sql-read

name	sql-read
description	Use when a user needs live read-only Postgres or SQLite inspection; configure a named read-only target once, then use the blanket-approvable `sql-read run` path for later queries.
metadata	{"short-description":"Read Postgres or SQLite safely"}

SQL Read

Use the sql-read helper from the active Nix profile to inspect Postgres or SQLite data in read-only mode.

Trigger this skill for:

• live schema inspection
• ad hoc read-only analysis queries
• fast table or column lookups when the answer lives in a database

Do not use it for:

• writes, migrations, or backfills
• admin tasks
• general interactive shell work

Quick Start

The active Nix profile puts sql-read on PATH, so call it directly.

1. Use the default Claude state dir: "$CLAUDE_CONFIG_DIR/skills/sql-read/state".
2. Export the connection input in its own command.
   • export PROD_READONLY_URL='postgresql://user:pass@host:5432/dbname'
3. Persist the named target in a separate command.
   • sql-read target upsert --state-dir "$CLAUDE_CONFIG_DIR/skills/sql-read/state" --name prod-readonly --engine postgres --dsn-env-var PROD_READONLY_URL
4. Run read-only queries against the stored target in a separate command.
   • sql-read run --state-dir "$CLAUDE_CONFIG_DIR/skills/sql-read/state" --target prod-readonly --file "$CLAUDE_CONFIG_DIR/skills/sql-read/assets/queries/postgres-schema-overview.sql" --format json
5. Use the same split workflow for SQLite targets.
   • export LOCAL_APP_DB='/absolute/path/to/app.sqlite3'
   • sql-read target upsert --state-dir "$CLAUDE_CONFIG_DIR/skills/sql-read/state" --name local-app --engine sqlite --sqlite-db-path-env-var LOCAL_APP_DB
   • sql-read run --state-dir "$CLAUDE_CONFIG_DIR/skills/sql-read/state" --target local-app --file "$CLAUDE_CONFIG_DIR/skills/sql-read/assets/queries/sqlite-schema-overview.sql" --format json
6. Inspect or remove configured targets with explicit target-management commands.
   • sql-read target list --state-dir "$CLAUDE_CONFIG_DIR/skills/sql-read/state"
   • sql-read target remove --state-dir "$CLAUDE_CONFIG_DIR/skills/sql-read/state" --name local-app

Workflow

1. Start with schema inspection.
   • Use the bundled query templates in assets/queries/.
   • Load the relevant engine notes from references/postgres.md or references/sqlite.md only when needed.
2. Configure the target once.
   • Use sql-read target upsert --state-dir <dir> --name <target> ... to persist the resolved read-only DSN or SQLite path under the skill state dir.
   • Keep env-var assignment and target upsert in separate commands.
   • Do target setup in its own command; it is not part of the blanket-approved surface.
3. Keep queries narrow.
   • Ask for aggregates, counts, or explicit filters before scanning large tables.
   • Add an explicit limit in the SQL unless the query is already aggregate-only.
4. Prefer JSON output.
   • Use --format json by default so the agent sees compact, structured rows.
   • Switch to table only when the user explicitly wants a human-readable table.
5. Run read-only queries through the stored target.
   • sql-read run --state-dir <dir> --target <name> ... is the only blanket-approved execution path.
   • Later run calls do not require the env var to still exist.

Output

Default JSON output is compact and stable:

{"engine":"postgres","target":{"mode":"named-target","name":"prod-readonly"},"columns":["id","email"],"rows":[[1,"[e-mail]"]],"row_count":1,"truncated":false,"duration_ms":38}

Bundled Resources

• sql-read
  • The active Nix profile provides this command on PATH.
  • run is the blanket-approvable read-only execution path.
  • target upsert|list|remove manages persisted named targets.
• scripts/
  • Contains the Rust source package that the Nix profile builds into sql-read.
• references/postgres.md
  • Catalog-query patterns and Postgres-specific caveats.
• references/sqlite.md
  • SQLite schema-inspection patterns and read-only caveats.
• assets/queries/
  • Reusable schema-inspection and table-introspection query templates.

Gotchas

• Rust is a control and safety tool here, not a speed win over psql.
• If sql-read is missing, reapply the profile so the packaged helper is rebuilt and activated.
• Blanket approval should target sql-read run, not sql-read target.
• Keep env-var assignment, target upsert, and run as separate commands.
• run does not accept raw DSNs, raw SQLite paths, or env-var target flags; configure the target first.
• Read-only enforcement in the helper is defense-in-depth, not a substitute for least-privilege credentials.
• Postgres queries with duplicate column names are rejected; alias duplicate columns before running them.