Agent Skills
pairmux embeds one canonical SKILL.md package. It teaches agents when to use pairmux, the
new -> run -> wait/send/log loop, secret handoff, and how to interpret pairmux.v1 envelopes.
Install
pairmux skill install --target agents --dry-run
pairmux skill install --target opencode
pairmux skill install --target all
all is conservative: it writes only where an agent's configuration directory already exists.
| target | installation directory |
|---|---|
claude-code | ~/.claude/skills/pairmux/ |
codex | ~/.agents/skills/pairmux/ |
gemini | ~/.gemini/skills/pairmux/ |
cursor | project-relative .cursor/skills/pairmux/ |
opencode | ~/.config/opencode/skills/pairmux/ |
copilot | ~/.copilot/skills/pairmux/ |
windsurf | ~/.codeium/windsurf/skills/pairmux/ |
kiro | ~/.kiro/skills/pairmux/ |
amp | ~/.config/amp/skills/pairmux/ |
agents | ~/.agents/skills/pairmux/ (compatibility alias for the shared standard) |
codex and agents intentionally resolve to the same open Agent Skills directory; all writes that
destination once. The companion pairmux-skills
repository is the source of truth and includes the full install map. The binary contains a
release-synchronized embedded copy.
Cross-agent evals
The S01-S10 suite covers finite and slow commands, large logs, confirmations, secret handoff, REPLs, pagers, a dev server, Ctrl-C recovery, and human notes. Its runner isolates every episode in a copied worktree with a unique tmux endpoint and records exact pairmux argv through a PATH proxy. The eval runner is not part of this pairmux checkout: clone the companion repository and run these commands from its root (see its eval README):
git clone https://github.com/treeleaves30760/pairmux-skills.git
cd pairmux-skills
python3 evals/run.py --agent opencode --model opencode/big-pickle --scenario S01-S10 --repeat 3
python3 evals/run.py --agent claude --scenario S01-S09
python3 evals/run.py --agent codex --scenario S01-S06,S08
Each run writes JSONL episode results plus JSON and Markdown summaries with pass rate, step count, wall time, failure class, tool/model versions, and artifact paths. OpenCode Big Pickle is the minimum cross-agent canary; repeated episodes matter because that model alias can change over time.