# Agent Bundle Fixtures
Compact scenario fixtures for the CCD agent-bundle evaluation.
These files are consumed by the external lab harness at
`/Users/emanuele/devel/ccd-agent-bundle-lab/scripts/run_eval/run_eval.py`,
but the fixture corpus itself lives in `ccd-guide` so the evidence set can be
reviewed alongside the runtime specs.
Each fixture mirrors the subset of runtime-state export that the lab needs plus
a deterministic scoring contract.
Most fixtures keep the compact narrative projection used by the broader planning
evaluation.
Start-specific fixtures may also include:
- `runtime_export.export.bundle_projection`: the committed bundle snapshot used by
the `ccd-start` adapter evaluation
- `start_report`: a compact subset of `ccd --output json start --refresh --path .`
- `start_expectations`: scoring targets for the `ccd-start` adapter workflow
Current corpus goals:
- broaden evidence beyond the original 3 scenarios
- include policy, runtime, session, backlog, and memory tasks
- include one concrete "real workflow" integration scenario via `ccd-start`
- include a dedicated `ccd-start` subset for clean, noisy, and backlog-heavy starts
Example run:
```bash
python3 /Users/emanuele/devel/ccd-agent-bundle-lab/scripts/run_eval/run_eval.py \
--fixture-dir /Users/emanuele/devel/ccd-guide/benches/fixtures/agent_bundle \
--output-dir /tmp/ccd-agent-bundle-eval \
--api-base http://100.92.91.40:1234/v1 \
--model qwen3-coder-next-mlx \
--temperature 0 \
--max-tokens 600 \
--repeat-count 5
```