[Gujiassh]

Summary

Closes #295.

LLMAgent.run passed input and trace_id positionally to TaskRecorder. Inherited dataclass fields consumed these values as completion/error state, leaving the returned task, input and trace metadata empty even when the model run succeeded.

• Initialize the existing task/input/trace_id fields by keyword.
• Preserve string tasks and original response-item lists; list inputs retain the existing empty task default without guessing a task string.
• Add actual Runner regressions for string/list input and supplied/generated trace IDs, including output, trajectory and pure trajectory-model conversion.

No API, schema, persistence or streaming changes.

Validation

Python3.12, real LLMAgent/Runner with a local SDK Model; no remote provider or database writes:

• Four new regressions fail on main and pass after the fix; 10 bounded agent/common/parser tests pass.
• Standalone real-run reproduction now returns the correct metadata and legal inherited state while preserving the answer/trajectory.
• Changed-file Ruff and format checks pass.
• Focused mypy has 4 pre-existing implicit-Optional errors; exact-main comparison has those same 4 plus 2 incorrect-constructor diagnostics. No new test/changed-line errors.

Full suite, live providers, database persistence and streaming were not validated.