/**
 * copy-reference-run.ts - Copy Harbor job trials into a task's reference-runs directory.
 *
 * Usage:
 *   npx tsx scripts/copy-reference-run.ts <trial-path> [trial-path...]
 *
 * Examples:
 *   # Copy a single trial
 *   npx tsx scripts/copy-reference-run.ts harbor-jobs/2026-04-02__11-43-55/my-task-slug__3Df3Bjr
 *
 *   # Copy all trials from a job
 *   npx tsx scripts/copy-reference-run.ts harbor-jobs/2026-04-02__11-43-55/my-task-slug__*
 *
 * What gets copied:
 *   - verifier/agent-output/ (answer.md, etc.)
 *   - verifier/reward.txt + reward-correctness.txt (behavioural + correctness scores)
 *   - verifier/reward.json (the split grader's machine-readable dual-score record)
 *   - verifier/signals-status.txt (whether every deterministic check actually ran,
 *     i.e. whether the correctness score is signal-backed)
 *   - verifier/grade.md + every grade-<N>.md grader sample
 *   - verifier/grader-result(-<N>).json, grader-stderr(-<N>).log, grader-samples.txt
 *   - agent/claude-code.txt or agent/codex.txt (the harness's own log), agent/trajectory.json
 *   - session.jsonl — the resumable session log, hoisted to the top of the run dir so
 *     `view-harbor-session.ts <run-dir>/session.jsonl` can load it without further
 *     indirection. Its location is per harness: Claude Code writes
 *     `agent/sessions/projects/-workspace/<id>.jsonl` with the id printed in
 *     `agent/claude-code.txt`; codex writes `agent/sessions/<Y>/<M>/<D>/rollout-*.jsonl`.
 *   - config.json, result.json, trial.log
 *   - input-checksums.json — sha256 checksums of the task inputs the run was
 *     generated against (prompt, session snapshot, workspace patch, gitref),
 *     so submit-task.ts can warn when the run goes stale. Copied from the
 *     trial dir when harbor-run stamped one at launch time (capturedBy:
 *     'run' — immune to edits made between the run and this copy);
 *     otherwise captured here at copy time as a fallback (capturedBy:
 *     'copy').
 *
 * What is NOT copied:
 *   - agent/sessions/, agent/setup/ (workspace data — the resumable JSONL
 *     is hoisted out as `session.jsonl` above; everything else here is
 *     untyped workspace state)
 *   - artifacts/
 */

import './lib/check-devcontainer';

import {
  chmodSync,
  copyFileSync,
  existsSync,
  lstatSync,
  mkdirSync,
  readdirSync,
  readFileSync,
  readlinkSync,
  rmSync,
  statSync,
  symlinkSync,
  writeFileSync,
} from 'fs';
import { basename, join } from 'path';

import {
  captureTaskInputs,
  INPUT_CHECKSUMS_FILENAME,
  readTaskInputChecksums,
} from './lib/input-checksums';
import { manualRepairHint, normalizeTreePermissions } from './lib/tree-permissions';
import { readSessionId } from './session-id';

/**
 * Copy helpers that stand in for `cpSync`, which this script must not call.
 *
 * `cpSync`'s work happens in a C++ helper (Node 22+) that a macOS docker bind mount does
 * not satisfy: `cpSyncCopyDir` fails EACCES for every directory copy into one, and
 * `cpSyncOverrideFile` fails EACCES when a single-file copy would overwrite an existing
 * destination. Workers run this from the toolkit, whose harbor-jobs and harbor-tasks both
 * live on that mount — so `copy-reference-run` died partway through agent-output/, leaving
 * a half-copied reference run behind. mkdir/copyFile/chmod on those same paths all work.
 */
function copyPath(src: string, dest: string) {
  const st = lstatSync(src);
  if (st.isSymbolicLink()) {
    rmSync(dest, { force: true });
    symlinkSync(readlinkSync(src), dest);
    return;
  }
  if (st.isDirectory()) {
    copyTree(src, dest);
    return;
  }
  // Unlink first: copyFileSync onto an existing file keeps that file's mode.
  rmSync(dest, { force: true });
  copyFileSync(src, dest);
  chmodSync(dest, statSync(src).mode & 0o777);
}

function copyTree(src: string, dest: string) {
  mkdirSync(dest, { recursive: true });
  for (const entry of readdirSync(src, { withFileTypes: true })) {
    copyPath(join(src, entry.name), join(dest, entry.name));
  }
}

const HARBOR_TASKS_DIR = 'harbor-tasks';

function findTaskDir(trialPrefix: string): string | null {
  if (!existsSync(HARBOR_TASKS_DIR)) return null;
  const entries = readdirSync(HARBOR_TASKS_DIR, { withFileTypes: true });
  for (const entry of entries) {
    if (entry.isDirectory() && entry.name.startsWith(trialPrefix)) {
      return join(HARBOR_TASKS_DIR, entry.name);
    }
  }
  return null;
}

/** Newest `rollout-*.jsonl` anywhere under a codex `sessions/` tree, or null. */
function newestRollout(sessionsDir: string): string | null {
  if (!existsSync(sessionsDir)) return null;
  const found: Array<{ path: string; mtime: number }> = [];
  const walk = (dir: string) => {
    for (const entry of readdirSync(dir, { withFileTypes: true })) {
      const full = join(dir, entry.name);
      if (entry.isDirectory()) walk(full);
      else if (entry.name.startsWith('rollout-') && entry.name.endsWith('.jsonl')) {
        found.push({ path: full, mtime: statSync(full).mtimeMs });
      }
    }
  };
  walk(sessionsDir);
  if (found.length === 0) return null;
  found.sort((a, b) => b.mtime - a.mtime);
  return found[0].path;
}

/**
 * Resolve the task dir from the trial's result.json `task_name` — the FULL,
 * unambiguous slug. Harbor TRUNCATES long task names in the trial DIRNAME, so two
 * distinct tasks sharing a truncated prefix (e.g. `foo--hash` and `foo--hash-2`,
 * both truncating to `foo--ha`) collide: a dirname-prefix scan returns whichever
 * sorts first and misroutes the other's trials (observed in the wild as base +
 * `-2` reference-runs sharing trial IDs). result.json is written per-trial with
 * the real task_name, so it disambiguates exactly. Returns null when result.json
 * is absent/unparseable or names a task dir that doesn't exist (caller then falls
 * back to the prefix scan).
 */
function findTaskDirByResultJson(trialPath: string): string | null {
  const resultPath = join(trialPath, 'result.json');
  if (!existsSync(resultPath)) return null;
  let taskName: unknown;
  try {
    taskName = (JSON.parse(readFileSync(resultPath, 'utf-8')) as { task_name?: unknown }).task_name;
  } catch {
    return null;
  }
  if (typeof taskName !== 'string' || taskName.length === 0) return null;
  // Hub-published task_names are org-prefixed (`<org>/<slug>`); the dir is bare.
  // Inlined (not the shared bareSlug helper) because this script ships in the
  // worker toolkit and must not import outside its shipped file set.
  const dir = join(HARBOR_TASKS_DIR, taskName.replace(/^[^/]+\//, ''));
  return existsSync(dir) ? dir : null;
}

function copyTrial(trialPath: string, destName?: string) {
  trialPath = trialPath.replace(/\/$/, '');

  if (!existsSync(trialPath)) {
    console.error(`Error: ${trialPath} does not exist`);
    process.exit(1);
  }

  // Repair the SOURCE before reading a byte of it. A trial can leave files
  // write-only, which locks out their own owner: everything below — reading
  // reward.txt, copying agent-output — fails on them, and any that do get
  // through land in the task dir, where harbor hashes every file on every
  // later trial and one unreadable path aborts the run.
  let sourcePerms = null;
  try {
    sourcePerms = normalizeTreePermissions(trialPath);
  } catch (err) {
    console.warn(`Warning: could not normalize permissions on ${trialPath}: ${String(err)}`);
    console.warn(`         If the copy below fails on permissions:`);
    console.warn(`           ${manualRepairHint(trialPath)}`);
  }
  if (sourcePerms && sourcePerms.failures.length > 0) {
    console.warn(
      `Warning: could not normalize permissions on ${sourcePerms.failures.length} path(s) under ${trialPath}.`
    );
    console.warn(`         If the copy below fails on permissions, run:`);
    console.warn(`           ${manualRepairHint(trialPath)}`);
  }

  const rewardPath = join(trialPath, 'verifier', 'reward.txt');
  if (!existsSync(rewardPath)) {
    console.error(`Error: No reward.txt found in ${trialPath}/verifier/`);
    process.exit(1);
  }

  const reward = readFileSync(rewardPath, 'utf-8').trim();
  const trialDir = basename(trialPath);

  // Trial dir format: <task-slug-truncated>__<trialId>
  const separatorIndex = trialDir.lastIndexOf('__');
  if (separatorIndex === -1) {
    console.error(
      `Error: Trial directory '${trialDir}' does not match expected format <slug>__<trialId>`
    );
    process.exit(1);
  }

  const trialPrefix = trialDir.substring(0, separatorIndex);
  const trialId = trialDir.substring(separatorIndex + 2);

  // Prefer the exact task_name from result.json (handles truncated-prefix
  // collisions like `foo--hash` vs `foo--hash-2`); fall back to the dirname
  // prefix scan only when result.json can't resolve it.
  const taskDir = findTaskDirByResultJson(trialPath) ?? findTaskDir(trialPrefix);
  if (!taskDir) {
    console.error(
      `Error: Could not find task directory matching prefix '${trialPrefix}' in ${HARBOR_TASKS_DIR}/`
    );
    console.error('Available tasks:');
    readdirSync(HARBOR_TASKS_DIR).forEach((d) => console.error(`  ${d}`));
    process.exit(1);
  }

  // destName (--dest-name) makes a RECORDED rollout id authoritative: the run
  // is copied to exactly that name instead of the minted reward-<r>-<id> —
  // used by the SxS pipeline so task.toml preference_rollouts, this dir, and
  // the publish-manifest run_id stay byte-identical by construction.
  const dest = join(taskDir, 'reference-runs', destName ?? `reward-${reward}-${trialId}`);

  if (existsSync(dest)) {
    console.warn(`Warning: ${dest} already exists, overwriting`);
    rmSync(dest, { recursive: true });
  }

  mkdirSync(dest, { recursive: true });

  // Copy verifier outputs. The grader writes more than one behavioural grade: the
  // behavioural reward (reward.txt) AND the separate correctness reward
  // (reward-correctness.txt, plus the machine-readable dual-score reward.json
  // from the split grader), signals-status.txt, its aggregate reasoning (grade.md)
  // PLUS one grade-<N>.md per grader sample, the structured grade(-<N>).json
  // (the source of truth grade.md/reward.txt are rendered from), the
  // machine-readable grader-result(-<N>).json, the grader-stderr(-<N>).log,
  // render-stderr(-<N>).log, and grader-samples.txt.
  // Copy the whole set — glob so it stays agnostic to the sample count. (Earlier
  // this took only reward.txt/grade.md/grader-stderr.log and silently dropped the
  // correctness score and every per-sample record.)
  //
  // signals-status.txt is what qualifies the correctness score: it records whether
  // every deterministic check actually produced a verdict ("ok") or one or more was
  // killed before finishing ("degraded" — the correctness score is then NOT
  // signal-backed). Without it a copied run is indistinguishable from a run whose
  // checks all passed, so it must travel with reward-correctness.txt.
  const verifierDir = join(trialPath, 'verifier');
  if (existsSync(verifierDir)) {
    for (const entry of readdirSync(verifierDir, { withFileTypes: true })) {
      if (!entry.isFile()) continue;
      const f = entry.name;
      if (
        f === 'reward.txt' ||
        f === 'reward-correctness.txt' ||
        f === 'reward.json' ||
        f === 'signals-status.txt' ||
        f === 'grader-samples.txt' ||
        /^grade(-\d+)?\.md$/.test(f) ||
        /^grade(-\d+)?\.json$/.test(f) ||
        /^grader-result(-\d+)?\.json$/.test(f) ||
        /^grader-stderr(-\d+)?\.log$/.test(f) ||
        /^render-stderr(-\d+)?\.log$/.test(f)
      ) {
        copyPath(join(verifierDir, f), join(dest, f));
      }
    }
  }

  const agentOutputDir = join(verifierDir, 'agent-output');
  if (existsSync(agentOutputDir)) {
    copyTree(agentOutputDir, join(dest, 'agent-output'));
  }

  // Copy agent session log and trajectory (not workspace)
  const agentDir = join(trialPath, 'agent');
  if (existsSync(agentDir)) {
    mkdirSync(join(dest, 'agent'), { recursive: true });
    // The agent's own log is named per harness (claude-code.txt / codex.txt); copying only
    // the Claude one left codex reference runs with nothing but the trajectory.
    for (const file of ['claude-code.txt', 'codex.txt', 'trajectory.json']) {
      const src = join(agentDir, file);
      if (existsSync(src)) copyPath(src, join(dest, 'agent', file));
    }
    // The grader (and downstream worldbench export / replay) reads
    // agent/trajectory.json. If it's missing, scream so we don't silently
    // ship a reference run that's only half-useful.
    if (!existsSync(join(agentDir, 'trajectory.json'))) {
      console.warn(
        `WARNING: ${trialPath}/agent/trajectory.json is missing. ` +
          `Harbor's adapter for this harness failed to write it (typically because ` +
          `the converter choked on the session log). Downstream consumers ` +
          `(grader replay, worldbench export) need this file — investigate ` +
          `before relying on this reference run.`
      );
    }
    // Hoist the resumable session JSONL to <run-dir>/session.jsonl. The id
    // gets embedded in claude-code.txt's first non-empty line; we use it to
    // locate the sibling JSONL Claude Code wrote in the same trial.
    // Layout is per harness, so each needs a case here — the same reason
    // harness-session.mjs has one finder per CLI. A harness with no case gets no hoisted
    // session, which is what codex got before this: nothing at all.
    const claudeCodeTxt = join(agentDir, 'claude-code.txt');
    if (existsSync(claudeCodeTxt)) {
      // Claude Code: the id is in claude-code.txt, the JSONL is its sibling.
      const sessionId = readSessionId(claudeCodeTxt);
      if (sessionId) {
        const jsonl = join(agentDir, 'sessions', 'projects', '-workspace', `${sessionId}.jsonl`);
        if (existsSync(jsonl)) copyPath(jsonl, join(dest, 'session.jsonl'));
      }
    } else {
      // codex: sessions/<Y>/<M>/<D>/rollout-<ts>-<id>.jsonl, newest wins.
      const rollout = newestRollout(join(agentDir, 'sessions'));
      if (rollout) copyPath(rollout, join(dest, 'session.jsonl'));
    }
  }

  // Copy top-level metadata
  for (const file of ['config.json', 'result.json', 'trial.log']) {
    const src = join(trialPath, file);
    if (existsSync(src)) copyPath(src, join(dest, file));
  }

  // Record the checksums of the task inputs this run was generated against
  // (prompt, session snapshot, workspace patch, gitref, grader guidance).
  // submit-task.ts re-captures at packaging time and warns when any of them
  // changed — the run then describes an older revision of the task than the
  // one being shipped. Preferred source: the launch-time stamp harbor-run
  // wrote into the trial dir (scripts/stamp-trial-inputs.ts, capturedBy:
  // 'run') — it records the inputs the agent actually ran against, so an
  // input edited BETWEEN harbor-run and this copy is still caught. Fallback
  // (trials from an older harbor-run, or a failed stamp): capture here at
  // copy time, marked capturedBy: 'copy' so staleness.json is honest about
  // the weaker evidence.
  const trialStampPath = join(trialPath, INPUT_CHECKSUMS_FILENAME);
  const trialStamp = readTaskInputChecksums(trialStampPath);
  // A stamp that names a DIFFERENT task got mis-routed (e.g. concurrent
  // harbor-runs sharing a cwd) — its hashes describe some other task's
  // inputs, so treat it as absent rather than importing false evidence.
  const stampSlug = trialStamp?.taskSlug;
  const stampMisrouted = typeof stampSlug === 'string' && stampSlug !== basename(taskDir);
  if (trialStamp && !stampMisrouted) {
    copyPath(trialStampPath, join(dest, INPUT_CHECKSUMS_FILENAME));
  } else {
    if (stampMisrouted) {
      console.warn(
        `Warning: ${trialStampPath} was captured for task '${stampSlug}', not ` +
          `'${basename(taskDir)}' — ignoring it and capturing at copy time instead`
      );
    }
    writeFileSync(
      join(dest, INPUT_CHECKSUMS_FILENAME),
      JSON.stringify(captureTaskInputs(taskDir, 'copy'), null, 2) + '\n'
    );
  }

  // Files captured from a run can land unreadable to you, which makes packaging
  // fail later. Fix that now. Guarded so it can never fail a copy that worked.
  let perms = null;
  try {
    perms = normalizeTreePermissions(dest);
  } catch (err) {
    console.warn(`Warning: could not normalize permissions on ${dest}: ${String(err)}`);
    console.warn(`         The run copied fine. If packaging later fails on permissions:`);
    console.warn(`           ${manualRepairHint(dest)}`);
  }
  if (perms && perms.failures.length > 0) {
    console.warn(
      `Warning: could not normalize permissions on ${perms.failures.length} path(s) under ${dest}.`
    );
    console.warn(
      `         If packaging later fails with 'Cannot stat: Permission denied', run:\n` +
        `           ${manualRepairHint(dest)}`
    );
  }

  console.log(`Copied to ${dest}`);
  console.log(`  reward: ${reward}`);
  console.log(`  task:   ${taskDir}`);
  console.log(`  trial:  ${trialId}`);
  const ownerFixed = (sourcePerms?.ownerFixed.length ?? 0) + (perms?.ownerFixed.length ?? 0);
  const modeFixed = (sourcePerms?.modeFixed.length ?? 0) + (perms?.modeFixed.length ?? 0);
  if (ownerFixed > 0 || modeFixed > 0) {
    console.log(`  perms:  normalized ${ownerFixed} owner / ${modeFixed} mode`);
  }
}

// Main
const rawArgs = process.argv.slice(2);
let destName: string | undefined;
const args: string[] = [];
for (let i = 0; i < rawArgs.length; i++) {
  if (rawArgs[i] === '--dest-name') {
    destName = rawArgs[++i];
    if (!destName) {
      console.error('Error: --dest-name requires a value');
      process.exit(1);
    }
  } else {
    args.push(rawArgs[i]);
  }
}

if (args.length === 0) {
  console.error(
    'Usage: npx tsx scripts/copy-reference-run.ts [--dest-name <run-id>] <trial-path> [trial-path...]'
  );
  process.exit(1);
}
if (destName && args.length !== 1) {
  console.error('Error: --dest-name applies to exactly one trial path');
  process.exit(1);
}

for (const trialPath of args) {
  copyTrial(trialPath, destName);
}