Roadmap¶
Reel is built and shipped in well-scoped milestones. Everything below has been delivered.
Shipped — v0.1.0¶
| Milestone | What landed |
|---|---|
| Foundation | Repo, license, CI matrix on Python 3.11 / 3.12 / 3.13, project scaffolding |
| Core proxy | OpenAI adapter, record / replay / auto modes, JSONL cassette format |
| Streaming | SSE capture with millisecond timing fidelity, --timing realtime / fast / slow=N |
| Multi-provider | Anthropic + Gemini adapters, path-based and explicit prefix routing |
| Smart matching | exact / normalized / ignore-fields / fuzzy (optional sentence-transformers) |
| Redaction | Capture-time secret + PII scrubbing, post-hoc reel redact, pre-commit hook |
| pytest plugin | Auto-registered via pytest11, reel_cassette fixture, @cassette decorator, --reel-mode |
| Analytics CLI | reel inspect / cost / diff / stats / doctor over cassette JSONL |
| Observability | Structured per-request JSON logs (--log-format json) |
| Web inspector | reel ui — local HTMX cassette browser |
| Distribution | PyPI publish via Trusted Publishing, docs site on GitHub Pages, Homebrew formula scaffold |
See the architecture overview for how the pieces fit together.
What's next (not yet committed to a release)¶
These are tracked as ideas; comment on a GitHub issue if you'd use any of them and want to push them up the queue.
- Zero-config onboarding —
reel up/reel down/eval "$(reel env)"so users never set env vars by hand. reel run -- <cmd>— wrapper that auto-spawns the proxy for one subprocess and tears it down on exit.- Embedding-cache mode — transparent dedupe of
/v1/embeddingscalls (model-keyed, cross-session). Pure win, zero risk. - MCP server — exposes
lookup_cassette/report_savings/embed_cachedas tools that agents can wire in deliberately. - TypeScript SDK + vitest plugin
- Auto-instrumentation packages for LangChain / LlamaIndex / AI SDK
- Tool-call multi-turn matching for agent loops
- Snapshot mode — assert response shape rather than exact content
- More provider adapters: Bedrock, Vertex, Azure OpenAI, Mistral, Groq, Cerebras, Together
- Reel Cloud (eventually) — team cassette sharing + dashboards, only if there's demand
Contributing¶
PRs welcome. See CONTRIBUTING.md.