@galda/cli 0.10.72 → 0.10.74

package/engine/lib.mjs

@@ -655,9 +655,25 @@ export function estimateTokenOptimization({ usage = null, weightedTokens = null,
   // layer did to save tokens, and it is the opposite: the model is the user's
   // own pick from the composer, and we deliberately never override it
   // (Masa 2026-07-22). Every remaining line is an action this layer took.
+
+  // Masa 2026-07-26: a positive `baseline` percent is a comparison against past
+  // goals, not proof any technique was applied THIS run — a small goal that
+  // followed a run of large ones scores a big "saved" number purely from the
+  // size mix, with nothing the user actually did behind it. Only call it
+  // "saved" when a real technique signal fired this run (cache reuse, a fresh
+  // handoff, /compact, /clear); otherwise it is a true fact (this run was
+  // smaller/larger than usual) but not a causal claim, so we relabel it
+  // `baseline-untied` and the renderer must phrase it as a comparison, not a
+  // "you saved X% by doing Y".
+  const techniqueDriven = cacheRatio > 0 || usedHandoff || sessionMode === 'cold-handoff' || usedCompact || sessionMode === 'compact' || usedClear;
+  if (basis === 'baseline' && !techniqueDriven) basis = 'baseline-untied';
+
   if (basis === 'baseline' && percent > 0) {
     techniques.push(`${Math.round(perReq / 1000)}k tokens/requirement vs your ${Math.round(baseline / 1000)}k average`);
   }
+  if (basis === 'baseline-untied') {
+    techniques.push(`${Math.round(perReq / 1000)}k tokens/requirement vs your ${Math.round(baseline / 1000)}k average`);
+  }
   if (!techniques.length) techniques.push(`${Math.round(raw / 1000)}k tokens this run (no cache reuse yet)`);
 
   return { rawTokens: raw, weightedTokens: weighted, savedTokens, percent, basis, techniques: [...new Set(techniques)] };
@@ -1261,6 +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 &&
@@ -2254,6 +2302,21 @@ export function isNothingVerifiable({ changedFiles = [], hasPassCondition = fals
   return hasPassCondition && changedFiles.length === 0 && !hasAnyProof;
 }
 
+// goal-751 (2026-07-26): a wantsPR goal is an explicit ask for a PR, so a worker
+// that ends with zero changed files and no proof did not deliver what was asked —
+// most often a lane-conflict report ("空き次第自動再開します") that cannot actually
+// happen, since a `claude -p` worker is a one-shot batch run with no way to resume
+// itself later. taskStatusAfterVerify still reads a clean exit (code===0, no
+// verdict) as 'done', so this reached Review looking exactly like a normal
+// completed PR. Same shape as isNothingVerifiable above but keyed on wantsPR
+// instead of passCondition (a wantsPR goal has no passCondition of its own — the
+// PR ask itself is the promise). Advisory only, like proofNote/verificationInfra
+// Error below: the caller sets a note, it never blocks (Failedを作らない方針, see
+// docs/HANDOFF-2026-07-25-verify-gate-default-off.md).
+export function isNothingVerifiableForPR({ wantsPR = false, changedFiles = [], hasAnyProof = false } = {}) {
+  return Boolean(wantsPR) && changedFiles.length === 0 && !hasAnyProof;
+}
+
 // チェックリストUIで要件が全件Approveされたら goal を review→done へ進める。
 // 以前は「PRありgoalはマージが唯一の完了トリガー」として弾いていたが、それだと
 // 承認しても done が永続化されず次のrefreshで review に戻る不具合になっていた
@@ -2542,10 +2605,10 @@ export function hasGitChanges(porcelainStatus) {
   return Boolean(String(porcelainStatus ?? '').trim());
 }
 
-export function resolveWorkspaceMode(input) {
-  if (input === 'direct-workspace' || input === 'direct') return 'direct-workspace';
-  if (input === 'isolated-worktree' || input === 'worktree' || input === true) return 'isolated-worktree';
-  if (input === false) return 'direct-workspace';
+// Direct workspace is retired (2026-07-26): two accidents where a stray client-side
+// toggle edited a shared checkout directly. This always returns isolated-worktree
+// regardless of what a client sends, so a stale/tampered client can't force direct.
+export function resolveWorkspaceMode(_input) {
   return 'isolated-worktree';
 }
 
@@ -3151,6 +3214,13 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
     // (research, summaries, answers) simply skip this — the worker decides,
     // never us.
     `- 人が実際に動かして結果を見られる依頼なら、最後にプロジェクト直下へ ${RUN_DECL_FILE} を書く: {"cmd": "<起動コマンド>", "url": "<開くURL>", "note": "<何が見えるか1行>", "check": "<人に何を見て判断してほしいか1行>", "title": "<この作業のチケット名。対象が分かる程度に説明的で動作を含む短い名前。例「看板モードの追加」「ヘッダのズレの修正」。依頼文をそのまま写さない>"}。Manager はこれを使って人に「動かして見る」と「何を見ればいいか」を出す。動かして見せるものが無い依頼（調査・要約・質問への回答など）では書かなくてよい。`,
+    // `file` (goal #744):起動不要な静的成果物（比較HTML、レポート等）専用。この
+    // クラスの依頼は cmd/url を使う理由がなく、worker が唯一持っている手（file://
+    // URL）は http(s) しか許さないバリデーションに毎回弾かれ、しかも弾かれた理由が
+    // 「JSONが壊れている」という誤診断メッセージだったため、同じ file:// を書き直
+    // して再度失敗するループが実際に2回連続で起きた（goal #744）。ここで file を
+    // 明示し、file:// URLを書く選択肢自体を塞ぐ。
+    `- 起動不要な静的ファイル（HTML比較ページ・レポート等、npm run devのようなコマンドを必要としないもの）を見せたい場合は、cmd/url でなく "file": "<プロジェクト内の相対パス>" を使う。例: {"file": "docs/design/options.html", "title": "3案の比較"}。file:// URLは絶対に書かない（urlはhttp(s)以外だと宣言ごと弾かれる）。`,
     // Root-cause prevention (goal #733, 2026-07-26): the worker Wrote this file,
     // then Edit'd it, and the edit broke the JSON mid-edit — Manager silently
     // dropped the whole declaration (fixed separately: it now surfaces an honest
@@ -3186,6 +3256,22 @@ export function workerPrompt(task, goal, lastFailure, agentName = 'claude-code',
     // the request is a change-something request, and most requests are not.
     agentName === 'codex' ? '' : `- 同じ画面の「変える前」と「変えた後」を撮ったなら、${RUN_DECL_FILE} に \`"compare": ["<前の画像>.png", "<後の画像>.png"]\` と書いてよい（その2枚が同じ場所の前後だと分かるのは撮った本人だけ）。人には並べて、違う部分に印を付けて見せる。3案を並べた等、前後の関係が無い画像には書かない。`,
     `- 最後に${replyLang}で2〜4行：何を変えたか・どのファイルか・テスト結果・人間が確認すべき点。`,
+    // Root-cause fix (Masa 2026-07-26, goal #727 review): a worker fixed a real bug
+    // (the To-Do lane's Doing ring had zero animation) but its report also claimed
+    // the project-rail dot was "static" before — it actually already had a working
+    // pulse animation (fsrundot), just not a spin. The worker never re-read the
+    // pre-change code to check its own diagnosis; it narrated a plausible-sounding
+    // "before" state instead of a verified one, and nothing downstream caught it.
+    '- 「〜は動いていなかった/静止していた」のように直す前の状態を断定する場合は、その断定をする前に修正前の該当コード（該当CSS/関数）を実際に読んで確認すること。読まずに推測で「動いていない」と書かない。読んで確認した箇所は、報告に該当コードを短く引用すること。',
+    // goal #751 (2026-07-26): you are a one-shot batch process — once this run exits,
+    // nothing resumes you. If something outside your control blocks you (a lane/file
+    // lock held by another agent, a missing permission, etc.), do NOT report that you
+    // will "resume automatically once it frees up" or similar — that is not something
+    // you or anything in this run can do. Say plainly that you made no code changes and
+    // 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。',
   ].filter(Boolean).join('\n');
 }
 
@@ -3203,6 +3289,37 @@ 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';
 
+// 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
+// file. Conflating them (goal #744) sent the worker back a message that only
+// describes JSON corruption ("was not usable JSON"), so a worker whose actual
+// mistake was a file:// url read that as "my JSON broke" and rewrote the exact
+// same file:// url a second time. This return shape lets the caller tell them
+// apart: `null` stays "nothing usable to report" (unparseable / genuinely
+// empty), an `{ error, message }` object carries a reason a human — or the
+// worker reading its own feedback — can act on.
+function runDeclError(error, message) {
+  return { error, message };
+}
+
+// `file` (goal #744, this PR): a static, already-built artifact — an HTML
+// comparison page, a report — that needs no process and has no http(s) URL of
+// its own. Declared as a path relative to the project so the Manager can copy
+// it out of the worktree (same pattern as SHOT_DIR) and serve it from a URL
+// that survives after the worktree is gone. Only a syntactic check lives here
+// (no absolute path, no ".." segment, no backslash/drive letter, no NUL) —
+// server.mjs still re-resolves against the worktree root before it reads
+// anything, because a syntax check alone doesn't rule out a symlink.
+export function isSafeRunDeclFilePath(p) {
+  if (typeof p !== 'string' || !p) return false;
+  if (p.includes('\0')) return false;
+  if (p.startsWith('/') || p.startsWith('\\')) return false;
+  if (/^[a-zA-Z]:[\\/]/.test(p)) return false; // Windows drive letter
+  const norm = p.replace(/\\/g, '/');
+  return !norm.split('/').some((seg) => seg === '..');
+}
+
 export function parseRunDeclaration(text) {
   let raw;
   try { raw = JSON.parse(String(text ?? '')); } catch { return null; }
@@ -3210,6 +3327,7 @@ export function parseRunDeclaration(text) {
   const str = (v, max) => (typeof v === 'string' && v.trim() ? v.trim().slice(0, max) : null);
   const cmd = str(raw.cmd, 500);
   const url = str(raw.url, 500);
+  const file = str(raw.file, 500);
   // `compare` (Masa 2026-07-22): which TWO of the worker's own screenshots are
   // the same view before and after. Only the agent that did the work knows that
   // two images relate — a request for three design options has no "after" at
@@ -3219,9 +3337,16 @@ export function parseRunDeclaration(text) {
   const compare = parseComparePair(raw.compare);
   const title = str(raw.title, 60);
   // Nothing to start, nothing to open, nothing named, and no pair to show = nothing to look at.
-  if (!cmd && !url && !compare && !title) return null;
-  // Only http(s): the button opens this in the user's browser.
-  if (url && !/^https?:\/\//i.test(url)) return null;
+  if (!cmd && !url && !file && !compare && !title) return null;
+  // Only http(s): the button opens this in the user's browser. A file:// (or
+  // any other scheme) is a worker mistake with a specific fix — declare `file`
+  // instead — not a JSON parse failure, so it gets its own reason.
+  if (url && !/^https?:\/\//i.test(url)) {
+    return runDeclError('bad-url-scheme', "url must start with http:// or https:// — a file:// path can't be opened from a remote browser session; declare an artifact file instead of a file:// url");
+  }
+  if (file && !isSafeRunDeclFilePath(file)) {
+    return runDeclError('bad-file-path', 'file must be a path inside the project — no leading /, no .. segments');
+  }
   // `check` (Masa 2026-07-22): what to look at, in the working AI's own words. It used
   // to come from the intake AI's plan — this layer inventing a pass condition for a
   // request it only read. The AI that did the work states it here instead, in the same
@@ -3232,7 +3357,7 @@ export function parseRunDeclaration(text) {
   // the user's own sentence, repeated back at them in the one place that is
   // supposed to say what the work WAS. The agent that just did it can name it in
   // one line, and it is the only party that knows what actually got done.
-  return { cmd, url, note: str(raw.note, 300), check: str(raw.check, 300), compare, title };
+  return { cmd, url, file, note: str(raw.note, 300), check: str(raw.check, 300), compare, title };
 }
 
 // Exactly two shot file names, in [before, after] order. Anything else is

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, 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';
@@ -2406,6 +2406,17 @@ function readRunDeclaration(task, workDir) {
   try { raw = readFileSync(file, 'utf8'); } catch { return; }
   try { unlinkSync(file); } catch { /* best effort — never fail a task over cleanup */ }
   const run = parseRunDeclaration(raw);
+  // A declaration that parsed as JSON but broke a rule (bad url scheme, unsafe
+  // file path) is a worker mistake with a specific, sayable fix — NOT the same
+  // failure as unparseable JSON. Goal #744: conflating the two sent the worker
+  // "was not usable JSON" for a file:// url, so it "fixed" the JSON (which was
+  // never broken) and wrote the exact same file:// url a second time. `run` is
+  // `{ error, message }` here, distinct from the `null` case below.
+  if (run && run.error) {
+    task.runDeclError = { reason: run.error, message: run.message, snippet: String(raw ?? '').slice(0, 200) };
+    sendAct(task, `run declaration ignored — ${run.message}`);
+    return;
+  }
   if (!run) {
     // The worker DID try to declare something (the file exists) but it came out
     // unusable — e.g. a Write followed by an Edit that broke the JSON mid-edit
@@ -2417,17 +2428,62 @@ function readRunDeclaration(task, workDir) {
     // actually written (raw content is deleted above and never otherwise
     // recoverable) so both the activity log and the Review card can say WHY,
     // instead of the human having to guess or dig through file loss.
-    task.runDeclError = { snippet: String(raw ?? '').slice(0, 200) };
+    task.runDeclError = { reason: 'invalid-json', snippet: String(raw ?? '').slice(0, 200) };
     sendAct(task, `run declaration ignored — ${RUN_DECL_FILE} was not usable JSON: ${task.runDeclError.snippet.slice(0, 80)}`);
     return;
   }
   task.run = { ...run, dir: workDir };
-  const how = run.cmd ?? run.url;
+  collectRunDeclFileArtifact(task, workDir);
+  const how = task.run.cmd ?? task.run.url;
   if (how) sendProcessAct(task, `Worker declared how to see the result: ${how}`);
   if (run.title) sendProcessAct(task, `Worker named the work: ${run.title}`);
   if (run.compare) sendProcessAct(task, `Worker declared a before/after pair: ${run.compare.join(' → ')}`);
 }
 
+// `file` (goal #744): a static, already-built artifact that needs no process
+// and has no http(s) URL of its own — an HTML comparison page, a report. Same
+// pattern as collectWorkerShots: copy it out of the worktree (which is about
+// to be diffed/discarded) into logDir/uploads and give it a URL that survives.
+// parseRunDeclaration already checked the path is syntactically inside the
+// project (no leading /, no ..); this re-resolves against the worktree root
+// before reading anything, because a syntax check alone doesn't rule out a
+// symlink pointing outside it.
+function collectRunDeclFileArtifact(task, workDir) {
+  const rel = task.run?.file;
+  if (!rel) return;
+  const root = resolve(workDir) + '/';
+  const abs = resolve(workDir, rel);
+  if (!(abs + '/').startsWith(root) && abs !== root.slice(0, -1)) {
+    task.runDeclError = { reason: 'file-outside-project', message: `file "${rel}" resolves outside the project`, snippet: rel };
+    sendAct(task, `run declaration ignored — ${task.runDeclError.message}.`);
+    task.run = { ...task.run, file: null };
+    return;
+  }
+  let data;
+  try { data = readFileSync(abs); } catch (err) {
+    task.runDeclError = { reason: 'file-unreadable', message: `file "${rel}" could not be read: ${String(err?.code ?? err).slice(0, 80)}`, snippet: rel };
+    sendAct(task, `run declaration's file "${rel}" could not be read: ${String(err?.code ?? err).slice(0, 80)}`);
+    task.run = { ...task.run, file: null };
+    return;
+  }
+  const outDir = join(logDir, 'uploads');
+  mkdirSync(outDir, { recursive: true });
+  const safe = basename(rel).replace(/[^\w.-]/g, '_').slice(-80);
+  const outFile = `${Date.now()}-artifact-${task.id}-${safe}`;
+  try {
+    writeFileSync(join(outDir, outFile), data);
+  } catch (err) {
+    task.runDeclError = { reason: 'file-attach-failed', message: `file "${rel}" could not be attached: ${String(err?.code ?? err).slice(0, 80)}`, snippet: rel };
+    sendAct(task, `run declaration's file "${rel}" could not be attached: ${String(err?.code ?? err).slice(0, 80)}`);
+    task.run = { ...task.run, file: null };
+    return;
+  }
+  // No process to start for a static file — the Review card renders this as a
+  // plain link (like a PR link), not the "Start it" button.
+  task.run = { ...task.run, url: `/upload/${outFile}`, cmd: null };
+  sendProcessAct(task, `Worker attached a file to open: ${rel}`);
+}
+
 // A direct/shared workspace can still contain a declaration left by an older
 // run (for example when that run was interrupted before readRunDeclaration).
 // Remove it before this task starts so only a file created by THIS worker can
@@ -2865,6 +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);
@@ -3532,6 +3601,18 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
     hasPassCondition: siblings.some((t) => t.passCondition),
     hasAnyProof: siblings.some((t) => t.proof),
   });
+  // goal-751: wantsPR but the worker changed nothing and proved nothing (typically a
+  // lane-conflict report promising an automatic resume that a one-shot `claude -p`
+  // run can never deliver). Advisory only — see isNothingVerifiableForPR in lib.mjs.
+  const prNothingVerifiable = isNothingVerifiableForPR({
+    wantsPR: goal.wantsPR,
+    changedFiles,
+    hasAnyProof: siblings.some((t) => t.proof),
+  });
+  goal.prNote = prNothingVerifiable
+    ? 'workerはコード変更をせずに終了しました（PRを求めていたのに0ファイル変更）。実装は行われていない可能性があります。'
+    : null;
+  if (goal.prNote) goalAct(`PR advisory (not blocking): ${goal.prNote}`);
   if (testsFailing) {
     const failure = recordGoalFailure(goal, {
       kind: 'test',
@@ -3608,7 +3689,7 @@ async function verifyGate(goal, prProject, siblings, baseProject, activityTask =
     const preNote = goal.testResult?.preExistingOnly
       ? ` (${goal.testResult.preExisting?.length ?? 0} pre-existing test failure(s) ignored as unrelated to this change)`
       : '';
-    goalAct(`Verification gate passed; moving goal to ${goal.status}.${preNote}${goal.proofNote ? ` [proof advisory: ${goal.proofNote}]` : ''}`);
+    goalAct(`Verification gate passed; moving goal to ${goal.status}.${preNote}${goal.proofNote ? ` [proof advisory: ${goal.proofNote}]` : ''}${goal.prNote ? ` [PR advisory: ${goal.prNote}]` : ''}`);
   }
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@galda/cli",
-  "version": "0.10.72",
+  "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": {