install_skills: read back a global rollback like #429's project path; M4 host removal receipt

[seathatflowsinourveins]

What

Follow-up to #464 (M4) on #429's convention. On this host, the M4 removal of the trial skill verification-before-completion exposed a gap. The scoped native remove, skills remove <name> -g -y -a claude-code codex, exits 0, yet it leaves the canonical ~/.agents/skills/<name> folder and its lock entry in place whenever another detected universal agent resolves to that folder. Codex keeps reading the folder, so the removal is only partial.

#429 already reads back the project rollback. The global rollback still reported rolled-back without checking.

• tools/adoption/install_skills.py:
  • the global rollback now reads back three things: the canonical folder, the lock entry and the Claude link;
  • it reports Runtime-worker skills program: project-scoped installs and a 137-skill manifest #429's error state with Runtime-worker skills program: project-scoped installs and a 137-skill manifest #429's reason strings ("rollback retained, in use by another agent" / "rollback incomplete");
  • a remove that cannot run now returns error instead of raising;
  • the docstring and the SKILL_AGENTS comment are updated.
• tests/test_install_skills.py: the new cases run on main's fake-CLI retain_after_remove fixtures. Failing-first against bdf25d2: 6 of 52 fail.
• Receipt evidence/receipts/skills-m4-host-removal-20260928.json, with artifacts: the scoped-remove reproduction, the host read-back and the redacted override-drop script. It records:
  • the scoped remove kept the canonical folder and the lock entry;
  • the set was asserted as the only copy before the unscoped remove, which completed;
  • the profile install read-back;
  • skills_status over 28 skills;
  • the dropped skillOverrides key.
• Records:
  • a correction addendum to the trial decision's M4 host step;
  • one anti-pattern row in docs/harness-defaults.md, cross-linked with Runtime-worker skills program: project-scoped installs and a 137-skill manifest #429's rollback row;
  • the receipt registered in manifests/evidence.json.

Review repair round (bb3c88f)

The cross-family review at 6744c2e (codex exec, gpt-6-astra, effort max) returned MERGE-AFTER-FIXES. The independent stack-verifier confirmed all 8 claims. The repair covers every finding:

• Major: the global read-back looked for the Claude link only under ~/.claude. It now resolves it as skills 1.7.0 does: $CLAUDE_CONFIG_DIR/skills when that is set and not blank (dist/cli.mjs L1398, L1511).
• Major: a malformed lock that still held the entry confirmed the rollback. lock_retains now reads the lock strictly.
• Minor: read-back I/O errors raised. Any OSError or ValueError during the read-back is now error, "rollback could not be verified", in both modes.
• Nit: the addendum said "settings". It now says skillOverrides.
• Verifier gap: no test covered a missing binary. Now one does.
• Native checks:
  • m4_claude_config_dir_probe.py: the pinned CLI links only under CLAUDE_CONFIG_DIR. With an unrelated ~/.claude copy present, the pre-repair installer reported a false error; the repaired one reports rolled-back.
  • The unchanged reproduction, re-run with the repaired installer, produced output identical to its first run.

Final round (after the re-check at bb3c88f)

The GPT-6 re-check found the lock, error and nit findings fixed and the CLAUDE_CONFIG_DIR finding partial. A second stack-verifier confirmed 9 of 10 claims and found the same gaps:

• Python str.strip() differs from JS trim(): it keeps U+FEFF and strips U+001C-U+001F and U+0085.
• pathlib keeps .., which Node path.join collapses.
• One line reference was stale.

The fix:

• JS_TRIM_CHARS holds the 25 code points node's trim() removed over every code point. A test compares it with node when node is on PATH.
• The global path is normalized as path.join does.
• The fake CLI mirrors both rules.
• 59 tests pass. Against bb3c88f's installer, 3 subtests fail and 1 test errors.
• The native probe has two new arms. A …/missing/../x value links under x and never creates missing. A U+FEFF-only value falls back to ~/.claude.

SOTA sources

• vercel-labs/skills@7407f389 src/remove.ts: isStillUsed keeps the canonical folder and the lock entry while another detected agent's path is canonical. Installed 1.7.0 dist/cli.mjs L6853-6907; getAgentBaseDir L2214; isUniversalAgent L2180.
• skills@1.7.0 npm dist/cli.mjs L1398 (CLAUDE_CONFIG_DIR?.trim()) and L1511 (join(claudeHome, "skills")); ECMA-262 String.prototype.trim (WhiteSpace and LineTerminator), measured with node v24.21.0.
• openai/codex rust-v0.157.1 codex-rs/ext/skills/src/host_roots.rs L103-108: Codex reads ~/.agents/skills.
• Runtime-worker skills program: project-scoped installs and a 137-skill manifest #429 (bdf25d2): the project read-back convention and the reason strings that are reused here.

Evidence classes

Class	What
our-integration	A1 tests/test_install_skills.py: 59 OK (final-round tests fail against bb3c88f's installer; 5 repair-round tests fail against 6744c2e's; 6 of the original 52 failed against bdf25d2)
our-integration	A2 36 affected modules: 2,049 OK, skipped=43
our-integration	A3 python3 scripts/validate.py: passed, 165 receipts
our-integration	A4 registry tests: 3 OK
native-measurement	the host removal and read-back on this host only; historical for any other host
native-measurement, synthetic-fixture	the scratch-HOME reproduction (both runs) and the CLAUDE_CONFIG_DIR probe

The project-mode native reproduction was not run. The malformed-lock, unreadable-lock and missing-binary cases are fake-CLI fault injections only; the receipt records both limits.

🤖 Generated with Claude Code