsx)(`span`,{children:h(e,r)}),e.logRef&&n&&(0,_.jsx)(`button`,{type:`button`,onClick:n,"aria-expanded":!!t,children:t?a(r,`Ẩn log`,`Hide log`):a(r,`Xem log`,`View log`)})]})]})}function w({snapshot:e,sessionRef:t,terminalActivities:n=[],liveActivities:r}){let{locale:s}=i(),[l,d]=(0,u.useState)({}),f=e?.activity.running??[],m=e?.activity.recent??[],{rows:h,terminalCount:g,runningCount:v}=p(f,m,n,s,r),y=e=>{if(l[e.activityRef]){d(t=>({...t,[e.activityRef]:void 0}));return}d(t=>({...t,[e.activityRef]:{state:`loading`}})),(t?o(t,e.activityRef):c(e.activityRef)).then(t=>d(n=>({...n,[e.activityRef]:{state:`ready`,value:t}}))).catch(()=>d(t=>({...t,[e.activityRef]:{state:`error`}})))};return(0,_.jsxs)(`section`,{className:`activity-panel surface`,"aria-labelledby":`activity-title`,children:[(0,_.jsxs)(`header`,{className:`panel-heading`,children:[(0,_.jsxs)(`div`,{children:[(0,_.jsx)(`p`,{className:`section-kicker`,children:`Activity`}),(0,_.jsx)(`h2`,{id:`activity-title`,children:a(s,`Tool, command và kết quả lượt chạy`,`Tools, commands, and operation outcomes`)})]}),(0,_.jsxs)(`div`,{children:[(0,_.jsxs)(`span`,{className:`running-count`,children:[v,` `,a(s,`đang chạy`,`running`)]}),(0,_.jsxs)(`span`,{children:[m.length+g,` `,a(s,`gần đây`,`recent`)]})]})]}),(0,_.jsxs)(`div`,{className:`activity-list`,children:[h.map(e=>(0,_.jsx)(C,{activity:e,preview:l[e.activityRef],onPreview:e.logRef?()=>y(e):void 0,locale:s},e.activityRef)),t&&(0,_.jsx)(x,{sessionRef:t,running:!!r,live:r,locale:s,empty:h.length===0?a(s,`Chưa có tool, command hoặc kết quả lượt chạy nào trong timeline hiện tại.`,`No tools, commands, or operation outcomes in the current timeline.`):void 0}),h.length===0&&!t&&(0,_.jsx)(`div`,{className:`activity-empty`,children:a(s,`Chưa có tool, command hoặc kết quả lượt chạy nào trong timeline hiện tại.`,`No tools, commands, or operation outcomes in the current timeline.`)})]}),e?.activity.page.truncated&&(0,_.jsx)(`footer`,{className:`panel-footnote`,children:a(s,`Timeline đã được rút gọn theo giới hạn local.`,`The timeline was truncated to local limits.`)})]})}export{S as n,w as t};