project-work/worker-toolkit-stocks-in-the-future/scripts/lib/input-checksums.ts

/**
 * input-checksums.ts — capture and compare sha256 checksums of the task inputs
 * that reference runs and detector reports depend on.
 *
 * A reference run is only meaningful for the task inputs it actually ran
 * against: the prompt (instruction.md), the snapshot session
 * (environment/session.jsonl), the workspace patch
 * (environment/workspace.patch), and the gitref the workspace is built from
 * (task.toml `[metadata].commit`). Detector reports likewise assess a specific
 * revision of instruction.md + tests/grader-guidance.md. When any of those
 * change after the artifact was produced, the artifact is stale — it describes
 * an older revision of the task than the one being packaged.
 *
 * This module is the single source of truth for WHAT gets checksummed and how
 * captures are compared. Capture sites (copy-reference-run.ts,
 * record-detector-inputs.ts) write a {@link TaskInputChecksums} record next to
 * the artifact; submit-task.ts re-captures at packaging time and diffs.
 * Content hashes rather than mtimes: a re-clone / whole-tree touch can fake or
 * mask an mtime, but can't change a sha256.
 *
 * Lives in the worker toolkit's shipped file set (static/scripts/lib/), so in
 * a packed toolkit it sits at scripts/lib/ next to both consumers. Repo-side
 * callers go through the scripts/lib/input-checksums.ts re-export shim, which
 * occupies the same relative path there (mirroring the check-devcontainer
 * pattern).
 *
 * Related but deliberately separate: `computeDeliveryHash` (repo-side
 * delivery script — grep the internal repo for it; not shipped with the
 * toolkit) hashes an overlapping input set for delivery idempotency. It is
 * NOT built on this module because its hash format is load-bearing (a
 * changed hash re-delivers every task); if you change WHAT counts as a task
 * input here, check whether the delivery hash needs the same change.
 */

import { createHash } from 'node:crypto';
import { existsSync, readFileSync } from 'node:fs';
import { join } from 'node:path';

/** Bump when the record shape changes incompatibly. */
export const INPUT_CHECKSUMS_VERSION = 1;

/** Filename of the record inside a reference-run directory. */
export const INPUT_CHECKSUMS_FILENAME = 'input-checksums.json';

/**
 * Where in the artifact lifecycle a capture happened. The moment matters for
 * how much a "fresh" verdict can be trusted:
 *
 *   - 'run'    — at trial launch (scripts/harbor-run stamps the trial dir).
 *                The strongest evidence: the record is what the agent ran
 *                against, whatever got edited afterwards.
 *   - 'copy'   — at copy-reference-run time, the fallback when a trial carries
 *                no run-time stamp. An input edited between harbor-run and the
 *                copy is recorded at its post-edit state, so a stale run can
 *                read fresh.
 *   - 'stamp'  — record-detector-inputs.ts, right after a detector skill
 *                writes its report.
 *   - 'mirror' — fetch-detectors --write-dir, when canonical detector
 *                reports are re-materialized to disk from their remote
 *                store (repo-side flow only; never written in a worker
 *                checkout).
 *
 * Absent on records written before this field existed.
 */
export type TaskInputCaptureMethod = 'run' | 'copy' | 'stamp' | 'mirror';

/**
 * The checksums of a task's inputs as they stood at capture time. Every hash
 * field is a sha256 hex digest, or `null` when the file didn't exist (a null
 * that later becomes a hash — or vice versa — is a change like any other).
 * `gitref` is the raw `[metadata].commit` string, recorded verbatim rather
 * than hashed so a mismatch message can show it.
 */
export interface TaskInputChecksums {
  version: number;
  /** ISO-8601 timestamp of the capture. */
  capturedAt: string;
  /** Lifecycle point of the capture ({@link TaskInputCaptureMethod}). */
  capturedBy?: TaskInputCaptureMethod;
  /**
   * The task slug the capture was taken from (harbor-tasks/<slug>). Written
   * by launch-time captures so stamping can be scoped to the right task's
   * trial dirs when several harbor-runs share a cwd — and so a mis-routed
   * stamp is detectable after the fact.
   */
  taskSlug?: string;
  /**
   * Set when the finalize flow re-stamped the prompt/graderGuidance hashes
   * after the harbor-path scrub deliberately rewrote those docs
   * (scripts/restamp-task-inputs.ts) — the run/report still reflects the
   * task; only the doc bytes were normalized.
   */
  restampedAt?: string;
  inputs: {
    prompt: string | null;
    graderGuidance: string | null;
    sessionJsonl: string | null;
    workspacePatch: string | null;
    gitref: string | null;
    /**
     * tests/grader-guidance-consolidated.md — the consolidated-standard
     * guidance. Absent (undefined) on records captured before the field
     * existed; comparisons skip a field the record predates, so old captures
     * stay fresh until they are re-stamped. Last in field order because the
     * task-checksum digest serializes fields in this order and appends new
     * fields at the end (see scripts/lib/grader-run-checksums.ts).
     */
    graderGuidanceConsolidated?: string | null;
  };
}

export type TaskInputName = keyof TaskInputChecksums['inputs'];

/** Human-readable component names, used verbatim in staleness warnings. */
export const INPUT_LABELS: Record<TaskInputName, string> = {
  prompt: 'prompt (instruction.md)',
  graderGuidance: 'grader guidance (tests/grader-guidance.md)',
  graderGuidanceConsolidated:
    'consolidated grader guidance (tests/grader-guidance-consolidated.md)',
  sessionJsonl: 'session snapshot (environment/session.jsonl)',
  workspacePatch: 'workspace patch (environment/workspace.patch)',
  gitref: 'gitref (task.toml commit)',
};

/**
 * The inputs that shape what the AGENT saw and did. Changing any of them means
 * a captured reference run no longer reflects the task being packaged, and
 * only re-running the agent can fix that. The grader-guidance files (legacy
 * and consolidated) are deliberately NOT in this set: editing the rubric
 * stales the run's GRADE, not the run itself, and `scripts/harbor-regrade`
 * re-derives grades without re-running the agent.
 */
export const REFERENCE_RUN_INPUTS: readonly TaskInputName[] = [
  'prompt',
  'sessionJsonl',
  'workspacePatch',
  'gitref',
];

/**
 * The inputs a detector report assesses — instruction.md plus whichever
 * grader-guidance files the task carries (the legacy pair is what the old
 * mtime comparison in submit-task.ts watched; the consolidated file joined
 * when grading moved to the consolidated standard). Compared by content.
 */
export const DETECTOR_REPORT_INPUTS: readonly TaskInputName[] = [
  'prompt',
  'graderGuidance',
  'graderGuidanceConsolidated',
];

/** sha256 hex digest of a file's bytes, or null when it doesn't exist. */
function sha256File(filePath: string): string | null {
  if (!existsSync(filePath)) return null;
  return createHash('sha256').update(readFileSync(filePath)).digest('hex');
}

/**
 * The gitref (`[metadata].commit`) from a task.toml, or null when the file is
 * missing, unreadable, or has no commit line — a read failure downgrades to
 * "absent" rather than crashing a capture or a validation sweep.
 *
 * Deliberately a regex, not a TOML parser: this module ships in the worker
 * toolkit, where a new runtime dep would break packaging for every worker
 * whose container predates the dep (npm install runs only on container
 * create, and containers survive toolkit upgrades). build-workspace.sh reads
 * the same key with the same grep-a-`commit`-line approach. The one `commit`
 * key in a task.toml is `[metadata].commit`, so anchoring to the first
 * `commit = "…"` line is exact in practice.
 */
function readGitref(taskDir: string): string | null {
  const tomlPath = join(taskDir, 'task.toml');
  if (!existsSync(tomlPath)) return null;
  try {
    const match = /^\s*commit\s*=\s*(?:"([^"\n]+)"|'([^'\n]+)')\s*(?:#.*)?$/m.exec(
      readFileSync(tomlPath, 'utf-8')
    );
    const commit = match?.[1] ?? match?.[2];
    return commit && commit.length > 0 ? commit : null;
  } catch {
    return null;
  }
}

/** Checksum the task inputs as they stand right now under `taskDir`. */
export function captureTaskInputs(
  taskDir: string,
  capturedBy?: TaskInputCaptureMethod
): TaskInputChecksums {
  return {
    version: INPUT_CHECKSUMS_VERSION,
    capturedAt: new Date().toISOString(),
    ...(capturedBy ? { capturedBy } : {}),
    inputs: {
      prompt: sha256File(join(taskDir, 'instruction.md')),
      graderGuidance: sha256File(join(taskDir, 'tests', 'grader-guidance.md')),
      sessionJsonl: sha256File(join(taskDir, 'environment', 'session.jsonl')),
      workspacePatch: sha256File(join(taskDir, 'environment', 'workspace.patch')),
      gitref: readGitref(taskDir),
      graderGuidanceConsolidated: sha256File(
        join(taskDir, 'tests', 'grader-guidance-consolidated.md')
      ),
    },
  };
}

/**
 * Read a previously captured record. Returns null when the file is missing or
 * doesn't look like a capture (pre-tracking artifact, hand-edited JSON, a
 * future incompatible version) — callers treat null as "staleness unknowable",
 * never as an error. No zod here: this module ships in the worker toolkit,
 * whose dependency set stays minimal, so the guard is manual.
 */
export function readTaskInputChecksums(filePath: string): TaskInputChecksums | null {
  if (!existsSync(filePath)) return null;
  let parsed: unknown;
  try {
    parsed = JSON.parse(readFileSync(filePath, 'utf-8'));
  } catch {
    return null;
  }
  if (typeof parsed !== 'object' || parsed === null) return null;
  const record = parsed as TaskInputChecksums;
  if (record.version !== INPUT_CHECKSUMS_VERSION) return null;
  if (typeof record.inputs !== 'object' || record.inputs === null) return null;
  for (const name of Object.keys(INPUT_LABELS) as TaskInputName[]) {
    const value = record.inputs[name];
    // undefined = the record predates this input field; still a valid capture.
    if (value !== undefined && value !== null && typeof value !== 'string') return null;
  }
  return record;
}

/**
 * Which of `names` changed between a recorded capture and the current state?
 * Returns the human-readable labels ({@link INPUT_LABELS}) of every component
 * whose value differs — including absent→present and present→absent flips. A
 * field the recorded capture predates (the key is not in the record at all)
 * is skipped: freshness on that axis is unknowable, and flagging every old
 * record the moment a new axis ships would drown the real signal.
 */
export function diffTaskInputs(
  recorded: TaskInputChecksums,
  current: TaskInputChecksums,
  names: readonly TaskInputName[]
): string[] {
  return names
    .filter((name) => name in recorded.inputs)
    .filter((name) => recorded.inputs[name] !== current.inputs[name])
    .map((name) => INPUT_LABELS[name]);
}