Personal agent skills. One tree, every harness.
~/Github/dotagents/skills/<name>/SKILL.md
Shared T0 operational skills (investigate, pnl-review, release-service,
breakglass, ...) live in
T0Trade/agent-skills, not here.
The installer clones it to ~/Github/agent-skills when missing, fast-forwards
a clean main checkout, and links its skills next to these. A name in both
trees is an error.
Install (or refresh) the per-harness links (rebuild runs this too):
nu ~/Github/dotagents/scripts/install-skills.nu| Harness | Link farm |
|---|---|
| Claude | ~/.claude/skills/<name> |
| Codex | ~/.codex/skills/<name> and ~/.codex-2/skills/<name> (plus .system → ~/.codex/system-skills) |
| Grok | ~/.grok/skills/<name> and [skills].paths in ~/.grok/config.toml |
| Antigravity | ~/.gemini/config/skills/<name> and ~/.gemini/antigravity-cli/skills/<name> |
Multi-lab orchestration (review-loop, review-pr, critique-loop,
implement-issue, plan-issue) follows skills/panel-runtime.md.
Shared runtime files live in data/ (reports, teach logs, handoffs,
skill state, goal-loop) and hooks/ (goal-loop Stop hook). Skills
persist there — never under a harness home dir. The installer also
registers that Stop hook on Grok, Codex, and Agy (Claude is already
in ~/.claude/settings.json). All changes land on main.