AI Session Export

Export AI coding session transcripts from multiple tools into a unified Markdown archive for browsing, search, and downstream workflows.

Supported Sources

Source	Data Location
OpenCode	~/.local/share/opencode/opencode.db
Claude Code	~/.claude/projects/**/*.jsonl
Codex	~/.codex/sessions/**/*.jsonl, ~/.codex/archived_sessions/*.jsonl
Google Antigravity 2.0	~/.gemini/antigravity/brain/*/.system_generated/logs/transcript_full.jsonl
Google Antigravity IDE	~/.gemini/antigravity-ide/brain/*/.system_generated/logs/transcript_full.jsonl
Google Antigravity CLI	~/.gemini/antigravity-cli/brain/*/.system_generated/logs/transcript_full.jsonl
Cursor	~/Library/Application Support/Cursor/User/globalStorage/state.vscdb
DeepSeek Harness	~/.dsh/sessions/*/*/session.jsonl*
Gemini CLI	~/.gemini/tmp/*/chats/session-*.json or .jsonl
Grok Build	~/.grok/sessions/*/*/updates.jsonl and summary.json
Second Mind	second_mind_export.json

Quick Start

# Install
uv pip install -e '.[dev]'

# Export all sources (incremental)
python export_sessions.py

# Export specific source
python export_sessions.py --source antigravity
python export_sessions.py --source codex
python export_sessions.py --source cursor
python export_sessions.py --source dsh
python export_sessions.py --source gemini
python export_sessions.py --source grok

# Full re-export (ignore incremental state)
python export_sessions.py --full

# Export only recent sessions
python export_sessions.py --since-date 2026-06-01

# Dry run
python export_sessions.py --dry-run

By default, output and incremental state are stored under ~/.local/share/ai-session-export/. Use --base-dir and --state-file to target another private archive. Never write real session exports into a public repository.

--source antigravity scans all three Antigravity surfaces by default. They share the antigravity/ output directory and source: antigravity, while optional frontmatter surface identifies "2", "ide", or "cli". The legacy --antigravity-dir /path/to/brain override scans one IDE-compatible root.

The DeepSeek Harness source reads each session's append-only event log (session.jsonl.zstd, or plain session.jsonl when compression is disabled; session directory ids are not a single namespace, so discovery accepts any id shape). It keeps user/message and assistant/message events (final assembled turns, not streaming chunks), takes titles from session/title events, and attributes models per message from assistant/message source (falling back to request/header), back-filling the user turns that triggered each response. It drops subagent-child sessions and <system-reminder> instruction injections, tolerates torn trailing records (including a truncated final Zstandard frame), and rewrites one stable file per growing live session. Decompression shells out to the zstd binary.

Gemini CLI

Export sessions created by the Gemini CLI (google-gemini/gemini-cli).

• CLI Options: --source gemini --gemini-dir <path> (specifies the temporary project root containing recordings).
• Implementation Details:
  • Reads session JSON and JSONL records, utilizing chatRecordingService.ts and chatRecordingTypes.ts schema structures.
  • Uses displayContent when supplied.
  • Replays message replacements, metadata checkpoints, and rewind actions to reconstruct the final state.
  • Skips internal thoughts, tools, and subagents.
  • Fills the User model from the next response metadata.
  • Does not attempt to convert project hashes back to a local working directory (cwd).

Grok Build

Export sessions created by the Grok Build platform (xai-org/grok-build).

• CLI Options: --source grok --grok-sessions-dir <path> (specifies the directory holding Grok sessions).
• Implementation Details:
  • Parses updates.jsonl and summary.json based on session/storage/mod.rs and session/persistence.rs format mappings.
  • Supports both timestamped method/params payloads and legacy event types.
  • Stitches streaming text chunks into unified messages.
  • Follows session rewinds correctly to maintain clean chronological outputs.
  • Omits thoughts, tool execution steps, hidden host prompts, and nested subagent sessions.
  • Uses the summary model as session-level inventory only; never assigns it to earlier turns.

Output Format

Each session is exported as a Markdown file with YAML frontmatter:

---
source: opencode
session_id: "ses-example"
title: "Fix the bug in auth.py"
date: "2026-06-29"
message_count: 3
turn_models: ["gpt-example", "gpt-example", "gpt-example"]
---
# Fix the bug in auth.py

## User [16:38]

Fix the bug in auth.py

## Assistant [16:38]

I'll look at the auth.py file first.

## Assistant [16:39]

The bug is on line 42.

When a source can attribute models per turn, turn_models is a JSON array aligned one-to-one with the rendered User and Assistant sections. Unknown entries are null; the field is omitted when every turn is unknown. models_used remains a session-level inventory and must not be used to guess per-turn attribution. Antigravity records additionally include surface: "2", surface: "ide", or surface: "cli".

Installation as a Coding Agent Skill

This project is designed to be used as a skill by AI coding agents (Codex, Claude Code, Cursor, OpenCode, etc.).

1. Clone or download this repository.
2. Point your AI agent at skill.md in the project root — it contains the workflow instructions.
3. If your workspace has a skills index (e.g., rules/skills/INDEX.md), add an entry pointing to this project's skill.md.

Testing

# Unit + integration tests
python -m pytest tests/ -v

# Live end-to-end tests (requires real local data)
AI_SESSION_EXPORT_LIVE=1 python -m pytest tests/ -v -m live_e2e

License

MIT