Improve Agent Harness

Improve Agent Harness

Use this skill to keep the repository's agent-facing documentation useful, current, and concise.

Scope

• Primary Codex-facing guidance:
  • AGENTS.md
  • .agents/skills/**/SKILL.md
  • .agents/skills/**/agents/openai.yaml
• Historical or tool-specific reference:
  • .claude/skills/**/SKILL.md
  • .claude/rules/**/*.md
  • CLAUDE.md
  • .claude/settings*.json

Workflow

1. Read current guidance before editing:

   sed -n '1,240p' AGENTS.md
   find .agents .claude -maxdepth 4 -type f
   

2. Identify stale assumptions:
   • old test layout names
   • outdated public API names
   • commands that ignore .venv/bin/python
   • duplicated prose that should live in one skill instead of AGENTS.md
   • Claude-only metadata copied into Codex skills
3. Keep AGENTS.md short. Put repeatable procedures into skills and leave AGENTS.md as the routing and project baseline.
4. For skill changes, keep YAML frontmatter to only name and description. Put UI metadata in agents/openai.yaml.
5. Validate changed skills:

   .venv/bin/python /home/b/b36291/.codex/skills/.system/skill-creator/scripts/quick_validate.py .agents/skills/<skill-name>
   

6. If CLAUDE.md intentionally mirrors AGENTS.md, update both. If they diverge by design, state that in the final summary.

Guidelines

• Prefer specific trigger descriptions over long bodies.
• Do not add scripts, references, or assets directories unless the skill needs them.
• Remove stale instructions instead of annotating around them.
• Commit harness-only changes separately from solver implementation changes when the user allows commits.

Source: Nkzono99/vdist-solver-fortran — distributed by TomeVault.