Skip to content

feat(pumas): consume the Pumas v0.8 typed HTTP API - #2

Open
MrScripty wants to merge 2 commits into
mainfrom
feat/pumas-v080-api
Open

MrScripty wants to merge 2 commits into
mainfrom
feat/pumas-v080-api

Conversation

@MrScripty

Copy link
Copy Markdown
Owner

Adopts the Pumas v0.8 consumer API.

  • Discovers embedding and decision models through GET /v1/served-models instead of hard-coded configuration.
  • Adds EmbeddingGemma query/document prefixes.
  • Adds a native decision_native_v1 adapter for Kev through the Pumas decision capability. It keeps complete native probability vectors for Brier scoring and labels them uncalibrated.
  • Ignores response fields, capabilities and error codes it does not use, and classifies failures by outcome, showing the Pumas message.

Tests: 53/53 pass locally.

Pumas side: MrScripty/Pumas-Library#76.

🤖 Generated with Claude Code

MrScripty and others added 2 commits November 2, 2026 09:37
- Discover each role's served alias from GET /v1/served-models; the profile
  is sent only when Pumas reports profile_required or the operator pins it,
  and ambiguous aliases fail with the listed profiles.
- Parse responses tolerantly: unknown fields, descriptors, availability
  states, formats, option names, error codes and list entries are ignored;
  capabilities are selected by name and only used fields are validated.
- Examine typed errors before correlation: not_admitted rejections (null
  request_id, empty-body 400/412/503 fence refusals, unknown codes) are no
  longer reported as uncertain, and Pumas' error.message is surfaced.
- Add configurable embedding query/document prefixes (EmbeddingGemma 2),
  applied to every query and stored text and recorded in the space identity.
- Add the decision_native_v1 adapter on Pumas' native decision capability
  (Kev): three choice questions over the app's labels, relationship
  confidence and per-question probabilities retained as uncalibrated
  evidence. The path follows what the model advertises; a mismatch with the
  operator qualification fails without falling back between adapters.
- Retain complete native relationship probabilities for Brier scoring.
- Simulators mirror v0.8 shapes and cover discovery, prefixes, not_admitted
  errors, decision results and adapter mismatches.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Add config.models.example.json with EmbeddingGemma 2 task prefixes and a
Kev decision role, and describe served-model discovery, tolerant parsing,
not_admitted/unknown error reporting, prefix space identity and the
decision_native_v1 qualification in the README and docs.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@coderabbitai

coderabbitai Bot commented Nov 3, 2026

Copy link
Copy Markdown

Important

  • Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

Run configuration
  • Source mode: Auto
  • Repository: Unavailable
  • Branch: Unavailable
  • File: No configuration file from this repository applied
  • Commit: Unavailable
  • Layers: defaults
  • Review profile: CHILL
  • Plan: Advanced
  • Run ID: c80c5673-20aa-41cc-bedb-6d1311327e0d
  • Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

Share

CodeRabbit Verify

  • Run CodeRabbit Verify to test the changes in this pull request.

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