Make debug artifacts ATIF-compatible

[mmprotest]

Motivation

• Ensure Villani’s generic debug artifacts contain package and model/provider metadata sufficient for later conversion to an external ATIF/packaging format without adding any benchmark-specific paths or logic.
• Provide a stable separation between API compatibility (OpenAI-compatible request format) and the actual inference provider (e.g., LM Studio) in run metadata.
• Emit a generic ATIF-compatible trajectory artifact at run start (and preserve partial trajectories on failure) using Villani’s existing artifact directory and filenames.

Description

• Add small metadata helpers and populate run/session metadata with the package version and structured model metadata by introducing _agent_metadata, _model_metadata, _api_compatibility, and _inference_provider, and use villani_code.__version__ for the agent version. (files changed: villani_code/debug_recorder.py, villani_code/atif_export.py, villani_code/trace_summary.py)
• Include model_metadata in the generic transcript artifact and add inference_provider + api_compatibility into the ATIF trajectory.json agent extra fields while preserving existing schema fields. (files changed: villani_code/atif_export.py)
• Ensure a Villani-owned generic ATIF-compatible trajectory.json is written for every run that starts successfully by writing a trajectory at run startup and again on model-request failure; this reuses existing debug artifact locations and filenames and does not add any benchmark-specific path or naming. (files changed: villani_code/state.py, villani_code/debug_recorder.py)
• Surface the new agent and model_metadata into summary aggregation by loading session_meta.json when building the run summary, without changing runner behavior. (files changed: villani_code/trace_summary.py)
• Tests adjusted/added to cover the new metadata and trajectory behavior; modified test expectations only (no behavior changes). (files changed: tests/test_regular_trace_artifacts.py, tests/test_debug_mode_and_recorder.py)

Files changed: villani_code/debug_recorder.py, villani_code/atif_export.py, villani_code/state.py, villani_code/trace_summary.py, tests/test_regular_trace_artifacts.py, tests/test_debug_mode_and_recorder.py.

Testing

• Ran pytest tests/test_regular_trace_artifacts.py tests/test_debug_mode_and_recorder.py and the suite passed (24 passed).
• Ran the focused test selection: pytest tests/test_regular_trace_artifacts.py::test_openai_compatible_metadata_uses_lmstudio_inference_provider tests/test_regular_trace_artifacts.py::test_regular_trajectory_exists_after_model_request_failure tests/test_regular_trace_artifacts.py::test_normal_debug_writes_generic_trajectory_only tests/test_debug_mode_and_recorder.py::test_debug_metadata_separates_api_compatibility_from_inference_provider and they passed.
• Verified compiled modules with python -m compileall and checked git diff --check; both succeeded without warnings.

Sample artifact snippets produced by the patch (examples):

• session_meta.json / run-start payload (generic run metadata):

  {
  "agent": { "name": "villani-code", "version": "0.1.0rc1" },
  "model_metadata": { "identifier": "local-model", "inference_provider": "lmstudio", "api_compatibility": "openai" }
  }

• ATIF trajectory.json agent block (extra fields):

  {
  "schema_version": "ATIF-v1.7",
  "agent": {
  "name": "villani-code",
  "version": "0.1.0rc1",
  "model_name": "local-model",
  "extra": {
  "provider": "openai",
  "inference_provider": "lmstudio",
  "api_compatibility": "openai",
  "mode": "regular"
  }
  },
  "extra": { "status": "completed", "termination_reason": "completed", "transcript_artifact": "transcript.full.json" }
  }

• Failed run (partial trajectory) trajectory.json extra fields show failure:

  {
  "schema_version": "ATIF-v1.7",
  "extra": { "status": "failed", "termination_reason": "model_request_failed", "transcript_artifact": "transcript.full.json" },
  "steps": [ { "source": "system", ... }, { "source": "user", ... } ]
  }

Confirmations

• No runner task-solving, planning, repair, validation, tool policy, retry, or execution behavior was changed; only metadata and artifact emission were added or enriched.
• No Terminal-Bench-specific paths, filenames, directory layouts, benchmark branching, or submission-mode logic were introduced; artifacts are written to Villani’s existing debug artifact directory and filenames.
• Patch is intentionally minimal and restricted to metadata enrichment and writing a generic ATIF-compatible trajectory artifact that external packagers can later adapt for submission.

---

Codex Task