@galda/cli 0.10.75 → 0.10.76

package/app/index.html

@@ -3354,12 +3354,14 @@ const withKey = (u) => {
 };
 
 // Theme names v3 (§6.5): English registered names only. Migrate persisted values
-// (old internal codenames → v3 names; dark→midnight; archived glass→gradient default).
+// (old internal codenames → v3 names; dark→midnight; archived glass→midnight default).
 const THEMES = ['studio', 'gradient', 'banff', 'midnight', 'cinema'];
 const THEME_MIGRATE = { w4: 'studio', arc: 'gradient', nature: 'banff', glow: 'midnight', comp: 'cinema', dark: 'midnight', glass: 'gradient' };
 const savedTheme = (() => {
   const s = localStorage.getItem('theme');
-  const v = THEME_MIGRATE[s] ?? (THEMES.includes(s) ? s : 'gradient');
+  // New installations start in Midnight. Existing valid values are preserved;
+  // only a missing/invalid persisted value takes this default.
+  const v = THEME_MIGRATE[s] ?? (THEMES.includes(s) ? s : 'midnight');
   if (v !== s) localStorage.setItem('theme', v);
   return v;
 })();
@@ -8702,6 +8704,11 @@ async function postOutboxItem(item){
       render();
       return; // finally still runs posting.delete(); skip land()/upsert()
     }
+    // This optimistic turn was only for the short interval before the server
+    // classified the input. A real goal is represented by the To Do lane and
+    // its live Implementing log, never as a second message below the composer.
+    const turns = state.chatTurns[item.projectId];
+    if (turns) state.chatTurns[item.projectId] = turns.filter((turn) => turn.oid !== item.oid);
     goal.imageUrls = item.imageUrls;
     outbox.land(item.oid, goal.id);
     upsert(state.goals, goal);
@@ -8875,8 +8882,9 @@ async function send(){
     const fullText = batt.length
       ? `${text}${text ? '\n\n' : ''}添付画像（workerはこのパスをReadで確認できます）: ${batt.map((a) => a.path).join(', ')}`
       : text;
-    // same housekeeping as a normal send: the words appear in the feed at once.
-    (state.chatTurns[state.active] ||= []).push({ role: 'you', text, ts: Date.now() });
+    // A bound send belongs to this goal's persisted thread.  Do not also add it
+    // to the center conversation feed: that made a task reply appear below the
+    // composer as an unrelated "You" turn.
     state.attach = [];
     state.msg = { mode: null, model: null, effort: null, skill: null, later: false, review: false };
     renderMsgHint();
@@ -8925,9 +8933,10 @@ async function send(){
     status: 'sending', ts: Date.now(),
   };
   outbox.add(item); // saved locally BEFORE any network call
-  // Pattern A (H23): the user's own turn enters the center feed immediately, so
-  // the composer reads like a conversation; if it's chat, a reply turn follows.
-  (state.chatTurns[state.active] ||= []).push({ role: 'you', text, ts: Date.now() });
+  // The server decides whether this is a conversation or a real task. Show an
+  // optimistic conversation turn while that decision is pending, then remove it
+  // if the request becomes a task (tasks belong in To Do / Implementing only).
+  (state.chatTurns[state.active] ||= []).push({ role: 'you', text, oid: item.oid, ts: Date.now() });
   state.attach = [];
   // mode/skill/later/review are per-message → clear after send (model stays sticky).
   state.msg = { mode: null, model: null, effort: null, skill: null, later: false, review: false };

package/engine/lib.mjs

@@ -1306,6 +1306,10 @@ export function looksIncomplete(text) {
 export function isSuspiciouslyIncompleteDone({ status, changedFiles = [], result = '', hasVerdict = false } = {}) {
   if (status !== 'done' || hasVerdict) return false;
   if ((changedFiles ?? []).length > 0) return false;
+  // A worker that only asks for permission to start a local server did not
+  // produce a locally checkable result. Treating its clean exit as success
+  // re-opened the old PR/review summary and hid this real blocker (#788).
+  if (/\b(permission|approve running|approval required)\b|許可.*(必要|ください)|実行してよろしい/i.test(String(result))) return true;
   return looksIncomplete(result);
 }
 
@@ -2382,65 +2386,6 @@ export function dismissGoal({ goalStatus, outcome = 'continue' } = {}) {
   return { ok: true, outcome: 'continue', status: 'running', spawnsWorker: true };
 }
 
-// 誰が決めるか＝**AI本人** (docs/design/ONE-CONVERSATION.md §1.4)
-//
-// 返事が来た時、それが「このまま直して」なのか「考え直したい」なのか「別件」なのか
-// 「ただ聞いただけ」なのかを、**我々の層で当てない**。キーワードでも類似度でもなく、
-// 人の言葉を1回だけ AI に渡して、返ってきた答えのとおりに配線する。0.72 の自動合流を
-// 消したのと同じ理由——この層が人の意図を推測し始めると、外れた時に黙って別のことを
-// する（今日そこを1つずつ潰してきた）。
-//
-// 迷った時だけ人に聞く: AI が unsure と答えたら**ゴールに触らず**選択肢を返す。人は
-// モードを普段選ばない。迷った時だけ選ぶ。
-export const REPLY_INTENTS = [...REPLY_OUTCOMES, 'unsure'];
-export function buildReplyIntentPrompt({ goalText, reply } = {}) {
-  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 (the usual case).',
-    '- "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 — they want to know something.',
-    '- "unsure": you genuinely cannot tell. Prefer this over guessing: the person will be shown the choices.',
-    'Judge the reply on its own words. Do not treat a question mark as proof of "answer", or an imperative as proof of "continue" — "why is this red? make it blue" asks for a change.',
-    'Output ONLY one line of raw JSON: {"outcome":"continue|rescope|split|answer|unsure"}. No markdown, no code fence, no other text.',
-    '',
-    'The request that was worked on:',
-    String(goalText ?? '').slice(0, 2000),
-    '',
-    'Their reply:',
-    String(reply ?? '').slice(0, 4000),
-  ].join('\n');
-}
-
-// Tolerant parser — never throws. It returns TWO things, and the difference
-// between them is the whole point:
-//
-//   read: true  … the AI answered, and we understood it. 'unsure' here means the
-//                 AI genuinely could not tell → ask the person (§1.4).
-//   read: false … we could not get an answer at all (no agent installed, offline,
-//                 rate-limited, garbled output). The AI never said anything, so
-//                 there is nobody to have been unsure.
-//
-// Treating those two the same is a real bug, found by an existing test on
-// 2026-07-22: on a machine with no agent, a reply stopped doing ANYTHING — it put
-// the four choices on screen and waited. But if the agent cannot be reached, no
-// answer can be written either; the only useful thing a reply can still do is
-// what it always did — go back as rework, and run when the agent returns.
-// So the caller falls back to 'continue' when read is false. That is not a guess
-// about what the person meant: it is the documented default this endpoint has
-// always had when no outcome is given.
-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;   // answered, but not in our words
-  return { outcome: obj.outcome, 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

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 } from './lib.mjs';
-import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession } from './lib.mjs';
+import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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 } from './lib.mjs';
 import { createSerialQueue } from './lib.mjs';
 import { isPrClosed } from './lib.mjs';
 import { collectProjectRules } from './lib.mjs';
@@ -5214,24 +5214,6 @@ const server = createServer(async (req, res) => {
     return json(res, 200, payload);
   }
 
-// 返事の行き先を AI 本人に決めてもらう (ONE-CONVERSATION.md §1.4)。
-//
-// 判定は1往復だけの軽い仕事なので、作業した本人ではなくユーティリティのモデルに
-// 聞く（答えを書く answer 本体は本人に聞く——あちらは中身を知っている必要がある）。
-// 呼べなかった時は 'unsure'。「たぶん continue」で走り出すより、人に1タップ選ばせる
-// ほうが安い。
-async function judgeReplyIntent(goal, reply) {
-  let parsed;
-  try {
-    const r = await runUtility({ prompt: buildReplyIntentPrompt({ goalText: goal.text, reply }), cwd: ROOT, tools: 'Read' });
-    parsed = parseReplyIntent(r.result);
-  } catch { parsed = { outcome: 'unsure', read: false }; }
-  // Could not ask at all → do what a reply has always done (continue). Blocking the
-  // reply on four choices would leave a machine with no reachable agent unable to
-  // send anything back, and an answer could not have been written there anyway.
-  return parsed.read ? parsed.outcome : 'continue';
-}
-
 // 「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
@@ -5285,26 +5267,20 @@ async function answerGoalQuestion(goal, question) {
     let body = '';
     req.on('data', (d) => { body += d; });
     req.on('end', async () => {
-      let text = '', outcome = 'continue', auto = false, requested = {};
+      let text = '', outcome = 'continue', requested = {};
       try {
         requested = JSON.parse(body || '{}');
         text = String(requested.text ?? '').trim();
         if (REPLY_OUTCOMES.includes(requested.outcome)) outcome = requested.outcome;
-        auto = requested.outcome === 'auto';
       } catch {}
       const worker = applyReplyWorker(goal, requested);
       if (!worker.ok) return json(res, 409, { error: worker.error });
 
-      // outcome:'auto' = 決めるのは AI 本人 (§1.4)。人の言葉を1回だけ渡して、返って
-      // きた行き先のとおりに配線する。我々はキーワードも類似度も見ない。
-      // 迷ったと言われたら**ゴールに触らずに**選択肢を返す——ここで勝手に continue に
-      // 倒すと、質問しただけの人のコードが動き出す（それが直したかった壊れ方）。
-      if (auto) {
-        if (!text) return json(res, 400, { error: 'text required' });
-        const judged = await judgeReplyIntent(goal, text);
-        if (judged === 'unsure') return json(res, 200, { goal, outcome: 'unsure', choices: REPLY_OUTCOMES });
-        outcome = judged;
-      }
+      // A review reply is transport, not a classification request. Legacy clients
+      // send outcome:'auto'; it deliberately means the deterministic default:
+      // preserve the exact words and resume this goal's agent session. Specialized
+      // outcomes work only when an explicit UI action sends one of REPLY_OUTCOMES.
+      if (requested.outcome === 'auto' && !text) return json(res, 400, { error: 'text required' });
       // 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。
       const result = dismissGoal({ goalStatus: goal.status, outcome });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.75",
+  "version": "0.10.76",
   "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": {