devin-conversation-recovery

Devin Conversation Recovery

Devin runs as an ACP provider inside Windsurf ("Devin Local"). Conversations are NOT in Hermes's session DB — they live in two SQLite stores on disk:

• Titles/session metadata: ~/Library/Application Support/Devin/User/globalStorage/state.vscdb

  → ItemTable keys windsurf.acp.sessioninfo.session.* and the windsurf.acp.messageStore.index key mapping sessionId → message-db uuid.

• Message bodies: ~/Library/Application Support/Devin/User/acp-messages/<uuid>.db

  → table messages(position, kind, payload); payload is JSON with content as streaming text chunks, plus tool_call rows carrying rawInput (commands, file reads) and exit codes.

Quick path: use the helper script

Prefer the Rust twin ~/.ciel/bin/ciel-convo (same arguments); the Python find_devin_convo.py remains as fallback while the parity soak runs.

ciel-convo --list
ciel-convo "blender-mcp"          # titles + first messages
ciel-convo "search text" --deep   # grep ALL payloads (slow)
ciel-convo --extract <uuid> -o /tmp/convo.txt
# fallback: python <skill-dir>/scripts/find_devin_convo.py <same args>

The extract output is plain text (=== [position] kind sections): agent thoughts/ messages joined from streaming chunks, tool commands with $ cmd (exit N), and truncated tool output. Full transcripts can be 1M+ chars — dump to a file and page through with read_file rather than printing.

Manual procedure (if script unavailable)

1. List sessions:

import sqlite3, json
con = sqlite3.connect("~/Library/Application Support/Devin/User/globalStorage/state.vscdb")
rows = con.execute("select key,value from ItemTable where key like 'windsurf.acp.sessioninfo.session.%'").fetchall()
idx = json.loads(con.execute("select value from ItemTable where key='windsurf.acp.messageStore.index'").fetchone()[0])

# idx["acp/devin-cli/<name>"]["uuid"] gives the acp-messages db filename

1. Search bodies across all dbs: iterate User/acp-messages/*.db,