edo-tensei

name	edo-tensei
description	Transfers AI session context across IDEs (Claude Code, Claude.ai Cowork, Copilot, Cursor, Kiro, Windsurf, Trae, Antigravity, Codex) by reading local session files directly. Works in any environment with file system access — no VS Code or MCP required. Use this skill when the user wants to continue work from another AI tool, summarize a recent session, or generate a structured handoff prompt.
argument-hint	[claude|cowork|copilot|cursor|codex|kiro|windsurf|trae|antigravity]

Edo Tensei

Use this skill to take over work from another AI IDE, or recover the latest session context without manually pasting the whole conversation.

This skill works directly via file system access — no VS Code extension or MCP server required.

Primary Goal

Find the most relevant recent session file for the current project, read it selectively, and return a concise handoff summary:

• Goal
• Completed / attempted steps
• Current blocker
• Proposed next steps

Output Contract

Always return:

## Session Handoff Summary
**Source IDE**: <ide>
**Session Path**: <path>
**Last Activity**: <timestamp>

### Goal
<1-2 sentences>

### Completed / Attempted
- <item>

### Current Blocker
<description or "None">

### Proposed Next Steps
- [ ] <actionable item>

If the user asks for Chinese, output the same structure in Chinese.

---

Layer 0: Scope Gate

Before reading any large file:

1. Confirm the current working directory / workspace name.
2. Prefer the most recent 20% of a session or the last 60–250 lines/messages.
3. Filter to human-facing turns only:
   • role === "user" or role === "assistant"
   • content blocks with type === "text"
4. Skip tool-call noise unless it explains the blocker.

---

Layer 1: Search Strategy

Search the target IDE first if the user specified one. Otherwise try in this order:

1. Claude Code
2. Claude.ai Cowork
3. GitHub Copilot
4. Cursor
5. OpenAI Codex CLI
6. Kiro
7. Windsurf
8. Trae
9. Antigravity

Claude Code

Paths

OS	Path
Windows	%USERPROFILE%\.claude\projects\{slug}\*.jsonl
Linux / macOS	~/.claude/projects/{slug}/*.jsonl

Slug: workspace absolute path with slashes/colons replaced by -
e.g. C:\Users\user\myproject → c--users-user-myproject

Find

# Windows
$slug = (Get-Location).Path -replace '\\','-' -replace ':','-'
Get-ChildItem "$env:USERPROFILE\.claude\projects\*$slug*\*.jsonl" |
  Sort-Object LastWriteTime -Descending | Select-Object -First 5

# Linux/macOS
PROJECT_NAME=$(basename "$PWD")
ls -lt ~/.claude/projects/*${PROJECT_NAME}*/*.jsonl 2>/dev/null | head -5

Read: JSONL — filter by top-level type === "user" or "assistant". Extract message.content[].text (skip tool_result items) and message.content[].thinking (reasoning). cwd field on some lines gives the actual workspace path. Full details: session-claude.md

---

Claude.ai Cowork

Paths

OS	Path
macOS	~/Library/Application Support/Claude/local-agent-mode-sessions/
Linux	~/.config/Claude/local-agent-mode-sessions/
Windows	%APPDATA%\Claude\local-agent-mode-sessions\

Structure: {session-uuid}/{conversation-uuid}/agent/local_ditto_{conversation-uuid}/audit.jsonl
Session metadata (title, cwd, timestamps): {conversation-uuid}/local_{child-uuid}.json
Child sub-sessions (tool/MCP calls): {conversation-uuid}/local_{child-uuid}/audit.jsonl

Find

# macOS — newest session → newest conversation → main transcript path
BASE=~/Library/Application\ Support/Claude/local-agent-mode-sessions
SESSION=$(ls -t "$BASE" | head -1)
CONV=$(ls -t "$BASE/$SESSION" | head -1)
echo "$BASE/$SESSION/$CONV/agent/local_ditto_${CONV}/audit.jsonl"

Read: JSONL — filter by top-level type === "user" or type === "assistant".
User: message.content is a plain string.
Assistant: message.content[] array — extract type === "text" → text; optionally type === "thinking" → thinking.
Skip: type === "system", "rate_limit_event", "result".
Timestamps on _audit_timestamp field. Full details: session-claude-cowork.md

---

GitHub Copilot

Paths

OS	Path
Windows	%APPDATA%\Code\User\globalStorage\emptyWindowChatSessions\
Linux / macOS	~/.config/Code/User/globalStorage/emptyWindowChatSessions/

Read: JSON or JSONL. JSONL has two sub-formats:

• Old: kind=0 lines contain full v.requests[]; take last per sessionId
• New: kind=0 has empty requests; real data in kind=2 k="requests" lines (cumulative append); response patches in kind=2 k=["requests",N,"response"] lines

Full details: session-copilot.md

---

Cursor

Paths

OS	Path
Windows	%USERPROFILE%\.cursor\projects\{slug}\agent-transcripts\{uuid}\{uuid}.jsonl
Linux / macOS	~/.cursor/projects/{slug}/agent-transcripts/{uuid}/{uuid}.jsonl

Slug: lowercase drive + single dashes (e.g. C:\Users\user\MyProject → c-Users-user-MyProject). Each {uuid} folder contains exactly one {uuid}.jsonl file.

Find

# Windows
Get-ChildItem "$env:USERPROFILE\.cursor\projects" -Directory |
  Sort-Object LastWriteTime -Descending | Select-Object -First 3 |
  ForEach-Object { Get-ChildItem "$($_.FullName)\agent-transcripts\*\*.jsonl" |
    Sort-Object LastWriteTime -Descending | Select-Object -First 2 }

Read: JSONL — each line is one message: role ("user"/"assistant") and message.content[].text where type === "text". Read from the end for recent turns. Full details: session-cursor.md

---

OpenAI Codex CLI

Paths

OS	Path
Windows	%USERPROFILE%\.codex\sessions\rollout-*.jsonl
Linux / macOS	~/.codex/sessions/rollout-*.jsonl

Read: JSONL. Relevant types: session_meta (has payload.cwd for workspace path) and response_item (when payload.type === "message"). Content at payload.content[].text. Skip developer/system roles and Codex-injected prefixes (<permissions instructions>, <collaboration_mode>, etc.). Full details: session-codex.md

---

Kiro

Two formats — Format B (newer) takes priority:

Format B (has workspace path): kiro.kiroagent\workspace-sessions\{Base64URL(path)}\{uuid}.json
→ { "workspaceDirectory": "...", "history": [{ "message": { "role": "user"|"assistant", "content": "..." | ContentPart[] } }] }

Format A (legacy): kiro.kiroagent\{hex32}\*.chat
→ { "chat": [{ "role": "user"|"bot"|"human"|"tool", "content": "..." }] } — skip tool role and bot-only-ack messages ("On it.", "Understood.")

Full details: session-kiro.md

---

Windsurf / Trae

• Windsurf (%APPDATA%\Windsurf\User\globalStorage\chatSessions\): binary format — fall back to path-based handoff only. See session-windsurf.md
• Trae (%APPDATA%\Trae\User\globalStorage\.ckg\storage\{hash}\*_codekg.db): SQLite binary — currently unsupported. See session-trae.md

---

Antigravity

Paths

OS	Path
Windows	%USERPROFILE%\.gemini\antigravity\brain\{uuid}\.system_generated\logs\transcript.jsonl (or overview.txt)
Linux / macOS	~/.gemini/antigravity/brain/{uuid}/.system_generated/logs/transcript.jsonl (or overview.txt)

Read: JSONL.

• Note: Modern versions save to transcript.jsonl while older versions use overview.txt. Both use the same JSONL schema.
• User: source === "USER" or "USER_EXPLICIT" → input/content/text; unwrap <USER_REQUEST>...</USER_REQUEST> if present
• Assistant: source === "MODEL" + type === "PLANNER_RESPONSE" → content/text or tool_calls[{name:"reply"}].args.content
• ⚠️ Content truncated at ~900 chars; full history is cloud-only. Full details: session-antigravity.md

---

Layer 2: Reading Rules

• Read the tail / most recent messages first. Only go deeper if the recent section is insufficient.
• Use keywords to locate context faster: error, TODO, next, blocker, failed, plan, fix
• Do not claim you read the whole session if you only read the tail.
• Distinguish confirmed facts from inference.

---

Layer 3: Halt / Fallback Rules

Stop and report instead of guessing when:

• The file cannot be read (permissions, binary format, remote-only).
• Multiple candidate sessions conflict and workspace/time can't resolve it.

Required message:

I can't read that local session file from this environment. Please paste the last few user/assistant turns, or let me know which recent session to prioritize.

---

MCP Enhancement (Optional)

If the Edo Tensei MCP server is connected in your environment, you can use these tools instead of manual file search for better results:

• list_ide_sources — check which IDEs have sessions available
• scan_all_sessions — list all sessions across IDEs, sorted newest first
• scan_project_sessions — filter sessions by workspace path
• get_session / get_session_messages — read a specific session
• search_sessions — search by keyword, time, IDE, or workspace
• generate_handoff_prompt — generate structured handoff (path or fullText mode, 5 languages)
• export_session — export session to .edo_tensei/ as Markdown

To set up the MCP server:

1. Open VS Code Command Palette (Ctrl+Shift+P)
2. Run: Edo Tensei: Show MCP Config
3. Follow the setup instructions for your AI client