claude-code-cast

name	claude_code_cast
description	Parse Claude Code trajectories from the terminal-bench recording format. Extracts tool-use steps from sessions/claude_code.log (LiteLLM proxy log containing full Claude API request JSON) and metadata from results.json. Captures thinking blocks, parallel tool calls, and the full Claude Code tool taxonomy.
fingerprints	["sessions/claude_code.log","commands.txt","results.json"]
priority	50
metadata	{"version":"2.0","source":"seed"}

Claude Code Cast Trajectory Parser

Directory Layout

run_dir/
  # Often contains a nested trial dir:
  #   task_name/task_name.1-of-1.YYYY-MM-DD__HH-MM-SS/
  results.json            # trial metadata: id, instruction, is_resolved, ...
  commands.txt            # tmux send-keys log (fallback parser)
  sessions/
    claude_code.log       # LiteLLM request/response log (primary source)
    agent.cast            # asciinema recording of agent session
    agent.log             # (optional) agent output log
    tests.cast            # asciinema recording of test run
    tests.log             # (optional) test output log
  panes/
    pre-agent.txt         # terminal state before agent
    post-agent.txt        # terminal state after agent
    post-test.txt         # terminal state after test
  agent-logs/             # (usually empty)

Step Extraction Logic

1. Read sessions/claude_code.log line by line, parsing JSON objects
2. Collect all requests containing both messages and tools keys (these are the main Claude Code agent conversation turns)
3. Take the last such request (it contains the full accumulated conversation)
4. Walk through messages, extracting:
   • assistant messages with tool_use content blocks -> action steps
   • tool messages matched by tool_call_id -> observations
   • thinking content blocks -> attached to the first step in the message
5. Map tool names to human-readable actions (see Tool Taxonomy below)
6. Task description from results.json -> instruction

Parallel Tool Call Detection

When an assistant message contains multiple tool_use blocks, all resulting steps are assigned the same parallel_group ID. This allows downstream analysis to evaluate them as a batch rather than sequentially.

Thinking Block Capture

Claude Code may include type: "thinking" content blocks in assistant messages. The thinking text is captured in StepRecord.thinking on the first step of the message. This preserves the agent's reasoning chain for analysis.

Tool Taxonomy

The parser formats the following Claude Code tools:

Core Tools

• Bash - Shell command execution
• Read - File reading
• Write - File creation/overwrite
• Edit - Targeted file edits (old_string -> new_string)
• Glob - File pattern matching
• Grep - Content search (ripgrep-based)

Agent & Planning Tools

• Agent - Subagent spawning (with subagent_type)
• EnterPlanMode / ExitPlanMode - Plan mode transitions
• TodoWrite / TodoRead - Task tracking
• SendMessage - Inter-agent messaging
• AskUserQuestion - User interaction

Notebook & Web Tools

• NotebookEdit - Jupyter notebook cell editing
• WebSearch - Web search
• WebFetch - URL content fetching

Workspace & Environment Tools

• EnterWorktree / ExitWorktree - Git worktree isolation
• Skill - Skill invocation
• CronCreate / CronDelete / CronList - Scheduled tasks

Task Management Tools

• Task - Task subagent management
• TaskOutput / TaskStop - Background task control
• BashOutput - Background bash output retrieval

Other Tools

• PowerShell - Windows shell execution
• FileRead / FileWrite / FileEdit - Alternative file tool names

Format Fingerprint

Presence of sessions/claude_code.log file alongside results.json.