@galda/cli 0.10.91 → 0.10.93

package/engine/completion-contract.mjs

@@ -1,5 +1,39 @@
-const CHANGE_REQUEST_RE = /(?:修正|直して|解決して|改善して|一致させ|統一して|実装して|追加して|削除して|変更して|作って|fix\b|solve\b|resolve\b|implement\b|change\b|make\b)/i;
-const REFUSAL_RESULT_RE = /(?:直せません|修正できません|解決できません|実現できません|不可能です|cannot\s+(?:be\s+)?(?:fix(?:ed)?|solv(?:e|ed)|implement(?:ed)?|done)|can't\s+(?:fix|solve|implement|do)|unable\s+to\s+(?:fix|solve|implement))/i;
+// Which user requests are "change requests" whose ONLY acceptable outcome is the
+// change itself. Deliberately wide on Japanese request forms: the first cut only
+// matched the 「〜して」 imperative stem, so ordinary asks like 「〜してほしい」
+// 「〜できる？」「〜消せますか？」「〜にして」 never qualified and their refusals
+// sailed straight into Review as completed work (2026-07-29 audit).
+const CHANGE_REQUEST_RE = new RegExp([
+  '修正', '直して', '直せ', 'なおして', '解決', '改善', '一致させ', '統一',
+  '実装', '追加', '削除', '変更', '作って', '対応して', '消して', '消せ',
+  '出して', '表示して', '止めて', '揃え', '直したい', '外して', '戻して',
+  'できるように', 'にして', 'して(?:ほしい|下さい|ください)', 'られるように',
+  '\\bfix\\b', '\\bsolve\\b', '\\bresolve\\b', '\\bimplement\\b', '\\bchange\\b',
+  '\\bmake\\b', '\\bremove\\b', '\\badd\\b', '\\bhide\\b', '\\bshow\\b',
+].join('|'), 'i');
+
+// What counts as "the agent closed the request by declaring it impossible".
+// The earlier list missed the exact phrasings Masa reported —「仕様なので
+// 変えられません」「対応できません」— plus the English "by design / working as
+// intended" family, which is the same move in a different register.
+const REFUSAL_RESULT_RE = new RegExp([
+  '直せません', '直せない', '修正できません', '修正できない',
+  '解決できません', '解決できない', '実現できません', '実現できない',
+  '変えられません', '変えられない', '変更できません', '変更できない',
+  '対応できません', '対応できない', '削除できません', '消せません',
+  '実装できません', '実装できない', '不可能です', '仕様(?:のため|なので|上)[^。\\n]{0,20}(?:できません|変えられません|変更できません)',
+  'cannot\\s+(?:be\\s+)?(?:fix(?:ed)?|solv(?:e|ed)|implement(?:ed)?|chang(?:e|ed)|remov(?:e|ed)|done)',
+  "can't\\s+(?:fix|solve|implement|change|remove|do)",
+  'unable\\s+to\\s+(?:fix|solve|implement|change|remove)',
+  'not\\s+possible\\s+to\\s+(?:fix|change|implement|remove)',
+  'by\\s+design[^.\\n]{0,30}(?:cannot|can\'t|won\'t)',
+  'working\\s+as\\s+intended',
+].join('|'), 'i');
+
+// An explicit, resolved answer beats the refusal phrase: a worker that says
+// "直せませんでしたが、代わりに X を実装しました" DID deliver. Only the shapes
+// where the refusal is the whole outcome should be blocked.
+const RESOLVED_ANYWAY_RE = /(?:代わりに|その代わり|instead[, ]|workaround|回避策)/i;
 
 export const COMPLETION_CONTRACT_PROMPT = [
   'MANAGER COMPLETION CONTRACT:',
@@ -8,10 +42,14 @@ export const COMPLETION_CONTRACT_PROMPT = [
   '- Do not report a change request as complete with “cannot fix”, “直せません”, or equivalent. If a real external blocker remains, state the concrete blocker and leave the work explicitly unfinished instead of presenting it for Review as completed.',
 ].join('\n');
 
-export function completionContractIssue({ request = '', followUp = '', result = '', status = '' } = {}) {
+export function completionContractIssue({ request = '', followUp = '', result = '', status = '', changedFiles = [] } = {}) {
   if (status !== 'done') return null;
   const requested = `${request}\n${followUp}`;
-  if (!CHANGE_REQUEST_RE.test(requested) || !REFUSAL_RESULT_RE.test(String(result))) return null;
+  const text = String(result);
+  if (!CHANGE_REQUEST_RE.test(requested) || !REFUSAL_RESULT_RE.test(text)) return null;
+  // Evidence beats phrasing, in both directions: if the run actually changed
+  // files, or explicitly delivered an alternative, it is not a closed door.
+  if ((changedFiles ?? []).length > 0 || RESOLVED_ANYWAY_RE.test(text)) return null;
   return {
     kind: 'unresolved-refusal',
     reason: '変更・解決依頼に対して「対応不能」で終了しており、要求した結果が実現されていません。',

package/engine/lib.mjs

@@ -1346,9 +1346,18 @@ export function looksIncomplete(text) {
   const t = String(text ?? '').trim();
   if (!t) return true;
   if (t.length >= 200) return false;
-  // A short blurb ending mid-word/mid-clause (no sentence-final punctuation
-  // in any of ja/en) reads as truncated, not as a deliberate one-liner.
-  if (!/[.!?。！？」）)]$/.test(t)) return true;
+  // A short blurb ending mid-clause reads as truncated. This used to be "no
+  // sentence-final punctuation → truncated", which was far too wide: a terse
+  // but FINISHED one-liner ("worker done", "対応済み") has no period either, and
+  // that false positive marked the task `interrupted` → the whole goal `partial`,
+  // so a legitimate 0-file answer/investigation goal never reached Review
+  // (2026-07-29 audit; it is also what deterministically broke
+  // auto-test-rework.test.mjs). Require an actual DANGLING tail instead: a
+  // trailing connector/particle, a comma/colon/opening bracket, or a hyphen
+  // left mid-word. The fragment this rule was written for (goal-783/task-448)
+  // is still caught — by the future-intent opener below, which is the real
+  // signal there.
+  if (!/[.!?。！？」）)]$/.test(t) && /(?:[,、:：\-—(（「【]|\b(?:and|but|or|so|then|to|the|a|an|of|for|with|because|that|which|while|after|before|from|into)|[でをにがはとしてやりつつ])$/i.test(t)) return true;
   // goal-783/task-448's actual fragment ("I'll wait for this background
   // full-suite run to complete rather than poll.") IS a grammatically
   // complete sentence, so the check above misses it — but it's a one-shot
@@ -2987,12 +2996,14 @@ export function replayQueueLog(rawText, reviewDefinitions = {}) {
   }
   for (const t of tasks) {
     if (t.status !== 'running') continue;
+    const parent = goals.find((g) => g.id === t.goalId);
     // A Manager restart is an infrastructure interruption, not a request for
     // human judgment. Resume the same task once from the files/session already
-    // on disk. The bounded counter prevents a crash-looping task from being
-    // relaunched forever; a second interrupted boot remains visibly retryable.
+    // on disk, but only while its parent goal is genuinely in flight. A stale
+    // running task under a partial/review/done goal must never resurrect work
+    // the customer has already moved past. The bounded counter prevents loops.
     const recoveries = Number(t.restartRecoveries ?? 0);
-    if (recoveries < 1) {
+    if (parent?.status === 'running' && recoveries < 1) {
       t.status = 'queued';
       t.restartRecoveries = recoveries + 1;
       t.recoveredFromRestart = true;

package/engine/server.mjs

@@ -2813,6 +2813,7 @@ async function runTask(task) {
       followUp: task.detail,
       result: task.result,
       status: task.status,
+      changedFiles: task.changedFiles,
     });
     if (completionIssue) {
       task.status = 'interrupted';
@@ -3047,6 +3048,7 @@ async function runTask(task) {
     followUp: task.reply ? task.detail : '',
     result: task.result,
     status: task.status,
+    changedFiles: task.changedFiles,
   });
   if (completionIssue) {
     task.status = 'interrupted';
@@ -3105,13 +3107,21 @@ function testEnvironmentPreflight(dir, dependencySourceDir = null) {
   const sources = nodePath ? nodePath.split(process.platform === 'win32' ? ';' : ':') : [];
   const hasPuppeteer = sources.some((p) => existsSync(join(p, 'puppeteer-core', 'package.json')));
   const npm = spawnSync('npm', ['--version'], { encoding: 'utf8' });
+  // BLOCKING vs ADVISORY (2026-07-29, docs/INCIDENT-2026-07-25-29-INTERFACE-AUDIT.md).
+  // Only `node`/`npm` make `npm test` impossible to START. `dependencies` and
+  // `puppeteer` are properties of THIS repo's own browser suite — most user
+  // projects legitimately have neither, and treating them as blocking meant
+  // runTestsOnce returned "0 failed" for a suite it never ran, which
+  // runProjectTests then published as ok:true. A missing puppeteer-core is
+  // already discounted downstream by countInfraFlakes, so the honest thing is
+  // to run the suite and report the advisory alongside the real result.
   const checks = [
-    { id: 'node', ok: Boolean(process.execPath), detail: process.version },
-    { id: 'npm', ok: npm.status === 0, detail: npm.status === 0 ? npm.stdout.trim() : 'npm command unavailable' },
-    { id: 'dependencies', ok: sources.length > 0, detail: sources.length ? `${sources.length} node_modules source(s)` : 'node_modules not found' },
-    { id: 'puppeteer', ok: hasPuppeteer, detail: hasPuppeteer ? 'puppeteer-core available' : 'puppeteer-core missing; run npm install in the project directory' },
+    { id: 'node', ok: Boolean(process.execPath), detail: process.version, blocking: true },
+    { id: 'npm', ok: npm.status === 0, detail: npm.status === 0 ? npm.stdout.trim() : 'npm command unavailable', blocking: true },
+    { id: 'dependencies', ok: sources.length > 0, detail: sources.length ? `${sources.length} node_modules source(s)` : 'node_modules not found', blocking: false },
+    { id: 'puppeteer', ok: hasPuppeteer, detail: hasPuppeteer ? 'puppeteer-core available' : 'puppeteer-core missing; run npm install in the project directory', blocking: false },
   ];
-  return { ok: checks.every((c) => c.ok), checks };
+  return { ok: checks.every((c) => c.ok), canRun: checks.every((c) => c.ok || !c.blocking), checks };
 }
 function runTestsOnce(dir, dependencySourceDir = null) {
   // The timeout is env-overridable so tests can force the killed-run path fast;
@@ -3119,7 +3129,12 @@ function runTestsOnce(dir, dependencySourceDir = null) {
   const runTimeout = Number(process.env.MANAGER_TEST_TIMEOUT_MS) || 240000;
   return testRunQueue.run(() => new Promise((res) => {
     const preflight = testEnvironmentPreflight(dir, dependencySourceDir);
-    if (!preflight.ok) return res({ passed: 0, failed: 0, preflight, timedOut: false, realAssertion: false, infraOnly: true, infraFlakes: 0, detail: preflight.checks.filter((c) => !c.ok).map((c) => c.detail).join('\n'), failingNames: [] });
+    // Never report an unrun suite as a run with zero failures — runProjectTests
+    // turns `failed === 0` into ok:true, and that is the shape a caller reads as
+    // "verification passed". A suite that could not START is `notRun`.
+    if (!preflight.canRun) {
+      return res({ notRun: true, passed: 0, failed: 0, preflight, timedOut: false, realAssertion: false, infraOnly: true, infraFlakes: 0, detail: preflight.checks.filter((c) => !c.ok && c.blocking).map((c) => c.detail).join('\n'), failingNames: [] });
+    }
     const nodePath = testNodePath(dir, dependencySourceDir);
     execFile('npm', ['test'], { cwd: dir, timeout: runTimeout, env: { ...process.env, ...(nodePath ? { NODE_PATH: nodePath } : {}) }, maxBuffer: 8 * 1024 * 1024 }, (e, out = '', err = '') => {
       const text = `${out}\n${err}`;
@@ -3158,6 +3173,12 @@ async function runProjectTests(dir, dependencySourceDir = null) {
   try { hasTest = !!JSON.parse(readFileSync(join(dir, 'package.json'), 'utf8')).scripts?.test; } catch { /* no package.json */ }
   if (!hasTest) return { ran: false };
   let r = await runTestsOnce(dir, dependencySourceDir);
+  // The suite could not be started at all (node/npm missing). Report it as
+  // NOT RUN — the same honest shape as the documentation-only / gate-off skips
+  // — instead of ran:true + ok:true, which reads as "verification passed".
+  if (r.notRun) {
+    return { ran: false, skipped: true, skippedReason: `verification preflight blocked: ${r.detail || 'test environment unavailable'}`, preflight: r.preflight ?? null, detail: r.detail ?? '' };
+  }
   // Flake resistance: the integration suite spawns real servers; under load a
   // startup can time out. Re-run once and take the fewer failures — a real
   // failure fails both runs, a flake clears on the retry — so a good goal is
@@ -3701,10 +3722,17 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
   if (goal.testResult.ran && goal.testResult.infraOnly && !goal.testResult.inconclusive) {
     goalAct(`Project tests: all ${goal.testResult.infraFlakes || goal.testResult.failed || ''} failure(s) were infra flakes (a real server could not boot / a wait timed out under machine load), not code regressions — not blocking. Re-run when the machine is idle to confirm.`.replace(/\s{2,}/g, ' '));
   }
-  if (goal.testResult.ran && goal.testResult.preflight && !goal.testResult.preflight.ok) {
+  // Preflight is an ADVISORY note, not a verdict (2026-07-29). It used to force
+  // inconclusive=true, which silently disabled the gate for every project that
+  // lacks this repo's puppeteer suite — i.e. essentially every user project —
+  // while still publishing testResult.ok === true. Now: say what is missing, and
+  // let the REAL run decide. A suite that could not start at all never reaches
+  // here (runProjectTests returns ran:false with skippedReason instead).
+  if (goal.testResult.preflight && !goal.testResult.preflight.ok) {
     const missing = goal.testResult.preflight.checks.filter((c) => !c.ok).map((c) => c.id).join(', ');
-    goal.testResult.inconclusive = true;
-    goalAct(`Project tests were not started: verification preflight failed (${missing}). This is an environment issue, not an implementation failure. Fix the listed dependency or tool, then re-run verification.`);
+    goalAct(goal.testResult.ran
+      ? `Verification environment note: ${missing} missing. The suite still ran; browser-dependent tests in it may fail for environment reasons (those are discounted as infra flakes).`
+      : `Project tests were not started: verification preflight failed (${missing}). This is an environment issue, not an implementation failure. Fix the listed dependency or tool, then re-run verification.`);
   }
   // Baseline diff (Masa 2026-07-19): block ONLY on failures THIS change
   // introduced. A test already red on the base — e.g. persist-failure on clean

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.91",
+  "version": "0.10.93",
   "type": "module",
   "description": "Galda - hand off work to Claude Code or Codex, get proof back. Runs on your existing subscription, no extra API cost.",
   "scripts": {