@galda/cli 0.10.83 → 0.10.85

package/app/theme.css

@@ -21,9 +21,9 @@
   --tint:20,23,31;
   --hair:rgba(var(--tint),.10); --hair2:rgba(var(--tint),.16);
   --glass:rgb(255 255 255 / .72);
-  --green-rgb:10,154,106; --blue-rgb:56,116,214; --amber-rgb:170,128,40; --danger-rgb:200,44,44;
+  --green-rgb:10,154,106; --blue-rgb:56,116,214; --review-rgb:92,196,255; --amber-rgb:170,128,40; --danger-rgb:200,44,44;
   --purple-rgb:150,90,220; --glow-rgb:120,150,230;
-  --green:rgb(var(--green-rgb)); --amber:rgb(var(--amber-rgb)); --blue:rgb(var(--blue-rgb)); --danger:rgb(var(--danger-rgb));
+  --green:rgb(var(--green-rgb)); --amber:rgb(var(--amber-rgb)); --blue:rgb(var(--blue-rgb)); --review:rgb(var(--review-rgb)); --danger:rgb(var(--danger-rgb));
   --iri:linear-gradient(112deg,#b794f6 0%,#7cc4ff 38%,#5ce0c0 72%,#ffd27a 100%);
   --on-accent:#0a0a0b; --invink:#ffffff;
   --send-bg:#1d1d1f; --send-ink:#ffffff;
@@ -44,7 +44,7 @@ body[data-theme="banff"]{
   --tint:255,255,255;
   --hair:rgba(var(--tint),.09); --hair2:rgba(var(--tint),.16);
   --glass:rgba(12,13,17,.6);
-  --green-rgb:79,215,156; --blue-rgb:124,196,255; --amber-rgb:228,198,120; --danger-rgb:240,115,111;
+  --green-rgb:79,215,156; --blue-rgb:3,155,229; --amber-rgb:228,198,120; --danger-rgb:240,115,111;
   --purple-rgb:183,148,246; --glow-rgb:146,180,255;
   --green:rgb(var(--green-rgb)); --amber:rgb(var(--amber-rgb)); --blue:rgb(var(--blue-rgb)); --danger:rgb(var(--danger-rgb));
   --on-accent:#0a0a0b; --invink:#0a0a0b;
@@ -62,7 +62,7 @@ body[data-theme="dark"],body[data-theme="midnight"],body[data-theme="cinema"]{
   --tint:255,255,255;
   --hair:rgba(var(--tint),.07); --hair2:rgba(var(--tint),.12);
   --glass:rgba(22,22,27,.55);
-  --green-rgb:52,211,153; --blue-rgb:124,196,255; --amber-rgb:228,198,120; --danger-rgb:255,120,120;
+  --green-rgb:52,211,153; --blue-rgb:3,155,229; --amber-rgb:228,198,120; --danger-rgb:255,120,120;
   --purple-rgb:183,148,246; --glow-rgb:146,180,255;
   --on-accent:#0a0a0b; --invink:#0a0a0b;
   --send-bg:var(--iri); --send-ink:#0a0a0b;
@@ -92,7 +92,7 @@ body[data-theme="glass"]{
   --tint:255,255,255;
   --hair:rgba(255,255,255,.16); --hair2:rgba(255,255,255,.24);
   --glass:rgba(12,13,17,.6);
-  --green-rgb:102,230,198; --blue-rgb:143,208,255; --amber-rgb:232,206,134; --danger-rgb:255,138,134;
+  --green-rgb:102,230,198; --blue-rgb:3,155,229; --amber-rgb:232,206,134; --danger-rgb:255,138,134;
   --purple-rgb:183,148,246; --glow-rgb:146,180,255;
   --green:rgb(var(--green-rgb)); --amber:rgb(var(--amber-rgb)); --blue:rgb(var(--blue-rgb)); --danger:rgb(var(--danger-rgb));
   --on-accent:#0a0a0b; --invink:#0a0a0b;
@@ -117,7 +117,7 @@ body[data-theme="gradient"]{
   --tint:20,23,31;
   --hair:rgba(var(--tint),.10); --hair2:rgba(var(--tint),.16);
   --glass:rgb(255 255 255 / .72);
-  --green-rgb:79,215,156; --blue-rgb:37,99,235; --amber-rgb:178,140,40; --danger-rgb:240,115,111;
+  --green-rgb:79,215,156; --blue-rgb:3,155,229; --amber-rgb:178,140,40; --danger-rgb:240,115,111;
   --purple-rgb:150,90,220; --glow-rgb:120,150,230;
   --green:rgb(var(--green-rgb)); --amber:rgb(var(--amber-rgb)); --blue:rgb(var(--blue-rgb)); --danger:rgb(var(--danger-rgb));
   --on-accent:#0a0a0b; --invink:#ffffff;

package/engine/lib.mjs

@@ -215,7 +215,7 @@ export const INTERNAL_QUALITY_EVENT_DEFINITIONS = Object.freeze([
 
 const INTERNAL_TERMINAL_TASK_STATUSES = new Set(['done', 'failed', 'interrupted', 'skipped']);
 const INTERNAL_FAILURE_GOAL_STATUSES = new Set(['failed', 'blocked', 'interrupted', 'partial']);
-const INTERNAL_OPEN_GOAL_STATUSES = new Set(['stacked', 'planning', 'running', 'needsInput', 'blocked', 'partial', 'retesting']);
+const INTERNAL_OPEN_GOAL_STATUSES = new Set(['stacked', 'planning', 'running', 'needsInput', 'needsApproval', 'blocked', 'partial', 'retesting']);
 
 function timeMs(value) {
   if (!value) return null;
@@ -1151,6 +1151,34 @@ export function parseGoalMessages(raw) {
   }).filter(Boolean);
 }
 
+// A Review describes one implementation attempt. Preserve the user's words
+// and measured worker output together, without deciding whether the request
+// was satisfied — that judgment remains with the reviewer.
+export function buildReviewAttemptLedger({ goal = null, tasks = [] } = {}) {
+  const time = (task) => Date.parse(task?.finishedAt || task?.startedAt || task?.createdAt || 0) || 0;
+  const own = (tasks ?? []).filter((task) => task?.goalId === goal?.id);
+  const settled = own.filter((task) => ['done', 'failed', 'interrupted', 'skipped'].includes(task?.status));
+  const outcome = [...settled].sort((a, b) => time(b) - time(a))[0] ?? null;
+  const requirements = [
+    goal?.text ? { id: `goal-${goal.id}-initial`, kind: 'initial', text: String(goal.text), at: goal.createdAt ?? null } : null,
+    ...own.filter((task) => task?.reply && task?.detail).map((task) => ({
+      id: task.attemptId || `task-${task.id}`,
+      kind: 'reply', text: String(task.detail), at: task.createdAt ?? null,
+    })),
+  ].filter(Boolean);
+  return {
+    attemptId: outcome?.attemptId || (outcome?.id != null ? `task-${outcome.id}` : null),
+    taskId: outcome?.id ?? null,
+    completedAt: outcome?.finishedAt ?? null,
+    requirements,
+    activeRequirement: requirements.at(-1) ?? null,
+    result: String(outcome?.result ?? '').slice(0, 8000),
+    changedFiles: [...new Set((outcome?.changedFiles ?? []).filter(Boolean))],
+    verification: { proof: outcome?.proof ?? null, tests: goal?.testResult ?? null },
+    run: outcome?.run ?? null,
+  };
+}
+
 // ---- which goal states allow EDIT / DELETE from the board ------------------
 // Masa (dogfood): 「急に入ったやつ(ゴール)が編集したり削除したりできない。
 // できるようにして」。Expanded beyond the original stacked/pending.
@@ -1176,7 +1204,7 @@ export function canEditGoal(status, startedAt) {
 // and dropping its queued tasks (server side). A goal already in review or
 // finished (done) is NOT deleted here — that's the Ledger's Archive/Revert
 // territory, and 'done' keeps its 409 as before.
-export const DELETABLE_GOAL_STATUSES = ['stacked', 'sending', 'pending', 'planning', 'needsInput', 'running', 'partial', 'failed', 'interrupted', 'blocked'];
+export const DELETABLE_GOAL_STATUSES = ['stacked', 'sending', 'pending', 'planning', 'needsInput', 'needsApproval', 'running', 'partial', 'failed', 'interrupted', 'blocked'];
 export function canDeleteGoal(status) {
   return DELETABLE_GOAL_STATUSES.includes(status);
 }
@@ -1474,7 +1502,7 @@ export function buildGoalPrBody({ lang, goal, goalTasks, files, owner, branch, p
 // 修正/追加/改善/調査 so an "追加したい" request never becomes "…の修正", and (c) a
 // failed/missing reply falls back to goal.text (no regression) — all without an
 // LLM call. The app's saReviewTitle reads reviewSummary.title first (PR-B).
-export function buildReviewSummaryPrompt({ lang, goalText, changed, reports }) {
+export function buildReviewSummaryPrompt({ lang, goalText, changed, reports, latestFeedback = '' }) {
   return [
     `あるゴールの実装が完了しました。レビューする人が一目で判断できるよう、やさしい${lang === 'ja' ? '日本語' : lang}で書いてください。`,
     '専門用語・長い説明は禁止。小学生でも分かる短い言い方。JSONのみ出力（前後に文章を書かない）:',
@@ -1483,6 +1511,7 @@ export function buildReviewSummaryPrompt({ lang, goalText, changed, reports }) {
       + '例:「Xの不具合」→「Xの修正」/「Yを追加したい」→「Yの追加」/「Zを速くしたい」→「Zの改善」/「なぜWか調べて」→「Wの調査」。',
     '',
     `ゴール: ${(goalText ?? '').slice(0, 300)}`,
+    latestFeedback ? `最新のユーザー返信（この再レビューで満たすべき条件。初回の結果ではなく、これへの対応をchanged/checkに書く）:\n${String(latestFeedback).slice(0, 1000)}` : null,
     `変更ファイル: ${(changed ?? []).join(', ') || '(なし)'}`,
     `ワーカー報告:\n${(reports ?? '').slice(0, 1500)}`,
   ].join('\n');
@@ -1650,6 +1679,8 @@ export function countInfraFlakes(text = '') {
     /Navigation timeout of \d+\s*ms exceeded/g,        // puppeteer page.goto
     /test timed out after \d+\s*ms/g,                  // node --test per-test timeout
     /\bTarget closed\b/g,                              // puppeteer: browser died mid-test
+    /puppeteer-core is not installed next to the Manager/g, // worktree/node_modules missing
+    /Cannot find module ['"]puppeteer-core['"]/g,          // same dependency failure, alternate Node wording
   ];
   return pats.reduce((n, re) => n + (String(text ?? '').match(re) || []).length, 0);
 }
@@ -1906,6 +1937,7 @@ export function goalGroupStatus(goalStatus, startedAt) {
   if (goalStatus === 'folded') return 'review'; // PRD §6.5: rendered inside the Review section, not Done
   if (goalStatus === 'blocked') return 'attn'; // §4E: blocked is Attention, not Done
   if (goalStatus === 'needsInput') return 'todo'; // smart intake: awaiting an answer
+  if (goalStatus === 'needsApproval') return 'needsYou'; // paused for a human go/no-go, never Review
   if (goalStatus === 'rejected') return 'rejected'; // 07-20c: Reject を溜める独立欄（Done ではない）
   return 'done'; // includes 'reverted' (terminal)
 }
@@ -1954,9 +1986,32 @@ export function goalChip(goalStatus, columns, startedAt) {
   if (goalStatus === 'folded') return ['FOLDED', '']; // PRD §6.5: settled, not failed — no red/badge class, matches 'done'/'reverted'
   if (goalStatus === 'blocked') return ['BLOCKED', 'blocked']; // §4E: distinct state (reason + reverify), not FAIL/Done
   if (goalStatus === 'needsInput') return ['確認待ち', 'blocked']; // smart intake: awaiting an answer
+  if (goalStatus === 'needsApproval') return ['NEEDS YOU', 'blocked'];
   return ['FAIL', ''];
 }
 
+// Workers can pause without pretending their unfinished work is reviewable.
+// The marker is deliberately explicit and machine-readable; ordinary prose
+// containing words such as "confirm" must never stop a goal by accident.
+// Invalid/missing fields fail open as a normal worker result.
+export function parseApprovalRequest(text) {
+  const source = String(text ?? '');
+  const match = source.match(/<manager-approval>\s*([\s\S]*?)\s*<\/manager-approval>/i);
+  if (!match) return null;
+  try {
+    const raw = JSON.parse(match[1]);
+    const question = String(raw.question ?? '').trim().slice(0, 500);
+    const action = String(raw.action ?? '').trim().slice(0, 120);
+    if (!question || !action) return null;
+    const scope = Array.isArray(raw.scope)
+      ? raw.scope.map((x) => String(x).trim().slice(0, 300)).filter(Boolean).slice(0, 20)
+      : [];
+    const resumeInstruction = String(raw.resumeInstruction ?? '').trim().slice(0, 1000)
+      || `承認されたため「${action}」を実行し、元のタスクを最後まで完了してください。`;
+    return { type: 'approval_required', action, question, scope, resumeInstruction };
+  } catch { return null; }
+}
+
 // Group a flat task list under their parent goal for the Tasks sidebar, so
 // a goal like "この体験を実現したい" shows as one collapsible unit instead of
 // its tasks scattered flat in the list. Each group carries a done/total
@@ -2035,6 +2090,9 @@ export function buildReviewDigest({ goals, tasks } = {}) {
       testResult: g.testResult ?? null,
       verificationInfraError: g.verificationInfraError ?? null,
       proof: goalTasks.find((t) => t.proof)?.proof ?? null,
+      // Worker-declared partial/not-addressed requirements are not a failed
+      // verdict. Count them so a human does not miss them in the compact list.
+      incompleteRequirementCount: goalTasks.reduce((n, t) => n + (t.requirementEvidence?.requirements ?? []).filter((r) => ['partial', 'not-addressed'].includes(r.status)).length, 0),
       plan,
       planText: plan ? (g.planText ?? '') : null,
     };
@@ -2047,6 +2105,7 @@ export function buildReviewDigest({ goals, tasks } = {}) {
       if (existing.proof == null) existing.proof = row.proof;
       if (existing.testResult == null) existing.testResult = row.testResult;
       if (existing.verificationInfraError == null) existing.verificationInfraError = row.verificationInfraError;
+      existing.incompleteRequirementCount += row.incompleteRequirementCount;
       continue;
     }
     if (row.pr) byPr.set(row.pr, row);
@@ -2804,6 +2863,13 @@ export function buildEphemeralSeedLines(at = '2026-01-01T00:00:00.000Z') {
     ...reqTasks,
     { kind: 'goal', id: 16, projectId: 'default', text: '[verify fixture] 失敗したTo Do行', status: 'partial', wantsPR: false, model: 'sonnet', images: [], plan: ['失敗行のアクション確認'], pr: undefined, createdAt: at },
     { kind: 'task', id: 17, num: 9, goalId: 16, projectId: 'default', title: '失敗行のアクション確認', detail: '', passCondition: null, model: 'sonnet', status: 'failed', createdAt: at, result: 'テスト失敗 4 件', changedFiles: ['app/index.html'], secs: 1, proof: null, usage: null, finishedAt: at },
+    // タスク461: 「レビューの青丸を確認できるようなデータ」— 'default' プロジェクトは
+    // MANAGER_SEED_RUNNING_FIXTURE の実行中ゴールも持つため、レールの丸は常にスピン
+    // (running優先)になり、青いレビュー丸を単体では確認できない。'default-2' には
+    // running なゴールが無いので、ここにレビュー待ちゴールを置けば、そのプロジェクトの
+    // レール丸は spin せず素直にReview Bになる。
+    { kind: 'goal', id: 18, projectId: 'default-2', text: '[verify fixture] レビュー待ちゴール(running無し・レール青丸確認用)', status: 'review', wantsPR: true, model: 'sonnet', images: [], plan: ['fixture task'], pr: 'https://github.com/example/verify-fixture/pull/3', createdAt: at },
+    { kind: 'task', id: 19, num: 1, goalId: 18, projectId: 'default-2', title: 'fixture task', detail: '', passCondition: '', model: 'sonnet', status: 'done', createdAt: at, result: 'ok', changedFiles: ['fixture.txt'], secs: 1, proof: { pass: true, detail: 'ok', gif: null, png: null, mp4: null }, usage: null, finishedAt: at },
   ];
 }
 
@@ -3120,9 +3186,20 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
     // TASK/DETAIL below is a paraphrase from the planner — when they differ,
     // follow the user's words.
     goal?.text ? `USER'S REQUEST (verbatim — the source of truth; deliver exactly this):\n${goal.text}` : '',
+    // A review reply is a new, explicit acceptance condition. It used to be
+    // present only as TASK DETAIL below the inherited session context, which
+    // made it easy for a resumed agent to keep describing the first attempt.
+    // Put it next to the original request and make the precedence explicit:
+    // the interface transports the user's words, it does not reinterpret them.
+    task?.reply && task?.detail ? `CURRENT USER FOLLOW-UP (verbatim — these are the acceptance conditions for THIS re-review):\n${task.detail}` : '',
+    task?.reply && task?.userRequirementHistory?.length ? `EARLIER USER REQUIREMENTS (verbatim task records; Manager-internal retry messages are excluded):\n${task.userRequirementHistory.map((r, i) => `${i + 1}. [${r.status}] ${r.text}`).join('\n')}` : '',
+    // Falls back to the unfiltered attempt log only when the user-authored list
+    // is absent, so the worker never sees the same items twice.
+    task?.reply && !task?.userRequirementHistory?.length && task?.requirementHistory?.length ? `REQUIREMENT HISTORY (facts from earlier attempts; do not silently drop an item):\n${task.requirementHistory.map((r, i) => `${i + 1}. [${r.status}] ${r.text}`).join('\n')}` : '',
     runContext ? (typeof runContext === 'string' ? runContext : buildWorkerContextBlock(runContext)) : '',
     '',
     `TASK ${task.num}: ${task.title}`,
+    `REQUIREMENT ID FOR THIS RUN: task-${task.id ?? task.num}.`,
     task.detail ? `DETAIL: ${task.detail}` : '',
     `WORKSPACE MODE: ${workspaceMode}.`,
     goal?.images?.length ? `ATTACHED IMAGES (absolute paths, use the Read tool to view them): ${goal.images.join(' ')}` : '',
@@ -3144,6 +3221,9 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
     // instead of expanding scope"): the request, not our fear of scope, sets
     // the size of the work.
     '- Deliver what the request actually asks for. If it asks for multiple options, N variations, a visual, an artifact, or a full exploration, produce ALL of them — do not reduce it to one, and do not cut it short to finish faster.',
+    '- At the end of your final report, add exactly one line: MANAGER_REQUIREMENT_EVIDENCE: {"requirements":[{"id":"task-<this run id>","status":"addressed|partial|not-addressed","evidence":"file, test, or Open it evidence","files":["changed/path"]}]}. This labels evidence for Review; it is not self-approval. Report partial or not-addressed honestly.',
+    '- For a review follow-up, implement and report against CURRENT USER FOLLOW-UP above. Do not claim the first attempt as this attempt\'s result. Do not change shared theme tokens, seed/demo data, or unrelated indicators merely to make a manual check easier unless the user explicitly requested those changes; put test fixtures in test files only.',
+    '- Requirement history is context, not a Manager priority decision. Keep prior items visible; if an item remains unaddressed, say so plainly rather than omitting it.',
     '- デザイン案・UI案・レイアウト案を求められた場合は、文章だけで完了にしない。画像を使わない案でも、比較して確認できるHTML等のartifactをプロジェクト内に作り、実際に開ける .manager-run.json を残す。',
     '- Keep the work bounded to the request: inspect only files needed; avoid broad repository scans unless they are necessary.',
     "- You may use the user's installed skills (Skill tool, e.g. /galda-doc) when one clearly fits the request.",
@@ -3157,6 +3237,9 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
     // stop re-checking; but the deliverables above must still be produced in full.
     '- 動くと確認できたら十分（無駄な再検証でログを漁り続けない）。ただし依頼された成果物は削らず出し切ること。',
     '- Do not commit (git is handled by the Manager).',
+    '- 原則は合理的な仮定で最後まで進める。どうしても人の明示許可なしに実行できない操作が残った場合だけ、完成物のReviewとして提出せず、最終出力に次の機械可読マーカーを1個含めて停止する:',
+    '<manager-approval>{"action":"短い操作ID","question":"人に確認する1問","scope":["対象"],"resumeInstruction":"承認後に同じセッションで行う具体的な続き"}</manager-approval>',
+    '- 上のマーカーは、作業完了後の見た目確認・通常のレビュー・自分で安全に仮定できる質問には使わない。Manager が承認後に同じゴールを再開する。',
     // How the human actually looks at the result. We never guess the run
     // command ourselves — guessing means branching on project type (Next.js vs
     // Python vs …), and that is the "smart PM" behaviour this layer must not
@@ -3222,7 +3305,7 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
     // why, so the goal reads as unfinished/blocked rather than as a completed PR. A
     // human or the CTO will re-dispatch the work (e.g. via manager_reply) once the
     // blocker clears.
-    '- あなたは1回きりのバッチ実行であり、このプロセスが終了した後に自動で再開することはない。レーン/ファイルの競合など自分では解けない理由で着手できない場合、「空き次第自動で再開します」のような、この実行の外では実現できない約束を書かないこと。何をどう変更できなかったか（コード変更0件であること）を正直に書き、ゴールが完了したPRではなく未完了/ブロックとして扱われるように終えること。再開は人間かCTOが改めて指示する。',
+    '- この実行は終了後に勝手には再開しない。レーン/ファイルの競合など承認では解けないブロッカーに対して「空き次第自動で再開します」と約束しないこと。人の明示許可で解ける場合だけ上記 manager-approval を使い、それ以外は何を変更できなかったかを正直に報告する。',
   ].filter(Boolean).join('\n');
 }
 
@@ -3240,6 +3323,32 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
 // button that opens the wrong thing is the fabricated-proof bug all over again.
 export const RUN_DECL_FILE = '.manager-run.json';
 
+// Structured evidence is a transport record, not a Manager verdict. A worker
+// may append it to its normal final report so each requirement remains visible
+// with the attempt that addressed it.
+export function parseRequirementEvidence(raw) {
+  const m = String(raw ?? '').match(/MANAGER_REQUIREMENT_EVIDENCE:\s*(\{[^\n]+\})/);
+  if (!m) return null;
+  let value;
+  try { value = JSON.parse(m[1]); } catch { return null; }
+  if (!Array.isArray(value?.requirements)) return null;
+  const requirements = value.requirements.slice(0, 20).map((entry) => {
+    const id = typeof entry?.id === 'string' ? entry.id.trim().slice(0, 120) : '';
+    const status = ['addressed', 'partial', 'not-addressed'].includes(entry?.status) ? entry.status : '';
+    const evidence = typeof entry?.evidence === 'string' ? entry.evidence.trim().slice(0, 500) : '';
+    // `files` is optional: a worker that names no files gets no key, so the
+    // record stays exactly what was declared.
+    const files = Array.isArray(entry?.files) ? [...new Set(entry.files.map((f) => typeof f === 'string' ? f.trim().slice(0, 300) : '').filter(Boolean))].slice(0, 30) : [];
+    return id && status ? { id, status, evidence, ...(files.length ? { files } : {}) } : null;
+  }).filter(Boolean);
+  return requirements.length ? { requirements } : null;
+}
+
+export function unmappedChangedFiles(changedFiles = [], evidence = null) {
+  const claimed = new Set((evidence?.requirements ?? []).flatMap((r) => r.files ?? []));
+  return [...new Set((changedFiles ?? []).filter(Boolean))].filter((file) => !claimed.has(file));
+}
+
 // A declaration that parsed as JSON but breaks a rule (bad url scheme, an
 // unsafe `file` path) is NOT the same failure as unparseable JSON — the first
 // is a worker mistake with a specific, sayable reason; the second is a broken
@@ -3742,6 +3851,7 @@ export function inboxKind(goal) {
   if (isPlanReview(goal)) return 'plan';
   if (goal.status === 'review') return goal.reviewOnly ? 'decision' : 'review';
   if (goal.status === 'needsInput') return 'question';
+  if (goal.status === 'needsApproval') return 'approval';
   if (goal.status === 'blocked') return 'blocked';
   return null;
 }
@@ -3754,6 +3864,7 @@ export function inboxKind(goal) {
 // 質問し続ける（running に戻す分岐にも入らない）。弾かれるより悪い＝黙って違うことを
 // するので、ここで行き先を固定する。
 export const INBOX_ACTIONS = {
+  approval: 'decide in the Manager UI, or call manager_decide_approval(decision). Proceed/edit resumes only the approved action; cancel stops it. This is NOT Review.',
   plan: 'manager_approve → runs the approved plan (does NOT file to Done) / manager_dismiss → sends it back',
   decision: 'manager_approve → files it to Done / manager_dismiss → drops it (no worker runs either way)',
   review: 'manager_approve → files it to Done / manager_dismiss(text) → sends it back to the worker with your fix note',
@@ -3784,6 +3895,11 @@ export function buildReviewInbox(goals = [], { project } = {}) {
             options: Array.isArray(opts) ? opts.map((o) => String(o).replace(/\s+/g, ' ').trim()).slice(0, 8) : [],
           }
           : null,
+        approval: kind === 'approval' ? {
+          action: String(g.approvalRequest?.action ?? '').slice(0, 120),
+          question: String(g.approvalRequest?.question ?? g.text ?? '').replace(/\s+/g, ' ').trim().slice(0, 200),
+          scope: Array.isArray(g.approvalRequest?.scope) ? g.approvalRequest.scope.slice(0, 20) : [],
+        } : null,
         blockedReason: kind === 'blocked'
           ? String(g.blocked?.reason ?? 'blocked').replace(/\s+/g, ' ').trim().slice(0, 120)
           : null,
@@ -3792,6 +3908,7 @@ export function buildReviewInbox(goals = [], { project } = {}) {
 }
 
 const INBOX_KIND_HEADING = {
+  approval: 'NEEDS YOUR APPROVAL',
   question: 'NEEDS YOUR ANSWER',
   plan: 'PLAN — approve to execute',
   decision: 'DECISION — no worker ran',
@@ -3799,7 +3916,7 @@ const INBOX_KIND_HEADING = {
   blocked: 'BLOCKED',
 };
 // 止まっているものから先に。question/plan は答えるまで前に進まないので上。
-const INBOX_KIND_ORDER = ['question', 'plan', 'decision', 'review', 'blocked'];
+const INBOX_KIND_ORDER = ['approval', 'question', 'plan', 'decision', 'review', 'blocked'];
 
 // kind ごとにまとめ、「次の一手」は1グループ1回だけ書く。これを item ごとに繰り返すと、
 // レビューが20件溜まった時に同じ一文を20回読ませることになる——読み手は Claude なので
@@ -3822,6 +3939,11 @@ export function formatReviewInbox(items = []) {
           lines.push(`  Q: ${it.question.text}`);
           if (it.question.options.length) lines.push(`  options: ${it.question.options.join(' | ')}`);
         }
+        if (it.approval) {
+          lines.push(`  Q: ${it.approval.question}`);
+          lines.push(`  action: ${it.approval.action || 'continue'}`);
+          if (it.approval.scope.length) lines.push(`  scope: ${it.approval.scope.join(' | ')}`);
+        }
       }
       return lines.join('\n');
     });

package/engine/mcp.mjs

@@ -37,6 +37,7 @@ const fmtGoal = (g, tasks) => {
   const gt = tasks.filter((t) => t.goalId === g.id);
   const lines = gt.map((t) => `  - [${t.status}] ${t.num} ${t.title}${t.secs ? ` (${t.secs}s)` : ''}${t.proof ? ` proof:${t.proof.pass ? 'PASS' : 'FAIL'}` : ''}`);
   return [`goal ${g.id} [${g.status}] (${g.projectId}) ${g.text.replace(/\s+/g, ' ').slice(0, 120)}`,
+    g.approvalDecision ? `  decision: ${g.approvalDecision.decision}${g.approvalDecision.instruction ? ` — ${g.approvalDecision.instruction}` : ''}` : '',
     g.pr ? `  PR: ${g.pr}` : '', ...lines].filter(Boolean).join('\n');
 };
 
@@ -58,7 +59,7 @@ const TOOLS = [
   },
   {
     name: 'manager_request_review',
-    description: 'Put an item in the Manager Review lane for a human to look at / approve — WITHOUT running any worker. Use for "please confirm / decide / eyeball this" that needs human judgment, not implementation (a design call, a go/no-go, a thing you already did and want signed off). It appears in the Review column; the human taps Approve (files it to Done) or Dismiss (drops it). No Claude worker is spawned, no PR is opened. Prefer this over manager_create_goal whenever you only need a decision, not work.',
+    description: 'Put a FINISHED deliverable in Review for acceptance. Use only after the work/result exists and can be evaluated. Do not use for permission to continue an operation; use manager_request_approval for that.',
     inputSchema: { type: 'object', properties: {
       project: { type: 'string', description: 'project id (e.g. galda1, galda2)' },
       text: { type: 'string', description: 'the one-line ask / what to decide (shown as the requirement)' },
@@ -70,6 +71,23 @@ const TOOLS = [
       return `Filed for review: goal ${g.id} [${g.status}] in ${g.projectId}. No worker started — waiting on a human Approve/Dismiss.`;
     },
   },
+  {
+    name: 'manager_request_approval',
+    description: 'Pause for a human go/no-go before continuing an operation. Creates a Needs you card, never a Review item. Poll manager_status for proceed/edit/cancel, then continue only within the approved scope.',
+    inputSchema: { type: 'object', properties: {
+      project: { type: 'string', description: 'project id (e.g. galda1, galda2)' },
+      question: { type: 'string', description: 'one short approval question shown to the human' },
+      action: { type: 'string', description: 'stable short action id, e.g. git_commit_rebase_create_pr' },
+      scope: { type: 'array', items: { type: 'string' }, description: 'exact files, branch, systems, or effects covered by the approval' },
+      resumeInstruction: { type: 'string', description: 'what the requesting agent will do after proceed' },
+    }, required: ['project', 'question', 'action'] },
+    run: async (a) => {
+      const g = await api('/api/tasks', { method: 'POST', headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({ projectId: a.project, text: a.question, approval: true, action: a.action,
+          scope: a.scope, resumeInstruction: a.resumeInstruction, source: 'mcp' }) });
+      return `Approval requested: goal ${g.id} [${g.status}] in ${g.projectId}. Do not continue yet; poll manager_status until decision is proceed, edit, or cancel.`;
+    },
+  },
   {
     name: 'manager_status',
     description: 'List goals and their tasks (status, proof, PR links). Optionally filter by project id.',
@@ -145,6 +163,20 @@ const TOOLS = [
       return `Answered: goal ${a.goalId} is now [${g.status}] in ${g.projectId}.`;
     },
   },
+  {
+    name: 'manager_decide_approval',
+    description: 'Resolve a Needs you approval request. proceed resumes the exact requested action, edit replaces its allowed scope/instruction and resumes, cancel stops it. Never use this for a finished Review item.',
+    inputSchema: { type: 'object', properties: {
+      goalId: { type: 'number' },
+      decision: { type: 'string', enum: ['proceed', 'edit', 'cancel'] },
+      instruction: { type: 'string', description: 'required for edit: the revised allowed scope/instruction' },
+    }, required: ['goalId', 'decision'] },
+    run: async (a) => {
+      const g = await api(`/api/goals/${a.goalId}/approval`, { method: 'POST', headers: { 'content-type': 'application/json' },
+        body: JSON.stringify({ decision: a.decision, instruction: a.instruction }) });
+      return `Approval ${a.decision}: goal ${a.goalId} is now [${g.status}]${g.replyTaskId ? `; continuation task ${g.replyTaskId} queued` : ''}.`;
+    },
+  },
 ];
 
 function reply(id, result) { process.stdout.write(JSON.stringify({ jsonrpc: '2.0', id, result }) + '\n'); }