signoz-modifying-dashboards
Этот навык вызывает инструменты сервера SigNoz MCP (signoz:signoz_get_dashboard, signoz:signoz_update_dashboard, signoz:signoz_list_dashboards, signoz:signoz_list_metrics). Перед запуском рабочего процесса убедитесь, что инструменты signoz:signoz_* доступны. Если их нет, сервер SigNoz MCP не установлен или не настроен — остановитесь и укажите пользователю на необходимость его настройки: https://signoz.io/docs/ai/signoz-mcp-server/. Не используйте прямые HTTP-вызовы и не редактируйте JSON панелей вручную без инструментов MCP.
npx skills add https://github.com/signoz/agent-skills --skill signoz-modifying-dashboardsDashboard Modify
Prerequisites
This skill calls SigNoz MCP server tools (signoz_get_dashboard,
signoz_patch_dashboard, signoz_update_dashboard, signoz_list_dashboards,
signoz_list_metrics, signoz_get_field_keys, signoz_get_field_values,
signoz_execute_builder_query).
Before running the workflow, confirm the signoz_* tools are available.
If they are not, the SigNoz MCP server is not installed or configured —
run signoz-mcp-setup first to initialize or repair the MCP connection. Do not
fall back to raw HTTP calls or hand-edit dashboard JSON without the MCP tools.
When to use
Use this skill when the user asks to:
- Add, remove, or edit panels/widgets on an existing dashboard
- Change a panel's query, title, type, or display settings
- Add, remove, or edit dashboard variables
- Rename or re-describe a dashboard
- Rearrange panel layout or resize panels
- Change a panel type (e.g., graph to table, value to graph)
- Add or modify thresholds on a panel
- Update tags on a dashboard
Do NOT use when:
- User wants to understand what a dashboard shows →
signoz-explaining-dashboards - User wants to create a new dashboard →
signoz-creating-dashboards
Instructions
Step 1: Identify the target dashboard
Use an id supplied directly or by dashboard resource context. A name is not an
ID: resolve every name-only request with signoz_list_dashboards, narrowing with
the filter argument (see signoz://dashboard/list-filter-guide) and paging by
offset until you have covered total. Match on spec.display.name. If multiple
dashboards match, present them and ask which one to modify; if one matches, use
its id.
Step 2: Fetch the current dashboard state
Call signoz_get_dashboard with the dashboard id to retrieve its full
configuration. This is mandatory — signoz_update_dashboard requires the
complete post-update state, and a patch still needs the real panel ids and grid
item indices. Never skip this step.
Examine the response to understand:
- Current panels and their ids (
spec.panelsis a map keyed by panel id) - Current grid item positions (x, y, width, height in the 12-column grid)
- Current variables
- Current query on each panel (exactly one per panel)
- The
spec.layoutsstructure — one Grid per section, each with its own items
Step 3: Plan the modification
Based on the user's request, plan the changes.
Confirm with the user before applying if:
- The modification is destructive — removing panels, deleting variables,
replacing an entire query with a different one, changing a query's
signal(e.g., traces → logs), or fundamentally altering what data is shown (changing aggregation from p99 to avg, removing groupBy dimensions) - The request is ambiguous — multiple panels could match "the latency panel"
- The change is large — restructuring sections, adding many panels at once
Destructive means data loss or silent behavior change. Even if the user says "just do it quickly," a brief confirmation ("I'll remove 'Memory Fragmentation' permanently — OK?") takes seconds and prevents irreversible mistakes. User urgency does not override this guardrail.
Non-destructive changes need no destructive confirmation: renaming, adding a panel or variable, changing a unit or panel type, adjusting layout, and adding thresholds. Variable additions still require the panel-applicability prompt below.
Compound modifications: When a request involves multiple changes (e.g., remove a panel + add a panel + rename), plan all changes against the fetched state and apply them as a single write — one patch array may carry many ops. Do not apply and re-fetch between changes.
Step 4: Apply the modification
Pick the write tool first. Default to signoz_patch_dashboard (RFC 6902):
it sends only the changed ops, so it cannot drop a panel that was never in
question. Read signoz://dashboard/patch-instructions for the JSON Pointer
paths. Reach for signoz_update_dashboard — a full replacement merged into the
Step 2 state — only when most of the dashboard changes. The rules below apply to
both; each names its patch path.
Ops apply in sequence, against the document as the previous ops left it. When one patch removes several entries from the same array — grid items, layouts, variables, tags, or a composite's member queries — order those removals by descending index. Otherwise each removal reindexes the entries after it, and a stale index quietly drops the wrong entry or no-ops (remove on a missing path is not an error). Indices planned against the Step 2 state are only valid for the first removal from each array.
Modification rules:
-
Time range and refresh are viewer controls. Dashboard payloads do not persist a default time range or refresh interval. If asked to change one, explain that panels follow the viewer-selected global range; do not invent
timeRange,defaultTimeRange, orrefreshfields. -
Preserve supported mutable state. Copy the fetched dashboard, change only what the user requested, and compare semantics after MCP normalization. Do not drop unrelated panels, variables, grid items, or sections.
-
Preserve panel/grid-item identity. Every panel id in
spec.panelsneeds exactly one grid item whosecontent.$refnames it; add or remove both together. Reuse existing panel ids verbatim — a rename is a display change, not a new id. -
Read schemas before every write. Read all required and applicable conditional resources named by the write tool. For Query Builder, also read
signoz://metrics-aggregation-guide,signoz://traces/query-builder-guide, orsignoz://logs/query-builder-guidefor the signal. MCP is the source of truth. -
Adding a panel:
- Build the panel with a short, stable id and
addit at/spec/panels/<id>. When the dashboard already holds a panel of the target type, start from that panel's JSON in the Step 2 response instead of composing one — the same diff-and-merge principle the update path follows, and the server's own shape is known-valid. Re-point every field that identifies the data (aggregation,groupBy,filter.expression,order,legend, alias,signal); a clone with one stale field is well-formed and will be accepted, so the mandatory dry-run below is what catches it. Compose fromsignoz://dashboard/widgets-examplesonly when no panel of that type exists. - Pick the Grid the user meant: sections are separate entries in
spec.layouts, each with its ownspec.display.titleand its own item coordinates, so adding to an earlier section touches only that Grid. adda grid item at/spec/layouts/<n>/spec/items/-whosecontent.$refis#/spec/panels/<id>, obeying the bounds below. Unless side-by-side placement is explicit, place it atx: 0,y: max(y + height)within that Grid. A panel with no grid item is accepted and never renders.
- Build the panel with a short, stable id and
-
Removing a panel: Remove the grid item (0-based index within its Grid) and the
/spec/panels/<id>entry together; an item left pointing at a removed panel is rejected. Do not try to auto-compact or shiftypositions of remaining panels — the SigNoz frontend grid engine handles gap-closing automatically. Simply remove the two references (panel entry, grid item) and leave all other positions unchanged. -
Editing a panel's query: Replace the query at
/spec/panels/<id>/spec/queries/0and keep all other panel fields intact — never append a second one, since a panel holds exactly one query. -
Changing panel type: Update the plugin
kindand the query envelopekindtogether (signoz/TimeSeriesPanel+time_series→signoz/TablePanel+scalar): follow the target type's complete shape inwidgets-examples. Preserve the existing query and signal; change only visualization-specific fields. -
Adding/editing variables:
- For ambiguous or version-sensitive attributes, call
signoz_get_field_keysand optionallysignoz_get_field_valueswith the relevant signal andfieldContext=resource. Trust the discovered key (for example,deployment.environmentversusdeployment.environment.name). - Show the panel list and ask whether the variable applies to all panels or a selected subset.
- Use a
ListVariablewith asignoz/DynamicVariableplugin for an attribute-backed dropdown, appended at/spec/variables/-. Itsspec.nameis the$handle;plugin.spec.nameis the attribute key. - Add
$<name>only to the selected panels'filter.expression, preserve unselected panels, and dry-run every query changed by the variable.
- For ambiguous or version-sensitive attributes, call
-
Rearranging layout / side-by-side placement: patch the grid items at
/spec/layouts/<n>/spec/items/<i>; the panels map is untouched.- SigNoz uses a 12-column grid, never 24: every entry must satisfy
0 <= x < 12,1 <= width <= 12, andx + width <= 12. - Two panels side-by-side: each gets
width: 6, first atx: 0, second atx: 6, sameyandheight. - Three panels in a row:
width: 4atx: 0,x: 4,x: 8. - When resizing an existing panel to make room, update its
widthandx, then place the new panel in the freed space at the samey. - Common heights:
height: 6for graphs/tables,height: 2–3for value panels.
- SigNoz uses a 12-column grid, never 24: every entry must satisfy
Dry-run modified panels (mandatory). For every added or changed query-bearing
panel, read the compact
dashboard-to-query-builder-v5 reference.
The panel already stores the execution spec, so lift it into the outer envelope and
call signoz_execute_builder_query with that payload. Dry-run over a short
absolute Unix-ms window — usually the last 30-60 minutes, never the panel's
display range by reflex; apply the reference's dry-run hygiene rules before
widening or retrying after a timeout. Use representative variable values in the
dry-run copy and keep $var in saved state.
Preserve or add explicit result bounds on every changed builder query/formula:
the saved spec carries a positive limit plus non-empty order, and the dry-run
sends those same values unchanged. List
and trace-request panels default to 100 rows ordered by timestamp desc (raw logs
also id desc), but preserve a deliberate smaller positive list limit.
Standalone aggregate queries and formula outputs default to
100 groups. Every base query referenced by a formula uses 10000 because base
limits are applied before formula evaluation; raise an existing smaller bound
unless it was an intentional pre-formula top-N selection. Find the complete
base-query set by inspecting every formula expression, including formulas with
disabled: true, and following formula references to all builder_query
leaves. This dependency walk chooses bounds only; it does not establish
deterministic formula-to-formula evaluation order, so dry-run the complete
composite payload. Base queries order by their primary aggregation and formulas
by __result, in the saved panel and the dry-run alike. For time series, this
top-N is chosen over the whole window, so a short-lived local spike can be
omitted. Narrow filters/grouping if formula-input cardinality can exceed 10000.
If the reference's safety gate finds an unsupported execution field, report the panel as unvalidated and continue only after explicit acceptance. Server or validation errors and unexpected empty results block unless explicitly accepted.
Call signoz_patch_dashboard with the id and an op array; the result is
validated after the ops apply, and locked dashboards are rejected:
signoz_patch_dashboard({
"id": "<dashboard-id>",
"patch": [{"op": "replace", "path": "/spec/display/name", "value": "New title"}]
})
If the write fails, never resend the same payload. A client-side
InputValidationError never reached SigNoz, so an identical retry cannot
succeed — and working out why it failed does not license another attempt.
That reasoning is what turns two failures into twenty. Allow at most two
attempts per approach, then change approach: clone-and-re-point rather than
author, or split one patch into several. If two approaches have failed, stop
and report the exact ops and error text to the user instead of trying a third.
This bounds client-side validation failures only — a dry-run timeout is a
server-side condition where retrying over a narrower window is correct.
For a full replacement, signoz_update_dashboard takes the merged state flat
beside id, not nested under a dashboard key. Merge into the data object of the
Step 2 response — passing the {status, data} envelope itself is rejected. Send the
fetched name — an immutable machine label — back unchanged, and drop the read-only
fields the GET returns (createdAt, updatedBy, orgId, webUrl, and friends):
signoz_update_dashboard({
"id": "<dashboard-id>",
"schemaVersion": "v6",
"name": "<the fetched name, unchanged>",
"tags": [...],
"spec": <complete merged spec from signoz_get_dashboard>
})
Step 5: Report the result
Briefly tell the user what was changed. Offer further modifications if relevant.
Guardrails
- Patch first; full state on update: prefer
signoz_patch_dashboardfor targeted edits. When a full replacement is warranted,signoz_update_dashboardtakes the complete dashboard flat besideid(schemaVersion,name,tags,spec) — never nested under adashboardkey. Always callsignoz_get_dashboardfirst, merge into that response'sdataobject, round-tripnameunchanged, and never construct a payload from scratch. - Preserve what you don't change: Preserve supported mutable semantics for panels, variables, and grid items outside the request. Diff-and-merge; do not rebuild or promise byte-for-byte equality after MCP normalization.
- Confirm destructive changes: Before removing panels, replacing queries, or deleting variables, confirm with the user — even if they say "just do it" or express urgency. Additions, renames, type changes, and variable additions do not need confirmation.
- Validate changed queries Follow the mandatory dry-run step above before the write.
- Valid JSON only: Follow the schema documented in the
signoz://dashboard/*MCP resources (instructions,widgets-instructions,widgets-examples,query-builder-example). Never generate malformed queries or layouts. - OTel attribute names: Use
service.namenotserviceandhost.namenothost, but discover version-sensitive keys such asdeployment.environmentversusdeployment.environment.nameinstead of forcing one form. - No metric guessing: If adding or changing queries and you are not sure what
metrics are available, ask the user or call
signoz_list_metricsto discover available metrics. Wrong metric names produce empty panels. - Paginate dashboard listing: When searching for a dashboard by name, always
page through the whole
signoz_list_dashboardsresult before concluding a dashboard does not exist. - Identifiers: Give new panels short, stable ids and reference them from the
grid item's
content.$ref; keep each variable'sspec.nameidentical to the$handleits queries use, and keep query names such asA,B, andF1stable. - Real dashboard IDs only: Never send a sentinel such as
"unused"as a dashboard id. Resolve it throughsignoz_list_dashboardsandsignoz_get_dashboardfirst. - Scope boundary: This skill modifies existing dashboards. Hand new-dashboard
requests to
signoz-creating-dashboards.
Examples
User: "Add an error rate panel to my Redis dashboard"
Agent:
- Calls
signoz_list_dashboardswithfilter="redis"— finds "Redis Overview" dashboard with idabc-123. - Calls
signoz_get_dashboardwith idabc-123— gets full configuration with 8 existing panels across two Grid sections. - Calls
signoz_list_metricsto find available Redis error metrics. - Creates a new timeseries panel (a
signoz/CompositeQueryholding the two disabled base queries and the error-rate formula) and a grid item for it below the last item in the Overview Grid. - Dry-runs the new panel with
signoz_execute_builder_query; fixes any error or unexpected empty result. - Calls
signoz_patch_dashboardwith two ops —add /spec/panels/error-rateandadd /spec/layouts/0/spec/items/-— rather than re-sending all 9 panels. - Reports: "Added an 'Error Rate' timeseries panel to your Redis Overview dashboard under the Overview section. Want me to adjust anything?"
User: "Change the latency panel from a graph to a table on my API dashboard"
Agent:
- Resolves "API Monitoring" through a filtered
signoz_list_dashboards, then callssignoz_get_dashboardwith its id. - Finds the panel titled "Request Latency" — if multiple panels could match, confirms with user: "I found 'Request Latency'. Convert that one to a table?"
- Changes the plugin
kindtosignoz/TablePaneland the querykindtoscalar, matches the table shape inwidgets-examples, and keeps the query intact. - Dry-runs the table execution shape because panel type affects request shape.
- Calls
signoz_patch_dashboardwith replace ops on that panel's plugin and query kind. - Reports: "Changed 'Request Latency' from a graph to a table. Want me to adjust column widths or add column units?"
User: "Remove the CPU panel and rename the dashboard to 'Service Health'"
Agent:
- Fetches the dashboard via
signoz_get_dashboard. - Finds the "CPU Usage" panel. Confirms: "I'll remove the 'CPU Usage' panel and rename the dashboard to 'Service Health'. Proceed?" (Removal is destructive — always confirm.)
- User confirms.
- Removes the panel entry and its grid item. Leaves all other panel positions
unchanged (the frontend grid closes gaps automatically). Renames via
/spec/display/name, never the immutable top-levelname. - Calls
signoz_patch_dashboardwith all three ops in one array. - Reports: "Removed the 'CPU Usage' panel and renamed the dashboard to 'Service Health'. Anything else to adjust?"