# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Commands

```bash
# Install
pip install -e ".[dev]"

# Run dev server (port 8001)
uvicorn repo_registry.web_api.app:app --reload --port 8001

# Run tests
pytest
pytest -k "test_scanner"        # filter by keyword
pytest tests/test_web_api.py    # single file

# Health check
curl http://127.0.0.1:8001/health
```

Note: `AGENTS.md` shows `src.repo_registry.app:app` but the correct module path is `repo_registry.web_api.app:app` (as installed via `src/`).

## Architecture

The service maps Git repositories to reviewable scope maps using a fixed hierarchy:

```
Scope → Ability → Capability → Feature → Evidence → ObservedFact
```

**Data flow for an analysis run:**

1. `POST /repos/{id}/analysis-runs` triggers the pipeline in `RegistryService.run_analysis()`
2. `GitIngestionService` clones or resolves the repo path
3. `RepositoryMetadataExtractor` reads pyproject.toml / package.json / README
4. `DeterministicScanner` produces `ObservedFact` objects (files, languages, manifests, APIs, etc.)
5. `ContentExtractor` chunks files into searchable segments
6. `CandidateGraphGenerator` builds a draft ability→capability→feature→evidence tree from facts
7. Optionally, `LLMCandidateExtractor` proposes additional candidates (requires `REPO_REGISTRY_LLM_ENABLED=true`)
8. Candidates are stored; humans or agents review them via `POST .../candidate-graph/approve`
9. Approved characteristics feed `ScopeGenerator` to produce `SCOPE.md`

**Key source locations:**

| Component | Path |
|-----------|------|
| FastAPI routes + DI | `src/repo_registry/web_api/app.py` |
| Orchestration | `src/repo_registry/core/service.py` |
| Frozen dataclasses | `src/repo_registry/core/models.py` |
| Deterministic scanner | `src/repo_registry/repo_scanning/scanner.py` |
| Candidate graph builder | `src/repo_registry/candidate_graph/generator.py` |
| SQLite store | `src/repo_registry/storage/sqlite.py` |
| Schema migration | `migrations/0001_initial.sql` |

**Storage:** SQLite at `var/repo-registry.sqlite3` (auto-created). Schema migrations run at startup. Dynamic columns are added to support evidence relationships, classification, and expectation gaps.

**LLM extraction** is optional and disabled by default. Enable with `REPO_REGISTRY_LLM_ENABLED=true` plus `REPO_REGISTRY_LLM_PROVIDER` and `REPO_REGISTRY_LLM_MODEL`. The `llm-connect` sibling package provides the adapter abstraction.

**Semantic search** uses `HashingEmbeddingProvider` by default — deterministic, no external service required.

## Environment Variables

| Variable | Default | Purpose |
|----------|---------|---------|
| `REPO_REGISTRY_DATABASE_PATH` | `var/repo-registry.sqlite3` | SQLite file |
| `REPO_REGISTRY_CHECKOUT_ROOT` | `var/checkouts` | Git clone cache |
| `REPO_REGISTRY_LLM_ENABLED` | `true` | Enable LLM extraction |
| `REPO_REGISTRY_LLM_PROVIDER` | — | e.g. `gemini`, `anthropic` |
| `REPO_REGISTRY_LLM_MODEL` | — | e.g. `gemini-2.5-flash` |
| `REPO_REGISTRY_STATE_HUB_BASE_URL` | `http://127.0.0.1:8000` | State Hub for coordination |

## State Hub & Workplans

Active work is tracked in `workplans/RREG-WP-*.md` — these files are the source of truth (ADR-001). The Custodian State Hub caches this state; workplan files take precedence.

Session protocol (see `AGENTS.md` for full curl examples):
- **Start:** check `workplans/` status headers and State Hub inbox
- **Close:** update task statuses in workplan files, then `POST /progress/` and sync via `POST /repos/repo-scoping/sync`

Workplan sync warns on C-17 (unpushed commits) — that's normal. A `"result": "fail"` needs investigation.

## Docs

Design decisions and terminology live in `docs/`:
- `docs/terminology.md` — characteristic model definitions
- `docs/scope-md-spec.md` — SCOPE.md format
- `docs/characteristic-evidence-model.md` — evidence target kinds
- `docs/classification-strategy.md` — how characteristics are classified