@galda/cli 0.10.41 → 0.10.42

package/app/index.html

@@ -1218,6 +1218,14 @@
   .gearpop .srow.dis{opacity:.38;pointer-events:none}
   .gearpop .srow .sl{flex:1;min-width:0;font-size:12.5px;color:var(--ink)}
   .gearpop .srow .sd{display:block;font-size:10.5px;color:var(--ink3);margin-top:2px}
+  /* Project rules read-only view (§1.5a): a quiet framed panel under its row —
+     mono filename label + the file body, scrollable. Reuses --canvas/--hair so it
+     matches the MCP box; no new surface invented. */
+  .gearpop .rulesview{margin:2px 0 9px;display:flex;flex-direction:column;gap:8px}
+  .gearpop .rulesfile{background:var(--canvas);border-radius:8px;overflow:hidden}
+  .gearpop .rulesname{font:600 9px/1 var(--mono);letter-spacing:.08em;text-transform:uppercase;color:var(--ink3);padding:8px 10px 0}
+  .gearpop .rulestext{margin:6px 0 0;padding:0 10px 10px;max-height:220px;overflow:auto;font:400 11px/1.6 var(--mono);color:var(--ink2);white-space:pre-wrap;word-break:break-word}
+  .gearpop .rulesempty{color:var(--ink3);font-style:italic}
   /* MCP tab: live status dot + connect command + quiet explanation */
   .gearpop .mcpstat{display:flex;align-items:center;gap:8px;font:600 11px/1 var(--ui);color:var(--ink);padding:2px 0 10px}
   .gearpop .mcpstat i{width:7px;height:7px;border-radius:50%;background:var(--green);box-shadow:0 0 6px rgba(var(--green-rgb),.5)}
@@ -3276,7 +3284,7 @@ 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: 'default', editing: null, attach: [], connectCommand: '', models: ['sonnet'], agentModels: null, agentEfforts: null, agents: ['claude-code', 'codex'], expanded: new Set(), openThreads: new Set(), queueOrder: {}, externalActivity: [], selectedGoal: null, workflowColumns: {}, wfDraft: null, reviewDefinitions: {}, rdDraft: 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,
+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: 'default', editing: null, attach: [], connectCommand: '', models: ['sonnet'], agentModels: null, agentEfforts: null, agents: ['claude-code', 'codex'], expanded: new Set(), openThreads: new Set(), queueOrder: {}, externalActivity: [], selectedGoal: null, workflowColumns: {}, wfDraft: null, reviewDefinitions: {}, rdDraft: 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,
   // per-message send overrides set via slash-commands (reset after each send,
   // except model which is sticky in localStorage). palette = open command list.
   msg: { mode: null, model: null, effort: null, skill: null, later: false, review: false },
@@ -10526,8 +10534,44 @@ function gearSetHTML(){
     <button class="optchip${lang === 'en' ? ' on' : ''}" data-gslang="en">EN</button>
     <button class="optchip${lang === 'ja' ? ' on' : ''}" data-gslang="ja">JA</button></div>
   ${billingRow}
+  ${gearRulesHTML()}
   <div class="srow dis"><div class="sl">Data folder<span class="sd">Runs are stored in the engine's chat-runs folder on the host</span></div></div>`;
 }
+// ONE-CONVERSATION §1.5a: the project's working rules — the file(s) the connected
+// AI actually reads (Claude Code → CLAUDE.md, Codex → AGENTS.md). Read-only: Galda
+// does NOT own these rules, it only surfaces them (editing is done in an editor or
+// by asking the AI in chat). §3 段階開示 — when the project has no such file the row
+// is absent entirely (never an empty container; the "ask once" flow is §1.5b).
+function gearRulesHTML(){
+  const r = state.projectRules[state.active];
+  if (!r || !r.files || !r.files.length) return '';
+  const names = r.files.map((f) => f.name).join(' · ');
+  const view = state.rulesOpen
+    ? `<div class="rulesview">${r.files.map((f) =>
+        `<div class="rulesfile"><div class="rulesname">${esc(f.name)}</div>`
+        + `<pre class="rulestext">${esc(f.content) || '<span class="rulesempty">(empty file)</span>'}</pre></div>`).join('')}</div>`
+    : '';
+  return `
+  <div class="srow"><div class="sl">Project rules<span class="sd">What this project tells the AI — read-only · ${esc(names)}</span></div>
+    <button class="optchip${state.rulesOpen ? ' on' : ''}" data-grules="1">${state.rulesOpen ? 'Hide' : 'View'}</button></div>
+  ${view}`;
+}
+// Fetch the active project's rules files once (cached per project), then rebuild
+// the settings panel so the row appears. The browser can't read the host FS, so
+// this is the only way in — GET /api/projects/:id/rules (engine, §1.5a).
+let rulesFetching = null;
+async function loadProjectRules(){
+  const pid = state.active;
+  if (state.projectRules[pid] || rulesFetching === pid) return;
+  rulesFetching = pid;
+  try {
+    const r = await fetch(withKey(`/api/projects/${pid}/rules`));
+    state.projectRules[pid] = r.ok ? await r.json() : { files: [] };
+  } catch { state.projectRules[pid] = { files: [] }; }
+  finally { rulesFetching = null; }
+  // Only rebuild if the settings tab is still the one on screen.
+  if ($('gearPop') && custTab === 'settings') buildGearPop();
+}
 // MCP (v46): connection status dot + connectCommand + COPY + quiet explanation.
 // The dot is the app's own live engine link (SSE up/down) — the only real signal here.
 function gearMcpHTML(){
@@ -10578,6 +10622,9 @@ function buildGearPop(){
       try { await fetch(withKey('/api/signout'), { method: 'POST' }); recentlySignedOut = true; }
       finally { await refresh(); if ($('gearPop')) buildGearPop(); }
     };
+    const rulesBtn = pop.querySelector('[data-grules]');
+    if (rulesBtn) rulesBtn.onclick = () => { state.rulesOpen = !state.rulesOpen; buildGearPop(); };
+    loadProjectRules();
     return;
   }
   if (custTab === 'mcp') {
@@ -10711,6 +10758,7 @@ function custPlace(){ // flagship: top-anchored band beside the rail (v45 §2);
 }
 function custOpen(tab){
   custTab = tab || 'theme'; // v46 deep-link (flagship __custOpen(tab)): menu entries land on their tab
+  state.rulesOpen = false; // rules panel starts collapsed each time the gear opens
   custSnap = { cust: { ...cust }, theme: state.theme, band: fsUI.band };
   buildGearPop(); custPlace();
   $('gearPop').classList.add('open');

package/engine/lib.mjs

@@ -1135,6 +1135,26 @@ export function classifyTestGate({ current = [], baseline = [] } = {}) {
   return { newFailures, preExisting, blocked: newFailures.length > 0 };
 }
 
+// A manager test-run that was KILLED before it could finish is inconclusive, not
+// a red. When the machine is saturated (prod + several worker goals all standing
+// up real servers + headless Chrome), `npm test` gets killed at its execFile
+// timeout BEFORE node --test can print its "ℹ pass N / ℹ fail N" summary — so the
+// parse sees 0 passed and the killed process is counted as a single phantom
+// failure. That is the machine being too busy to run the suite, not the goal's
+// code breaking, and blocking a finished goal (with a PR) on it strands good work
+// (dogfood #602/#604, 2026-07-23: passed:0, "1 failing", every listed failure a
+// 30s timeout, no summary line).
+//
+// The signal is `timedOut` (execFile killed the process — killed/SIGTERM/ETIMEDOUT),
+// NOT the mere absence of a summary: a small suite that printed `not ok 1` and
+// exited 1 on its own is a REAL failure and must still block. We also require
+// passed===0 (a killed run that had already logged passes is unusual, but if any
+// test passed the suite clearly ran) and fail CLOSED on a real AssertionError that
+// shares the killed run.
+export function isInconclusiveTestRun({ ran = false, timedOut = false, passed = 0, realAssertion = false } = {}) {
+  return Boolean(ran) && Boolean(timedOut) && !realAssertion && passed === 0;
+}
+
 // The verification-gate decision (Masa 2026-07-19: "検証/proof は Failed を作らない").
 // PROOF IS ADVISORY: a UI goal whose screenshot could not be captured
 // (proofMissing) or whose pixel heuristic could not auto-confirm the target
@@ -1979,6 +1999,28 @@ export function hasGitChanges(porcelainStatus) {
   return Boolean(String(porcelainStatus ?? '').trim());
 }
 
+// ---- project rules files (ONE-CONVERSATION §1.5a) --------------------------
+// The files that hold a project's working rules — the instructions the connected
+// AI actually reads (Claude Code reads CLAUDE.md from cwd; Codex reads AGENTS.md).
+// Galda does NOT own these rules: it only surfaces whichever files exist,
+// read-only. Order is fixed so the settings row is stable, and both are shown
+// when both exist (never show one and let the other read as "missing").
+export const PROJECT_RULES_FILENAMES = ['CLAUDE.md', 'AGENTS.md'];
+
+// Given a project dir and a reader (dir,name)->string|null, return the rules
+// files that exist, in PROJECT_RULES_FILENAMES order. A file the reader can't
+// return (absent → null, or throws) is omitted, so the UI shows only what is
+// really there rather than an empty rules pane.
+export function collectProjectRules(dir, readFile) {
+  const files = [];
+  for (const name of PROJECT_RULES_FILENAMES) {
+    let content = null;
+    try { content = readFile(dir, name); } catch { content = null; }
+    if (typeof content === 'string') files.push({ name, content });
+  }
+  return files;
+}
+
 // ---- merge-time conflict detection -----------------------------------------
 // Parallel goals never corrupt each other's WORKING files (each runs in its
 // own worktree) — but two goals that edited the same repo path will still

package/engine/server.mjs

@@ -21,8 +21,9 @@ import { homedir } from 'node:os';
 import { fileURLToPath } from 'node:url';
 import { pathToFileURL } from 'node:url';
 import { needsProjectFolder, folderLabel, buildFolderQuestion, resolveFolderAnswer, classifyFolderTarget, buildFolderInitConfirm, isFolderInitConfirmed, needsReviewPreference, buildReviewPreferenceQuestion, resolveReviewPreferenceAnswer, routeQuestionAnswer, notUnderstoodNote, parseRelocalizedQuestion, classifyAuthProbe, authBannerText, makeSigninChallenge, parseClaimResponse } from './lib.mjs';
-import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, parseGitLog, diffNewCommits, replayQueueLog, reconcileOrphanGoals, reorderQueue, sortQueueByPriority, TASK_PRIORITIES, validateWorkflowColumns, DEFAULT_WORKFLOW_COLUMNS, validateReviewDefinition, DEFAULT_REVIEW_DEFINITION, buildEphemeralSeedLog, buildEphemeralSeedProjects, trimTaskActivityForState, buildAskContext, WORKER_TOOLS, workerPrompt, verifyCfAccessJwt, resolveIdentity, goalVisibleTo, clampParallelLimit, canStartMore, nextRunnableTasks, goalsConflict, detectConflicts, pickFoldTarget, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, classifyVerifyGate, buildExecutionPlan, buildContextHandoffSummary, buildFailurePostmortem, appendFailureMemory, latestFailurePolicy, classifyChangeRisk, checkRunBudget, usageBudgetTokens, isRateLimited, nextResumeDelay, checkFreeTierLimit, resolveEntitlement, resolveCachedEntitlement, FREE_TIER_LIMITS, QUEUED_GOAL_STATUSES, verifyLicenseToken, shouldEmitSetupCompleted, pickAnalyticsUid, shouldEmitFreeExhausted, detectRequestLanguage, isNothingVerifiable, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, RUN_DECL_FILE, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages } from './lib.mjs';
+import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, parseGitLog, diffNewCommits, replayQueueLog, reconcileOrphanGoals, reorderQueue, sortQueueByPriority, TASK_PRIORITIES, validateWorkflowColumns, DEFAULT_WORKFLOW_COLUMNS, validateReviewDefinition, DEFAULT_REVIEW_DEFINITION, buildEphemeralSeedLog, buildEphemeralSeedProjects, trimTaskActivityForState, buildAskContext, WORKER_TOOLS, workerPrompt, verifyCfAccessJwt, resolveIdentity, goalVisibleTo, clampParallelLimit, canStartMore, nextRunnableTasks, goalsConflict, detectConflicts, pickFoldTarget, hasTestRelevantChanges, parseFailingTestNames, classifyTestGate, isInconclusiveTestRun, classifyVerifyGate, buildExecutionPlan, buildContextHandoffSummary, buildFailurePostmortem, appendFailureMemory, latestFailurePolicy, classifyChangeRisk, checkRunBudget, usageBudgetTokens, isRateLimited, nextResumeDelay, checkFreeTierLimit, resolveEntitlement, resolveCachedEntitlement, FREE_TIER_LIMITS, QUEUED_GOAL_STATUSES, verifyLicenseToken, shouldEmitSetupCompleted, pickAnalyticsUid, shouldEmitFreeExhausted, detectRequestLanguage, isNothingVerifiable, classifyComposerIntentHeuristic, detectPauseIntent, buildIntentPrompt, parseIntentResponse, buildBoardSnapshot, validateBoardSnapshot, boardIsEmpty, decideBoardPull, resolveConnectedAgents, pickAvailableAgent, pickUtilityAgent, utilityModel, liveTakeoverDecision, shouldOpenBrowserOnBoot, goalTaskTitle, buildGoalTask, RUN_DECL_FILE, parseRunDeclaration, SHOT_DIR, collectShotNames, parseGoalAddress, REPLY_OUTCOMES, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages } from './lib.mjs';
 import { createSerialQueue } from './lib.mjs';
+import { collectProjectRules } from './lib.mjs';
 import { openPR } from './pr.mjs';
 import { runVerification, exerciseUi } from './verify.mjs';
 
@@ -2465,8 +2466,11 @@ const testRunQueue = createSerialQueue();
 // project's own `npm test` in the goal's worktree and parse node --test's
 // "ℹ pass N / ℹ fail N" (or TAP "# pass/# fail"). Returns {ran,passed,failed,ok}.
 function runTestsOnce(dir) {
+  // The timeout is env-overridable so tests can force the killed-run path fast;
+  // production keeps the 4-minute default.
+  const runTimeout = Number(process.env.MANAGER_TEST_TIMEOUT_MS) || 240000;
   return testRunQueue.run(() => new Promise((res) => {
-    execFile('npm', ['test'], { cwd: dir, timeout: 240000, env: { ...process.env }, maxBuffer: 8 * 1024 * 1024 }, (e, out = '', err = '') => {
+    execFile('npm', ['test'], { cwd: dir, timeout: runTimeout, env: { ...process.env }, maxBuffer: 8 * 1024 * 1024 }, (e, out = '', err = '') => {
       const text = `${out}\n${err}`;
       const n = (re) => { const m = text.match(re); return m ? Number(m[1]) : null; };
       const passed = n(/ℹ pass (\d+)/) ?? n(/# pass (\d+)/) ?? n(/(\d+) passing/) ?? 0;
@@ -2480,9 +2484,15 @@ function runTestsOnce(dir) {
       const timeoutFlakes = (text.match(/did not report ready in time/g) || []).length;
       const realAssertion = /AssertionError|Expected values to be|expected .+ (?:to |but )/i.test(text);
       const failed = realAssertion ? failedTotal : Math.max(0, failedTotal - timeoutFlakes);
+      // Was the run KILLED at the execFile timeout (machine too saturated to
+      // finish), rather than exiting on its own? execFile sets killed=true and
+      // signal='SIGTERM' on timeout. That is the inconclusive signal — a normal
+      // non-zero exit (a small suite that printed `not ok` and exited 1) is NOT
+      // killed and must still be treated as a real result (isInconclusiveTestRun).
+      const timedOut = Boolean(e && (e.killed || e.signal === 'SIGTERM' || e.code === 'ETIMEDOUT'));
       // Capture the failing test NAMES too (not just the count) so the gate can
       // diff them against the base's failures and block only on NEW ones.
-      res({ passed, failed, detail: testOutputExcerpt(text), failingNames: parseFailingTestNames(text) });
+      res({ passed, failed, timedOut, realAssertion, detail: testOutputExcerpt(text), failingNames: parseFailingTestNames(text) });
     });
   }));
 }
@@ -2501,11 +2511,16 @@ async function runProjectTests(dir) {
     r = {
       passed: Math.max(r.passed, r2.passed),
       failed: Math.min(r.failed, r2.failed),
+      // Only an ALL-killed pair is inconclusive: if EITHER attempt exited on its
+      // own (timedOut=false) the result is conclusive. realAssertion is OR-ed so
+      // a real red in either run still fails closed.
+      timedOut: Boolean(r.timedOut && r2.timedOut),
+      realAssertion: r.realAssertion || r2.realAssertion,
       detail: kept.detail || r.detail || r2.detail || '',
       failingNames: kept.failingNames ?? r.failingNames ?? [],
     };
   }
-  return { ran: true, passed: r.passed, failed: r.failed, ok: r.failed === 0, detail: r.detail ?? '', failingNames: r.failingNames ?? [] };
+  return { ran: true, passed: r.passed, failed: r.failed, ok: r.failed === 0, timedOut: !!r.timedOut, realAssertion: !!r.realAssertion, detail: r.detail ?? '', failingNames: r.failingNames ?? [] };
 }
 
 // Failing tests on the BASE (pre-change) checkout — the set the review gate
@@ -2946,6 +2961,16 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
   // Manager-shaped repos.
   const proofFailing = siblings.some((t) => t.proof && t.proof.pass === false);
   let testsFailing = goal.testResult.ran && goal.testResult.failed > 0;
+  // Inconclusive run (2026-07-23): the suite was killed before it could finish
+  // (0 passed, no summary, no real assertion) — the machine was too saturated to
+  // run the tests, not the code failing. Blocking a finished goal (with a PR) on
+  // that stranded #602/#604. Treat it like the infra-flake discount above: don't
+  // block, note it, and let a human re-run when the machine is idle.
+  if (testsFailing && isInconclusiveTestRun(goal.testResult)) {
+    testsFailing = false;
+    goal.testResult.inconclusive = true;
+    goalAct('Project tests could not complete — the suite was killed before finishing (0 passed, no real assertion), which means the machine was too saturated to run them, not a code failure. Treating as inconclusive, not blocking. Re-run the tests when the machine is idle to confirm.');
+  }
   // Baseline diff (Masa 2026-07-19): block ONLY on failures THIS change
   // introduced. A test already red on the base — e.g. persist-failure on clean
   // origin/main — is pre-existing noise and must not knock a proof-verified
@@ -3410,6 +3435,22 @@ const server = createServer(async (req, res) => {
     return json(res, 200, { repos: detectRepos(HOME), home: HOME });
   }
 
+  // The project's working rules (ONE-CONVERSATION §1.5a): the settings panel
+  // shows a read-only view of whichever of CLAUDE.md / AGENTS.md exist in the
+  // project dir — the instructions the connected AI actually reads. The browser
+  // can't touch the filesystem, so it reads them here. Filenames are fixed
+  // (collectProjectRules), so `project.dir` is the only path input and there is
+  // no traversal surface.
+  const rulesMatch = url.pathname.match(/^\/api\/projects\/([\w-]+)\/rules$/);
+  if (rulesMatch && req.method === 'GET') {
+    const project = projects.find((p) => p.id === rulesMatch[1]);
+    if (!project) return json(res, 404, { error: 'project not found' });
+    const files = collectProjectRules(project.dir, (dir, name) => {
+      try { return readFileSync(join(dir, name), 'utf8'); } catch { return null; }
+    });
+    return json(res, 200, { dir: project.dir, files });
+  }
+
   if (url.pathname === '/api/projects' && req.method === 'POST') {
     let body = '';
     req.on('data', (d) => { body += d; });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.41",
+  "version": "0.10.42",
   "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": {