@iowarp/clio-coder 0.3.4 → 0.3.7

package/src/domains/dispatch/types.ts

@@ -6,12 +6,15 @@
  * 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 { DispatchIntent } from "./intent.js";
 import type { RouteDecisionV1 } from "./route-decision.js";
 import type { RoutingIntent } from "./routing-intent.js";
 
@@ -39,7 +42,8 @@ export type RunOutcomeCode =
 	| "worker_tool_call_cap_exhausted"
 	| "loop_guard_tools_disabled_exhausted"
 	| "result_contract_exhausted"
-	| "worker_final_output_missing";
+	| "worker_final_output_missing"
+	| "host_verification_rejected";
 
 export function isRunOutcomeCode(value: unknown): value is RunOutcomeCode {
 	return (
@@ -47,7 +51,8 @@ export function isRunOutcomeCode(value: unknown): value is RunOutcomeCode {
 		value === "worker_tool_call_cap_exhausted" ||
 		value === "loop_guard_tools_disabled_exhausted" ||
 		value === "result_contract_exhausted" ||
-		value === "worker_final_output_missing"
+		value === "worker_final_output_missing" ||
+		value === "host_verification_rejected"
 	);
 }
 
@@ -236,6 +241,13 @@ export interface RunGateProvenance {
 	verdict?: "pass" | "fail" | "revise";
 }
 
+export interface RunCouncilProvenance {
+	group: string;
+	label: string;
+	color?: string;
+	round: number;
+}
+
 /**
  * Plan-approval provenance for multi-task, compete, or remote dispatch.
  * `approval: "operator"` records that a supervised autonomy level parked the
@@ -246,7 +258,7 @@ export interface RunGateProvenance {
  */
 export interface RunPlanProvenance {
 	hash: string;
-	topology: "parallel" | "sequential" | "pipeline" | "review" | "compete" | "detached" | "fleet";
+	topology: "parallel" | "sequential" | "pipeline" | "review" | "compete" | "council" | "detached" | "fleet";
 	taskCount: number;
 	approval: "operator" | "full-auto";
 	source: null | {
@@ -309,7 +321,7 @@ export interface RunPhaseDurations {
  * bumping one without the other is a compile error.
  */
 export interface RunReceiptIntegrity {
-	version: 15;
+	version: 19;
 	algorithm: "sha256";
 	digest: string;
 }
@@ -418,6 +430,8 @@ 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. */
@@ -445,6 +459,7 @@ export interface RunEnvelope {
 	pipeline?: RunPipelineProvenance;
 	/** Review/compete gate provenance; present only on runs of a gated dispatch. */
 	gate?: RunGateProvenance;
+	council?: RunCouncilProvenance;
 	/** Plan-approval provenance; present only on runs of an approval-gated plan. */
 	plan?: RunPlanProvenance;
 	/** Ad-hoc specialist provenance; present only when a persona override composed the stable prompt. */
@@ -615,9 +630,10 @@ 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;
 }
 
 /**
@@ -669,6 +685,25 @@ export interface RunReceiptVerification {
 		| "receipt-unavailable";
 }
 
+export interface RunHostVerificationCheck {
+	check: string;
+	argv: string[];
+	cwd: string;
+	exitCode: number;
+	durationMs: number;
+	memo: boolean;
+	outputTail: string;
+	artifactPath?: string;
+	/** Original evidence owner on a memo hit. */
+	evidenceRunId?: string;
+}
+
+export interface RunHostVerification {
+	status: "verified" | "rejected" | "skipped";
+	reason?: string;
+	checks: RunHostVerificationCheck[];
+}
+
 export interface RunReceipt {
 	runId: string;
 	agentId: string;
@@ -677,6 +712,10 @@ export interface RunReceipt {
 	agentAudience?: AgentAudience;
 	requestOrigin?: DispatchRequestOrigin;
 	task: string;
+	/** Normalized dispatch intent admitted before worker execution. */
+	intent?: DispatchIntent;
+	/** 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. */
@@ -707,8 +746,11 @@ export interface RunReceipt {
 	pipeline?: RunPipelineProvenance;
 	/** Review/compete gate provenance; present only on runs of a gated dispatch. */
 	gate?: RunGateProvenance;
+	council?: RunCouncilProvenance;
 	/** Plan-approval provenance; present only on runs of an approval-gated plan. */
 	plan?: RunPlanProvenance;
+	/** Version 5 fleet gate artifact authored by this run. */
+	fleetGate?: { path: string; pathHash: string };
 	/** Ad-hoc specialist provenance; present only when a persona override composed the stable prompt. */
 	personaOverride?: RunPersonaOverride;
 	/** Effective project-context tier for this run; absent on receipts written before this field landed. */
@@ -757,6 +799,17 @@ export interface RunReceipt {
 	toolActivity?: ToolActivitySummary;
 	/** Integrity-sealed evidence confidence. Sealed on every receipt. */
 	verification: RunReceiptVerification;
+	/** Orchestrator-executed declared checks. Worker self-report never populates this field. */
+	hostVerification?: RunHostVerification;
+	/** Isolated writer checkout and its collect-time application result. */
+	worktree?: {
+		path: string;
+		branch: string;
+		diffHash: string;
+		apply: "merge" | "preserve";
+		applied: boolean;
+		reason?: string;
+	};
 	/** Required normalized routing request, sealed without task or prompt data. */
 	routingIntent: RoutingIntent;
 	/** Required routing-quality facts known at receipt finalization. */
@@ -779,8 +832,7 @@ export interface RunReceipt {
 	 * What this run contributed to its dispatch's agent ledger. Optional and
 	 * absent unless the run had a ledger, following validationGrounding and
 	 * capabilityMismatch: receiptDigestFields skips undefined, so a receipt
-	 * without one digests exactly as it did before this landed and
-	 * RUN_RECEIPT_INTEGRITY_VERSION stays where it is.
+	 * without one omits the field from canonical serialization.
 	 */
 	ledgerContribution?: RunLedgerContribution;
 	/**

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,
@@ -18,6 +19,7 @@ import {
 } from "./execution-handoff.js";
 import { isExecutionRole } from "./execution-role.js";
 import { COMPETE_STANCES, type CompeteStance } from "./gate-role-prompts.js";
+import { type DispatchIntent, isDispatchIntent } from "./intent.js";
 import { isRoutingIntent, type RoutingIntent } from "./routing-intent.js";
 import type {
 	DispatchRequestOrigin,
@@ -76,6 +78,8 @@ 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
@@ -118,6 +122,10 @@ export interface JobSpec {
 	 */
 	denyTools?: ReadonlyArray<string>;
 	cwd?: string;
+	/** Run this writer in an isolated git worktree. */
+	worktree?: true;
+	/** How a successful isolated writer is applied to its parent checkout. */
+	apply?: "merge" | "preserve";
 	memorySection?: string;
 	noSkills?: boolean;
 	skillPaths?: ReadonlyArray<string>;
@@ -131,6 +139,8 @@ export interface JobSpec {
 	 * per-tool calls (subprocess).
 	 */
 	writeRoots?: ReadonlyArray<string>;
+	/** Normalized model-declared path, output, and host verification intent. */
+	intent?: DispatchIntent;
 	requestOrigin?: DispatchRequestOrigin;
 	/**
 	 * Threaded output from the previous pipeline step. Set only by the dispatch
@@ -156,6 +166,8 @@ export interface JobSpec {
 	autonomy?: AutonomyLevel;
 	/** Review/compete gate provenance sealed into the run's receipt. */
 	gate?: RunGateProvenance;
+	/** Council grouping projected to receipts and fleet surfaces. */
+	council?: { group: string; label: string; color?: string; round: number };
 	/** Plan-approval provenance sealed into the run's receipt. */
 	plan?: RunPlanProvenance;
 	/**
@@ -180,6 +192,7 @@ const KNOWN_KEYS = new Set([
 	"target",
 	"model",
 	"thinkingLevel",
+	"budget",
 	"node",
 	"routingIntent",
 	"failover",
@@ -191,18 +204,22 @@ const KNOWN_KEYS = new Set([
 	"toolProfile",
 	"denyTools",
 	"cwd",
+	"worktree",
+	"apply",
 	"memorySection",
 	"noSkills",
 	"skillPaths",
 	"trustProjectCompatRoots",
 	"responseSchema",
 	"writeRoots",
+	"intent",
 	"requestOrigin",
 	"pipelineInput",
 	"predecessorHandoffs",
 	"lineage",
 	"autonomy",
 	"gate",
+	"council",
 	"plan",
 	"competeStance",
 ]);
@@ -378,6 +395,15 @@ 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[]");
@@ -386,7 +412,7 @@ export function validateJobSpec(spec: unknown): Validated {
 
 	if ("toolProfile" in spec && spec.toolProfile !== undefined) {
 		if (typeof spec.toolProfile !== "string" || !isToolProfileName(spec.toolProfile)) {
-			errors.push("toolProfile must be one of: minimal-local|science-local|full-agent");
+			errors.push("toolProfile must be one of: minimal-local|science-local|full-agent|council-read-only");
 		}
 	}
 
@@ -401,6 +427,13 @@ export function validateJobSpec(spec: unknown): Validated {
 			errors.push("cwd must be a non-empty string");
 		}
 	}
+	if ("worktree" in spec && spec.worktree !== undefined && spec.worktree !== true) {
+		errors.push("worktree must be true when present");
+	}
+	if ("apply" in spec && spec.apply !== undefined && spec.apply !== "merge" && spec.apply !== "preserve") {
+		errors.push("apply must be merge or preserve");
+	}
+	if (spec.apply !== undefined && spec.worktree !== true) errors.push("apply requires worktree: true");
 
 	if ("memorySection" in spec && spec.memorySection !== undefined) {
 		if (typeof spec.memorySection !== "string") {
@@ -444,6 +477,15 @@ export function validateJobSpec(spec: unknown): Validated {
 			errors.push("writeRoots must be a non-empty array of non-empty strings");
 		}
 	}
+	if ("intent" in spec && spec.intent !== undefined && !isDispatchIntent(spec.intent)) {
+		errors.push("intent must be a normalized dispatch intent");
+	}
+	if (isDispatchIntent(spec.intent) && spec.intent.writeRoots.length > 0 && Array.isArray(spec.writeRoots)) {
+		const jobCwd = typeof spec.cwd === "string" && spec.cwd.length > 0 ? spec.cwd : process.cwd();
+		const legacy = [...new Set(spec.writeRoots.map((root) => path.resolve(jobCwd, String(root))))].sort();
+		const declared = [...new Set(spec.intent.writeRoots.map((root) => path.resolve(jobCwd, root)))].sort();
+		if (JSON.stringify(legacy) !== JSON.stringify(declared)) errors.push("intent_write_roots_contradiction");
+	}
 
 	if ("requestOrigin" in spec && spec.requestOrigin !== undefined) {
 		if (typeof spec.requestOrigin !== "string" || !VALID_REQUEST_ORIGINS.has(spec.requestOrigin)) {
@@ -530,24 +572,32 @@ 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));
 	}
 	if (typeof spec.toolProfile === "string" && isToolProfileName(spec.toolProfile)) out.toolProfile = spec.toolProfile;
 	if (Array.isArray(spec.denyTools)) out.denyTools = spec.denyTools.map((tool) => String(tool));
 	if (typeof spec.cwd === "string") out.cwd = spec.cwd;
+	if (spec.worktree === true) out.worktree = true;
+	if (spec.apply === "merge" || spec.apply === "preserve") out.apply = spec.apply;
 	if (typeof spec.memorySection === "string") out.memorySection = spec.memorySection;
 	if (typeof spec.noSkills === "boolean") out.noSkills = spec.noSkills;
 	if (Array.isArray(spec.skillPaths)) out.skillPaths = spec.skillPaths.map((p) => String(p));
 	if (typeof spec.trustProjectCompatRoots === "boolean") out.trustProjectCompatRoots = spec.trustProjectCompatRoots;
 	if (responseSchema) out.responseSchema = responseSchema;
-	if (
+	if (isDispatchIntent(spec.intent)) out.intent = structuredClone(spec.intent);
+	const effectiveWriteRoots =
 		Array.isArray(spec.writeRoots) &&
 		spec.writeRoots.length > 0 &&
 		spec.writeRoots.every((root) => typeof root === "string" && root.length > 0)
-	) {
+			? spec.writeRoots
+			: isDispatchIntent(spec.intent) && spec.intent.writeRoots.length > 0
+				? spec.intent.writeRoots
+				: undefined;
+	if (effectiveWriteRoots !== undefined) {
 		const jobCwd = typeof spec.cwd === "string" && spec.cwd.length > 0 ? spec.cwd : process.cwd();
-		out.writeRoots = spec.writeRoots.map((root) => path.resolve(jobCwd, String(root)));
+		out.writeRoots = effectiveWriteRoots.map((root) => path.resolve(jobCwd, String(root)));
 	}
 	if (typeof spec.requestOrigin === "string" && VALID_REQUEST_ORIGINS.has(spec.requestOrigin)) {
 		out.requestOrigin = spec.requestOrigin as DispatchRequestOrigin;
@@ -558,6 +608,7 @@ export function validateJobSpec(spec: unknown): Validated {
 	if (isValidLineage(spec.lineage)) out.lineage = spec.lineage;
 	if (isAutonomyLevel(spec.autonomy)) out.autonomy = spec.autonomy;
 	if (isValidGate(spec.gate)) out.gate = cloneGate(spec.gate);
+	if (isValidCouncil(spec.council)) out.council = { ...spec.council };
 	if (typeof spec.competeStance === "string" && VALID_COMPETE_STANCES.has(spec.competeStance)) {
 		out.competeStance = spec.competeStance as CompeteStance;
 	}
@@ -640,9 +691,18 @@ function isValidProtectedArtifactRemap(value: unknown): value is ProtectedArtifa
 	);
 }
 
-const VALID_GATE_ROLES = new Set(["builder", "reviewer", "candidate", "judge"]);
+const VALID_GATE_ROLES = new Set(["builder", "reviewer", "candidate", "judge", "member", "synthesis"]);
 const VALID_GATE_VERDICTS = new Set(["pass", "fail", "revise"]);
-const VALID_PLAN_TOPOLOGIES = new Set(["parallel", "sequential", "pipeline", "review", "compete", "detached", "fleet"]);
+const VALID_PLAN_TOPOLOGIES = new Set([
+	"parallel",
+	"sequential",
+	"pipeline",
+	"review",
+	"compete",
+	"council",
+	"detached",
+	"fleet",
+]);
 const VALID_PLAN_APPROVALS = new Set(["operator", "full-auto"]);
 
 function isValidGate(value: unknown): value is RunGateProvenance {
@@ -674,6 +734,20 @@ function cloneGate(gate: RunGateProvenance): RunGateProvenance {
 	};
 }
 
+function isValidCouncil(value: unknown): value is NonNullable<JobSpec["council"]> {
+	return (
+		isPlainObject(value) &&
+		typeof value.group === "string" &&
+		value.group.length > 0 &&
+		typeof value.label === "string" &&
+		value.label.length > 0 &&
+		(value.color === undefined || typeof value.color === "string") &&
+		typeof value.round === "number" &&
+		Number.isInteger(value.round) &&
+		value.round >= 1
+	);
+}
+
 function isValidPlan(value: unknown): value is RunPlanProvenance {
 	if (!isPlainObject(value)) return false;
 	if (typeof value.hash !== "string" || value.hash.length === 0) return false;

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
 				? {}