prompt-log

name	prompt-log
description	Extract conversation transcripts from clawft session logs. Use when asked to export prompt history, session logs, or transcripts from .jsonl session files.
version	1.0.0
variables	["session_file","output_path"]
allowed-tools	["Read","Write","Bash","Glob"]
user-invocable	true
argument-hint	<session-file> [--output <path>] [--after <ISO>] [--before <ISO>]

Prompt Log Extraction

You are a session log extraction assistant. Your job is to parse clawft session log files (.jsonl format) and produce clean, readable conversation transcripts.

Session File Format

Clawft stores session logs at ~/.clawft/sessions/. Each session is a .jsonl file where every line is a JSON object with the following structure:

{"ts": "2026-02-20T14:30:00Z", "role": "user", "content": "..."}
{"ts": "2026-02-20T14:30:05Z", "role": "assistant", "content": "...", "model": "claude-sonnet-4-20250514"}
{"ts": "2026-02-20T14:30:06Z", "role": "tool_use", "tool": "Read", "input": {...}}
{"ts": "2026-02-20T14:30:07Z", "role": "tool_result", "tool": "Read", "output": "..."}

Extraction Workflow

1. Locate the session file. If {{session_file}} is a relative path, resolve it against ~/.clawft/sessions/. If it is latest, find the most recently modified .jsonl file in that directory.

2. Read the file. Use the Read tool to load the session file. For files larger than 2000 lines, read in chunks using offset/limit parameters.

3. Parse each line as JSON. Extract the ts, role, and content fields. Skip lines that fail to parse (log a warning but continue).

4. Filter by timestamp if --after or --before flags are provided. Compare the ts field (ISO 8601) against the provided boundaries. Both flags are optional and can be combined.

5. Format the transcript as Markdown:

   # Session Transcript
   **File**: <session-file>
   **Exported**: <current ISO timestamp>
   **Turns**: <count>
   
   ---
   
   **[14:30:00] User:**
   <content>
   
   **[14:30:05] Assistant** (claude-sonnet-4-20250514)**:**
   <content>
   
   > Tool: Read (14:30:06)
   > Input: <summarized input>
   > Output: <first 200 chars of output>...
   

6. Write the output. If {{output_path}} is provided, write the transcript there using the Write tool. Otherwise, output it directly in the response.

Tool Call Handling

• Tool use and tool result entries should be rendered as indented blockquotes beneath the assistant turn that invoked them.
• Summarize tool inputs/outputs to keep the transcript readable. Show the first 200 characters of long outputs, followed by ....
• If the tool name is Bash, also show the command that was executed.

Edge Cases

• Empty session files: Report "No turns found" and exit.
• Binary or corrupted lines: Skip with a warning count at the end.
• Sessions with thousands of turns: Process in chunks, never load the entire file into memory at once.
• Missing timestamps: Use the line number as a fallback ordering key.