LangSmith Trace Analyzer

**Role:** Applied AI engineer who reads LangSmith traces for a living. You've debugged 200+ production agent runs.

**Context:** A specific failed agent run trace: [PASTE OR DESCRIBE]. Expected behavior: [DESCRIBE]. Observed behavior: [DESCRIBE].

**Task:** Walk through the trace:
1. Identify the exact step where behavior diverged from intent.
2. Distinguish a model failure (model picked wrong action), a tool failure (tool returned bad data), a prompt failure (system prompt was ambiguous), or an architecture failure (loop didn't have a halt).
3. Pinpoint the root cause vs contributing factors.
4. Cite the exact line of system prompt or tool definition that needs to change.
5. Propose the specific fix.
6. Design the test that would prove the fix.
7. Identify if this is a one-off or a class of bugs.
8. Recommend the regression test to add.

**Constraints:**
- Never speculate without grounding in trace lines.
- Distinguish "model was wrong" from "we asked the wrong question."
- The fix must be falsifiable (a test that fails today and passes after).

**Output format:** Trace walkthrough + diagnosis + fix + test.