fix(loop): enrich trace entries with call_id for tool-call correlation

[zwrong]

Summary

• Add call_id to tool_call and tool_result trace entries, enabling clear correlation between tool calls and their results in trace.jsonl via call_id
• Remove 200-char arg truncation in trace writes for better post-hoc debugging

Why

When replaying agent sessions from trace.jsonl, there is currently no way to correlate a tool_call entry with its corresponding tool_result. The existing iter field is not sufficient because _execute_parallel runs multiple tool calls within the same iteration:

iteration=3:
  tool_call   { iter: 3, tool: "web_search", call_id: "call_abc" }
  tool_call   { iter: 3, tool: "read_file",  call_id: "call_def" }
  tool_call   { iter: 3, tool: "web_search", call_id: "call_ghi" }   ← same iter, same tool
  tool_result { iter: 3, tool: "read_file",  call_id: "call_def" }
  tool_result { iter: 3, tool: "web_search", call_id: "call_abc" }
  tool_result { iter: 3, tool: "web_search", call_id: "call_ghi" }

In this scenario, iter alone (or even iter + tool) cannot distinguish the two web_search calls. call_id (the LLM-assigned tool-call identifier) provides a 1:1 join key for request ↔ response correlation regardless of parallelism or duplicate tool names.

Separately, the 200-char arg truncation in trace makes post-hoc debugging difficult for tools with large payloads (e.g. code generation).

Changes

agent/src/agent/loop.py (3 lines changed):

Location	Change
_execute_parallel (L749)	Add call_id: tc.id to trace, remove [:200] on args
_execute_single (L794)	Same as above
_finalize_tool_result (L873)	Add call_id: tc.id to tool_result trace entry

Before

trace.write({"type": "tool_call", ..., "args": {k: str(v)[:200] for ...}})
trace.write({"type": "tool_result", ..., "preview": result[:200]})

After

trace.write({"type": "tool_call", ..., "call_id": tc.id, "args": {k: str(v) for ...}})
trace.write({"type": "tool_result", ..., "call_id": tc.id, "preview": result[:200]})

Note: black reformatting of loop.py is a pre-existing issue on main, not addressed in this PR to keep the diff minimal.

Test Plan

• Existing tests pass (pytest --ignore=agent/tests/e2e_backtest --tb=short -q) — 2845 passed, 2 skipped, 0 failed. The 2 skipped tests are pre-existing (alpha101_096 NaN ratio, TUSHARE_TOKEN not set), unrelated to this PR.
• New tests added (if applicable) — N/A, no new behavior, only trace metadata enrichment
• Tested manually: ran an agent session, inspected trace.jsonl and confirmed call_id appears on both tool_call and tool_result entries, and args are untruncated in trace

Checklist

• No changes to protected areas (src/agent/, src/session/, src/providers/) without prior discussion
• No hardcoded values (API keys, file paths, magic numbers)
• Code follows CONTRIBUTING.md guidelines
• Documentation updated (if user-facing change) — N/A, internal trace format only

[warren618]

Thanks for spotting this, @zwrong — correlating tool_call ↔ tool_result via call_id is a genuinely useful trace improvement. We've landed an equivalent change in abbbab5, crediting you via Co-authored-by. The only tweak from your patch: we kept the existing [:200] truncation on the serialized tool args, since dropping it would write full untruncated inputs into every trace entry and bloat trace files for tools with large arguments. Closing in favor of that commit — thanks again for the contribution! 🙏