feat(swarm): record provider/model on SwarmRun for cost audit + debug

[Teerapat-Vatpitak]

Summary

• Adds two optional fields to SwarmRun — provider and model — and populates them from the active LANGCHAIN_PROVIDER / LANGCHAIN_MODEL_NAME env vars at run-creation time.
• Closes a small but useful observability gap: every .swarm/runs/<id>/run.json written from this point forward records which LLM the run was actually executed against.
• Backward-compatible: existing on-disk run.json files lack the keys and continue to deserialise (default None).

Why

run.json currently has no field describing which provider / model the swarm was launched against. SwarmAgentSpec.model_name carries per-agent overrides but defaults to None ("use the global config"), and that resolved config is never captured anywhere.

This makes three things harder than they should be:

1. Cost auditing — the persisted total_input_tokens / total_output_tokens are meaningless without knowing the model's price-per-token.
2. Side-by-side experiments — comparing kimi-k2.5 vs gpt-4o on the same preset requires manually correlating timestamps with .env history.
3. Post-hoc bug triage — "this run produced garbage" is much faster to diagnose if the run state shows it ran on the wrong model.

Changes

agent/src/swarm/models.py

Two Optional[str] fields on SwarmRun, both defaulting to None:

provider: str | None = None  # lowercased LANGCHAIN_PROVIDER
model: str | None = None     # raw LANGCHAIN_MODEL_NAME (case-sensitive)

The docstring spells out the scoping rule: this is the run-level default; per-agent overrides remain visible on SwarmAgentSpec.model_name.

agent/src/swarm/runtime.py

In SwarmRuntime.start_run, read the same env vars src/providers/llm.py:136,195 reads when constructing ChatLLM, and assign them on the run before _store.create_run persists run.json. Empty / unset env produces None, not the empty string.

run.provider = (os.getenv("LANGCHAIN_PROVIDER") or "").strip().lower() or None
run.model = (os.getenv("LANGCHAIN_MODEL_NAME") or "").strip() or None

agent/tests/test_swarm_run_metadata.py (new)

Six unit tests covering the new behaviour:

• Construction + JSON round-trip with explicit values
• Default-to-None behaviour
• Backward compat: legacy run.json (no provider / model keys) still parses, untouched fields preserved
• Parametrised acceptance of multiple provider names (anthropic, deepseek, openrouter)

Out of scope

• Per-call (per-iteration) model tracking — workers can already override via SwarmAgentSpec.model_name. That field is unchanged.
• Real token counting — worker.py:_estimate_tokens is still a character-count heuristic. That's a separate gap worth addressing but it doesn't depend on this PR.
• Frontend display — the new fields are now in run.json; surfacing them in the UI is a follow-up.

Test plan

• pytest agent/tests/test_swarm_run_metadata.py -v — 6/6 pass.
• pytest --ignore=agent/tests/e2e_backtest --tb=line -q — full sweep: same baseline pass count + 6 new tests, no regressions.
• pytest agent/tests/test_swarm_preset_inspect.py agent/tests/test_swarm_presets_packaging.py agent/tests/test_models.py -v — adjacent swarm tests still green (17/17).

Checklist

• No changes to protected areas (agent/src/agent/, agent/src/session/, agent/src/providers/).
• No hardcoded values; reuses the existing LANGCHAIN_* env-var contract.
• Code follows CONTRIBUTING.md (Conventional Commit prefix, Google-style docstrings, regression test added).
• Backward-compatible — existing on-disk run.json parses without modification.

[warren618]

Thanks for adding the run-level provider/model metadata. I verified the new serialization/backward-compatibility tests in a combined merge gate, and the change is small and useful for later cost/debug auditing. Merged.