expansion-agent-selection

name	expansion-agent-selection
description	Select the best bundled developer agent for automated expansion leaves and document fallback decisions.
version	1.0.0
category	core
internal	true
triggers	expansion agent selection, assigned agent, automated leaves
metadata	{"gobby":{"audience":"all","depth":0}}

expansion-agent-selection - Automated Leaf Agent Selection

Use this skill during the expander's Agent Selection step after leaf generation and before the compiled spec is returned or applied.

Automated Leaf Categories

Assign an agent for every automated leaf in these categories:

• code
• config
• docs
• refactor
• test

Reject category: manual, category: planning, and category: research leaves. Those categories are valid for direct task creation, but automated expansion manifests must compile to deterministic development-forward leaf work.

Label Vocabulary

Use labels and task text as weak signals. They should guide the decision, not override clear file paths or task descriptions.

Frontend indicators

Prefer frontend-developer for work involving UI components, CSS, browser behavior, accessibility, routing, client state, React, Vue, Svelte, Vite, Webpack, Storybook, Playwright UI flows, Lighthouse, or frontend linting such as ESLint.

Backend indicators

Prefer backend-developer for APIs, CLIs, storage, migrations, schedulers, workers, servers, daemon code, Python modules, pytest, mypy, Ruff, SQL, psql, MCP tools, workflow engines, and infrastructure tests.

Fullstack indicators

Mixed-stack tasks that touch both UI and backend should be split when the plan permits. If splitting would violate the plan's atomic task shape, pick the primary agent for the riskiest surface and document the cross-concern surface in the leaf description.

Heuristics

1. Read the leaf title, description, category, labels, and affected files.
2. Compare those signals with available agent names and descriptions from list_agent_definitions.
3. Prefer a specialized matching agent over a general fallback.
4. Keep additional_skills empty unless the leaf needs a skill beyond the agent's baseline competence.
5. Do not escalate on ambiguity.

Confidence Threshold

A confident match requires a clear category or content signal and an available agent whose name or description matches the surface. If the signal is weak, mixed, or no available agent crosses the threshold, use the fallback.

Fallback

Default to backend-developer when no agent matches above the confidence threshold. Append a ## Agent Selection marker to the leaf description that explains the default, including the leaf category and the ambiguous signals.

Example marker:

## Agent Selection

No specialized agent matched this docs leaf above the confidence threshold.
Defaulting to `backend-developer` for auditability.

The fallback is intentional audit data, not an error path. Repeated fallback markers are the signal to add a more specialized agent later.