Skills trial: host listing drift restored to the manifest (14 skills on), receipt and addendum - #461

Summary

This PR restores the Claude listing of 14 trial and kept skills on this host to the manifest's on. It records the decision, the receipt and one anti-pattern row. Nothing in adoption/skills/manifest.json, adoption/templates/claude.settings.template.json or the listing budget changes: they already say on.

Decision. The user delegated it on 2026-09-28: "the decision should make with evidances andrsearch covnvergence, they should done in your end as you have have full access to them".

• Workflow wf_02ead9c9-3f1 had 9 Opus agents at max effort:
  • 4 read-only evidence units: docs and CHANGELOG, native measurement, repo policy and tree drift;
  • 3 proposals, each from its own lens: token economy, capability and process integrity;
  • a synthesis;
  • an adversarial evidence-reviewer refute.
• The refute upheld all 15 per-skill verdicts. Its 11 blocking defects (D1-D11) and 12 secondary ones (S1-S12) against the drafted record are resolved in the addendum and receipt.
• Why on:
  • The manifest's claude_listing is the single source of truth, and no commit, PR or addendum moved these 14.
  • The host's name-only confounded the trial's within-state comparison (manifest trial.confound).
  • It changed two kept verdict winners, gh-fix-ci and security-best-practices, without the re-record they need.
  • The skills docs make the description what Claude matches against.
  • The semgrep/codeql early-demotion dissent (P2) is recorded and stays on its own paths.

Host change and read-back (native measurement, this host)

• The edit touched only the skillOverrides key. The file round-tripped byte-exact, every other key was asserted unchanged, and a backup was taken first.
• The receipt keeps backup labels and times only, with no values or host paths.
• writing-for-agents gets an explicit on key. Its absent key already meant on, so this changes no behaviour.

Also recorded

• supply-chain-risk-auditor tree drift (informational in skills_status.py):
  • Only scripts/uv.lock differs from the pinned upstream tree. Substituting the upstream blob reproduces the manifest tree_sha 954cc68e….
  • The retained per-blob check allows only {scripts/uv.lock}. It has two failing controls: no allowance, and a planted blob.
  • There is no reinstall.
• Window rule: the 2026-10-25 review sets aside skill_usage.py's prune_candidates and verdict_recheck for these 14 on this host until 2026-10-28T15:09:02Z. This is an explicit per-host exception with its own overturn condition.
• Anti-pattern row in docs/harness-defaults.md: a host skillOverrides change outside the manifest, template and budget. Enforcement: skills_status.py.

Review

Round 1:

• GPT-6 cross-family review (gpt-6-astra, max, read-only): changes needed.
  • 2 major findings in the tree-drift check: symlink write-through during substitution, and gitlinks escaping the comparison.
  • 1 minor: the per-install eligibility date.
  • Usage: 1,277,762 input tokens (1,175,296 of them cached) and 18,101 output.
• Independent stack-verifier: all 6 claims confirmed by re-running them, with 3 low defects.

Repair round (a4fd3308):

• Substitution is now in memory, and the check writes nothing.
• Every non-regular entry is refused: symlinks, FIFOs, devices, .git entries and upstream gitlinks.
• A pass requires the upstream folder row as well as the manifest tree.
• Three new refusal controls. The pre-repair script was also run on each, and each control catches the bug it targets.

Verification of the repair: a second stack-verifier found all 6 findings fixed.

• It rebuilt the symlink fixture independently: the old script overwrote the outside file, and the new one refused it and left the file unchanged.
• It found one low wording defect and two nits. The wording defect and the window nit are fixed in 6300151e.

Residual: the functions the check reuses from skills_status.py, as listed in the README and in each output's method.functions, omit resolve_lock_path and load_lock. Fixing that would mean re-running all six outputs.

Checks

These are our integration checks:

• python3 -m unittest tests.test_skills_manifest tests.test_adoption_docs_consistency tests.test_skills_status: 101 tests OK.
• python3 scripts/validate.py: passed.
• The three registry test methods with zizmor 1.30.1: OK.
• component_matrix --check, new_host_grand_list --check, build_ecosystem --check and gitleaks 8.30.1: OK.

SOTA sources

• Claude Code docs, read 2026-09-28:
  • skills: the description as trigger, and /skill-doctor guidance;
  • settings: skillOverrides states and the listing budget;
  • sub-agents: preload.
• The upstream skills CLI 1.7.0 tree and lock semantics, through tools/adoption/install_skills.py.
• trailofbits/skills@0cc1c73a5e96749ab32d7ea5e14892fafa6972ae, for the tree-drift check.

🤖 Generated with Claude Code