Commit Graph

4 Commits

Author SHA1 Message Date
Claire Novotny 72495c1c14 Restructure as Codex plugin 2026-04-23 20:15:02 -04:00
Matt Van Horn e8105df4fd fix: v3.0.9 - engine refuses Class 1 keyword traps, delete stale SKILL.md files, reinforce LAW 1 over WebSearch
Five Opus 4.7 self-debugs on v3.0.8 (3 passing, 2 failing runs) converged
on four fixes:

1. Engine refuses Class 1 demographic-shopping queries at main() front-door.
   Birthday-gift failure mode becomes structurally impossible - the pipeline
   never runs on a doomed query. Exit code 2 with a REFUSE message on stderr
   pointing the model to ask for hobbies/relationship/budget. Escape hatch:
   LAST30DAYS_SKIP_PREFLIGHT=1 for "just run it" overrides.

2. Delete stale `.agents/skills/last30days/SKILL.md` (1382 lines, April 13
   snapshot) and `.hermes-plugin/SKILL.md` (269 lines, April 13 snapshot).
   Peter Steinberger's self-debug named the first file as the one it read
   instead of the real SKILL.md. One SKILL.md per plugin, at the plugin root.
   Sync script simplified: Hermes now always uses main SKILL.md.

3. render_compact() appends an explicit END-OF-CANONICAL-OUTPUT boundary
   with pass-through instruction. The model had the canonical body in its
   buffer on the Peter run and discarded it; the boundary makes pass-through
   the path of least resistance.

4. LAW 1 gains a verbatim-pattern override clause naming the exact WebSearch
   tool-result reminder ("CRITICAL REQUIREMENT: MUST include Sources:
   section") that caused Peter's trailing Sources leak. No more ambiguity
   at synthesis time.

Tests: tests/test_preflight.py, 29 scenarios covering Class 1 matches
(birthday gift, best-for-demographic, what-to-buy-relationship), qualifier
skips (budget, hobbies, activity after year-old), and the REFUSE message
shape.

Validation gate before merging to main: re-run the 5 debug topics
(Peter Steinberger, birthday gift for 40 year old, Kanye West, Garry Tan,
OpenClaw vs Paperclip vs Hermes) on v3.0.9 and confirm 5/5 canonical
compliance. Rollback to v3.0.8 if any previously-passing topic regresses.

Plan: docs/plans/2026-04-18-015-fix-engine-refuse-keyword-traps-delete-stale-skillmd-files-plan.md

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-18 13:30:33 -07:00
shalomma 8b67378964 Bump SKILL.md version header from v2.9.5 to v3.0.0
The SKILL.md prompt header still said v2.9.5 while pyproject.toml
and the rest of the codebase are on v3.0.0.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-04-13 12:31:43 +03:00
Matt Van Horn 71e0492840 feat: make skill discoverable by OpenAI Codex CLI (#219)
Adds a Codex CLI skill integration by creating the two files Codex's real
loader actually reads:

- .agents/skills/last30days/SKILL.md (real file, not a symlink - Codex's
  loader skips symlinked files per codex-rs/core-skills/src/loader.rs)
- .codex-plugin/plugin.json with {"name": "last30days"} as a namespace
  marker, per codex-rs/utils/plugins/src/plugin_namespace.rs

When Codex CLI runs in a checkout of this repo, it walks .agents/skills/
from CWD up to the project root, picks up .agents/skills/last30days/SKILL.md,
and walks ancestors looking for .codex-plugin/plugin.json to resolve the
plugin namespace. The skill registers as last30days:last30days.

The SKILL.md is a verbatim copy of the root SKILL.md at this point to
avoid content drift during the rollout. A future PR can slim the Codex copy
or introduce a sync mechanism.

Verified against Codex CLI's own source by running codex exec from the
repo CWD and having it trace the loader logic.

Replaces PR #153, which used a fake $schema URL
(https://openai.com/codex/plugin.schema.json returns 404) and put a
misunderstanding of Codex's plugin manifest (Codex only reads the `name`
field - all other fields like version, description, author, skills[] are
silently ignored).

This contribution was developed with AI assistance (Codex + Claude Code).

Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
2026-04-11 09:02:28 -04:00