Skip to content

roadmap.md

Latest commit

 

History

History
149 lines (128 loc) · 7.55 KB

roadmap.md

File metadata and controls

149 lines (128 loc) · 7.55 KB
Raw

Roadmap

Where lm15-python is, and what is planned before and after the stable 1.0 release. Dates are intentions, not promises; everything here follows the same discipline as the code — see How lm15 is specified.

Where we are (October 2026)

  • 2.0.0 is the current release (2026-10-31): every usage counter means one thing on every provider (input includes cached tokens, output includes reasoning) and InferencePricing.estimate(usage) is exact or None, with prices for long prompts. Breaking for code that read Anthropic input or Gemini/xAI output counts; see the changelog. On top of 1.5.0 (2026-10-15): bounding boxes, a judgment that answers where something is in a picture, on every provider, on top of 1.4.1 (2026-10-15: a saved cloud login's project and location are used, found and fixed by Rolf Fredheim), on top of 1.4.0 (2026-10-14: Jev reads pictures), on top of 1.3.0 (2026-10-12, never published on its own: the Meta login reaches the Muse Code subscription, Muse Wonder Full included), 1.2.3 (2026-10-10: a structured answer reads through response.text/.json, and a thinking budget alone is enough), 1.2.2 (2026-10-10: a refused key is AuthError on Gemini and xAI too, and a key passed as credential= is never printed), 1.2.1 (2026-10-06: long streamed replies and tools without a description fixed), 1.2.0 (2026-09-30: the Claude Code release as a setting, and Claude's own output ceiling as the default max_tokens), 1.1.0 (2026-09-26: four open-model hosts and a Google Cloud pass) and 1.0.1 (2026-09-25), the first stable one (1.0.0 was used by a June upload that was removed; PyPI never reuses a number). The chat core — canonical types, serde, errors, request building, response parsing, streaming — is checked against the pinned language-neutral contract, and frozen for 2.x.
  • Non-chat endpoints, stored-cache resources, live sessions and Chat Completions ingest ship as provisional. Contract tests cover recorded behavior; they do not promise that every provider/account works live. The stability boundary below applies throughout 1.x.
  • The model-string router is available. Tool derivation from functions was removed before 1.0 (2026-09-23): a FunctionTool is written out, in every language. The contract's API-family playbook governs the shared public names; an older router-portability proposal is not the authority for the current implementation.
  • Python, Rust and TypeScript pass the shared corpus at their recorded pins. That is evidence of agreement on the recorded cases, not a release approval or a claim about untested runtime behavior.

What ships in 1.0, and what is stable

Already decided: provisional features ship in 1.0, clearly labeled. This follows the ratified contract's spec/SCOPE.md, not a new release-policy decision.

  • Frozen: the canonical chat types, serialization, errors, request and response mapping, streaming, credential resolution and model listing. Removing or changing frozen behavior requires a major release and maintainer ratification. Additions follow the contract's change process.
  • Provisional: non-chat endpoints (files, batches, image/speech/video generation), stored-cache resources, live sessions and Chat Completions ingest. These are included, but may change incompatibly during 1.x with a contract changes/ entry. Additive changes are preferred, not guaranteed. Pin an exact package version and review change entries before upgrading applications that use them.
  • Out of scope: embeddings and canonical provider-executed computer use. The contract describes the limits of provider-specific passthrough; it is not a stability promise for those features.

The stability of a feature is separate from its test coverage. In particular, typed media inside chat belongs to the frozen chat model; the standalone media-generation endpoints are provisional.

After 1.0

1.0.1 was released on 2026-09-25, ahead of the documentation plan it was waiting for, by the maintainer's decision. Still to do, all additive:

  1. Complete documentation site — guides, API reference, specification pages; several guides and the reference are still marked unfinished.
  2. User-experience review pass — read the docs as a new user would. Changes to the frozen chat core must be additive; provisional surfaces may still move.
  3. Scope labels — keep the provisional notices on the relevant guides and release notes.
  4. Release engineering, in place — tag-driven publishing via PyPI trusted publishing (OIDC); CI on Python 3.10–3.14 × Linux, macOS and Windows; a type-checking gate (typecheck/README.md).

Provider coverage

Today, with identical canonical behavior and live-receipt fixtures:

  • OpenAI (Responses API) and OpenAI Codex, including GA Realtime (live) sessions, Sora video, images, and speech
  • Anthropic and Claude Code
  • Google Gemini, including Live (WebSocket) sessions, Veo video, images, and speech
  • xAI (Grok), including subscription OAuth, images, and grok-imagine video
  • Any Chat Completions–compatible server through one dialect adapter with typed compatibility policies — Groq, OpenRouter, DeepSeek, vLLM, SGLang, Ollama

Azure, Bedrock and Vertex already have contract coverage; see cloud hosts for their distinct credentials and settings. Provider/model availability and live verification remain separate from passing recorded cases. Additional providers require evidence-backed fixtures before implementation.

Layers above the foundation

lm15 is deliberately low-level: no automatic tool loop, no retries, no cost ledger, no policy routing (the shipped router is a lookup table — no fallbacks, no ranking). Several companion pieces are under consideration once the foundation's user experience is validated — each as a separate package built on the frozen core, none of them contract-governed:

  • An ergonomic layer — a concise call()-style interface, automatic tool loops, retry/fallback patterns, for people who want three lines and sensible defaults.
  • A model catalog — maintained pricing, context-window, and capability metadata via the entry-point protocol already specified in model-hydration, enabling cost estimation and routing.
  • Recipes — cookbook pages for everything the core deliberately omits (retries, fallback, budget caps, proxying), so each "lm15 doesn't do X" has a one-page answer.

Multi-language

  • Keep Python, Rust and TypeScript aligned with their contract pins and runtime tests. Other ports must meet the same gates before claiming parity; the corpus grows, so use the current reports rather than a historical fixed check count.
  • Publish them (crates.io, Go module, npm) once they pass the full corpus — never before.
  • The promise stays the same in every language: byte-identical wire requests, identical canonical parses, one spec.

Ecosystem and community

  • Integration examples: a FastAPI service, an agent loop, notebooks, and migration guides from other clients.
  • A fixture-first "add a provider" contributor path (see CONTRIBUTING).
  • Benchmarks stay machine-generated and re-run on a schedule — numbers in the README and on this site are never hand-edited.