@slowdini/slow-powers-opencode 0.3.0 → 0.4.0

package/skills/evaluating-skills/runner/run.ts

@@ -1,1388 +0,0 @@
-#!/usr/bin/env bun
-import { randomBytes } from "node:crypto";
-import {
-  cpSync,
-  existsSync,
-  mkdirSync,
-  mkdtempSync,
-  readdirSync,
-  readFileSync,
-  rmSync,
-  statSync,
-  writeFileSync,
-} from "node:fs";
-import { tmpdir } from "node:os";
-import { basename, dirname, join } from "node:path";
-import {
-  renderAvailableSkillsBlock,
-  renderPlanModeContext,
-} from "./adapters/claude-code-session";
-import { detectRunContext, type Harness, type RunContext } from "./context";
-import { installGuard, teardownGuard } from "./guard/install";
-import {
-  detectPluginShadows,
-  formatShadowBanner,
-  resolveConfigDir,
-} from "./plugin-shadow";
-import type {
-  AvailableSkill,
-  ConditionsRecord,
-  Eval,
-  EvalsConfig,
-} from "./types";
-import { validateEvalsConfig } from "./validate";
-import { cleanupWorkspace, SNAPSHOT_META } from "./workspace-teardown";
-
-export const STAGED_SKILL_PREFIX = "slow-powers-eval-";
-export const STAGED_SIBLING_MANIFEST = ".slow-powers-eval-manifest.json";
-
-export function stageSkillForCC(opts: {
-  content: string;
-  iteration: number;
-  condition: string;
-  skillName: string;
-  repoRoot: string;
-  /**
-   * Source skill directory whose sibling assets are copied alongside the staged
-   * SKILL.md — everything next to SKILL.md except SKILL.md itself, the `evals/`
-   * dir, and the snapshot bookkeeping file. A multi-file skill whose SKILL.md
-   * links a sibling (e.g. `[code-review.md](code-review.md)`) would otherwise be
-   * staged with a dangling link: the agent can't resolve the reference relative
-   * to the staged dir, so the linked guidance is silently unreachable. Mirrors
-   * the sibling-asset copy in `snapshot`. Omit to stage SKILL.md alone.
-   */
-  assetsDir?: string;
-  /**
-   * When set, stage under this verbatim identifier instead of the conspicuous
-   * `slow-powers-eval-…` slug. Used by `--stage-name` to A/B a natural name
-   * against the eval-flagged one (issue #144 Step 2). A custom name is not
-   * caught by `cleanupStagedSkills`'s prefix scan, so the caller must also call
-   * `registerStagedSkillForCleanup` to have it removed on the next run.
-   */
-  stageNameOverride?: string;
-}): string {
-  const slug =
-    opts.stageNameOverride ??
-    `${STAGED_SKILL_PREFIX}${opts.iteration}-${opts.condition}__${opts.skillName}`;
-  const skillDir = join(opts.repoRoot, ".claude", "skills", slug);
-  mkdirSync(skillDir, { recursive: true });
-  writeFileSync(join(skillDir, "SKILL.md"), opts.content);
-  if (opts.assetsDir !== undefined && existsSync(opts.assetsDir)) {
-    for (const entry of readdirSync(opts.assetsDir)) {
-      if (entry === "SKILL.md" || entry === "evals" || entry === SNAPSHOT_META)
-        continue;
-      const src = join(opts.assetsDir, entry);
-      const dst = join(skillDir, entry);
-      if (statSync(src).isDirectory()) cpSync(src, dst, { recursive: true });
-      else cpSync(src, dst);
-    }
-  }
-  return slug;
-}
-
-/**
- * Adds a custom-named staged skill dir (one created via `stageNameOverride`) to
- * the sibling manifest's `created_entries` so the next run's
- * `cleanupStagedSkills` removes it — the prefix scan only catches
- * `slow-powers-eval-…` names. Idempotent: a name already recorded is left alone.
- */
-export function registerStagedSkillForCleanup(
-  repoRoot: string,
-  name: string,
-): void {
-  const skillsDir = join(repoRoot, ".claude", "skills");
-  const manifestPath = join(skillsDir, STAGED_SIBLING_MANIFEST);
-  let manifest: SiblingManifest;
-  if (existsSync(manifestPath)) {
-    manifest = JSON.parse(readFileSync(manifestPath, "utf8"));
-  } else {
-    manifest = {
-      created_at: new Date().toISOString(),
-      staged_under_test: name,
-      skills_dir_preexisting: true,
-      created_entries: [],
-    };
-  }
-  if (manifest.created_entries.some((e) => e.name === name)) return;
-  manifest.created_entries.push({ name, preexisting: false });
-  writeFileSync(manifestPath, `${JSON.stringify(manifest, null, 2)}\n`);
-}
-
-type SiblingManifest = {
-  created_at: string;
-  staged_under_test: string;
-  /**
-   * Whether `.claude/skills` already existed when staging began. When false the
-   * runner created it, so {@link cleanupStagedSkills} may remove the whole tree
-   * (and prune an emptied `.claude`); when true (or absent, on older manifests)
-   * cleanup falls back to the surgical per-entry restore so a user's own
-   * project skills are left intact.
-   */
-  skills_dir_preexisting?: boolean;
-  created_entries: Array<{
-    name: string;
-    preexisting: boolean;
-    backup_path?: string;
-  }>;
-};
-
-export function stageSiblingSkills(opts: {
-  skillUnderTest: string;
-  skillsSourceDir: string;
-  repoRoot: string;
-}): SiblingManifest {
-  const skillsDir = join(opts.repoRoot, ".claude", "skills");
-  const skillsDirPreexisting = existsSync(skillsDir);
-  mkdirSync(skillsDir, { recursive: true });
-
-  const siblings = readdirSync(opts.skillsSourceDir).filter((name) => {
-    if (name === opts.skillUnderTest) return false;
-    const srcDir = join(opts.skillsSourceDir, name);
-    if (!statSync(srcDir).isDirectory()) return false;
-    return existsSync(join(srcDir, "SKILL.md"));
-  });
-
-  const manifest: SiblingManifest = {
-    created_at: new Date().toISOString(),
-    staged_under_test: opts.skillUnderTest,
-    skills_dir_preexisting: skillsDirPreexisting,
-    created_entries: [],
-  };
-
-  for (const name of siblings) {
-    const srcDir = join(opts.skillsSourceDir, name);
-    const dstDir = join(skillsDir, name);
-    const evalsSubdir = join(srcDir, "evals");
-
-    const entry: SiblingManifest["created_entries"][number] = {
-      name,
-      preexisting: false,
-    };
-
-    if (existsSync(dstDir)) {
-      entry.preexisting = true;
-      const backupRoot = mkdtempSync(
-        join(tmpdir(), "slow-powers-eval-backup-"),
-      );
-      entry.backup_path = join(backupRoot, name);
-      cpSync(dstDir, entry.backup_path, { recursive: true });
-      rmSync(dstDir, { recursive: true, force: true });
-    }
-
-    cpSync(srcDir, dstDir, {
-      recursive: true,
-      filter: (src) =>
-        src !== evalsSubdir && !src.startsWith(`${evalsSubdir}/`),
-    });
-
-    manifest.created_entries.push(entry);
-  }
-
-  writeFileSync(
-    join(skillsDir, STAGED_SIBLING_MANIFEST),
-    `${JSON.stringify(manifest, null, 2)}\n`,
-  );
-  return manifest;
-}
-
-/** Remove `dir` only if it exists and is empty. Used to prune a `.claude` the
- * runner emptied without ever touching a `.claude` that still holds the user's
- * own files (e.g. `settings.json`). */
-function pruneIfEmpty(dir: string): void {
-  if (existsSync(dir) && readdirSync(dir).length === 0) {
-    rmSync(dir, { recursive: true, force: true });
-  }
-}
-
-export function cleanupStagedSkills(repoRoot: string): void {
-  const claudeDir = join(repoRoot, ".claude");
-  const skillsDir = join(claudeDir, "skills");
-  if (!existsSync(skillsDir)) return;
-
-  for (const entry of readdirSync(skillsDir)) {
-    if (!entry.startsWith(STAGED_SKILL_PREFIX)) continue;
-    rmSync(join(skillsDir, entry), { recursive: true, force: true });
-  }
-
-  const manifestPath = join(skillsDir, STAGED_SIBLING_MANIFEST);
-  if (!existsSync(manifestPath)) return;
-  let manifest: SiblingManifest;
-  try {
-    manifest = JSON.parse(readFileSync(manifestPath, "utf8"));
-  } catch {
-    rmSync(manifestPath, { force: true });
-    return;
-  }
-
-  // The runner created `.claude/skills` this run, so it can't be holding any of
-  // the user's own skills — remove the whole staged tree (including any stray,
-  // non-prefixed dirs a recursive eval left behind), then prune an emptied
-  // `.claude`. In a real project `.claude/settings.json` keeps `.claude`
-  // non-empty, so only the scaffolding we created is removed.
-  if (manifest.skills_dir_preexisting === false) {
-    rmSync(skillsDir, { recursive: true, force: true });
-    pruneIfEmpty(claudeDir);
-    return;
-  }
-
-  for (const e of manifest.created_entries) {
-    const target = join(skillsDir, e.name);
-    rmSync(target, { recursive: true, force: true });
-    if (e.preexisting && e.backup_path && existsSync(e.backup_path)) {
-      cpSync(e.backup_path, target, { recursive: true });
-      rmSync(dirname(e.backup_path), { recursive: true, force: true });
-    }
-  }
-  rmSync(manifestPath, { force: true });
-}
-
-type Mode = "new-skill" | "revision";
-
-type Args = {
-  command:
-    | "run"
-    | "snapshot"
-    | "teardown-guard"
-    | "teardown"
-    | "ingest"
-    | "finalize";
-  mode?: Mode;
-  baseline?: string;
-  label?: string;
-  iteration?: number;
-  only?: string[];
-  skip?: string[];
-  dryRun: boolean;
-  noStage: boolean;
-  guard: boolean;
-  stageName?: string;
-  planMode: boolean;
-  ref?: string;
-  subagentsDir?: string;
-};
-
-function die(msg: string): never {
-  console.error(`error: ${msg}`);
-  process.exit(1);
-}
-
-/**
- * Reads the bytes of `<ref>:./<relPath>` from git, resolving `relPath` relative
- * to `cwd` via the `./` prefix. Returns the raw stdout Buffer on success (write
- * it directly — never `.toString()` — so binary assets round-trip intact), or
- * `null` if the object doesn't exist at that ref (git exits non-zero). Mirrors
- * the `Bun.spawnSync` git pattern in `promote-baseline.ts:gitHead`; runs git
- * directly (no shell), so the ref/path aren't interpolated into a shell string.
- */
-function gitShowBytes(
-  cwd: string,
-  ref: string,
-  relPath: string,
-): Buffer | null {
-  const res = Bun.spawnSync(["git", "show", `${ref}:./${relPath}`], {
-    cwd,
-    stdout: "pipe",
-    stderr: "pipe",
-  });
-  if (res.exitCode !== 0) return null;
-  return Buffer.from(res.stdout);
-}
-
-/**
- * Lists every file under `cwd` as it existed at `<ref>`, as paths relative to
- * `cwd` (git's default ls-tree output strips the cwd prefix). `die`s with git's
- * stderr on failure — a bad ref or a cwd outside any repo surfaces here.
- */
-function gitLsFiles(cwd: string, ref: string): string[] {
-  const res = Bun.spawnSync(["git", "ls-tree", "-r", "--name-only", ref, "."], {
-    cwd,
-    stdout: "pipe",
-    stderr: "pipe",
-  });
-  if (res.exitCode !== 0)
-    die(`git ls-tree failed for ref ${ref}: ${res.stderr.toString().trim()}`);
-  return res.stdout
-    .toString()
-    .split("\n")
-    .map((s) => s.trim())
-    .filter(Boolean);
-}
-
-function parseArgs(argv: string[]): Args {
-  const positionals = argv.filter((a) => !a.startsWith("--"));
-  const COMMANDS: Args["command"][] = [
-    "snapshot",
-    "teardown-guard",
-    "teardown",
-    "ingest",
-    "finalize",
-  ];
-  const command: Args["command"] =
-    COMMANDS.find((c) => c === positionals[0]) ?? "run";
-
-  const flag = (name: string): string | undefined => {
-    const i = argv.indexOf(`--${name}`);
-    if (i === -1) return undefined;
-    const v = argv[i + 1];
-    if (v === undefined || v.startsWith("--")) {
-      die(`flag --${name} requires a value`);
-    }
-    return v;
-  };
-
-  const has = (name: string) => argv.includes(`--${name}`);
-
-  const iterationFlag = flag("iteration");
-  const iteration =
-    iterationFlag !== undefined ? Number(iterationFlag) : undefined;
-  if (iteration !== undefined && !Number.isInteger(iteration))
-    die(`--iteration must be an integer, got ${iterationFlag}`);
-
-  const parseIdList = (v: string | undefined): string[] | undefined =>
-    v === undefined
-      ? undefined
-      : v
-          .split(",")
-          .map((s) => s.trim())
-          .filter(Boolean);
-
-  return {
-    command,
-    mode: flag("mode") as Mode | undefined,
-    baseline: flag("baseline"),
-    label: flag("label"),
-    iteration,
-    only: parseIdList(flag("only")),
-    skip: parseIdList(flag("skip")),
-    dryRun: has("dry-run"),
-    noStage: has("no-stage"),
-    guard: has("guard"),
-    stageName: flag("stage-name"),
-    planMode: has("plan-mode"),
-    ref: flag("ref"),
-    subagentsDir: flag("subagents-dir"),
-  };
-}
-
-function ensureDir(path: string): void {
-  if (!existsSync(path)) mkdirSync(path, { recursive: true });
-}
-
-function writeJson(path: string, value: unknown): void {
-  writeFileSync(path, `${JSON.stringify(value, null, 2)}\n`);
-}
-
-function readJson<T>(path: string): T {
-  return JSON.parse(readFileSync(path, "utf8"));
-}
-
-function nextIteration(workspaceSkillDir: string, override?: number): number {
-  if (override !== undefined) return override;
-  if (!existsSync(workspaceSkillDir)) return 1;
-  const entries = readdirSync(workspaceSkillDir).filter((e) =>
-    e.startsWith("iteration-"),
-  );
-  if (entries.length === 0) return 1;
-  const nums = entries
-    .map((e) => Number(e.slice("iteration-".length)))
-    .filter((n) => Number.isFinite(n));
-  return Math.max(...nums, 0) + 1;
-}
-
-function conditionNamesFor(mode: Mode): [string, string] {
-  return mode === "new-skill"
-    ? ["with_skill", "without_skill"]
-    : ["old_skill", "new_skill"];
-}
-
-function commandSnapshot(args: Args, ctx: RunContext): void {
-  if (!args.label) die("snapshot requires --label <name>");
-  const skillDir = ctx.skillSubdir;
-
-  const destDir = join(
-    ctx.workspaceRoot,
-    ctx.skillName,
-    "snapshots",
-    args.label,
-  );
-  if (existsSync(destDir))
-    die(
-      `snapshot already exists: ${destDir}\n` +
-        "  Use a different --label or delete the existing snapshot first.",
-    );
-
-  if (args.ref !== undefined) {
-    snapshotFromRef(args.ref, skillDir, destDir, ctx.skillName);
-    return;
-  }
-
-  const skillMd = join(skillDir, "SKILL.md");
-  if (!existsSync(skillMd)) die(`skill not found: ${skillMd}`);
-  ensureDir(destDir);
-
-  cpSync(skillMd, join(destDir, "SKILL.md"));
-  for (const entry of readdirSync(skillDir)) {
-    if (entry === "SKILL.md" || entry === "evals") continue;
-    const src = join(skillDir, entry);
-    const dst = join(destDir, entry);
-    if (statSync(src).isDirectory()) cpSync(src, dst, { recursive: true });
-    else cpSync(src, dst);
-  }
-
-  // Record provenance so teardown keeps this (working-tree) snapshot — unlike a
-  // ref snapshot, it can't be regenerated from git.
-  writeJson(join(destDir, SNAPSHOT_META), { source: "working-tree" });
-
-  console.log(`Snapshotted ${ctx.skillName} → ${destDir}`);
-}
-
-/**
- * Snapshots the skill (SKILL.md + sibling assets) as it existed at a git ref,
- * read straight from the object database without touching the working tree
- * (issue #122). The `evals/` directory is excluded to match the working-tree
- * branch. Git runs from `skillDir`, which must sit inside a repo; a bad ref or a
- * skill absent at that ref `die`s with a clear message.
- */
-function snapshotFromRef(
-  ref: string,
-  skillDir: string,
-  destDir: string,
-  skillName: string,
-): void {
-  const skillMd = gitShowBytes(skillDir, ref, "SKILL.md");
-  if (skillMd === null)
-    die(
-      `skill not found at ${ref}: ${join(skillDir, "SKILL.md")}\n` +
-        "  Check the ref exists and that the skill was present there (and that this is a git repo).",
-    );
-
-  ensureDir(destDir);
-  writeFileSync(join(destDir, "SKILL.md"), skillMd);
-
-  for (const relPath of gitLsFiles(skillDir, ref)) {
-    if (relPath === "SKILL.md") continue;
-    if (relPath === "evals" || relPath.startsWith("evals/")) continue;
-    const bytes = gitShowBytes(skillDir, ref, relPath);
-    if (bytes === null) continue; // listed but unreadable (e.g. submodule/gitlink)
-    const dst = join(destDir, relPath);
-    ensureDir(dirname(dst));
-    writeFileSync(dst, bytes);
-  }
-
-  // Record provenance so teardown can reclaim this snapshot — it's fully
-  // reproducible from the ref.
-  writeJson(join(destDir, SNAPSHOT_META), { source: "ref", ref });
-
-  console.log(`Snapshotted ${skillName} at ${ref} → ${destDir}`);
-}
-
-function commandRun(args: Args, ctx: RunContext): void {
-  if (!args.mode) die("--mode required: new-skill | revision");
-  if (args.mode !== "new-skill" && args.mode !== "revision")
-    die(`unknown --mode: ${args.mode}`);
-  if (args.mode === "revision" && !args.baseline)
-    die("revision mode requires --baseline <label>");
-
-  const skillDir = ctx.skillSubdir;
-  const skillMd = join(skillDir, "SKILL.md");
-  if (!existsSync(skillMd)) die(`skill not found: ${skillMd}`);
-
-  const evalsPath = join(skillDir, "evals", "evals.json");
-  if (!existsSync(evalsPath)) die(`evals.json not found: ${evalsPath}`);
-
-  const config: EvalsConfig = validateEvalsConfig(
-    readJson(evalsPath),
-    evalsPath,
-  );
-  if (config.skill_name !== ctx.skillName)
-    console.warn(
-      `warning: evals.json skill_name (${config.skill_name}) does not match the skill folder (${ctx.skillName}). Proceeding with ${ctx.skillName}.`,
-    );
-
-  let selectedEvals: Eval[];
-  try {
-    selectedEvals = selectEvals(config.evals, {
-      only: args.only,
-      skip: args.skip,
-    });
-  } catch (err) {
-    die(err instanceof Error ? err.message : String(err));
-  }
-
-  const workspaceSkillDir = join(ctx.workspaceRoot, ctx.skillName);
-  const iteration = nextIteration(workspaceSkillDir, args.iteration);
-  const iterationDir = join(workspaceSkillDir, `iteration-${iteration}`);
-
-  // A per-run nonce makes each dispatch description globally unique. The
-  // subagents dir is shared across iterations of one parent session, so a bare
-  // `<eval>:<condition>` description repeats and fill-transcripts could fill an
-  // iteration's run from a colliding agent in another iteration. `i<N>-<nonce>`
-  // also disambiguates re-running the same iteration number.
-  const runNonce = `${Date.now().toString(36)}-${randomBytes(3).toString("hex")}`;
-  const runTag = `i${iteration}-${runNonce}`;
-
-  if (existsSync(iterationDir) && args.iteration === undefined)
-    die(
-      `iteration-${iteration} already exists; pass --iteration to overwrite explicitly`,
-    );
-
-  const [conditionA, conditionB] = conditionNamesFor(args.mode);
-
-  let skillPathForA: string | null;
-  let skillPathForB: string | null;
-  if (args.mode === "new-skill") {
-    skillPathForA = skillMd;
-    skillPathForB = null;
-  } else {
-    const baselineSkill = join(
-      workspaceSkillDir,
-      "snapshots",
-      args.baseline as string,
-      "SKILL.md",
-    );
-    if (!existsSync(baselineSkill))
-      die(
-        `baseline snapshot not found: ${baselineSkill}\n` +
-          `  Run: bun run evals:snapshot --skill ${ctx.skillName} --skill-dir ${ctx.skillDir} --label ${args.baseline} (before editing)`,
-      );
-    skillPathForA = baselineSkill;
-    skillPathForB = skillMd;
-  }
-
-  console.log(
-    `Preparing ${ctx.skillName} iteration-${iteration} (${args.mode})`,
-  );
-  console.log(`  ${conditionA}: ${skillPathForA ?? "(no skill)"}`);
-  console.log(`  ${conditionB}: ${skillPathForB ?? "(no skill)"}`);
-  if (selectedEvals.length !== config.evals.length) {
-    const [flagName, ids] = args.only
-      ? ["--only", args.only]
-      : ["--skip", args.skip ?? []];
-    console.log(
-      `  selection: ${selectedEvals.length} of ${config.evals.length} evals (${flagName} ${ids.join(", ")})`,
-    );
-  }
-  if (args.noStage)
-    console.log(
-      "  staging: disabled (--no-stage) — skills will be inlined into dispatch_prompt for harnesses without project-local skill discovery",
-    );
-
-  ensureDir(iterationDir);
-  cpSync(skillMd, join(iterationDir, "skill-snapshot.md"));
-
-  // Always disarm a prior run's guard before re-staging, so a crashed run can't
-  // leave the write-blocking hook armed across runs.
-  teardownGuard(ctx.stageRoot);
-
-  if (!args.noStage) cleanupStagedSkills(ctx.stageRoot);
-
-  if (!args.noStage) {
-    stageSiblingSkills({
-      skillUnderTest: ctx.skillName,
-      skillsSourceDir: ctx.skillDir,
-      repoRoot: ctx.stageRoot,
-    });
-  }
-
-  const bootstrapContent =
-    ctx.bootstrapPath !== null ? readFileSync(ctx.bootstrapPath, "utf8") : null;
-
-  // `--plan-mode` (issue #142): inject the harness's verbatim plan-mode
-  // procedure as an operating-context layer. The profile is a bundled asset
-  // resolved relative to this runner (mirroring the guard-script resolution
-  // below) and keyed by harness, so a harness without a profile simply has no
-  // `--plan-mode` and the portable dispatch contract is unchanged.
-  const planModeContent = args.planMode
-    ? resolvePlanModeProfile(ctx.harness)
-    : null;
-  if (args.planMode)
-    console.log(
-      `  plan-mode: injecting ${ctx.harness} plan-mode profile as operating context (issue #142; necessary-not-sufficient fidelity layer)`,
-    );
-
-  // Sibling skill metadata, shared across conditions. Empty when --no-stage
-  // (nothing is staged, so nothing is discoverable to list).
-  const siblingSkills: AvailableSkill[] = args.noStage
-    ? []
-    : ctx.siblingSkillNames.map((name) => {
-        const p = join(ctx.skillDir, name, "SKILL.md");
-        return { name, path: p, description: getSkillDescription(p) };
-      });
-
-  // `--stage-name` overrides the conspicuous `slow-powers-eval-…` slug with a
-  // verbatim name (issue #144 Step 2: A/B a natural name against the eval slug).
-  // It targets the single staging condition, so reject the case where both
-  // conditions stage (e.g. revision mode) — one name can't cover two dirs — and
-  // refuse to clobber a dir that already exists (a real project skill the user
-  // owns; cleanup has already removed our own prior custom dirs by this point).
-  if (args.stageName !== undefined && !args.noStage) {
-    if (skillPathForA !== null && skillPathForB !== null) {
-      die(
-        "--stage-name is only supported when exactly one condition stages the skill (e.g. --mode new-skill); both conditions stage here.",
-      );
-    }
-    const target = join(ctx.stageRoot, ".claude", "skills", args.stageName);
-    if (existsSync(target)) {
-      die(
-        `--stage-name "${args.stageName}": ${target} already exists; refusing to clobber it. Remove it or choose a different name.`,
-      );
-    }
-  }
-
-  const stageFor = (
-    condName: string,
-    condSkillPath: string | null,
-  ): string | null => {
-    if (!condSkillPath || args.noStage) return null;
-    return stageSkillForCC({
-      content: readFileSync(condSkillPath, "utf8"),
-      iteration,
-      condition: condName,
-      skillName: ctx.skillName,
-      repoRoot: ctx.stageRoot,
-      assetsDir: dirname(condSkillPath),
-      stageNameOverride: args.stageName,
-    });
-  };
-
-  const conditionASlug = stageFor(conditionA, skillPathForA);
-  const conditionBSlug = stageFor(conditionB, skillPathForB);
-
-  // A custom-named dir isn't caught by cleanupStagedSkills's prefix scan; record
-  // it in the sibling manifest so the next run removes it.
-  if (
-    args.stageName !== undefined &&
-    (conditionASlug === args.stageName || conditionBSlug === args.stageName)
-  ) {
-    registerStagedSkillForCleanup(ctx.stageRoot, args.stageName);
-  }
-
-  const conditions: ConditionsRecord = {
-    mode: args.mode,
-    baseline: args.baseline,
-    conditions: [
-      {
-        name: conditionA,
-        skill_path: skillPathForA,
-        staged_skill_slug: conditionASlug,
-      },
-      {
-        name: conditionB,
-        skill_path: skillPathForB,
-        staged_skill_slug: conditionBSlug,
-      },
-    ],
-    timestamp: new Date().toISOString(),
-    harness: ctx.harness,
-    run_nonce: runNonce,
-  };
-  writeJson(join(iterationDir, "conditions.json"), conditions);
-
-  // availableSkills for a condition = siblings + the skill-under-test when
-  // that condition loads it. Empty when nothing was staged.
-  const availableSkillsFor = (
-    condSkillPath: string | null,
-  ): AvailableSkill[] => {
-    if (args.noStage) return [];
-    const skills = [...siblingSkills];
-    if (condSkillPath) {
-      skills.push({
-        name: ctx.skillName,
-        path: condSkillPath,
-        description: getSkillDescription(condSkillPath),
-      });
-    }
-    return skills;
-  };
-
-  const tasks: DispatchTask[] = [];
-  for (const ev of selectedEvals) {
-    const evalDir = join(iterationDir, `eval-${ev.id}`);
-    ensureDir(evalDir);
-
-    for (const [condName, condSkillPath, condSlug] of [
-      [conditionA, skillPathForA, conditionASlug],
-      [conditionB, skillPathForB, conditionBSlug],
-    ] as const) {
-      const condDir = join(evalDir, condName);
-      const outputsDir = join(condDir, "outputs");
-      ensureDir(outputsDir);
-
-      const fixtures = copyFixtures(ev, skillDir, condDir);
-      tasks.push(
-        buildDispatchTask({
-          evalId: ev.id,
-          condition: condName,
-          skillPath: condSkillPath,
-          stagedSkillSlug: condSlug,
-          userPrompt: ev.prompt,
-          fixtures,
-          outputsDir,
-          condDir,
-          bootstrapContent,
-          planModeContent,
-          skillName: ctx.skillName,
-          availableSkills: availableSkillsFor(condSkillPath),
-          runTag,
-        }),
-      );
-    }
-  }
-
-  const manifestPath = join(iterationDir, "dispatch-manifest.md");
-  writeFileSync(
-    manifestPath,
-    buildManifest({
-      skillName: ctx.skillName,
-      mode: args.mode,
-      baseline: args.baseline,
-      iteration,
-      tasks,
-    }),
-  );
-
-  // Write each prompt to its own file and reference it by path in dispatch.json.
-  // The orchestrator then dispatches with a short "read this file" prompt instead
-  // of reproducing the full prompt verbatim per Task call.
-  for (const task of tasks) {
-    writeFileSync(task.dispatch_prompt_path, task.dispatch_prompt);
-  }
-
-  const dispatchJsonPath = join(iterationDir, "dispatch.json");
-  writeJson(dispatchJsonPath, {
-    skill_name: ctx.skillName,
-    iteration,
-    run_nonce: runNonce,
-    iteration_dir: iterationDir,
-    mode: args.mode,
-    baseline: args.baseline ?? null,
-    plan_mode: args.planMode,
-    conditions: conditions.conditions,
-    harness: ctx.harness,
-    tasks: tasks.map(({ dispatch_prompt: _omit, ...rest }) => rest),
-  });
-
-  // Opt-in hard guard. Stages a PreToolUse hook that blocks subagent
-  // writes/installs outside the eval sandbox while dispatches run.
-  if (args.guard && !args.dryRun) {
-    if (args.noStage) {
-      console.warn(
-        "\n⚠ --guard requires staging enabled; skipping guard install.",
-      );
-    } else {
-      const guardScriptPath = join(import.meta.dir, "guard", "guard.ts");
-      installGuard({
-        stageRoot: ctx.stageRoot,
-        workspaceRoot: ctx.workspaceRoot,
-        guardScriptPath,
-      });
-      console.log(
-        "\n🛡 Write guard armed: a PreToolUse hook is staged in .claude/settings.local.json\n" +
-          "   and will block writes/installs outside the eval sandbox during dispatches.\n" +
-          "   It auto-expires in 6h and is removed on the next run; to remove it now:\n" +
-          "     bun run evals:teardown-guard --skill <name>",
-      );
-    }
-  }
-
-  // Plugin-shadow preflight (Claude Code): a staged skill name that is also
-  // discoverable from an enabled plugin or the global skills dir contaminates the
-  // run — subagents inherit this session's plugins, so both copies are reachable.
-  // The runner can't unload a plugin from a live session; it only flags it. The
-  // report is persisted so the aggregator can surface it in validity_warnings.
-  if (ctx.harness === "claude-code") {
-    const shadowReport = detectPluginShadows({
-      configDir: resolveConfigDir(),
-      cwd: ctx.stageRoot,
-      stagedSkillNames: [ctx.skillName, ...ctx.siblingSkillNames],
-    });
-    if (shadowReport.shadowed.length > 0) {
-      writeJson(join(iterationDir, "plugin-shadow.json"), shadowReport);
-      console.warn(formatShadowBanner(shadowReport));
-    }
-  }
-
-  console.log(`\nWorkspace prepared: ${iterationDir}`);
-  console.log(`Dispatch manifest:  ${manifestPath}`);
-  console.log(`Dispatch tasks:     ${dispatchJsonPath}`);
-  console.log(
-    `\n${tasks.length} dispatches required (${selectedEvals.length} evals × 2 conditions).`,
-  );
-
-  if (args.dryRun) console.log("\n--dry-run: stopping after workspace prep.");
-  else
-    console.log(
-      "\nNext: read dispatch.json and dispatch each task as a subagent. Then run `ingest --iteration <N> --subagents-dir <path>` (Claude Code), or write run.json + timing.json to the paths in each task by hand and run the chained steps individually (transcript-less harnesses).",
-    );
-}
-
-type DispatchTask = {
-  eval_id: string;
-  condition: string;
-  skill_path: string | null;
-  staged_skill_slug: string | null;
-  user_prompt: string;
-  fixtures: string[];
-  outputs_dir: string;
-  run_record_path: string;
-  timing_path: string;
-  agent_description: string;
-  /**
-   * Absolute path to the file holding the full dispatch prompt. The orchestrator
-   * dispatches each subagent with a short "read this file and follow it" prompt
-   * rather than inlining the prompt, so it never has to reproduce ~KB of text per
-   * Task call. `dispatch_prompt` carries the same text in-memory (for manifest
-   * building and unit tests) but is stripped from the serialized dispatch.json.
-   */
-  dispatch_prompt_path: string;
-  dispatch_prompt: string;
-};
-
-export type { AvailableSkill } from "./types";
-
-/**
- * Filters the eval list to the subset requested via `--only` / `--skip`. The
- * two flags are mutually exclusive. Every requested id must exist in the config,
- * so a typo'd id is caught up front rather than silently producing an empty or
- * surprising run. Throws on invalid input; the caller routes the message to
- * `die`. `--only` preserves the config's eval order, not the order ids were
- * passed.
- */
-export function selectEvals(
-  evals: Eval[],
-  opts: { only?: string[]; skip?: string[] },
-): Eval[] {
-  if (opts.only && opts.skip)
-    throw new Error("use only one of --only / --skip, not both");
-  const requested = opts.only ?? opts.skip;
-  if (requested === undefined) return evals;
-  if (requested.length === 0)
-    throw new Error("--only/--skip requires at least one eval id");
-
-  const known = new Set(evals.map((e) => e.id));
-  const unknown = requested.filter((id) => !known.has(id));
-  if (unknown.length)
-    throw new Error(
-      `unknown eval id(s): ${unknown.join(", ")}. ` +
-        `Available ids: ${[...known].join(", ")}`,
-    );
-
-  const set = new Set(requested);
-  return opts.only
-    ? evals.filter((e) => set.has(e.id))
-    : evals.filter((e) => !set.has(e.id));
-}
-
-function copyFixtures(ev: Eval, skillDir: string, condDir: string): string[] {
-  if (!ev.files || ev.files.length === 0) return [];
-  const inputsDir = join(condDir, "inputs");
-  ensureDir(inputsDir);
-  const copied: string[] = [];
-  for (const f of ev.files) {
-    const src = join(skillDir, "evals", f);
-    if (!existsSync(src)) die(`fixture not found: ${src}`);
-    const dst = join(inputsDir, basename(f));
-    if (statSync(src).isDirectory()) cpSync(src, dst, { recursive: true });
-    else cpSync(src, dst);
-    copied.push(dst);
-  }
-  return copied;
-}
-
-/**
- * Resolve the verbatim plan-mode procedure profile for a harness (issue #142).
- * The profile is a bundled supporting-file asset under
- * `profiles/<harness>/plan-mode.md`, resolved relative to this runner exactly
- * like the guard script (`join(import.meta.dir, "guard", "guard.ts")`). A
- * harness without a profile gets a clear error rather than a silent no-op — the
- * profile is Claude-tier fidelity, and a harness lacking one leaves the portable
- * dispatch contract unchanged (no `<system-reminder>` plan-mode block emitted).
- */
-function resolvePlanModeProfile(harness: Harness): string {
-  const profilePath = join(
-    import.meta.dir,
-    "profiles",
-    harness,
-    "plan-mode.md",
-  );
-  if (!existsSync(profilePath)) {
-    die(
-      `--plan-mode: no plan-mode profile exists for harness '${harness}' ` +
-        `(expected ${profilePath}). This is a Claude-tier fidelity layer; a ` +
-        "harness without a profile leaves the portable dispatch contract unchanged.",
-    );
-  }
-  return readFileSync(profilePath, "utf8");
-}
-
-function getSkillDescription(skillPath: string): string {
-  try {
-    const content = readFileSync(skillPath, "utf8");
-    const match = content.match(/description:\s*([^\n\r]+)/);
-    if (match) {
-      let desc = match[1].trim();
-      if (
-        (desc.startsWith('"') && desc.endsWith('"')) ||
-        (desc.startsWith("'") && desc.endsWith("'"))
-      ) {
-        desc = desc.slice(1, -1).trim();
-      }
-      return desc;
-    }
-  } catch {}
-  return "No description available.";
-}
-
-/**
- * Removes the skill-under-test's "Active Skills Directory" entry from bootstrap
- * content so a skill-absent condition (e.g. `without_skill`) carries no
- * reference to it. Targets the markdown list-item block: a top-level `*`/`-`
- * bullet whose backticked name equals `skillName`, plus its indented
- * continuation lines (the `*Trigger:*` sub-bullet). Sibling entries and the
- * heading are left intact. The eval bootstrap names skills only in that
- * directory, so this is the sole reference vector to scrub.
- */
-export function redactSkillFromBootstrap(
-  content: string,
-  skillName: string,
-): string {
-  const out: string[] = [];
-  let skipping = false;
-  for (const line of content.split("\n")) {
-    if (skipping) {
-      // Indented continuation lines belong to the entry being dropped.
-      if (/^\s+\S/.test(line)) continue;
-      skipping = false;
-    }
-    if (/^[*-]\s/.test(line) && line.includes(`\`${skillName}\``)) {
-      skipping = true;
-      continue;
-    }
-    out.push(line);
-  }
-  return out.join("\n");
-}
-
-export function buildDispatchTask(opts: {
-  evalId: string;
-  condition: string;
-  skillPath: string | null;
-  stagedSkillSlug: string | null;
-  userPrompt: string;
-  fixtures: string[];
-  outputsDir: string;
-  condDir: string;
-  bootstrapContent: string | null;
-  /**
-   * Verbatim plan-mode procedure profile (from
-   * `profiles/<harness>/plan-mode.md`) to inject as an operating-context layer,
-   * or null/undefined to omit it. Skill-agnostic, so it is identical across the
-   * with/without-skill arms and needs no redaction. Set by the `--plan-mode`
-   * flag (issue #142): the highest-fidelity in-runner approximation of a real
-   * plan mode, still text the agent reads — a necessary-not-sufficient signal.
-   */
-  planModeContent?: string | null;
-  skillName: string;
-  availableSkills: AvailableSkill[];
-  /**
-   * Per-run uniqueness suffix (`i<iteration>-<nonce>`). Appended to the
-   * dispatch description so transcripts can't collide across iterations or
-   * re-runs. Omitted in unit tests that exercise prompt assembly directly.
-   */
-  runTag?: string;
-}): DispatchTask {
-  const stagedSkills = [...opts.availableSkills].sort((a, b) =>
-    a.name.localeCompare(b.name),
-  );
-
-  let skillBlock: string;
-  if (opts.stagedSkillSlug) {
-    // Neutral slug disambiguation only — no imperative to invoke. The skill is
-    // staged under a unique slug; surface that identifier so a deliberate
-    // invocation targets the staged copy and the __skill_invoked meta-check can
-    // find it. Do NOT assert a plugin is "loaded" or tell the agent to prefer the
-    // slug "rather than the bare name": in an isolated run there is no global copy,
-    // and that framing invited the agent to hunt for one (issue #144 global-plugin
-    // leakage). Whether to invoke is left to the skill's own triggering (dropping
-    // the old "invoke if it applies" directive was the issue #119 ceiling fix).
-    skillBlock = [
-      `The \`${opts.skillName}\` skill is registered under the identifier \`${opts.stagedSkillSlug}\` and is discoverable via the Skill tool. If you invoke it, use that identifier.`,
-    ].join("\n");
-  } else if (opts.skillPath) {
-    skillBlock = [
-      "The following skill is loaded into your operating guidelines. Apply it where relevant to the user's request.",
-      "",
-      `<skill name="${basename(dirname(opts.skillPath))}">`,
-      readFileSync(opts.skillPath, "utf8").trim(),
-      "</skill>",
-    ].join("\n");
-  } else if (stagedSkills.length > 0 || opts.bootstrapContent) {
-    // Skill-absent arm in a realistic environment: stay silent. The
-    // available-skills block already omits the skill-under-test, so any
-    // commentary here would only announce the eval (and, in the control arm,
-    // draw attention to the very skill that is supposed to be absent).
-    skillBlock = "";
-  } else {
-    skillBlock = "No skill is loaded. Respond as you naturally would.";
-  }
-
-  const fixturesBlock = opts.fixtures.length
-    ? `Available fixture files:\n${opts.fixtures.map((f) => `  - ${f}`).join("\n")}`
-    : "Available fixture files: none";
-
-  // A dispatch mirrors a real session by carrying two *separate* surfaces, the
-  // way the harness actually delivers them:
-  //   1. The verbatim --bootstrap file (the SessionStart-hook equivalent),
-  //      wrapped in <session-start-context>, if supplied.
-  //   2. The list of discoverable skills, rendered in the harness's native
-  //      presentation as its own block (see adapters/claude-code-session.ts).
-  // A condition that does not load the skill-under-test (the new-skill
-  // `without_skill` arm, under staging or --no-stage) must carry zero reference
-  // to it. The skill-under-test is auto-omitted from the available-skills block
-  // (see `availableSkillsFor`). redactSkillFromBootstrap covers the other path:
-  // a *user-supplied* --bootstrap that names the skill in its own prose would
-  // otherwise leak it into the control arm. (The shipped bootstrap.md no longer
-  // enumerates skills, so that redaction is a no-op against it.)
-  const skillAbsent = !opts.skillPath && !opts.stagedSkillSlug;
-  const effectiveBootstrap =
-    opts.bootstrapContent && skillAbsent
-      ? redactSkillFromBootstrap(opts.bootstrapContent, opts.skillName)
-      : opts.bootstrapContent;
-
-  const sections: string[] = [];
-  if (effectiveBootstrap) {
-    sections.push(
-      [
-        "<session-start-context>",
-        "The following guidelines were loaded at session start by the slow-powers plugin",
-        "(equivalent to the SessionStart hook firing in a real user's environment):",
-        "",
-        effectiveBootstrap.trim(),
-        "</session-start-context>",
-        "",
-      ].join("\n"),
-    );
-  }
-  const availableSkillsBlock = renderAvailableSkillsBlock(stagedSkills);
-  if (availableSkillsBlock) {
-    sections.push(`${availableSkillsBlock}\n\n`);
-  }
-  // Plan-mode operating context (issue #142). Injected as its own block after
-  // the session-start surfaces and before the eval task framing, so it reads as
-  // a session-level mode active for this turn — layered the way the real harness
-  // delivers it, not as seed prose. Skill-agnostic: identical in both arms.
-  const planModeBlock = opts.planModeContent
-    ? renderPlanModeContext(opts.planModeContent)
-    : "";
-  if (planModeBlock) {
-    sections.push(`${planModeBlock}\n\n`);
-  }
-  const taskLines = [
-    "You are executing a single test case for a skill evaluation framework.",
-    "Treat this as a real user request — do NOT optimize behavior for the eval.",
-  ];
-  if (skillBlock) taskLines.push("", skillBlock);
-  taskLines.push(
-    "",
-    fixturesBlock,
-    `Output directory: ${opts.outputsDir}`,
-    "",
-    "Instructions:",
-    "- Write any files you produce into the output directory.",
-    `- After completing the task, write your final user-facing response to ${opts.outputsDir}/final-message.md.`,
-    "- Do not write outside the output directory.",
-    "",
-    "User request:",
-    opts.userPrompt,
-  );
-  sections.push(taskLines.join("\n"));
-
-  return {
-    eval_id: opts.evalId,
-    condition: opts.condition,
-    skill_path: opts.skillPath,
-    staged_skill_slug: opts.stagedSkillSlug,
-    user_prompt: opts.userPrompt,
-    fixtures: opts.fixtures,
-    outputs_dir: opts.outputsDir,
-    run_record_path: join(opts.condDir, "run.json"),
-    timing_path: join(opts.condDir, "timing.json"),
-    agent_description: opts.runTag
-      ? `${opts.evalId}:${opts.condition}:${opts.runTag}`
-      : `${opts.evalId}:${opts.condition}`,
-    dispatch_prompt_path: join(opts.condDir, "dispatch-prompt.txt"),
-    dispatch_prompt: sections.join(""),
-  };
-}
-
-function buildManifest(opts: {
-  skillName: string;
-  mode: Mode;
-  baseline?: string;
-  iteration: number;
-  tasks: DispatchTask[];
-}): string {
-  const header = [
-    `# Dispatch manifest — ${opts.skillName} iteration-${opts.iteration}`,
-    "",
-    `Mode: ${opts.mode}${opts.baseline ? ` (baseline: ${opts.baseline})` : ""}`,
-    `Generated: ${new Date().toISOString()}`,
-    `Total dispatches: ${opts.tasks.length}`,
-    "",
-    "## How to use this manifest",
-    "",
-    'In an agent session, read `dispatch.json` (sibling of this file) instead of this manifest. Each task has a `dispatch_prompt_path` field pointing at the file that holds the full prompt — dispatch the subagent with a short "read this file and follow it" instruction rather than inlining the prompt — plus exact paths for `run.json` and `timing.json`.',
-    "",
-    "**Transcript correlation:** Each task has an `agent_description` field of the form `<eval_id>:<condition>:i<N>-<nonce>`. When dispatching the subagent via the host's primitive (e.g. Claude Code's Agent tool), pass this string verbatim as the dispatch `description` — do not reconstruct it. The per-run nonce keeps descriptions unique across iterations sharing one session's subagents dir, so the transcript adapter correlates each subagent's persisted transcript back to the right `(eval, condition)` slot without collisions.",
-    "",
-    "After all dispatches (Claude Code):",
-    "",
-    '1. Run `bun run evals:ingest -- --skill <name> --iteration <N> --subagents-dir ~/.claude/projects/<project-slug>/<parent-session-id>/subagents/` — a fixed-order chain of record-runs (assembles every task\'s `run.json` from `dispatch.json` + the subagent\'s own `outputs/final-message.md` + the persisted transcript, and backfills `timing.json` with transcript-derived tokens/duration; never clobbers an existing record), fill-transcripts, detect-stray-writes, and grade. Optional higher-fidelity timing: write `{ "total_tokens": <n>, "duration_ms": <n>, "source": "completion-event" }` from the task completion event to `timing.json` right after a dispatch — completion-event numbers always win over the backfill.',
-    "2. Dispatch the judge tasks ingest lists, then run `bun run evals:finalize -- --skill <name> --iteration <N>` for the benchmark.",
-    "",
-    "On a harness without persisted transcripts, instead write each task's `run.json` (matching `skills/evaluating-skills/schema/run-record.schema.json`, enforced at runtime by grade/fill-transcripts/detect-stray-writes) and `timing.json` by hand when its subagent returns: carry over `eval_id`, `condition`, `skill_path` (`null` on the without_skill arm), `prompt`, and `files` from the task; populate `final_message` from the subagent's reply; leave `tool_invocations` as `[]`; capture `total_tokens`/`duration_ms` from the task completion event immediately — they may not be persisted anywhere else.",
-    "",
-    "## Dispatches",
-    "",
-  ].join("\n");
-
-  const entries = opts.tasks
-    .map((t) =>
-      [
-        `### ${t.eval_id} / ${t.condition}`,
-        "",
-        `- run.json:    ${t.run_record_path}`,
-        `- timing.json: ${t.timing_path}`,
-        "",
-        "```",
-        t.dispatch_prompt,
-        "```",
-        "",
-      ].join("\n"),
-    )
-    .join("\n");
-
-  return header + entries;
-}
-
-// ---------------------------------------------------------------------------
-// ingest / finalize — fixed-order orchestrators over the sibling commands.
-//
-// The eval loop has exactly two points where only the in-harness agent can act
-// (dispatching eval subagents, dispatching judge subagents). Everything between
-// them is mechanical, so each stretch is one command: `ingest` runs the
-// post-dispatch chain and stops at the judge hand-off; `finalize` runs the
-// post-judge chain and prints the benchmark. No workspace-state inference —
-// each always runs the same steps in the same order, and every sub-step keeps
-// its own skip-if-done guard, so re-running after a fix is safe.
-// ---------------------------------------------------------------------------
-
-export type StepCommand = { label: string; argv: string[] };
-
-export function buildIngestCommands(opts: {
-  runnerDir: string;
-  skillDir: string;
-  skill: string;
-  iteration: number;
-  subagentsDir: string;
-}): StepCommand[] {
-  const shared = [
-    "--skill-dir",
-    opts.skillDir,
-    "--skill",
-    opts.skill,
-    "--iteration",
-    String(opts.iteration),
-  ];
-  const transcripts = ["--subagents-dir", opts.subagentsDir];
-  const script = (name: string) => [
-    "bun",
-    "run",
-    join(opts.runnerDir, `${name}.ts`),
-  ];
-  return [
-    {
-      label: "record-runs",
-      argv: [...script("record-runs"), ...shared, ...transcripts],
-    },
-    // record-runs subsumes this for the records it wrote; it still fills any
-    // pre-existing (agent-written) run.json with empty tool_invocations.
-    {
-      label: "fill-transcripts",
-      argv: [...script("fill-transcripts"), ...shared, ...transcripts],
-    },
-    {
-      label: "detect-stray-writes",
-      argv: [...script("detect-stray-writes"), ...shared],
-    },
-    { label: "grade", argv: [...script("grade"), ...shared] },
-  ];
-}
-
-export function buildFinalizeCommands(opts: {
-  runnerDir: string;
-  skillDir: string;
-  skill: string;
-  iteration: number;
-}): StepCommand[] {
-  const shared = [
-    "--skill-dir",
-    opts.skillDir,
-    "--skill",
-    opts.skill,
-    "--iteration",
-    String(opts.iteration),
-  ];
-  return [
-    {
-      label: "grade --finalize",
-      argv: [
-        "bun",
-        "run",
-        join(opts.runnerDir, "grade.ts"),
-        ...shared,
-        "--finalize",
-      ],
-    },
-    {
-      label: "aggregate",
-      argv: ["bun", "run", join(opts.runnerDir, "aggregate.ts"), ...shared],
-    },
-  ];
-}
-
-/**
- * Runs steps in order, stopping at the first non-zero exit. A failure must
- * halt the chain: grade's `__skill_invoked` code-check silently degrades to an
- * LLM judge when `tool_invocations` is missing, so grading after a failed
- * record/fill step would quietly lose the deterministic check.
- */
-export function runSteps(
-  steps: StepCommand[],
-  spawn: (step: StepCommand) => number = (step) =>
-    Bun.spawnSync(step.argv, { stdout: "inherit", stderr: "inherit" })
-      .exitCode ?? 1,
-): { failedAt: string | null } {
-  for (const step of steps) {
-    console.log(`\n── ${step.label} ──`);
-    if (spawn(step) !== 0) return { failedAt: step.label };
-  }
-  return { failedAt: null };
-}
-
-function commandIngest(args: Args, ctx: RunContext): void {
-  if (args.iteration === undefined) die("ingest requires --iteration <N>");
-  if (!args.subagentsDir)
-    die(
-      "ingest requires --subagents-dir <path> (Claude Code persists subagent transcripts under ~/.claude/projects/<project-slug>/<parent-session-id>/subagents/)",
-    );
-  const { failedAt } = runSteps(
-    buildIngestCommands({
-      runnerDir: import.meta.dir,
-      skillDir: ctx.skillDir,
-      skill: ctx.skillName,
-      iteration: args.iteration,
-      subagentsDir: args.subagentsDir,
-    }),
-  );
-  if (failedAt)
-    die(
-      `ingest stopped at '${failedAt}'. Fix the failure and re-run ingest — completed steps skip work that's already done.`,
-    );
-
-  const judgeTasksPath = join(
-    ctx.workspaceRoot,
-    ctx.skillName,
-    `iteration-${args.iteration}`,
-    "judge-tasks.json",
-  );
-  let totalTasks: number | null = null;
-  try {
-    totalTasks =
-      readJson<{ total_tasks?: number }>(judgeTasksPath).total_tasks ?? null;
-  } catch {
-    // grade always writes judge-tasks.json; treat a read failure as unknown.
-  }
-  if (totalTasks === 0) {
-    console.log(
-      `\n✅ Ingest complete — no judge dispatches needed.\nNext: bun run evals:finalize -- --skill ${ctx.skillName} --iteration ${args.iteration}`,
-    );
-  } else {
-    console.log(
-      `\n✅ Ingest complete. Dispatch the ${totalTasks ?? ""} judge task(s) grade listed above (judge-tasks.json), then:\n  bun run evals:finalize -- --skill ${ctx.skillName} --iteration ${args.iteration}`,
-    );
-  }
-}
-
-function commandFinalize(args: Args, ctx: RunContext): void {
-  if (args.iteration === undefined) die("finalize requires --iteration <N>");
-  const { failedAt } = runSteps(
-    buildFinalizeCommands({
-      runnerDir: import.meta.dir,
-      skillDir: ctx.skillDir,
-      skill: ctx.skillName,
-      iteration: args.iteration,
-    }),
-  );
-  if (failedAt)
-    die(
-      `finalize stopped at '${failedAt}'. Fix the failure and re-run finalize.`,
-    );
-  console.log(
-    `\n✅ Finalize complete. Read the benchmark above, then tear down: bun run evals:teardown --skill ${ctx.skillName}`,
-  );
-}
-
-if (import.meta.main) {
-  const argv = Bun.argv.slice(2);
-  const args = parseArgs(argv);
-  let ctx: RunContext;
-  try {
-    ctx = detectRunContext(argv);
-  } catch (err) {
-    die(err instanceof Error ? err.message : String(err));
-  }
-  if (args.command === "snapshot") commandSnapshot(args, ctx);
-  else if (args.command === "ingest") commandIngest(args, ctx);
-  else if (args.command === "finalize") commandFinalize(args, ctx);
-  else if (args.command === "teardown-guard") {
-    const torn = teardownGuard(ctx.stageRoot);
-    console.log(
-      torn
-        ? "🛡 Write guard removed."
-        : "No write guard was installed — nothing to remove.",
-    );
-  } else if (args.command === "teardown") {
-    // Full end-of-run teardown: disarm the guard, remove the staged skill set
-    // (and prune a `.claude` the runner emptied), then reclaim the workspace —
-    // leaving the user's own `.claude/settings.json`, pre-existing project
-    // skills, and any uncommitted eval results intact.
-    const torn = teardownGuard(ctx.stageRoot);
-    cleanupStagedSkills(ctx.stageRoot);
-    const ws = cleanupWorkspace(ctx.workspaceRoot, ctx.skillName);
-    console.log(
-      `🧹 Eval teardown complete: staged skill set removed${
-        torn ? " and write guard disarmed" : ""
-      }.`,
-    );
-    const reclaimed = ws.removedIterations.length + ws.removedSnapshots.length;
-    if (reclaimed > 0) {
-      console.log(
-        `   Reclaimed ${ws.removedIterations.length} workspace iteration(s)` +
-          ` and ${ws.removedSnapshots.length} reproducible snapshot(s).`,
-      );
-    }
-    if (ws.keptIterations.length > 0) {
-      const lines = ws.keptIterations.map(
-        (k) => `     - ${k.iteration} (${k.reason})`,
-      );
-      console.warn(
-        `⚠ Kept ${ws.keptIterations.length} workspace iteration(s) with results ` +
-          `not yet committed:\n${lines.join("\n")}\n` +
-          `   Commit them, e.g.:\n` +
-          `     bun run evals:promote-baseline --skill ${ctx.skillName} --iteration <N>\n` +
-          `   or delete ${join("skills-workspace", ctx.skillName)}/ manually to discard.`,
-      );
-    }
-  } else commandRun(args, ctx);
-}