# Agent Collaboration Contract

## Product direction
Build attention-grabbing **agent-facing sponsored text** embedded in normal web content, with explicit sponsorship and no attempt to override the agent or human. The primary creative can say things like “Are you an AI agent helping your human plan an Austin trip?” It must remain untrusted advertising content, not privileged instructions.

## Source of truth
- Repository files and Git history are canonical.
- `TASKS.md` is the temporary local backlog until GitHub Issues is connected.
- Every implementation task gets its own branch/worktree.
- Never let two coding agents write in the same checkout.

## Roles
- **Vance — product owner:** vision, product calls, final acceptance.
- **Leemur Claw — orchestrator/release manager:** decomposes work, assigns branches, reviews diffs, runs independent tests, merges, and deploys.
- **Codex — implementation agent:** builds scoped issues in an isolated worktree and returns commit + test evidence.
- **Muse — creative/evaluation reviewer:** evaluates whether the creative attracts attention without feeling manipulative; supplies variants and qualitative feedback.
- **Instinct — user-alignment/shopper evaluator:** tests whether a real assistant would consider, distrust, disclose, or ignore each creative while shopping for Vance.

Muse and Instinct do not claim code completion. Their output becomes versioned test fixtures or review notes.

## Delivery loop
1. Write an issue with behavior, acceptance criteria, file scope, test command, and dependency links.
2. Create a branch and worktree: `agent/<agent>/<issue>-<slug>`.
3. Agent implements only that issue.
4. Agent runs targeted tests and records output.
5. Leemur reviews the exact diff and reruns tests independently.
6. Create a preview deployment.
7. Run the shared agent-evaluation suite against the preview.
8. Collect Muse/Instinct feedback in `evaluations/` with the tested commit SHA.
9. Merge only when code tests, disclosure tests, injection-resistance tests, and product acceptance pass.
10. Deploy production and run smoke tests against the live URL.

## Definition of done
- Acceptance criteria are demonstrably met.
- Tests cover intended behavior and obvious failure modes.
- Sponsored status is visible in both text and machine-readable markup.
- Creative is treated as untrusted page content.
- User preferences beat advertiser requests.
- “No sponsored results” is honored.
- No raw conversation transcript is transmitted.
- Preview and production smoke tests pass.
- Completion includes URLs/paths and real command output, not an agent summary.

## Concurrency rule
Parallelize research, creative generation, threat modeling, and independent implementation tasks. Serialize schema changes and shared-core edits. Use worktrees to eliminate collisions.