agent-discovery

To automate agent discovery, scan project directories for .claude/agents/*.md files to build a roster, parse YAML frontmatter for domains and roles, and score candidates by domain overlap to select the right specialist agent.

YAML frontmatter parsing extracts agent metadata like name, domains, role, invoked_during, trigger_contexts, and skills from Markdown files, then filters by role and scores candidates by domain overlap and trigger-context adjustments for ranking.

Yes, you can apply workflow.yaml overrides to enforce priority rules and validation checks during agent selection, ensuring discovered agents meet specific workflow criteria before task dispatch.

The best way to dynamically find domain experts is running file-based agent discovery during the prepare phase, which scans available agents and returns a ranked roster based on domain overlap without requiring manual configuration.

You need file-based agent discovery during setup to avoid duplicate agent generation and during the prepare phase to select eligible agents for design, implementation, and review sessions.

When agent discovery finds no eligible reviewers matching the required domains and roles, it applies fallback behavior during the prepare phase to handle the absence of specialist agents gracefully.