@piagent/platform 1.48.4 → 1.49.0

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

@@ -1,10 +1,10 @@
-import fs from 'node:fs';
-import path from 'node:path';
 import { reviewUnavailableReason } from './patch-snapshot.mjs';
 import { failureCode, unnamedFailureText } from '../runtime/managed-failure.mjs';
 import { TsDiagnostics } from './ts-diagnostics.mjs';
 import { checkpointsOf } from './checkpoints.mjs';
-import { discoverDeliveryContext } from './delivery-context.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 { LANE_LIMITS, patchStats, changeLane, lanePolicy } from './risk-lanes.mjs';
 import { briefCounts, decisionsText, ledgerGuard, ledgerOf, ledgerWords, reviewFocus, saveLedger, setStatus, settleLedger } from './ledger.mjs';
 
@@ -22,83 +22,6 @@ export function workflowPolicy(manifest) {
   return { plan: mode(w.plan), verify: mode(w.verify), review: config.review ? mode(w.review) : 'off', maxFixLoops: loops, lanes: w.lanes !== false };
 }
 
-// Checks come from the repository: a "## Checks" section in AGENTS.md (code
-// block lines or `command` bullets), else what its manifests imply.
-const HEADING = /^#{1,6}\s+(checks?|verification|kiểm tra)\s*$/i;
-export function declaredChecks(agentsFiles = []) {
-  const out = [];
-  for (const file of agentsFiles) {
-    let inside = false, fence = false;
-    for (const line of String(file.content ?? '').split('\n')) {
-      if (!fence && /^#{1,6}\s/.test(line)) { inside = HEADING.test(line.trim()); continue; }
-      if (!inside) continue;
-      if (/^\s*(```|~~~)/.test(line)) { fence = !fence; continue; }
-      const command = (fence ? line : line.match(/^\s*[-*]\s+`([^`]+)`/)?.[1] ?? '').trim();
-      if (command && !command.startsWith('#') && command.length <= 200 && !out.includes(command) && out.length < 10) out.push(command);
-    }
-  }
-  return out;
-}
-export function detectedChecks(dir) {
-  const has = name => { try { const s = fs.lstatSync(path.join(dir, name)); return s.isFile() && s.size <= 1024 * 1024; } catch { return false; } };
-  const out = [];
-  if (has('package.json')) {
-    try {
-      const scripts = JSON.parse(fs.readFileSync(path.join(dir, 'package.json'), 'utf8')).scripts ?? {};
-      const runner = has('pnpm-lock.yaml') ? 'pnpm' : has('yarn.lock') ? 'yarn' : has('bun.lock') || has('bun.lockb') ? 'bun' : 'npm';
-      for (const name of ['typecheck', 'type-check', 'lint', 'test', 'test:unit', 'test:integration', 'test:api', 'test:contract', 'test:e2e']) if (typeof scripts[name] === 'string') out.push(name === 'test' && runner === 'npm' ? 'npm test' : `${runner} run ${name}`);
-    } catch { /* not JSON: nothing implied */ }
-  }
-  if (has('go.mod')) out.push('go vet ./...', 'go test ./...');
-  if (has('Package.swift')) out.push('swift build', 'swift test');
-  if (has('Cargo.toml')) out.push('cargo check', 'cargo test');
-  if (has('pyproject.toml') || has('pytest.ini')) out.push(has('uv.lock') ? 'uv run pytest' : 'pytest');
-  if (has('pom.xml')) out.push(has('mvnw') ? './mvnw verify' : 'mvn verify');
-  if (has('build.gradle') || has('build.gradle.kts')) out.push(has('gradlew') ? './gradlew test' : 'gradle test');
-  return out.slice(0, 16);
-}
-export function repositoryChecks(agentsFiles, cwd, top = cwd) {
-  const declared = declaredChecks(agentsFiles);
-  if (declared.length) return { source: 'agents', commands: declared };
-  const detected = detectedChecks(cwd);
-  // Include package/repository-local checks even when the workspace root has a
-  // lint command. These are candidates, not a request to test unrelated repos.
-  const quote = value => "'" + value.replaceAll("'", "'\\''") + "'";
-  const nested = discoverDeliveryContext(cwd).units.filter(unit => unit.path !== '.').flatMap(unit =>
-    detectedChecks(path.join(cwd, unit.path)).map(command => `cd ${quote(unit.path)} && ${command}`));
-  const commands = [...new Set([...detected, ...nested])].slice(0, 64);
-  return { source: 'detected', commands: commands.length || cwd === top ? commands : detectedChecks(top) };
-}
-// A command that verifies code: a declared check, or a known test, type,
-// lint or build runner. Text tools that merely mention one do not count.
-const RUNNER = /\b(npm|pnpm|yarn|bun)\s+(run\s+)?(test|lint|typecheck|type-check|check|build|verify)\b|\bnode\s+--test\b|\b(tsc|vitest|jest|eslint|biome|ruff|mypy|pyright|pytest|rspec|phpunit)\b|\bplaywright\s+test\b|\bpython3?\s+-m\s+(pytest|unittest|mypy)\b|\bgo\s+(test|vet|build)\b|\bcargo\s+(test|check|clippy|build)\b|\bswift\s+(test|build)\b|\bxcodebuild\b.*\b(test|build)\b|\bmake\s+(test|check|lint)\b|\b(gradle|gradlew|mvn|mvnw)\b.*\b(test|check|verify|compile|package|build)\b|\bdotnet\s+(test|build)\b|\bdeno\s+(test|check|lint)\b/;
-// Files no check covers: documentation, pictures and documents,
-// and Spec Kit's .specify/ state. A turn that changed only these changed no
-// code: no check is required of it (one is still reported if it ran).
-const NON_CODE = /\.(md|mdx|markdown|rst|adoc|asciidoc|org|png|jpe?g|gif|webp|bmp|ico|pdf|docx?|xlsx?|pptx?|odt|ods|odp|rtf)$/i;
-export function nonCodePath(file) {
-  const name = String(file ?? '').replaceAll('\\', '/');
-  return !!name && (NON_CODE.test(name) || /(^|\/)\.specify\//.test(name) || /(^|\/)(LICENSE|NOTICE|AUTHORS|CODEOWNERS)$/.test(name));
-}
-export const onlyNonCode = paths => { const list = [...(paths ?? [])]; return list.length > 0 && list.every(nonCodePath); };
-
-// A command that only reads: each part starts with a reading tool and nothing
-// is written to a file. Anything else may change files. Git may name the
-// repository first (git -C <folder>): a workspace folder holding several
-// repositories is read that way.
-const GIT_READ = String.raw`git(\s+(-C\s+('[^']*'|"[^"]*"|\S+)|-c\s+\S+|--no-pager))*\s+(status|diff|log|show|branch|ls-files|rev-parse|rev-list|blame|describe|shortlog|cat-file|for-each-ref|reflog(?!\s+(expire|delete))|remote(\s+(-v|--verbose|show|get-url)\b|$))`;
-const READER = new RegExp(String.raw`^(ls|cat|head|tail|wc|grep|egrep|rg|find|fd|sed|awk|cut|sort|uniq|tr|echo|printf|pwd|which|stat|file|du|df|tree|jq|diff|cmp|basename|dirname|realpath|date|true|cd|${GIT_READ})(\b|$)`);
-export function isReadOnlyCommand(command) {
-  const text = String(command ?? '').trim();
-  if (!text || />(?!>|\s*&\d|\s*\/dev\/null)|\btee\b|\bsed\s+(-[a-zA-Z]*i|--in-place)|\s-(delete|exec|execdir|fprint|ok)\b|\bxargs\b|\$\(|`/.test(text)) return false;
-  return text.split(/&&|\|\||;|\||\n/).every(part => !part.trim() || READER.test(part.trim()));
-}
-export function isCheckCommand(command, declared = []) {
-  const text = String(command ?? '').replace(/\s+/g, ' ').trim();
-  if (!text || /^(echo|printf|grep|rg|cat|sed|awk|head|tail|less|git|ls|find)\b/.test(text)) return false;
-  return declared.some(d => text.includes(d)) || RUNNER.test(text);
-}
-
 // update_plan: the agent's checklist, kept as a session entry so it survives
 // compaction and "continue", and shown to the member beside the conversation.
 export const PLAN_ENTRY = 'agent-watch-plan';
@@ -148,7 +71,7 @@ export function workflowPrompt(policy, checks) {
   if (policy.plan !== 'off') lines.push(`- Plan: for work with several steps keep a checklist with update_plan and update it as you go. When the member asks for several items (plans, files, tickets), put every item in the checklist and work through all of them; do not end your turn to report between items.${policy.plan === 'require' ? ' For a complex task, file edits are refused until a plan exists. A checklist left with pending steps sends you back to the next one until every step is completed or deferred: if you cannot go on without the member, ask with ask_user. Mark a step deferred only when the member asked to leave it for later or it is no longer needed; never mark completed a step you did not do.' : ''}`);
   if (policy.verify !== 'off') {
     const list = checks.commands.length ? `${checks.commands.map(c => '`' + c + '`').join(', ')}${checks.source === 'detected' ? ' (detected; the repository declares none in AGENTS.md "## Checks")' : ''}` : 'none declared: choose the tests, type check or build that cover the change';
-    lines.push(`- Verify: after changing code, run the repository checks that cover the change and fix failures before you finish. Repository checks: ${list}.${policy.verify === 'require' ? ' The turn does not end until a check has passed on the final code.' : ''} Only the member may skip the checks: when they ask to, call waive_checks with their words; never skip them on your own.`);
+    lines.push(`- Verify: after changing code, run the tests that cover the change (one file, a filter, one package) and fix failures before you finish; after a failure rerun the part that failed. Do not choose the whole test suite yourself: it can take long, so the harness asks the member before a full run and before a push. Repository checks: ${list}.${policy.verify === 'require' ? ' The turn does not end until a check has passed on the final code.' : ''} Only the member may skip the checks: when they ask to, call waive_checks with their words; never skip them on your own.`);
   }
   // What each severity asks: blocking is fixed or argued, major decided, minor only listed.
   const severities = ' Blocking findings: fix them, or reject one with resolve_objection and the reviewer judges your reason; a disagreement that remains goes to the member. Major findings: fix them or reject them with resolve_objection; no new review follows. Minor findings are listed for the member and need nothing.';
@@ -251,6 +174,8 @@ export class RunProcess {
       // 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,
+      // 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.
       lane: null, laneStats: null, laneDigest: null, webLookups: 0 });
   }
@@ -273,18 +198,27 @@ export class RunProcess {
     this.unplannedChange = true; this.planSkipped = true;
   }
   async afterTool(name, args, ok, digestOf) {
+    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;
-      this.lastCheck = { ok, digest: await digestOf(), mutations: this.mutations };
+      // 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;
     // 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.
     } else if (['write', 'edit'].includes(name) && ok || name === 'bash' && (this.startDigest !== null || !isReadOnlyCommand(args?.command))) this.mutations += 1;
   }
-  verifiedOn(digest) {
+  passedOn(digest) {
     const c = this.lastCheck;
     return !!c?.ok && (digest !== null && c.digest !== null ? c.digest === digest : c.mutations === this.mutations);
   }
+  partOnlyOn(digest) { return this.passedOn(digest) && this.lastCheck.scope === 'partial'; }
+  suitePassedOn(digest) { return this.passedOn(digest) && this.lastCheck.scope === 'suite'; }
+  // A partial run counts unless the member wants the whole suite.
+  verifiedOn(digest) { return this.passedOn(digest) && !(this.partOnlyOn(digest) && (this.fullAsked || this.scopeChoice === 'full')); }
   // Version 2 (a broker advertising "process-v2") adds whether code was
   // changed without the required plan and how many calls named no tool.
   // Version 3 adds helper objections, those the agent answered and the
@@ -334,7 +268,7 @@ export async function completionGate(managed, run, signal) {
   let digest = null, interrupted = false, verifyGaveUp = false;
   for (let step = 0; step < 16; step += 1) {
     digest = await digestOf();
-    run.changed = run.startDigest !== null && digest !== null ? digest !== run.startDigest : run.mutations > 0;
+    run.changed = run.startDigest !== null && digest !== null ? digest !== run.startDigest : run.startDigest !== null ? run.edited === true : run.mutations > 0; // code that cannot be read now: only a real edit says it changed
     // A check that passed still counts when the turn was cut off after it.
     if (run.changed) run.verified = run.verifiedOn(digest);
     run.docsOnly = run.changed && onlyNonCode(managed.ownPaths);
@@ -350,10 +284,26 @@ export async function completionGate(managed, run, signal) {
       // The member declined the approval a check needed (a JDK, Docker, the
       // network): asking again would only repeat the same answer.
       if (policy.verify === 'require' && !run.verified && run.approvalDeclined && run.fixLoops > 0) verifyGaveUp = true;
-      if (policy.verify === 'require' && !run.verified && !run.docsOnly && !verifyGaveUp && run.waivedDigest !== digest) {
+      // "Do not run" covered the code it was given for; new code is asked about again.
+      if (run.scopeChoice === 'skip' && run.waivedDigest !== digest) run.scopeChoice = null;
+      // How much to test is the member's call; asked once, not while a check is failing.
+      if (policy.verify === 'require' && !run.docsOnly && !verifyGaveUp && run.waivedDigest !== digest && !run.scopeChoice && !run.fullAsked && run.lastCheck?.ok !== false && (!run.verified || run.partOnlyOn(digest))) {
+        const answer = await askCheckScope(managed, { partPassed: run.partOnlyOn(digest), signal });
+        if (stopped()) { interrupted = true; break; }
+        if (answer) {
+          run.scopeChoice = answer.choice; run.scopeWords = answer.words ?? null;
+          if (answer.choice === 'skip') run.waivedDigest = digest; // the same record as waive_checks: this exact code, reported unverified
+          run.verified = run.verifiedOn(digest);
+        }
+      }
+      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';
-        await send(`Harness process check: code changed in this turn, but no check has passed on the current code. Run the checks that cover your change (${list}) and fix what fails. If a check cannot run because the machine lacks something (a runtime such as a JDK, Docker, a database or another service, the network), request the approval that makes it run (run_with_network, run_with_docker, start_docker, run_on_host) or ask the member with ask_user, then run it; install or start nothing without that approval. If it fails for a reason unrelated to your change, say so plainly.`,
+        const words = run.scopeWords; run.scopeWords = null;
+        const ask = words ? `The member answered about testing: "${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.'
+            : `Run the checks that cover your change (${list}) and fix what fails.${run.fullAsked || run.scopeChoice === 'full' ? ' The member wants the whole suite: a partial run does not count.' : ''}`;
+        await send(`Harness process check: code changed in this turn, but no check has passed on the current code. ${ask} If a check cannot run because the machine lacks something (a runtime such as a JDK, Docker, a database or another service, the network), request the approval that makes it run (run_with_network, run_with_docker, start_docker, run_on_host) or ask the member with ask_user, then run it; install or start nothing without that approval. If it fails for a reason unrelated to your change, say so plainly.`,
           { phase: 'verify', loop: run.fixLoops, maxLoops: policy.maxFixLoops, lastCheckFailed: run.lastCheck?.ok === false });
         continue;
       }
@@ -362,7 +312,7 @@ export async function completionGate(managed, run, signal) {
         const answer = run.sentAt !== null ? answerSince(run.sentAt) + decisionsText(run, 'review') : '';
         // After a fix the reviewer checks its blocking findings and the files changed since.
         const focus = run.sentAt !== null ? await reviewFocus(managed) : null, previous = (review?.findings ?? []).filter(f => f.severity === 'blocking').map((f, i) => `${i + 1}. ${findingText(f)}`).join('\n');
-        const brief = `Review the current patch against the member's request.\n\nChecks: ${run.verified ? 'a repository check passed on this code' : 'no check has passed on this code'}.${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}` : ''}` : ''}`;
+        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;
         try { await managed.delegate({ role: 'review', task: brief }, signal, { harness: true }); }
         catch (error) {
@@ -442,8 +392,8 @@ export async function completionGate(managed, run, signal) {
   if (!run.changed && !run.disputes.length && !ledgerOf(run).length) return;
   const finalPlan = currentPlan(session.sessionManager), planOpen = run.planUpdated ? finalPlan?.plan?.filter(p => p.status === 'pending' || p.status === 'in_progress').length ?? 0 : 0;
   const status = { verified: run.verified, reviewed: run.reviewed, blockingOpen: run.blockingOpen, checks: run.checksRun, fixLoops: run.fixLoops,
-    ...(planOpen ? { planOpen } : {}), ...(run.planSkipped ? { planSkipped: true } : {}), ...(!run.verified && run.waivedDigest === digest ? { checksWaived: true } : {}), ...(run.disputes.length ? { disputes: run.disputes.length } : {}), ...(run.changed ? {} : { unchanged: true }) };
-  const words = [...(run.changed ? [run.verified ? 'a check passed on the final code' : policy.verify === 'off' ? null : run.docsOnly ? 'only documentation changed: no check needed' : run.waivedDigest === digest ? 'no check passed on the final code: the member waived the checks' : 'no check passed on the final code',
+    ...(planOpen ? { planOpen } : {}), ...(run.planSkipped ? { planSkipped: true } : {}), ...(!run.verified && run.waivedDigest === digest ? { checksWaived: true } : {}), ...(run.verified && run.partOnlyOn(digest) ? { partOnly: true } : {}), ...(run.disputes.length ? { disputes: run.disputes.length } : {}), ...(run.changed ? {} : { unchanged: true }) };
+  const words = [...(run.changed ? [run.partOnlyOn(digest) && run.verified ? 'only part of the test suite passed on the final code' : run.verified ? 'a check passed on the final code' : policy.verify === 'off' ? null : run.docsOnly ? 'only documentation changed: no check needed' : run.waivedDigest === digest ? 'no check passed on the final code: the member waived the checks' : 'no check passed on the final code',
     run.reviewed ? (run.blockingOpen ? `review: ${run.blockingOpen} blocking issue(s) still open` : 'review: no blocking issue') : policy.review === 'off' ? null : run.reviewUnavailable ? `review could not run (${run.reviewUnavailableReason})` : 'not reviewed',
     run.planSkipped ? 'code was changed before the required plan existed' : null, planOpen ? `plan: ${planOpen} step(s) not completed` : null] : []),
     run.disputes.length ? `${run.disputes.length} disagreement(s) between the main agent and a subagent go to the member` : null, ledgerWords(run)].filter(Boolean);
@@ -464,6 +414,7 @@ export function processEditTools(session, tools) {
       session.run?.beforeEdit(tool.name, plan, session.mainAnswers);
       // An open premise objection is resolved before files change (ledger.mjs).
       if (tool.name !== 'bash' || !isReadOnlyCommand(args[1]?.command)) ledgerGuard(session.run, session.mainAnswers);
+      if (tool.name === 'bash') await beforeCheckOrPush(session, args[1]?.command, args[4], args[2]);
       // While a required plan is missing, a command that changes files
       // counts as an edit: the code is compared after each command.
       const watch = tool.name === 'bash' && !!session.run?.planMissing(plan) && !session.run.unplannedChange && session.run.startDigest !== null;

package/packages/piagent-core/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@piagent/core",
-  "version": "1.48.4",
+  "version": "1.49.0",
   "type": "module",
   "private": true,
   "license": "MIT",