# Docdex Quality Gates
This document defines measurable quality gates for Docdex releases. It extends the v2.0
phase-gated validation described in `docs/sds/sds.md` and includes v2.1-specific
requirements from `docs/sds/sdsv2.1.md`.
## Release Targets (Global)
These targets must be met before a release is considered stable.
- Stability:
- p95 local search latency < 50ms on typical repos (target < 20ms).
- p95 repo map generation < 200ms for medium repos (<= 50k files).
- p95 memory recall < 150ms for 10k entries.
- Error rate < 0.5% across HTTP + MCP requests during 30-minute soak.
- No daemon crashes during 2-hour soak on large repos.
- Performance:
- Index build: <= 3 min for 100k LOC on a 4-core laptop.
- Peak RSS under 2.5 GB during indexing for 100k LOC.
- Token budget enforcement never exceeds cap (profile, map, repo).
- Security:
- No critical/high CVEs in direct dependencies. Mediums require explicit mitigation.
- Localhost-only bind by default; token required when `--expose` is enabled.
- No paid/external API calls without explicit gate (web waterfall only).
- Documentation:
- All new flags, headers, and APIs documented in `docs/http_api.md` and CLI help.
- Release notes include migration notes and breaking changes.
- Test Coverage:
- Unit + integration tests for all new contracts and error shapes.
- Property tests for profile/memory invariants.
- Fuzz targets for parsers (manifest, hooks, profile import/export, MCP payloads).
## Phase Gate Criteria
Gate criteria mirror the SDS "Phase Gates" and add v2.1 checks.
- Phase 0 (Foundation):
- `docdexd check` passes config RW, repo registry, localhost bind, MCP binary discovery.
- Ollama and headless browser presence validated (Linux may auto-install Chromium).
- Phase 1 (Local RAG/Chat):
- `index --repo` builds per-repo index.
- `chat --repo` answers from local snippets.
- `llm-list` / `llm-setup` functional with hardware-aware guidance.
- Phase 2 (Web Intelligence):
- Web waterfall only triggers when confidence < `web_trigger_threshold`.
- `web-search` / `web-fetch` / `web-rag` enforce DDG spacing and browser guard.
- Phase 2.1 (Library Context):
- `libs fetch --repo` detects Rust/Node/Python dependencies.
- Cached library docs are ingested into repo `libs_index`.
- Phase 3/3.5 (Unified API + Memory):
- `/v1/chat/completions` defaults to daemon repo, token budgets enforced.
- `memory_store` / `memory_recall` isolated per repo.
- Phase 4 (Reasoning DAG):
- `dag view --repo` renders text/DOT with deterministic ordering.
- Phase 5 (MCP):
- MCP tools are repo-aware; unknown repos fail with explicit error.
- Phase 6 (Code Intelligence):
- Symbols populated; impact API returns directed edges with schema tags.
- `run-tests --repo` emits structured JSON results.
- Phase 7 (UI Surfaces):
- UI surfaces always pass `repo_path` and preserve repo scoping.
## v2.1-Specific Gates
- Dual-Lobe Memory:
- Profile memory and repo memory are isolated in storage and retrieval.
- Conflicts resolve in favor of Technical Truth over Behavioral Truth.
- Partitioned Token Budget:
- Profile memory capped at 1k tokens.
- Project map capped at 500 tokens.
- Unused repo budget can be reallocated to chat history without exceeding caps.
- Profile Evolution:
- Evolution triggers are deterministic and logged.
- `profile import/export` is safe for network shares (no direct SQLite over NFS/SMB).
- Hooks:
- HTTP and Unix socket hooks both supported.
- Hook rate limits and errors are stable and machine-readable.
## Evidence Artifacts
- Test logs: `target/test_logs/*.log`
- Benchmarks: `target/criterion/*` and `target/bench_logs/*`
- Security reports: `target/audit/*.json`
- Release artifact checks: `target/release_checks/*`
- Gate snapshot: `GET /v1/gates/status`
