Roadmap (repository)

Rendered from docs/roadmap.md at commit b9d69f7299ed. The repository file is authoritative.

Later milestones, each with its own tests before being described as supported:

Native decision-model adapters (Ollama /v1/systemone: clef, nimble, laya) as routing classifiers and as a /v1/decisions-style endpoint.
Audio/video inputs; multimodal embeddings (EmbeddingGemma 2 image/video/audio); OCR pipeline (GLM-OCR with layout stage) and guardrail pipelines (Granite Guardian) as separate results.
Responses API and Anthropic Messages with fail-closed translation (currently explicit 501).
Exact tokenization for more runtimes; Ollama served-context detection improvements.
Managed model lifecycle (keep-alive/unload) under leases and locks; composition with llama-swap.
Cascades/ensembles; distributed scheduling across hosts.
Opt-in adaptive learning from outcomes with versioning, holdout evaluation, privacy controls and rollback.
Optional LLM-judge scorers in the evaluation harness (identity, bias and cost disclosed).
PyPI package and container image publication (requires separate authorization).