[fidecastro]

The problem

Follow-up to #3. The extra_body clash check ran per request, so a key jevify sets on only one path was refused only when that path ran. logit_bias is set by the equal-bias rung and the probe: a recipe with extra_body: {logit_bias: ...} passed ask and eval (warm and top-k readout never set it) and failed inside probe. Which command failed depended on which request happened to set the key first.

The change

• RESERVED_REQUEST_FIELDS in jevify/adapters/endpoint/client.py: every top-level request field jevify sets, across the readout requests of every dialect and rung, the probe, the tokenize and apply-template routes, and the rerank and embeddings bodies.
• The client refuses a reserved key when it is built, so probe, ask, eval and serve fail the same way, before any request. build_backend is the one path all four take.
• The per-request check from endpoint.extra_body: top-level request fields, so Ollama can turn thinking off #3 stays as the backstop for a field the registry does not know yet.
• Docs: SKILL.md (when the refusal happens; the fields also reach the tokenize and template routes) and the ADR-0004 D2 row.

Tests

• tests/guards/test_request_fields.py: a ratchet guard in the style of the other structure guards. It drives every adapter kind and dialect through the fakes the way an operator does (probe, apply_probe, warm, evaluate each question type) and fails naming any request field the registry lacks. It caught top_n in the rerank body on its first run, a field my read of the adapters had missed.
• test_extra_body_naming_a_field_jevify_sets_is_refused_before_any_request, parametrized over max_tokens (set on every request) and logit_bias (the case above). Replaces the per-request test from endpoint.extra_body: top-level request fields, so Ollama can turn thinking off #3, whose scenario it covers. Both cases were red before the change.
• The fake records tokenize and apply-template bodies in template_requests, apart from requests, so request counts in existing tests are unchanged.

Evidence

• uv run pytest -m "not live and not encoder": 138 passed (137 before).
• ruff check and ruff format --check: clean.
• mypy jevify: 20 errors, the same 20 as on main.
• Removing top_n from the registry makes the guard fail naming top_n and the rerank recipe.
• jevify probe on a recipe with extra_body: {logit_bias: ...} and an unreachable base_url prints endpoint.extra_body may not set logit_bias: jevify sets it and never connects.

Not exercised

• No live server. The registry was verified against the fakes, which cover every route the adapters call; a field a real server would see that the fakes do not is still caught by the per-request backstop.
• Encoder-kind recipes have no endpoint, so nothing to check.

🤖 Generated with Claude Code