dosql · git:20261108.6af37d0 · 2026-11-08 · sha256 8cd32182cc51ca18

dosql git:20261108.6af37d0A

Immutable. This exact content never changes and is served at /api/v1/blob/8cd32182cc51ca18.

No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_jlxplg73y3fa2ldq/8cd32182cc51ca189e74dd5ce9f5a5da9040700168100e68146df0d72d639cf2.json

Then /plugin install dosql@markdownregistry-jlxplg73y3fa2ldq in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.

---
name: dosql
description: DoSql 数据库变更控制助手。用于通过 doagent/CLI 对 MySQL、PostgreSQL、MongoDB 执行受控查询、变更计划、迁移脚本生成、多环境迁移和审计留痕。
metadata:
  family: executable
  runtime: node
  entrypoint: scripts/dosql-agent.mjs
---

# DoSql Agent

DoSql is a skill inside the DoOps Agent runtime. The user authenticates once to
the DoOps Gateway, selects a Gateway target, and then asks doagent to route the
database task here. DoSql does not implement login, target grants, Gateway
token parsing or a second user authorization system.

The canonical skill name and DoOps domain are both `dosql`. Load `Skill name=dosql`.
Use the bound resource name or `credential_id` as `credential_ref`; `grant_id` and
`version_id` are audit metadata, not credential selectors.

The supported CLI entrypoint is `doops sql`. A generic `doops ask` prompt that
mentions DoSql does not guarantee the `dosql` skill or `mcp_dosql_*` tools are
bound. If the target is not upgraded/configured for DoSql, fail explicitly;
never replace it with `mcp_doops_k8s_read`, `kubectl`, `psql`, or a local
connection file.

The core rule is simple: do not bypass the DoSql proxy path. Every database
query, DDL, DML and MongoDB command must create an operation record before it is
executed.

## Operation Routing

Use exactly one route for each database task. Do not silently switch routes:

| User intent | Public route | What it actually does |
| --- | --- | --- |
| Change data/schema/admin state | `doops sql --mode apply --sql/--file` | Requires target exec authority and the target's bound database credential. Statements run verbatim with optional `--verify`, and every real apply takes a version in the per-database journal. Rollback-capable batches use one transaction; MySQL/Oracle DDL commits implicitly, rejects dry-run and DDL/DML mixing, and cannot be undone by a later verification failure. There is no model-driven or signed-approval mutation path. |
| Copy whole tables between two bound databases | `doops sql --mode apply --copy-from <src> --database <dst> --tables a,b` | The Agent reads and writes in batches, outside the model. |
| Plan a change | `doops sql --mode plan` | Produces a human-readable change plan, impact, risk and verification steps before the apply. Version/state claims require live evidence; supplied examples and fixture versions are not real database state. |
| Find database instances/assets | `doops sql --mode discover` or `mcp_dosql_read operation=discover` | Uses target Kubernetes Pod/Service/Deployment/StatefulSet metadata; never reads Secret values or connects to the database. |
| Inspect tables/columns/indexes/schema | `doops sql --mode query` or `mcp_dosql_read operation=query` | Runs one asset-bound catalog statement through the database client. `discover` is not a schema query. |
| Read business data | `doops sql --mode query` or `mcp_dosql_read operation=query` | Runs one statement and records bounded result/evidence. |

`register-database`, `discover-resources`, `compare-databases` and structure
snapshot commands consume supplied probe/snapshot/repository metadata. They are
not a live database schema inspection fallback. A standalone `inspect-schema`
command is not currently implemented; do not claim it exists or invoke an
unregistered MCP operation. If the user asks for structure, use a `query` with
the correct engine catalog/information-schema statement and the selected
`database_asset_id`.

For MongoDB, use a structured `mongoOperation`; raw JavaScript is not accepted.
Changes go through `apply`, not through `query`, `discover`, a generic
`doops ask`, Bash, `psql`, or a local connection file.

Direct MySQL apply owns the transaction boundary: caller transaction controls,
autocommit changes and executable comments are rejected. Dry-run rejects
implicit-commit and unsupported side-effecting syntax before execution; rollback
still requires transactional tables and does not prove that stored functions or
triggers have no external effects. Version journals identify the declared
engine, endpoint list and database, preserving history across credential
rotation. An ambiguous legacy journal blocks writes until its ownership is
explicitly reconciled; never bypass it or silently start another history.
Connection URI restrictions and the legacy migration procedure are owned by
[the usage guide](../../../docs/DOOPS_USAGE_GUIDE.md#dosql-直接执行-sql).

## Intended Capabilities

- Resolve the target project, environment and database asset.
- Discover candidate databases from target probe output, then ask the user to
  confirm business names and aliases.
- Classify the operation as read, schema change, data change, admin or unknown.
- Decide whether the operation enters the formal change document.
- Generate MySQL/PostgreSQL SQL migration scripts.
- Generate MongoDB migration scripts or controlled command documents.
- Execute read queries and migrations through the DoSql CLI.
- Use `scripts/dosql-agent.mjs` for the first implemented Agent commands:
  `classify`, `query`, `discover-databases`, `discover-resources`, `register-database`, `resolve-database`,
  `compare-databases`, `derive-change-plan-from-comparison`,
  `render-forward-change-artifacts`, `execute-forward-artifacts`,
  `render-forward-evidence-metadata-commit`, `scan`, `record-execution`,
  `inspect-journal`, `migrate-journal`,
  `replay-timeline`, timeline lookup, rollback planning,
  restore-node baseline derivation.
- Record SQL/MongoDB execution lifecycle into append-only JSONL journals before
  importing the same events into the future control-plane database.
- Return affected rows, schema snapshots, dry-run output and verification
  evidence.
- Explain each mutating change in user-readable language before execution.
- Scan database health, extract SQL logs and detect SQL injection indicators
  through read-only commands.

## Safety Rules

- Schema changes, data changes, admin operations and unknown operations always
  enter the change document and take a version.
- Queries are audited and do not enter a change document by default.
- Production changes are admitted by the Gateway target grant and
  recorded as a version in the per-database journal. There is no signed
  approval artifact.
- MongoDB commands must use structured `mongoOperation` documents. Raw
  JavaScript is never executed, and queries require a database read-only role.
- Scanning modules are read-only and must still produce operation records and
  evidence.
- A mutating execution must have JSONL lifecycle events for planned, running,
  succeeded or failed, and verified or verification_failed.
- Users confirm business intent, target, impact, risk and verification.
  Existing authorization for that scope is sufficient; do not ask again
  merely because a plan or SQL artifact was generated. See
  `docs/user-confirmation.md` for uncertainty and version evidence boundaries.
- If a database name resolves to multiple assets, ask the user to choose the
  environment or database. Do not guess.
- 不允许用兼容分支或静默降级掩盖真实失败。

## Target And Credential Boundary

- The selected `cluster/instance` comes from the DoOps Gateway target selected
  by the user. Do not read a skill-local default target or invent one from a
  product URL.
- Discover database assets from the selected target and require an explicit
  choice when more than one asset matches. `config/resources.json` is optional
  metadata for labels and examples; it is not a target registry or permission
  source.
- `config/assets.json` is the metadata-only discovered asset inventory. It is
  generated by `discover-resources` and is not an authorization source; exact
  target, database credential scope and Gateway target grant checks remain required.
- `mcp_dosql_read operation=discover` scans non-secret Kubernetes metadata
  (Pods, Services, Deployments and StatefulSets) and returns database
  candidates with a stable `database_asset_id`, engine, namespace, service,
  port and any database name visible in non-Secret configuration. It never
  reads Secret values. Candidates without a database name remain unresolved
  until the user selects the asset and provides its bound database credential.
- During a conversation, a user may reference selected database assets and
  code repositories by metadata-only IDs or `repo_id@branch` references. The
  references are task context, not authorization; Gateway target grant and
  credential scope checks remain authoritative.
- `discover-resources` may scan a repository workspace named in the task,
  including application code, connection references, Helm values and Kubernetes
  manifests. It merges PostgreSQL, MySQL and MongoDB metadata into the DoSql
  inventory, adopts an explicit Gateway `database_asset_id` from target
  discovery, deduplicates by database identity, and can project the same asset
  into `config/resources.json`. It ignores credential-like filenames and never
  persists URI userinfo, passwords, tokens or other secret payloads.
- Changes are not made through an MCP tool. `doops sql --mode apply
  --sql/--file` runs them on the Agent's direct path with a Gateway-bound
  database credential and the version journal. Transaction and rollback
  guarantees depend on the engine and statement kind; see the direct SQL
  boundary in `docs/agent-cli-contract.md`.
- Database credentials are DoOps managed credentials. Pass only credential
  resource names to the Gateway; never put payloads, connection strings or
  passwords in prompts, artifacts or logs.
- `state=active` on a credential resource does not mean it has an active
  credential version or an active target grant. The required lifecycle is
  `create -> put (staged) -> grant -> verify -> promote`. Query/execute is
  admitted only after all four are complete. The payload for PostgreSQL and
  MongoDB is opaque JSON with `data.connection_uri`; MySQL uses the exact
  datasource keys in `CREDENTIAL_RESOURCES.md`.
- `credential list --view managed` reports resource metadata only and does not
  load grant details. Its grant field must not be interpreted as proof that a
  target grant is absent; use the target view after an active version exists or
  inspect the explicit grant response.
- Gateway target grant checks happen before the skill runs. DoSql's operation
  journal and version records protect database change evidence after
  admission and must not duplicate Gateway authentication.

## Detailed Contract

- `docs/install-management.md`
- `docs/agent-cli-contract.md`
- `docs/journal-recovery.md`
- `docs/user-confirmation.md`
- `docs/scanning.md`
- `docs/mysql.md`
- `docs/postgres.md`
- `docs/mongodb.md`