feat: app corrections ingestion endpoint for CF menagerie #47

New issue

Closed

opened 2026-05-01 12:12:32 -07:00 by pyr0ball · 1 comment

pyr0ball commented

2026-05-01 12:12:32 -07:00

Owner

Context

Corrections currently come from manual SFT import and benchmark runs. CF apps (Peregrine, Kiwi, Linnet, etc.) should be able to push correction pairs directly to avocet to feed the training flywheel.

Work

Add POST /api/data/corrections/ingest to app/data/corrections.py
Auth: shared secret via AVOCET_INGESTION_SECRET env var (Bearer token, same pattern as cf-orch coordinator auth)
Schema: {source: str, task_type: str, prompt: str, response: str, correction: str, label: str | null}
Batch endpoint: POST /api/data/corrections/ingest/batch accepting array
Store in corrections queue (pending human review before export)
Add AVOCET_INGESTION_SECRET to .env.example
Document integration pattern for product repos

Acceptance

Peregrine can POST a correction pair and it appears in Corrections view for review
Unauthenticated requests return 401
Invalid schema returns 422 with clear error

## Context Corrections currently come from manual SFT import and benchmark runs. CF apps (Peregrine, Kiwi, Linnet, etc.) should be able to push correction pairs directly to avocet to feed the training flywheel. ## Work - Add `POST /api/data/corrections/ingest` to `app/data/corrections.py` - Auth: shared secret via `AVOCET_INGESTION_SECRET` env var (Bearer token, same pattern as cf-orch coordinator auth) - Schema: `{source: str, task_type: str, prompt: str, response: str, correction: str, label: str | null}` - Batch endpoint: `POST /api/data/corrections/ingest/batch` accepting array - Store in corrections queue (pending human review before export) - Add `AVOCET_INGESTION_SECRET` to `.env.example` - Document integration pattern for product repos ## Acceptance - Peregrine can POST a correction pair and it appears in Corrections view for review - Unauthenticated requests return 401 - Invalid schema returns 422 with clear error