Skip to content

feat: v0.8.0 consumer API, desktop contract validation and release version - #76

Merged
MrScripty merged 49 commits into
mainfrom
feat/v080-consumer-api
Nov 3, 2026
Merged

MrScripty merged 49 commits into
mainfrom
feat/v080-consumer-api

Conversation

@MrScripty

@MrScripty MrScripty commented Nov 3, 2026 •

Copy link
Copy Markdown
Owner

The v0.8.0 consumer API: typed model operations that explain every rejection, decision and multimodal-embedding capabilities, served-model discovery, and a desktop contract validated end to end from the backend's declaration. It also merges the previously local fix/main-test-validation work (download recovery, deletion and the installed-file checksum policy) and applies the Dependabot updates from #74 and #75.

Sets the source version to 0.8.0 (manifests, workspace lock entries, 0.8.0 attribution). No tag or release is created; the pre-tag release QA in RELEASING.md follows the merge.

HTTP API (/v1/model-operations, /v1/capabilities)

  • Every rejection carries a human-readable message naming the field, limit, model or unavailable reason. The caller's request_id is echoed whenever it can be read, and the HTTP status matches the error code. Provider failure reasons are forwarded with credential-shaped text redacted.
  • One request grammar. capability, semantic_task and options are optional selectors, and every request accepted before is still accepted.
  • Responses may grow (new fields, capabilities and enum values) within contract version 1. Consumers must ignore what they don't use.
  • GET /v1/served-models lists each model and profile once, with capabilities and context size. /v1/models no longer repeats IDs.
  • No flat limits on the typed path. max_tokens follows the served context, and ONNX batch limits apply only to ONNX. Image decode-safety limits stay.
  • New decision capability: llama.cpp decision models (Kev, OpenJev) through /v1/systemone, offered only when llama.cpp itself reports the model as a decision model.
  • New multimodal_embedding capability: image and mixed text/image items for projector-backed embedders (EmbeddingGemma 2).
  • Vision-language models (Gemma 3/4, Qwen VL) serve chat and image-to-text instead of reporting model_task_mismatch.

Desktop

  • RPC and outcome errors carry their failure reason (bounded and redacted) instead of timing out or reporting a generic failure.
  • The remaining hand-written preload validators are replaced by validators generated from the Rust contract or the Electron-owned schemas:
    • library-upgrade confirmation and model-import selection;
    • the Torch install options, saved probe, startup trial and alternative-wheel responses;
    • the five update streams (model library, downloads, runtime profiles, serving status, status telemetry).
  • An update a subscriber can't accept is reported through a required onError, and the view reloads its RPC snapshot instead of silently dropping the update. Rejections name the failed field and rule, never a field's value.

Merged from fix/main-test-validation

Download recovery, deletion during unrelated downloads, and partial preservation on resume. Also the checksum policy: supplied hashes are verified once at acquisition, and installed runtimes are not rehashed. 38ec86f9's installed-byte rehash for audio is backed out to follow that policy, and 20160752 updates the release test that still required the rehash.

Serving

  • llama.cpp embedding and reranking sessions size their micro-batch from the profile's served context, capped at the model's trained context from its GGUF header; otherwise an 8,192-token startup default applies. The trained context alone can't be the default: EmbeddingGemma 2 and Kev record 262,144 tokens, and a batch that size made llama.cpp try to allocate 202 GB.

Native QA fixes

  • The staged pip install test pins a Linux target, because macOS CPU wheels have no +cpu suffix.
  • The Windows flush test matches the checksum policy: the first verification of a read-only file is denied, and reusing its receipt needs read access only.

Dependencies

Electron 43.8.1, Vite 6.4.5, undici 6.29.1/7.30.1, zip 2.6.1, rustls 0.23.46 and rand 0.10, at the exact versions proposed in #74 and #75. rand 0.10 renames the range helpers to RngExt; the two imports are updated, since #74 did not compile. Both 0.8.0-rc.1 attribution sets are regenerated.

Verification

  • Rust workspace (excluding pumas_rustler): 3,167 tests pass, and fmt and clippy (all features) are clean.
    • owned_audio_route_refuses_unsupported_requests_and_stale_selection_without_replay can exceed its fixed 30 s budget when the whole workspace runs on a loaded machine. It passes on its own and in a full pumas-rpc run (429/429).
  • Frontend and Electron: lint, types, tests and build pass, along with desktop-contract generation, the generator test, conformance, and the launcher tests.
  • Release: dependency ownership, release versions, attribution (base and S3), dependency features, and the release script tests (37/37) pass.
  • Live checks:
    • Kev through Pumas against llama.cpp b11429: decision discovered, and choice, score and yes/no answers returned with normalized probabilities.
    • EmbeddingGemma 2 has not run live. b11429 cannot load its gemma-embedding2 architecture; b11454 is the first published build that can, and the release notes now say so.
  • Not run here: native Windows/macOS packaging, pretrained ONNX/vision models, and the Python sidecar tests that need numpy and uvicorn.

Consumers

Matching branches named feat/pumas-v080-api exist in phased-memory, Tuldok, Lanternwake and Pumas-Playground. They read responses tolerantly, use served-model discovery, and show error messages.

Closes #74, closes #75.

🤖 Generated with Claude Code

Loading
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant