@iowarp/clio-coder 0.3.4 → 0.3.6

package/src/domains/dispatch/extension.ts

@@ -19,6 +19,7 @@ import { GUARDRAIL_DEFAULTS } from "../../core/guardrails.js";
 import { readClioVersion, readPiMonoVersion } from "../../core/package-root.js";
 import { canonicalizeExistingPath } from "../../core/path-canonical.js";
 import { protectedResidencyModels } from "../../core/residency-protection.js";
+import { responseModelIdObservationFromRecord } from "../../core/response-model-id.js";
 import {
 	responseSchemaConflictsWithTools,
 	runtimeSpeaksResponseSchemaDialect,
@@ -134,6 +135,7 @@ import {
 	registerDetachedBatch,
 } from "./batch-store.js";
 import { type BatchState, createBatch, onRunComplete, snapshotBatch } from "./batch-tracker.js";
+import { type RunToolBudgetEnvelope, resolveToolBudgetEnvelope } from "./budget-envelope.js";
 import { assessCapabilityMismatch, type CapabilityMismatch } from "./capability-match.js";
 import { capacityLeaseUsage, createNodeLeaseUsageReader } from "./capacity-lease.js";
 import type {
@@ -307,6 +309,7 @@ interface ActiveRun {
 	requestOrigin?: DispatchRequestOrigin;
 	agentId: string;
 	task: string;
+	budget?: RunToolBudgetEnvelope;
 	cwd: string;
 	/** Fleet node this run was placed on; null when no placement resolved it. */
 	node: RunNodeIdentity | null;
@@ -1287,6 +1290,7 @@ interface DispatchLifecycleStage {
 	capabilityMismatch: CapabilityMismatch | null;
 	effectiveAutonomy: AutonomyLevel;
 	budget: WorkerBudget;
+	budgetEnvelope: RunToolBudgetEnvelope;
 	settings?: Readonly<ReturnType<ConfigContract["get"]>>;
 }
 
@@ -1411,7 +1415,7 @@ function assertWriteRootsEnforceable(runtime: RuntimeDescriptor, writeRoots: Rea
 function assertWorkerBudgetEnforceable(runtime: RuntimeDescriptor, hasDeclaredBudget: boolean): void {
 	if (!hasDeclaredBudget || runtime.kind !== "subprocess") return;
 	throw new Error(
-		`dispatch: runtime '${runtime.id}' cannot enforce an explicit agent budget because subprocess workers do not expose per-tool mediation; choose a native or claude-sdk worker`,
+		`dispatch: runtime '${runtime.id}' cannot enforce an explicit dispatch budget because subprocess workers do not expose per-tool mediation; choose a native or claude-sdk worker`,
 	);
 }
 
@@ -1543,26 +1547,36 @@ function assertPostRuntimeToolCompatibility(
 	);
 }
 
+function workerToolCallHardCap(settings: EffectiveSettings): number {
+	const rawEnvCap = process.env.CLIO_CODER_WORKER_TOOL_CALL_CAP?.trim();
+	const parsedEnvCap = rawEnvCap && /^[1-9]\d*$/.test(rawEnvCap) ? Number(rawEnvCap) : Number.NaN;
+	return Number.isSafeInteger(parsedEnvCap)
+		? parsedEnvCap
+		: (settings?.guardrails.workerToolCallCap ?? GUARDRAIL_DEFAULTS.workerToolCallCap);
+}
+
 function resolveEffectiveWorkerBudget(input: {
+	req: DispatchRequest;
+	recipeId: string;
 	declared: ReturnType<typeof normalizeAgentSpec>["budget"];
 	allowedTools: ReadonlyArray<ToolName>;
 	settings: EffectiveSettings;
-}): WorkerBudget {
-	const rawEnvCap = process.env.CLIO_CODER_WORKER_TOOL_CALL_CAP?.trim();
-	const parsedEnvCap = rawEnvCap && /^[1-9]\d*$/.test(rawEnvCap) ? Number(rawEnvCap) : Number.NaN;
-	const hardCap = Number.isSafeInteger(parsedEnvCap)
-		? parsedEnvCap
-		: (input.settings?.guardrails.workerToolCallCap ?? GUARDRAIL_DEFAULTS.workerToolCallCap);
+}): RunToolBudgetEnvelope {
+	const hardCap = workerToolCallHardCap(input.settings);
 	const declared = input.declared ?? {
 		toolCalls: hardCap,
 		readReserve: Math.min(5, Math.max(0, hardCap - 1)),
 		synthesis: true,
 	};
-	const toolCalls = Math.min(declared.toolCalls, hardCap);
-	const readReserve = input.allowedTools.includes(ToolNames.Read)
-		? Math.min(declared.readReserve, Math.max(0, toolCalls - 1))
-		: 0;
-	return { toolCalls, readReserve, synthesis: declared.synthesis, hardCap };
+	return resolveToolBudgetEnvelope({
+		recipeId: input.recipeId,
+		policy: declared,
+		...(input.req.budget === undefined ? {} : { request: input.req.budget }),
+		hardCap,
+		hasReadTool: input.allowedTools.includes(ToolNames.Read),
+		retry: (input.req.lineage?.attempt ?? 0) > 0,
+		revision: input.req.gate?.role === "builder" && input.req.gate.cycle > 1 && input.req.gate.verdict === "revise",
+	});
 }
 
 function frozenProtectedArtifactState(state: ProtectedArtifactState | undefined): ProtectedArtifactState {
@@ -3193,7 +3207,7 @@ export function createDispatchBundle(
 		);
 		enforceCapabilityGate(target.target.id, target.modelCapabilities, req.requiredCapabilities);
 		assertRuntimeCanHonorWorkerPermissionMode(target.runtime, settings?.workers.onPermission ?? "deny");
-		assertWorkerBudgetEnforceable(target.runtime, spec.budget !== null);
+		assertWorkerBudgetEnforceable(target.runtime, spec.budget !== null || req.budget !== undefined);
 
 		const cwd = req.cwd ?? process.cwd();
 		const sessionAutonomy = settings?.autonomy ?? "auto-edit";
@@ -3231,7 +3245,13 @@ export function createDispatchBundle(
 			workingContextPaths: workerWorkingContextPaths(req),
 		});
 		const systemPrompt = compiledWorkerPrompt.systemPrompt;
-		const budget = resolveEffectiveWorkerBudget({ declared: spec.budget, allowedTools: effectiveTools, settings });
+		const budgetEnvelope = resolveEffectiveWorkerBudget({
+			req,
+			recipeId: recipe.id,
+			declared: spec.budget,
+			allowedTools: effectiveTools,
+			settings,
+		});
 		// Fetch structured project context only for tiers that receive it, so
 		// read-only scouts never pay the CLIO-CODER.md read. The tier is spec policy
 		// (capability-class default, recipe frontmatter override).
@@ -3291,7 +3311,8 @@ export function createDispatchBundle(
 			operatorProfileApplied: compiledWorkerPrompt.operatorProfileApplied ?? false,
 			capabilityMismatch,
 			effectiveAutonomy,
-			budget,
+			budget: budgetEnvelope.effective,
+			budgetEnvelope,
 			...(settings ? { settings } : {}),
 		};
 	}
@@ -3473,6 +3494,7 @@ export function createDispatchBundle(
 					role?: string;
 					usage?: unknown;
 					model?: unknown;
+					responseModelIdObservation?: unknown;
 					responseModel?: unknown;
 					responseId?: unknown;
 					stopReason?: unknown;
@@ -3522,12 +3544,16 @@ export function createDispatchBundle(
 				tokenMeter.cacheReadTokens += typeof u.cacheRead === "number" ? u.cacheRead : 0;
 				tokenMeter.cacheWriteTokens += typeof u.cacheWrite === "number" ? u.cacheWrite : 0;
 				tokenMeter.reasoningTokens += extractReasoningTokenCount(u);
-				const model = readStringOrNull(event.message.model);
-				const responseModel = readStringOrNull(event.message.responseModel);
-				const responseId = readStringOrNull(event.message.responseId);
-				if (model !== null || responseModel !== null || responseId !== null) {
-					upstreamResponses.push({ model, responseModel, responseId });
-				}
+				const requestedModelId = readStringOrNull(event.message.model);
+				const responseModelIdObservation = responseModelIdObservationFromRecord(event.message, "not-observed");
+				const differingResponseModelId = readStringOrNull(event.message.responseModel);
+				const providerResponseId = readStringOrNull(event.message.responseId);
+				upstreamResponses.push({
+					requestedModelId,
+					responseModelIdObservation,
+					differingResponseModelId,
+					providerResponseId,
+				});
 				if (event.message.stopReason === "error") {
 					const message = readStringOrNull(event.message.errorMessage);
 					if (message !== null) failureMessage = message;
@@ -4142,6 +4168,11 @@ export function createDispatchBundle(
 			return handle;
 		};
 		if (req.delegationAgentId) {
+			if (req.budget !== undefined) {
+				throw new Error(
+					"dispatch: budget envelopes cannot be enforced on an ACP delegation target; dispatch to a native or claude-sdk worker",
+				);
+			}
 			assertPlannedNodeIdentity(req, { id: "local", kind: "local" });
 			assertProtectedArtifactsEnforceable("acp-delegation", false, protectedArtifactState);
 			if (req.responseSchema !== undefined) {
@@ -4372,6 +4403,7 @@ export function createDispatchBundle(
 					role?: string;
 					usage?: unknown;
 					model?: unknown;
+					responseModelIdObservation?: unknown;
 					responseModel?: unknown;
 					responseId?: unknown;
 					stopReason?: unknown;
@@ -4477,12 +4509,16 @@ export function createDispatchBundle(
 				tokenMeter.cacheReadTokens += typeof u.cacheRead === "number" ? u.cacheRead : 0;
 				tokenMeter.cacheWriteTokens += typeof u.cacheWrite === "number" ? u.cacheWrite : 0;
 				tokenMeter.reasoningTokens += extractReasoningTokenCount(u);
-				const model = readStringOrNull(event.message.model);
-				const responseModel = readStringOrNull(event.message.responseModel);
-				const responseId = readStringOrNull(event.message.responseId);
-				if (model !== null || responseModel !== null || responseId !== null) {
-					upstreamResponses.push({ model, responseModel, responseId });
-				}
+				const requestedModelId = readStringOrNull(event.message.model);
+				const responseModelIdObservation = responseModelIdObservationFromRecord(event.message, "not-observed");
+				const differingResponseModelId = readStringOrNull(event.message.responseModel);
+				const providerResponseId = readStringOrNull(event.message.responseId);
+				upstreamResponses.push({
+					requestedModelId,
+					responseModelIdObservation,
+					differingResponseModelId,
+					providerResponseId,
+				});
 				if (event.message.stopReason === "error") {
 					const message = readStringOrNull(event.message.errorMessage);
 					if (message !== null) failureMessage = message;
@@ -4573,6 +4609,7 @@ export function createDispatchBundle(
 				agentAudience: lifecycle.agentAudience,
 				requestOrigin: lifecycle.requestOrigin,
 				task: req.task,
+				budget: lifecycle.budgetEnvelope,
 				targetId: lifecycle.target.target.id,
 				wireModelId: lifecycle.target.wireModelId,
 				runtimeId: lifecycle.target.runtime.id,
@@ -4638,6 +4675,7 @@ export function createDispatchBundle(
 			// reroute lineage depth, and the model context window for the
 			// per-worker meter. All optional so consumers degrade to local/plain.
 			const fleetIdentity = {
+				budget: lifecycle.budgetEnvelope,
 				...(placement !== undefined && placement !== null ? { node: placement.node.id } : {}),
 				...(req.gate !== undefined ? { gate: { role: req.gate.role, cycle: req.gate.cycle } } : {}),
 				...(req.reroutes !== undefined && req.reroutes.length > 0 ? { rerouteCount: req.reroutes.length } : {}),
@@ -4743,6 +4781,7 @@ export function createDispatchBundle(
 			requestOrigin: lifecycle.requestOrigin,
 			agentId: req.agentId,
 			task: req.task,
+			budget: lifecycle.budgetEnvelope,
 			cwd: lifecycle.cwd,
 			node: placement?.node ?? null,
 			aborted: false,
@@ -4817,6 +4856,7 @@ export function createDispatchBundle(
 				agentAudience: lifecycle.agentAudience,
 				requestOrigin: lifecycle.requestOrigin,
 				task: req.task,
+				budget: lifecycle.budgetEnvelope,
 				targetId: lifecycle.target.target.id,
 				wireModelId: lifecycle.target.wireModelId,
 				runtimeId: lifecycle.target.runtime.id,
@@ -5306,6 +5346,11 @@ export function createDispatchBundle(
 		req = validation.restore(validated.spec);
 
 		if (req.delegationAgentId) {
+			if (req.budget !== undefined) {
+				throw new Error(
+					"dispatch: budget envelopes cannot be enforced on an ACP delegation target; dispatch to a native or claude-sdk worker",
+				);
+			}
 			const protectedArtifactState = getProtectedArtifactState();
 			assertProtectedArtifactsEnforceable("acp-delegation", false, protectedArtifactState);
 			if (req.responseSchema !== undefined) {
@@ -5369,6 +5414,7 @@ export function createDispatchBundle(
 		);
 		assertPostRuntimeToolCompatibility(req.agentId, agentSpec, effectiveTools, target);
 		assertRuntimeCanHonorWorkerPermissionMode(target.runtime, settings?.workers.onPermission ?? "deny");
+		assertWorkerBudgetEnforceable(target.runtime, agentSpec.budget !== null || req.budget !== undefined);
 		assertResponseSchemaEnforceable(target.runtime, target.modelCapabilities, req.responseSchema, effectiveTools.length);
 		assertWriteRootsEnforceable(target.runtime, req.writeRoots);
 		assertProtectedArtifactsEnforceable(
@@ -5376,6 +5422,13 @@ export function createDispatchBundle(
 			target.runtime.kind !== "subprocess",
 			getProtectedArtifactState(),
 		);
+		resolveEffectiveWorkerBudget({
+			req,
+			recipeId: recipe.id,
+			declared: agentSpec.budget,
+			allowedTools: effectiveTools,
+			settings,
+		});
 		return {
 			agentId: req.agentId,
 			specFingerprint: agentSpecFingerprint(agentSpec),
@@ -5773,6 +5826,7 @@ export function createDispatchBundle(
 				runId: run.runId,
 				agentId: run.agentId,
 				task: run.task,
+				...(run.budget !== undefined ? { budget: run.budget } : {}),
 				runtimeKind: run.runtimeKind,
 				outcomePhase: run.stallKilled ? "terminating" : run.aborted ? "aborting" : "running",
 				heartbeat,

package/src/domains/dispatch/orphan-recovery.ts

@@ -55,6 +55,7 @@ function envelopeFromReceipt(receipt: RunReceipt, status: RunStatus, receiptPath
 		...(receipt.agentAudience !== undefined ? { agentAudience: receipt.agentAudience } : {}),
 		...(receipt.requestOrigin !== undefined ? { requestOrigin: receipt.requestOrigin } : {}),
 		task: receipt.task,
+		...(receipt.budget !== undefined ? { budget: receipt.budget } : {}),
 		...(receipt.briefing !== undefined ? { briefing: receipt.briefing } : {}),
 		...(receipt.steering !== undefined ? { steering: receipt.steering } : {}),
 		targetId: receipt.targetId,

package/src/domains/dispatch/receipt-integrity.ts

@@ -76,6 +76,8 @@ export const RECEIPT_INTEGRITY_FIELD_COVERAGE = {
 	agentAudience: true,
 	requestOrigin: true,
 	task: true,
+	// Historical receipts omit this field, so their receipt and ledger payloads remain unchanged.
+	budget: true,
 	targetId: true,
 	wireModelId: true,
 	runtimeId: true,
@@ -179,6 +181,7 @@ function ledgerDigestFields(envelope: RunEnvelope): Record<string, unknown> {
 		agentId: envelope.agentId,
 		executionRole: envelope.executionRole,
 		task: envelope.task,
+		...(envelope.budget !== undefined ? { budget: envelope.budget } : {}),
 		targetId: envelope.targetId,
 		wireModelId: envelope.wireModelId,
 		runtimeId: envelope.runtimeId,
@@ -356,6 +359,7 @@ function firstLedgerMismatch(receipt: RunReceipt, envelope: RunEnvelope): string
 		if (!Object.is(receiptValue, ledgerValue)) return field;
 	}
 	if (canonicalJson(receipt.briefing ?? null) !== canonicalJson(envelope.briefing ?? null)) return "briefing";
+	if (canonicalJson(receipt.budget ?? null) !== canonicalJson(envelope.budget ?? null)) return "budget";
 	if (canonicalJson(receipt.steering ?? null) !== canonicalJson(envelope.steering ?? null)) return "steering";
 	return null;
 }

package/src/domains/dispatch/state.ts

@@ -175,6 +175,7 @@ export function openLedger(opts?: LedgerOptions): Ledger {
 				...(input.agentAudience !== undefined ? { agentAudience: input.agentAudience } : {}),
 				...(input.requestOrigin !== undefined ? { requestOrigin: input.requestOrigin } : {}),
 				task: input.task,
+				...(input.budget !== undefined ? { budget: structuredClone(input.budget) } : {}),
 				...(input.briefing !== undefined ? { briefing: structuredClone(input.briefing) } : {}),
 				...(input.steering !== undefined ? { steering: structuredClone(input.steering) } : {}),
 				targetId: input.targetId,

package/src/domains/dispatch/types.ts

@@ -6,11 +6,13 @@
  * are pure data: no class methods, no engine refs.
  */
 
+import type { ResponseModelIdObservation } from "../../core/response-model-id.js";
 import type { SkillActivation } from "../../core/skill-activation.js";
 import type { ToolProfileName } from "../../tools/profiles.js";
 import type { AgentAudience } from "../agents/spec.js";
 import type { EvidenceTag } from "../evidence/index.js";
 import type { CostProvenance, RuntimeTargetSnapshot } from "../providers/index.js";
+import type { RunToolBudgetEnvelope } from "./budget-envelope.js";
 import type { ExecutionRole, GateTopologyRole } from "./execution-role.js";
 import type { RouteDecisionV1 } from "./route-decision.js";
 import type { RoutingIntent } from "./routing-intent.js";
@@ -418,6 +420,8 @@ export interface RunEnvelope {
 	agentAudience?: AgentAudience;
 	requestOrigin?: DispatchRequestOrigin;
 	task: string;
+	/** Immutable recipe policy, invocation request, effective phase, and admission reasons. */
+	budget?: RunToolBudgetEnvelope;
 	/** Present only when a bounded parent briefing was sent as dynamic task data. */
 	briefing?: RunBriefingProvenance;
 	/** Sent steering provenance in stable per-run sequence order; prose is never persisted. */
@@ -615,9 +619,10 @@ export interface RunReceiptReproducibility {
 }
 
 export interface RunReceiptUpstreamResponse {
-	model: string | null;
-	responseModel: string | null;
-	responseId: string | null;
+	requestedModelId: string | null;
+	responseModelIdObservation: ResponseModelIdObservation;
+	differingResponseModelId: string | null;
+	providerResponseId: string | null;
 }
 
 /**
@@ -677,6 +682,8 @@ export interface RunReceipt {
 	agentAudience?: AgentAudience;
 	requestOrigin?: DispatchRequestOrigin;
 	task: string;
+	/** Integrity-sealed recipe policy, invocation request, effective phase, and admission reasons. */
+	budget?: RunToolBudgetEnvelope;
 	/** Proof of briefing content without copying its prose into the receipt. */
 	briefing?: RunBriefingProvenance;
 	/** Sent steering provenance in stable per-run sequence order; prose is never persisted. */

package/src/domains/dispatch/validation.ts

@@ -11,6 +11,7 @@ import { cloneValidatedResponseSchema } from "../../core/response-schema.js";
 import { isToolProfileName, type ToolProfileName } from "../../tools/profiles.js";
 import type { AgentProduct } from "../agents/spec.js";
 import { type AutonomyLevel, isAutonomyLevel } from "../safety/autonomy.js";
+import { cloneDispatchBudgetRequest, type DispatchBudgetRequest } from "./budget-envelope.js";
 import {
 	EXECUTION_HANDOFF_MAX_ITEMS,
 	EXECUTION_HANDOFF_MAX_TEXT_BYTES,
@@ -76,6 +77,8 @@ export interface JobSpec {
 	target?: string;
 	model?: string;
 	thinkingLevel?: JobThinkingLevel;
+	/** Invocation phase and optional preauthorized retry or revision ceiling. */
+	budget?: DispatchBudgetRequest;
 	/**
 	 * What this run delivers, which decides what its reserve window admits.
 	 * "orientation" adds `code_nav` to the delivery tools kept live inside the
@@ -180,6 +183,7 @@ const KNOWN_KEYS = new Set([
 	"target",
 	"model",
 	"thinkingLevel",
+	"budget",
 	"node",
 	"routingIntent",
 	"failover",
@@ -378,6 +382,15 @@ export function validateJobSpec(spec: unknown): Validated {
 		}
 	}
 
+	let budget: DispatchBudgetRequest | undefined;
+	if ("budget" in spec && spec.budget !== undefined) {
+		try {
+			budget = cloneDispatchBudgetRequest(spec.budget);
+		} catch (error) {
+			errors.push(error instanceof Error ? error.message : String(error));
+		}
+	}
+
 	if ("requiredCapabilities" in spec && spec.requiredCapabilities !== undefined) {
 		if (!Array.isArray(spec.requiredCapabilities) || spec.requiredCapabilities.some((c) => typeof c !== "string")) {
 			errors.push("requiredCapabilities must be a string[]");
@@ -530,6 +543,7 @@ export function validateJobSpec(spec: unknown): Validated {
 		out.reroutes = spec.reroutes.map((hop) => ({ ...hop }));
 	}
 	if (typeof spec.thinkingLevel === "string") out.thinkingLevel = spec.thinkingLevel as JobThinkingLevel;
+	if (budget !== undefined) out.budget = budget;
 	if (Array.isArray(spec.requiredCapabilities)) {
 		out.requiredCapabilities = spec.requiredCapabilities.map((c) => String(c));
 	}

package/src/domains/dispatch/worker-spawn.ts

@@ -133,6 +133,16 @@ export interface WorkerProcessOptions {
 	approvedIdentity?: ApprovedWorkerIdentity;
 	/** How long bulk output may precede the announce before the peer is refused. */
 	attestationGraceMs?: number;
+	/**
+	 * Clock behind the heartbeat stamp, defaulting to `Date.now`. The stamp stays
+	 * a wall-clock instant in production because the ledger, the receipt, and the
+	 * evidence record all serialize it as ISO-8601 for an operator to read. That
+	 * leaves a test with no honest way to assert the stamp advanced: comparing
+	 * two wall-clock reads taken either side of an interval orders them backward
+	 * on any host that resyncs its clock in between. A test injects the steppable
+	 * clock from tests/harness/clock.ts here and asserts on that instead.
+	 */
+	now?: () => number;
 }
 
 /**
@@ -181,6 +191,7 @@ export function spawnWorkerProcess(
 ): SpawnedWorker {
 	const shutdownGraceMs = opts?.shutdownGraceMs ?? DEFAULT_SHUTDOWN_GRACE_MS;
 	const attestationGraceMs = opts?.attestationGraceMs ?? DEFAULT_ATTESTATION_GRACE_MS;
+	const now = opts?.now ?? Date.now;
 	const approved = opts?.approvedIdentity ?? approvedIdentityForSpec(spec);
 
 	const child: ChildProcess = spawn(command, [...args], {
@@ -193,7 +204,7 @@ export function spawnWorkerProcess(
 	});
 	const pid = child.pid ?? null;
 
-	const heartbeatAt = { current: Date.now() };
+	const heartbeatAt = { current: now() };
 
 	const queue = createBoundedEventQueue();
 	const waiters: Array<(r: IteratorResult<unknown>) => void> = [];
@@ -222,7 +233,7 @@ export function spawnWorkerProcess(
 	}
 
 	function push(value: unknown): void {
-		heartbeatAt.current = Date.now();
+		heartbeatAt.current = now();
 		const w = waiters.shift();
 		if (w) {
 			w({ value, done: false });
@@ -367,7 +378,7 @@ export function spawnWorkerProcess(
 			if (!announceAccepted) failAttestation(`Invalid worker attestation: ${frame.reason}`);
 			return;
 		}
-		heartbeatAt.current = Date.now();
+		heartbeatAt.current = now();
 		if (frame.value.kind === "announce") {
 			if (announceAccepted || announceFailed) return;
 			const verdict = verifyWorkerAttestation(frame.value.attestation, approved);

package/src/domains/eval/metrics/evidence.ts

@@ -64,119 +64,3 @@ export function evidenceMetricsFromReceipt(receipt: RunReceipt): Record<string,
 		...(typeof receipt.costUsd === "number" && Number.isFinite(receipt.costUsd) ? { "cost.usd": receipt.costUsd } : {}),
 	};
 }
-
-/**
- * Count terminal dispatch tool calls in a `clio-coder run --json` stream: the
- * decomposition checkpoint. Mirrors the dedup rules of the tool-call metric
- * fold (canonical clio finishes win over pi execution ends when both exist).
- * An absent or eventless stream counts zero.
- */
-export function dispatchCountFromJsonl(stdout: string): number {
-	let executionEnds = 0;
-	let canonicalFinishes = 0;
-	const seenExecutionEnds = new Set<string>();
-	const seenCanonicalFinishes = new Set<string>();
-	let sawCanonicalFinishEvent = false;
-	for (const event of jsonlEvents(stdout)) {
-		if (event.type === "tool_execution_end") {
-			if (stringField(event, "toolName") !== "dispatch") continue;
-			const callId = stringField(event, "toolCallId");
-			if (callId !== undefined) {
-				if (seenExecutionEnds.has(callId)) continue;
-				seenExecutionEnds.add(callId);
-			}
-			executionEnds += 1;
-			continue;
-		}
-		if (event.type !== "clio_tool_finish" || !isRecord(event.payload)) continue;
-		sawCanonicalFinishEvent = true;
-		if (stringField(event.payload, "tool") !== "dispatch") continue;
-		const callId = stringField(event.payload, "toolCallId") ?? stringField(event, "toolCallId");
-		if (callId !== undefined) {
-			if (seenCanonicalFinishes.has(callId)) continue;
-			seenCanonicalFinishes.add(callId);
-		}
-		canonicalFinishes += 1;
-	}
-	return sawCanonicalFinishEvent ? canonicalFinishes : executionEnds;
-}
-
-/** Count model-authored dispatch calls whose normalized task set includes Scout. */
-export function scoutDispatchCountFromJsonl(stdout: string): number {
-	const seen = new Set<string>();
-	let count = 0;
-	for (const event of jsonlEvents(stdout)) {
-		if (event.type !== "tool_execution_start" || stringField(event, "toolName") !== "dispatch") continue;
-		if (!isRecord(event.args) || !dispatchArgsIncludeAgent(event.args, "scout")) continue;
-		const callId = stringField(event, "toolCallId");
-		if (callId !== undefined) {
-			if (seen.has(callId)) continue;
-			seen.add(callId);
-		}
-		count += 1;
-	}
-	return count;
-}
-
-function dispatchArgsIncludeAgent(args: Record<string, unknown>, expectedAgent: string): boolean {
-	const sharedAgent = stringField(args, "agent") ?? stringField(args, "agent_id") ?? stringField(args, "agentId");
-	let tasks: unknown = args.tasks;
-	if (typeof tasks === "string") {
-		const trimmed = tasks.trim();
-		if (trimmed.startsWith("[") || trimmed.startsWith("{")) {
-			try {
-				tasks = JSON.parse(trimmed) as unknown;
-			} catch {
-				// A plain task string keeps using the shared agent below.
-			}
-		}
-	}
-	const taskList: unknown[] = Array.isArray(tasks) ? tasks : tasks === undefined ? [{ task: args.task }] : [tasks];
-	return taskList.some((task: unknown) => {
-		if (typeof task === "string") return sharedAgent?.toLowerCase() === expectedAgent;
-		if (!isRecord(task)) return false;
-		const agent =
-			stringField(task, "agent") ?? stringField(task, "agent_id") ?? stringField(task, "agentId") ?? sharedAgent;
-		return agent?.toLowerCase() === expectedAgent;
-	});
-}
-
-/**
- * Stale-wiki behavioral checkpoint: true when the stream shows a live source
- * read (read/grep start) AFTER a `code_nav mode=wiki` start, i.e. the model
- * did not answer from the wiki alone. Absent either half → false.
- */
-export function wikiStaleAcknowledgedFromJsonl(stdout: string): boolean {
-	let sawWikiLookup = false;
-	for (const event of jsonlEvents(stdout)) {
-		if (event.type !== "tool_execution_start") continue;
-		const toolName = stringField(event, "toolName");
-		if (toolName === "code_nav" && isRecord(event.args) && event.args.mode === "wiki") {
-			sawWikiLookup = true;
-			continue;
-		}
-		if (sawWikiLookup && (toolName === "read" || toolName === "grep")) return true;
-	}
-	return false;
-}
-
-function* jsonlEvents(stdout: string): Generator<Record<string, unknown>> {
-	for (const line of stdout.split(/\r?\n/)) {
-		if (line.trim().length === 0) continue;
-		try {
-			const parsed: unknown = JSON.parse(line);
-			if (isRecord(parsed)) yield parsed;
-		} catch {
-			// Non-JSON noise (including the pretty receipt tail) is not an event.
-		}
-	}
-}
-
-function stringField(record: Record<string, unknown>, field: string): string | undefined {
-	const value = record[field];
-	return typeof value === "string" && value.length > 0 ? value : undefined;
-}
-
-function isRecord(value: unknown): value is Record<string, unknown> {
-	return typeof value === "object" && value !== null && !Array.isArray(value);
-}

package/src/domains/eval/metrics/invariants.ts

@@ -2,7 +2,7 @@
  * Invariant metrics: the promises Clio makes about its own machinery, read
  * from the journal one eval item left behind.
  *
- * These invert what a benchmark measures. A weak model that never solved the
+ * These invert what a task outcome measures. A weak model that never solved the
  * task leaves every one of them intact; a strong model that solved it leaves
  * them broken the moment Clio failed to seal, to agree with itself, or to
  * write a receipt its own ledger can authenticate. The task outcome and the

package/src/domains/eval/runners/clio-run.ts

@@ -1,11 +1,5 @@
 import { shellQuote } from "../../../core/shell-quote.js";
-import {
-	dispatchCountFromJsonl,
-	evidenceMetricsFromReceipt,
-	receiptFromRunJsonStdout,
-	scoutDispatchCountFromJsonl,
-	wikiStaleAcknowledgedFromJsonl,
-} from "../metrics/evidence.js";
+import { evidenceMetricsFromReceipt, receiptFromRunJsonStdout } from "../metrics/evidence.js";
 import { streamInvariantMetrics } from "../metrics/invariants.js";
 import { tokenMetricEntries } from "../metrics/token-stream.js";
 import type { EvalRunnerV2, EvalSuiteTargetV2 } from "../schema/suite.js";
@@ -58,9 +52,6 @@ export async function runClioRunRunner(
 			"tools.failed": tools.failed,
 			"tools.blocked": tools.blocked,
 			"verifier.exitCode": result.exitCode,
-			"dispatch.count": dispatchCountFromJsonl(toolMetricStream),
-			"dispatch.scoutCount": scoutDispatchCountFromJsonl(toolMetricStream),
-			"wiki.staleAcknowledged": wikiStaleAcknowledgedFromJsonl(toolMetricStream),
 			...(receipt === null ? {} : evidenceMetricsFromReceipt(receipt)),
 			...(receipt === null
 				? {}