Agent trace studio

Visualize Claude Code trajectories as search behavior.

Load native Claude Code JSONL or Galapagos derived exports and see the run as a readable timeline: prompts, assistant turns, tool calls, evaluator feedback, and SFT-ready conversation slices.

Supported inputs
Native
claude_sessions/projects/**/*.jsonl
Events
derived/claude_code_events.jsonl
SFT
derived/claude_code_sft.jsonl

Load a trajectory

Paste or upload Claude Code native JSONL, Galapagos event JSONL, or SFT JSONL.

4
Messages

Claude Code native JSONL

6
Events

native

2
Tool calls

1 result events

1
Sessions

claude-fable-5

Timeline

11:16:16 PM11:17:00 PM

native
Selected eventline 1 · 11:16:16 PM
{
  "type": "user",
  "sessionId": "demo-session",
  "timestamp": "2026-07-02T23:16:16.560Z",
  "message": {
    "role": "user",
    "content": "# Galapagos: Optimization Task\n\nImprove `$GALAPAGOS_WORKSPACE/solution.py` and maximize `combined_score`."
  }
}

Conversation

4 messages
user11:16:16 PM
# Galapagos: Optimization Task

Improve `$GALAPAGOS_WORKSPACE/solution.py` and maximize `combined_score`.
assistant11:16:20 PM
I'll inspect the current solution and run the baseline evaluation.

<tool_call name="Read" id="toolu_read">
{
  "file_path": "$GALAPAGOS_WORKSPACE/solution.py"
}
</tool_call>
user11:16:25 PM
<tool_result id="toolu_read" error="false">
def construct_packing(): ...
</tool_result>
assistant11:17:00 PM
The baseline is weak. I will replace it with a structured optimizer.

<tool_call name="Write" id="toolu_write">
{
  "file_path": "$GALAPAGOS_WORKSPACE/solution.py",
  "content": "# EVOLVE-BLOCK-START\n..."
}
</tool_call>