Transcript Scraper

Overview

The transcript scraper reads Claude Code's JSONL transcript files and feeds structured data to the skilltracer agent. It extracts tool calls, results, user prompts, and assistant text into a format the agent can evaluate for skill-worthy work.

Prompt-Cycle Windowing

The transcript is windowed to the previous prompt cycle — the activity between the last two user prompts. Only the most recently completed task is evaluated on each invocation. This keeps the scraper focused and avoids re-evaluating old work.

Boundary Tracking

The scraper is boundary-aware via the last_traced_boundary field stored in the project's .skilltrace marker file. This tracks where the last successful scrape ended.

When consecutive user prompts trigger no tool use and evaluation is skipped, the window automatically expands to cover all missed work. After a successful scrape, the boundary advances to the start of the current prompt so nothing is lost.

What Gets Captured

• Tool calls — name and truncated parameters
• Tool results and error outputs
• User prompts
• Assistant text responses
• Subagent actions — appear with role:"subagent"
• Workflow actions — appear with role:"workflow"

MCP Tool Support

MCP tools (prefixed with mcp__) receive enhanced capture: up to five parameters are included in the scraper output, giving the skilltracer enough context to understand MCP-driven workflows.

Secret Redaction

Sensitive values such as API keys, tokens, and credentials are automatically redacted from scraper output before it reaches the skilltracer agent.

What Gets Filtered

• Skilltrace-internal tool calls are excluded from results
• System prompts and internal hook outputs

This ensures the skilltracer only sees your actual work, not Skilltrace's own bookkeeping operations.

Privacy

The transcript scraper operates entirely locally. No data leaves your machine. Transcripts are read from Claude Code's local storage and processed in-memory.