| name | harness |
| description | Use when maintaining, reviewing, or improving this repository's agent harness. |
Harness Skill
Use this skill for changes to agent instructions, startup files, state,
verification routing, harness docs, lifecycle handoff, or harness evidence.
Model
RunHaven uses a lightweight five-part harness:
- Instructions:
AGENTS.md plus repo-local .agents/skills/ loaded on demand
- Tools: shell, git, file edits, and
init.sh
- Environment: manifests, lockfiles,
rust-toolchain.toml, and pins.toml
- State:
feature_list.json and current-state.md
- Feedback: focused checks plus
./init.sh when the change needs it
Startup
- Read
references/repo-harness.md.
- Read only the startup files or focused harness docs named by the task.
- Do not bulk-load
docs/harness/.
Rules
- Keep root instructions as a map, not a manual.
- Prefer deleting stale guidance, merging duplicates, or tightening a check
before adding a new harness file.
- Do not require HarnessForge or any optional owner tool for ordinary
contributor verification.
- Do not add new machine-local paths, personal tool mandates, credential
rotation, cloud-cost actions, pushes, PRs, or broad automation. Existing
repo-approved Persona and local Codex evidence paths in
AGENTS.md are the
narrow exception for Codex TUI source-first work.
- Treat structural scores as advisory signals, not proof of agent
effectiveness.
- Record only current state and meaningful verification. Keep raw logs and long
history out of startup files.
Verification
For harness-only edits, normally run:
cargo run --locked --bin runhaven-check-pins
python3 -m json.tool feature_list.json >/dev/null
git diff --check
Add a local Markdown link check when docs links change. Run broader checks only
when the harness change affects runtime commands, dependency pins, security
boundaries, release controls, or startup scripts.