@galda/cli 0.10.93 → 0.10.94

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';
@@ -5522,29 +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'
@@ -5608,16 +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": "0.10.93",
+  "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": {