feat(evals): register shipped skills for hermetic PTY children

Hermetic children get a config dir that deliberately seeds no skills —
right for children that install their own, fatal for the PTY family that
TYPES /office-hours or /plan-ceo-review: claude rejects the command as
Unknown before any model turn, so the plan-family gate smokes measure
nothing. hermeticSkillsConfigDir() is a second, opt-in config dir under
the same runRoot that mirrors ./setup's registration exactly (real dir
per registry name, SKILL.md + sections/ symlinks, frontmatter-name
resolution, _gstack-command root alias), driven by the shared
skill-census so connect-chrome's dir symlink collapses the same way
setup's idempotent overwrite does.

Ported from fork commit 03c4eca2, tree walk rewritten for the upstream
layout (top-level <skill>/SKILL.md dirs, no skills/ tree). Unit tests
are new: seed shape, census parity, symlink resolution, connect-chrome
collapse, idempotence, no-API-key seed.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
(cherry picked from commit 93dae6107b30ce453a07c2d342b60262bba6ce0b)

test/helpers/hermetic-env.ts

@@ -36,7 +36,8 @@ import * as fs from 'fs';
 import * as path from 'path';
 import * as os from 'os';
 import { promotedEnv } from '../../lib/conductor-env-shim';
-import { isProcessAlive } from '../../browse/src/error-handling';
+import { isProcessAlive, safeUnlink } from '../../browse/src/error-handling';
+import { skillCensus, frontmatterName } from './skill-census';
 
 /** Exact env names a hermetic child keeps. Everything not listed (or matched
  * by a prefix rule below) is dropped. */
@@ -225,6 +226,73 @@ export function getHermeticDirs(): HermeticDirs {
   return cachedDirs;
 }
 
+let cachedSkillsConfigDir: string | null = null;
+
+/**
+ * A hermetic CLAUDE_CONFIG_DIR with the repo's shipped skills REGISTERED in
+ * user scope, mirroring ./setup's registration exactly: each discovered skill
+ * gets a REAL directory `<configDir>/skills/<registryName>/` containing a
+ * SYMLINK to that skill's SKILL.md (absolute path), plus a `sections/`
+ * symlink when the skill has one. registryName is the frontmatter `name:`
+ * (dir-name fallback), NO gstack- prefix; the root SKILL.md router registers
+ * as `_gstack-command`. skillCensus().registryEntries is the authoritative
+ * set of what must appear here.
+ *
+ * The default hermetic dir deliberately seeds no skills — correct for
+ * children that install their own or probe setup behavior — but a PTY test
+ * that TYPES `/office-hours` needs the slash command to exist, or claude
+ * rejects it as Unknown command before any model turn and the gate measures
+ * nothing. Separate dir, same runRoot: opt-in per session, never contaminates
+ * the default-config children, and the existing exit teardown + pid-aware GC
+ * cover it. Ends in `/.claude` for the same plan-path anchoring reason as
+ * HermeticDirs.configDir.
+ *
+ * Two intentional non-hermetic edges:
+ * - Seeding reads the LIVE repo tree BY DESIGN — the skills ARE the subject
+ *   under test; a snapshot would measure stale copies.
+ * - HOME is not hermeticized, so the ~64 absolute
+ *   `~/.claude/skills/gstack/...` preamble references inside each SKILL.md
+ *   still resolve to the operator install (same limitation as CI).
+ */
+export function hermeticSkillsConfigDir(): string {
+  if (cachedSkillsConfigDir) return cachedSkillsConfigDir;
+  const { runRoot } = getHermeticDirs();
+  const configDir = path.join(runRoot, 'with-skills', '.claude');
+  const skillsDir = path.join(configDir, 'skills');
+  fs.mkdirSync(skillsDir, { recursive: true });
+  fs.writeFileSync(
+    path.join(configDir, '.claude.json'),
+    JSON.stringify(buildSeedConfig({
+      apiKey: process.env.ANTHROPIC_API_KEY ?? process.env.GSTACK_ANTHROPIC_API_KEY,
+      trustedDirs: [repoRoot()],
+    }), null, 2),
+  );
+  const root = repoRoot();
+  for (const rel of skillCensus(root).physicalSkillFiles) {
+    const skillMd = path.join(root, rel);
+    const skillDir = path.dirname(rel);
+    const registryName = rel === 'SKILL.md'
+      ? '_gstack-command'
+      : frontmatterName(skillMd) || skillDir;
+    const target = path.join(skillsDir, registryName);
+    // Idempotent overwrite mirrors setup's re-link: connect-chrome (a dir
+    // symlink to open-gstack-browser) shares its target's frontmatter name,
+    // so the two walk entries collapse to one registry dir.
+    fs.mkdirSync(target, { recursive: true });
+    safeUnlink(path.join(target, 'SKILL.md'));
+    fs.symlinkSync(skillMd, path.join(target, 'SKILL.md'));
+    if (rel !== 'SKILL.md') {
+      const sections = path.join(root, skillDir, 'sections');
+      if (fs.existsSync(sections)) {
+        safeUnlink(path.join(target, 'sections'));
+        fs.symlinkSync(sections, path.join(target, 'sections'));
+      }
+    }
+  }
+  cachedSkillsConfigDir = configDir;
+  return configDir;
+}
+
 /** A dir younger than this is never GC'd even if its pid looks dead — guards
  * against PID reuse deleting a freshly-created dir whose original pid exited
  * and was recycled to an unrelated live process between create and GC. */