Same scoring philosophy as the OpenAI variant — drop skip tiers from the average so model-limit constraints don't penalize the relay. See prot

Combine per-tier outcomes into one score / status / human summary. Tier entry kinds: - pass/partial/fail: probed with a real request

Flag provider-adapter fingerprints inside OpenAI usage objects. Official OpenAI Chat Completions usage carries `prompt_tokens`, `completion_t

Filled circle with a white check mark inside.

Big ring-style score badge — thick colored border, white interior.

One of the four bottom metric boxes.

One synthetic sentence ~80–180 chars. Mix of templates for variety (~20–45 tokens each) so the haystack converges on the target length in 500–

Try `/models`, then `/v1/models` for providers that require it.

Render a per-sub-check value snippet, e.g. (ns=58, stream=30, diff=28).

Concat all text blocks from an Anthropic Messages content array.

Lightweight .env loader (avoid python-dotenv dep for M1). Project-local .env values override shell environment — matches standard dotenv tool

See OpenAI long_context variant — same logic, same markers plus Anthropic's `overloaded_error` which is also transient and not a truncation si

Curated dropdown — 4 most-tested model names. Free-form input still accepts anything; lookup_model() prefix-matches snapshot/alias forms. We

Return candidate model-list URLs for OpenAI-compatible roots. Most relays ask users to paste a `/v1` API root, so `/models` is enough. Some p

Minimal SSE parser for OpenAI-shape Chat Completions streams. Treats `data: [DONE]` as a normal chunk with `_done=True` so callers can observ

Delegate to the protocol package's pick_default_model. Lazy-imported per call to keep web/probe protocol-agnostic at module load time (and to

Ask Anthropic exactly how many input_tokens the request will be, without sending it. Returns None on any failure (relay doesn't imple

Read each protocol's _PREFERRED_DEFAULTS tuple. Empty on errors.

Sleep ~75s and retry once on rate_limited — see OpenAI variant for full rationale. Anthropic's RPM window is also 60s sliding, so the

Run a tier; on rate_limited, sleep ~75s and retry once. ThrottledClient's built-in retry caps Retry-After at MAX_BACKOFF_S = 30s (fin

Render a side-by-side comparison table to the terminal.

Human-readable report notes for the result page. These are deliberately plain Chinese explanations. The raw detector JSON is useful for debug

Flatten results into the order/labels the result template expects.