@galda/cli 0.10.73 → 0.10.74

package/app/index.html

@@ -576,7 +576,8 @@
   .riskfiles{font-family:var(--mono);font-size:10px;color:var(--ink3)}
   /* task 161/162: Review-row quick actions — aligned to the right of the row,
      vertically centered; icon (no emoji) reply; subtle inline reply box. */
-  .revquick{display:flex;gap:4px;flex:0 0 auto;align-items:center;align-self:center;margin-left:8px}
+  .revquick{display:flex;gap:2px;flex:0 0 auto;align-items:center;align-self:center;margin-left:8px;opacity:0;transition:opacity .15s}
+  .task:hover .revquick,.digestrow:hover .revquick,.revquick:focus-within{opacity:1}
   .qbtn{width:26px;height:26px;flex:0 0 26px;border:1px solid var(--hair2);border-radius:7px;background:transparent;cursor:pointer;font-size:12px;line-height:0;display:grid;place-items:center;color:var(--ink2);padding:0}
   .qbtn svg{width:14px;height:14px}
   .qbtn.ok{color:var(--green);border-color:rgba(var(--green-rgb),.4)}
@@ -584,6 +585,15 @@
   .qbtn.no{color:var(--danger);border-color:rgba(var(--danger-rgb),.4)}
   .qbtn.no:hover{background:rgba(var(--danger-rgb),.12)}
   .qbtn:hover{color:var(--ink)}
+  /* Review行だけ: flagshipの .act/.b パターン(枠なし・ホバーで出現・面を塗らずグリフ色のみ) */
+  .revquick .qbtn{border:0;border-radius:6px}
+  .revquick .qbtn.ok{color:var(--ink3);border-color:transparent;background:transparent}
+  .revquick .qbtn.ok:hover{color:var(--green);background:transparent}
+  .revquick .qbtn.no{color:var(--ink3);border-color:transparent;background:transparent}
+  .revquick .qbtn.no:hover{color:var(--danger);background:transparent}
+  .revquick .qbtn.reply{color:var(--ink3)}
+  .revquick .qbtn.reply:hover{color:var(--ink)}
+  .revquick .qbtn:active{transform:scale(.92)}
   .qreplyrow{display:flex;gap:6px;margin-top:8px;width:100%}
   .qreplyrow[hidden]{display:none}
   .qreplyrow input{flex:1;min-width:0;height:30px;border:1px solid var(--hair);border-radius:8px;background:transparent;color:var(--ink);padding:0 10px;font:inherit;font-size:12px}
@@ -3418,6 +3428,8 @@ const FINISHED = (s) => ['done', 'partial', 'failed', 'interrupted'].includes(s)
 const ICONS = {
   pencil: '<svg viewBox="0 0 24 24"><path d="M4 20l1-4L16.5 4.5a2.1 2.1 0 013 3L8 19l-4 1z"/></svg>',
   reply: '<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M9 17l-5-5 5-5"/><path d="M4 12h11a4 4 0 0 1 4 4v3"/></svg>',
+  check: '<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><path d="M5 12.5l4.2 4.2L19 7"/></svg>',
+  close: '<svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2.2" stroke-linecap="round" stroke-linejoin="round"><path d="M6 6l12 12M18 6L6 18"/></svg>',
   up: '<svg viewBox="0 0 24 24"><path d="M12 19V5M5 12l7-7 7 7"/></svg>',
   down: '<svg viewBox="0 0 24 24"><path d="M12 5v14M5 12l7 7 7-7"/></svg>',
   trash: '<svg viewBox="0 0 24 24"><path d="M4 7h16M9 7V5a1 1 0 011-1h4a1 1 0 011 1v2M6.5 7l1 13h9l1-13M10 11v6M14 11v6"/></svg>',
@@ -4097,8 +4109,8 @@ function reviewDigestHtml(rows){
         <div class="digestmeta">${testBadge}${prLink}</div>
       </div>
       <div class="revquick">
-        <button type="button" class="qbtn ok" data-qapprove="${ids}" title="Approve">✓</button>
-        <button type="button" class="qbtn no" data-qdismiss="${ids}" title="Send back">✕</button>
+        <button type="button" class="qbtn ok" data-qapprove="${ids}" title="Approve">${ICONS.check}</button>
+        <button type="button" class="qbtn no" data-qdismiss="${ids}" title="Send back">${ICONS.close}</button>
       </div>
     </div>`;
   };
@@ -4686,8 +4698,8 @@ function renderTasks(){
      <div class="cond">${g.pr ? `<a class="prbig" href="${esc(g.pr)}" target="_blank">PR #${esc((g.pr.match(/pull\/(\d+)/) ?? [,''])[1])} ↗</a>` : ''}</div>
      <div class="qreplyrow" data-qreplyrow="${g.id}"${state.qreplyOpen.has(g.id) ? '' : ' hidden'}><input type="text" data-qreplyinput="${g.id}" placeholder="直したい点を書いて送信（Todoに戻して再作業）"><button type="button" class="qreplysend" data-qreplysend="${g.id}" aria-label="Send">${ICONS.up}</button></div></div>
      <div class="revquick">
-       <button type="button" class="qbtn ok" data-qapprove="${g.id}" title="Approve">✓</button>
-       <button type="button" class="qbtn no" data-qdismiss="${g.id}" title="Send back">✕</button>
+       <button type="button" class="qbtn ok" data-qapprove="${g.id}" title="Approve">${ICONS.check}</button>
+       <button type="button" class="qbtn no" data-qdismiss="${g.id}" title="Send back">${ICONS.close}</button>
        <button type="button" class="qbtn reply" data-qreply="${g.id}" title="Reply">${ICONS.reply}</button>
      </div></div>`;
   }).join('')
@@ -8358,9 +8370,12 @@ async function refresh(){
 // events — especially worker activity — don't saturate the main thread and drop
 // keystrokes while typing.
 let _renderT = null, _renderP = false;
-function scheduleRender(){ if (_renderT) { _renderP = true; return; } render(); _renderT = setTimeout(() => { _renderT = null; if (_renderP) { _renderP = false; scheduleRender(); } }, 120); }
+// A thrown render() must not stall the throttle loop for good (live feed freeze,
+// e.g. on review reply) — catch it, log it, and keep scheduling so the next SSE
+// event still gets a paint attempt.
+function scheduleRender(){ if (_renderT) { _renderP = true; return; } try { render(); } catch (e) { console.error('[render]', e); } _renderT = setTimeout(() => { _renderT = null; if (_renderP) { _renderP = false; scheduleRender(); } }, 120); }
 let _actT = null, _actP = false;
-function scheduleActRender(){ if (_actT) { _actP = true; return; } renderActPanel(); if (state.layout === 'flagship') renderFsFeed(); _actT = setTimeout(() => { _actT = null; if (_actP) { _actP = false; scheduleActRender(); } }, 200); }
+function scheduleActRender(){ if (_actT) { _actP = true; return; } try { renderActPanel(); if (state.layout === 'flagship') renderFsFeed(); } catch (e) { console.error('[actRender]', e); } _actT = setTimeout(() => { _actT = null; if (_actP) { _actP = false; scheduleActRender(); } }, 200); }
 // (connection refused = fatal, no auto-retry), so we own the reconnect loop.
 let esRetryMs = 1000;
 let sseUp = false; // live engine link — the honest signal behind the MCP tab's status dot (v46)

package/engine/lib.mjs

@@ -1277,6 +1277,38 @@ export function taskStatusAfterVerify({ verdict, code, timedOut }) {
   return code === 0 ? 'done' : 'failed';
 }
 
+// task 450 (goal #783, task 448): a worker can exit 0 with changedFiles: []
+// and a result that reads as a thought cut off mid-sentence (e.g. "I'll wait
+// for this background full-suite run to complete rather than poll.") — no
+// work happened, but taskStatusAfterVerify (decided from `code` alone, before
+// changedFiles even exists) already called it 'done'. This is the same
+// principle the product already applies to verification (don't trust
+// self-report, judge by evidence) applied to task completion itself.
+// Deliberately narrow, per the same 0-file-legitimate-report carve-out as
+// isNothingVerifiable above: a real investigation/report task also has zero
+// changed files, but its result reads as a normal finished sentence, so only
+// text that also LOOKS cut off trips this.
+export function looksIncomplete(text) {
+  const t = String(text ?? '').trim();
+  if (!t) return true;
+  if (t.length >= 200) return false;
+  // A short blurb ending mid-word/mid-clause (no sentence-final punctuation
+  // in any of ja/en) reads as truncated, not as a deliberate one-liner.
+  if (!/[.!?。！？」）)]$/.test(t)) return true;
+  // goal-783/task-448's actual fragment ("I'll wait for this background
+  // full-suite run to complete rather than poll.") IS a grammatically
+  // complete sentence, so the check above misses it — but it's a one-shot
+  // batch worker narrating a future intent ("I'll…"/"let me…"), never a
+  // report of something actually done. That opener is itself the signal.
+  return /^(i'll|i will|i'm going to|i am going to|let me)\b/i.test(t);
+}
+
+export function isSuspiciouslyIncompleteDone({ status, changedFiles = [], result = '', hasVerdict = false } = {}) {
+  if (status !== 'done' || hasVerdict) return false;
+  if ((changedFiles ?? []).length > 0) return false;
+  return looksIncomplete(result);
+}
+
 // A task counts as complete for GOAL status when it cleanly finished
 // (done/skipped) OR it was force-stopped (interrupted) but its change is STRONGLY
 // verified — an independent passCondition verification PASSED (proof.verified &&

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 } from './lib.mjs';
-import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, 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, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession } from './lib.mjs';
+import { parseStreamEvents, parseCodexEvents, buildCodexArgs, parsePlan, refinePlanTasks, canEditGoal, canDeleteGoal, shouldAskClarification, workerExitReason, classifyWorkerFailure, isForcedStop, taskStatusAfterVerify, workerResultText, taskCountsAsComplete, resolveWantsPR, resolveGoalSource, buildGoalSummary, 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, REPLY_INTENTS, buildReplyIntentPrompt, parseReplyIntent, buildGoalMessage, serializeGoalMessage, parseGoalMessages, computeInternalQualityMetrics, classifyInternalQualityTaskError, resolveWorkspaceMode, goalSessionFor, setGoalAgentSession, switchGoalAgentSession } from './lib.mjs';
 import { createSerialQueue } from './lib.mjs';
 import { isPrClosed } from './lib.mjs';
 import { collectProjectRules } from './lib.mjs';
@@ -2921,6 +2921,19 @@ async function runTask(task) {
   const after = await gitChanges(workDir);
   task.changedFiles = after.filter((l) => !before.includes(l)).map((l) => l.slice(3));
   sendProcessAct(task, `Detected ${task.changedFiles.length} changed file${task.changedFiles.length === 1 ? '' : 's'} from this task.`);
+  // task 450: taskStatusAfterVerify decided 'done' from `code` alone, before
+  // changedFiles existed to check against. Don't trust a clean exit code by
+  // itself — a worker that changed nothing and left a cut-off-looking result
+  // (goal #783 task 448: "I'll wait for this background full-suite run to
+  // complete rather than poll.") did not actually finish. Downgrade to
+  // 'interrupted' (retryable, surfaces in Attention) instead of silently
+  // shipping it as done. Never touches a real 0-file report/investigation
+  // task, whose result reads as a normal finished sentence.
+  if (isSuspiciouslyIncompleteDone({ status: task.status, changedFiles: task.changedFiles, result: task.result, hasVerdict: Boolean(verdict) })) {
+    task.status = 'interrupted';
+    task.result = `[未完了の疑い: 変更ファイル0件かつ応答が途中で切れています] ${task.result}`.slice(0, RESULT_MAX);
+    sendProcessAct(task, 'Result looked incomplete (0 changed files, cut-off text) — marked interrupted instead of done for human review.');
+  }
   task.finishedAt = new Date().toISOString();
   task.resultAttemptId = task.attemptId;
   saveTask(task);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.73",
+  "version": "0.10.74",
   "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": {