Query

A Query is a request to run work on an Ark resource. The CRD ships an input plus exactly one of target (a single agent / team / model / tool) or selector (picks a single target by label — first match wins, not a fan-out). For a task-oriented walkthrough see User Guide → Run queries / chat with agents and teams; this page is the field-by-field reference.

Spec


apiVersion: ark.mckinsey.com/v1alpha1
kind: Query
metadata:
  name: example-query
spec:
  # --- Required ----------------------------------------------------------
  input: "What's the weather in {{.city}}?"   # user message (string)
 
  # Exactly ONE of target or selector is required.
  target:
    type: agent                               # agent | team | model | tool
    name: weather-agent
 
  # OR pick a single target by label (first match wins; not a fan-out).
  # selector:
  #   matchLabels:
  #     role: reviewer
 
  # --- Common optional fields --------------------------------------------
  type: user                                  # only 'user' is accepted (default)
  parameters:                                 # template values for {{.name}} in input
    - name: city
      value: Boston
  conversationId: chat-alice-001              # thread queries into one conversation
  memory:
    name: broker                              # Memory resource that stores history
  sessionId: user-session-123                 # telemetry / tracking key (separate from conversationId)
  serviceAccount: my-sa                       # RBAC for execution
  timeout: 5m                                 # default '5m'
  ttl: 720h                                   # default from ArkConfig/default (720h fallback)
  cancel: false                               # set to true to cancel mid-flight
  overrides:                                  # HTTP header overrides for downstream model/MCP calls
    - resourceType: model
      headers:
        - name: X-Trace-ID
          value:
            value: trace-xyz789

Fields

Field	Type	Required	Description
`spec.input`	string	yes	The user message. Supports Go template substitution via `{{.name}}` against `spec.parameters`.
`spec.target`	object	conditional	Single execution target. Either `target` or `selector` must be set — the validation webhook rejects queries with neither.
`spec.target.type`	enum (`agent`, `team`, `model`, `tool`)	yes (when `target` set)	Resource kind to dispatch to.
`spec.target.name`	string	yes (when `target` set)	Name of the target resource in the same namespace.
`spec.selector`	`LabelSelector`	conditional	Pick a single target by label (alternative to `target`). Resolves to the first matching resource, checked in order Agent → Team → Model → Tool. Not a fan-out — it does not run against multiple matches. Empty match set fails with `no matching resources found for selector`.
`spec.type`	enum (`user`)	no	Only `user` is accepted (default). The previous `messages` type is deprecated; the mutating webhook migrates it to `user` and stamps a `migration-warning-input-type` annotation.
`spec.parameters[]`	list	no	Template values substituted into `input` (and into any agent prompt that references `queryParameterRef`). Each entry has a `name` and either a `value` or `valueFrom` (`configMapKeyRef`, `secretKeyRef`).
`spec.memory.name`	string	no	Name of a `Memory` resource that stores conversation history keyed by `conversationId`. If unset, the completions executor falls back to a `Memory` named `default` in the same namespace; if that doesn’t exist either, it uses a no-op (in-process) store.
`spec.sessionId`	string	no	Identifier for grouping queries in tracking / telemetry. Auto-derived from the Query UID if unset. Not the same as `conversationId` — `sessionId` is observability, `conversationId` is memory continuity.
`spec.conversationId`	string	no	Conversation threading identifier. Sent to execution engines as the A2A `contextId`. With memory active, ties subsequent queries to the same history. Auto-generated by the memory backend if unset (a new conversation is created and the ID written back to `status.conversationId`).
`spec.serviceAccount`	string	no	Kubernetes ServiceAccount used for RBAC during execution.
`spec.timeout`	duration	no	Execution timeout (`30s`, `5m`, `1h`). Defaults to `5m`. Applies to all targets in the query.
`spec.ttl`	duration	no	How long the Query resource lingers after completion. Mutating webhook resolves the default from `ArkConfig/default.spec.queryTTL` (falls back to `720h`).
`spec.cancel`	bool	no	When `true`, indicates intent to cancel a running query. The controller transitions the query to `phase: canceled`.
`spec.overrides[]`	list	no	HTTP header overrides applied when the query’s target calls its model or any MCP server. See Overrides.

Targets

A query dispatches to exactly one target — named directly via spec.target, or picked by label via spec.selector. Supported type values: agent, team, model, tool.


spec:
  input: "Summarise the news."
  target:
    type: team
    name: news-team

Pick by label (resolves to a single target — first match across Agent → Team → Model → Tool):


spec:
  input: "Should we ship this feature?"
  selector:
    matchLabels:
      role: reviewer

The webhook enforces that one of target / selector is set, and that the named resource exists when target is used.

Parameters

Parameters are substituted into spec.input (and into agent prompts that declare queryParameterRef) before dispatch. Sources:


spec:
  input: "Hit {{.endpoint}} with token {{.key}}, mode {{.mode}}."
  parameters:
    - name: mode
      value: "production"                     # direct value
    - name: endpoint
      valueFrom:
        configMapKeyRef:                      # from ConfigMap
          name: api-config
          key: url
    - name: key
      valueFrom:
        secretKeyRef:                         # from Secret
          name: api-credentials
          key: api-key

Template syntax is Go templates: {{.parameter_name}}. Names must match the parameters[].name field.

Wiring parameters into the agent

Agents can pull the same query parameters into their own prompt or system message using queryParameterRef:


apiVersion: ark.mckinsey.com/v1alpha1
kind: Agent
metadata:
  name: dynamic-agent
spec:
  prompt: |
    You are running in {{.mode}} mode against {{.environment}}.
  parameters:
    - name: mode
      valueFrom:
        queryParameterRef:
          name: operation_mode
    - name: environment
      valueFrom:
        queryParameterRef:
          name: target_env

The matching query supplies them:


apiVersion: ark.mckinsey.com/v1alpha1
kind: Query
metadata:
  name: parameterised-q1
spec:
  input: "Run a health check."
  parameters:
    - name: operation_mode
      value: "production"
    - name: target_env
      value: "us-west-2"
  target:
    type: agent
    name: dynamic-agent

See Agents → Pass parameter values at query time for the dashboard walkthrough.

Conversations + memory

conversationId threads queries together; Memory stores the history. By default the completions executor uses the Memory named default in the query’s namespace when spec.memory is unset — most Ark installs ship one backed by the in-cluster broker. Two queries with the same conversationId share context as long as they hit the same memory store:


apiVersion: ark.mckinsey.com/v1alpha1
kind: Query
metadata:
  name: chat-q1
spec:
  input: "Hi, my name is Alice."
  conversationId: chat-alice-001
  memory:
    name: broker
  target:
    type: agent
    name: assistant
---
apiVersion: ark.mckinsey.com/v1alpha1
kind: Query
metadata:
  name: chat-q2
spec:
  input: "What's my name?"
  conversationId: chat-alice-001            # same conversation
  memory:
    name: broker
  target:
    type: agent
    name: assistant

sessionId is a separate identifier for telemetry — same telemetry session can contain many distinct conversations.

If you set conversationId without spec.memory, the executor still resolves the default Memory (so continuity works out of the box on a standard install). The conversationId also flows through to execution engines as the A2A contextId — named engines (Claude Agent SDK, LangChain, …) can use it for their own session model.

If you set memory without conversationId, the controller auto-generates a fresh conversationId on each query.

Timeout

spec.timeout caps execution. Default is 5m. Format follows Go’s time.ParseDuration (30s, 5m, 1h).

Completes within timeout → status.phase: done.
Exceeds timeout → status.phase: error with a timeout error message in status.response.content.

For A2A agents the timeout flows through as a Go context deadline. A2A servers can also enforce their own per-server timeouts via A2AServer.spec.timeout — see Building A2A Servers → Timeout configuration.

Cancel

To cancel a running query, patch spec.cancel: true:


kubectl patch query slow-q1 --type='merge' -p='{"spec":{"cancel": true}}'

The controller transitions the query to phase: canceled once the in-flight dispatch acknowledges. The dashboard chat panel does the same when you click Stop on an embedded chat.

Deletion and cleanup

When a Query is deleted, a controller finalizer removes the messages and operation events that query wrote to the broker, so deleting a Query does not leave orphaned history behind.

Messages go through the Memory contract:

Memory resolution mirrors execution: the controller uses spec.memory if set, otherwise a Memory named default in the query’s namespace. If neither exists, cleanup is skipped.
The controller resolves the backend address from the Memory resource (status.lastResolvedAddress, falling back to spec.address) and issues DELETE /queries/{queryId}/messages against it — see the Memory delete endpoint.

Operation events are not part of the Memory contract, so they use a separate path:

The controller discovers the broker endpoint from the ark-config-broker ConfigMap and issues DELETE /events/{queryId} directly against it. If no broker is configured for the namespace, cleanup is skipped.
Events are keyed by the Query’s UID, not its name (unlike messages), so {queryId} here is the UID.

Both cleanups run independently and either can fail without blocking the other. A backend that does not implement a delete endpoint returns 404/405, which the controller treats as “not supported” and skips. Scope: messages and events only — chunks and traces are not affected. Failure policy: if a backend is unreachable the finalizer retries every 15s; after a 5-minute grace period it gives up and removes itself, so a Query is never stuck in Deleting.

This is complementary to spec.ttl: TTL expires history on a timer, while deletion cleans it up immediately when the Query is removed.

Status


status:
  phase: done
  conditions:
    - type: Completed
      status: "True"
      reason: QuerySucceeded
      message: Query completed successfully
  conversationId: ctx-86e022b0-1f4f-49f5-bc4b-e0d6437a789b
  duration: 1.82s
  response:
    target:
      type: agent
      name: weather-agent
    content: "It's 72°F and sunny in New York"
    raw: '[{"content":"...","role":"assistant","name":"weather-agent"}]'
    phase: done
    a2a:
      contextId: ctx-86e022b0-1f4f-49f5-bc4b-e0d6437a789b
      taskId: task-xyz789
  tokenUsage:
    promptTokens: 97
    completionTokens: 20
    totalTokens: 117

Status fields

Field	Description
`status.phase`	One of `pending`, `provisioning`, `running`, `done`, `error`, `canceled`.
`status.conditions[]`	Standard Kubernetes conditions. The `Completed` condition flips to `True` on success or failure.
`status.response`	Single response object (singular — there is no `responses[]` array). A selector resolves to one target, so there is always exactly one response.
`status.response.content`	The final assistant message.
`status.response.raw`	The full conversation as a JSON-encoded array of messages — what the dashboard’s Debug tab renders.
`status.response.phase`	Per-response phase (mirrors top-level `phase`).
`status.response.target`	The target the response is for.
`status.response.a2a.contextId`	A2A context ID returned by the engine. When `conversationId` is set on the spec, the engine receives it as the `contextId` and echoes it back here.
`status.response.a2a.taskId`	Set when the target is an A2A agent and a task was created. References an `A2ATask` resource — see A2ATask.
`status.conversationId`	Effective conversation ID — either the one you supplied, or the auto-generated one when memory is set without a `conversationId`.
`status.duration`	Total wall-clock execution time.
`status.tokenUsage`	`promptTokens`, `completionTokens`, `totalTokens` from the underlying model call.

Phases

Phase	Meaning
`pending`	Resource accepted by admission, controller hasn’t picked it up yet.
`provisioning`	Controller is preparing the dispatch (parameter resolution, target lookup).
`running`	Dispatched to the executor; waiting for the response.
`done`	Execution finished successfully.
`error`	Execution failed; details in `status.response.content`.
`canceled`	`spec.cancel: true` was honoured.

Print columns

kubectl get queries prints Type, Phase, Duration, Age. Pipe through -o jsonpath for the content/raw fields.

Webhook validation

Enforced by ark/internal/validation/query.go:

Exactly one of spec.target and spec.selector is required.
When spec.target is set, the named resource must exist in the namespace and target.type must be one of agent, team, model, tool.
Parameter resolution must succeed (referenced ConfigMaps and Secrets must exist).
spec.overrides[] must reference real resources for any valueFrom source.

Mutating webhook

Enforced by ark/internal/validation/defaults.go:

spec.type: messages is deprecated. The mutating webhook extracts the last user-message text, sets spec.type: user, replaces spec.input with that text, and stamps an ark.mckinsey.com/migration-warning-input-type annotation. Use conversationId for multi-turn instead.
spec.ttl is defaulted from ArkConfig/default.spec.queryTTL (or 720h if absent).

User Guide → Run queries / chat with agents and teams — task-oriented walkthrough.
Agents — target type agent.
Teams — target type team.
Memory — backing store for spec.memory.
Overrides — header overrides via spec.overrides.
A2ATask — task resource referenced by status.response.a2a.taskId.