gpt-4o-mini has 128k context — the 200k tier must be skipped, not failed. Otherwise a non-fraudulent OpenAI key would always show false-positive

Default standard mode (only include_long_context=True) keeps hardcoded (32k, 100k, 200k) tiers — backward compatibility for everyone who liked

opus-4-8 必须在册且 thinking 支持开启,否则 thinking_signature 会对最新模型静默跳过(皇冠级检测失灵)。自适应思维,同 4-7。

The exact issue we saw on router.8864k.com: id is a UUID, not msg_*.

Models without a deprecation rule keep their temperature.

Opus 4.7 rejects `temperature` (deprecated). Client must strip it.

Opus 4.8 rejects `temperature` (deprecated). Client must strip it.

Values verified against official docs 2026-05-05 — see _MODEL_CONTEXT_LIMITS docstring for sources. Update both the table and this test togeth

The bundled test PDF must load, look like a PDF, and be non-trivial. We can't grep for MAGIC in raw bytes — reportlab encodes content streams

Auth fail is the ONE upstream error the UI should treat as blocking (red pill) — every other failure path is graceful.

Common on private relays. Frontend renders this as neutral, NOT red, and submission stays enabled.

Even when /models is 404 the response shape must include the new field so the frontend doesn't NPE on data.best_by_protocol[protocol].

A relay that only carries non-preferred SKUs should still get a suggestion (first available) — strictly better UX than null.

When the relay only lists snapshots (e.g. gpt-4o-mini-2024-07-18), pick_default_model should still match via prefix and pick that snapshot.

Don't re-hit the relay if the user just blurs/re-blurs api_key.

Smoke: a model that responds 200 to a 4-token chat call is alive.

A 4xx (model_not_found) must surface the upstream body so the user sees the real reason, not just a generic exception.

best_by_protocol should follow each protocol's pick_default_model ordering, not just the relay's natural sort. For OpenAI that means gpt-4o-mi

Providers like B.AI document a host root but only allow /v1/models.

Some relays serve an HTML landing page at /models but JSON at /v1/models.

The dropdown shows the relay's actual whitelist; preferred (cheap + stable) models must come first so users default-select toward the healthy

We deliberately do not penalize missing anthropic-request-id — the official Anthropic API doesn't return it either (empirically verified).

UUID-form id is technically a non-empty string, so ProtocolDetector should not flag it. (MessageIDDetector handles prefix conventions.)

Sanity check the formula: prior_value=50, prior_weight=5.

Empty relay returns 50 (neutral prior) — sort doesn't crash on it.

Users frequently type `claude-opus-4.7` (dotted) in the web form. The deprecation strip must still kick in — otherwise temperature leaks throu

Replicate a minimal Anthropic-like SSE stream and verify event parsing.

ThrottledClient.total_usage absorbs usage from each non-stream response.

ThrottledClient absorbs message_start input + message_delta output (cumulative final value) across a streamed response.

Bug B fix: passive detectors must observe streaming requests too. ThrottledClient builds a final response dict from message_start + message_de

First call returns 429, second returns 200. ThrottledClient must retry.