db-connector

dsh-db-connector

中文说明

A database connector bundle for DeepSeek Harness (dsh, the "everything is a plugin" framework built on Cordis).

It gives your agent safe, audited access to SQLite, PostgreSQL and MySQL through five model-facing tools, plus a human /db command, with a deliberate focus on read-only safety, schema intelligence, write approval, and a durable SQL audit trail — an area the official dsh tool set does not cover.

This bundle is an independent, from-scratch implementation. It does not

reuse any existing open-source dsh tool code; the tool names, parameter

shapes, and result formats are this bundle's own design and are not tied to

any official dsh schema.

Feature overview

Minimum requirements

• Node.js ≥ 22.13 (uses the built-in, unflagged node:sqlite).
• A running dsh profile for the tools to show up on ctx.tools.
• Optional server drivers (peer dependencies, only needed for those engines):
• PostgreSQL: npm i pg
• MySQL: npm i mysql2

SQLite needs nothing extra. Development (npm test) is easiest on Node ≥ 23.6, which runs the type-stripped test files natively.

How it plugs into dsh (bundle format)

This package is a bundle: package.json declares dsh.bundle.patch → ./cordis.patch.yml, and the patch inserts one plugin row that resolves the package by name. The entry module exports the standard name / inject / apply(ctx, config) contract.

Install the bundle into a profile:

dsh plugin --profile <name> add /path/to/dsh-db-connector

The package is also on npm for the standalone tooling / programmatic API:

npm install -g dsh-db-connector   # global CLI-style usage of the engine
npm install dsh-db-connector      # or add it as a local dependency

(Equivalently: add it to the profile's dependencies ("dsh-db-connector": "link:/path/to/dsh-db-connector") and append "dsh-db-connector" to dsh.profile.bundles.)

On boot, apply(ctx, config):

1. reads the plugin config (second apply argument, with $DSH_HOME

overrides and DSH_DB_CONNECTOR_* environment fallbacks);

1. pre-registers any connections from config (opened lazily on first use);
2. registers the five tools on ctx.tools;
3. registers the /db command on ctx.commands when that service exists;
4. on context disposal, closes every connection and flushes the audit log.

All registrations are effect-based; unloading the row unwinds them.

Plugin configuration

Configuration is the config: block of the plugin row (or a later patch layer overriding row id db-connector).

- insert:
    - id: db-connector
      name: 'dsh-db-connector'
      config:
        # Connections pre-registered at boot; each is opened lazily.
        connections:
          appdata:
            driver: sqlite
            database: ./data/app.db
          warehouse:
            driver: postgres
            host: db.internal
            database: warehouse
            user: readonly
            passwordEnv: WAREHOUSE_PG_PASSWORD   # env var NAME, not the value
            # ...or passwordRef: WAREHOUSE_PG_PASSWORD (dsh credentials service)

        audit:
          enabled: true
          path: .dsh-db/audit.jsonl        # default; supports ${ENV}

        query:
          maxRows: 1000                    # result row cap
          timeoutMs: 30000                 # per-statement deadline (ms)
          maxSqlChars: 512                 # statement text kept per audit record

        schema:
          ttlMs: 60000                     # snapshot cache lifetime

        defaultAllowWrite: false           # write gate global default (see below)

Environment overrides: DSH_DB_CONNECTOR_AUDIT_PATH, DSH_DB_CONNECTOR_MAX_ROWS, DSH_DB_CONNECTOR_TIMEOUT_MS. Any f"${VAR}" placeholder in a string value is expanded from the environment at connect time.

Credentials — never in logs

Security posture: secrets are referenced by name, never embedded.

• passwordEnv: PG_PASSWORD reads the value from the environment with that name.
• passwordRef: MY_REF resolves through the dsh credentials service

(ctx.credentials), falling back to a plain env lookup.

• Any field may use a ${VAR} placeholder.
• Inline password is accepted but strongly discouraged.

The bundle never logs connection configs, passwords, or connection strings: connection summaries carry only name / driver / host:port / database plus auth=env|credentials|inline|none. Error messages report environment variable names, not values. Audit records carry statement digests/summaries and counts — never connection config.

The tools

All five tools return a canonical JSON value and are rendered as formatted text to the model.

db_connect

Register, open, list or close a named connection.

{ "action": "connect", "name": "app",
  "config": { "driver": "sqlite", "database": "./data/app.db" } }

{ "action": "connect", "name": "wh",
  "config": { "driver": "postgres", "host": "db.local", "database": "wh",
              "user": "readonly", "passwordEnv": "WH_PASSWORD" } }

{ "action": "list" }
{ "action": "close", "name": "app" }

Connecting a name that is already defined (config-provided or previously connected) reopens it instead of failing; connecting a brand-new name fails loudly if the database is unreachable. list returns redacted status.

db_schema

Introspect a connection. Returns tables/views, columns, indexes and foreign keys, cached per connection for schema.ttlMs; refresh: true bypasses the cache and filter narrows to table/view names containing the given substring.

{ "name": "app", "refresh": false, "filter": "user" }

db_query

Run a read-only query (SELECT / EXPLAIN / DESCRIBE / SHOW).

• Enforced read-only: any write or DDL statement is rejected with

READ_ONLY_VIOLATION and recorded as a denied audit entry — before any driver is touched. This covers INSERT/UPDATE/DELETE/DDL/PRAGMA, and also the sneaky forms: EXPLAIN ANALYZE <dml> (which executes its statement) and data-modifying CTEs (WITH x AS (DELETE ...) SELECT ...).

• Result cap: rows are capped at limit (or query.maxRows). For a plain

top-level SELECT without its own LIMIT a guard LIMIT is appended.

• Timeouts: timeoutMs (or the default) is enforced through an

AbortSignal; SQLite genuinely terminates via child-process teardown.

{ "name": "app", "sql": "SELECT id, email FROM users WHERE age >= ? AND age < ?",
  "params": [26, 40], "limit": 100 }

Values can be positional (params array for ?) or named (namedParams object for :name); both are bound as parameters. Placeholder count vs value count mismatches are a friendly INVALID_PARAMS error.

db_exec

Execute a statement that may write (INSERT / UPDATE / DELETE / DDL, and anything the classifier can't read).

• The write approval gate is ON by default: you must pass

"allowWrite": true — the explicit confirmation — for anything that is not a pure read. Otherwise it is denied with WRITE_NOT_ALLOWED and audited.

• The statement runs inside a transaction: **COMMIT on success, ROLLBACK on

failure** (no partial rows survive an error). The result includes affected rows and a rollback explanation.

• DDL invalidates that connection's schema cache automatically.

{ "name": "app", "sql": "UPDATE users SET age = age + 1 WHERE id = ?",
  "params": [1], "allowWrite": true }

db_audit

Read the audit trail back (metadata + statement digests/summaries; credentials never appear). Filters: name (connection), kind, since (ISO), limit (newest first, default 200).

{ "kind": "denied", "limit": 50 }

The human command (/db)

When a commands service is mounted (dsh-base mounts one), the bundle registers one slash command mirroring the tools through the same engine and gates.

/db status
/db connect <name> --driver sqlite --db ./data/app.db
/db connect pg --driver postgres --host h --database d --user u --password-env PG_PASSWORD
/db close <name>
/db schema <name> [--refresh] [--filter sub]
/db query <name> --sql "SELECT ..." [--limit n] [--timeout ms] [--params a,b,c]
/db exec <name> --sql "UPDATE ..." --allow-write [--params a,b,c] [--timeout ms]
/db audit [name] [--kind k] [--limit n] [--since ISO]
/db help

The SQL audit trail

Each entry is one JSON line:

{"id":"m2x3st-abc123-1","ts":"2026-08-20T00:00:00.000Z","connection":"app",
 "kind":"write","way":"tool",
 "statement":{"summary":"UPDATE users SET age = age + 1 WHERE id = ?",
              "digest":"<sha256>","chars":47},
 "rows":1,"durationMs":14,"status":"ok"}

kind is query | write | ddl | read | schema | denied; status is ok | error | denied. Denied and failed calls are recorded too. The file is append-only JSONL at audit.path (default .dsh-db/audit.jsonl under the working directory; DSH_DB_CONNECTOR_AUDIT_PATH / ${VAR} work here too).

Driver notes

• SQLite — built-in node:sqlite, no install needed. Each connection owns

a persistent child process so a runaway synchronous statement can be hard-terminated on timeout (a worker thread stuck in native SQLite code cannot be joined, which would hang the host). File-backed databases survive a timeout teardown; :memory: connections are intentionally best-effort for testing.

• PostgreSQL — npm i pg (peer, optional). Reads run inside

BEGIN TRANSACTION READ ONLY … ROLLBACK, writes inside BEGIN/COMMIT/ROLLBACK, all parameterized with $1..$n. AbortSignal is forwarded to the client. Note: the JSONB ? operator is indistinguishable from a ? placeholder; prefer #>, ->, or @> for JSONB expressions.

• MySQL — npm i mysql2 (peer, optional). Reads run inside a

READ ONLY transaction; writes inside beginTransaction/commit/rollback, using server-side prepared statements. Cancellation destroys the connection; it reconnects on the next use.

Development

npm install            # dev deps (typescript, @types/node); optional drivers via npm i pg / mysql2
npm run build          # tsc -> dist/
npm test               # build + full suite (node --test) — 90 tests
npm run check          # build + typecheck both src and test

The test suite covers: statement classification / read-only rejection, the write approval gate, transaction rollback, parameter-injection safety, result caps and SELECT guard-LIMIT, timeouts, the audit trail, connection lifecycle, the tools, the /db command, and the plugin entry.

Limitations and notes

• One SQL statement per call; multi-statement input is refused.
• The SQL scanner follows ANSI string escaping ('') and handles

double-quoted identifiers, backtick identifiers, PostgreSQL dollar-quoted strings, -- / /* */ comments, and :: / :=. Backslash-escaped quotes inside MySQL single-quoted strings are recognized only approximately; because classification is used to *reject* writes rather than to allow them, this cannot widen the write surface.

• truncated: true can also be reported when a query naturally returns exactly

limit rows while the guard LIMIT applied (documented ambiguity).

• PostgreSQL/MySQL introspection and execution are implemented but exercised

only against real servers; the SQLite path is fully covered by the tests.

License

MIT — see LICENSE. Found a bug or want a new driver? Open an issue at github.com/JohnXu22786/db-connector.