Added
- Compact human and JSON help projections so agents can discover the public verb surface without loading the full command reference.
- Batch search now preserves per-query intent, caller, run-context, source, and retrieval-mode telemetry using the same bounded local log as individual searches.
- Repeatable local
make publish-ready, verification, help-size, and isolated smoke gates for contributors and release preparation.
Measured retrieval
| Benchmark | Queries | Mode | Hit@1 | Hit@5 | MRR | Reproducible by you |
|---|---|---|---|---|---|---|
| Full fixture suite | 258 | HyDE + hybrid + LLM rerank | 69.0% | 92.2% | 0.786 | No — maintainer's private corpus mirror |
| Sample corpus | 24 | BM25 / FTS5 only | 95.8% | 100% | 0.979 | Yes — pinned public pages, no API key |
Reproduce the sample corpus baseline in a few minutes with no account:
corpus="$(mktemp -d)"
docs-puller pull --from eval/sample-corpus/sources.md --out "$corpus"
docs-puller reindex --out "$corpus"
docs-puller eval --fixture eval/sample-corpus/fixture.yaml --out "$corpus" \
--diff eval/sample-corpus/baseline-2026-07-03.jsonVerification
make publish-ready · make help-sizes · make smoke · CI green
Full changelog: v0.4.0...v0.5.0