AI Action

AI Action allows you to integrate AI capabilities into your playbooks by leveraging the rich context that Mission Control maintains about your infrastructure. When executed against configs or components, the AI action automatically injects:

Component/config manifests and specifications
Related configurations with configurable relationship depth
Historical analysis data within the specified time period
Change history across related infrastructure within defined time ranges

This comprehensive context enables AI models to provide more informed analysis and insights about your infrastructure state and relationships. For example, when a Kubernetes pod fails, it examines the pod spec, ConfigMap changes, service logs together, revealing patterns that single-component analysis might overlook.

context-provider-playbook.yaml
# Source: mission-control-playbooks-ai/templates/recommend-playbooks.yaml
apiVersion: mission-control.flanksource.com/v1
kind: Playbook
metadata:
  name: recommend-playbook
spec:
  title: Recommend Playbooks
  description: Diagnoses the health of a resource using AI, and then recommends playbooks to fix the issue, sending the results to Slack
  icon: bot
  category: AI
  configs:
    - name: '*'
  parameters:
    - name: prompt
      label: Prompt
      default: Find out why $(.config.name) is unhealthy
      properties:
        multiline: 'true'
  actions:
    - name: analyse
      ai:
        formats:
          - recommendPlaybook
        recommendPlaybooks:
          - search: category!=AI
        connection: 'connection://mission-control/anthropic'
        systemPrompt: 'You are an experienced Kubernetes engineer and diagnostic expert. Your task is to analyze Kubernetes resources and provide a comprehensive diagnosis of issues with unhealthy resources. You will be given information about various Kubernetes resources, including manifests and related components.

          Please follow these steps to diagnose the issue:

          1. Thoroughly examine the manifest of the unhealthy resource. 2. Consider additional related resources provided (e.g., pods, replica sets, namespaces) to gain a comprehensive understanding of the issue. 3. Analyze the context and relationships between different resources. 4. Identify potential issues based on your expertise and the provided information. 5. Formulate clear and precise diagnostic steps. 6. Provide a comprehensive diagnosis that addresses the issue without requiring follow-up questions.

          Before providing your final diagnosis, show your thought process and break down the information. This will ensure a thorough interpretation of the data and help users understand your reasoning.

          - Identify the unhealthy resource(s). - Examine relationships between resources, noting any dependencies or conflicts. - Consider common Kubernetes issues and check if they apply to this situation. - Formulate hypotheses about potential root causes. '
        prompt: '$(.params.prompt)'
        changes:
          since: 24h
        analysis:
          since: 1d
        relationships:
          - depth: 3
            direction: outgoing
            changes:
              since: 24h
            analysis:
              since: 1d
          - depth: 5
            direction: incoming
            changes:
              since: 24h
            analysis:
              since: 1d
    - name: send recommended playbooks
      notification:
        connection: 'connection://mission-control/slack'
        title: Recommended playbooks
        message: '$(getLastAction.result.recommendedPlaybooks)'

Field	Description	Scheme
`name*`	Step Name	`string`
`ai`	AI Action	AI
`delay`	A delay before running the action e.g. `8h`	`Duration` or CEL with Playbook Context
`filter`	Conditionally run an action	CEL with Playbook Context
`runsOn`	Which runner (agent) to run the action on	`[]Agent`
`templatesOn`	Where templating (and secret management) of actions should occur	`host` or `agent`
`timeout`	Timeout on this action.	Duration

AI

Field	Description	Scheme
`prompt*`	Main prompt	`string`
`systemPrompt*`	Context-setting system prompt	`string`
`analysis.since`	Select the analysis of the playbook resource to feed into the AI	Duration
`apiKey`	AI service API key	EnvVar
`apiURL`	Custom API endpoint (applicable when behind a proxy or using Ollama)	`string`
`awsAccessKeyId`	AWS access key ID. (applicable when the backend is `bedrock`)	EnvVar
`awsRegion`	AWS region. (applicable when the backend is `bedrock`)	`string`
`awsSecretAccessKey`	AWS secret access key. (applicable when the backend is `bedrock`)	EnvVar
`backend`	LLM provider	`openai` \| `anthropic` \| `ollama` \| `gemini` \| `bedrock`
`changes.since`	Select the changes of the playbook resource to feed into the AI	Duration
`config`	The config item to build the context from. Defaults to the config the playbook is running on	`string`
`connection`	Connection string for the LLM	`string`
`dryRun`	When enabled, the prompt is simply saved without passing it on to the LLM	`boolean`
`formats`	Output format	`markdown` \| `slack` \| `recommendPlaybook`
`model`	LLM model (e.g. gpt-4)	`string`
`outputSchema`	A JSON schema that the response must conform to	OutputSchema
`playbooks`	List of playbooks to execute and use as context	[]ContextProviderPlaybook
`recommendPlaybooks`	Specify selectors for playbooks. The LLM will recommend the best suited playbooks	[]ResourceSelector
`relationships`	Select the related configs and their changes and analysis to feed into the AI	[]Relationship
`skills`	Skill libraries that the model can load on demand	[]Skill

Output schema

Constrain the model's response to a JSON schema. The schema can be given inline, read from a configMap or secret, or fetched from a git repository.

Field	Description	Scheme
`checkout`	Fetch the schema from a git repository	Checkout
`value`	Inline JSON schema	`string`
`valueFrom`	Read the schema from a configMap or a secret	EnvVar

Output schema checkout

Field	Description	Scheme
`connection*`	Git connection to use e.g. `connection://github/my-org`	`string`
`path*`	Path of the JSON schema file within the repository	`string`
`branch`	Branch or tag to checkout. Defaults to the repository's default branch	`string`

Skills

Skills point at a directory of skill libraries, each with its own SKILL.md. The model loads them on demand rather than receiving them all upfront.

Field	Description	Scheme
`path*`	Path to the directory that contains the skill sub-directories. This is the parent of the skill directories, not a skill directory itself	`string`
`branch`	Branch or tag to checkout. Defaults to the repository's default branch	`string`
`connection`	Git connection to fetch the skills from e.g. `connection://github/my-org`. When empty, the path is read from the local filesystem	`string`

Context provider playbook

These playbooks are executed concurrently and their output is used as context for the AI action. If any of these playbooks fail, it does not affect the execution of the main playbook - the AI action will continue with whatever context is available from the successful playbooks.

Field	Description	Scheme
`name*`	Name of the playbook	`string`
`namespace*`	Namespace of the playbook	`string`
`if`	If is a CEL expression that decides if this playbook should be included in the context	CEL
`params`	Parameters to pass to the playbook	`[map[string]string]`

Relationship

The AI Action can maintain relationships with other elements:

Field	Description	Scheme
`analysis.since`	Select the analysis of the related resources to feed into the AI	Duration
`changes.since`	Select the changes of the related resources to feed into the AI	Duration
`depth`	Depth of the relationship	`integer`
`direction`	Direction of the relationship	`all` \| `incoming` \| `outgoing`

Artifacts

When Mission Control has a configured artifact store, the AI Action automatically stores its prompts as artifacts. These artifacts are:

Available in the Playbook Runs page
Downloadable for reference and analysis

Templating

CEL Expressions

The following variables can be used within the CEL expressions of filter, if, delays and parameters.default:

Field	Description	Schema
`config`	Config passed to the playbook	`ConfigItem`
`check`	Canary Check passed to the playbook	`Check`
`playbook`	Playbook passed to the playbook	`Playbook`
`run`	Current run	`Run`
`params`	User provided parameters to the playbook	`map[string]any`
`request`	Webhook request	`Webhook Request`
`env`	Environment variables defined on the playbook	`map[string]any`
`user.name`	Name of the user who invoked the action	`string`
`user.email`	Email of the user who invoked the action	`string`
`agent.id`	ID of the agent the resource belongs to.	`string`
`agent.name`	Name of the agent the resource belongs to.	`string`

Conditionally Running Actions

Playbook actions can be selectively executed based on CEL expressions. These expressions must either return

a boolean value (true indicating run the action & skip the action otherwise)
or a special function among the ones listed below

Function	Description
`always()`	run no matter what; even if the playbook is cancelled/fails
`failure()`	run if any of the previous actions failed
`skip()`	skip running this action
`success()`	run only if all previous actions succeeded (default)
`timeout()`	run only if any of the previous actions timed out

delete-kubernetes-pod.yaml
---
apiVersion: mission-control.flanksource.com/v1
kind: Playbook
metadata:
  name: notify-send-with-filter
spec:
  parameters:
    - name: message
      label: The message for notification
      default: '{{.config.name}}'
  configs:
    - types: 
        - Kubernetes::Pod
  actions:
    - name: Send notification
      exec:
        script: notify-send "{{.config.name}} was created"
    - name: Bad script
      exec:
        script: deltaforce
    - name: Send all success notification
      if: success() # this filter practically skips this action as the second action above always fails
      exec:
        script: notify-send "Everything went successfully"
    - name: Send notification regardless
      if: always()
      exec:
        script: notify-send "a Pod config was created"

Defaulting Parameters

delete-kubernetes-pod.yaml
apiVersion:
  mission-control.flanksource.com/v1
  kind: Playbook
  metadata:
    name: edit
  spec:
    title: 'Edit Kustomize Resource'
    icon: flux
    parameters:
    - default: 'chore: update $(.config.type)/$(.config.name)'
      name: commit_message

Go Templating

When templating actions with Go Templates, the context variables are available as fields of the template's context object . eg .config, .user.email

Templating Actions

delete-kubernetes-pod.yaml
apiVersion: mission-control.flanksource.com/v1
kind: Playbook
metadata:
  name: scale-deployment
spec:
  description: Scale Deployment
  configs:
    - types:
        - Kubernetes::Deployment
  parameters:
    - name: replicas
      label: The new desired number of replicas.
  actions:
    - name: kubectl scale
      exec:
        script: |
          kubectl scale --replicas={{.params.replicas}} \
            --namespace={{.config.tags.namespace}} \
            deployment {{.config.name}}

Functions

Function	Description	Return
`getLastAction()`	Returns the result of the action that just run	Action Specific
`getAction({action})`	Return the result of a specific action	Action Specific

Printing out Results

Reusing Action Results

action-results.yaml
apiVersion: mission-control.flanksource.com/v1
kind: Playbook
metadata:
  name: use-previous-action-result
spec:
  description: Creates a file with the content of the config
  configs:
    - types:
        - Kubernetes::Pod
  actions:
    - name: Fetch all changes
      sql:
        query: SELECT id FROM config_changes WHERE config_id = '{{.config.id}}'
        driver: postgres
        connection: connection://postgres/local
    - name: Send notification
      if: 'last_result().count > 0'
      notification:
        title: 'Changes summary for {{.config.name}}'
        connection: connection://slack/flanksource
        message: |
          {{$rows:=index last_result "count"}}
          Found {{$rows}} changes

AI​

Output schema​

Output schema checkout​

Skills​

Context provider playbook​

Relationship​

Artifacts​

Templating​

CEL Expressions​

Go Templating​

Functions​

AI