Roadmap
Each phase names what it delivers and how it is judged done.
| Phase | Delivers | Status |
|---|---|---|
| L0 | Repository, CI, the llm contract, the usage record, the adkv1 and adkv2 adapter modules | done. adkv1 was deprecated in v0.6.6 and later deleted: both products are on ADK v2 and adkv2 |
| L1 | OpenAI Chat Completions, serving the built-in profiles for Vertex AI partner models, vLLM, SGLang, Ollama and any OpenAI-compatible endpoint; mast adopts it | done: the adapter, coremodels.Open and a recorded corpus (v0.1.0); mast adopted it (mast #510); GLM 5.2 holds tool-calling parity with Claude on mast’s judged corpus over three runs; vLLM on GKE runs end to end (deploy/gke-vllm). See tested models |
| L2 | OpenAI Responses; OpenAI and xAI profiles | done: dialect/openairesp and the openai and xai profiles. Recorded live from OpenAI (gpt-5-mini, including the encrypted-reasoning round trip) and xAI (grok-4.20-0309-reasoning over Chat Completions and Responses), plus Gemma 4 on our vLLM’s /v1/responses |
| L3 | KV-cache statistics from self-hosted servers; core-agent adopts L1–L3 | done: core-agent adopted provider profiles (core-agent #1294); kvmetrics samples vLLM (verified live on 0.31), SGLang and llama.cpp metrics endpoints; mast’s /usage reports a fleet-level prefix-cache window per turn and per session, and its judged eval board per incident (mast #543), live against our vLLM at 85–95% hit rates |
| L4 | Anthropic adapter moves here, with the best of both products’ versions | done: dialect/anthropic and the anthropic and anthropic-vertex profiles (v0.6.0); mast (mast #521) and core-agent (core-agent #1296) both run on it, with judged-eval and live parity against their previous adapters. Claude is recorded live on Anthropic’s first-party API and on Vertex AI |
| L5 | Gemini adapter and Vertex context caching move here | done: dialect/gemini, dialect/gemini/vertexcache and the gemini, vertex and vertex-express profiles (v0.6.0–v0.6.3); mast and core-agent both run on it, with judged-eval and live parity, including web search and the Vertex context cache |
| L6 | One price catalog for both products | done: pricing (v0.11.0) merges both products’ catalogs, with backend-keyed built-in rows, long-context rates, a profile-declared layer and the regen tool. mast prices from it (mast #544), and its own introductory-rate check and reviewed pin bumps keep the safeguards. core-agent prices from it too (core-agent #1328) |
| Later | Claude on Bedrock, Azure OpenAI, more built-in profiles | planned |
The design document has the exit criterion for each phase.
core-models is pre-1.0, so a minor release may break the API. The CHANGELOG says so under API Change. The road to v1.0 is tracked in the v1.0 milestone (tracking issue). The criteria are proposals for the maintainer to decide:
- A deliberate public API, with CI blocking unlabelled breaking changes. See the API surface inventory.
- A decision on the genai/ADK type coupling in the
llmcontract. - One typed error classification across every dialect.
- A versioning rule for price data and pricing behavior.
- Answers to the open questions that shape the API: tiers, price dimensions, governance and retry posture.
- Optionally, a public conformance kit.
- A six-week stability soak, ending in the v1.0.0 tag.
The P1 providers (Claude on Bedrock, Azure OpenAI, more built-in profiles) don’t gate v1.0. They can land in v1.x.