# Agent Instructions

This repository uses the GH Token Furnace Harness.

## Required Flow

1. Work from a GitHub issue or an explicit local job contract.
2. Use a dedicated branch. Do not work directly on `main` or `master`.
3. Keep changes inside the issue scope.
4. Prefer reading, analysis, tests, and documentation over broad code mutation.
5. Produce reviewable artifacts before asking for owner decisions.

## Allowed Work

- Read code and docs extensively.
- Run tests and verifier commands.
- Create audit, test, review, decision, and risk artifacts.
- Make small scoped fixes when the issue explicitly permits them.

## Repo-Specific Boundaries

- Treat `tower.yaml`, `focus.md`, `decisions.md`, and `runs/` as human-owned operator state unless an issue explicitly allows touching them.
- Store review artifacts under `artifacts/issue-<id>/` so parallel agent runs do not collide.

## Forbidden Work

- No open-ended feature building.
- No large rewrites.
- No new dependencies without owner approval and a documented alternative.
- No public API, auth, secrets, deployment, production config, or migration changes without a human gate.
- No architecture changes without a decision log.

## Required End State

Every agent run should leave an artifact directory such as `artifacts/issue-123/` containing:

- `AGENT_RESULT.md`
- Diff summary
- Test or verifier output
- Open risks
- Concrete follow-up issues

Stop and document `NEEDS_HUMAN` if the scope is unclear, a dependency or architecture decision is needed, more than 300 LOC would change, or behavior cannot be verified safely.