Results / truncation.reasoning-only
Reasoning exhausts the output limit before an answer
The model spends the output token limit on reasoning and stops without sending answer text.
Bug class: Truncation not reported as truncation
The provider stopped at the output token limit, but the library reports a normal finish, a generic parse error, or a partial result as success.
Agents accept partial output or retry without raising the limit, and structured pipelines ingest incomplete records.
Requirements
| Level | Id | Check | Requirement | Providers |
|---|---|---|---|---|
| MUST | cause-is-truncation | truncation_is_signalled | The caller can tell that the output stopped at the token limit. | All |
| MUST | text-reported | text_matches_truth | The provider sent no answer text, so none may be reported. | All |
| SHOULD | usage-reported | usage_matches_truth | Token usage is reported and matches the provider's counts. | All |
Fixtures
| Provider | Tier | any-llm | litellm | pydantic-ai | tanstack-ai | vercel-ai |
|---|---|---|---|---|---|---|
| anthropic | documented | ✓ PASS | ✓ PASS | ✓ PASS | – UNSUPPORTED | ✓ PASS |
| gemini | recorded | ✓ PASS | ✓ PASS | ✓ PASS | – UNSUPPORTED | ✓ PASS |
| ollama | recorded | ✓ PASS | ✓ PASS | – UNSUPPORTED | – UNSUPPORTED | – UNSUPPORTED |
| ollama-openai | recorded | ✓ PASS | ✓ PASS | ✓ PASS | – UNSUPPORTED | ✓ PASS |
| openai-chat | recorded | ✓ PASS | ✓ PASS | ✓ PASS | – UNSUPPORTED | ✓ PASS |
| openai-responses | recorded | ✓ PASS | ✓ PASS | ✓ PASS | – UNSUPPORTED | ✓ PASS |