1. Parse requirements → restate objective
2. Unsure about tables/fields → first query schema (information_schema or list_tables)
3. Plan SQL → present to user
4. Read-only → execute directly
5. Write operation → confirm before execution → verify affected rows → report result

runs/<workflow>/active/<run_id>/
├── proposal.md                # Requirements / objective
├── context.json               # Known tables/fields/IDs
├── tasks.md                   # Checklist + approval gate
├── evidence/sql.md            # SQL to execute (write ops written here first)
├── evidence/result.md         # Conclusion + SQL + results
└── logs/events.jsonl          # Optional tool call summary (no sensitive data)

✅ Query complete: 142 new users in the last 7 days

Executed SQL:
SELECT DATE(created_at) as date, COUNT(*) as count 
FROM user_profiles 
WHERE created_at > NOW() - INTERVAL '7 days'
GROUP BY DATE(created_at) ORDER BY date DESC;

| date       | count |
|------------|-------|
| 2025-01-09 | 23    |
| 2025-01-08 | 31    |
| ...        | ...   |

User: Get registered user count for the last 7 days, by day

1. Confirm table user_profiles, field created_at
2. Execute aggregation SQL
3. Return: conclusion + numbers + SQL + table

User: Find projects that have runs but all failed

1. Confirm projects, runs tables and status field
2. Present JOIN + aggregation SQL
3. Execute and return results (mask email)

User: Create a new run for project xxx

1. First check if project exists
2. Present INSERT SQL + expected impact: 1 row
3. Await confirmation → execute → return new record id

User: Change run abc's status to completed

1. First SELECT to verify current state
2. Present UPDATE SQL + WHERE id = 'abc'
3. Confirm → execute → SELECT again to verify

User: Delete all runs where status = 'failed'

1. First SELECT count(*) WHERE status = 'failed'
2. Present count + DELETE SQL
3. If > 100 rows, force double confirmation
4. After confirmation execute → report deleted row count

User: Clear the runs table

❌ Refuse to execute
→ Prompt: DELETE without WHERE condition, this will delete all data
→ Suggest: Use TRUNCATE (requires migration) or add explicit condition

-- List all tables
SELECT table_name FROM information_schema.tables 
WHERE table_schema = 'public';

-- View table structure
SELECT column_name, data_type, is_nullable 
FROM information_schema.columns 
WHERE table_name = '<table_name>';