Agent Manifest Reference¶

Agent manifests describe every AI capability the platform can expose. They live alongside other configuration-as-code assets inside .kiket/agents/ within your repository.
Looking for the companion configuration-as-code format? See the Agent Definition DSL for authoring repositories of reusable agents.

File Locations¶

Bundled defaults: config/kiket/agents/*.yml
Template repositories: definitions/<template>/.kiket/agents/*.yml
Customer repositories: storage/workflow_sources/<org>/<repo>/.kiket/agents/*.yml

Files are loaded in that order. Duplicate id values are rejected so each agent is defined exactly once.

Schema Overview¶

Every manifest uses schema version 1.0 and must include the following keys:

Key	Type	Notes
`model_version`	String	Use `1.0` for the current schema.
`id`	String	Lowercase slug (`a-z`, `0-9`, `_`, `-`, `.`). Must be unique.
`version`	String	Semantic version for rollout tracking.
`name`	String	Human-friendly title for dashboards and admin tooling.
`description`	String (optional)	Short explanation of what the agent does.
`prompt`	String	ID of the prompt template to render.
`capabilities`	Array\<String>	Tags such as `summarize`, `classify`, `assign`.
`context.required`	Array\<String>	Context providers that must be present (e.g. `issues`, `organization_settings`).
`context.optional`	Array\<String>	Additional providers used when available.
`human_in_loop`	Hash (optional)	Approval policy (`required`, `escalation_strategy`, `notes`, `reason`).
`confidence_threshold`	Float (optional)	Value between `0.0`–`1.0` used to trigger human review.
`metadata`	Hash (optional)	Free-form data for UI hints or rollout flags.

All strings are stripped of whitespace and blank entries are removed from arrays at load time.

Validating Manifests¶

Use the kiket agents lint command before committing. It validates manifests against the schema, checks for required fields, validates ID formats, and reports errors for unknown keys. Run from your project root:

kiket agents lint                    # Lint agents in current directory
kiket agents lint path/to/project    # Lint agents in specific path
kiket agents lint --fail-fast        # Stop on first error

Example Manifest¶

model_version: "1.0"
id: triage.classifier
version: 1.2.0
name: Incident Triage Classifier
description: Suggests severity and the right responder for new incidents.
prompt: ai.prompts.triage
capabilities:
  - classify
  - summarize
context:
  required:
    - issues
    - organization_settings
  optional:
    - analytics_snapshot
human_in_loop:
  required: true
  escalation_strategy: escalate_to_oncall
  notes: "Escalate if confidence < 0.6 or priority is critical."
confidence_threshold: 0.65
metadata:
  audience: support

Authoring Tips¶

Keep manifests declarative—business logic should live in extensions or KiketScript actions.
Increment version whenever behaviour changes to help identify rollout state.
Pair every manifest with an integration or system test whenever you roll out a new capability.