@tea-agent/loop-agent 0.38.3 → 0.38.4

package/dist/commands/inspect-next.js

@@ -0,0 +1,8 @@
+import { collectNextInfo, formatNextInfo, } from "../../scripts/next-info.mjs";
+export async function runInspectNext(repoRoot, args = []) {
+    const json = args.includes("--json");
+    const result = collectNextInfo({ cwd: repoRoot });
+    process.stdout.write(formatNextInfo(result, { json }));
+    if (!result.ok)
+        process.exitCode = result.exitCode || 1;
+}

package/dist/executors/dag-pi-executor.js

@@ -9,7 +9,7 @@ import { cleanupPlaywrightCliDefaultSession, createPlaywrightCliTool, PI_COMMAND
 import { dagCommandPolicyAllows, resolveDagCommandPolicy, } from "../workflows/dag/types.js";
 import { redactPromptForLog, truncateOutput, } from "../shared/output-truncation.js";
 import { GitStatusUnavailableError, pathsChangedDuringRun, readGitStatusPorcelain, recoverRootNulArtifact, snapshotGitStatusPathFingerprints, snapshotGitStatusPorcelain, validateShellWriteGuard, } from "./shell-write-guard.js";
-import { isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, } from "../workflows/dag/retry-policy.js";
+import { isTargetTemplateTransientRetryNode, isWriterEmptyDiffRetryCandidate, isWriterTransportRetryCandidate, INCOMPLETE_WRITE_SET_RETRY_CATEGORY, WRITER_CLEAN_TIMEOUT_RETRY_CATEGORY, WRITER_EMPTY_DIFF_RETRY_CATEGORY, } from "../workflows/dag/retry-policy.js";
 import { assessBackendTestMdPlanCompleteness, assessBackendTestMdWriterCompleteness, assessBackendTestPytestPlanCompleteness, assessBackendTestPytestWriterCompleteness, assessBackendTestShardChildCompleteness, backendTestWriterProgressRoleForTask, classifyBackendTestWriterCompletenessFailure, isBackendTestCompletenessRetryCandidate, isBackendTestMdPlanTask, isBackendTestPytestPlanTask, isBackendTestShardChildTask, writeBackendTestWriterProgressArtifacts, } from "../workflows/dag/backend-test-writer-completeness.js";
 import { resolveBackendTestLayout } from "../workflows/dag/backend-test-layout.js";
 import { redactSecrets, truncateUtf8Preview } from "../shared/preview.js";
@@ -942,22 +942,31 @@ export async function executeDagPiNode(input, meta, piStepFn = executePiStep, wr
     if (meta.writeGuardAttribution === "best-effort") {
         stderrParts.push("write guard note: concurrent rank writers use best-effort per-node attribution; keep same-rank writeSet entries disjoint");
     }
-    // Pure provider stall/timeout with zero workspace side effects is eligible
-    // for a single writer transport retry (implement-pi flake resilience).
-    // Partial writes, tool calls, or write-guard failures keep raw `timeout`.
-    const writerCleanTimeout = !mapped.ok &&
-        mapped.failureCategory === "timeout" &&
-        writeGuardOk &&
+    const rawFailureCategory = mapped.failureCategory;
+    const writeToolCallCount = readWriterThinkingExhaustionEvidence(result).writeToolCallCount;
+    const zeroSideEffects = writeGuardOk &&
         !completenessFailure &&
         !writerThinkingExhausted &&
-        isWriterTransportRetryCandidate(input.task) &&
+        !writerOutcomeViolation &&
         changeManifestChangedFiles !== undefined &&
         changeManifestChangedFiles.length === 0 &&
-        readWriterThinkingExhaustionEvidence(result).writeToolCallCount === 0;
+        writeToolCallCount === 0;
+    const targetCleanTransport = !mapped.ok &&
+        isTargetTemplateTransientRetryNode(input.task) &&
+        rawFailureCategory !== undefined &&
+        ["timeout", "network", "rate-limit", "unavailable"].includes(rawFailureCategory) &&
+        zeroSideEffects;
+    const frontendCleanTimeout = !mapped.ok &&
+        rawFailureCategory === "timeout" &&
+        isWriterTransportRetryCandidate(input.task) &&
+        !isTargetTemplateTransientRetryNode(input.task) &&
+        zeroSideEffects;
+    const writerCleanTimeout = targetCleanTransport || frontendCleanTimeout;
     return {
         ...mapped,
         ok: false,
         stderr: stderrParts.filter(Boolean).join("\n\n"),
+        ...(rawFailureCategory ? { rawFailureCategory } : {}),
         failureCategory: mapped.ok
             ? writeGuardOk
                 ? completenessFailure

package/dist/executors/pi-playwright-cli-tool.js

@@ -4,7 +4,7 @@ import { appendFile, lstat, mkdir, readFile, realpath, stat, } from "node:fs/pro
 import path from "node:path";
 import { processTreeSpawnOptions, terminateProcessTree, } from "./process-tree.js";
 import { resolvePlaywrightCliLauncher } from "./playwright-cli-launcher.js";
-import { PLAYWRIGHT_CLI_ALLOWED_COMMANDS, isPlaywrightCliCommand, } from "../shared/playwright-cli-command-policy.js";
+import { PLAYWRIGHT_CLI_ALLOWED_COMMANDS, isPlaywrightCliCommand, parsePlaywrightCliFindMatchCount, } from "../shared/playwright-cli-command-policy.js";
 export { PLAYWRIGHT_CLI_ALLOWED_COMMANDS, isPlaywrightCliCommand, };
 /** Logical capability name only; execution always uses the verified JS launcher. */
 export const PLAYWRIGHT_CLI_EXECUTABLE = "playwright-cli";
@@ -111,6 +111,13 @@ export async function readPlaywrightCliReceipts(runDir, nodeId) {
     }
 }
 export function summarizePlaywrightCliReceipts(receipts) {
+    // Sequence numbers are assigned at dispatch time, which is the controller's
+    // causal order. Receipt lines are appended when each command COMPLETES, so
+    // concurrent commands (e.g. parallel find calls issued in one turn) persist
+    // out of sequence order. Judge chain contiguity on the dispatch order by
+    // sorting on `sequence` first; a genuinely missing receipt (gap) or a
+    // duplicated sequence still fails the contiguity check below.
+    const orderedReceipts = [...receipts].sort((a, b) => a.sequence - b.sequence);
     let hasSuccessfulOpen = false;
     let hasInteractionOrAssertion = false;
     let hasMeaningfulAssertion = false;
@@ -122,7 +129,7 @@ export function summarizePlaywrightCliReceipts(receipts) {
     let hasContiguousControllerSequence = true;
     let previousSequence = 0;
     const caseIds = new Set();
-    for (const receipt of receipts) {
+    for (const receipt of orderedReceipts) {
         caseIds.add(receipt.caseId);
         const succeeded = receipt.exitCode === 0 && !receipt.timedOut;
         const ordered = receipt.sequence === previousSequence + 1;
@@ -137,8 +144,10 @@ export function summarizePlaywrightCliReceipts(receipts) {
         if (PLAYWRIGHT_CLI_INTERACTION_COMMANDS.has(receipt.command) && succeeded) {
             hasInteractionOrAssertion = true;
         }
-        // A successful find is the only model-visible semantic assertion in v1.
-        if (receipt.command === "find" && succeeded) {
+        // A successful find (or find-unique, its semantic wrapper) is the only
+        // model-visible semantic assertion in v1.
+        if ((receipt.command === "find" || receipt.command === "find-unique") &&
+            succeeded) {
             hasMeaningfulAssertion = true;
             if (ordered && sawOpen)
                 sawMeaningfulAssertionAfterOpen = true;
@@ -523,6 +532,17 @@ export function preparePlaywrightCliArgv(input, ctx) {
     else if (command === "pdf" || command === "snapshot") {
         args = normalizePdfOrSnapshotArgs(command, rawArgs, ctx);
     }
+    else if (command === "find-unique") {
+        // find-unique is a controller-side semantic wrapper: the underlying CLI
+        // only knows `find`; uniqueness (matchCount === 1) is judged by the
+        // executor after parsing the find output. Require exactly one text
+        // literal — a unique-control assertion is meaningless with multiple.
+        const literals = rawArgs.filter((arg) => !arg.startsWith("-"));
+        if (rawArgs.length !== 1 || literals.length !== 1) {
+            throw new PlaywrightCliPolicyError("find-unique-args", "find-unique requires exactly one text literal and no flags");
+        }
+        args = literals;
+    }
     else {
         args = rewriteArgsForPaths(command, rawArgs, ctx);
     }
@@ -818,7 +838,10 @@ export async function executePlaywrightCliCommand(input, ctx, signal) {
         };
     }
     let executable = PLAYWRIGHT_CLI_EXECUTABLE;
-    let argv = [command, ...args];
+    // find-unique executes as the underlying CLI `find` command; uniqueness is
+    // a controller-side judgment on the parsed match count.
+    const cliCommand = command === "find-unique" ? "find" : command;
+    let argv = [cliCommand, ...args];
     try {
         await validatePlaywrightCliOutputPath(command, args, ctx);
         await validatePlaywrightCliInputFiles(command, args, ctx);
@@ -827,7 +850,7 @@ export async function executePlaywrightCliCommand(input, ctx, signal) {
         if (!ctx.runCommand) {
             const launcher = resolvePlaywrightCliLauncher(ctx.repoRoot);
             executable = launcher.executable;
-            argv = [...launcher.argvPrefix, command, ...args];
+            argv = [...launcher.argvPrefix, cliCommand, ...args];
         }
     }
     catch (error) {
@@ -887,6 +910,19 @@ export async function executePlaywrightCliCommand(input, ctx, signal) {
     }
     const finishedAt = new Date();
     const combined = `${runResult.stdout}\n${runResult.stderr}`;
+    // find-family commands report their match count in stdout with exit code
+    // 0 even on zero matches — parse the count so receipts carry the real
+    // assertion signal for finalize-side audits.
+    const matchCount = command === "find" || command === "find-unique"
+        ? parsePlaywrightCliFindMatchCount(runResult.stdout || "")
+        : undefined;
+    // find-unique fails closed unless exactly one match was found (including
+    // zero and unparseable outputs).
+    const findUniqueFailed = command === "find-unique" && matchCount !== 1 && !runResult.timedOut;
+    const effectiveExitCode = findUniqueFailed && runResult.exitCode === 0 ? 1 : runResult.exitCode;
+    const effectiveStderr = findUniqueFailed
+        ? `${runResult.stderr}${runResult.stderr ? "\n" : ""}find-unique: expected exactly 1 match, got ${matchCount === undefined ? "unparseable output" : `${matchCount}`}`
+        : runResult.stderr;
     const receipt = {
         schemaVersion: 1,
         caseId: ctx.caseId,
@@ -895,25 +931,28 @@ export async function executePlaywrightCliCommand(input, ctx, signal) {
         argsRedacted: redactPlaywrightCliArgs(command, args),
         startedAt: startedAt.toISOString(),
         finishedAt: finishedAt.toISOString(),
-        exitCode: runResult.exitCode,
+        exitCode: effectiveExitCode,
         durationMs: finishedAt.getTime() - startedAt.getTime(),
         outputSha256: createHash("sha256").update(combined).digest("hex"),
         toolVersion: ctx.toolVersion ?? "unknown",
         timedOut: runResult.timedOut || undefined,
+        ...(matchCount !== undefined ? { matchCount } : {}),
         errorClass: runResult.timedOut
             ? "timeout"
             : runResult.signal === "abort"
                 ? "abort"
-                : runResult.exitCode === 0
+                : effectiveExitCode === 0
                     ? undefined
-                    : "nonzero-exit",
+                    : findUniqueFailed
+                        ? "find-unique-count-mismatch"
+                        : "nonzero-exit",
     };
     await appendReceipt(ctx, receipt);
     return {
-        ok: runResult.exitCode === 0 && !runResult.timedOut,
-        exitCode: runResult.exitCode,
+        ok: effectiveExitCode === 0 && !runResult.timedOut,
+        exitCode: effectiveExitCode,
         stdout: runResult.stdout,
-        stderr: runResult.stderr,
+        stderr: effectiveStderr,
         timedOut: runResult.timedOut,
         receipt,
         errorClass: receipt.errorClass,

package/dist/executors/shell-executor.js

@@ -123,6 +123,39 @@ function shouldReuseSuccessfulShellCommands(shell) {
 const DEFAULT_SHELL_TIMEOUT_MS = 300_000;
 const SUMMARY_STDOUT_MAX = 4_000;
 const SUMMARY_STDERR_MAX = 2_000;
+const TEST_TIMEOUT_CLASSIFY_MAX_CHARS = 8_000;
+const TEST_TIMEOUT_MARKERS = [
+    /Test timed out in \d+ms/i,
+    /Timeout of \d+ms exceeded/i,
+];
+const DETERMINISTIC_NON_TIMEOUT_FAILURE_MARKERS = [
+    /AssertionError/i,
+    /expected:?\s/i,
+    /received:?\s/i,
+    /Snapshot /i,
+    /toMatchSnapshot/i,
+    /SyntaxError/i,
+    /TS\d{4}/,
+    /error TS/i,
+    /eslint/i,
+    /compile error/i,
+    /Failed to compile/i,
+];
+function isIdentifiedTestCommand(command) {
+    return /(?:^|[\s"'`])(?:npx\s+)?vitest(?:\s|$)|(?:^|[\s"'`])npm(?:\.cmd)?\s+test(?:\s|$)|(?:^|[\s"'`])pnpm(?:\.cmd)?\s+test(?:\s|$)|(?:^|[\s"'`])yarn(?:\.cmd)?\s+test(?:\s|$)/i.test(command);
+}
+function classifyShellCommandFailure(input) {
+    if (input.processCategory !== "nonzero-exit")
+        return input.processCategory;
+    if (!isIdentifiedTestCommand(input.command))
+        return "nonzero-exit";
+    const sample = `${input.stdout}\n${input.stderr}`.slice(0, TEST_TIMEOUT_CLASSIFY_MAX_CHARS);
+    const hasTimeoutMarker = TEST_TIMEOUT_MARKERS.some((marker) => marker.test(sample));
+    if (!hasTimeoutMarker)
+        return "nonzero-exit";
+    const mixedDeterministicFailure = DETERMINISTIC_NON_TIMEOUT_FAILURE_MARKERS.some((marker) => marker.test(sample));
+    return mixedDeterministicFailure ? "nonzero-exit" : "test-timeout";
+}
 const NODE_STDOUT_MAX = 12_000;
 const NODE_STDERR_MAX = 8_000;
 const MAX_BUFFERED_OUTPUT_BYTES = 64 * 1024;
@@ -329,16 +362,21 @@ export async function executeShellCommand(input) {
             });
         });
         child.on("close", (code) => {
-            let failureCategory = "nonzero-exit";
+            let processCategory = "nonzero-exit";
             if (timedOut) {
-                failureCategory = "timeout";
+                processCategory = "timeout";
             }
             else if (code === 0) {
-                failureCategory = "success";
+                processCategory = "success";
             }
             finish({
                 exitCode: code,
-                failureCategory,
+                failureCategory: classifyShellCommandFailure({
+                    command: input.command,
+                    processCategory,
+                    stdout: formatBoundedOutput(stdout),
+                    stderr: formatBoundedOutput(stderr),
+                }),
                 ok: code === 0 && !timedOut,
             });
         });
@@ -2348,7 +2386,76 @@ async function executeFrontendLintBaseline(input, meta) {
         };
     }
 }
-export async function executeDagShellNode(input, meta) {
+async function writeShellPresetProcessReceipt(input) {
+    await writeDagNodeJsonArtifact(input.runDir, input.nodeId, "process-receipt.json", {
+        schemaVersion: 1,
+        kind: "shell-preset",
+        preset: input.preset,
+        ok: input.ok,
+        durationMs: input.durationMs,
+        steps: input.steps,
+    });
+}
+async function ensureShellPresetProcessReceipt(input) {
+    const receiptPath = path.join(input.runDir, input.nodeId, "process-receipt.json");
+    if (existsSync(receiptPath))
+        return;
+    if (existsSync(path.join(input.runDir, input.nodeId, "shell-command-receipt.json"))) {
+        return;
+    }
+    const detail = input.result.ok ? input.result.stdout : input.result.stderr;
+    await writeShellPresetProcessReceipt({
+        runDir: input.runDir,
+        nodeId: input.nodeId,
+        preset: input.preset,
+        ok: input.result.ok,
+        durationMs: input.result.durationMs,
+        steps: [
+            {
+                id: "run",
+                label: input.preset,
+                ok: input.result.ok,
+                detail: detail?.slice(0, 8 * 1024),
+                durationMs: input.result.durationMs,
+            },
+        ],
+    });
+}
+function namedShellHandler(shell) {
+    if (!shell)
+        return null;
+    const record = shell;
+    for (const key of [
+        "finalWriteSetApprovalGate",
+        "frontendTestStandardScenarios",
+        "frontendTestEnvironmentProbe",
+        "frontendBrowserToolPreflight",
+        "frontendPrewriteGate",
+        "frontendLintBaseline",
+        "frontendVerificationBundle",
+        "frontendReviewContext",
+        "frontendTestCaseChecklist",
+        "frontendTestCaseManifest",
+        "frontendTestEvidenceValidation",
+        "frontendTestResultFinalize",
+        "frontendTestL5Report",
+        "frontendTestHtmlReport",
+        "frontendTestReports",
+        "backendTestPipeline",
+        "jsonArtifactGate",
+        "projectGovernanceGate",
+        "requirementCoverageGate",
+        "verdictGate",
+        "verifyEvidence",
+        "repairArtifactGate",
+        "jacocoCoverage",
+    ]) {
+        if (record[key] != null)
+            return key;
+    }
+    return null;
+}
+async function executeDagShellNodeInner(input, meta) {
     const shell = input.task.shell;
     if (shell?.finalWriteSetApprovalGate) {
         const started = Date.now();
@@ -2395,28 +2502,81 @@ export async function executeDagShellNode(input, meta) {
     }
     if (shell?.frontendBrowserToolPreflight) {
         const started = Date.now();
+        const steps = [];
+        const finish = async (result) => {
+            await writeShellPresetProcessReceipt({
+                runDir: meta.runDir,
+                nodeId: input.task.id,
+                preset: "frontendBrowserToolPreflight",
+                ok: result.ok,
+                durationMs: result.durationMs,
+                steps: steps.length > 0
+                    ? steps
+                    : [
+                        {
+                            id: "run",
+                            label: "浏览器工具能力探测",
+                            ok: result.ok,
+                            detail: result.ok ? result.stdout : result.stderr,
+                            durationMs: result.durationMs,
+                        },
+                    ],
+            });
+            return result;
+        };
         try {
             if (resolvePiBackend() === "cli-only") {
+                steps.push({
+                    id: "pi-backend",
+                    label: "Pi SDK 后端可用",
+                    ok: false,
+                    detail: "CODE_AGENT_PI_BACKEND=cli-only",
+                });
                 throw new Error("browser-command-capability-unavailable: CODE_AGENT_PI_BACKEND=cli-only rejects SDK-only playwright_cli");
             }
+            steps.push({
+                id: "pi-backend",
+                label: "Pi SDK 后端可用",
+                ok: true,
+            });
             const sdkCapability = await checkPiSdkCustomToolCapability(input.cwd);
             if (!sdkCapability.ok) {
+                steps.push({
+                    id: "sdk-custom-tool",
+                    label: "Pi SDK custom-tool 能力",
+                    ok: false,
+                    detail: sdkCapability.detail,
+                });
                 throw new Error(`browser-command-capability-unavailable: ${sdkCapability.detail}`);
             }
+            steps.push({
+                id: "sdk-custom-tool",
+                label: "Pi SDK custom-tool 能力",
+                ok: true,
+            });
             const toolProbe = await createPlaywrightCliTool({
                 repoRoot: input.cwd,
                 runDir: meta.runDir,
                 nodeId: input.task.id,
                 caseId: "frontend-browser-preflight",
                 evidenceDir: "testcase/frontend/evidence/frontend-browser-preflight",
-                // Capability-only registration probe. The tested URL is selected later by
-                // retrieve-frontend-test-context-pi and validated by the environment shell.
                 baseUrl: "http://localhost:5173/",
                 inputRoot: "testcase/frontend/fixtures",
             });
             if (toolProbe.name !== "playwright_cli") {
+                steps.push({
+                    id: "structured-tool",
+                    label: "注册 playwright_cli",
+                    ok: false,
+                    detail: toolProbe.name,
+                });
                 throw new Error("browser-command-capability-unavailable: structured playwright_cli tool registration failed");
             }
+            steps.push({
+                id: "structured-tool",
+                label: "注册 playwright_cli",
+                ok: true,
+            });
             const launcher = resolvePlaywrightCliLauncher(input.cwd);
             const result = await defaultPlaywrightCliCommandRunner({
                 executable: launcher.executable,
@@ -2429,8 +2589,19 @@ export async function executeDagShellNode(input, meta) {
             if (result.exitCode !== 0 ||
                 result.timedOut ||
                 !["open", "close", "find", "snapshot", "click"].every((command) => new RegExp(`\\b${command}\\b`, "i").test(help))) {
+                steps.push({
+                    id: "playwright-cli-help",
+                    label: "playwright-cli --help 合同",
+                    ok: false,
+                    detail: help.slice(0, 8 * 1024),
+                });
                 throw new Error("playwright-cli-contract-incompatible");
             }
+            steps.push({
+                id: "playwright-cli-help",
+                label: "playwright-cli --help 合同",
+                ok: true,
+            });
             const capability = {
                 schemaVersion: 1,
                 status: "ready",
@@ -2441,22 +2612,28 @@ export async function executeDagShellNode(input, meta) {
                 targetUrlAuthority: "retrieve-context-and-environment-probe",
             };
             await writeDagNodeJsonArtifact(meta.runDir, input.task.id, "frontend-browser-capability.json", capability);
-            return {
+            steps.push({
+                id: "capability-artifact",
+                label: "写出 frontend-browser-capability.json",
+                ok: true,
+                detail: JSON.stringify(capability, null, 2),
+            });
+            return await finish({
                 ok: true,
                 stdout: JSON.stringify(capability),
                 stderr: "",
                 failureCategory: "success",
                 durationMs: Date.now() - started,
-            };
+            });
         }
         catch (error) {
-            return {
+            return await finish({
                 ok: false,
                 stdout: "",
                 stderr: `playwright-cli-unavailable: ${error instanceof Error ? error.message : String(error)}`,
                 failureCategory: "tool-policy",
                 durationMs: Date.now() - started,
-            };
+            });
         }
     }
     if (shell?.frontendPrewriteGate) {
@@ -2878,6 +3055,7 @@ export async function executeDagShellNode(input, meta) {
     const cwd = resolveShellCwd(input.cwd, shell.cwd);
     const timeoutMs = shell.timeoutMs ?? DEFAULT_SHELL_TIMEOUT_MS;
     const failFast = shell.failFast !== false;
+    const attemptNumber = input.attempt ?? 1;
     const results = [];
     const executedCommandKeys = new Set();
     let beforeStatus;
@@ -2973,8 +3151,8 @@ export async function executeDagShellNode(input, meta) {
             dagRunMeta: { runDir: meta.runDir, runId: meta.runId },
             reportActivity: input.reportActivity,
             outputArtifacts: {
-                stdoutPath: path.join(meta.runDir, input.task.id, "commands", `${commandNumber}.stdout.txt`),
-                stderrPath: path.join(meta.runDir, input.task.id, "commands", `${commandNumber}.stderr.txt`),
+                stdoutPath: path.join(meta.runDir, input.task.id, `attempt-${attemptNumber}`, "commands", `${commandNumber}.stdout.txt`),
+                stderrPath: path.join(meta.runDir, input.task.id, `attempt-${attemptNumber}`, "commands", `${commandNumber}.stderr.txt`),
             },
         });
         results.push(result);
@@ -3033,6 +3211,7 @@ export async function executeDagShellNode(input, meta) {
         schemaVersion: 1,
         runId: meta.runId,
         nodeId: input.task.id,
+        attempt: attemptNumber,
         workspaceFingerprint: workspaceFingerprint ?? null,
         failFast,
         commands: results.map((result) => ({
@@ -3051,6 +3230,7 @@ export async function executeDagShellNode(input, meta) {
             subchecks: parseCheckRepoSubcheckDurations(result),
         })),
     };
+    await writeDagNodeJsonArtifact(meta.runDir, input.task.id, path.posix.join(`attempt-${attemptNumber}`, "shell-command-receipt.json"), receipt);
     await writeDagNodeJsonArtifact(meta.runDir, input.task.id, "shell-command-receipt.json", receipt);
     const summaryMarkdown = buildShellResultSummaryMarkdown({
         nodeId: input.task.id,
@@ -3102,8 +3282,24 @@ export async function executeDagShellNode(input, meta) {
         stdout: `${recordPrefix}${reuseNote}${aggregate.stdout}`,
         stderr: stderrParts.filter(Boolean).join("\n\n"),
         failureCategory,
+        ...(firstFailure?.failureCategory
+            ? { rawFailureCategory: firstFailure.failureCategory }
+            : {}),
         durationMs: Date.now() - started,
         // extra field consumed by node-execution if supported; also embedded in stdout marker
         ...{ commandResults },
     };
 }
+export async function executeDagShellNode(input, meta) {
+    const handler = namedShellHandler(input.task.shell);
+    const result = await executeDagShellNodeInner(input, meta);
+    if (handler) {
+        await ensureShellPresetProcessReceipt({
+            runDir: meta.runDir,
+            nodeId: input.task.id,
+            preset: handler,
+            result,
+        });
+    }
+    return result;
+}

package/dist/shared/dag-prompt-override.js

@@ -0,0 +1,28 @@
+/**
+ * Internal switch for the experimental "rerun failed node with an edited
+ * prompt" capability. Default is OFF; nothing may depend on it being enabled.
+ * The switch is deliberately env-based (no harness.json surface) while the
+ * feature is internal.
+ */
+export const DAG_RERUN_PROMPT_OVERRIDE_ENV = "HARNESS_DAG_RERUN_PROMPT_OVERRIDE";
+/** Upper bound for an override input body (characters). */
+export const MAX_DAG_RERUN_PROMPT_OVERRIDE_CHARS = 65_536;
+export function isDagRerunPromptOverrideEnabled(env = process.env) {
+    const value = env[DAG_RERUN_PROMPT_OVERRIDE_ENV];
+    return value === "1" || value === "true";
+}
+export const DAG_RERUN_PROMPT_ADDENDUM_TAG = "operator_addendum";
+/**
+ * Compose the final subtask_prompt for a rerun override.
+ * - "replace": the operator text becomes the whole prompt (advanced mode).
+ * - "append" (recommended): the frozen base prompt is kept verbatim and the
+ *   operator addendum is appended in a clearly marked block, mirroring the
+ *   rerunFeedback / retry_instruction augmentation semantics — small diffs,
+ *   no risk of dropping template scaffolding.
+ */
+export function composeDagPromptOverridePrompt(input) {
+    if (input.mode === "replace") {
+        return input.text;
+    }
+    return `${input.basePrompt}\n\n<${DAG_RERUN_PROMPT_ADDENDUM_TAG}>\n${input.text.trim()}\n</${DAG_RERUN_PROMPT_ADDENDUM_TAG}>\n`;
+}

package/dist/shared/operator/capabilities.js

@@ -1155,6 +1155,19 @@ export function buildOperatorCapabilitiesDocument() {
             modelCallable: "always",
             humanConfirmation: "none",
         },
+        {
+            action: "dagPromptOverrideConfig",
+            cli: "console dag-prompt-override-config",
+            kind: "read",
+            inputSchemaVersion: 1,
+            resultSchemaVersion: 1,
+            envelopeSchemaVersion: 1,
+            requiredErrorCodes: [],
+            description: "Read-only probe of the experimental prompt-override rerun switch (env-gated, default off). The browser Recovery UI hides the failed-node prompt editor unless this returns enabled=true. Never mutates.",
+            inputParams: [],
+            modelCallable: "always",
+            humanConfirmation: "none",
+        },
         {
             action: "dagRerunPlan",
             cli: "loop-agent dag rerun --run-id <id> --from-node <node> --plan --json",
@@ -1177,6 +1190,12 @@ export function buildOperatorCapabilitiesDocument() {
                     required: true,
                     description: "node id to rerun from (prefer failed/attention node from dagReport primaryFailure.nodeId)",
                 },
+                {
+                    name: "promptOverrides",
+                    type: "array",
+                    required: false,
+                    description: "experimental (switch off by default): [{nodeId, prompt, mode?}] — mode append (recommended) keeps the frozen base prompt and appends the text as a marked addendum block, replace swaps the whole subtask_prompt; only FAILED nodes inside the reset closure are admitted and the planHash binds the composed change",
+                },
             ],
             modelCallable: "always",
             humanConfirmation: "none",
@@ -1219,6 +1238,12 @@ export function buildOperatorCapabilitiesDocument() {
                     required: true,
                     description: "why continue from this node (e.g. provider flake; resume from review-pi)",
                 },
+                {
+                    name: "promptOverrides",
+                    type: "array",
+                    required: false,
+                    description: "experimental (switch off by default): must match the dagRerunPlan promptOverrides exactly (same mode/text); the planHash already binds them, any drift is rejected",
+                },
             ],
             // always: fresh eligible dagRerun plans are model-callable (2026-08-11).
             // The CLI keeps enforcing parent lifecycle, fromNode, planHash,
@@ -1822,6 +1847,13 @@ export const OPERATOR_COMMAND_COVERAGE = Object.freeze([
         source: "loop-agent",
         exclusionReason: "Not exposed to Operator Chat in M5.",
     },
+    {
+        command: "loop-agent init model-catalog",
+        coverage: "excluded",
+        action: null,
+        source: "loop-agent",
+        exclusionReason: "Not exposed to Operator Chat in M5.",
+    },
     {
         command: "loop-agent init doctor",
         coverage: "excluded",

package/dist/shared/playwright-cli-command-policy.js

@@ -9,6 +9,7 @@ export const PLAYWRIGHT_CLI_ALLOWED_COMMANDS = [
     "goto",
     "snapshot",
     "find",
+    "find-unique",
     "click",
     "dblclick",
     "fill",
@@ -35,6 +36,20 @@ export const PLAYWRIGHT_CLI_ALLOWED_COMMANDS = [
     "requests",
     "request",
 ];
+/**
+ * Parse the match count from a playwright-cli find output. The CLI prints
+ * `Found N match(es) for "...":` on success and `No matches found for "...".`
+ * on zero — both with exit code 0, so the count is the only assertion signal.
+ * Returns undefined when the output shape is unrecognized (e.g. error pages).
+ */
+export function parsePlaywrightCliFindMatchCount(output) {
+    const found = /Found\s+(\d+)\s+match(?:es)?\s+for\b/i.exec(output);
+    if (found)
+        return Number(found[1]);
+    if (/No matches found for\b/i.test(output))
+        return 0;
+    return undefined;
+}
 const allowedCommandSet = new Set(PLAYWRIGHT_CLI_ALLOWED_COMMANDS);
 export function isPlaywrightCliCommand(value) {
     return allowedCommandSet.has(value);