Frozen outbound request snapshots for every provider adapter, captured offline by tests/test_provider/test_request_goldens.py through tests/test_provider/golden/_harness.py (httpx.MockTransport; no network, no real credentials).

Each <backend>/<kind-or-provider>__<case>.json records the request one adapter builds today: method, url, auth_style (a redacted marker — bearer / x-api-key / none; never the key itself), content_type, and body (the parsed JSON payload), serialized with json.dumps(..., indent=2, sort_keys=True) plus a trailing newline.

Freeze policy

These files are a behavior contract. Any diff in a golden means the outbound request changed — a refactor of the request builders (reasoning-toggle ladder, compat model-id sets, schema keyword strips, URL joins, auth style) must leave every file byte-identical. A failing golden is never "fixed" by blind regeneration: treat the diff as a deliberate behavior change that needs review.

Regenerating

OPENSQUILLA_REGEN_GOLDENS=1 uv run pytest tests/test_provider/test_request_goldens.py -q
uv run pytest tests/test_provider/test_request_goldens.py -q   # must pass afterwards

Regeneration rewrites files in place; renamed or removed cases leave stale files behind that test_golden_tree_matches_case_matrix flags — delete them by hand. test_matrix_covers_every_compat_policy_kind fails when a new provider kind is added to _POLICIES_BY_KIND until its goldens land.

Scope notes

openai_codex is excluded: its OAuth flow (Codex CLI credentials) is out of scope for offline request capture.
All data is synthetic: fake key sk-test-000 (asserted absent from every golden), dummy messages, and a synthetic base URL for kinds registered without a default one.