@galda/cli 0.10.92 → 0.10.94

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
@@ -2477,59 +2486,14 @@ export function dismissGoal({ goalStatus, outcome = 'continue' } = {}) {
 // remains unsure.  Returning unsure is safe because the UI already presents
 // these exact choices without moving the goal.
 export const REPLY_INTENTS = [...REPLY_OUTCOMES, 'unsure'];
-export function buildReplyIntentPrompt({ goalText, reply, currentProject = null, projects = [] } = {}) {
-  const projectLines = (projects ?? []).map((project) =>
-    `- ${project.id}: ${project.name || project.id} | root=${project.dir || '(unknown)'} | scope=${project.note || '(not specified)'}`);
-  return [
-    'You route replies in Agent Manager: a person delegated a coding task, an AI did it, and it came back for review. The person has now replied.',
-    'Decide what the reply ASKS FOR — exactly one of:',
-    '- "continue": fix or change the work that was just done.',
-    '- "rescope": stop and rethink; the request itself should now say something different.',
-    '- "split": a separate piece of work that should not touch this one.',
-    '- "answer": a question about the work. Nothing should be built or changed.',
-    '- "unsure": you genuinely cannot tell. Prefer this over guessing.',
-    'Judge the full meaning. A question followed by a requested change is "continue".',
-    'Also decide which project must receive the requested work.',
-    '- Keep targetProjectId equal to the current project for ordinary fixes to the reviewed work.',
-    '- If the current project is design/docs/prototype-only and the reply asks to implement or launch the chosen design in the real product, select the matching production/code project.',
-    '- Never infer a project that is not listed. If the target is ambiguous, use outcome "unsure" and targetProjectId null.',
-    'Output ONLY one line of raw JSON: {"outcome":"continue|rescope|split|answer|unsure","targetProjectId":"listed-id-or-null"}.',
-    '',
-    `Current project: ${currentProject?.id || '(unknown)'}`,
-    'Available projects:',
-    ...(projectLines.length ? projectLines : ['- (none)']),
-    '',
-    'The request that was worked on:',
-    String(goalText ?? '').slice(0, 2000),
-    '',
-    'Their reply:',
-    String(reply ?? '').slice(0, 4000),
-  ].join('\n');
-}
-
-export function parseReplyIntent(raw) {
-  const unread = { outcome: 'unsure', read: false };
-  if (!raw || typeof raw !== 'string') return unread;
-  const m = raw.match(/\{[\s\S]*\}/);
-  if (!m) return unread;
-  let obj;
-  try { obj = JSON.parse(m[0]); } catch { return unread; }
-  if (!REPLY_INTENTS.includes(obj?.outcome)) return unread;
-  const targetProjectId = typeof obj.targetProjectId === 'string' && obj.targetProjectId.trim()
-    ? obj.targetProjectId.trim() : null;
-  return { outcome: obj.outcome, targetProjectId, read: true };
-}
-
-// 宛先（docs/design/ONE-CONVERSATION.md §1.2）
-//
-// チャットで「#12 の件だけど」と書けたら、そのチケット宛て。**決定論で、我々の推測は
-// ゼロ**。これが要るのは、今まで「同じ話かどうか」を我々の層が文字列の似かた（0.72）で
-// 当てずっぽうに決めていたから——当たれば黙って合流し、外れれば別のカードが生えた。
-// 人が番号を書けるなら、推測する理由が無い。
-//
-// 最初の `#N` だけを見る。番号が実在しなければ宛先なしとして扱う（＝ふつうに新しい
-// チケットになる）ので、`#` を含むだけの文章がどこかへ黙って吸い込まれることはない。
-// 本文は削らない: 言われた言葉をそのまま作業AIへ渡す（P0 の「原文100%」と同じ約束）。
+// buildReplyIntentPrompt / parseReplyIntent lived here: the prompt and parser for a
+// utility LLM that decided what a Review reply MEANT before the goal's own agent saw
+// it. They are gone with their caller. Its "answer" bucket — "a question about the
+// work. Nothing should be built or changed" — swallowed every politely-phrased
+// Japanese instruction, because there is no other way to ask politely in Japanese.
+// Claude Code and Codex already tell a question from an instruction; a second opinion
+// in front of them could only ever be worse-informed. See the `auto` branch of
+// /dismiss in engine/server.mjs and engine/test/reply-goes-to-the-agent.test.mjs.
 export function parseGoalAddress(text) {
   // `(?![\w-])` ＝「番号の直後に英数字やハイフンが続くなら、それは番号ではない」:
   // `#1f6f4a`(色) や `#12-3` を宛先と読み違えない。日本語は \w に入らないので

package/engine/server.mjs

@@ -21,7 +21,7 @@ import { homedir } from 'node:os';
 import { fileURLToPath } from 'node:url';
 import { pathToFileURL } from 'node:url';
 import { needsProjectFolder, folderLabel, chooseFolderScript, parseChosenFolder, buildFolderQuestion, resolveFolderAnswer, classifyFolderTarget, buildFolderInitConfirm, isFolderInitConfirmed, needsReviewPreference, buildReviewPreferenceQuestion, resolveReviewPreferenceAnswer, routeQuestionAnswer, notUnderstoodNote, parseRelocalizedQuestion, classifyAuthProbe, authBannerText, makeSigninChallenge, parseClaimResponse, isCommandOnPath, parseRequestedWorkspaceDir } from './lib.mjs';
-import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, buildReviewAttemptLedger, latestGoalOutcomeTask, buildDirtyWorkspacePrBlock, buildGoalProofMd, buildGoalPrBody, buildReviewSummaryPrompt, parseReviewSummary, buildReviewRuleSection, nextGoalStatus, isPrMerged, isPrApproved, reviewToDoneStatus, advanceReviewGoal, revertReviewGoal, approveGoal, dismissGoal, rejectGoal, reopenGoal, permissionModeFor, isPlanReview, nextAfterPlanApprove, GOAL_MODES, parseSkillFrontmatter, collectSkills, retestGoal, cancelRetestGoal, computeRetestOutcome, revertGoal, planRevertActions, archiveGoal, unarchiveGoal, undismissGoal, parseNumstat, truncateDiffText, sumUsage, resolveEntryUrl, rebasePreviewUrl, shouldCreateGoalPR, normalizePullRequestUrl, parseGitLog, diffNewCommits, replayQueueLog, reconcileOrphanGoals, reorderQueue, sortQueueByPriority, TASK_PRIORITIES, validateWorkflowColumns, DEFAULT_WORKFLOW_COLUMNS, validateReviewDefinition, DEFAULT_REVIEW_DEFINITION, resolveTestGate, buildEphemeralSeedLog, buildEphemeralSeedProjects, trimTaskActivityForState, buildAskContext, WORKER_TOOLS, workerPrompt, verifyCfAccessJwt, resolveIdentity, goalVisibleTo, clampParallelLimit, canStartMore, nextRunnableTasks, goalsConflict, detectConflicts, pickFoldTarget, autoFoldReviewsEnabled, parseGitUnifiedDiffLocations, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, classifyPrSafety, isInconclusiveTestRun, countInfraFlakes, classifyRunFailures, classifyVerifyGate, buildVerificationInfraError, verificationInfraErrorFrom, buildExecutionPlan, buildContextHandoffSummary, buildFailurePostmortem, appendFailureMemory, latestFailurePolicy, classifyChangeRisk, checkRunBudget, usageBudgetTokens, isRateLimited, nextResumeDelay, checkFreeTierLimit, resolveEntitlement, resolveCachedEntitlement, FREE_TIER_LIMITS, QUEUED_GOAL_STATUSES, verifyLicenseToken, licenseTokenPayload, shouldRefreshLicense, shouldEmitSetupCompleted, pickAnalyticsUid, shouldEmitFreeExhausted, detectRequestLanguage, testFailureReason, manualTestRetryPrompt, isNothingVerifiable, isNothingVerifiableForPR, isSuspiciouslyIncompleteDone, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, shouldUsePlannerForGoal, RUN_DECL_FILE, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession, parseApprovalRequest } from './lib.mjs';
+import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, buildReviewAttemptLedger, latestGoalOutcomeTask, buildDirtyWorkspacePrBlock, buildGoalProofMd, buildGoalPrBody, buildReviewSummaryPrompt, parseReviewSummary, buildReviewRuleSection, nextGoalStatus, isPrMerged, isPrApproved, reviewToDoneStatus, advanceReviewGoal, revertReviewGoal, approveGoal, dismissGoal, rejectGoal, reopenGoal, permissionModeFor, isPlanReview, nextAfterPlanApprove, GOAL_MODES, parseSkillFrontmatter, collectSkills, retestGoal, cancelRetestGoal, computeRetestOutcome, revertGoal, planRevertActions, archiveGoal, unarchiveGoal, undismissGoal, parseNumstat, truncateDiffText, sumUsage, resolveEntryUrl, rebasePreviewUrl, shouldCreateGoalPR, normalizePullRequestUrl, parseGitLog, diffNewCommits, replayQueueLog, reconcileOrphanGoals, reorderQueue, sortQueueByPriority, TASK_PRIORITIES, validateWorkflowColumns, DEFAULT_WORKFLOW_COLUMNS, validateReviewDefinition, DEFAULT_REVIEW_DEFINITION, resolveTestGate, buildEphemeralSeedLog, buildEphemeralSeedProjects, trimTaskActivityForState, buildAskContext, WORKER_TOOLS, workerPrompt, verifyCfAccessJwt, resolveIdentity, goalVisibleTo, clampParallelLimit, canStartMore, nextRunnableTasks, goalsConflict, detectConflicts, pickFoldTarget, autoFoldReviewsEnabled, parseGitUnifiedDiffLocations, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, classifyPrSafety, isInconclusiveTestRun, countInfraFlakes, classifyRunFailures, classifyVerifyGate, buildVerificationInfraError, verificationInfraErrorFrom, buildExecutionPlan, buildContextHandoffSummary, buildFailurePostmortem, appendFailureMemory, latestFailurePolicy, classifyChangeRisk, checkRunBudget, usageBudgetTokens, isRateLimited, nextResumeDelay, checkFreeTierLimit, resolveEntitlement, resolveCachedEntitlement, FREE_TIER_LIMITS, QUEUED_GOAL_STATUSES, verifyLicenseToken, licenseTokenPayload, shouldRefreshLicense, shouldEmitSetupCompleted, pickAnalyticsUid, shouldEmitFreeExhausted, detectRequestLanguage, testFailureReason, manualTestRetryPrompt, isNothingVerifiable, isNothingVerifiableForPR, isSuspiciouslyIncompleteDone, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, shouldUsePlannerForGoal, RUN_DECL_FILE, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession, parseApprovalRequest } from './lib.mjs';
 import { createSerialQueue } from './lib.mjs';
 import { parseRequirementEvidence, unmappedChangedFiles } from './lib.mjs';
 import { isPrClosed } from './lib.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
@@ -5494,29 +5522,13 @@ const server = createServer(async (req, res) => {
 // 「聞いてるだけ」への返事 (ONE-CONVERSATION.md §1.3 answer)。
 //
 // 答えるのは**作業した AI 本人**＝その agent の session を resume する。状況一覧しか
-// Resolve the app's `outcome:auto` contract before changing goal state. The
-// utility call is classification-only; the goal's own agent/session still
-// writes an answer or performs the requested change. If routing is unavailable,
-// return unsure so a customer chooses explicitly instead of accidentally
-// starting a ten-minute implementation worker for a question.
-async function judgeReplyIntent(goal, reply) {
-  try {
-    const currentProject = projects.find((project) => project.id === goal.projectId) ?? null;
-    const r = await runUtility({
-      prompt: buildReplyIntentPrompt({ goalText: goal.text, reply, currentProject, projects }),
-      cwd: ROOT,
-      tools: 'Read',
-    });
-    const parsed = parseReplyIntent(r.result);
-    if (!parsed.read) return { outcome: 'unsure', targetProjectId: null };
-    if (parsed.targetProjectId && !projects.some((project) => project.id === parsed.targetProjectId)) {
-      return { outcome: 'unsure', targetProjectId: null };
-    }
-    return { outcome: parsed.outcome, targetProjectId: parsed.targetProjectId };
-  } catch {
-    return { outcome: 'unsure', targetProjectId: null };
-  }
-}
+// judgeReplyIntent lived here: a utility-LLM pass that sorted a Review reply into
+// continue / rescope / split / answer BEFORE the goal's own agent saw it. It is
+// gone — see the `auto` branch of /dismiss for the measured reason. Its stated
+// worry was "accidentally starting a ten-minute implementation worker for a
+// question", but the trade it actually made was the opposite one: it answered
+// real instructions, and the person paid several round trips to get the work
+// started at all. Deciding that is the agent's job.
 
 // 知らない別AIでは「なぜそうしたか」に答えられない（Masa確定 2026-07-22）。
 // コードは1文字も変わらない: ツールは Read だけ、permissionMode は 'plan'
@@ -5580,16 +5592,33 @@ async function answerGoalQuestion(goal, question) {
 
       if (auto) {
         if (!text) return json(res, 400, { error: 'text required' });
-        const routed = await judgeReplyIntent(goal, text);
-        outcome = routed.outcome;
-        requested.targetProjectId = routed.targetProjectId;
-        if (outcome === 'unsure') return json(res, 200, {
-          goal,
-          outcome,
-          choices: REPLY_OUTCOMES,
-          projectChoices: projects.map((project) => ({ id: project.id, name: project.name, current: project.id === goal.projectId })),
-        });
-        if (requested.targetProjectId && requested.targetProjectId !== goal.projectId) outcome = 'split';
+        // The reply goes to the agent exactly as written. Galda does not decide
+        // first what the person meant.
+        //
+        // This used to call a utility LLM (judgeReplyIntent) to sort the reply
+        // into continue / rescope / split / answer before the agent ever saw it.
+        // Its definition of "answer" was "a question about the work. Nothing
+        // should be built or changed" — and Japanese has no other way to make a
+        // polite request. Measured on the operator's own instance 2026-07-25..29:
+        // 「案3で実装出来ますか？」「Aの方針で実装できますか！？」「あれ、prいるよね？」
+        // 「okなんでprなげれますか？」「案がみえないのでローカルで確認できるように貰える？」
+        // were all real instructions, and every one of them could be read as a
+        // question. When it read one that way it created NO task at all, so the
+        // request was answered instead of done and the person had to ask two or
+        // three more times.
+        //
+        // The fix is not a better classifier. Claude Code and Codex already
+        // decide correctly whether a message wants an answer or a change — that
+        // is what they are. Galda's job is the interface around them, not a
+        // second opinion in front of them. Note the `answer` route spawned an
+        // agent too (Read-only, plan mode), so passing the reply through costs
+        // no extra run — it removes one utility-LLM call per reply.
+        //
+        // Explicit routes are untouched: when the PERSON picks rescope / split /
+        // answer from the chips, `requested.outcome` carries it and this branch
+        // never runs. A human choosing where their words go is the interface
+        // working; Galda guessing is not.
+        outcome = 'continue';
       }
       // 返事の行き先4種 (ONE-CONVERSATION.md §1.3). 既定は今までどおり continue なので、
       // outcome を送ってこない今の UI の挙動は1ミリも変わらない。

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.92",
+  "version": "0.10.94",
   "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": {