@galda/cli 0.10.108 → 0.10.110

package/CLAUDE.md

@@ -26,12 +26,53 @@
   - 新しいロジックは可能な限り純関数として `engine/lib.mjs` 等に切り出し、`engine/test/` にテストを追加する。
   - テストが書けない変更（純UI等）は、その理由と手動確認手順を報告に書く。
 
+### テストの書き方（Masa体感「変に判断せずつなぐだけになった」を守る・2026-07-30 追加）
+
+1日で main を2回赤にした型・3日間気付かれなかった型が実在する。以下は**好みでなく再発防止**。
+
+- 🔴 **`app/index.html` のソース文字列を正規表現で照合するテストを新しく書かない。**
+  `align-items:center` や `outcome?.requirementEvidence?.…` という**式そのもの**を要求していたテストが、
+  実装の書き換えだけで赤くなった（2026-07-29 に PR #574／#584 の2回）。
+  **正規表現は「書き方を変えた」と「壊した」を区別できない。**
+  → **実際に開いて、描かれたものを読む**。ヘルパは `engine/test/helpers/serve-app.mjs`＋`helpers/browser.mjs`、
+  実例は `engine/test/review-delivery-label.test.mjs`／`review-requirement-coverage-ui.test.mjs`。
+- 🔴 **テストを足したら、製品を壊して赤くなることを必ず確認する**（1〜3通り）。落ちないテストは何も証明しない。
+  確認したことと壊し方を PR 本文に書く。
+- 🔴 **テストだけコミットして実装が無い状態を作らない。** 実例2件:
+  `5802056`（Reconnect のテストだけ main に入り実装が存在しなかった）／
+  **PR #396 が統合時に engine と app の実装を落とし、テストだけ入れた**（main が3日間、
+  存在しない機能のテストで赤かった。復元＝PR #586）。
+  → 複数レーンの成果を1本に統合する PR では、**テストだけでなく実装が入っているかを機械で照合する**
+  （`node tools/conflict-check.mjs` の `test-without-implementation`）。
+- **transient な状態をポーリングで待たない。** 偽ワーカーは数msで返るので `running` は掴めない。
+  **残る証跡**（verdict・キューの中身・保存された状態）を待つ。時計を待つテストは飽和で揺れる。
+- **赤を見つけたら、まず単体で再走**して「飽和フレーク」と「本物」を分ける。飽和で31秒かけて落ちるものが実在する
+  （`agent-picker-availability`）。素の `origin/main` と比べて自分の変更のせいかを判定する。
+- **古い open PR をマージする前に、今日の `main` に載せて実機で確認する。**
+  main が既に別実装で解決済みなのに残っていた PR が実在（#425/#435/#509・全てクローズ）。
+  → `node tools/conflict-check.mjs --pr <n>` の `already-in-main` 検査。
+
 ## デザイン（Masa指示 2026-07-06・必読）
 
 - **UI（app/index.html・app/theme.css 等）に触る前に、必ず `docs/design/DESIGN-RULES.md` を読むこと**。特に §1.5 レイアウトの憲法（10条）と §5.5 美学。条文に反する実装はせず、変えたい場合は理由を添えて仲田さんの決定を仰ぐ。
 - app/ のUIファイル編集時はhook（tools/design_rules_hook.sh）が自動でリマインドする。リマインドが出たら読んだ上で作業すること。
 - 実装時の詳細契約は `docs/design/CTO-HANDOFF.md`（CDO→CTO引き継ぎ書）に従う。**既存パーツで作れない新しい画面・コンポーネントが必要になったら、自作せず実装前にCDOへ依頼すること。**
 
+## プロダクトの不変条件（実装前に読む・2026-07-30 追加）
+
+**正典 = `docs/PRD.md` §0（薄いインターフェース＋4決定）と §0.1（信頼性の不変条件）。** 実装で迷ったらここに戻る。
+Galda の役割は**依頼を Claude Code / Codex に繋ぐこと**で、途中に自前の AI 判断を挟まない。要約すると:
+
+- **測っていない割合を主張しない**（キャッシュヒット率は節約ではない）。基準線が無い間は黙る。
+- **判定は必ず最新の試行のもの**（文章が最新・絵や判定が最初、を作らない。同じ形のバグが3か所にあった）。
+- **出荷していないものを成功として祝わない**（未マージでクローズされた PR を done にしない）。
+- **人が頼んでいないものを要求として数えない**（Manager 発の返信を要求カバレッジに載せない）。
+- **止まっているものは「なぜ」を言う**（`stacked` は動いても失敗してもいない＝外からは壊れて見える）。
+- **このマシンにできないことは先に言う**（欠けている実行ファイル名でなく、何ができなくなるかを言う）。
+- **助言と停止判断を混ぜない**（`npm test` を開始できないものだけが blocking。混ぜると「走らせていないスイートが0 failed」になる）。
+
+新しく「Galda が判断する」挙動を足したくなったら、**足す前に Masa の決定を仰ぐ**（§0 決定3の自動選択禁止と同じ理由）。
+
 ## 実装規律
 
 - 変更は最小限。周辺コードの流儀に合わせる。

package/app/index.html

@@ -2279,19 +2279,6 @@
   /* Changed/Tests aligns with the request TEXT, not the number/chevron — same 18px
      the fold body below uses, so both sit under the same column (Masa 2026-07-29). */
   .redesign .rqit .rqev{font-size:12px;color:var(--ink3);margin:5px 0 0;padding-left:18px}
-  /* Canonical requirement coverage: one compact ledger, not three competing
-     interpretations. Status is text + a small rule/mark (no badge surface),
-     and Manager checks stay visually distinct from worker claims. */
-  .redesign .rqcov{display:flex;flex-direction:column;margin-top:2px}
-  .redesign .rqcov-row{padding:11px 0;border-top:1px solid var(--hair)}
-  .redesign .rqcov-row:first-child{padding-top:0;border-top:0}
-  .redesign .rqcov-main{display:grid;grid-template-columns:minmax(0,1fr) auto;gap:14px;align-items:baseline}
-  .redesign .rqcov-text{font-size:13.5px;line-height:1.55;color:var(--ink)}
-  .redesign .rqcov-state{font:600 10px/1.4 var(--mono);letter-spacing:.02em;color:var(--ink3);white-space:nowrap}
-  .redesign .rqcov-state.verified{color:var(--green)}
-  .redesign .rqcov-state.failed{color:var(--danger)}
-  .redesign .rqcov-detail{margin-top:5px;font-size:11.5px;line-height:1.55;color:var(--ink3)}
-  .redesign .rqcov-detail b{font-weight:600;color:var(--ink2)}
   /* A long ask folds with the SAME .lacc pulldown Activity/Conversation use — chevron on
      the peek line, body is the genuine rest (no duplicate, no trailing "…": Masa
      2026-07-29 flagged both the vertical anchor line and the "…" as unnecessary — the
@@ -3498,7 +3485,7 @@ const savedTheme = (() => {
   if (v !== s) localStorage.setItem('theme', v);
   return v;
 })();
-const state = { repos: null, repoHome: '', /* GET /api/repos — the folders a project can point at (Context Bar) */ projects: [], goals: [], tasks: [], act: {}, actAt: {}, todos: {}, chatTurns: {}, active: localStorage.getItem('sel:project') || 'default', editing: null, attach: [], connectCommand: '', models: ['sonnet'], agentModels: null, agentEfforts: null, agents: ['claude-code', 'codex'], auth: null, expanded: new Set(), openThreads: new Set(), queueOrder: {}, externalActivity: [], selectedGoal: null, composerDetachedGoal: null, workflowColumns: {}, wfDraft: null, reviewDefinitions: {}, rdDraft: null, runningCounts: {}, parallelLimits: null, connectedApp: localStorage.getItem('sel:agent') || 'claude-code', reviewOpenGoal: null, reviewApprovals: {}, goalDetailOpen: null, qreplyOpen: new Set(), logCollapsed: localStorage.getItem('logCollapsed') !== '0', logEntries: [], layout: localStorage.getItem('layout') || 'flagship', theme: savedTheme, boardskin: localStorage.getItem('boardskin') || 'auto', summaryPattern: Number(localStorage.getItem('summaryPattern')) || 3, peraSpot: 0, skill: null, skillsCache: null, projectRules: {}, rulesOpen: false,
+const state = { repos: null, repoHome: '', /* GET /api/repos — the folders a project can point at (Context Bar) */ projects: [], goals: [], tasks: [], act: {}, actAt: {}, todos: {}, chatTurns: {}, active: localStorage.getItem('sel:project') || 'default', editing: null, attach: [], connectCommand: '', models: ['sonnet'], agentModels: null, agentEfforts: null, agents: ['claude-code', 'codex'], auth: null, expanded: new Set(), openThreads: new Set(), queueOrder: {}, externalActivity: [], selectedGoal: null, composerDetachedGoal: null, workflowColumns: {}, wfDraft: null, reviewDefinitions: {}, rdDraft: null, runningCounts: {}, parallelLimits: null, hostReadiness: [], connectedApp: localStorage.getItem('sel:agent') || 'claude-code', reviewOpenGoal: null, reviewApprovals: {}, goalDetailOpen: null, qreplyOpen: new Set(), logCollapsed: localStorage.getItem('logCollapsed') !== '0', logEntries: [], layout: localStorage.getItem('layout') || 'flagship', theme: savedTheme, boardskin: localStorage.getItem('boardskin') || 'auto', summaryPattern: Number(localStorage.getItem('summaryPattern')) || 3, peraSpot: 0, skill: null, skillsCache: null, projectRules: {}, rulesOpen: false,
   // per-message send overrides set via slash-commands (reset after each send,
   // except model which is sticky in localStorage). palette = open command list.
   // Scoped per project (msgByProject) so a skill/mode picked in one project can't leak
@@ -8790,6 +8777,7 @@ async function refresh(){
     }
     state.connectCommand = s.connectCommand ?? ''; state.models = s.models ?? ['sonnet']; state.agentModels = s.agentModels ?? null; state.agentEfforts = s.agentEfforts ?? null; state.agents = s.agents ?? ['claude-code', 'codex']; state.uiVersion = s.uiVersion;
     state.billing = s.billing ?? { licensed: false, email: null };
+    state.hostReadiness = s.hostReadiness ?? [];
     renderAppSelect(); // picker follows the server's installed-agent set; may correct a stale connectedApp
     renderAuthBanner(s.auth); // only the selected worker's auth can raise a banner
     setAgentModel(state.connectedApp);
@@ -9553,7 +9541,7 @@ function closeSummaryPera(){ $('summaryOverlay').classList.remove('show'); }
 // 4 archived / 5 reverted — session-local; the server round-trip is what
 // settles it (see saSync's "returned to review" reset).
 // ============================================================================
-const SA = { open: false, items: [], diffs: {}, actLog: {}, msgs: {}, convOpen: new Set(), shotBox: {}, detOpen: new Set(), diffOpen: new Set(), attachments: {}, focusGid: null, focusScrolled: false, focusTimer: null, cur: null, solo: null, lastMarkup: null };
+const SA = { open: false, items: [], diffs: {}, actLog: {}, msgs: {}, msgCounts: {}, convOpen: new Set(), shotBox: {}, detOpen: new Set(), diffOpen: new Set(), attachments: {}, focusGid: null, focusScrolled: false, focusTimer: null, cur: null, solo: null, lastMarkup: null };
 // Same convention as the existing ?dev=1 devbar flag — off unless someone opts in by URL.
 const SA_DEV = new URLSearchParams(location.search).get('dev') === '1';
 // task-971: 白黒(モノクロ)版のレールを ?railmono=1 で見比べられるようにする。
@@ -10118,7 +10106,13 @@ function saReqStackRows(it, attemptEvidence){
   const gk = it.goalId;
   const reqMsgs = SA.msgs[gk];
   const reqAsks = Array.isArray(reqMsgs) ? reqMsgs.filter((m) => m?.from === 'you' && String(m.text || '').trim()) : [];
-  const reqItems = reqAsks.length ? reqAsks : (it.goalText ? [{ text: it.goalText }] : []);
+  // The frozen attempt ledger is task-backed and available on first paint. Use it
+  // before the asynchronously loaded conversation so Request #2/#3 cannot vanish
+  // because the browser still holds an older messages cache.
+  const ledgerAsks = Array.isArray(it.attempt?.requirements)
+    ? it.attempt.requirements.filter((r) => ['initial', 'reply'].includes(r?.kind) && String(r.text || '').trim())
+    : [];
+  const reqItems = ledgerAsks.length ? ledgerAsks : reqAsks.length ? reqAsks : (it.goalText ? [{ text: it.goalText }] : []);
   const row = (m, idx) => {
     const text = String(m.text || '').trim();
     const isLatest = idx === reqItems.length - 1;
@@ -10275,18 +10269,6 @@ function saItemHtml(it, i){
   const summarySec = summaryText
     ? `<section class="rsec rsum"><div class="rsec-h">Summary</div><p class="rbody v">${esc(summaryText.slice(0, 200))}</p></section>`
     : '';
-  const coverage = (it.requirementCoverage || []).filter((r) => r.text);
-  const coverageRows = coverage.map((r) => {
-    const labels = { verified: 'Verified', failed: 'Failed', 'not-applicable': 'Not applicable', unverified: 'Not verified' };
-    const details = [
-      r.legacy ? 'Legacy item — no canonical evidence was recorded' : '',
-      r.checkFacts?.length ? `Manager: ${r.checkFacts.join(' · ')}` : '',
-      r.claims?.length ? `Worker claim: ${r.claims.join(', ')}` : '',
-      r.workerEvidence?.length ? r.workerEvidence.join(' · ') : '',
-    ].filter(Boolean);
-    return `<div class="rqcov-row" data-requirement-id="${esc(r.id)}"><div class="rqcov-main"><div class="rqcov-text">${esc(r.text)}</div><span class="rqcov-state ${esc(r.coverage)}">${esc(labels[r.coverage] || labels.unverified)}</span></div>${details.length ? `<div class="rqcov-detail">${details.map((detail, index) => index === 0 ? esc(detail) : `<br>${esc(detail)}`).join('')}</div>` : ''}</div>`;
-  }).join('');
-  const coverageSec = coverageRows ? `<section class="rsec"><div class="rsec-h">Requirement coverage</div><div class="rqcov">${coverageRows}</div></section>` : '';
   // ── Open it (Masa 2026-07-22): a plain section in the card's own typography — the same
   // heading + body shape as Request and Result, not a chip in another vocabulary. The
   // address the worker declared IS the control (press it → the process starts on demand →
@@ -10534,7 +10516,6 @@ function saItemHtml(it, i){
     ${reqSec}
     ${S('Open it', openBody + infraBody)}
     ${S('What to check', checkBody)}
-    ${coverageSec}
     ${S('Run state', interruptedBody)}
     ${S('PR', prRepairBody)}
     ${S('Also carries', carriesBody)}
@@ -10583,8 +10564,11 @@ async function saLoadActivityLog(taskId){
 // request writes an honest line instead of an empty box — the same rule the
 // implementation log follows.
 async function saLoadMessages(goalId){
-  if (!Number.isFinite(goalId) || SA.msgs[goalId] !== undefined) return;
+  if (!Number.isFinite(goalId)) return;
+  const expectedCount = Number(SA.items.find((x) => x.goalId === goalId)?.messageCount) || 0;
+  if (SA.msgs[goalId] !== undefined && SA.msgCounts[goalId] === expectedCount) return;
   SA.msgs[goalId] = null; // in flight — a second toggle must not re-fetch
+  SA.msgCounts[goalId] = expectedCount;
   let msgs = [];
   try {
     const r = await fetch(withKey(`/api/goals/${goalId}/messages`));
@@ -12540,6 +12524,7 @@ function gearSetHTML(){
   ${billingRow}
   ${stability}
   ${gearRulesHTML()}
+  ${hostReadinessRow()}
   <div class="srow dis"><div class="sl">Data folder<span class="sd">Runs are stored in the engine's chat-runs folder on the host</span></div></div>`;
 }
 async function loadStability(){
@@ -12552,6 +12537,25 @@ async function loadStability(){
 // does NOT own these rules, it only surfaces them (editing is done in an editor or
 // by asking the AI in chat). §3 段階開示 — when the project has no such file the row
 // is absent entirely (never an empty container; the "ask once" flow is §1.5b).
+// Two capabilities degrade silently depending on the machine: without Chrome
+// nothing can screenshot or verify in a browser, and without a native picker
+// "Open folder…" cannot open Finder. The server states them in /api/state
+// (hostReadiness); this is where a person can read them before spending an hour
+// treating the absence as a bug.
+//
+// Same row shape as Data folder — a host fact, no chip, nothing to press (§SETTINGS
+// 行=ラベル+説明). When everything is available it stays dimmed like its neighbour;
+// when something is missing the row is NOT dimmed, because a dimmed row at .38 is
+// exactly the thing nobody reads. No colour and no fill either way (§1.5-8: an
+// exception is a small inline element, never a painted surface) — this is a
+// capability note, not an error.
+function hostReadinessRow(){
+  const checks = state.hostReadiness || [];
+  if (!checks.length) return ''; // nothing measured → no row (§1.5-3)
+  const missing = checks.filter((c) => !c.ok);
+  const detail = (missing.length ? missing : checks).map((c) => c.detail).join(' · ');
+  return `<div class="srow${missing.length ? '' : ' dis'}"><div class="sl">This machine<span class="sd">${esc(detail)}</span></div></div>`;
+}
 function gearRulesHTML(){
   const r = state.projectRules[state.active];
   if (!r || !r.files || !r.files.length) return '';

package/engine/lib.mjs

@@ -1174,7 +1174,7 @@ export function buildReviewAttemptLedger({ goal = null, tasks = [] } = {}) {
   const outcome = [...settled].sort((a, b) => time(b) - time(a))[0] ?? null;
   const requirements = [
     goal?.text ? { id: `goal-${goal.id}-initial`, kind: 'initial', text: String(goal.text), at: goal.createdAt ?? null } : null,
-    ...own.filter((task) => task?.reply && task?.detail).map((task) => ({
+    ...own.filter((task) => task?.reply && task?.detail && (!task.replyFrom || task.replyFrom === 'you')).map((task) => ({
       id: task.attemptId || `task-${task.id}`,
       kind: 'reply', text: String(task.detail), at: task.createdAt ?? null,
     })),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.108",
+  "version": "0.10.110",
   "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": {