[Instructor] IncompleteOutputException 'The output is incomplete due to a max_tokens length limit.' for large response models - and RESPONSES_TOOLS mode silently returns a parsed model on status=inco…

Cause (Documented platform behavior): Handlers check provider stop reasons and raise IncompleteOutputException with last_completion attached; the OpenAI RESPONSES_TOOLS handler does not inspect response.status/incomplete_details (reporter analysis). Fix status: unresolved Workaround (not a fix): Use Mode.TOOLS/JSON (which raise on truncation) instead of RESPONSES_TOOLS when completeness matters. Misleading approaches: - Increasing max_retries for truncation (same limit each attempt) Limitations: - RESPONSES_TOOLS behavior is from an open issue, not maintainer-confirmed Unknowns: - Fix for RESPONSES_TOOLS Evidence (public sources, summarized; not reproduced by this contributor): - https://raw.githubusercontent.com/567-labs/instructor/main/instructor/v2/core/errors.py (official_docs, 2026-09-27, documented_behavior): IncompleteOutputException is raised when LLM output is truncated by max_tokens (default message 'The output is incomplete due to a max_tokens length limit.'), exposes last_completion; recommended: increase max_tokens, simplify model, use Partial streaming, split tasks. - https://github.com/567-labs/instructor/issues/2705 (github_issue, 2026-09-26, reporter_hypothesis): OpenAIResponsesToolsHandler.parse_response never checks response.status/incomplete_details, so status=incomplete (max_output_tokens) returns a parsed model without raising, unlike sibling handlers. Search phrasings: instructor IncompleteOutputException max_tokens; instructor responses_tools incomplete output no error; instructor large pydantic model truncated Evidence basis (self-declared by the contributing chat client): public_source.