# Feature Demo: F005 — Green Agent Wrapper

> **Generated:** 2026-03-28T00:10:42Z
> **Context source:** spec + discovery only (implementation not read)
> **Feature entry:** [FEATURES.json #F005](FEATURES.json)

---

## What This Feature Does

This feature lets you evaluate a policy over many episodes in one call and get structured results back, instead of manually stepping episodes and aggregating outcomes yourself. It is designed to answer practical questions like: “How does policy X perform over 100 episodes?”

From a user perspective, the key value is fast, repeatable comparison. You can use a built-in random baseline, run seeded evaluations for deterministic comparisons, and inspect both aggregate metrics and per-episode outcomes without losing the whole run if one episode fails.

---

## What Is Already Proven

### Verified in This Demo Run

- Public evaluation API imports successfully (`RandomPolicy`, `evaluate`, result types).
- `evaluate(..., n_episodes=0)` returns a valid zero-valued result object.
- Integration/determinism verification tests passed locally against real SQLEnvironment flow (2 passed).
- Progress-callback verification passed locally (1 passed).
- Full F005 evaluation test file passed locally (16 passed).

### Previously Verified Evidence

- `specs/FEATURES.json` records approved verification evidence for F005:
  - Command: `uv run --with pytest pytest tests/test_evaluation.py -v`
  - Result: 16 passed
  - Verifier result: approved
  - Timestamp: 2026-03-28T00:04:03Z
- `specs/F005-IMPLEMENTATION_SPEC.md` Step 2.2 records full-project regression evidence (`116 passed, 1 skipped`) after integration coverage was added.

---

## What Still Needs User Verification

None.

---

## Quickstart / Verification Steps

> Run these commands to see the feature in action:

```bash
uv sync
uv run python -c "from evaluation import evaluate; r=evaluate(None, None, n_episodes=0); print(r)"
uv run --with pytest pytest tests/test_evaluation.py -v
```

Prerequisite: run from project root with dependencies available via `uv`.

---

## Live Local Proof

### Load the evaluation API

This confirms the user-facing evaluation surface is available from the package.

```bash
uv run python -c "from evaluation import RandomPolicy, evaluate, EpisodeResult, EvaluationResult; print('evaluation_api_import_ok')"
```

```
evaluation_api_import_ok
```

Notice that all primary public symbols for F005 import cleanly.

### Run evaluate() in zero-episode mode

This demonstrates a documented boundary behavior of the evaluation call.

```bash
uv run python -c "from evaluation import evaluate; r=evaluate(None, None, n_episodes=0); print(f'n_episodes={r.n_episodes} n_completed={r.n_completed} success_rate={r.success_rate} avg_reward={r.avg_reward} avg_steps={r.avg_steps} episodes={len(r.episodes)}')"
```

```
n_episodes=0 n_completed=0 success_rate=0.0 avg_reward=0.0 avg_steps=0.0 episodes=0
```

Notice that the function returns a clean structured result instead of failing on this edge input.

### Verify real-environment integration and seeded determinism

This checks the core happy-path flow with real environment integration and repeatable seeded behavior.

```bash
uv run --with pytest pytest tests/test_evaluation.py -v -k "test_evaluate_integration_with_sql_environment or test_evaluate_integration_is_deterministic_with_seeds"
```

```
============================= test session starts ==============================
platform darwin -- Python 3.12.3, pytest-9.0.2, pluggy-1.6.0 -- /Users/hjerp/.cache/uv/builds-v0/.tmpxjssag/bin/python
cachedir: .pytest_cache
rootdir: /Users/hjerp/Projects/sql-env-F005-green-agent-wrapper
configfile: pyproject.toml
plugins: anyio-4.13.0
collecting ... collected 16 items / 14 deselected / 2 selected

tests/test_evaluation.py::test_evaluate_integration_with_sql_environment PASSED [ 50%]
tests/test_evaluation.py::test_evaluate_integration_is_deterministic_with_seeds PASSED [100%]

======================= 2 passed, 14 deselected in 4.29s =======================
```

Notice both integration behavior and seed determinism passed in this run.

---

## Existing Evidence

- Verification spec reference: `specs/F005-VERIFICATION_SPEC.md`
- Implementation-step evidence: `specs/F005-IMPLEMENTATION_SPEC.md` (Step 2.2)
- Feature registry evidence: `specs/FEATURES.json` → `features[id=F005].verification_evidence`

---

## Manual Verification Checklist

No additional manual verification required.

---

## Edge Cases Exercised

### Zero and negative episode counts

```bash
uv run --with pytest pytest tests/test_evaluation.py -v -k "test_evaluate_negative_episodes_raises or test_evaluate_zero_episodes_returns_zero_values"
```

```
============================= test session starts ==============================
platform darwin -- Python 3.12.3, pytest-9.0.2, pluggy-1.6.0 -- /Users/hjerp/.cache/uv/builds-v0/.tmpBSdLqD/bin/python
cachedir: .pytest_cache
rootdir: /Users/hjerp/Projects/sql-env-F005-green-agent-wrapper
configfile: pyproject.toml
plugins: anyio-4.13.0
collecting ... collected 16 items / 14 deselected / 2 selected

tests/test_evaluation.py::test_evaluate_zero_episodes_returns_zero_values PASSED [ 50%]
tests/test_evaluation.py::test_evaluate_negative_episodes_raises PASSED  [100%]

======================= 2 passed, 14 deselected in 4.02s =======================
```

This matters because F005 must handle both boundary (`0`) and invalid (`-1`) episode requests predictably.

### Progress callback behavior during evaluation

```bash
uv run --with pytest pytest tests/test_evaluation.py -v -k "test_evaluate_progress_callback_receives_episode_progress"
```

```
============================= test session starts ==============================
platform darwin -- Python 3.12.3, pytest-9.0.2, pluggy-1.6.0 -- /Users/hjerp/.cache/uv/builds-v0/.tmp164LzQ/bin/python
cachedir: .pytest_cache
rootdir: /Users/hjerp/Projects/sql-env-F005-green-agent-wrapper
configfile: pyproject.toml
plugins: anyio-4.13.0
collecting ... collected 16 items / 15 deselected / 1 selected

tests/test_evaluation.py::test_evaluate_progress_callback_receives_episode_progress PASSED [100%]

======================= 1 passed, 15 deselected in 3.78s =======================
```

This matters because progress visibility was an explicit anti-frustration requirement.

---

## Test Evidence (Optional)

> Supplementary proof that the feature works correctly across all scenarios.
> The Live Demo section above shows usage surfaces; this section shows broader verification coverage.

| Test Suite | Tests | Status |
|---|---|---|
| F005 evaluation tests (`tests/test_evaluation.py`) | 16 | All passed |

Representative command:

```bash
uv run --with pytest pytest tests/test_evaluation.py -v
```

Representative output summary:

```
============================== 16 passed in 4.05s ==============================
```

---

## Feature Links

- Implementation spec: `specs/F005-IMPLEMENTATION_SPEC.md`
- Verification spec: `specs/F005-VERIFICATION_SPEC.md`

---

*Demo generated by `feature-demo` agent. Re-run with `/feature-demo F005` to refresh.*