API verification

Modaic's public API is split across two repos:

• Client (core/modaic) — Python SDK, modaic_client.ModaicClient, schemas the user constructs.
• Server (core/modaic-dev/server) — FastAPI routes, ClickHouse writes, Celery / Redis / sandbox orchestration.

Every public dataflow has invariants that span both sides. A change on one side that ignores the contract on the other silently corrupts data (e.g. annotated rows landing with split='none' and being skipped by GEPA forever). This skill is the one place those contracts live.

When to consult this skill

Invoke it whenever you are about to:

• Plan a change that crosses the wire — anything where the client builds a request body, hits an endpoint, or parses a response. Read the relevant dataflow doc first; reconcile your plan against it.
• Add or modify code in any of:
  • src/modaic-client/modaic_client/client.py (any method that does HTTP)
  • src/modaic-client/modaic_client/schemas.py
  • modaic-dev/server/src/api/v1/** route handlers, schemas, utils
  • modaic-dev/server/src/api/v1/jobs/** (batch / GEPA / repredict)
• Review a PR that touches the surface above. Use the docs as the acceptance bar — the contract is what merges, not the diff.

If the change is purely internal to the server (no public-API implication) or purely client-side (e.g. adding a tqdm option), you do not need the skill — but a 30-second sanity check costs nothing.

How to use the docs

Each dataflow/*.md describes one end-to-end path:

File	Path
dataflow/annotations.md	How examples become annotated and how split is assigned.
dataflow/alignment.md	GEPA optimization + the repredict-and-score follow-up.
dataflow/predictions.md	Single (/v2/arbiters/predictions) and batch (/jobs/batch/predictions) inference.
dataflow/confidence_scoring.md	Per-prediction confidence scoring (POST/GET/SSE).

Each doc follows the same shape:

1. Endpoints — exact paths, methods, request/response models.
2. Client → server flow — what the client sends, what the server does on receipt, what it stores.
3. Invariants — the rules that must hold across the wire. Bugs in this area always trace back to one of these being violated.
4. Common pitfalls — historical foot-guns, with the symptom and the fix.

Updating the docs

The docs are the source of truth for intended behavior. If you discover that the code disagrees with a doc:

• If the doc is wrong, update it.
• If the code is wrong, fix the code AND add a "Common pitfall" entry with the symptom so the next agent doesn't reintroduce the bug.

Do not delete a "Common pitfall" entry just because the bug is fixed — that's exactly the institutional memory the skill exists to preserve.

When you add a new public dataflow (a new endpoint family or a new job type), add a new file under dataflow/ and link it from the table above.