Bug Report: Kanban SQLite database corruption under rapid task creation

Summary

The kanban SQLite database (~/.hermes/kanban.db) becomes corrupted (database disk image is malformed) when creating ~9-10 tasks in rapid succession via the kanban_create tool API. This has happened 3 times in 2 days under normal orchestrator workflow.

Environment

• Hermes Agent version: v0.14.0 (2026.5.16)
• Python: 3.11.15
• OS: Ubuntu 22.04 (Linux 6.8.0-117-generic)
• SQLite: bundled with Python 3.11

Steps to Reproduce

1. Start gateway: hermes gateway start
2. Create tasks via kanban_create tool API in a loop (or rapid succession)
3. After ~9-10 tasks, the next kanban_create call fails with:

   {"error": "kanban_create: database disk image is malformed"}
   

4. All subsequent kanban operations fail with the same error

Observed Behavior

• kanban.db-wal file becomes 0 bytes after corruption
• The database file itself appears intact in size but is unreadable by SQLite
• Previously created tasks are lost
• Gateway continues running but cannot dispatch tasks

Expected Behavior

• Creating 10+ tasks sequentially should not corrupt the database
• WAL mode should handle concurrent access safely
• If corruption occurs, it should be recoverable without full re-initialization

Recovery Steps (currently required)

hermes gateway stop
cp ~/.hermes/kanban.db ~/.hermes/kanban.db.backup.$(date +%Y%m%d_%H%M%S)
rm -f ~/.hermes/kanban.db-shm ~/.hermes/kanban.db-wal
mv ~/.hermes/kanban.db ~/.hermes/kanban.db.corrupted.$(date +%Y%m%d_%H%M%S)
hermes kanban init
hermes gateway start

Additional Context

• The issue occurs when using the tool API (kanban_create), not CLI commands
• We added 1-second delays between kanban_create calls as a workaround, but this is not a fix
• The dispatcher holds an open DB connection; concurrent writes from tool API calls may race with WAL checkpointing
• Previous corruption incidents: 2025-05-23 (twice), 2025-05-24 (once)

Suggested Investigation

1. Check if the kanban DB connection uses proper transaction isolation
2. Verify WAL checkpoint behavior under rapid writes
3. Consider adding an application-level write queue or mutex for kanban operations
4. Add automatic WAL recovery on startup if -wal or -shm files are stale

Attachments

• Will attach kanban.db and kanban.db-wal from next corruption incident if helpful