Stub that records each chat_completions_create call and returns a configurable response. Used to drive the detector through pass/ partial/fail

Raw httpx-based client. Returns dicts unchanged from the wire.

Semaphore + global-backoff wrapper around AnthropicClient. - max_concurrent caps simultaneous in-flight requests to the relay. - 429/503 trig

One server-sent event from the streaming Messages API.

Raw httpx client for OpenAI-compatible endpoints.

Raised on non-2xx HTTP response. Holds status + body for inspection.

Raw httpx client for the Gemini OpenAI-compatible Chat Completions endpoint.

One fact embedded in the haystack that the model must recall. `position_pct` is where the sentence is inserted in the haystack (0.0 = start,

Per-protocol stats for one relay domain.

A compact description of an official OpenAI response envelope.

Aggregated stats for one relay domain across all protocols + checks.

Stand-in for the real GeminiClient used by ThrottledGeminiClient. Returns deterministic responses keyed off prompt content and tools so a ful

One row in a per-domain detail page's history table.

Anthropic-compatible runner preserving the v1 constructor.

Semaphore + backoff wrapper. Mirrors the OpenAI throttled client shape. Detectors call .chat_completions_create / .chat_completions_stream; passi

Stand-in for runner.run() return — minimal shape jobs._run reads.

Issues its own API requests.

Observes (request, response) tuples from active detectors and accumulates. Does NOT issue its own requests — zero token cost. Called by Throttled