Add serialized two-lane chat campaign coordinator

This commit is contained in:
Dotta committed 2026-10-27 04:42:56 -05:00
1 parent d476ef75bf
commit 49e06605d8
2 files changed
+535

No files matched your search

@@ -0,0 +1,255 @@
import { createHash } from "node:crypto";
import { mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync, existsSync, unlinkSync } from "node:fs";
import { tmpdir } from "node:os";
import path from "node:path";
import { afterEach, expect, it } from "vitest";
import { ChatCampaignCoordinator, type CampaignCoordinatorOptions, type CampaignLaunchPermit } from "./chat-campaign-coordinator.js";
import type { CampaignAttempt, CampaignScenario, ChatCampaign } from "./chat-campaign.js";
const roots: string[] = [];
afterEach(() => { for (const root of roots.splice(0)) rmSync(root, { recursive: true, force: true }); });
function fixture(fault?: CampaignCoordinatorOptions["fault"]) {
const root = mkdtempSync(path.join(tmpdir(), "chat-two-lane-")); roots.push(root);
const baseline = "a".repeat(40), startedAt = "2026-10-27T10:00:00.000Z";
const selections = ["fixture.first", "fixture.second", "fixture.third", "fixture.fourth"].map(scenarioId => {
const attempt: CampaignAttempt = {
ordinal: 1, outcome: "running", executionId: `chat-features.codex.local.${scenarioId}`, startedAt, finishedAt: null,
sourceRevision: baseline, workingTreeDigest: "bytes", definitionDigest: `definition-${scenarioId}`, suiteDefinitionDigest: "suite",
graderDigest: `grader-${scenarioId}`, model: "synthetic", profile: "synthetic", environment: "local", automaticRetries: 0,
correction: null, classification: null, observedFailure: null, evidencePaths: ["tests/runner-e2e/results/synthetic/selection.json"],
usage: { costUsd: null, tokens: null, coverage: "not run" }, cleanup: "pending", corePrDependencies: [], databaseLane: { lane: "embedded-postgres" },
productIdentity: { kind: "source-build", productionSourceRevision: baseline, snapshotSourceRevision: baseline,
nativeBinarySha256: "b".repeat(64), workspaceBuildDigest: "c".repeat(64), receiptPaths: ["tests/runner-e2e/results/synthetic/build.json"], coreCandidate: false },
};
return { scenarioId, attempt };
});
const scenarios: CampaignScenario[] = selections.map(({ scenarioId, attempt }) => ({
id: scenarioId, feature: "fixture", sourceFile: "fixture.yaml", sourceSha256: "source", scenarioSha256: "scenario", capabilities: [],
context: "synthetic only", query: "synthetic", expected_outcome: "synthetic", owner: "fixture", readiness: "implemented", authoringStatus: "complete",
qualifyingDefinitionDigest: attempt.definitionDigest, qualifyingGraderDigest: attempt.graderDigest, executionId: attempt.executionId, blocker: null,
productPolicy: { lane: "source-build", sourceBaselineRevision: baseline }, attempts: [],
}));
const manifest: ChatCampaign = { version: 1, campaignId: "fixture", frozenAt: startedAt, baselineRevision: baseline, inventoryDigest: "inventory",
defaultsSha256: "defaults", denominator: 4, areas: 1, capabilities: 0, requiredPasses: 4, corePrs: [], evalPr: null, scenarios };
const manifestPath = path.join(root, "manifest.json"); writeFileSync(manifestPath, JSON.stringify(manifest));
const options: CampaignCoordinatorOptions = { manifestPath, controlDirectory: path.join(root, "control"), batchId: "synthetic-batch", selections, fault, evidenceRoot: root, now: () => startedAt };
return { root, options, ledger: () => JSON.parse(readFileSync(manifestPath, "utf8")) as ChatCampaign, open: () => ChatCampaignCoordinator.open(options) };
}
function completion() {
return { outcome: "failed" as const, finishedAt: "2026-10-27T10:00:01.000Z", definitionDigest: "", graderDigest: "",
evidencePaths: ["tests/runner-e2e/results/synthetic/result.json"], cleanup: "passed" as const,
usage: { costUsd: null, tokens: 123, coverage: "synthetic test receipt; no model" }, classification: "synthetic_behavior_failure",
observedFailure: "Synthetic substantive failure", corePrDependencies: [] };
}
function result(f: ReturnType<typeof fixture>, p: CampaignLaunchPermit) {
const plan = f.options.selections.find(x => x.scenarioId === p.scenarioId)!;
return { ...completion(), definitionDigest: plan.attempt.definitionDigest, graderDigest: plan.attempt.graderDigest };
}
const retainFailure = async (_permit: unknown, error: unknown) => ({ evidencePath: "tests/runner-e2e/results/synthetic/failure.json", summary: error instanceof Error ? error.message : "Synthetic failure" });
function deferred() { let release!: () => void; const promise = new Promise<void>(resolve => { release = resolve; }); return { promise, release }; }
it("serializes racing reservation/completion writes, preserves both rows and exactly matching receipts", async () => {
const f = fixture(), c = f.open();
const [a, b, duplicate] = await Promise.all([c.reserve("fixture.first"), c.reserve("fixture.second"), c.reserve("fixture.first")]);
expect(a).toEqual(duplicate); expect(a.reservationId).not.toBe(b.reservationId);
expect(f.ledger().scenarios.slice(0, 2).map(x => x.attempts.length)).toEqual([1, 1]);
await expect(c.reserve("fixture.third")).rejects.toThrow("Two lanes");
let launches = 0;
await Promise.all([a, b].map(p => c.launchOnce(p, async permit => {
launches++; expect(permit.args).toEqual(["--id", p.executionId, "--max-parallel", "1", "--max-automatic-retries", "0"]);
})));
await expect(c.launchOnce(duplicate, async () => { launches++; })).rejects.toThrow("not admissible");
await Promise.all([c.complete(a, result(f, a)), c.complete(b, result(f, b)), c.complete(a, result(f, a))]);
expect(launches).toBe(2);
expect(f.ledger().scenarios.slice(0, 2).map(x => x.attempts[0]?.usage.tokens)).toEqual([123, 123]);
await expect(c.complete(a, { ...result(f, a), observedFailure: "wrong bytes" })).rejects.toThrow("content conflict");
await c.close();
});
it("rejects a second owner even with a different batch directory and never steals stale PID locks", async () => {
const f = fixture(), c = f.open();
expect(() => ChatCampaignCoordinator.open({ ...f.options, controlDirectory: path.join(f.root, "other") })).toThrow();
const filename = `${f.options.manifestPath}.coordinator-lock`, lock = JSON.parse(readFileSync(filename, "utf8"));
writeFileSync(filename, JSON.stringify({ ...lock, pid: 2147483647, createdAt: "1900-01-01" }));
expect(() => f.open()).toThrow();
// Restoring this test's exact owner does not model authorizing recovery of real processes.
writeFileSync(filename, JSON.stringify(lock)); await c.close();
});
it.each(["executionId", "definitionDigest", "graderDigest", "ordinal", "productIdentity", "databaseLane"] as const)("rejects changed reviewed %s before admission", async field => {
const f = fixture();
const manifest = f.ledger();
if (field === "definitionDigest") manifest.scenarios[0]!.qualifyingDefinitionDigest = "changed";
else if (field === "graderDigest") manifest.scenarios[0]!.qualifyingGraderDigest = "changed";
else if (field === "executionId") manifest.scenarios[0]!.executionId = "foreign";
else if (field === "ordinal") f.options.selections[0]!.attempt.ordinal = 2;
else if (field === "productIdentity") f.options.selections[0]!.attempt.productIdentity = { ...f.options.selections[0]!.attempt.productIdentity!, kind: "source-build", productionSourceRevision: "f".repeat(40) } as CampaignAttempt["productIdentity"];
else f.options.selections[0]!.attempt.databaseLane = { lane: "owned-docker-postgres", engine: "foreign" } as unknown as CampaignAttempt["databaseLane"];
writeFileSync(f.options.manifestPath, JSON.stringify(manifest));
const c = f.open(); await expect(c.reserve("fixture.first")).rejects.toThrow();
expect(f.ledger().scenarios.every(x => !x.attempts.length)).toBe(true); await c.close();
});
it("uses fresh ledger reads and rejects changed reserved identity or tampered matching-completion ledger", async () => {
const f = fixture(), c = f.open(), p = await c.reserve("fixture.first");
await c.launchOnce(p, async () => undefined);
const m = f.ledger(); m.scenarios[1]!.blocker = "independent metadata preserved"; writeFileSync(f.options.manifestPath, JSON.stringify(m));
await c.complete(p, result(f, p));
expect(f.ledger().scenarios[1]!.blocker).toBe("independent metadata preserved");
const changed = f.ledger(); changed.scenarios[0]!.attempts[0]!.usage.tokens = 999; writeFileSync(f.options.manifestPath, JSON.stringify(changed));
await expect(c.complete(p, result(f, p))).rejects.toThrow("differs from receipt"); await c.close();
});
it.each(["intent", "manifest", "state"] as const)("fails closed at the real filesystem %s crash boundary without replaying a launch", async stage => {
const f = fixture(actual => { if (actual === stage) throw new Error("synthetic crash"); }), c = f.open();
await expect(c.reserve("fixture.first")).rejects.toThrow("synthetic crash");
expect(() => f.open()).toThrow(); // crash lock is retained
expect(() => JSON.parse(readFileSync(f.options.manifestPath, "utf8"))).not.toThrow();
// Synthetic owner-death reconciliation: no child was invoked in this test.
unlinkSync(`${f.options.manifestPath}.coordinator-lock`); f.options.fault = undefined;
const restarted = f.open(), state = await restarted.snapshot();
expect(state.stop?.reason).toBe("coordinator_failure");
await expect(restarted.reserve("fixture.second")).rejects.toThrow("stopped");
expect(f.ledger().scenarios[0]!.attempts.length).toBe(stage === "intent" ? 0 : 1);
await restarted.close();
});
it("retains launch uncertainty both before spawn and after a child returns, never automatically relaunches", async () => {
const f = fixture(), c = f.open(), p = await c.reserve("fixture.first");
let launches = 0;
await expect(c.launchOnce(p, async () => { launches++; throw new Error("spawn boundary uncertain"); })).rejects.toThrow("uncertain");
await expect(c.launchOnce(p, async () => { launches++; })).rejects.toThrow();
await c.close();
const restarted = f.open(); expect((await restarted.snapshot()).stop?.reason).toBe("coordinator_failure");
await expect(restarted.launchOnce(p, async () => { launches++; })).rejects.toThrow();
expect(launches).toBe(1); await restarted.close();
});
it("runs at most two independent children and drains the entire wave before stop-before-refill", async () => {
const f = fixture(), c = f.open(), gate = deferred(), secondStarted = deferred();
const launched: string[] = []; let active = 0, maximum = 0;
const run = c.run({
retainFailure, launchOnly: async p => { launched.push(p.scenarioId); maximum = Math.max(maximum, ++active);
if (p.scenarioId === "fixture.first") await secondStarted.promise;
if (p.scenarioId === "fixture.second") { secondStarted.release(); await gate.promise; } active--; return p.scenarioId; },
readCompletion: async (p, id) => ({ completion: result(f, p), ...(id === "fixture.first" ? { stop: { reason: "secret_leak" as const, detail: "actual synthetic secret match" } } : {}) }),
});
await secondStarted.promise;
await expect(c.close()).rejects.toThrow("in flight");
gate.release(); const state = await run;
expect(maximum).toBe(2); expect(launched.sort()).toEqual(["fixture.first", "fixture.second"]);
expect(Object.values(state.entries).every(x => x.phase === "completed")).toBe(true);
expect(state.stop?.reason).toBe("secret_leak"); expect(f.ledger().scenarios[2]!.attempts).toEqual([]);
await c.close(); const restarted = f.open(); await expect(restarted.reserve("fixture.third")).rejects.toThrow("stopped"); await restarted.close();
});
it("does not classify a caller-verified safe bound-only ZIP omission as a secret leak", async () => {
const f = fixture(), c = f.open(); let launches = 0;
const state = await c.run({ retainFailure, launchOnly: async () => { launches++; }, readCompletion: async p => ({ completion: { ...result(f, p), observedFailure: "Reporter privacy violation: independently verified safe bound-only ZIP omission, no secret match" } }) });
expect(launches).toBe(4); expect(state.stop).toBeNull(); await c.close();
});
it("durably stops on cleanup failure before completion commit or admission, including a crash", async () => {
let fail = false;
const f = fixture(stage => { if (fail && stage === "intent") throw new Error("completion crash"); });
const c = f.open(), p = await c.reserve("fixture.first"); await c.launchOnce(p, async () => undefined); fail = true;
await expect(c.complete(p, { ...result(f, p), cleanup: "failed" })).rejects.toThrow("completion crash");
expect(JSON.parse(readFileSync(path.join(f.options.controlDirectory, "state.json"), "utf8")).stop.reason).toBe("cleanup_failure");
expect(existsSync(path.join(f.options.controlDirectory, "pending.json"))).toBe(true);
});
it("does not relaunch completed entries when a clean parent resumes remaining selections", async () => {
const f = fixture(), c = f.open(), p = await c.reserve("fixture.first");
await c.launchOnce(p, async () => undefined); await c.complete(p, result(f, p)); await c.close();
const next = f.open(), launched: string[] = [];
await next.run({ retainFailure, launchOnly: async permit => { launched.push(permit.scenarioId); }, readCompletion: async permit => ({ completion: result(f, permit) }) });
expect(launched.sort()).toEqual(["fixture.fourth", "fixture.second", "fixture.third"]); await next.close();
});
it("rejects an outside writer during commit and retains the recovery lock instead of overwriting", async () => {
const f = fixture(stage => { if (stage === "intent") {
const manifest = f.ledger(); manifest.evalPr = "unexpected writer"; writeFileSync(f.options.manifestPath, JSON.stringify(manifest));
} });
const c = f.open(); await expect(c.reserve("fixture.first")).rejects.toThrow("outside coordinator");
expect(f.ledger().evalPr).toBe("unexpected writer"); expect(f.ledger().scenarios[0]!.attempts.length).toBe(0);
unlinkSync(`${f.options.manifestPath}.coordinator-lock`); f.options.fault = undefined;
expect(() => f.open()).toThrow("Unrecognized manifest");
expect(existsSync(`${f.options.manifestPath}.coordinator-lock`)).toBe(true);
});
it("rejects a forged permit and modified native reservation before accepting result evidence", async () => {
const f = fixture(), c = f.open(), p = await c.reserve("fixture.first");
await expect(c.launchOnce({ ...p, args: [...p.args, "--all"] }, async () => undefined)).rejects.toThrow("not admissible");
await c.launchOnce(p, async () => undefined);
const m = f.ledger(); m.scenarios[0]!.attempts[0]!.model = "wrong model"; writeFileSync(f.options.manifestPath, JSON.stringify(m));
await expect(c.complete(p, result(f, p))).rejects.toThrow("Reserved attempt changed"); await c.close();
});
it("persists a queued explicit stop before admitting the next queued reservation", async () => {
const f = fixture(), c = f.open();
const stopped = c.stop("shared_infrastructure", "synthetic real host failure");
await expect(c.reserve("fixture.first")).rejects.toThrow("stopped"); await stopped;
expect(f.ledger().scenarios.every(x => !x.attempts.length)).toBe(true); await c.close();
});
it("retains the original passing measurement while invalidating only qualification from a hash-bound closure review", async () => {
const f = fixture(), c = f.open(), p = await c.reserve("fixture.first"); await c.launchOnce(p, async () => undefined);
const original = f.options.selections[0]!.attempt;
const previous = { definitionDigest: original.definitionDigest, graderDigest: original.graderDigest };
const next = { definitionDigest: "new-normalized-selected-closure", graderDigest: "new-normalized-grader" };
const evidencePath = "tests/runner-e2e/results/synthetic/identity-review.json";
const filename = path.join(f.root, evidencePath); mkdirSync(path.dirname(filename), { recursive: true });
const bytes = JSON.stringify({ id: p.scenarioId, ordinal: p.ordinal, previous, next }); writeFileSync(filename, bytes);
const review = { evidencePath, receiptSha256: createHash("sha256").update(bytes).digest("hex"), previous, next };
const passed = { ...result(f, p), outcome: "passed" as const, evidencePaths: [evidencePath] };
await expect(c.complete(p, passed, undefined, { ...review, next: previous })).rejects.toThrow("Invalid qualifying");
await expect(c.complete(p, passed, undefined, { ...review, receiptSha256: "wrong" })).rejects.toThrow("receipt mismatch");
await c.complete(p, passed, undefined, review);
const row = f.ledger().scenarios[0]!;
expect(row.qualifyingDefinitionDigest).toBeNull(); expect(row.qualifyingGraderDigest).toBeNull();
expect(row.attempts[0]).toMatchObject({ outcome: "passed", definitionDigest: original.definitionDigest, graderDigest: original.graderDigest });
await c.close();
});
it("recovers a committed completion after interruption without duplicating the attempt or restarting any child", async () => {
let fault = false;
const f = fixture(stage => { if (fault && stage === "manifest") throw new Error("after completion commit"); });
const c = f.open(), p = await c.reserve("fixture.first"); await c.launchOnce(p, async () => undefined); fault = true;
await expect(c.complete(p, result(f, p))).rejects.toThrow("completion commit");
unlinkSync(`${f.options.manifestPath}.coordinator-lock`); f.options.fault = undefined;
const restarted = f.open(); await restarted.complete(p, result(f, p));
expect(f.ledger().scenarios[0]!.attempts).toHaveLength(1);
expect((await restarted.snapshot()).stop?.reason).toBe("coordinator_failure");
await expect(restarted.launchOnce(p, async () => { throw new Error("must not launch"); })).rejects.toThrow("not admissible");
await restarted.close();
});
it("stamps each real reservation time rather than preparation time, retaining exact later-wave identity", async () => {
const f = fixture(); let clock = "2026-10-27T10:00:00.100Z";
f.options.now = () => clock;
const c = f.open(), first = await c.reserve("fixture.first");
expect(f.ledger().scenarios[0]!.attempts[0]!.startedAt).toBe(clock);
await c.launchOnce(first, async () => undefined); await c.complete(first, result(f, first));
clock = "2026-10-27T10:15:00.000Z";
const later = await c.reserve("fixture.third");
expect(f.ledger().scenarios[2]!.attempts[0]!.startedAt).toBe(clock);
const entry = (await c.snapshot()).entries[later.reservationId]!;
expect(entry.reservedAttempt.startedAt).toBe(clock);
await c.launchOnce(later, async () => undefined);
await c.complete(later, { ...result(f, later), finishedAt: "2026-10-27T10:16:00.000Z" }); await c.close();
// Regenerated preparation timestamps are not a new reviewed selection identity.
for (const plan of f.options.selections) plan.attempt.startedAt = "2026-10-27T09:00:00.000Z";
const restarted = f.open(); expect((await restarted.snapshot()).entries[later.reservationId]!.reservedAttempt.startedAt).toBe(clock);
await restarted.close();
});
it("retains actionable sanitized completion failure evidence while draining the peer and stopping refill", async () => {
const f = fixture(), c = f.open(); const retained: string[] = [];
const state = await c.run({ launchOnly: async p => p.scenarioId,
readCompletion: async (p, id) => { if (id === "fixture.first") throw new Error("Synthetic result source digest differs from reserved digest"); return { completion: result(f, p) }; },
retainFailure: async (p, error, phase) => { retained.push(`${p?.scenarioId}:${phase}`); return retainFailure(p, error); },
});
expect(retained).toEqual(["fixture.first:completion"]);
expect(state.failures[0]).toMatchObject({ scenarioId: "fixture.first", phase: "completion", summary: "Synthetic result source digest differs from reserved digest" });
expect(state.stop?.detail).toContain("failure.json"); expect(f.ledger().scenarios[2]!.attempts).toEqual([]); await c.close();
});
@@ -0,0 +1,280 @@
import { createHash, randomUUID } from "node:crypto";
import { closeSync, existsSync, fsyncSync, mkdirSync, openSync, readFileSync, realpathSync, renameSync, unlinkSync, writeFileSync } from "node:fs";
import path from "node:path";
import { campaignProductIdentityEligible, completeCampaignAttempt, reserveCampaignAttempt, type CampaignAttempt, type ChatCampaign } from "./chat-campaign.js";
type Completion = Parameters<typeof completeCampaignAttempt>[1];
export type CoordinatorStopReason = "cleanup_failure" | "secret_leak" | "shared_infrastructure" | "coordinator_failure";
export interface CampaignQualificationReview {
evidencePath: string; receiptSha256: string;
previous: { definitionDigest: string; graderDigest: string };
next: { definitionDigest: string; graderDigest: string };
}
export interface ReviewedCampaignSelection { scenarioId: string; attempt: CampaignAttempt }
export interface CampaignLaunchPermit {
reservationId: string; scenarioId: string; ordinal: 1 | 2; executionId: string;
campaignId: string; args: readonly string[]; maxParallel: 1; automaticRetries: 0;
}
interface Entry {
permit: CampaignLaunchPermit; reservationDigest: string; reservedAttempt: CampaignAttempt;
phase: "reserved" | "launch_uncertain" | "completed";
completionDigest?: string;
}
interface State {
version: 1; selectionDigest: string; entries: Record<string, Entry>;
stop: { reason: CoordinatorStopReason; detail: string } | null;
failures: Array<{ scenarioId: string; phase: "reserve" | "launch" | "completion"; evidencePath: string; summary: string }>;
}
interface Pending { before: string; after: string; manifest: string; state: State }
export interface CampaignCoordinatorOptions {
manifestPath: string; controlDirectory: string; batchId: string;
/** Defaults to the repository root above tests/runner-e2e. */
evidenceRoot?: string;
selections: readonly ReviewedCampaignSelection[];
/** Controlled filesystem fault points for offline crash tests only. */
now?: () => string;
fault?: (stage: "intent" | "manifest" | "state") => void;
}
const hash = (value: string) => createHash("sha256").update(value).digest("hex");
const json = (value: unknown) => `${JSON.stringify(value, null, 2)}\n`;
function reviewedIdentity(selections: readonly ReviewedCampaignSelection[]) {
return selections.map(({ scenarioId, attempt }) => ({ scenarioId, attempt: { ...attempt, startedAt: null } }));
}
function read<T>(filename: string): T { return JSON.parse(readFileSync(filename, "utf8")); }
function atomicWrite(filename: string, bytes: string) {
const temporary = `${filename}.${randomUUID()}.tmp`;
const fd = openSync(temporary, "wx", 0o600);
try { writeFileSync(fd, bytes); fsyncSync(fd); } finally { closeSync(fd); }
renameSync(temporary, filename);
const directory = openSync(path.dirname(filename), "r");
try { fsyncSync(directory); } finally { closeSync(directory); }
}
/** One ledger owner, two independent one-case launchers. Never steals a stale lock.
* A crashed owner requires independent process/resource reconciliation before an
* operator removes its lock. Restart never replays a launch, even if no PID was saved.
* Reconciliation/manual ledger writers must remain paused for this owner's lifetime.
*/
export class ChatCampaignCoordinator {
private tail: Promise<unknown> = Promise.resolve();
private readonly plans: ReviewedCampaignSelection[];
private readonly statePath: string;
private readonly intentPath: string;
private readonly lockPath: string;
private readonly owner = randomUUID();
private poisoned = false;
private closed = false;
private inFlight = 0;
private constructor(private readonly options: CampaignCoordinatorOptions) {
this.plans = structuredClone([...options.selections]);
this.statePath = path.join(options.controlDirectory, "state.json");
this.intentPath = path.join(options.controlDirectory, "pending.json");
// Lock belongs to the manifest, not a caller-chosen batch directory.
this.lockPath = `${options.manifestPath}.coordinator-lock`;
}
static open(options: CampaignCoordinatorOptions) {
if (!/^[a-zA-Z0-9_-]{1,100}$/.test(options.batchId) || !options.selections.length
|| new Set(options.selections.map(x => x.scenarioId)).size !== options.selections.length)
throw new Error("Invalid reviewed campaign selection");
for (const { scenarioId, attempt } of options.selections) {
if (!/^[a-z0-9-]+\.[a-z0-9-]+$/.test(scenarioId) || attempt.environment !== "local"
|| attempt.automaticRetries !== 0 || attempt.outcome !== "running" || !attempt.productIdentity
|| attempt.productIdentity.kind !== "source-build" || !attempt.databaseLane)
throw new Error("Two-lane coordinator requires reviewed local source-build provenance");
}
mkdirSync(options.controlDirectory, { recursive: true, mode: 0o700 });
options = { ...options, manifestPath: realpathSync(options.manifestPath), controlDirectory: realpathSync(options.controlDirectory) };
const coordinator = new ChatCampaignCoordinator(options);
const fd = openSync(coordinator.lockPath, "wx", 0o600);
try { writeFileSync(fd, json({ version: 1, owner: coordinator.owner, pid: process.pid,
batchId: options.batchId, controlDirectory: path.resolve(options.controlDirectory), createdAt: new Date().toISOString() })); fsyncSync(fd); }
finally { closeSync(fd); }
try {
const selectionDigest = hash(json({ batchId: options.batchId, manifestPath: path.resolve(options.manifestPath), selections: reviewedIdentity(coordinator.plans) }));
if (!existsSync(coordinator.statePath)) atomicWrite(coordinator.statePath, json({ version: 1, selectionDigest, entries: {}, stop: null, failures: [] } satisfies State));
const state = read<State>(coordinator.statePath);
if (state.version !== 1 || state.selectionDigest !== selectionDigest) throw new Error("Coordinator selection changed");
coordinator.recover();
return coordinator;
} catch (error) {
// Retain ownership on recovery failure: no second writer may guess what happened.
coordinator.poisoned = true;
throw error;
}
}
private assertOwner() {
if (this.closed || this.poisoned || read<{ owner: string }>(this.lockPath).owner !== this.owner)
throw new Error("Coordinator ownership unavailable");
}
private serial<T>(operation: () => T): Promise<T> {
const result = this.tail.then(() => { this.assertOwner(); return operation(); });
this.tail = result.catch(() => undefined);
return result;
}
private recover() {
let state = read<State>(this.statePath);
if (existsSync(this.intentPath)) {
const pending = read<Pending>(this.intentPath), actual = hash(readFileSync(this.options.manifestPath, "utf8"));
if (actual !== pending.before && actual !== pending.after) throw new Error("Unrecognized manifest after interrupted transaction");
// Never replay an uncommitted reservation or completion on restart. If the
// manifest commit landed, recover its corresponding state, then stop.
if (actual === pending.after) state = pending.state;
state.stop ??= { reason: "coordinator_failure", detail: "Interrupted ledger transaction; independent reconciliation required" };
atomicWrite(this.statePath, json(state)); unlinkSync(this.intentPath);
}
if (Object.values(state.entries).some(x => x.phase !== "completed")) {
state.stop ??= { reason: "coordinator_failure", detail: "Reserved or possibly launched execution; no automatic recovery launch" };
atomicWrite(this.statePath, json(state));
}
}
private commit(manifestBefore: string, manifest: ChatCampaign, state: State) {
const bytes = json(manifest);
try {
atomicWrite(this.intentPath, json({ before: hash(manifestBefore), after: hash(bytes), manifest: bytes, state } satisfies Pending));
this.options.fault?.("intent");
if (hash(readFileSync(this.options.manifestPath, "utf8")) !== hash(manifestBefore)) throw new Error("Manifest changed outside coordinator");
atomicWrite(this.options.manifestPath, bytes); this.options.fault?.("manifest");
atomicWrite(this.statePath, json(state)); this.options.fault?.("state");
unlinkSync(this.intentPath);
} catch (error) { this.poisoned = true; throw error; }
}
reserve(scenarioId: string): Promise<CampaignLaunchPermit> {
return this.serial(() => {
const state = read<State>(this.statePath), plan = this.plans.find(x => x.scenarioId === scenarioId);
if (!plan) throw new Error("Scenario not reviewed");
if (state.stop) throw new Error("Admissions stopped");
const reservationId = hash(json({ batchId: this.options.batchId, scenarioId, ordinal: plan.attempt.ordinal }));
const reservationDigest = hash(json(reviewedIdentity([plan])));
if (state.entries[reservationId]) {
if (state.entries[reservationId]!.reservationDigest !== reservationDigest) throw new Error("Reservation identity conflict");
return structuredClone(state.entries[reservationId]!.permit);
}
if (Object.values(state.entries).filter(x => x.phase !== "completed").length >= 2) throw new Error("Two lanes already reserved");
const before = readFileSync(this.options.manifestPath, "utf8"), manifest = JSON.parse(before) as ChatCampaign;
const row = manifest.scenarios.find(x => x.id === scenarioId);
if (!row || row.executionId !== plan.attempt.executionId || row.qualifyingDefinitionDigest !== plan.attempt.definitionDigest
|| row.qualifyingGraderDigest !== plan.attempt.graderDigest
|| !campaignProductIdentityEligible(row, plan.attempt, manifest.baselineRevision)) throw new Error("Reviewed selection differs from authoritative ledger");
const reservedAttempt = { ...structuredClone(plan.attempt), startedAt: this.options.now?.() ?? new Date().toISOString() };
if (!Number.isFinite(Date.parse(reservedAttempt.startedAt))) throw new Error("Invalid reservation clock");
reserveCampaignAttempt(row, reservedAttempt);
const permit: CampaignLaunchPermit = { reservationId, scenarioId, ordinal: plan.attempt.ordinal, executionId: plan.attempt.executionId,
campaignId: `${this.options.batchId}-${reservationId.slice(0, 20)}`, maxParallel: 1, automaticRetries: 0,
args: ["--id", plan.attempt.executionId, "--max-parallel", "1", "--max-automatic-retries", "0"] };
state.entries[reservationId] = { permit, reservationDigest, reservedAttempt: structuredClone(reservedAttempt), phase: "reserved" };
this.commit(before, manifest, state);
return structuredClone(permit);
});
}
async launchOnce<T>(permit: CampaignLaunchPermit, launchOnly: (permit: CampaignLaunchPermit) => Promise<T>): Promise<T> {
await this.serial(() => {
const state = read<State>(this.statePath), entry = state.entries[permit.reservationId];
if (state.stop || !entry || entry.phase !== "reserved" || json(entry.permit) !== json(permit)) throw new Error("Launch is not admissible");
// Persist uncertainty BEFORE calling spawn. A crash in this gap consumes
// the permit; it can never be replayed based on absent PID alone.
entry.phase = "launch_uncertain"; atomicWrite(this.statePath, json(state)); this.inFlight++;
});
try { return await launchOnly(structuredClone(permit)); }
finally { this.inFlight--; }
}
complete(permit: CampaignLaunchPermit, completion: Completion, stop?: { reason: CoordinatorStopReason; detail: string }, qualificationReview?: CampaignQualificationReview): Promise<void> {
return this.serial(() => {
const state = read<State>(this.statePath), entry = state.entries[permit.reservationId];
if (!entry || json(entry.permit) !== json(permit)) throw new Error("Completion identity conflict");
const completionDigest = hash(json({ completion, stop: stop ?? null, qualificationReview: qualificationReview ?? null }));
if (entry.phase === "completed") {
if (entry.completionDigest !== completionDigest) throw new Error("Completion content conflict");
const current = read<ChatCampaign>(this.options.manifestPath).scenarios.find(x => x.id === permit.scenarioId)?.attempts.at(-1);
if (!current || current.ordinal !== permit.ordinal || current.executionId !== permit.executionId
|| Object.entries(completion).some(([key, value]) => json(current[key as keyof CampaignAttempt]) !== json(value)))
throw new Error("Completed ledger differs from receipt");
return;
}
if (entry.phase !== "launch_uncertain") throw new Error("No admitted launcher for completion");
const before = readFileSync(this.options.manifestPath, "utf8"), manifest = JSON.parse(before) as ChatCampaign;
const row = manifest.scenarios.find(x => x.id === permit.scenarioId), plan = this.plans.find(x => x.scenarioId === permit.scenarioId)!;
const actual = row?.attempts.at(-1);
if (!row || !actual || actual.ordinal !== permit.ordinal || json(actual) !== json(entry.reservedAttempt)) throw new Error("Reserved attempt changed");
completeCampaignAttempt(row, structuredClone(completion));
if (qualificationReview) {
const review = qualificationReview;
if (completion.outcome !== "passed" || !completion.evidencePaths.includes(review.evidencePath)
|| !review.evidencePath.startsWith("tests/runner-e2e/results/") || review.evidencePath.split("/").includes("..")
|| review.previous.definitionDigest !== plan.attempt.definitionDigest || review.previous.graderDigest !== plan.attempt.graderDigest
|| json(review.previous) === json(review.next) || !review.next.definitionDigest || !review.next.graderDigest)
throw new Error("Invalid qualifying identity review");
const bytes = readFileSync(path.join(this.options.evidenceRoot ?? path.resolve(path.dirname(this.options.manifestPath), "../.."), review.evidencePath), "utf8");
const retained = JSON.parse(bytes);
if (hash(bytes) !== review.receiptSha256 || retained.id !== permit.scenarioId || retained.ordinal !== permit.ordinal
|| json(retained.previous) !== json(review.previous) || json(retained.next) !== json(review.next))
throw new Error("Qualifying identity review receipt mismatch");
row.qualifyingDefinitionDigest = null; row.qualifyingGraderDigest = null;
}
state.stop ??= stop ?? (completion.cleanup !== "passed" ? { reason: "cleanup_failure", detail: "Measured cleanup did not pass" } : null);
// Stop is durable before the completion/summary commit or any lane refill.
if (state.stop) atomicWrite(this.statePath, json(state));
entry.phase = "completed"; entry.completionDigest = completionDigest;
this.commit(before, manifest, state);
});
}
stop(reason: CoordinatorStopReason, detail: string): Promise<void> {
return this.serial(() => { const state = read<State>(this.statePath); state.stop ??= { reason, detail }; atomicWrite(this.statePath, json(state)); });
}
snapshot(): Promise<Readonly<State>> { return this.serial(() => read<State>(this.statePath)); }
/** Finite waves avoid refilling while an already settled peer awaits grading.
* readCompletion owns authentic result/usage and normalized selected-closure
* fingerprint validation (not unrelated raw descriptor/manifest bytes). It must
* distinguish actual secret matches from safe, bound-only ZIP omissions.
*/
async run<T>(input: {
launchOnly: (permit: CampaignLaunchPermit) => Promise<T>;
/** Parent redacts and retains the real failure privately before returning a safe reference. */
retainFailure: (permit: CampaignLaunchPermit | null, error: unknown, phase: "reserve" | "launch" | "completion") => Promise<{ evidencePath: string; summary: string }>;
readCompletion: (permit: CampaignLaunchPermit, result: T) => Promise<{ completion: Completion; stop?: { reason: CoordinatorStopReason; detail: string }; qualificationReview?: CampaignQualificationReview }>;
}) {
const unretained: unknown[] = [];
const existing = await this.snapshot();
const remaining = this.plans.filter(plan => !Object.values(existing.entries).some(entry => entry.permit.scenarioId === plan.scenarioId && entry.phase === "completed"));
for (let index = 0; index < remaining.length; index += 2) {
if ((await this.snapshot()).stop) break;
const settled = await Promise.allSettled(remaining.slice(index, index + 2).map(async plan => {
let permit: CampaignLaunchPermit | null = null;
let phase: "reserve" | "launch" | "completion" = "reserve";
try {
permit = await this.reserve(plan.scenarioId); phase = "launch";
const result = await this.launchOnce(permit, input.launchOnly); phase = "completion";
const observed = await input.readCompletion(permit, result);
await this.complete(permit, observed.completion, observed.stop, observed.qualificationReview);
} catch (error) {
try {
const retained = await input.retainFailure(permit, error, phase);
if (!retained.evidencePath.startsWith("tests/runner-e2e/results/") || retained.evidencePath.split("/").includes("..") || !retained.summary.trim())
throw new Error("Failure requires a sanitized retained evidence reference");
if (!this.poisoned) await this.serial(() => {
const state = read<State>(this.statePath);
state.failures.push({ scenarioId: plan.scenarioId, phase, ...retained });
state.stop ??= { reason: "coordinator_failure", detail: `${retained.summary} (${retained.evidencePath})` };
atomicWrite(this.statePath, json(state));
});
else unretained.push(error); // WAL/lock remain authoritative; expose original error to owner.
} catch (retentionError) {
unretained.push(error, retentionError);
if (!this.poisoned) await this.stop("coordinator_failure", "Failure evidence retention failed; inspect returned errors");
}
}
}));
for (const result of settled) if (result.status === "rejected") unretained.push(result.reason);
if (unretained.length) throw new AggregateError(unretained, "Coordinator failure requires independent reconciliation; no further admission");
if (this.poisoned) throw new Error("Coordinator ledger transaction interrupted");
if ((await this.snapshot()).stop) break;
}
return this.snapshot();
}
close(): Promise<void> {
return this.serial(() => {
if (this.inFlight) throw new Error("Cannot release coordinator with launchers in flight");
if (existsSync(this.intentPath)) throw new Error("Cannot release interrupted transaction");
unlinkSync(this.lockPath); this.closed = true;
});
}
}