callee-create-agent

name	callee-create-agent
description	Create project-defined Callee Role, Sequential, and Loop agents in Markdown or YAML. Use when the user asks to generate, scaffold, compose, or author a Callee agent or deterministic workflow.

Create a Callee agent

Use callee when available. Otherwise use the pinned fallback npx --yes @baldaworks/callee@0.16.0 for every command in the task.

Choose the kind and ID

Inspect the existing catalog before writing:

callee agent list --json

Choose exactly one supported kind:

• Role: one provider-backed leaf agent.
• Sequential: ordered child agents that each run once.
• Loop: ordered child agents repeated until an authorized Role escalates or the iteration limit is exhausted.

Do not create Parallel. Write below .callee/. Markdown is the default; use a complete .yaml or .yml object only when the user explicitly requests YAML. Directories such as roles/ and workflows/ are optional ID namespaces, not kind selectors. The agent ID is the relative path without its final supported extension.

Author a Role

When an embedded PromptKit template clearly fits the requested capability, inspect it and use the generator:

callee promptkit search "<intent>" --type template
callee promptkit show "<template>" --json
callee promptkit role create "<agent-id>" \
  --template "<template>" \
  --description "<capability description>" \
  --provider "<provider>" \
  --prompt-param "<runtime-input-parameter>"

Inspect metadata.mode in the selected template. When it is interactive, let the generator set spec.repl: true; keep its questions and confirmation gates for callee agent run and do not execute those phases while authoring the reusable Role. Use --repl only to force REPL behavior for a template that is not marked interactive.

Bind only values that are fixed when the agent is authored. Leave intended runtime values unbound so they become spec.params. Use --output when the requested ID must not live below the generator's default .callee/roles/ namespace.

If PromptKit does not fit, author the Role directly:

---
apiVersion: callee.metalagman.dev/v1alpha1
kind: Role
spec:
  description: <capability description>
  provider:
    type: <codex|claude|opencode|copilot|grok|cursor|generic_acp>
  permissions:
    mode: ask
  params:
    focus: What the agent should focus on
---
Task:
{{ .Input }}

Focus:
{{ .Params.focus }}

Keep provider configuration under spec.provider. Configure ACP permission handling separately with Role-only spec.permissions.mode: ask uses the controlling TTY, allow automatically selects a compatible allow option, and deny automatically selects a compatible reject option. Omission defaults to ask; do not choose allow unless the user explicitly requests unattended approval. Set spec.repl: true on a directly authored Role only when it must continue in the same provider session across operator turns. Keep exactly one unconditional bare {{ .Input }} insertion in a generated Role body. Use Go text/template syntax on every template surface.

Author a workflow

For every Sequential, Loop, or nested-composite request, read references/workflows.md completely before writing. Follow its placement, representation, child-wiring, state, parameter, loop-control, and output rules.

Resolve every referenced child with callee agent view "<child-id>" --json. If the request also needs new child Roles or workflows, create and validate those first. Then author the root workflow from the reference and validate the complete resolved tree.

Validate the result

Validate each written file, then resolve the complete tree and required runtime parameters:

callee agent validate "<written-agent-path>"
callee agent view "<agent-id>" --json

Use the actual generated .md, .yaml, or .yml path for validation. For a PromptKit template with metadata.mode: interactive, confirm that the resolved view reports repl: true. Confirm that every Role's authored and effective permission policy in agent view --json matches the request. Fix every schema, template, missing-child, duplicate-ID, and duplicate-alias error before reporting success. Do not add Gemini, legacy flat provider fields, a server transport, or thread persistence.