[codex] collect per-generation parsing errors

[Ki-Seki]

Summary

• add error_mode="raise" | "collect" to synchronous, asynchronous, and vLLM offline batch model interfaces
• add GenerationResult to preserve each raw response together with either its parsed Result or serializable error details
• isolate parsing and infill failures per candidate while keeping model request, configuration, and response-container failures as whole-call errors
• preserve existing default behavior and return types when error_mode is omitted
• document the behavior for OpenAI, vLLM server, and vLLM offline backends

Root cause

Multi-candidate and batch responses were passed through list-level infill helpers. A parsing exception from one generated string aborted the conversion of the entire response list, discarding otherwise valid candidates and queries.

Behavior

• error_mode="raise" remains the default and preserves fail-fast behavior.
• error_mode="collect" parses every generated string independently.
• single-query calls preserve the generator's scalar/list shape, replacing Result elements with GenerationResult elements.
• vLLM offline batch always returns a two-dimensional query/candidate structure.
• request failures, invalid arguments, invalid response containers, empty candidate groups, and batch shape mismatches still raise normally.
• existing response repair and non-strict infill behavior is unchanged.

Validation

• uv run pytest — 73 passed
• uv run mypy src — passed
• uv run ruff format src tests — passed
• uv run ruff check src tests — passed
• pre-commit hooks — passed

[github-actions]

PR Preview Action v1.8.1
🚀 View preview at https://SculptAI.github.io/GIMKit/pr-preview/pr-190/
Built to branch gh-pages at 2026-06-19 10:06 UTC. Preview will be ready when the GitHub Pages deployment is complete.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!