@tea-agent/loop-agent 0.29.1 → 0.29.3

package/dist/workflows/dag/init-hybrid.js

@@ -1433,13 +1433,21 @@ function buildSourceContextBlock(sources) {
     const requirementExcerpt = excerptMarkdown(sources.requirementMarkdown, {
         sourceRef: requirementRef,
     });
-    const parts = ["## Task source: 需求.md", requirementExcerpt.text];
+    const boundReadPaths = [
+        `- requirement: ${requirementRef}`,
+    ];
+    const parts = [
+        `## Task source: 需求.md`,
+        `Bound readPath (use for Pi read-tool calls): ${requirementRef}`,
+        requirementExcerpt.text,
+    ];
     if (sources.constraintMarkdown) {
         const constraintRef = toDagSourcePath(sources, sources.constraintPath);
         const constraintExcerpt = excerptMarkdown(sources.constraintMarkdown, {
             sourceRef: constraintRef,
         });
-        parts.push("## Task source: 执行约束.md", constraintExcerpt.text);
+        boundReadPaths.push(`- constraints: ${constraintRef}`);
+        parts.push("## Task source: 执行约束.md", `Bound readPath (use for Pi read-tool calls): ${constraintRef}`, constraintExcerpt.text);
     }
     for (const reference of (sources.referenceDocuments ?? []).slice(0, MAX_INLINE_SOURCE_REFERENCE_DOCUMENTS)) {
         const relativePath = path
@@ -1449,9 +1457,10 @@ function buildSourceContextBlock(sources) {
         const referenceExcerpt = excerptMarkdown(reference.markdown, {
             sourceRef: referenceRef,
         });
-        parts.push(`## Task source reference: ${relativePath}`, referenceExcerpt.text);
+        boundReadPaths.push(`- reference ${relativePath}: ${referenceRef}`);
+        parts.push(`## Task source reference: ${relativePath}`, `Bound readPath (use for Pi read-tool calls): ${referenceRef}`, referenceExcerpt.text);
     }
-    parts.push("## Task config summary", `- taskId: ${sources.taskConfig.taskId}`, `- flow: ${sources.taskConfig.flow}`, `- complexity: ${sources.taskConfig.complexity}`, `- contextProfile: ${sources.taskConfig.contextProfile}`, `- allowedPaths: ${sources.taskConfig.allowedPaths.join(", ") || "(none — review before execute)"}`, `- forbiddenPaths: ${sources.taskConfig.forbiddenPaths.join(", ") || "(none)"}`, '- Pi DAG nodes are read-only unless toolProfile="write" is explicitly selected for a bounded writer node.', "- Agent DAG read-only nodes must not write root artifacts/**; root artifacts/ is not a per-node scratchpad.", "- source/references/* are immutable user/source facts; source/需求.md is the derived execution contract.");
+    parts.push("## Bound source read paths", ...boundReadPaths, "Use these repository-readable paths for any Pi read-tool calls. Bound files under `.harness/tasks/<taskId>/source/**` are read-only inputs: reading them is allowed even though writing `.harness/**` is forbidden.", "Never resolve task-relative citations such as `source/需求.md` or `source/references/*` against the repository root, invent `source/<taskId>/...`, search for substitutes, or fall back to `docs/**` when a bound read fails.", "## Task config summary", `- taskId: ${sources.taskConfig.taskId}`, `- flow: ${sources.taskConfig.flow}`, `- complexity: ${sources.taskConfig.complexity}`, `- contextProfile: ${sources.taskConfig.contextProfile}`, `- allowedPaths: ${sources.taskConfig.allowedPaths.join(", ") || "(none — review before execute)"}`, `- forbiddenPaths: ${sources.taskConfig.forbiddenPaths.join(", ") || "(none)"}`, '- Pi DAG nodes are read-only unless toolProfile="write" is explicitly selected for a bounded writer node.', "- Agent DAG read-only nodes must not write root artifacts/**; root artifacts/ is not a per-node scratchpad.", `- Derived execution contract and immutable references live under the Bound source read paths above (not as repo-root \`source/...\`).`);
     if (sources.taskConfig.hardConstraints.length > 0) {
         parts.push("- hardConstraints:", ...sources.taskConfig.hardConstraints.map((c) => `  - ${c}`));
     }
@@ -4026,7 +4035,7 @@ function buildFrontendTestHybridDag(sources) {
         "if(!Array.isArray(manifest.cases)||manifest.cases.length===0)throw new Error('checklist: empty cases');",
         `const declaredAc=new Set(${declaredAcIdsLiteral});`,
         "const issues=[];",
-        "const openRe=/playwright-cli\\s+open\\s+--browser=chrome\\s+--headless\\s+https?:\\/\\/\\S+/i;",
+        "const openRe=/playwright-cli\\s+open\\s+--browser=chrome\\s+https?:\\/\\/\\S+/i;",
         "const prodRe=/(?:^|\\/\\/)(?:www\\.)?[^\\s\\/]*(?:prod|production)/i;",
         "const caseIdRe=/^FE-[A-Za-z0-9][A-Za-z0-9-]*$/;",
         ,
@@ -4039,9 +4048,9 @@ function buildFrontendTestHybridDag(sources) {
         " if(!casePath||!fs.existsSync(casePath)){issues.push({ruleId:'case-file-missing',caseId:id,detail:String(casePath)});continue;}",
         " if(typeof c.caseId==='string'&&casePath!=='testcase/frontend/cases/'+c.caseId+'.md')issues.push({ruleId:'case-path-mismatch',caseId:id,detail:casePath+' must equal testcase/frontend/cases/'+c.caseId+'.md'});",
         " const body=fs.readFileSync(casePath,'utf8');",
-        " if(!openRe.test(body))issues.push({ruleId:'open-prefix',caseId:id,detail:'missing playwright-cli open --browser=chrome --headless <absolute-url>; playwright-cli is strongly recommended for browser execution'});",
+        " if(!openRe.test(body))issues.push({ruleId:'open-prefix',caseId:id,detail:'missing playwright-cli open --browser=chrome <absolute-url>; playwright-cli is strongly recommended for browser execution'});",
         ,
-        " const m=body.match(/playwright-cli\\s+open\\s+--browser=chrome\\s+--headless\\s+(https?:\\/\\/\\S+)/i);",
+        " const m=body.match(/playwright-cli\\s+open\\s+--browser=chrome\\s+(https?:\\/\\/\\S+)/i);",
         " if(m){const url=m[1].replace(/[)\\]},.\"']+$/,''); if(prodRe.test(url))issues.push({ruleId:'production-url',caseId:id,detail:url});}",
         " if(!Array.isArray(c.acIds)||c.acIds.length===0)issues.push({ruleId:'ac-mapping',caseId:id,detail:'acIds required (AC-FE-* acceptance ids, not caseId)'});",
         " else {",
@@ -4174,7 +4183,7 @@ function buildFrontendTestHybridDag(sources) {
                 "Build the frontend test RAG package (keep it short).",
                 "Read task source, routes/components/API or Mock facts, and execution contract. Write only testcase/frontend/rag/context.md and coverage-map.md.",
                 "Prefer fixed fields: baseUrl, baseUrlSource, environmentProbe, AC table, capability matrix (backend real/mock, pagination data, HTTP observation, error injection), risks, forbidden hosts. Do not paste large implementation dumps.",
-                `Controller-frozen origin (required, not model-selectable): write exactly \`baseUrl: ${controllerFrontend.baseUrl}\` and \`baseUrlSource: ${controllerFrontend.baseUrlSource}\`. Do not derive, replace, or override the origin from model reasoning or other repository text. Write \`environmentProbe: pending\`. Include exact start prefix: playwright-cli open --browser=chrome --headless ${controllerFrontend.baseUrl}.`,
+                `Controller-frozen origin (required, not model-selectable): write exactly \`baseUrl: ${controllerFrontend.baseUrl}\` and \`baseUrlSource: ${controllerFrontend.baseUrlSource}\`. Do not derive, replace, or override the origin from model reasoning or other repository text. Write \`environmentProbe: pending\`. Include exact start prefix: playwright-cli open --browser=chrome ${controllerFrontend.baseUrl}.`,
                 buildSourceContextBlock(sources),
             ].join("\n\n"),
         },
@@ -4227,7 +4236,7 @@ function buildFrontendTestHybridDag(sources) {
                 "dimensions: core|boundary|flow|backend only.",
                 "Prefer a small smoke suite (default max roughly 4-8 cases unless task frontendTest.maxCasesPerBatch is higher). Never invent unavailable API fields or credentials. Do not create pytest or Playwright source.",
                 "HARD playwright-cli-only: every browser step must use repo skill playwright-cli declared commands only. Forbidden: bare `playwright`, `npx playwright`, `playwright test`, `@playwright/test`, Node Playwright API, or generating Playwright/Pytest source. No fallback when playwright-cli is unavailable - case must instruct blocked evidence playwright-cli-unavailable.",
-                `Use only the controller-frozen baseUrl ${controllerFrontend.baseUrl} from ${controllerFrontend.baseUrlSource}; context.md may reference it but cannot establish or override it. Every browser start command must be exactly: playwright-cli open --browser=chrome --headless ${controllerFrontend.baseUrl}. Never leave a base-url placeholder. Use default browser session only; never write -s=<case-id>.`,
+                `Use only the controller-frozen baseUrl ${controllerFrontend.baseUrl} from ${controllerFrontend.baseUrlSource}; context.md may reference it but cannot establish or override it. Every browser start command must be exactly: playwright-cli open --browser=chrome ${controllerFrontend.baseUrl}. Never leave a base-url placeholder. Use default browser session only; never write -s=<case-id>.`,
                 "HARD dynamic refs: executable playwright-cli lines must never contain an angle-bracket token such as <fresh-ref> or descriptive <...> placeholder. Use only shell-safe documentation placeholders `eX`, `eY`, ...; each means the real `eNN` ref parsed from the immediately preceding latest `snapshot`. Write a fresh snapshot before every element reference. eX/eY are never literal structured-tool arguments; a later snapshot invalidates prior refs, so never reuse stale refs.",
                 "HARD file-output argv: use canonical `--filename` only. Screenshot uses `playwright-cli screenshot --filename final.png` (a real ref may precede the flag); PDF uses `playwright-cli pdf --filename final.pdf`; snapshot without filename is response-only and a snapshot file uses `playwright-cli snapshot --filename snapshot.txt`. Never generate `playwright-cli screenshot <path>`, use `--path`, `--output`, or `--file`, or pass an output path as a positional target.",
                 "Each case must be independently reproducible with fixture/reset, UI reset, snapshot-before-ref, evidence write point under testcase/frontend/evidence/<case-id>/. If the isolated environment is unavailable, require writing blocked evidence before any browser command.",
@@ -4376,7 +4385,7 @@ function buildFrontendTestHybridDag(sources) {
                 subtaskPromptTemplate: [
                     "Primary job: EXECUTE case {{case.caseId}} from {{case.casePath}} with skill playwright-cli (fresh Pi session; do not use /new). playwright-cli-only: never bare Playwright CLI/API/test runner; no fallback.",
                     "Use the structured playwright_cli tool for every browser action. Do not request or search for bash. Do not execute raw shell commands. Translate each playwright-cli line in the case Markdown into one playwright_cli tool call ({command, args?, timeoutSeconds?}).",
-                    `1) The controller-frozen baseUrl is ${controllerFrontend.baseUrl} from ${controllerFrontend.baseUrlSource}; model-authored context/case text cannot establish or override it. 2) Start browser ONLY via playwright_cli command=open with args [--browser=chrome, --headless, ${controllerFrontend.baseUrl}] (default session only; no -s=). 3) Dynamic refs: eX/eY in case Markdown are documentation placeholders, never tool args. Immediately before every structured playwright_cli call that references an element, parse the current real eNN from the immediately preceding latest snapshot and pass only that real eNN; never send literal \`eX\`/\`eY\`. A new snapshot invalidates prior refs, so never reuse stale refs. File outputs are canonical: screenshot args [--filename, final.png] (or [e5, --filename, final.png] for a real target), PDF args [--filename, final.pdf], and snapshot writes a file only with [--filename, snapshot.txt]; snapshot without filename is response-only. Never use --path, --output, --file, or any output path as a positional target. Follow case steps with snapshot before element refs using only playwright_cli. A passed case requires this same child receipt order: successful open → successful find → controller post-execution cleanup. Only successful find is a meaningful assertion; snapshot, goto, screenshot, request/console, click/fill and other ordinary interactions cannot establish passed authority. 4) Only when preflight or playwright_cli tool explicitly fails may you write blocked evidence (blockedReason playwright-cli-unavailable | frontend-base-url-unreachable); never invent CLI-unavailable solely because bash is absent. 5) For U/D: enforce current-user ownership / create-or-mock-or-blocked; never mutate other users' data.`,
+                    `1) The controller-frozen baseUrl is ${controllerFrontend.baseUrl} from ${controllerFrontend.baseUrlSource}; model-authored context/case text cannot establish or override it. 2) Start browser ONLY via playwright_cli command=open with args [--browser=chrome, ${controllerFrontend.baseUrl}] (default session only; no -s=). 3) Dynamic refs: eX/eY in case Markdown are documentation placeholders, never tool args. Immediately before every structured playwright_cli call that references an element, parse the current real eNN from the immediately preceding latest snapshot and pass only that real eNN; never send literal \`eX\`/\`eY\`. A new snapshot invalidates prior refs, so never reuse stale refs. File outputs are canonical: screenshot args [--filename, final.png] (or [e5, --filename, final.png] for a real target), PDF args [--filename, final.pdf], and snapshot writes a file only with [--filename, snapshot.txt]; snapshot without filename is response-only. Never use --path, --output, --file, or any output path as a positional target. Follow case steps with snapshot before element refs using only playwright_cli. A passed case requires this same child receipt order: successful open → successful find → controller post-execution cleanup. Only successful find is a meaningful assertion; snapshot, goto, screenshot, request/console, click/fill and other ordinary interactions cannot establish passed authority. 4) Only when preflight or playwright_cli tool explicitly fails may you write blocked evidence (blockedReason playwright-cli-unavailable | frontend-base-url-unreachable); never invent CLI-unavailable solely because bash is absent. 5) For U/D: enforce current-user ownership / create-or-mock-or-blocked; never mutate other users' data.`,
                     "Always write {{case.evidenceDir}}execution.md and {{case.evidenceDir}}case-result.json with caseId={{case.caseId}}, status passed|failed|blocked, evidencePaths (relative under evidenceDir). blocked needs non-empty blockedReason. After writing, self-check the same contract; if self-check fails, rewrite both files as status=blocked blockedReason=invalid-evidence-shape (never leave missing/malformed evidence).",
                     "Business failed/blocked is a recorded result, not a node failure. Close browser via playwright_cli command=close. Return compact JSON (<=1200 chars): {caseId,status,evidencePaths,errorSummary,tokens}.",
                 ].join("\n\n"),
@@ -4453,7 +4462,7 @@ function buildFrontendTestHybridDag(sources) {
                     subtaskPromptTemplate: [
                         "RERUN attempt {{case.rerunAttempt}} for {{case.caseId}} (reason={{case.reason}}). Rewrite authoritative evidence; set rerunAttempt={{case.rerunAttempt}}.",
                         "Primary job: EXECUTE case {{case.caseId}} from {{case.casePath}} with skill playwright-cli (fresh Pi session). Use structured playwright_cli only; headless open.",
-                        `Start via playwright_cli command=open with args [--browser=chrome, --headless, ${controllerFrontend.baseUrl}]. Passed requires open → find → cleanup receipts.`,
+                        `Start via playwright_cli command=open with args [--browser=chrome, ${controllerFrontend.baseUrl}]. Passed requires open → find → cleanup receipts.`,
                         "Always write {{case.evidenceDir}}execution.md and {{case.evidenceDir}}case-result.json with caseId, status, evidencePaths, rerunAttempt={{case.rerunAttempt}}.",
                     ].join("\n\n"),
                 },
@@ -4543,7 +4552,7 @@ function buildFrontendTestHybridDag(sources) {
         "frontend-test-dag generates Markdown cases and browser evidence only; it must not generate pytest or Playwright test source code.",
         "Each browser case runs serially in a fresh Pi execution boundary. Persist its evidence before starting the next case.",
         "Use only the declared isolated test environment. Production URLs, real credentials, and unauthorized data are blocked.",
-        `Browser startup for generated cases must be playwright-cli open --browser=chrome --headless ${controllerFrontend.baseUrl}; this origin is frozen by the controller from ${controllerFrontend.baseUrlSource}, and model-authored files cannot establish or override it; generated operations stay in the default browser session and must not use unverified named-session flags.`,
+        `Browser startup for generated cases must be playwright-cli open --browser=chrome ${controllerFrontend.baseUrl}; this origin is frozen by the controller from ${controllerFrontend.baseUrlSource}, and model-authored files cannot establish or override it; generated operations stay in the default browser session and must not use unverified named-session flags.`,
         "Browser-tool preflight runs before any frontend-test Pi node; cli-only rollback, missing/incompatible Pi SDK custom-tool capability, missing verified playwright-cli launcher, or incompatible --help fails with zero Pi calls.",
         "Browser case children use commandPolicy capability-allowlist playwright-cli and structured playwright_cli tool; ordinary writers remain without Bash.",
         "Passed cases require same-child ordered controller receipts: successful open → successful find → successful post-execution cleanup. Pre-start cleanup, snapshot/goto/screenshot/request/console and ordinary interactions are insufficient; missing or unordered receipts convert to blocked (browser-command-evidence-missing).",

package/dist/workflows/dag/node-execution.js

@@ -8,6 +8,7 @@ import { writeNodeRecord, writeNodeSkillArtifacts } from "./run-store.js";
 import { resolveContextPolicy } from "./context-policy.js";
 import { buildDagNodePromptEnvelope, formatConvergenceFeedbackBlock, } from "./prompt.js";
 import { persistLongNodeOutputArtifacts } from "./upstream-artifacts.js";
+import { buildOutputLimitRecoverySection, loadBackendTestWriterProgressForRetry, } from "./backend-test-writer-completeness.js";
 import { computeBackoffDelayMs, isRetryablePiFailureCategory, isSafeReadOnlyPiRetryCandidate, isWriterEmptyDiffRetryCandidate, } from "./retry-policy.js";
 import { applyNodeActivity, evaluateNodeLiveness, resolveLivenessPolicy, } from "./liveness-policy.js";
 import { buildProtocolRetryInstruction, normalizeReviewVerdictAfterRetries, parseJsonReviewVerdict, validateOutputProtocol, } from "./output-protocol.js";
@@ -125,7 +126,7 @@ export function buildNodePrompt(spec, task, upstream, options) {
         convergenceFeedback: options?.convergenceFeedback,
     });
 }
-function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCategory, previousProtocolReason) {
+function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCategory, previousProtocolReason, recoveryTargetPaths) {
     if (attemptNumber <= 1)
         return basePrompt;
     if (previousFailureCategory === "protocol-invalid" &&
@@ -138,6 +139,22 @@ function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCate
         ].join("\n");
     }
     if (previousFailureCategory === "writer-empty-diff") {
+        const maxAttempts = task.retryPolicy?.maxAttempts ?? 3;
+        // When a completeness progress exists for this writer, fold the concrete
+        // target paths into the empty-diff retry so the model does not guess and
+        // does not need to read a forbidden `.harness/**` evidence file.
+        if (recoveryTargetPaths && recoveryTargetPaths.length > 0) {
+            return [
+                basePrompt,
+                "",
+                buildOutputLimitRecoverySection({
+                    attempt: attemptNumber,
+                    maxAttempts,
+                    reason: "T4",
+                    targetPaths: recoveryTargetPaths,
+                }),
+            ].join("\n");
+        }
         return [
             basePrompt,
             "",
@@ -150,6 +167,22 @@ function buildAttemptPrompt(task, basePrompt, attemptNumber, previousFailureCate
     }
     if (previousFailureCategory === "incomplete-write-set") {
         const maxAttempts = task.retryPolicy?.maxAttempts ?? 3;
+        // Embed concrete missing/broken paths from the run-owned progress facts
+        // so the continuation attempt is fully self-contained and never reads
+        // a forbidden `.harness/**` evidence file. When the loader finds no
+        // progress facts yet (rare), fall back to the path-pointing contract.
+        if (recoveryTargetPaths) {
+            return [
+                basePrompt,
+                "",
+                buildOutputLimitRecoverySection({
+                    attempt: attemptNumber,
+                    maxAttempts,
+                    reason: "T3_or_T5",
+                    targetPaths: recoveryTargetPaths,
+                }),
+            ].join("\n");
+        }
         return [
             basePrompt,
             "",
@@ -597,12 +630,18 @@ export async function executeDagNode(input) {
                 tasksById,
                 state,
             });
+            // For backend-test generation writers, load the most recent
+            // completeness progress from the run dir so the next attempt's
+            // prompt embeds concrete target paths. Undefined for non-writers
+            // or when no progress facts exist yet (no-op).
+            const recoveryProgress = await loadBackendTestWriterProgressForRetry(runDir, task.id);
+            const recoveryTargetPaths = recoveryProgress?.targetPaths;
             result = await executeNode({
                 task,
                 cwd,
                 model,
                 ...(thinking ? { thinking } : {}),
-                prompt: buildAttemptPrompt(task, prompt, attemptNumber, previousFailureCategory, previousProtocolReason),
+                prompt: buildAttemptPrompt(task, prompt, attemptNumber, previousFailureCategory, previousProtocolReason, recoveryTargetPaths),
                 attempt: attemptNumber,
                 reportActivity,
                 timeoutMs: livenessPolicy.absoluteMaxWallClockMs,

package/docs/templates/frontend-test-case-checklist.md

@@ -7,7 +7,7 @@ LLM review (when `frontendTest.reviewMode=blocking`) must not invent blocking ru
 
 | ruleId | Rule |
 |---|---|
-| `open-prefix` | Each case body includes `playwright-cli open --browser=chrome --headless <absolute-http(s)-url>` |
+| `open-prefix` | Each case body includes `playwright-cli open --browser=chrome <absolute-http(s)-url>` |
 | `production-url` | Open URL must not look like a production host |
 | `ac-mapping` | Manifest entry has non-empty `acIds` (**acceptance** ids, not case ids) |
 | `ac-id-shape` | Each `acIds[]` entry matches `AC-*` / `AC-FE-*` |

package/docs/templates/frontend-test-dag.generate-cases.prompt.md

@@ -18,7 +18,7 @@ Each case is independently executable and includes AC mapping (`acIds`), precond
 
 Every case must copy the **controller-frozen** absolute base URL supplied by the DAG. `testcase/frontend/rag/context.md` may reference that value, but model-authored context/case prose cannot establish or override the origin. Do not leave a base-url placeholder.
 
-Every case must start with `playwright-cli open --browser=chrome --headless` followed by the concrete controller-frozen URL supplied by the DAG. Do not copy an angle-bracket URL placeholder into a generated executable case line.
+Every case must start with `playwright-cli open --browser=chrome` followed by the concrete controller-frozen URL supplied by the DAG. Do not copy an angle-bracket URL placeholder into a generated executable case line.
 
 ### Dynamic element refs (hard)
 

package/docs/templates/frontend-test-dag.json

@@ -11,7 +11,7 @@
   "globalConstraints": [
     "Do not generate pytest or Playwright source code.",
     "Only use declared isolated test environments; production URLs and real credentials are blocked.",
-    "Every generated browser start command uses playwright-cli open --browser=chrome --headless followed by the concrete controller-resolved URL (from task source config.md when present, else http://localhost:5173); executable case lines never retain an angle-bracket URL/ref placeholder; subsequent commands stay in that default session and must not use unverified named-session flags.",
+    "Every generated browser start command uses playwright-cli open --browser=chrome followed by the concrete controller-resolved URL (from task source config.md when present, else http://localhost:5173); executable case lines never retain an angle-bracket URL/ref placeholder; subsequent commands stay in that default session and must not use unverified named-session flags.",
     "Case children execute serially. Persist each case result, logs and browser evidence before the next child starts.",
     "A token threshold is a post-case stop check, not a model hard token cap; unstarted cases must be recorded as blocked: token-budget-exhausted.",
     "Default pipeline acceptance is the final frontend-test-result-v1 plus testcase/frontend/reports/frontend-test-report.html",

package/docs/templates/frontend-test-dag.retrieve-context.prompt.md

@@ -10,7 +10,7 @@ The preflight controller resolves and freezes one absolute browser base URL from
 - Never use production hosts or credentials.
 - Write `environmentProbe: pending`. The environment preflight shell will replace this with `reachable`, `unreachable`, or `curl-unavailable` plus a structured `blockedReason` (for example `frontend-base-url-unreachable`).
 - Include the exact browser start prefix using the supplied controller-frozen URL:
-  `playwright-cli open --browser=chrome --headless <controller-frozen-base-url>`.
+  `playwright-cli open --browser=chrome <controller-frozen-base-url>`.
 
 Do not claim the environment is reachable until preflight completes. Preflight does not start the application.
 

package/docs/templates/frontend-test-dag.review-cases.prompt.md

@@ -2,4 +2,4 @@
 
 Read the RAG files and Markdown cases only. First line must be `VERDICT: pass` or `VERDICT: request-revision`. Report AC coverage, case independence, evidence completeness, unsafe environment/data dependencies, and manifest issues. This verdict is a deterministic safety gate: `request-revision` blocks manifest materialization and browser execution.
 
-Every case must retain the exact browser-start command prefix with the resolved absolute baseUrl from `testcase/frontend/rag/context.md` (prefer task source `config.md`, else `http://localhost:5173`): `playwright-cli open --browser=chrome --headless <resolved-base-url>`; session flags must not precede `open`, and subsequent commands must remain in its default session without `-s=` or assumed named-session binding. Verify every executable sub-scenario specifies fixture/reset, UI reset, fresh snapshot before element refs, a successful `playwright-cli find ...` semantic assertion after `open`, and an evidence write point. `snapshot`, `goto`, `screenshot`, `request`/`console`, and ordinary interactions cannot independently authorize passed. Verify each case requires both `execution.md` and `case-result.json` under its own evidence directory. The JSON result must contain matching `caseId`, `status` (`passed`, `failed`, or `blocked`) and `evidencePaths`; blocked cases must name a non-empty `blockedReason` and cannot count as passed.
+Every case must retain the exact browser-start command prefix with the resolved absolute baseUrl from `testcase/frontend/rag/context.md` (prefer task source `config.md`, else `http://localhost:5173`): `playwright-cli open --browser=chrome <resolved-base-url>`; session flags must not precede `open`, and subsequent commands must remain in its default session without `-s=` or assumed named-session binding. Verify every executable sub-scenario specifies fixture/reset, UI reset, fresh snapshot before element refs, a successful `playwright-cli find ...` semantic assertion after `open`, and an evidence write point. `snapshot`, `goto`, `screenshot`, `request`/`console`, and ordinary interactions cannot independently authorize passed. Verify each case requires both `execution.md` and `case-result.json` under its own evidence directory. The JSON result must contain matching `caseId`, `status` (`passed`, `failed`, or `blocked`) and `evidencePaths`; blocked cases must name a non-empty `blockedReason` and cannot count as passed.

package/docs/templates/init-managed-agents.md

@@ -32,10 +32,10 @@
 | --- | --- |
 | **允许** | 已发布 `loop-agent` / `agent-worker` CLI；只读 status/doctor/report/inspect/observe；准备 `source/*` 与 `task.json` 边界；human gate；shell 验证与 handoff。 |
 | **禁止** | 绕过 CLI 用宿主 Edit/Write/ApplyPatch 直接改业务实现；CLI/DAG 失败后「救火改文件」；用聊天自述代替 shell 验证。 |
-| **失败时** | `dag doctor` / `dag report` / `dag reconcile-run`（及适用 worker reconcile）；修正任务源/`task.json`/DAG 后经 CLI 重跑。 |
+| **失败时** | 先 `dag report` / `dag doctor`；优先 `dag rerun --from-node` 安全子图续跑；仅契约/源真变或 plan 不合格时同 task 重 advance / `dag rerun-task`；paused 用 approve→`dag resume`。 |
 | **实现写入** | 业务代码 **只** 经受治理 DAG writer（`implement-pi` / `repair-pi`）经 `task advance`（批准 writeSet gate 后长跑）。 |
 
-**永远不要**：`loop-agent` / `agent-worker` 失败 ⇒ 主会话直接改仓库实现。
+**永远不要**：`loop-agent` / `agent-worker` 失败 ⇒ 主会话直接改仓库实现；也勿因 provider 抖动或只读节点失败**新建无关 task-id**。
 
 ### 自然语言入口路由
 
@@ -47,6 +47,7 @@
 | 初始化对齐 / 升级后对齐 / reconcile 初始化 / loop-agent 初始化对齐 | 初始化升级闭环 | 直接运行 `init upgrade`，而非拼接低层 safe-update 命令 |
 | 初始化安全更新 / 应用初始化更新 / loop-agent 初始化安全更新 / loop agent初始化安全更新 | 初始化升级闭环 | 直接运行 `init upgrade`，由 controller 处理安全更新与验证 |
 | loop-agent 帮我完成 / 帮我实现 / 帮我修复 / 帮我开发 <需求>；使用 loop-agent 完成 <X>；按 loop-agent 流程处理 <X> | 通用需求实现 | 先提供 PRD 与 `allowedPaths` / `forbiddenPaths` / verify，再 `task advance` → 审查 writeSet gate → `task advance --approve-gate`；主会话不得直接修改业务实现 |
+| 从失败节点继续 / 续跑 / 不要重开任务 / 从 plan 或 code-review 接着跑 | DAG 中途恢复 | **不是**新 `task advance`：`dag report` → `dag rerun --from-node … --plan` → 带 `plan-hash` 执行（见失败默认恢复序） |
 
 **初始化升级闭环（写入型）**：主会话收到上述写入型表达时自动调用统一入口，并在 controller 返回的具体单文件 `allowedPaths` 内完成必要语义合并后使用 `--continue`。不得停在 `check-update`、`needs-safe-update`、`needs-model-merge` 或 `verification-pending`。默认安装/合并项目 `.opencode/plugins/`、`.pi/extensions/` 与 `.pi/settings.json`；Pi 必须 trust 项目才能加载，默认严禁写 `~/.pi/agent/settings.json`。
 
@@ -136,7 +137,9 @@ agent-worker console serve --repo . --port 8790   # 兼容入口，等价于上
 
 live run 先用 `loop-agent dag status --run-id <run-id>` 看 lifecycle 与 liveness；用 `loop-agent dag report --run-id <run-id> --markdown` 读 facts；失败/paused 用 `loop-agent dag doctor --run-id <run-id> --markdown`。生命周期对齐先只读运行 `loop-agent dag reconcile-run --run-id <run-id>`；只有 runner 已停止且 operator 明确提供 `--action supersede|abandon --reason "..."` 时才允许变更。失败 run 用 `loop-agent dag closeout-draft --run-id <run-id>` 生成 failure handoff，不要写成成功 closeout。
 
-Operator 须持续监控 live run，直到 controller 报告 run 已结束（节点/流程终态如 `FINISHED`、`FAILED` 或 `partial_failed`），或 Decision Gate **需要 approve**；不要在节点仍运行时假定完成。持续监视过程中，主会话可在合适节点（例如节点/rank 状态变化、进入 verify/closeout、出现 stall 嫌疑或需要 approve 时）向用户做简短进度汇报（当前节点、状态、是否有风险），避免长时间静默；汇报是告知，不是请求确认，不得因此停下流程。判活须组合 runner heartbeat、session events 与 `dag doctor` liveness/provider meaningful progress；Runner heartbeat 只证明 lease，alone ≠ progress，不得仅凭运行时长结束节点。exclusive writer（如 `implement-pi` / `repair-pi`）运行期间：主会话与其他 writer **不得并发修改工作区**，以免 write-guard 错误归因；只读 status/doctor/report/observe 与 approve/reject/resume CLI 仍允许。恢复：status/doctor/report → classify → reconcile/replan → CLI 重跑 → shell verify。**禁止**把主会话直接 Edit 业务代码当作恢复手段。
+**失败默认恢复序**：`dag report`/`dag doctor` → 优先 `dag rerun --from-node <node> --plan` 再带 `--plan-hash`（provider 抖动、plan/review/verify 安全下游；writer/decision/fingerprint 不合格勿硬跑）→ paused：approve→`dag resume` → 契约/源真变或 R1 不合格：同 task `task advance` / `dag rerun-task`（禁无理由新建 task-id）→ Worker-owned：`agent-worker task retry`。
+
+Operator 须监控 live run 至终态（FINISHED / FAILED / partial_failed）或 Decision Gate 需要 approve；可在节点/rank 变化、verify/closeout、stall 或需 approve 时简短汇报（告知非请求确认）。判活须组合 runner heartbeat、session events 与 `dag doctor` liveness/provider meaningful progress（heartbeat alone ≠ progress）。bounded writer 运行期间不得并发修改工作区（write guard / write-guard 会把越界 diff 错误归因到 writer）；只读 status/doctor/report 与 approve/reject/resume 仍允许。恢复：report/doctor → **优先** `dag rerun --from-node` → 必要时 rerun-task/同 task advance → shell verify。**禁止**把主会话直接 Edit 业务代码当作恢复手段。
 
 ### 运行态与验证
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tea-agent/loop-agent",
-  "version": "0.29.1",
+  "version": "0.29.3",
   "type": "module",
   "bin": {
     "loop-agent": "bin/loop-agent.js",

package/skills/playwright-cli/SKILL.md

@@ -58,7 +58,7 @@ request
 获取 snapshot，并只使用 snapshot 中可见的 ref 或已知安全 locator。
 
 ```text
-playwright-cli open --browser=chrome --headless http://localhost:5173
+playwright-cli open --browser=chrome http://localhost:5173
 playwright-cli snapshot
 playwright-cli find "Sign in"
 playwright-cli fill e1 "user@example.com"

package/skills/playwright-cli-case-generator/SKILL.md

@@ -41,7 +41,7 @@ description: 根据 FE-test RAG 知识包生成可由受限 playwright_cli runti
    `-s=<case-id>` 建立 named session。
 3. 可独立执行的命令序列：使用 RAG `context.md` 中已解析的绝对 `baseUrl`（优先来自任务源
    `config.md`；缺失时默认 `http://localhost:5173`），必须以
-   `playwright-cli open --browser=chrome --headless` 后接解析出的具体 base URL 开始；不得保留
+   `playwright-cli open --browser=chrome` 后接解析出的具体 base URL 开始；不得保留
    base-url 占位符，也不得使用 `-s=<case-id>` 或其他 named session；再按需登录、数据准备、
    `snapshot` 后优先使用元素引用和操作。每个拟通过场景必须在 `open` 后以成功的 `find` 作为
    唯一结构化 UI 语义断言，再由 controller 执行 post-execution cleanup。`snapshot`、`goto`、