dsh-grafana-query
用于 DeepSeek Harness 的只读 Grafana 指标和告警工具。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:maxmilian/dsh-grafana-query#18e9e4b4d6b5efb3181836a5bbb05bf6b0fbfd4eGrafana Query 以 dsh-grafana-query 发布，当前版本为 0.1.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
dsh-grafana-query is a free, open-source, read-only DeepSeek Harness plugin for Grafana.
It lets an agent run PromQL through Grafana's data source proxy and read the state of
Grafana unified alerting, without changing anything in Grafana.
Not to be confused with dsh-grafana on npm, which is a dashboard editor that writes
dashboard JSON back to Grafana. This plugin does the opposite job: read-only metric queries
and alert state. Dashboard and panel JSON are explicitly out of scope.
| Tool | Purpose |
|---|---|
grafana_health | Check that the instance is reachable and report its version. |
grafana_list_datasources | List data sources with uid, type, and access mode. Run this first. |
grafana_query | Run an instant PromQL query through the data source proxy. |
grafana_query_range | Run a range PromQL query with an enforced step and point budget. |
grafana_alert_state | Read the current state of unified alerting rules. |
grafana_list_alert_rules | List provisioned alert rule definitions. |
All tools are read-only. Version 0.1 never creates, edits, deletes, silences, acknowledges, or pauses anything in Grafana.
Every limit below is enforced by the plugin, not by Grafana. Whenever something is trimmed,
meta.truncated and the pre-truncation totals say so.
| Limit | Value |
|---|---|
Points per series (max_points) | 200 by default, 500 maximum. Prometheus returns both endpoints, so a range of n seconds at step s yields floor(n / s) + 1 points |
Range length (grafana_query_range) | 31 days |
| Total points in one range response | 20000; series past that are dropped whole, never cut in half |
| Series per query | maxSeries, 100 by default |
Alert rules per tool (grafana_alert_state, grafana_list_alert_rules) | 500 matching rules; the rest cannot be reached by paging, so use the filters |
| Alert instances per rule | 10 by default, 50 maximum |
| Page size | 20 by default, 100 maximum |
| Upstream error text | 200 characters, HTTP 400 only |
grafana_alert_state returns firing, pending, and unknown rules by default — inactive
rules are excluded unless you ask for them with state.
@deepseek-ai/dsh-tools APIs/api/datasources/proxy/uid/:uid/*)
is supported; the deprecated numeric-id path is notexport GRAFANA_URL='https://grafana.example.com'
export GRAFANA_TOKEN='glsa_your_service_account_token'
| Field | Environment variable | Default | Range |
|---|---|---|---|
baseUrl | GRAFANA_URL | required | HTTP(S) URL, no credentials, no query or fragment; a sub-path is fine |
token | GRAFANA_TOKEN | required | non-empty |
locale | — | en | en, zh-TW, zh-CN, ja |
requestTimeoutMs | — | 30000 | 1 – 300000 |
maxResponseBytes | — | 5242880 | 1 – 52428800 |
maxSeries | — | 100 | 1 – 1000 |
Plugin configuration takes precedence over the environment variables.
A Grafana service account token (recommended) or a legacy API key both work — they use the
same Authorization: Bearer header. A Grafana Cloud Access Policy token (glc_) is for the
Cloud data endpoints and does not work with this API.
The scope names in the table below are what Grafana checks internally — they are not what you tick in the UI. When you create the service account, the combination that works is:
datasources:read and datasources:query.alert.rules:read and
alert.provisioning:read.Verified on Grafana Cloud on 2026-08-27 with exactly that combination: all six tools worked. See the verification note.
A least-privilege token stays usable. Grafana filters GET /api/datasources by what the
token may reach rather than returning 403, so under a token granted Query on a single
data source, grafana_list_datasources returns just that one — measured 2026-08-27: 26
entries for a Viewer token, 1 for the restricted one. You never get a list full of data
sources that would fail when queried. (A per-data-source Query grant also implies metadata
read on that data source, so there is no "can query but cannot read" state to worry about.)
| Tool | Required permission |
|---|---|
grafana_health | none — /api/health needs no authentication, so this tool cannot tell you whether the token is valid. Use grafana_list_datasources for that. |
grafana_list_datasources | datasources:read |
grafana_query, grafana_query_range | datasources:query (plus datasources:read for pre-flight type checks) |
grafana_alert_state | alert.rules:read |
grafana_list_alert_rules | alert.provisioning:read |
Point baseUrl at the stack itself and use a service account token created in that stack:
export GRAFANA_URL='https://your-stack.grafana.net'
export GRAFANA_TOKEN='glsa_your_service_account_token'
Do not use a glc_ Access Policy token here. Cloud stacks ship many built-in data sources,
so use the type and name_contains filters of grafana_list_datasources to keep the list short.
bun add dsh-grafana-query
The package ships cordis.patch.yml, declared through dsh.bundle.patch in package.json,
so the DeepSeek Harness registry can load the plugin with its default configuration.
grafana_list_datasources with {"type": "prometheus"} to find the uid.grafana_query with {"datasource_uid": "prom-1", "query": "up"} for the current value.grafana_query_range with {"datasource_uid": "prom-1", "query": "rate(node_cpu_seconds_total[5m])", "start": "...", "end": "..."}
for the trend. Omit step and the plugin picks one so each series stays within max_points.grafana_alert_state with no arguments to see what is firing right now.Set locale to en, zh-TW, zh-CN, or ja to change the tool and parameter descriptions the
model sees. Tool names always stay in English, and error messages are always in English.
Authorization header, or a raw response body.error
field is passed through so the agent can fix its PromQL. It is capped at 200 characters and
runs through a redaction pass first. Every other status code returns a static message.maxResponseBytes, maxSeries, and a per-series point budget.
Whenever anything is trimmed, meta.truncated and the pre-truncation totals say so.bun install
bun run lint
bun run typecheck
bun run test
bun run build
MIT