Add CodeAct-style trajectory exporter for training data

[evansenter]

Background

From analyzing the Opus 4.5 training methods document:

"CodeActInstruct dataset (7,139 trajectories) includes critical filtering for self-improvement: trajectories where the model encounters errors but rectifies in later turns are preserved, explicitly promoting self-debugging capability."

The session logs already contain this data. We should be able to export it in a format suitable for training.

Proposal

Add a trajectory export feature that:

1. Extracts tool-use trajectories

Convert session logs to Thought-Action-Observation format:

Thought: I need to find the config file
Action: Read(path="config.json")
Observation: {"error": "File not found"}
Thought: Maybe it's in a subdirectory
Action: exec(command="find . -name config.json")
Observation: ./src/config.json
Action: Read(path="./src/config.json")
Observation: {"api_key": "..."}

2. Filters for error→recovery patterns

Identify trajectories where:

• An error occurred (tool returned error, assertion failed, etc.)
• The agent successfully recovered (subsequent success on related task)
• Mark these as high-value training examples

3. Outputs in standard format

# Export all trajectories from last 7 days
agent-session-analytics-cli export-trajectories --since 7d --format codeact

# Export only error-recovery patterns
agent-session-analytics-cli export-trajectories --filter error-recovery --format jsonl

# Export with metadata for training
agent-session-analytics-cli export-trajectories --include-metadata --output trajectories.jsonl

Output Format Options

CodeAct JSONL

{
  "session_id": "abc123",
  "trajectory": [
    {"role": "thought", "content": "..."},
    {"role": "action", "tool": "Read", "args": {...}},
    {"role": "observation", "content": "..."}
  ],
  "outcome": "success",
  "has_error_recovery": true,
  "error_recovery_spans": [[2, 5]]
}

Anthropic Messages Format

For direct use with Claude fine-tuning (if/when available).

Value

1. Training data generation: Build custom training sets from real usage
2. Quality analysis: Identify which sessions had good error recovery
3. Pattern mining: Find common failure modes and successful recovery strategies

Related Work

• CodeAct paper: https://arxiv.org/abs/2402.01030
• MATH-SHEPHERD (filtering for self-improvement): https://arxiv.org/abs/2312.08935