db-postgres

db-postgres

Query Postgres databases declared in .env. Connections follow the same numbered pattern as SOCIAL_YOUTUBE_N_* / SOCIAL_INSTAGRAM_N_* — one block per database, labelled for humans, picked by label or index at call time.

Setup — one-time, per database

Add a block to .env (gitignored). Increment the index per connection:

# ── Postgres: msgops-dev ─────────────────────────────
DB_POSTGRES_1_LABEL=msgops-dev
DB_POSTGRES_1_HOST=db.dev.internal
DB_POSTGRES_1_PORT=5432
DB_POSTGRES_1_DATABASE=msgops
DB_POSTGRES_1_USER=agent_readonly
DB_POSTGRES_1_PASSWORD=...                # raw; .env is gitignored
DB_POSTGRES_1_SSL_MODE=require            # disable | require | verify-ca | verify-full
# DB_POSTGRES_1_SSL_CA_PATH=/path/ca.pem  # optional, if verify-*
# DB_POSTGRES_1_ALLOW_WRITE=false         # default false
# DB_POSTGRES_1_QUERY_TIMEOUT=30          # seconds, default 30
# DB_POSTGRES_1_MAX_ROWS=1000             # default 1000

# ── Postgres: bms-prod (read-only replica) ───────────
DB_POSTGRES_2_LABEL=bms-prod
DB_POSTGRES_2_HOST=bms-ro.prod.internal
DB_POSTGRES_2_PORT=5432
DB_POSTGRES_2_DATABASE=bms
DB_POSTGRES_2_USER=agent_readonly
DB_POSTGRES_2_PASSWORD=...
DB_POSTGRES_2_SSL_MODE=require

Alternative — full DSN instead of components (DSN wins when both are set):

DB_POSTGRES_3_LABEL=evo-ai-dev
DB_POSTGRES_3_DSN=postgresql://agent_ro:***@evoai.dev.internal:5432/evoai?sslmode=require

LABEL is always required — it's how agents pick the connection.

Usage

All commands output a single JSON line on stdout (safe to pipe). Errors go to stderr as JSON and exit non-zero.

List configured connections

python3 .claude/skills/db-postgres/scripts/db_client.py accounts

Health-check a connection

python3 .claude/skills/db-postgres/scripts/db_client.py test msgops-dev

Run a read-only query

python3 .claude/skills/db-postgres/scripts/db_client.py query msgops-dev \
  "SELECT count(*) FROM users WHERE created_at > now() - interval '7 days'"

Explore schema

# List all tables (public + user schemas)
python3 .claude/skills/db-postgres/scripts/db_client.py tables msgops-dev

# Describe a table (columns, types, nullability, defaults)
python3 .claude/skills/db-postgres/scripts/db_client.py describe msgops-dev users
# Schema-qualified:
python3 .claude/skills/db-postgres/scripts/db_client.py describe msgops-dev analytics.events

Output shape

Successful query:

{
  "ok": true,
  "query_id": "uuid-v4",
  "label": "msgops-dev",
  "columns": ["id", "email", "created_at"],
  "rows": [[1, "a@b.com", "2026-04-22T12:00:00"]],
  "row_count": 1,
  "truncated": false,
  "full_result_path": null,
  "execution_time_ms": 12
}

When rows exceed MAX_ROWS, truncated is true and full_result_path points to a CSV in ADWs/logs/db-queries/<query_id>.csv — the agent can read that file directly instead of re-running the query.

Error:

{"ok": false, "error_code": "write_blocked", "error": "Write query blocked — connection 'msgops-dev' has ALLOW_WRITE=false. ...", "label": "msgops-dev"}

Error codes:

• no_connections — no DB_POSTGRES_N_* blocks in .env
• not_found — label/index doesn't match any block
• ambiguous — multiple blocks share the same label (use index instead)
• config_error — block present but required field missing (typically LABEL)
• driver_missing — psycopg2 not installed
• connection_failed — network, auth, TLS, or statement_timeout tripped
• write_blocked — write verb detected without ALLOW_WRITE=true
• multi_statement — more than one statement in a single call
• usage — wrong CLI args

Guardrails

• Write verbs (DELETE | UPDATE | INSERT | TRUNCATE | DROP | ALTER | CREATE | GRANT | REVOKE | COMMENT | VACUUM | REINDEX) — refused unless DB_POSTGRES_N_ALLOW_WRITE=true.
• Multi-statement queries (anything with ; that has content after it) — refused in v1.
• Query timeout — sets statement_timeout = <QUERY_TIMEOUT>s on the session before executing.
• Result size — fetchmany(MAX_ROWS + 1) to detect truncation; full result streamed to CSV when truncated so the agent context never holds >1000 rows.

Workflow

1. If the user doesn't specify a label, run accounts to see what's configured and pick the one that matches their intent. If ambiguous, ask.
2. Write the smallest SQL that answers the question — prefer aggregates, LIMIT, and EXPLAIN before dumping rows.
3. Run via query. Inspect the result.
4. For performance-sensitive queries, load the deep-dive references below.

Dependencies

• Python 3.10+
• psycopg2-binary (or psycopg2) — not pre-installed; uv pip install psycopg2-binary on first use.

Deep-dive references

These load the PlanetScale database-skills repo verbatim — same content the upstream authors ship for their own tooling.

• EXPLAIN analysis
• Index optimization
• Indexing fundamentals
• MVCC + VACUUM
• MVCC transactions
• Partitioning
• PGBouncer configuration
• Memory management / ops
• Monitoring
• Backup & recovery
• Optimization checklist

Upstream: planetscale/database-skills (MIT). Credit to PlanetScale for the reference content.