@galda/cli 0.10.93 → 0.10.95

package/engine/lib.mjs

@@ -2486,59 +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';
@@ -1425,8 +1425,19 @@ function gitStatusPaths(lines = []) {
 function newAttemptId(goal, task) {
   return `goal-${goal?.id ?? 'none'}-task-${task.id}-${Date.now()}-${randomBytes(3).toString('hex')}`;
 }
-function noteGoalTaskBaseline(goal, project, workDir, before) {
-  if (!goal || goal.prBaseline) return;
+// `refresh` re-measures a baseline that already exists. Without it the snapshot
+// is write-once: whatever the shared checkout happened to look like when the
+// goal's FIRST task started is what classifyPrSafety judges forever. So a goal
+// that began while some unrelated file was dirty is permanently unable to
+// produce a PR — "PR creation skipped: this goal started in a dirty shared
+// workspace" re-emits on every close-out even after the workspace was cleaned
+// minutes later, and no amount of asking can fix it. Measured on the operator's
+// instance 2026-07-25..29: goals #729 #733 #731 are all stuck this way.
+// Re-measuring is only correct while there is no PR yet — once a branch exists
+// it is the reference, and moving the baseline under it would compare the work
+// against the wrong starting point.
+function noteGoalTaskBaseline(goal, project, workDir, before, { refresh = false } = {}) {
+  if (!goal || (goal.prBaseline && !refresh)) return;
   const repoRoot = gitSyncOrNull(workDir, ['rev-parse', '--show-toplevel']);
   const branch = repoRoot ? gitSyncOrNull(workDir, ['branch', '--show-current']) : null;
   const head = repoRoot ? gitSyncOrNull(workDir, ['rev-parse', 'HEAD']) : null;
@@ -2731,7 +2742,9 @@ async function runTask(task) {
   task.attemptId = task.attemptId || newAttemptId(goal, task);
   if (goal) {
     goal.latestAttemptId = task.attemptId;
-    noteGoalTaskBaseline(goal, project, workDir, before);
+    // A reply is a fresh attempt: re-measure the workspace so a baseline taken
+    // when the checkout happened to be dirty cannot outlive the mess itself.
+    noteGoalTaskBaseline(goal, project, workDir, before, { refresh: !!task.reply && !goal.pr });
   }
   sendProcessAct(task, `Checked starting git state (${before.length} changed file${before.length === 1 ? '' : 's'} before this task).`);
   // Plan mode makes no edits, so there is nothing to verify — never author a
@@ -3886,7 +3899,15 @@ async function createGoalPR(goal, project, goalTasks, reviewDefinition = DEFAULT
   if (!shouldCreateGoalPR(goal, files)) {
     goal.pr = null;
     goal.prBranch = null;
-    goal.prError = null;
+    // Two different situations used to leave the same silent hole. "No PR was
+    // asked for" is nothing to explain — clear the error. But "a PR WAS asked
+    // for and none exists" is a fact the person needs, and nulling prError here
+    // erased the only field the Review card renders a reason from, so the card
+    // showed a PR-less goal with no explanation at all (goals #725 #751 on the
+    // operator's instance: status done, wantsPR true, pr null, prError empty).
+    goal.prError = goal.wantsPR === true
+      ? 'PRを作成できませんでした: 変更されたファイルが1つもありません（実装が行われていない可能性があります）/ Cannot open a PR: this run produced no file changes.'
+      : null;
     saveGoal(goal);
     goalAct(goal.wantsPR === false
       ? 'Deliverable is No PR, so PR creation is skipped even though files changed.'
@@ -5313,6 +5334,20 @@ const server = createServer(async (req, res) => {
           goal.blocked = null;
           saveGoal(goal);
         }
+        // "PR出して" has to count here too. This is the route EVERY follow-up reply
+        // takes (the review card switches to it once a card is in conversation) and
+        // the one MCP manager_reply uses, so until now a PR asked for on the second
+        // reply — or from an agent session — could never take effect: the deliverable
+        // was only ever upgraded in /dismiss. Measured on the operator's instance
+        // 2026-07-25..29: 「あれ、prいるよね？」then「PR作成して」(#790) and
+        // 「問題ないのでpr出して！」then「あれ、問題ないのでpr作ってください」(#800) —
+        // both asked twice and got no PR either time.
+        // Upgrade only: this can turn the deliverable on, never off, so a later
+        // unrelated reply cannot silently withdraw a PR the person already asked for.
+        if (resolveWantsPR('auto', text, false) && goal.wantsPR !== true) {
+          goal.wantsPR = true;
+          saveGoal(goal);
+        }
         json(res, 200, queueReplyTask(goal, text.trim(), { from: 'you', via: 'thread' }));
       } catch { json(res, 400, { error: 'bad json' }); }
     });
@@ -5391,6 +5426,41 @@ const server = createServer(async (req, res) => {
     return json(res, 200, goal);
   }
 
+  // PR再試行 / Retry PR. The Review card has offered this button all along and it
+  // has never had a route to call — fsRetryPrGoal POSTs here, the server had no
+  // match for it, and the press 404'd and showed an error. So the one repair the
+  // failure message advertises ("commit/stash/revert those unrelated changes,
+  // then rerun") was unreachable from the product: the person cleaned the
+  // workspace and had no way to say "now try again".
+  //
+  // Deliberately narrow — this only re-runs PR creation. It starts no worker and
+  // changes no code, so it is safe to press repeatedly: either the PR appears or
+  // the same honest reason comes back.
+  const retryPrMatch = url.pathname.match(/^\/api\/goals\/(\d+)\/retry-pr$/);
+  if (retryPrMatch && req.method === 'POST') {
+    const goal = goals.find((g) => g.id === Number(retryPrMatch[1]));
+    if (!goal) return json(res, 404, { error: 'goal not found' });
+    if (!requireGoalOwnership(req, res, goal)) return;
+    if (goal.pr) return json(res, 409, { error: 'this goal already has a pull request' });
+    const project = projects.find((p) => p.id === goal.projectId);
+    if (!project) return json(res, 404, { error: 'project not found' });
+    // Asking to retry the PR IS asking for a PR — a goal that was created as
+    // No PR but whose owner now presses this wants one.
+    goal.wantsPR = true;
+    // The frozen dirty-workspace verdict is exactly what this button exists to
+    // clear. Re-measure against the workspace as it is NOW, which is the state
+    // the person just repaired.
+    const wdir = (goalWorkDirs.get(goal.id) && existsSync(goalWorkDirs.get(goal.id))) ? goalWorkDirs.get(goal.id) : project.dir;
+    goal.prBaseline = null;
+    goal.prSafety = null;
+    goal.prError = null;
+    noteGoalTaskBaseline(goal, project, wdir, await gitChanges(wdir));
+    const siblings = tasks.filter((t) => t.goalId === goal.id);
+    await createGoalPR(goal, { ...project, dir: wdir }, siblings, getReviewDefinition(goal.projectId), null);
+    saveGoal(goal);
+    return json(res, 200, { goal, pr: goal.pr ?? null, prError: goal.prError ?? null });
+  }
+
   // Archive: retire an UNFINISHED goal (partial/failed/interrupted/blocked)
   // cleanly — spawning no rework. DELETE only accepts stacked/pending, and
   // dismiss on a partial goal queues a "差し戻し" rework task, so a goal that
@@ -5522,29 +5592,13 @@ 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'
@@ -5608,16 +5662,33 @@ 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。
@@ -5677,7 +5748,12 @@ async function answerGoalQuestion(goal, question) {
       // Manager close-out below owns commit/push/PR creation. Keep this
       // monotonic: an explicit later PR request upgrades a No-PR/auto goal,
       // while ordinary feedback never silently downgrades an existing PR goal.
-      if (result.spawnsWorker && resolveWantsPR('auto', text, false)) goal.wantsPR = true;
+      // Not gated on spawnsWorker: what the person wants DELIVERED is independent
+      // of where their words were routed. A PR asked for while parking the goal
+      // (rescope) or while asking a question (answer) is still a PR they asked
+      // for, and it must survive until close-out rather than being dropped
+      // because this particular reply happened not to wake a worker.
+      if (resolveWantsPR('auto', text, false)) goal.wantsPR = true;
       goal.status = result.status;
       saveGoal(goal);
       // rescope はワーカーを起こさない——「いったん止めて考え直す」なので、走り出したら

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.93",
+  "version": "0.10.95",
   "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": {