Company data in public schema is never modified by agents. Engineer
creates views, summary tables, and computed data in the dash schema
(monthly_mrr, customer_health_score, churn_risk, etc.).

Key changes:
- db/session.py: get_sql_engine() bootstraps dash schema with search_path
- Analyst: read-only SQLTools against public, reads dash views
- Engineer: full SQLTools scoped to dash schema, update_knowledge tool
- introspect_schema: shows both schemas with labels
- update_knowledge: Engineer records schema changes so Analyst discovers them
- Instructions rewritten for dual-schema world

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

The db_url user is `ai`, and PostgreSQL's search_path resolves to the
`ai` schema first (created by Agno). Explicitly set search_path=public
so company data lands in the right schema.

Three-schema layout:
- ai: Agno framework (knowledge, learnings, sessions)
- public: Company data (customers, subscriptions, etc.)
- dash: Agent-managed views and summaries

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…ies eval

- Analyst now uses a read-only PostgreSQL engine (default_transaction_read_only)
  so write protection is enforced by the database, not just instructions
- Leader instructions teach decomposition for multi-step queries while
  keeping simple questions single-shot
- Engineer instructions now embed source table metadata via semantic model
- Shared LearningMachine instance replaces 3 separate copies
- Introspect tool discovers views (not just tables), clamps sample limits,
  uses parameterized queries, and surfaces cleaner error messages
- Save-query tool rejects multi-statement SQL and uses regex keyword matching
- New boundaries eval category validates schema access controls
- Engine caching avoids redundant connections
- Module __init__ cleanup removes eager re-exports

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

- evals/smoke.py: 24 lightweight smoke tests across 8 groups (warmup,
  simple_data, metrics, data_quality, multistep, insight, engineering,
  edge_cases). Mirrors Pal's pattern — no judge model needed.
- evals/improve.py: Self-improvement loop that runs smoke tests, analyzes
  failures with GPT-5.4, applies targeted instruction changes, and verifies.
- evals/__main__.py: Add smoke and improve subcommands, load .env for local runs.
- Fix eval pass_rate check (100.0 not 1.0) and rephrase judge criteria so
  binary scoring correctly marks refusals as PASS.
- docs/IMPROVE_DASH.md: Claude Code prompt for running the improvement loop.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

…est steps

Add production deployment instructions alongside local Docker setup.
Fix ngrok port (7777, not 8000). Add Event Subscriptions verification
step that was causing silent failures. Add test instructions.

Show both ports in ngrok instructions: 8000 for Docker Compose (Quick
Start), 7777 for bare AgentOS. Update SLACK_CONNECT.md to match README:
add Railway deployment path, event subscription verification step, and
production URL guidance.

…Step 1

Replace full step-by-step procedure with a 4-step summary that links
to the canonical setup guide. All procedural detail (ngrok ports,
manifest, credentials, Railway flow, event verification, scopes) now
lives only in docs/SLACK_CONNECT.md.

docs: update Slack setup with production flow and event verification