@iowarp/clio-coder 0.3.3 → 0.3.6

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
 				? {}

package/src/domains/eval/runners/external-command.ts

@@ -1,13 +1,5 @@
 import { spawn } from "node:child_process";
 import { performance } from "node:perf_hooks";
-import {
-	addChaosObservations,
-	chaosMetricEntries,
-	createChaosFold,
-	EMPTY_CHAOS_OBSERVATION,
-	type EvalChaosObservation,
-	parseChaosMarker,
-} from "../metrics/chaos-stream.js";
 import {
 	addFleetLoopObservations,
 	createFleetLoopFold,
@@ -53,8 +45,6 @@ export interface ShellCommandResult {
 	usage: EvalTokenStreamUsage;
 	/** Wire-stream structural invariants folded live, for the same reason. */
 	streamInvariants: EvalStreamInvariants;
-	/** Strict SIGINT harness marker folded live from stdout. */
-	chaos: EvalChaosObservation;
 	fleetLoops: EvalFleetLoopObservation;
 	stderr: string;
 	wallTimeMs: number;
@@ -84,7 +74,6 @@ export async function runExternalCommandRunner(
 	// is reported as unmeasured, never as zero cost.
 	let usage = UNMEASURED_TOKEN_USAGE;
 	let streamInvariants = EMPTY_STREAM_INVARIANTS;
-	let chaos = EMPTY_CHAOS_OBSERVATION;
 	let fleetLoops = EMPTY_FLEET_LOOP_OBSERVATION;
 	for (const command of commands) {
 		const result = await runShellCommand(command, cwd, runner.timeoutMs ?? timeoutMs, env);
@@ -93,7 +82,6 @@ export async function runExternalCommandRunner(
 		wallTimeMs += result.wallTimeMs;
 		usage = addTokenStreamUsage(usage, result.usage);
 		streamInvariants = addStreamInvariants(streamInvariants, result.streamInvariants);
-		chaos = addChaosObservations(chaos, result.chaos);
 		fleetLoops = addFleetLoopObservations(fleetLoops, result.fleetLoops);
 		if (result.exitCode !== 0) {
 			return {
@@ -107,11 +95,10 @@ export async function runExternalCommandRunner(
 					"latency.wallMs": wallTimeMs,
 					...tokenMetricEntries(usage),
 					...streamInvariantMetrics(streamInvariants),
-					...chaosMetricEntries(chaos),
 					...fleetLoopMetricEntries(fleetLoops),
 					"verifier.exitCode": result.exitCode,
 				},
-				artifacts: chaosArtifacts(chaos),
+				artifacts: {},
 			};
 		}
 	}
@@ -126,11 +113,10 @@ export async function runExternalCommandRunner(
 			"latency.wallMs": wallTimeMs,
 			...tokenMetricEntries(usage),
 			...streamInvariantMetrics(streamInvariants),
-			...chaosMetricEntries(chaos),
 			...fleetLoopMetricEntries(fleetLoops),
 			"verifier.exitCode": 0,
 		},
-		artifacts: chaosArtifacts(chaos),
+		artifacts: {},
 	};
 }
 
@@ -150,7 +136,6 @@ export function runShellCommand(
 		const metricCapture = createJsonlMetricCapture();
 		const usageFold = createTokenUsageFold();
 		const streamFold = createStreamInvariantFold();
-		const chaosFold = createChaosFold();
 		const fleetLoopFold = createFleetLoopFold();
 		let timedOut = false;
 		let settled = false;
@@ -169,7 +154,6 @@ export function runShellCommand(
 			metricCapture.push(chunk);
 			usageFold.push(chunk);
 			streamFold.push(chunk);
-			chaosFold.push(chunk);
 			fleetLoopFold.push(chunk);
 		});
 		child.stderr.on("data", (chunk: string) => {
@@ -191,7 +175,6 @@ export function runShellCommand(
 				metricJsonl: metricCapture.finish(),
 				usage: usageFold.usage(),
 				streamInvariants: streamFold.invariants(),
-				chaos: chaosFold.observation(),
 				fleetLoops: fleetLoopFold.observation(),
 				stderr,
 				wallTimeMs: Math.round(performance.now() - started),
@@ -290,10 +273,6 @@ export function createJsonlMetricCapture(): JsonlMetricCapture {
 
 function compactMetricEvent(event: Record<string, unknown>): Record<string, unknown> | null {
 	const type = event.type;
-	if (type === "clio_soak_chaos") {
-		const marker = parseChaosMarker(event);
-		return marker === null ? null : { type, ...marker };
-	}
 	if (type === "tool_execution_start") {
 		const toolName = stringField(event, "toolName");
 		if (toolName !== "dispatch" && toolName !== "code_nav" && toolName !== "read" && toolName !== "grep") {
@@ -334,12 +313,6 @@ function compactMetricEvent(event: Record<string, unknown>): Record<string, unkn
 	};
 }
 
-function chaosArtifacts(observation: EvalChaosObservation): Record<string, string> {
-	return observation.markerCount === 1 && observation.marker !== null
-		? { chaosSeed: String(observation.marker.seed) }
-		: {};
-}
-
 function isRecord(value: unknown): value is Record<string, unknown> {
 	return typeof value === "object" && value !== null && !Array.isArray(value);
 }

package/src/domains/eval/schema/suite.ts

@@ -43,13 +43,6 @@ export const CANONICAL_METRICS = [
 	"evidence.verification",
 	"evidence.firstPassSuccess",
 	"cost.usd",
-	// Behavioral checkpoints for the bounded reconnaissance live suite.
-	"dispatch.count",
-	"wiki.staleAcknowledged",
-	// SIGINT chaos-harness facts folded live from its one strict stdout marker.
-	"chaos.seed",
-	"chaos.faultInjected",
-	"chaos.exitCode",
 	// Invariant checks: what Clio promised about its own machinery, read from
 	// the journal the item's run left behind. These are true or false about
 	// Clio and say nothing about whether the model solved the task. A check

package/src/domains/eval/suites/run.ts

@@ -126,7 +126,7 @@ async function runMatrixItem(
 			CLIO_CODER_ENTRY: clioEntry,
 		});
 		const patch = collectPatchMetrics(workspace.dir);
-		const receiptExitCode = receiptProcessExitCode(runner);
+		const receiptExitCode = runner.exitCode;
 		// Read after the runner returned and before the journal is removed: what
 		// Clio sealed for this item, judged against its own ledger, and whether
 		// the workers it attested are still running.
@@ -193,12 +193,6 @@ async function runMatrixItem(
 	}
 }
 
-/** A SIGINT harness exits cleanly after checking the nested Clio process; reconcile its receipt to that nested exit. */
-export function receiptProcessExitCode(runner: Pick<EvalRunnerOutput, "exitCode" | "metrics">): number {
-	const chaosExitCode = runner.metrics["chaos.exitCode"];
-	return typeof chaosExitCode === "number" && Number.isInteger(chaosExitCode) ? chaosExitCode : runner.exitCode;
-}
-
 /** Journal-derived invariants for one finished item, read from its isolated state directory. */
 function invariantMetrics(stateDir: string, runnerExitCode: number): Record<string, number | boolean> {
 	const journal = readRunJournal(stateDir);