@tea-agent/loop-agent 0.39.0-next.8 → 0.39.0

package/dist/executors/shell-write-guard.js

@@ -98,6 +98,13 @@ export function isEphemeralToolCachePath(filePath) {
     const normalized = normalizePath(filePath);
     if (!normalized)
         return false;
+    // CodeGraph extension sync side effect (plan 2026-08-21 D7): a loaded
+    // pi-codegraph extension keeps an existing .codegraph/ index fresh at
+    // session start; those index files are extension-owned run state, not
+    // writer output, so they never count as write-guard violations.
+    if (normalized === ".codegraph" || normalized.startsWith(".codegraph/")) {
+        return true;
+    }
     if (normalized === ".pytest_cache" ||
         normalized.startsWith(".pytest_cache/")) {
         return true;

package/dist/shared/dag-failure-category.js

@@ -0,0 +1,138 @@
+/**
+ * Pure DAG failure-category classification shared by the workflows kernel and
+ * the Operator Console projection (console sources may not import
+ * src/workflows — security-and-tools boundary test).
+ *
+ * No runtime imports: this module is a leaf safe for both the CLI kernel and
+ * the browser bundle.
+ */
+export const DAG_NORMALIZED_FAILURE_CATEGORY_VALUES = [
+    "success",
+    "validation",
+    "executor",
+    "write-guard",
+    "timeout",
+    "auth",
+    "human-required",
+    "human-rejected",
+    "shell-command",
+    "static-error",
+    "decision-envelope",
+    "skipped",
+    "unknown",
+];
+const RAW_TO_NORMALIZED = {
+    success: "success",
+    "write-guard": "write-guard",
+    timeout: "timeout",
+    // Clean implementer stall with zero writes — still a timeout class for report,
+    // but retryable under WRITER_TRANSPORT_RETRY_POLICY.
+    "writer-clean-timeout": "timeout",
+    "missing-api-key": "auth",
+    auth: "auth",
+    "human-rejected": "human-rejected",
+    "human-required": "human-required",
+    "decision-envelope-invalid": "decision-envelope",
+    "nonzero-exit": "shell-command",
+    "spawn-error": "shell-command",
+    "test-timeout": "timeout",
+    "static-error": "static-error",
+    "invalid-output": "validation",
+    "structured-output-truncated": "validation",
+    "output-too-large": "validation",
+    "structured-repair-exhausted": "validation",
+    "governance-blocked": "validation",
+    "path-violation": "validation",
+    "empty-output": "validation",
+    "verify-failure": "validation",
+    "test-failure": "validation",
+    "verification-plan-stale": "validation",
+    network: "executor",
+    quota: "executor",
+    "rate-limit": "executor",
+    unavailable: "executor",
+    // Deterministic capability/policy refusal: missing verified launcher JS
+    // entry, capability allowlist rejection, or missing bounded write tool
+    // profile. These are environment/config defects the operator repairs
+    // before rerunning; classify as executor so recovery recommends
+    // rerun-after-fix WITHOUT forcing a human-decision gate on
+    // standaloneTaskRerun (run-failed same-task rerun).
+    "tool-policy": "executor",
+    unknown: "unknown",
+};
+/** Status inputs are deliberately `string`: projection layers pass bounded
+ *  text from CLI JSON; unmatched values simply fall through to unknown. */
+function isFinishedStatus(status) {
+    return status === "FINISHED" || status === "finished";
+}
+function isInProgressStatus(status) {
+    return (status === "PENDING" ||
+        status === "RUNNING" ||
+        status === "pending" ||
+        status === "running");
+}
+function isFailureStatus(status) {
+    return (status === "ERROR" ||
+        status === "failed" ||
+        status === "partial_failed");
+}
+function mapRawCategory(raw) {
+    const direct = RAW_TO_NORMALIZED[raw];
+    if (direct)
+        return direct;
+    const lower = raw.toLowerCase();
+    if (lower.includes("timeout"))
+        return "timeout";
+    if (lower.includes("validation") ||
+        lower.includes("invalid") ||
+        lower.includes("verification-plan-stale") ||
+        lower.includes("plan-stale"))
+        return "validation";
+    if (lower.includes("auth") || lower.includes("api-key"))
+        return "auth";
+    if (lower.includes("write-guard") || lower.includes("write_guard"))
+        return "write-guard";
+    if (lower.includes("decision-envelope") || lower.includes("decision_envelope"))
+        return "decision-envelope";
+    if (lower.includes("human-rejected"))
+        return "human-rejected";
+    if (lower.includes("human-required"))
+        return "human-required";
+    if (lower.includes("static-error"))
+        return "static-error";
+    if (lower.includes("nonzero-exit") ||
+        lower.includes("spawn-error") ||
+        lower.includes("shell"))
+        return "shell-command";
+    return undefined;
+}
+export function isDagFailureCategory(value) {
+    return DAG_NORMALIZED_FAILURE_CATEGORY_VALUES.includes(value);
+}
+export function normalizeDagFailureCategory(raw, status) {
+    if (status === "SKIPPED") {
+        return "skipped";
+    }
+    if (raw !== undefined) {
+        const mapped = mapRawCategory(raw);
+        if (mapped) {
+            if (mapped === "success") {
+                return isFinishedStatus(status) ? "success" : "unknown";
+            }
+            return mapped;
+        }
+    }
+    if (raw === undefined && isFinishedStatus(status)) {
+        return "success";
+    }
+    if (isInProgressStatus(status)) {
+        return "unknown";
+    }
+    if (isFailureStatus(status)) {
+        return "unknown";
+    }
+    if (raw === "success" && isFinishedStatus(status)) {
+        return "success";
+    }
+    return "unknown";
+}

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

@@ -13,13 +13,45 @@ const CONTROLLER_PROTOCOL_CAPABILITIES = [
     "worker-attempt-revision-v1",
     "recovery-decision-v2",
 ];
-function officialModelCallable(coverage, kind) {
+function officialModelCallable(coverage, _kind) {
     if (coverage === "human-gated-required")
         return "prepare-only";
-    if (coverage === "advanced" && kind !== "read")
+    // 2026-08-18: advanced actions (read or mutation) stay OUT of the model
+    // tool surface — low-frequency/forensic commands with generic args[] params
+    // diluted model attention. They remain dispatchable via UI capabilities.
+    if (coverage === "advanced")
         return "never";
     return "always";
 }
+const OFFICIAL_ACTION_DESCRIPTIONS = {
+    dagDecisionInspect: "Inspect a pending DAG decision envelope (inputs, options, chosen branch). Use before advising the operator on approve/reject; read-only.",
+    dagDecisionValidate: "Validate a DAG decision envelope against its schema. Use when a decision payload looks malformed or a writer emitted a suspect envelope; read-only.",
+    dagCloseoutDraft: "Draft the closeout handoff document for a finished DAG run. Use when the run is FINISHED and you need the handoff text; read-only.",
+    planList: "List execution plans in the repository. Use to discover plan ids before plan create/complete/check.",
+    planCreate: "Create an execution plan from a summary. Use to record the next work block before starting it.",
+    planComplete: "Mark an execution plan complete with a summary of what was delivered. Use after a work block is done and verified.",
+    planCheck: "Check execution-plan health and drift. Use to audit plan state without mutating anything.",
+    docsAudit: "Audit docs freshness and links (document catalog policy). Use when asked whether docs are stale or need review.",
+    handoffCheck: "Check handoff document completeness against the verification matrix. Use before claiming a task is fully handed off.",
+    handoffCoverage: "Show verification coverage across tests/docs/links for the repo. Use when asked what is verified and what is not.",
+    coverageReport: "Produce the verification coverage report. Use for a structured view of tests vs requirements.",
+    stats: "Repository workflow statistics (tasks, runs, durations). Use for quantitative status questions.",
+    promoteRun: "Task advance alias; passes a task id and flags through to task advance to promote a run. Does not pre-verify run state here — confirm with dagReport that the run is FINISHED before calling.",
+    closeoutTask: "Task advance alias; passes a task id and flags through to task advance to close out a task. Confirm the task's work is delivered and verified before calling.",
+    knowledgeCurate: "Generate knowledge-base curation proposals (may write a proposal file when --output is passed). Write-capable; use via the operator CLI, not as a model tool.",
+    knowledgeQuery: "Query the repository knowledge base. Use instead of re-reading files when a curated answer may exist.",
+    workerFeatureReview: "Review worker feature status and checkpoints (agent-worker). Use to see where a feature line stands.",
+    workerFeatureDoctor: "Diagnose worker feature health (agent-worker). Use when a feature line looks stuck or inconsistent.",
+    workerTaskValidate: "Validate a worker task spec (agent-worker). Use before queuing or when a task looks malformed.",
+    workerBatchPlanReady: "Show which batch tasks are ready to run (agent-worker). Use for queue triage before workerBatchRunReady.",
+    workerReportMorning: "Morning report for overnight worker activity (agent-worker). Use at the start of a session to catch overnight failures.",
+    workerReportMetrics: "Worker throughput and failure metrics (agent-worker). Use for quantitative worker questions.",
+    workerAdmissionShow: "Show night admission queue state (agent-worker). Use before night-scheduler mutations.",
+    workerSchedulerList: "List night-scheduler schedules (agent-worker). Use to discover schedule ids.",
+    workerSchedulerStatus: "Night-scheduler status overview (agent-worker). Use to check whether overnight jobs are healthy.",
+    workerSchedulerLedger: "Night-scheduler harvest ledger (agent-worker). Use to audit what overnight runs produced.",
+    workerSchedulerDoctor: "Diagnose night-scheduler health (agent-worker). Use when overnight jobs fail to fire or harvest.",
+};
 const officialAction = (action, cli, kind, coverage) => ({
     action,
     cli,
@@ -38,7 +70,8 @@ const officialAction = (action, cli, kind, coverage) => ({
             "TRANSACTION_INCOMPLETE",
             "NOT_FOUND",
         ],
-    description: `Official CLI action coverage for ${cli}.`,
+    description: OFFICIAL_ACTION_DESCRIPTIONS[action] ??
+        `Official CLI action coverage for ${cli}.`,
     inputParams: [
         {
             name: "args",
@@ -63,8 +96,6 @@ const officialAction = (action, cli, kind, coverage) => ({
 });
 const OFFICIAL_ACTIONS = [
     ...[
-        ["dagDoctor", "loop-agent dag doctor", "read", "model-callable"],
-        ["dagStatus", "loop-agent dag status", "read", "model-callable"],
         [
             "dagDecisionInspect",
             "loop-agent dag decision inspect",
@@ -160,8 +191,8 @@ const OFFICIAL_ACTIONS = [
         [
             "knowledgeCurate",
             "loop-agent knowledge curate",
-            "read",
-            "model-callable",
+            "mutation",
+            "advanced",
         ],
         ["knowledgeQuery", "loop-agent knowledge query", "read", "model-callable"],
         [
@@ -284,7 +315,6 @@ const OFFICIAL_ACTIONS = [
             "long-running",
             "advanced",
         ],
-        ["workerPoolDoctor", "agent-worker pool doctor", "read", "model-callable"],
         [
             "workerPoolMarkFailed",
             "agent-worker pool mark-failed",
@@ -337,6 +367,12 @@ const OFFICIAL_ACTIONS = [
         ],
     ],
 ].map(([action, cli, kind, coverage]) => officialAction(action, cli, kind, coverage));
+/**
+ * Action ids produced by the officialAction() factory (the generic CLI
+ * coverage tail). Derived from the same tuple table so dispatcher / gate
+ * checks do not string-match descriptions.
+ */
+export const OFFICIAL_TAIL_ACTION_IDS = new Set(OFFICIAL_ACTIONS.map((entry) => entry.action));
 /** Shared constants so dag + dagSpecVersions cannot diverge. */
 const DAG_SPEC_SUPPORTED_PARSE = [1, 2, 3, 4];
 const DAG_SPEC_NEW_WRITER_DEFAULT = 4;
@@ -612,7 +648,7 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: [...COMMON_MUTATION_ERRORS],
-            description: "Create or advance a task via the single lifecycle mutation entry.",
+            description: "Create a new task (lifecycle entry). Prefer taskAdvance for the standard flow (PRD import, taskKind, boundaries, DAG generation); newTask only bootstraps a task id + title.",
             inputParams: [
                 {
                     name: "taskId",
@@ -688,7 +724,7 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: ["NOT_FOUND", "INVALID_INPUT"],
-            description: "Read source instruction contract for a task.",
+            description: "Read the task's source instruction contract (需求.md/执行约束.md projection: requirements, constraints, allowed/forbidden paths). Use before drafting or revising a contract to stay aligned with the source.",
             inputParams: [
                 {
                     name: "taskId",
@@ -838,8 +874,33 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: ["INVALID_INPUT"],
-            description: "Build Observe deep link from stable object identity.",
-            inputParams: [],
+            description: "Build Observe deep link from stable object identity. Pass exactly one of dagRunId (preferred), taskId (resolves to latest run), featureId, or workerRunId; omit all for the Console home link.",
+            inputParams: [
+                {
+                    name: "dagRunId",
+                    type: "string",
+                    required: false,
+                    description: "dag run id; deep-links to #/dag/<runId> (preferred when known)",
+                },
+                {
+                    name: "taskId",
+                    type: "string",
+                    required: false,
+                    description: "task id; resolves to the latest dag run when one exists, else #/task/<taskId>",
+                },
+                {
+                    name: "featureId",
+                    type: "string",
+                    required: false,
+                    description: "feature line id; deep-links to the feature view",
+                },
+                {
+                    name: "workerRunId",
+                    type: "string",
+                    required: false,
+                    description: "worker run id; deep-links to the worker run view",
+                },
+            ],
             modelCallable: "always",
             humanConfirmation: "none",
         },
@@ -851,7 +912,7 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: ["INVALID_INPUT"],
-            description: "Repo preflight (registry entry; envelope migration follow-up).",
+            description: "Preflight a repository for loop-agent: checks harness presence, task-pool layout, and reports what the Console can operate on. Use at session start on an unfamiliar repo, before init/upgrade decisions.",
             inputParams: [],
             modelCallable: "always",
             humanConfirmation: "none",
@@ -910,7 +971,7 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: ["NOT_FOUND", "INVALID_INPUT"],
-            description: "Task lifecycle status (read-only).",
+            description: "Read one task's lifecycle status: state, blockers, gates, contract revision, run associations. Use for task-level questions; use dagStatus for a specific DAG run's live node/liveness.",
             inputParams: [
                 {
                     name: "taskId",
@@ -930,7 +991,7 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: ["NOT_FOUND", "INVALID_INPUT"],
-            description: "Spine audit (registry entry).",
+            description: "Audit a task's minimal spec spine: source docs, path ownership, requirement ids, final verification coverage. Use before closing out a task to confirm every AC is verified.",
             inputParams: [
                 {
                     name: "taskId",
@@ -1023,7 +1084,7 @@ export function buildOperatorCapabilitiesDocument() {
                 "EXTERNALLY_MODIFIED",
                 "INVALID_INPUT",
             ],
-            description: "Generate + strict-validate DAG via task advance (stops at writeSet gate).",
+            description: "Generate + strict-validate a task's DAG (via task advance, stops at the writeSet gate). Legacy entry; the standard flow is taskAdvance then prepareDagExecution + runDag.",
             inputParams: [
                 {
                     name: "taskId",
@@ -1079,7 +1140,7 @@ export function buildOperatorCapabilitiesDocument() {
                 "CONTROLLER_MISMATCH",
                 "INVALID_INPUT",
             ],
-            description: "Consume a single-use execution receipt and execute the reviewed DAG (prefer task advance --approve-gate when gate token present). accepted/queued/running/operationId are NOT completion — supervise via operationGet/status/dagReport/dagDoctor/operationWait.",
+            description: "Consume a single-use executionId from prepareDagExecution and execute the reviewed DAG. Do not pass a DAG path and do not bash `task advance --approve-gate`. accepted/queued/running/operationId are NOT completion — supervise via operationGet/status/dagReport/dagDoctor/operationWait.",
             inputParams: [
                 {
                     name: "executionId",
@@ -1094,6 +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",
@@ -1102,7 +1176,7 @@ export function buildOperatorCapabilitiesDocument() {
             resultSchemaVersion: 1,
             envelopeSchemaVersion: 1,
             requiredErrorCodes: ["INVALID_INPUT", "NOT_FOUND"],
-            description: "Read-only R1 subgraph plan (no mutation). Call before dagRerun. Prefer primaryFailure.nodeId from dagReport. If eligible=false (writer/decision/fingerprint, unsafe shell, worker-managed, workspace/controller drift), do not force execute — Worker-owned runs use workerTaskRetry; standalone runs use standaloneTaskRerun or same-task advance after fixing the root cause (contract changes must be adopted first). Returns planHash required by execute.",
+            description: "Read-only check for failed-node recovery; it does not execute. Call after dagReport with primaryFailure.nodeId. eligible=true returns the exact planHash required by dagRerun. eligible=false covers writer/decision/fingerprint, unsafe shell, Worker ownership, or workspace/controller drift: use workerTaskRetry for Worker/Task Pool failures, or standaloneTaskRerun after any required contract adopt for standalone runs. Do not force dagRerun or invent a new task.",
             inputParams: [
                 {
                     name: "runId",
@@ -1116,6 +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",
@@ -1132,7 +1212,7 @@ export function buildOperatorCapabilitiesDocument() {
                 "BINDING_DRIFT",
                 "INVALID_INPUT",
             ],
-            description: "Execute R1 continuation from effective node. Requires planHash from a fresh dagRerunPlan (eligible=true). Prefer this over new task / standaloneTaskRerun for provider flake and safe downstream failures. Server enforces planHash/binding/fingerprint freshness.",
+            description: "Continue the same DAG from the effective failed node by creating a continuation run. Requires a fresh eligible dagRerunPlan and its exact planHash. Prefer this for provider, network, rate-limit, timeout, and other safe downstream failures. Not a full task regenerate: use standaloneTaskRerun for that. Server rejects stale planHash, binding, or fingerprint drift.",
             inputParams: [
                 {
                     name: "runId",
@@ -1158,6 +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,
@@ -1498,7 +1584,7 @@ export function buildOperatorCapabilitiesDocument() {
                 "INVALID_INPUT",
                 "HUMAN_CONFIRMATION_REQUIRED",
             ],
-            description: "Full standalone task regenerate → validate → execute with parent lineage. Only when node rerun plan is ineligible or primaryRecovery recommends rerun-task; the server checks read-only run facts before accepting. Regeneration reuses the CURRENT contract snapshot: if the failure requires contract changes (e.g. allowing a catalog index path), first update the contract (re-import PRD, or adopt externally-edited sources via `task contract adopt`) BEFORE calling this action, otherwise the regenerated DAG repeats the same failure.",
+            description: "Regenerate the whole standalone DAG from the CURRENT contract, then validate and execute a new run with parent lineage and the latest unresolved verify/review feedback when available. This is NOT continue-from-failed-node (use dagRerun for that). Server rejects the call when dagRerunPlan is eligible, primaryRecovery.humanRequired=true, or run facts do not recommend rerun-task. Feedback is recovery context, not a contract rewrite; adopt required contract changes before calling or the new DAG repeats the same failure.",
             inputParams: [
                 {
                     name: "runId",
@@ -1542,7 +1628,7 @@ export function buildOperatorCapabilitiesDocument() {
                 "INVALID_INPUT",
                 "HUMAN_CONFIRMATION_REQUIRED",
             ],
-            description: "Requeue failed Task Pool task (Failed → Ready) via in-package pool store. Only when read-only validation confirms the task exists in Failed state; the server checks this before accepting.",
+            description: "Requeue a Worker Task Pool task Failed → Ready. Not a DAG node rerun (use dagRerun) and not a standalone DAG regenerate (use standaloneTaskRerun). Confirm with workerPoolDoctor first; server accepts only Failed pool tasks. Pass featureId when known — it must match the pool record.",
             inputParams: [
                 {
                     name: "taskId",
@@ -1568,6 +1654,66 @@ export function buildOperatorCapabilitiesDocument() {
             modelCallable: "always",
             humanConfirmation: "none",
         },
+        {
+            action: "dagDoctor",
+            cli: "loop-agent dag doctor [--run-id <id>] --json",
+            kind: "read",
+            inputSchemaVersion: 1,
+            resultSchemaVersion: 1,
+            envelopeSchemaVersion: 1,
+            requiredErrorCodes: ["NOT_FOUND", "INVALID_INPUT"],
+            description: "Diagnose DAG run health. Omit runId for a repo-wide inventory (active/paused/completed + health issues). Pass runId for one run's liveness, heartbeat, failure node, and recovery hints. Use this to supervise stalls; it does not mutate the run.",
+            inputParams: [
+                {
+                    name: "runId",
+                    type: "string",
+                    required: false,
+                    description: "optional dag run id; omit to inventory all runs, pass to diagnose one run",
+                },
+            ],
+            modelCallable: "always",
+            humanConfirmation: "none",
+        },
+        {
+            action: "dagStatus",
+            cli: "loop-agent dag status --run-id <id>",
+            kind: "read",
+            inputSchemaVersion: 1,
+            resultSchemaVersion: 1,
+            envelopeSchemaVersion: 1,
+            requiredErrorCodes: ["NOT_FOUND", "INVALID_INPUT"],
+            description: "Read one DAG run's live status, current node, and liveness. Requires runId. Complements dagReport (terminal handoff) and dagDoctor (health diagnosis). Does not mutate the run.",
+            inputParams: [
+                {
+                    name: "runId",
+                    type: "string",
+                    required: true,
+                    description: "dag run id",
+                },
+            ],
+            modelCallable: "always",
+            humanConfirmation: "none",
+        },
+        {
+            action: "workerPoolDoctor",
+            cli: "agent-worker pool doctor --repo . --json",
+            kind: "read",
+            inputSchemaVersion: 1,
+            resultSchemaVersion: 1,
+            envelopeSchemaVersion: 1,
+            requiredErrorCodes: ["INVALID_INPUT", "NOT_FOUND"],
+            description: "Read-only Task Pool inventory (legacy + v2). Use before workerTaskRetry to confirm the task exists and status is Failed. Optional taskId filters the returned inventory to that task.",
+            inputParams: [
+                {
+                    name: "taskId",
+                    type: "string",
+                    required: false,
+                    description: "optional task id to focus the inventory",
+                },
+            ],
+            modelCallable: "always",
+            humanConfirmation: "none",
+        },
         ...OFFICIAL_ACTIONS,
     ];
     return {
@@ -1701,6 +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",
@@ -1947,12 +2100,12 @@ export const OPERATOR_COMMAND_COVERAGE = Object.freeze([
     },
     {
         command: "loop-agent knowledge curate",
-        coverage: "model-callable",
+        coverage: "advanced",
         action: "knowledgeCurate",
         source: "loop-agent",
+        note: "Writes knowledge proposals (conditional on --output); not model-callable.",
     },
-    {
-        command: "loop-agent knowledge query",
+    { command: "loop-agent knowledge query",
         coverage: "model-callable",
         action: "knowledgeQuery",
         source: "loop-agent",
@@ -2527,9 +2680,10 @@ export function assertCoverageManifestConsistent() {
                 capability.humanConfirmation !== "required"))
             throw new Error(`invalid required gate classification: ${entry.action}`);
         if (entry.coverage === "advanced") {
-            const expected = capability.kind === "read" ? "always" : "never";
-            if (capability.modelCallable !== expected) {
-                throw new Error(`invalid advanced classification: ${entry.action} (expected modelCallable=${expected} for kind=${capability.kind})`);
+            // 2026-08-18: advanced actions are never model-callable (read or
+            // mutation) — see officialModelCallable. UI dispatch stays available.
+            if (capability.modelCallable !== "never") {
+                throw new Error(`invalid advanced classification: ${entry.action} (expected modelCallable=never)`);
             }
         }
     }

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);