@themoltnet/agent-daemon 0.7.0 → 0.9.0

package/README.md

@@ -57,7 +57,7 @@ export ANTHROPIC_API_KEY=sk-ant-...
 # https://github.com/badlogic/pi-mono/blob/main/packages/ai/src/env-api-keys.ts
 ```
 
-To use an alternate auth-file path: `PI_AUTH_PATH=/abs/path/to/auth.json`.
+To use a non-default auth directory: `PI_CODING_AGENT_DIR=/abs/path/to/.pi/agent`.
 
 ### Observability
 
@@ -254,6 +254,41 @@ moltnet entry list --diary-id "$MOLTNET_DIARY_ID" --limit 10 \
   --credentials "$PWD/.moltnet/local-dev/moltnet.json"
 ```
 
+### 4b. Create a `pr_review` smoke task
+
+Use this path when you want to exercise the generic `pr_review` task type
+against the local e2e stack before the new schema exists on a deployed API.
+
+Start the daemon with `--task-types pr_review`:
+
+```bash
+pnpm --filter @themoltnet/agent-daemon dev poll \
+  --agent local-dev \
+  --team "$MOLTNET_TEAM_ID" \
+  --task-types pr_review \
+  --provider openai-codex \
+  --model gpt-5.4-codex \
+  --debug
+```
+
+Then, in another terminal, create the task:
+
+```bash
+pnpm exec tsx tools/src/tasks/create-pr-review.ts \
+  --agent local-dev \
+  --pr <number> \
+  --repo <owner/repo>
+```
+
+This helper stays imposer-only. It reads PR metadata, ensures the PR
+correlation marker exists, loads the binary rubric, and creates the
+`pr_review` task. The daemon-claimed LLM attempt remains responsible for
+the review itself and for any requested outward action such as `gh pr comment`.
+
+If you want an automated local check without real GitHub mutation, use the
+stubbed `pr_review` lifecycle coverage in
+`apps/agent-daemon/e2e/daemon.e2e.test.ts` instead of this manual smoke path.
+
 ### What to verify
 
 After the task completes, every entry produced **during the attempt** should:

package/dist/main.js

@@ -3045,10 +3045,10 @@ function validateRubricWeights(rubric) {
 * complementary places.
 *
 * Before this envelope existed, criteria were scattered: a vestigial
-* `criteriaCid` column nobody resolved, an `acceptanceCriteria: string[]`
-* field on `fulfill_brief.input` that was "interpreted by the claiming
-* agent," and inline `rubric` / `criteria[]` fields on judgment-task
-* inputs. None of those were machine-verifiable end-to-end.
+* `criteriaCid` column nobody resolved, free-form prose on
+* `fulfill_brief.input`, and inline `rubric` / `criteria[]` fields on
+* judgment-task inputs. None of those were machine-verifiable
+* end-to-end.
 *
 * This module defines a single, content-addressable envelope an imposer
 * attaches to any task type. It has four orthogonal sections — pick
@@ -4185,7 +4185,6 @@ var FULFILL_BRIEF_TYPE = "fulfill_brief";
 var FulfillBriefInput = Type$2.Object({
 	brief: Type$2.String({ minLength: 1 }),
 	title: Type$2.Optional(Type$2.String()),
-	acceptanceCriteria: Type$2.Optional(Type$2.Array(Type$2.String())),
 	successCriteria: Type$2.Optional(SuccessCriteria),
 	seedFiles: Type$2.Optional(Type$2.Array(Type$2.String())),
 	scopeHint: Type$2.Optional(Type$2.String())
@@ -4609,6 +4608,72 @@ async function onCreateJudgeEvalVariant(input, ctx) {
 	}];
 }
 //#endregion
+//#region ../../libs/tasks/src/task-types/pr-review.ts
+var PR_REVIEW_TYPE = "pr_review";
+var PrReviewSubject = Type$2.Object({
+	title: Type$2.String({ minLength: 1 }),
+	summary: Type$2.String({ minLength: 1 }),
+	resourceUrls: Type$2.Optional(Type$2.Array(Type$2.String({ minLength: 1 }))),
+	inspectionHints: Type$2.Optional(Type$2.Array(Type$2.String({ minLength: 1 })))
+}, {
+	$id: "PrReviewSubject",
+	additionalProperties: false
+});
+var PrReviewInput = Type$2.Object({
+	subject: PrReviewSubject,
+	taskPrompt: Type$2.Optional(Type$2.String({ minLength: 1 })),
+	successCriteria: SuccessCriteria
+}, {
+	$id: "PrReviewInput",
+	additionalProperties: false
+});
+var PrReviewScore = Type$2.Object({
+	criterionId: Type$2.String({ minLength: 1 }),
+	score: Type$2.Union([Type$2.Literal(0), Type$2.Literal(1)]),
+	rationale: Type$2.String({ minLength: 1 })
+}, {
+	$id: "PrReviewScore",
+	additionalProperties: false
+});
+var PrReviewOutput = Type$2.Object({
+	scores: Type$2.Array(PrReviewScore, { minItems: 1 }),
+	composite: Type$2.Number({
+		minimum: 0,
+		maximum: 1
+	}),
+	verdict: Type$2.String({ minLength: 1 })
+}, {
+	$id: "PrReviewOutput",
+	additionalProperties: false
+});
+function requireBooleanRubric(rubric) {
+	for (const criterion of rubric.criteria) if (criterion.scoring !== "boolean") return `pr_review requires boolean scoring for every rubric criterion; criterion "${criterion.id}" uses "${criterion.scoring}"`;
+	return null;
+}
+function validatePrReviewInput(input) {
+	const sc = input.successCriteria;
+	if (!sc) return "successCriteria is required for judgment tasks";
+	if (!sc.rubric) return "successCriteria.rubric is required for judgment tasks";
+	return validateRubricWeights(sc.rubric) ?? requireBooleanRubric(sc.rubric);
+}
+function validatePrReviewOutput(output, input) {
+	if (!input) return null;
+	const scores = output.scores;
+	const rubric = input.successCriteria.rubric;
+	if (!rubric) return null;
+	if (scores.length !== rubric.criteria.length) return `scores length ${scores.length} does not match rubric criteria length ${rubric.criteria.length}`;
+	let composite = 0;
+	for (let i = 0; i < rubric.criteria.length; i++) {
+		const criterion = rubric.criteria[i];
+		const score = scores[i];
+		if (score.criterionId !== criterion.id) return `scores[${i}] has criterionId "${score.criterionId}" but rubric expects "${criterion.id}" in that position`;
+		composite += criterion.weight * score.score;
+	}
+	const claimed = output.composite;
+	if (Math.abs(claimed - composite) > 1e-6) return `composite ${claimed} does not match weighted sum ${composite.toFixed(6)}`;
+	return null;
+}
+//#endregion
 //#region ../../libs/tasks/src/task-types/render-pack.ts
 /**
 * `render_pack` — turn a context pack into a signed rendered artefact.
@@ -4776,6 +4841,18 @@ var BUILT_IN_TASK_TYPES = {
 		validateInput: validateJudgmentInput,
 		validateInputAsync: validateAssessBriefInputAsync
 	},
+	[PR_REVIEW_TYPE]: {
+		name: PR_REVIEW_TYPE,
+		inputSchema: PrReviewInput,
+		outputSchema: PrReviewOutput,
+		outputKind: "judgment",
+		workspaceMode: "dedicated_worktree",
+		workspaceScope: "attempt",
+		sessionScope: "none",
+		requiresReferences: false,
+		validateInput: validatePrReviewInput,
+		validateOutput: validatePrReviewOutput
+	},
 	[CURATE_PACK_TYPE]: {
 		name: CURATE_PACK_TYPE,
 		inputSchema: CuratePackInput,
@@ -6169,78 +6246,6 @@ function isHelpFlag(args) {
 	return args.includes("--help") || args.includes("-h");
 }
 //#endregion
-//#region ../../libs/agent-runtime/src/subagent-output-contracts.ts
-var REGISTRY = /* @__PURE__ */ new Map();
-/**
-* Register a subagent output contract. Idempotent: re-registering the
-* same name with a different schema throws — contracts are meant to
-* be stable. Re-registering with the identical contract object (same
-* reference) is a no-op for HMR and test convenience.
-*
-* Typically called at module-init time alongside task-type
-* registration. See task-types/index.ts in @moltnet/tasks for the
-* conventional pattern.
-*/
-function registerSubagentOutputContract(contract) {
-	if (!contract.name || contract.name.trim().length === 0) throw new Error("subagent output contract name is required");
-	if (!/^[a-z][a-z0-9_]*$/.test(contract.name)) throw new Error(`subagent output contract name '${contract.name}' must be lower_snake_case (starts with a letter, then [a-z0-9_]+)`);
-	const existing = REGISTRY.get(contract.name);
-	if (existing && existing !== contract) {
-		if (existing.parametersSchema !== contract.parametersSchema) throw new Error(`subagent output contract '${contract.name}' is already registered with a different schema; refusing to override`);
-	}
-	REGISTRY.set(contract.name, contract);
-}
-/**
-* Resolve a subagent output contract by name. Returns `null` for
-* unknown names — callers (the subagent custom tool) decide whether
-* that's a tool error the parent LLM can recover from or a hard fail.
-*/
-function getSubagentOutputContract(name) {
-	return REGISTRY.get(name) ?? null;
-}
-/**
-* List all registered contracts. Useful for diagnostics and for the
-* subagent tool's parameter description so a parent LLM can see what
-* contracts are available without enumerating them in its prompt.
-*/
-function listSubagentOutputContracts() {
-	return [...REGISTRY.values()];
-}
-//#endregion
-//#region ../../libs/agent-runtime/src/built-in-contract-registrations.ts
-/**
-* Built-in subagent output contracts (#1087, #943).
-*
-* Why this is an exported function and not a module-init side
-* effect:
-*
-*   - The registry is process-global. Module-init registration
-*     fires exactly once per Node process (ESM modules are cached
-*     by URL). Tests that call `__resetSubagentOutputContractsForTests()`
-*     to start from an empty registry have no way to repopulate
-*     the built-ins without re-evaluating the module — which the
-*     cache prevents. PR #1101 review M4.
-*   - An explicit `registerBuiltInSubagentContracts()` lets the
-*     package index call it once at module load AND lets test
-*     setup hooks call it again after `__reset...`.
-*   - `registerSubagentOutputContract` is itself idempotent for
-*     identical re-registrations, so calling this function twice
-*     in the same process is safe.
-*
-* Adding a new built-in: extend the body of this function. Do not
-* call `registerSubagentOutputContract` from anywhere else in the
-* package — keeping all built-ins in one function makes the set
-* auditable.
-*/
-function registerBuiltInSubagentContracts() {
-	registerSubagentOutputContract({
-		name: "judge_eval_variant_result",
-		description: "Per-variant grading result produced by a subagent of judge_eval_variant: scores against the shared rubric, composite, and a 1-3 sentence verdict for a single variant.",
-		parametersSchema: JudgeEvalVariantResult
-	});
-}
-registerBuiltInSubagentContracts();
-//#endregion
 //#region ../../libs/agent-runtime/src/context-bindings.ts
 var PROMPT_SEPARATOR = "\n\n---\n\n";
 /**
@@ -6388,6 +6393,20 @@ function buildFinalOutputBlock(opts) {
 	return lines.join("\n");
 }
 //#endregion
+//#region ../../libs/agent-runtime/src/prompts/rubric-common.ts
+function renderRubricCriteriaList(rubric) {
+	return rubric.criteria.map((c, i) => `${i + 1}. **${c.id}** (weight ${c.weight}, scoring: \`${c.scoring}\`) — ${c.description}`).join("\n");
+}
+function renderRubricPreambleSection(rubric) {
+	if (!rubric.preamble) return null;
+	return [
+		"### Rubric preamble",
+		"",
+		rubric.preamble,
+		""
+	].join("\n");
+}
+//#endregion
 //#region ../../libs/agent-runtime/src/prompts/assess-brief.ts
 /**
 * Build the first user-message prompt for an `assess_brief` judge attempt.
@@ -6413,13 +6432,8 @@ function buildFinalOutputBlock(opts) {
 */
 function buildAssessBriefUserPrompt(input, ctx) {
 	const rubric = input.successCriteria.rubric;
-	const criteriaList = rubric.criteria.map((c, i) => `${i + 1}. **${c.id}** (weight ${c.weight}, scoring: \`${c.scoring}\`) — ${c.description}`).join("\n");
-	const preambleSection = rubric.preamble ? [
-		"### Rubric preamble",
-		"",
-		rubric.preamble,
-		""
-	].join("\n") : "";
+	const criteriaList = renderRubricCriteriaList(rubric);
+	const preambleSection = renderRubricPreambleSection(rubric) ?? "";
 	const workspaceSection = ctx.workspace?.mode === "dedicated_worktree" ? [
 		"### Workspace",
 		"",
@@ -6700,13 +6714,7 @@ function buildCuratePackUserPrompt(input, ctx) {
 * is told to inspect them itself.
 */
 function buildFulfillBriefUserPrompt(input, ctx) {
-	const { brief, title, acceptanceCriteria, seedFiles, scopeHint } = input;
-	const criteriaSection = acceptanceCriteria?.length ? [
-		"### Acceptance criteria",
-		"",
-		...acceptanceCriteria.map((c) => `- ${c}`),
-		""
-	].join("\n") : "";
+	const { brief, title, seedFiles, scopeHint } = input;
 	const seedSection = seedFiles?.length ? [
 		"### Seed files",
 		"",
@@ -6754,7 +6762,6 @@ function buildFulfillBriefUserPrompt(input, ctx) {
 		"",
 		brief,
 		"",
-		criteriaSection,
 		seedSection,
 		correlationSection,
 		workspaceSection,
@@ -6894,13 +6901,8 @@ function buildJudgeEvalVariantUserPrompt(input, ctx) {
 function buildJudgePackUserPrompt(input, ctx) {
 	const { renderedPackId, sourcePackId, successCriteria } = input;
 	const rubric = successCriteria.rubric;
-	const criteriaList = rubric.criteria.map((c, i) => `${i + 1}. **${c.id}** (weight ${c.weight}, scoring: \`${c.scoring}\`) — ${c.description}`).join("\n");
-	const preambleSection = rubric.preamble ? [
-		"### Rubric preamble",
-		"",
-		rubric.preamble,
-		""
-	].join("\n") : null;
+	const criteriaList = renderRubricCriteriaList(rubric);
+	const preambleSection = renderRubricPreambleSection(rubric);
 	return [
 		"# Judge Pack Agent",
 		"",
@@ -7016,6 +7018,112 @@ function buildJudgePackUserPrompt(input, ctx) {
 	].filter((l) => l !== null).join("\n");
 }
 //#endregion
+//#region ../../libs/agent-runtime/src/prompts/pr-review.ts
+function buildPrReviewUserPrompt(input, ctx) {
+	const rubric = input.successCriteria.rubric;
+	const criteriaList = renderRubricCriteriaList(rubric);
+	const preambleSection = renderRubricPreambleSection(rubric);
+	const taskPromptSection = input.taskPrompt ? [
+		"## Task-specific instructions",
+		"",
+		input.taskPrompt,
+		""
+	].join("\n") : "";
+	const resourceSection = input.subject.resourceUrls && input.subject.resourceUrls.length > 0 ? [
+		"### Resources",
+		"",
+		...input.subject.resourceUrls.map((url) => `- ${url}`),
+		""
+	].join("\n") : "";
+	const hintsSection = input.subject.inspectionHints && input.subject.inspectionHints.length > 0 ? [
+		"### Inspection hints",
+		"",
+		...input.subject.inspectionHints.map((hint) => `- ${hint}`),
+		""
+	].join("\n") : "";
+	const workspaceSection = ctx.workspace?.mode === "dedicated_worktree" ? [
+		"### Workspace",
+		"",
+		"This review attempt is running inside a dedicated disposable git",
+		"worktree. Inspect and reason inside this workspace only.",
+		ctx.workspace.branch ? `The current review branch is \`${ctx.workspace.branch}\`.` : "The current checkout is disposable and will be cleaned up when the task ends.",
+		""
+	].join("\n") : "";
+	return [
+		"# Review Agent",
+		"",
+		"You are an independent judge. You did NOT produce the subject under review.",
+		"Assess it strictly against the rubric below and emit a structured judgment.",
+		"You may inspect the local workspace and the referenced resources, but do NOT modify anything.",
+		"",
+		`Your diary ID is: ${ctx.diaryId}`,
+		`This task's id is: ${ctx.taskId}`,
+		"",
+		"## Subject",
+		"",
+		`**Title:** ${input.subject.title}`,
+		"",
+		input.subject.summary,
+		"",
+		resourceSection,
+		hintsSection,
+		workspaceSection,
+		"### Execution contract",
+		"",
+		"Treat the provided subject, resources, inspection hints, and any",
+		"task-specific instructions as the full",
+		"review contract for this task.",
+		"",
+		"If the task-specific instructions or inspection hints require an outward action tied to the review",
+		"(for example publishing the judgment somewhere), perform that action as",
+		"part of the task before reporting structured output.",
+		"",
+		"## Review workflow",
+		"",
+		"1. Read the subject summary, resources, inspection hints, and any",
+		"   task-specific instructions before scoring.",
+		"2. Inspect the target artefact directly using the tools and resources the",
+		"   task makes available.",
+		"3. If you are in a dedicated disposable worktree and need the review target",
+		"   checked out locally, do that work inside this disposable workspace only.",
+		"4. Apply the rubric strictly. This task is about complexity and",
+		"   reviewability, not correctness or feature desirability.",
+		"5. Perform any required outward action before emitting the final",
+		"   structured output.",
+		"",
+		taskPromptSection,
+		preambleSection,
+		"## Criteria",
+		"",
+		criteriaList,
+		"",
+		"### Scoring rules",
+		"",
+		"- Every criterion uses binary scoring only.",
+		"- Score `1` when the subject clearly clears the criterion.",
+		"- Score `0` when it does not, or when the evidence is ambiguous.",
+		"- `rationale` is REQUIRED for every score. Keep it concrete and audit-friendly.",
+		"- Compute `composite = Σ(weight_i × score_i)` exactly; the runtime rejects mismatches.",
+		"",
+		"Write a signed diary entry (tags: `judgment`, `pr_review`) capturing the rationale before reporting structured output.",
+		"",
+		buildFinalOutputBlock({
+			taskType: "pr_review",
+			outputSchemaName: "PrReviewOutput",
+			shapeSketch: [
+				"{",
+				"  \"scores\": [",
+				"    { \"criterionId\": \"...\", \"score\": 0, \"rationale\": \"...\" }",
+				"  ],",
+				"  \"composite\": <sum-of-weighted-binary-scores>,",
+				"  \"verdict\": \"<1-3 sentence overall>\"",
+				"}"
+			].join("\n"),
+			extraNotes: ["`scores` MUST stay in the same order as the rubric criteria.", "`score` MUST be exactly `0` or `1` for every criterion."]
+		})
+	].filter(Boolean).join("\n");
+}
+//#endregion
 //#region ../../libs/agent-runtime/src/prompts/render-pack.ts
 /**
 * Build the first user-message prompt for a `render_pack` task. Almost mechanical:
@@ -7198,6 +7306,16 @@ function buildTaskUserPrompt(task, ctx) {
 				diaryId: ctx.diaryId,
 				taskId: ctx.taskId
 			});
+		case PR_REVIEW_TYPE:
+			if (!Check(PrReviewInput, task.input)) {
+				const errors = [...Errors(PrReviewInput, task.input)];
+				throw new Error(`pr_review input failed validation: ${JSON.stringify(errors.slice(0, 3))}`);
+			}
+			return buildPrReviewUserPrompt(task.input, {
+				diaryId: ctx.diaryId,
+				taskId: ctx.taskId,
+				workspace: ctx.workspace
+			});
 		case JUDGE_EVAL_VARIANT_TYPE:
 			if (!Check(JudgeEvalVariantInput, task.input)) {
 				const errors = [...Errors(JudgeEvalVariantInput, task.input)];
@@ -13916,6 +14034,31 @@ function abortableSleep(ms, signal) {
 	});
 }
 //#endregion
+//#region ../../libs/agent-runtime/src/subagent-output-contracts.ts
+/**
+* Construct an immutable contract registry from a static list.
+*
+* The resulting registry is safe to share across sessions and
+* invocations — no mutation is possible after construction.
+*/
+function createSubagentContractRegistry(contracts) {
+	const lookup = /* @__PURE__ */ new Map();
+	for (const c of contracts) {
+		if (!c.name || c.name.trim().length === 0) throw new Error("subagent output contract name is required");
+		if (!/^[a-z][a-z0-9_]*$/.test(c.name)) throw new Error(`subagent output contract name '${c.name}' must be lower_snake_case (starts with a letter, then [a-z0-9_]+)`);
+		if (lookup.has(c.name)) throw new Error(`duplicate subagent output contract name '${c.name}' in constructor args`);
+		lookup.set(c.name, c);
+	}
+	return {
+		get(name) {
+			return lookup.get(name) ?? null;
+		},
+		list() {
+			return [...lookup.values()];
+		}
+	};
+}
+//#endregion
 //#region ../../libs/pi-extension/src/moltnet/render-phase6.ts
 function slugToTitle(value) {
 	return value.split(/[:/_-]+/).filter(Boolean).map((part) => part[0]?.toUpperCase() + part.slice(1)).join(" ");
@@ -14892,7 +15035,8 @@ function findMainWorktree() {
 function loadCredentials(agentDir) {
 	const moltnetJson = readFileSync(path.join(agentDir, "moltnet.json"), "utf8");
 	const agentEnvRaw = readFileSync(path.join(agentDir, "env"), "utf8");
-	const piAuthPath = process.env.PI_AUTH_PATH ?? path.join(process.env.HOME ?? "", ".pi", "agent", "auth.json");
+	const piAgentDir = process.env.PI_CODING_AGENT_DIR ?? path.join(process.env.HOME ?? "", ".pi", "agent");
+	const piAuthPath = path.join(piAgentDir, "auth.json");
 	const piAuthJson = existsSync(piAuthPath) ? readFileSync(piAuthPath, "utf8") : null;
 	const gitconfigPath = path.join(agentDir, "gitconfig");
 	const gitconfig = existsSync(gitconfigPath) ? readFileSync(gitconfigPath, "utf8") : null;
@@ -15404,7 +15548,7 @@ async function buildAgentSession(args) {
 		spanAttributes: args.otelSpanAttrs
 	});
 	const resourceLoader = new DefaultResourceLoader({
-		cwd: args.mountPath,
+		cwd: args.cwdPath,
 		agentDir: args.piAuthDir,
 		extensionFactories: [piOtelExtension],
 		appendSystemPrompt: args.appendSystemPrompt,
@@ -15412,12 +15556,12 @@ async function buildAgentSession(args) {
 	});
 	await resourceLoader.reload();
 	const sessionManager = args.sessionPersistence ? await resolvePersistentSessionManager({
-		cwd: args.mountPath,
+		cwd: args.cwdPath,
 		sessionDir: args.sessionPersistence.sessionDir
-	}) : SessionManager.inMemory(args.mountPath);
+	}) : SessionManager.inMemory(args.cwdPath);
 	return (await createAgentSession({
 		agentDir: args.piAuthDir,
-		cwd: args.mountPath,
+		cwd: args.cwdPath,
 		model: args.modelHandle,
 		customTools: args.customTools,
 		sessionManager,
@@ -15662,6 +15806,7 @@ var DEFAULT_SUBAGENT_TIMEOUT_MS = 300 * 1e3;
 */
 function createSubagentTool(args) {
 	const buildSession = args.buildAgentSession ?? buildAgentSession;
+	const { contractRegistry } = args;
 	let callCount = 0;
 	return {
 		tool: defineTool({
@@ -15672,8 +15817,8 @@ function createSubagentTool(args) {
 			async execute(_id, params) {
 				if (!Check(SubagentToolParameters, params)) return toolError(`subagent: invalid parameters: ${JSON.stringify([...Errors(SubagentToolParameters, params)].slice(0, 3))}`);
 				const { task, output_schema } = params;
-				const contract = getSubagentOutputContract(output_schema);
-				if (!contract) return toolError(`subagent: unknown output_schema "${output_schema}". Registered contracts: [${listSubagentOutputContracts().map((c) => c.name).join(", ")}]`);
+				const contract = contractRegistry.get(output_schema);
+				if (!contract) return toolError(`subagent: unknown output_schema "${output_schema}". Registered contracts: [${contractRegistry.list().map((c) => c.name).join(", ")}]`);
 				callCount += 1;
 				const callIndex = callCount;
 				let captured = null;
@@ -15703,6 +15848,7 @@ function createSubagentTool(args) {
 				});
 				const session = await buildSession({
 					mountPath: args.mountPath,
+					cwdPath: args.cwdPath ?? args.mountPath,
 					piAuthDir: args.piAuthDir,
 					modelHandle: args.modelHandle,
 					agentName: args.agentName,
@@ -16043,6 +16189,7 @@ function prepareTaskWorkspace(task, requestedMountPath, executionPlan) {
 	const branch = executionPlan?.worktreeBranch ?? null;
 	if (!branch) return {
 		mountPath: requestedMountPath,
+		cwdPath: requestedMountPath,
 		mode: "shared_mount",
 		branch: null,
 		cleanup: () => {}
@@ -16050,7 +16197,7 @@ function prepareTaskWorkspace(task, requestedMountPath, executionPlan) {
 	const mainRepo = findMainWorktree();
 	const worktreeDir = resolveTaskWorktreePath(mainRepo, executionPlan?.workspaceId ?? `task-${task.id}`);
 	const relMount = relative(mainRepo, requestedMountPath);
-	const mountPath = relMount === "" || relMount.startsWith("..") ? worktreeDir : join(worktreeDir, relMount);
+	const cwdPath = relMount === "" || relMount.startsWith("..") ? worktreeDir : join(worktreeDir, relMount);
 	const keepWorkspace = executionPlan?.workspaceScope === "session" && executionPlan.sessionKey !== null;
 	if (keepWorkspace) ensureReusableTaskWorktree(mainRepo, worktreeDir, branch);
 	else {
@@ -16058,7 +16205,8 @@ function prepareTaskWorkspace(task, requestedMountPath, executionPlan) {
 		addTaskWorktree(mainRepo, worktreeDir, branch);
 	}
 	return {
-		mountPath,
+		mountPath: mainRepo,
+		cwdPath,
 		mode: "dedicated_worktree",
 		branch,
 		cleanup: keepWorkspace ? () => {} : () => {
@@ -16171,15 +16319,24 @@ var noopTurnEventHandler = () => {};
 function createPiTaskExecutor(opts) {
 	let cachedCheckpoint = opts.checkpointPath ?? null;
 	return async (claimedTask, reporter) => {
-		if (!cachedCheckpoint) cachedCheckpoint = await ensureSnapshot({
-			config: opts.sandboxConfig?.snapshot,
-			onProgress: opts.onSnapshotProgress ?? ((m) => {
-				process.stderr.write(`[snapshot] ${m}\n`);
-			})
+		const reporterWasOpened = !reporter.cancelSignal.aborted;
+		if (reporterWasOpened) await reporter.open({
+			taskId: claimedTask.task.id,
+			attemptN: claimedTask.attemptN
 		});
 		return executePiTask(claimedTask, reporter, {
 			...opts,
-			checkpointPath: cachedCheckpoint
+			checkpointPath: cachedCheckpoint ?? void 0,
+			resolveCheckpointPath: async () => {
+				if (!cachedCheckpoint) cachedCheckpoint = await ensureSnapshot({
+					config: opts.sandboxConfig?.snapshot,
+					onProgress: opts.onSnapshotProgress ?? ((m) => {
+						process.stderr.write(`[snapshot] ${m}\n`);
+					})
+				});
+				return cachedCheckpoint;
+			},
+			reporterAlreadyOpened: reporterWasOpened
 		});
 	};
 }
@@ -16195,8 +16352,9 @@ async function executePiTask(claimedTask, reporter, opts) {
 	const startTime = Date.now();
 	const requestedMountPath = opts.mountPath ?? process.cwd();
 	const executionPlan = opts.makeExecutionPlan?.(claimedTask) ?? null;
-	const workspace = prepareTaskWorkspace(task, requestedMountPath, executionPlan);
-	const mountPath = workspace.mountPath;
+	let workspace = null;
+	let mountPath = requestedMountPath;
+	let cwdPath = requestedMountPath;
 	if (reporter.cancelSignal.aborted) return {
 		taskId: task.id,
 		attemptN,
@@ -16211,33 +16369,8 @@ async function executePiTask(claimedTask, reporter, opts) {
 			retryable: false
 		}
 	};
-	const checkpointPath = opts.checkpointPath ?? await ensureSnapshot({
-		config: opts.sandboxConfig?.snapshot,
-		onProgress: opts.onSnapshotProgress ?? ((m) => {
-			process.stderr.write(`[snapshot] ${m}\n`);
-		})
-	});
-	const mainRepoForRepair = findMainWorktree();
-	try {
-		execFileSync("git", [
-			"-C",
-			mainRepoForRepair,
-			"worktree",
-			"repair",
-			"--relative-paths"
-		], { stdio: "pipe" });
-	} catch {}
+	let reporterOpen = opts.reporterAlreadyOpened ?? false;
 	let managed = null;
-	managed = await resumeVm({
-		checkpointPath,
-		agentName: opts.agentName,
-		mountPath,
-		extraAllowedHosts: opts.extraAllowedHosts,
-		sandboxConfig: opts.sandboxConfig
-	});
-	const diaryId = task.diaryId ?? "";
-	const taskTeamId = task.teamId ?? "";
-	let reporterOpen = false;
 	let session = null;
 	let subagentHandle = null;
 	const finalUsage = emptyUsage(opts.provider, opts.model);
@@ -16256,41 +16389,103 @@ async function executePiTask(claimedTask, reporter, opts) {
 			retryable: false
 		}
 	});
+	let onTurnEvent;
+	if (opts.makeOnTurnEvent) try {
+		onTurnEvent = opts.makeOnTurnEvent(claimedTask);
+	} catch (err) {
+		process.stderr.write(`[emit] makeOnTurnEvent threw: ${err instanceof Error ? err.message : String(err)}\n`);
+		onTurnEvent = noopTurnEventHandler;
+	}
+	else onTurnEvent = opts.onTurnEvent ?? noopTurnEventHandler;
+	const emit = (kind, payload) => {
+		try {
+			onTurnEvent(kind, summarizePayloadForLog(kind, payload));
+		} catch (err) {
+			process.stderr.write(`[emit] onTurnEvent threw for kind="${kind}": ${err instanceof Error ? err.message : String(err)}\n`);
+		}
+		return reporter.record({
+			kind,
+			payload
+		});
+	};
+	const emitError = async (phase, message, extra = {}) => {
+		await emit("error", {
+			phase,
+			message,
+			...extra
+		});
+	};
 	try {
-		const mainRepo = findMainWorktree();
-		activateAgentEnv(managed.credentials.agentEnv, mainRepo);
-		await reporter.open({
+		if (!opts.reporterAlreadyOpened) await reporter.open({
 			taskId: task.id,
 			attemptN
 		});
 		reporterOpen = true;
-		let onTurnEvent;
-		if (opts.makeOnTurnEvent) try {
-			onTurnEvent = opts.makeOnTurnEvent(claimedTask);
+		let checkpointPath;
+		try {
+			checkpointPath = opts.checkpointPath ?? (opts.resolveCheckpointPath ? await opts.resolveCheckpointPath() : await ensureSnapshot({
+				config: opts.sandboxConfig?.snapshot,
+				onProgress: opts.onSnapshotProgress ?? ((m) => {
+					process.stderr.write(`[snapshot] ${m}\n`);
+				})
+			}));
+		} catch (err) {
+			const message = err instanceof Error ? err.message : String(err);
+			await emitError("snapshot", message);
+			return makeFailedOutput("snapshot_failed", message);
+		}
+		try {
+			workspace = prepareTaskWorkspace(task, requestedMountPath, executionPlan);
+			mountPath = workspace.mountPath;
+			cwdPath = workspace.cwdPath;
 		} catch (err) {
-			process.stderr.write(`[emit] makeOnTurnEvent threw: ${err instanceof Error ? err.message : String(err)}\n`);
-			onTurnEvent = noopTurnEventHandler;
+			const message = err instanceof Error ? err.message : String(err);
+			await emitError("worktree_setup", message);
+			return makeFailedOutput("worktree_setup_failed", message);
 		}
-		else onTurnEvent = opts.onTurnEvent ?? noopTurnEventHandler;
-		const emit = (kind, payload) => {
+		try {
+			const mainRepoForRepair = findMainWorktree();
 			try {
-				onTurnEvent(kind, summarizePayloadForLog(kind, payload));
-			} catch (err) {
-				process.stderr.write(`[emit] onTurnEvent threw for kind="${kind}": ${err instanceof Error ? err.message : String(err)}\n`);
-			}
-			return reporter.record({
-				kind,
-				payload
+				execFileSync("git", [
+					"-C",
+					mainRepoForRepair,
+					"worktree",
+					"repair",
+					"--relative-paths"
+				], { stdio: "pipe" });
+			} catch {}
+		} catch (err) {
+			const message = err instanceof Error ? err.message : String(err);
+			await emitError("worktree_setup", message);
+			return makeFailedOutput("worktree_setup_failed", message);
+		}
+		try {
+			managed = await resumeVm({
+				checkpointPath,
+				agentName: opts.agentName,
+				mountPath,
+				extraAllowedHosts: opts.extraAllowedHosts,
+				sandboxConfig: opts.sandboxConfig
 			});
-		};
+		} catch (err) {
+			const message = err instanceof Error ? err.message : String(err);
+			await emitError("vm_resume", message);
+			return makeFailedOutput("vm_resume_failed", message);
+		}
+		const diaryId = task.diaryId ?? "";
+		const taskTeamId = task.teamId ?? "";
+		const mainRepo = findMainWorktree();
+		activateAgentEnv(managed.credentials.agentEnv, mainRepo);
+		const activeWorkspace = workspace;
+		if (!activeWorkspace) throw new Error("task workspace not prepared");
 		await emit("info", {
 			event: "execute_start",
 			taskType: task.taskType,
 			teamId: task.teamId,
 			provider: opts.provider,
 			model: opts.model,
-			workspaceMode: workspace.mode,
-			workspaceBranch: workspace.branch
+			workspaceMode: activeWorkspace.mode,
+			workspaceBranch: activeWorkspace.branch
 		});
 		let taskPrompt;
 		try {
@@ -16298,8 +16493,8 @@ async function executePiTask(claimedTask, reporter, opts) {
 				diaryId,
 				taskId: task.id,
 				workspace: {
-					mode: workspace.mode,
-					branch: workspace.branch
+					mode: activeWorkspace.mode,
+					branch: activeWorkspace.branch
 				},
 				extras: opts.promptExtras
 			});
@@ -16351,7 +16546,7 @@ async function executePiTask(claimedTask, reporter, opts) {
 				getTeamId: () => taskTeamId,
 				getSessionErrors: () => [],
 				clearSessionErrors: () => {},
-				getHostCwd: () => mountPath,
+				getHostCwd: () => cwdPath,
 				hostExecBaseEnv: new Set([...HOST_EXEC_DEFAULT_BASE_ENV, ...Object.keys(managed.credentials.agentEnv)]),
 				hostExecAutoApprove: opts.hostExecAutoApprove ?? opts.sandboxConfig?.hostExec?.autoApprove ?? false,
 				getTaskContext: () => ({
@@ -16379,6 +16574,7 @@ async function executePiTask(claimedTask, reporter, opts) {
 			if (taskTypeUsesSubagents(task.taskType)) {
 				subagentHandle = createSubagentTool({
 					mountPath,
+					cwdPath,
 					piAuthDir,
 					modelHandle,
 					agentName: opts.agentName,
@@ -16387,12 +16583,14 @@ async function executePiTask(claimedTask, reporter, opts) {
 					parentTaskId: task.id,
 					parentTaskType: task.taskType,
 					parentAttemptN: attemptN,
+					contractRegistry: opts.subagentContractRegistry,
 					parentCancelSignal: reporter.cancelSignal
 				});
 				parentSubagentTools.push(subagentHandle.tool);
 			}
 			session = await buildAgentSession({
 				mountPath,
+				cwdPath,
 				piAuthDir,
 				modelHandle,
 				agentName: opts.agentName,
@@ -16477,6 +16675,10 @@ async function executePiTask(claimedTask, reporter, opts) {
 					is_error: event.isError,
 					result: event.isError ? truncateForWire(event.result) : void 0
 				}));
+				if (event.isError) track(emitError("tool_call_error", describeToolErrorMessage(event.result), {
+					tool: event.toolName,
+					result: truncateForWire(event.result)
+				}));
 				if (maxBashTimeouts > 0 && event.toolName === "bash" && event.isError && isBashTimeoutResult(event.result)) {
 					bashTimeoutCount += 1;
 					if (bashTimeoutCount >= maxBashTimeouts) triggerCapAbort("max_bash_timeouts_exceeded", `Aborted after ${bashTimeoutCount} bash timeouts in this attempt (cap ${maxBashTimeouts}).`);
@@ -16632,7 +16834,7 @@ async function executePiTask(claimedTask, reporter, opts) {
 			}
 		}
 		if (managed) await managed.vm.close();
-		try {
+		if (workspace) try {
 			workspace.cleanup();
 		} catch (err) {
 			const detail = err instanceof Error ? err.message : String(err);
@@ -16743,6 +16945,23 @@ function truncateForWire(value) {
 		};
 	}
 }
+function describeToolErrorMessage(result) {
+	if (typeof result === "string" && result.trim().length > 0) return result.trim();
+	if (result && typeof result === "object") {
+		const content = result.content;
+		if (Array.isArray(content)) {
+			for (const item of content) if (item && typeof item === "object" && typeof item.text === "string") {
+				const text = item.text.trim();
+				if (text.length > 0) return text;
+			}
+		}
+	}
+	try {
+		return JSON.stringify(truncateForWire(result));
+	} catch {
+		return "Tool call failed";
+	}
+}
 //#endregion
 //#region src/config.ts
 function loadConfig() {
@@ -17619,9 +17838,15 @@ async function runPolling(opts) {
 		maxPollIntervalMs
 	}, "agent-daemon.starting");
 	const outputs = [];
+	const subagentContractRegistry = createSubagentContractRegistry([{
+		name: "judge_eval_variant_result",
+		description: "Per-variant grading result produced by a subagent of judge_eval_variant: scores against the shared rubric, composite, and a 1-3 sentence verdict for a single variant.",
+		parametersSchema: JudgeEvalVariantResult
+	}]);
 	try {
 		const executeTask = createPiTaskExecutor({
 			agentName: common.agent,
+			subagentContractRegistry,
 			mountPath: sandbox.rootDir,
 			provider: common.provider,
 			model: common.model,
@@ -17667,11 +17892,19 @@ async function runPolling(opts) {
 			executeTask: async (claimedTask, reporter) => {
 				const executionPlan = executionPlans.getOrCreate(claimedTask);
 				const sessionDescriptor = executionPlan.descriptor;
-				const expired = slotRegistry.reapExpiredSlots();
-				if (expired.length > 0) rootLogger.info({
-					expiredCount: expired.length,
-					slotKeys: expired.map((item) => item.slot.slotKey)
-				}, "agent-daemon.daemon_slots_reaped");
+				let expired;
+				try {
+					expired = slotRegistry.reapExpiredSlots();
+					if (expired.length > 0) rootLogger.info({
+						expiredCount: expired.length,
+						slotKeys: expired.map((item) => item.slot.slotKey)
+					}, "agent-daemon.daemon_slots_reaped");
+				} catch (err) {
+					rootLogger.error({
+						phase: "daemon_slot_reap",
+						err: err instanceof Error ? err.message : String(err)
+					}, "agent-daemon.daemon_slot_reap_failed");
+				}
 				rootLogger.debug({
 					taskId: claimedTask.task.id,
 					taskType: claimedTask.task.taskType,
@@ -17866,9 +18099,15 @@ async function runOnce(argv) {
 	process.on("SIGTERM", () => {
 		onSignal("SIGTERM");
 	});
+	const subagentContractRegistry = createSubagentContractRegistry([{
+		name: "judge_eval_variant_result",
+		description: "Per-variant grading result produced by a subagent of judge_eval_variant: scores against the shared rubric, composite, and a 1-3 sentence verdict for a single variant.",
+		parametersSchema: JudgeEvalVariantResult
+	}]);
 	try {
 		const rawExecuteTask = createPiTaskExecutor({
 			agentName: opts.agent,
+			subagentContractRegistry,
 			mountPath: sandbox.rootDir,
 			provider: opts.provider,
 			model: opts.model,
@@ -17879,11 +18118,18 @@ async function runOnce(argv) {
 			maxBashTimeouts: opts.maxBashTimeouts
 		});
 		const executeTask = async (claimedTask, reporter) => {
-			const expired = slotRegistry.reapExpiredSlots();
-			if (expired.length > 0) rootLogger.info({
-				expiredCount: expired.length,
-				slotKeys: expired.map((item) => item.slot.slotKey)
-			}, "agent-daemon.daemon_slots_reaped");
+			try {
+				const expired = slotRegistry.reapExpiredSlots();
+				if (expired.length > 0) rootLogger.info({
+					expiredCount: expired.length,
+					slotKeys: expired.map((item) => item.slot.slotKey)
+				}, "agent-daemon.daemon_slots_reaped");
+			} catch (err) {
+				rootLogger.error({
+					phase: "daemon_slot_reap",
+					err: err instanceof Error ? err.message : String(err)
+				}, "agent-daemon.daemon_slot_reap_failed");
+			}
 			const executionPlan = executionPlans.getOrCreate(claimedTask);
 			if (executionPlan.slotKey && executionPlan.sessionPersistence) slotRegistry.beginSlot({
 				...slotIdentity,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@themoltnet/agent-daemon",
-  "version": "0.7.0",
+  "version": "0.9.0",
   "license": "AGPL-3.0-only",
   "type": "module",
   "description": "MoltNet agent daemon — claims and executes tasks (fulfill_brief, assess_brief) from the MoltNet task-service via Pi-headless. CLI: moltnet-agent.",
@@ -33,9 +33,9 @@
     "@opentelemetry/semantic-conventions": "^1.39.0",
     "pino": "^10.3.1",
     "pino-pretty": "^13.1.3",
-    "@themoltnet/pi-extension": "0.17.0",
-    "@themoltnet/agent-runtime": "0.15.1",
-    "@themoltnet/sdk": "0.102.0"
+    "@themoltnet/agent-runtime": "0.16.0",
+    "@themoltnet/sdk": "0.102.0",
+    "@themoltnet/pi-extension": "0.18.1"
   },
   "devDependencies": {
     "tsx": "^4.7.0",
@@ -43,9 +43,9 @@
     "vite": "^8.0.0",
     "vitest": "^3.0.0",
     "@moltnet/bootstrap": "0.1.0",
+    "@moltnet/crypto-service": "0.1.0",
     "@moltnet/database": "0.1.0",
-    "@moltnet/tasks": "0.1.0",
-    "@moltnet/crypto-service": "0.1.0"
+    "@moltnet/tasks": "0.1.0"
   },
   "nx": {
     "tags": [