Session Skill

Extracts human-readable conversation turns from AI coding session history files (.jsonl). Supports two formats:

• Claude Code, the JSONL format written by Anthropic's Claude Code CLI (~/.claude/projects/<hash>/<session-id>.jsonl)
• Codex / Cursor, the simpler {"role": ..., "content": ...} per-line format used by OpenAI Codex and Cursor IDE sessions

Format is detected automatically from the first parseable line.

What gets extracted

Only substantive conversation turns are kept:

Content type	Action
User text messages	Kept if ≥ 3 words
Assistant text responses	Kept if ≥ 20 words
Assistant thinking blocks	Skipped (internal reasoning, not final output)
Tool use / tool result blocks	Skipped (avoids leaking file contents or credentials)
Image / attachment blocks	Skipped
Sub-agent scaffolding (isSidechain: true)	Skipped (internal sub-agent turns)
Session metadata lines	Skipped (permission-mode, file-history-snapshot, system, last-prompt)

The extracted text is then passed through Synthadoc's standard pre-LLM source sanitizer (zero-width characters, bidi overrides, HTML comments, hidden CSS spans, base64 blobs, instruction-override phrases), exactly like PDF, DOCX, URL, and every other source type.

Output format

Each turn is labelled [USER] or [ASSISTANT] and separated by ---:

[USER]
How do I implement a sliding window algorithm?

---

[ASSISTANT]
A sliding window algorithm maintains a contiguous subarray (the "window") …

suggested_slug

The skill returns a suggested_slug in metadata derived from the session file's modification time and the first substantive user message:

session-2026-07-15-how-do-i-implement-a-sliding

Large sessions, chunking