Known model quirks
Per-model behaviors that the adapters absorb automatically. This file is the public catalog of quirks the static-catalog seeding inside each adapter pre-knows about. Runtime learning catches new ones; this file records what we've already absorbed so users can search before opening duplicate issues.
If you hit a model rejection or behavior not listed here, the adapter's runtime-learning warning will include a pre-filled GitHub issue URL. One click + submit and we add it to the next release.
@llm-ports/adapter-anthropic
Models that reject temperature
Anthropic deprecates temperature per-model on the newer reasoning lineup. The adapter strips the parameter automatically.
| Model pattern | Constraint | First seen |
|---|---|---|
claude-opus-4-5* | temperatureLocked | alpha.3 (2026-05) |
claude-sonnet-4-5* | temperatureLocked | alpha.3 (2026-05) |
To extend: edit KNOWN_TEMPERATURE_REJECTORS in packages/adapter-anthropic/src/capabilities.ts and ship a patch release. Or wait for runtime learning to catch the new model on first use; the adapter retries and learns automatically.
@llm-ports/adapter-openai
Models that reject temperature
OpenAI's reasoning models (o-series, gpt-5-nano) reject custom temperature values. Handled by the same runtime-learning machinery as adapter-anthropic, with OpenAI-specific error classifiers in packages/adapter-openai/src/capabilities.ts.
Detected via isTemperatureRejection matching:
code: "unsupported_value"+param: "temperature"in the OpenAI error response, OR- error message containing
/temperature/i + /(unsupported|not support|does not)/i
No static catalog yet (the adapter learns from the first call). If you want to skip the wasted round-trip, supply pricingOverrides[modelId].capabilities.temperatureLocked = true at adapter construction.
Models with reasoning (hidden chain-of-thought)
Reasoning models burn output tokens on internal reasoning before producing visible text. A naive call with a small maxOutputTokens returns empty visible text + finish_reason=length. The adapter detects this and retries once with a headroom-multiplier-expanded budget.
The static catalog KNOWN_REASONING_MODELS (in packages/adapter-openai/src/capabilities.ts) pre-seeds the learner so the first call against these models already uses the multiplier — no wasted round-trip.
| Model pattern | Provider example | First seen |
|---|---|---|
o1* / o3* / o4* | OpenAI native | alpha.4 (2026-05) |
gpt-5-nano* | OpenAI native | alpha.4 (2026-05) |
gpt-oss-* (case-insensitive) | Cerebras (baseURL=https://api.cerebras.ai/v1) | alpha.4 (2026-05) |
qwen3[._-]?6* (case-insensitive) | Clarifai (baseURL=https://api.clarifai.com/v2/ext/openai/v1); canonical ID Qwen3_6-35B-A3B-FP8 | alpha.4 (2026-05) |
minimax[-_]?m2[._]7* (case-insensitive) | SambaNova (baseURL=https://api.sambanova.ai/v1); canonical ID MiniMax-M2.7 | alpha.4 (2026-05) |
To extend: add a new KnownModelConstraint entry to KNOWN_REASONING_MODELS and ship a patch release. Or wait for runtime learning to catch the new model on first use; the adapter retries and learns automatically. Or supply pricingOverrides[modelId].capabilities.reasoningModel = true per call site.
Models that reject response_format: json_object
Some reasoning models don't support OpenAI's native JSON mode. The adapter falls back to prompted-JSON-plus-Zod-retry automatically.
Detected via isJsonModeRejection in packages/adapter-openai/src/capabilities.ts.
Models that reject a separate system message
Some reasoning models require the system content folded into the first user message. The adapter handles this transparently.
Detected via isSystemMessageRejection in packages/adapter-openai/src/capabilities.ts.
Reasoning-model starvation
Reasoning models (OpenAI o-series, gpt-5-nano, Cerebras gpt-oss-* via baseURL) can spend their entire output-token budget on hidden reasoning, returning empty visible text. The adapter detects this (empty text + finishReason === "length" + tokens consumed + caller-supplied maxOutputTokens) and retries once with REASONING_RETRY_MULTIPLIER × the budget.
@llm-ports/adapter-vercel
Reasoning-model starvation
Mirrors the adapter-openai behavior: empty visible text + finishReason === "length" + tokens consumed + maxOutputTokens set → adapter retries once with 4× budget.
Empty structured-output response
Vercel adapter's generateStructured throws a typed EmptyResponseError (from @llm-ports/core) instead of letting JSON.parse("") raise an uninterpretable SyntaxError. The registry can route on this error type.
@llm-ports/adapter-ollama
No model-specific quirks observed in the wild yet. Ollama runs locally; users control the model lineup. If you hit one, the warning + pre-filled GitHub URL will guide you to file.
How to read this file
Each entry has three things:
- Pattern: regex that matches model IDs the constraint applies to
- Constraint: the
ModelCapabilitiesflag the adapter learns / pre-seeds - First seen: which release added the entry
The static catalog is an optimization. If your model is NOT listed here, the adapter still catches the constraint on first call via runtime learning. The catalog just skips one round-trip on the first call for known cases.
How to add an entry
- Click the URL in the
console.warnyour adapter prints when it learns a constraint. - GitHub opens a New Issue page with title + body + labels pre-filled.
- Review the fields (especially the model ID, in case the warning got a partial match).
- Submit. We'll add the entry to this file + the relevant adapter's static catalog in the next patch release.