@galda/cli 0.10.43 → 0.10.44

package/app/index.html

@@ -4748,7 +4748,7 @@ function renderTasks(){
   }
   for (const inp of $('tasklist').querySelectorAll('[data-qreplyinput]')) {
     inp.onclick = (e) => e.stopPropagation();
-    inp.onkeydown = (e) => { if (e.key === 'Enter') { e.preventDefault(); sendQReply(inp.dataset.qreplyinput, inp); } };
+    inp.onkeydown = (e) => { if (e.key === 'Enter' && !e.isComposing) { e.preventDefault(); sendQReply(inp.dataset.qreplyinput, inp); } };
   }
   // Pending rows open the same detail dialog as Review (tap anywhere but the
   // Start/Delete buttons, which stop propagation in their own handlers).
@@ -4872,7 +4872,7 @@ function startGoalheadEdit(id, btn){
   const save = async () => { const t = inp.value.trim(); if (!t) { wrap.remove(); return; }
     const r = await fetch(withKey(`/api/goals/${id}`), { method: 'PUT', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ text: t }) });
     if (!r.ok) showErr('Cannot edit a running/finished goal — reply to append instead'); await refresh(); };
-  inp.onkeydown = (e) => { if (e.key === 'Enter') { e.preventDefault(); save(); } if (e.key === 'Escape') wrap.remove(); };
+  inp.onkeydown = (e) => { if (e.key === 'Enter' && !e.isComposing) { e.preventDefault(); save(); } if (e.key === 'Escape') wrap.remove(); };
 }
 // SL1 inline Reply: append an instruction to a started goal → POST /api/goals/:id/reply.
 function openGoalheadReply(id, btn){
@@ -4885,7 +4885,7 @@ function openGoalheadReply(id, btn){
   const send = async () => { const t = inp.value.trim(); if (!t) return;
     const r = await fetch(withKey(`/api/goals/${id}/reply`), { method: 'POST', headers: { 'content-type': 'application/json' }, body: JSON.stringify({ text: t }) });
     if (!r.ok) showErr('未開始のゴールは編集を使ってください'); await refresh(); };
-  inp.onkeydown = (e) => { if (e.key === 'Enter') { e.preventDefault(); send(); } if (e.key === 'Escape') wrap.remove(); };
+  inp.onkeydown = (e) => { if (e.key === 'Enter' && !e.isComposing) { e.preventDefault(); send(); } if (e.key === 'Escape') wrap.remove(); };
   snd.onclick = send;
 }
 
@@ -5778,7 +5778,7 @@ async function sendReviewReject(){
 $('reviewDismiss').onclick = () => sendReviewReject();
 $('reviewFixSend').onclick = () => { const t = $('reviewFixInput').value.trim(); if (t || reviewAttach.length) sendReviewDismiss(t); };
 $('reviewFixInput').addEventListener('keydown', (e) => {
-  if (e.key === 'Enter') { const t = $('reviewFixInput').value.trim(); if (t || reviewAttach.length) sendReviewDismiss(t); }
+  if (e.key === 'Enter' && !e.isComposing) { const t = $('reviewFixInput').value.trim(); if (t || reviewAttach.length) sendReviewDismiss(t); }
 });
 $('reviewFixImg').onclick = () => $('reviewFixFile').click();
 $('reviewFixFile').addEventListener('change', async () => {
@@ -6934,7 +6934,7 @@ function inlineFsRename(span, onCommit){
     if (changed) onCommit?.(v);
     span.removeEventListener('keydown', onKey); span.removeEventListener('keyup', onKeyUp); span.removeEventListener('blur', onBlur); };
   const onKey = (ev) => { ev.stopPropagation();
-    if (ev.key === 'Enter') { ev.preventDefault(); finish(true); span.blur(); }
+    if (ev.key === 'Enter' && !ev.isComposing) { ev.preventDefault(); finish(true); span.blur(); }
     else if (ev.key === 'Escape') { ev.preventDefault(); finish(false); span.blur(); } };
   const onKeyUp = (ev) => { ev.stopPropagation(); };
   const onBlur = () => finish(true);
@@ -9586,7 +9586,7 @@ $('ctxdirmenu').addEventListener('click', (e) => {
 });
 $('ctxdirmenu').addEventListener('keydown', (e) => {
   e.stopPropagation();
-  if (e.key === 'Enter' && e.target.id === 'ctxdirin') { e.preventDefault(); setProjectFolder(e.target.value); }
+  if (e.key === 'Enter' && !e.isComposing && e.target.id === 'ctxdirin') { e.preventDefault(); setProjectFolder(e.target.value); }
 });
 $('appbtn').onclick = (e) => { e.stopPropagation(); $('appmenu').classList.toggle('show'); };
 $('appmenu').addEventListener('click', (e) => {

package/engine/lib.mjs

@@ -1158,6 +1158,49 @@ export function isInconclusiveTestRun({ ran = false, timedOut = false, passed =
   return Boolean(ran) && Boolean(timedOut) && !realAssertion && passed === 0;
 }
 
+// Infra failures the review gate must NOT block real work on. A test that stands
+// up a real server/headless Chrome fails these ways under WHOLE-MACHINE
+// saturation (prod :4400 + several agent sessions + the baseline run all live at
+// once — which --test-concurrency=1 inside one suite cannot prevent, the load is
+// from OTHER processes): the server dies during boot ("server exited (code N)
+// before it was ready"), misses its window ("did not report ready in time"), or a
+// puppeteer wait/navigation times out. None is the goal's code regressing, and
+// they SHUFFLE to a different set of tests each run, so classifyTestGate's
+// name-diff can never subtract them. Count them so a run can discount them.
+export function countInfraFlakes(text = '') {
+  const pats = [
+    /server exited \(code \d+\) before it was ready/g, // test helper: server crashed on boot under load
+    /did not report ready in time/g,                   // test helper: server missed its startup window
+    /Waiting failed: \d+\s*ms exceeded/g,              // puppeteer waitForFunction/Selector
+    /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
+  ];
+  return pats.reduce((n, re) => n + (String(text ?? '').match(re) || []).length, 0);
+}
+
+// Decide how many of a test run's failures are REAL (should block Review) vs
+// infra flakes (countInfraFlakes above). The old runTestsOnce let a SINGLE
+// "expected … to" match anywhere in megabytes of output (realAssertion=true)
+// disable the infra discount for ALL failures — which is exactly why 93 boot
+// failures ("server exited (code 1) before it was ready", 650 passed) blocked
+// goals #616/#620 (2026-07-24): every failure was infra, yet realAssertion
+// tripped on stray text and none was discounted.
+//   - infraOnly: every counted failure is an infra flake (infraFlakes covers the
+//     whole count) → block on none, regardless of a stray assertion string. If a
+//     REAL assertion failure existed it would be an ADDITIONAL failure not covered
+//     by the infra count, so infraFlakes < failedTotal and we fall through.
+//   - a real assertion sharing a mixed run → fail CLOSED on the full count.
+//   - otherwise → discount the infra flakes from the total.
+export function classifyRunFailures({ failedTotal = 0, infraFlakes = 0, realAssertion = false } = {}) {
+  const ft = Math.max(0, Number(failedTotal) || 0);
+  const inf = Math.max(0, Number(infraFlakes) || 0);
+  if (ft <= 0) return { effectiveFailed: 0, infraOnly: false };
+  if (inf >= ft) return { effectiveFailed: 0, infraOnly: true };
+  if (realAssertion) return { effectiveFailed: ft, infraOnly: false };
+  return { effectiveFailed: Math.max(0, ft - inf), infraOnly: false };
+}
+
 // 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

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 } 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 { 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, countInfraFlakes, classifyRunFailures, 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';
@@ -2475,15 +2475,21 @@ function runTestsOnce(dir) {
       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;
       const failedTotal = n(/ℹ fail (\d+)/) ?? n(/# fail (\d+)/) ?? n(/(\d+) failing/) ?? (e ? 1 : 0);
-      // Don't block on INFRA flakes: the integration suite spawns real servers,
-      // and under load (prod + several worker goals running) a server can miss
-      // its startup window ("did not report ready in time"). That's not a code
-      // regression. BUT fail CLOSED: only discount those timeouts when NO real
-      // assertion failed. A genuine AssertionError sharing a run with a timeout
-      // must still block — never let arithmetic subtraction mask a real failure.
-      const timeoutFlakes = (text.match(/did not report ready in time/g) || []).length;
+      // Don't block on INFRA flakes: the integration suite spawns real servers +
+      // headless Chrome, and under WHOLE-MACHINE saturation (prod :4400 + several
+      // agent sessions + the baseline run) a server dies on boot ("server exited
+      // (code N) before it was ready"), misses its window ("did not report ready
+      // in time"), or a puppeteer wait times out — none a code regression, and
+      // they shuffle to different tests each run. countInfraFlakes tallies them;
+      // classifyRunFailures decides how many failures are REAL. Fail CLOSED: a
+      // real AssertionError sharing a MIXED run still blocks (infraFlakes <
+      // failedTotal). Only when EVERY failure is infra (infraOnly) do we discount
+      // despite a stray "expected … to" match — the coarse global realAssertion
+      // flag is exactly what stranded #616/#620 (93 boot failures, all discountable,
+      // yet realAssertion tripped on unrelated text and blocked them).
+      const infraFlakes = countInfraFlakes(text);
       const realAssertion = /AssertionError|Expected values to be|expected .+ (?:to |but )/i.test(text);
-      const failed = realAssertion ? failedTotal : Math.max(0, failedTotal - timeoutFlakes);
+      const { effectiveFailed: failed, infraOnly } = classifyRunFailures({ failedTotal, infraFlakes, realAssertion });
       // 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
@@ -2492,7 +2498,7 @@ function runTestsOnce(dir) {
       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, timedOut, realAssertion, detail: testOutputExcerpt(text), failingNames: parseFailingTestNames(text) });
+      res({ passed, failed, timedOut, realAssertion, infraOnly, infraFlakes, detail: testOutputExcerpt(text), failingNames: parseFailingTestNames(text) });
     });
   }));
 }
@@ -2516,11 +2522,16 @@ async function runProjectTests(dir) {
       // a real red in either run still fails closed.
       timedOut: Boolean(r.timedOut && r2.timedOut),
       realAssertion: r.realAssertion || r2.realAssertion,
+      // infra state from the run we keep (the one with fewer real failures): if
+      // its failures were all infra flakes (infraOnly → failed 0), that is what
+      // the gate should see, so no good goal is blocked on a saturated boot.
+      infraOnly: !!kept.infraOnly,
+      infraFlakes: kept.infraFlakes ?? 0,
       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, timedOut: !!r.timedOut, realAssertion: !!r.realAssertion, 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, infraOnly: !!r.infraOnly, infraFlakes: r.infraFlakes ?? 0, detail: r.detail ?? '', failingNames: r.failingNames ?? [] };
 }
 
 // Failing tests on the BASE (pre-change) checkout — the set the review gate
@@ -2982,6 +2993,15 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
     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.');
   }
+  // Infra-only run (root fix for #616/#620, 2026-07-24): the suite finished but
+  // EVERY failure was an infra flake (a real server could not boot — "server
+  // exited (code N) before it was ready" — or a wait timed out under whole-machine
+  // saturation), so runProjectTests already discounted them to 0. testsFailing is
+  // therefore already false; surface it as an advisory so the human knows the
+  // suite couldn't be trusted here (not that it ran green), and can re-run idle.
+  if (goal.testResult.ran && goal.testResult.infraOnly && !goal.testResult.inconclusive) {
+    goalAct(`Project tests: all ${goal.testResult.infraFlakes || goal.testResult.failed || ''} failure(s) were infra flakes (a real server could not boot / a wait timed out under machine load), not code regressions — not blocking. Re-run when the machine is idle to confirm.`.replace(/\s{2,}/g, ' '));
+  }
   // 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

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.43",
+  "version": "0.10.44",
   "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": {