Phase 2: Deterministic Core

[rmax]

Description

Implement the deterministic core of the harness — all components that don't require a live model. This includes goal contracts, state machine, policy engine, Git repository store, SQLite event store, artifact hashing, evaluator pipeline, fake sandbox, and scripted agent.

Key Deliverables

1. Goal Contract — typed YAML contract parser, validator, JSON Schema, contract hashing
2. State Machine — all states, transition decisions, invalid transition prevention
3. Policy Engine — path policy, command allowlist, dependency policy, approval gates
4. Git Repository Store — checkpoint, restore, diff, branch, worktree management
5. SQLite Event Store — SQLAlchemy models, Alembic migrations, repository pattern
6. Artifact Store — SHA-256 hashing, hash chain, tamper-evident manifest
7. Fake Sandbox — deterministic sandbox for unit tests
8. Scripted Agent — deterministic agent for testing without a model
9. Evaluator Pipeline — tiered evaluation, hard gates, evaluation vectors
10. Budget Tracking — iteration, token, cost, runtime budgets

Acceptance Criteria

• Goal contract parses from YAML and validates against schema
• Invalid contracts return actionable error messages
• State machine prevents invalid transitions by code
• Every transition is persisted before continuing
• Path policy rejects traversal and protected-file edits
• Command policy rejects forbidden commands
• Git store creates checkpoints and restores correctly
• Worktrees are isolated from source repository
• SQLite event store persists and queries events
• Alembic migrations run successfully
• Artifact hashing produces correct SHA-256 chains
• Fake sandbox returns deterministic results
• Scripted agent produces typed proposals
• Evaluator pipeline runs tiered checks with correct ordering
• Budget tracker enforces hard limits
• Ruff, mypy, and all unit tests pass

Depends on: #1, #2

[rmax]

Phase 2 Progress

✅ Completed (3/4 batches)

• Batch 1a: Domain models — contracts, runs, candidates ✅
• Batch 1b: Evaluation, evidence, budget models ✅
• Batch 1c: Approvals, events, state machine ✅ (15 unit tests)

🔄 In Progress

• Batch 2: Policy engine — paths, commands, dependencies, approvals

⏳ Queued

• Batch 3: Repository store, event store, artifact store
• Batch 4: Sandbox, agents, evaluator pipeline

All batches producing clean ruff + mypy + pytest results.

[rmax]

Phase 2 Progress

✅ Completed (4/6 batches)

• Batch 1a: Domain models — contracts, runs, candidates
• Batch 1b: Evaluation, evidence, budget models
• Batch 1c: Approvals, events, state machine
• Batch 2: Policy engine — paths, commands, dependencies (20 tests)

🔄 In Progress

• Batch 3: Repository store, event store, artifact store

⏳ Queued

• Batch 4: Sandbox, agents, evaluator pipeline