Spec Agent Command

"An agent without a clear role is a meeting without an agenda -- everyone is busy, nothing gets done."

Generates production-ready agent markdown files and command markdown files (canonical + alias pairs) from design documents. Enforces plugin subagent constraints: no hooks, mcpServers, or permissionMode in agent frontmatter. Validates kebab-case naming and cross-reference integrity.

---

Procedure

Step 1 -- Read Agent and Command Designs

• Read the design file at the path provided via argument-hint. [Read]
• Extract for each agent: name, description, role, tools, skills assigned, execution flows, operating principles.
• Extract for each command: name, description, skill it invokes, argument-hint, aliases.
• If the design references skills not yet defined, flag them as WARNING: Unresolved skill reference.

Step 2 -- Generate Agent .md File

Compose the agent file with these sections:

1. Frontmatter (YAML):

   • name: kebab-case agent identifier.
   • description: 1-2 sentence role description.
   • tools: list of tools the agent uses.
   • model: typically inherit.
   • MUST NOT contain: hooks, mcpServers, permissionMode (plugin subagent constraint).

2. Title -- H1 with the agent name in title case.

3. Role -- 2-4 sentences defining what the agent does, which movement it owns, and its relationship to other agents.

4. Responsibilities -- Numbered list of high-level duties.

5. Skills Assigned -- Table with columns: #, Skill, Trigger.

6. Execution Flows -- One subsection per command/workflow showing the skill execution sequence as an indented flow diagram.

7. Operating Principles -- Numbered list of behavioral rules (5-8 items). Include safety, idempotency, and cross-reference rules.

Step 3 -- Enforce Subagent Constraints

• Scan the generated agent frontmatter for forbidden fields: hooks, mcpServers, permissionMode.
• If any are present, remove them and add an INFO note explaining the constraint.
• Verify the tools list only contains valid tool names (Read, Write, Edit, Glob, Grep, Bash, etc.).

Step 4 -- Generate Command .md Files (Canonical)

For each canonical command:

1. Frontmatter (YAML):

   • description: 1 sentence describing what the command does.
   • user-invocable: true.
   • argument-hint: from design (use <> for required, [] for optional).

2. Body:

   • Usage section with the command syntax.
   • Execution Flow describing which skill(s) are invoked and in what order.
   • Output describing what the user receives.

Step 5 -- Generate Alias Command .md Files

For each alias:

1. Frontmatter (YAML):

   • alias-of: the canonical command name (without plugin prefix).
   • description: same as canonical.
   • user-invocable: true.

2. Body: Single line stating this is an alias. Example: Alias for [canonical-name]. See commands/{canonical-name}.md.

Step 6 -- Validate Naming

• Check all generated file names against kebab-case pattern: ^[a-z][a-z0-9]*(-[a-z0-9]+)*$.
• Check agent names, command names, and skill references for kebab-case compliance.
• Report WARNING for each violation with suggested correction.

Step 7 -- Validate Cross-References

• Collect all skill names referenced in agent files and command execution flows.
• Verify each referenced skill exists in the design or in the plugin's existing skills/ directory. [Glob]
• Collect all command names referenced in agent execution flows.
• Verify each referenced command exists in the design or in commands/. [Glob]
• Report WARNING for each unresolved reference.

---

Quality Criteria

• Agent frontmatter contains no forbidden fields (hooks, mcpServers, permissionMode).
• Every command has a matching alias file with correct alias-of pointer.
• All file names are kebab-case compliant.
• All skill and command cross-references resolve to existing or designed entities.
• Agent file includes at least: role, skills table, one execution flow, and operating principles.

Assumptions & Limits

• Requires a completed agent/command design document as input. Cannot infer agent roles or command routing from scratch -- that is design-agent's job.
• The subagent constraint check (Step 3) removes forbidden fields silently with an INFO note. It does not ask for confirmation because these fields are never valid for plugin agents.
• Alias generation assumes a 1:1 mapping (one alias per canonical command). Multiple aliases for the same canonical must be specified explicitly in the design.
• Cross-reference validation (Step 7) checks existence only, not semantic correctness. A command referencing validate-structure will pass even if the command has nothing to do with validation.
• Cannot validate that execution flows will actually work at runtime -- only that the referenced skills and commands exist.

Good vs Bad

Bad agent spec:

---
name: my-agent
description: Does stuff
hooks:
  PreToolUse: [...]
---
# My Agent
Handles validation.

Missing: forbidden hooks field, no tools list, no skills table, no execution flows, no operating principles.

Good agent spec:

---
name: plugin-qa-engineer
description: >
  Orchestrates plugin quality assurance across validation, audit, and reporting.
tools:
  - Read
  - Glob
  - Grep
  - Bash
maxTurns: 45
---
# Plugin QA Engineer
**Role**: Validates plugin quality by orchestrating 9 QA skills across 4 movements.
## Skills Assigned
| # | Skill | Movement |
|---|-------|----------|
| 1 | validate-structure | VALIDATE |
## Execution Flows
### /pqa:validate
1. Run validate-structure → 2. Run validate-manifest → ...
## Operating Principles
1. Read first, write never.

Includes: no forbidden fields, tool list, maxTurns, skills table, execution flow, operating principles.

Anti-Patterns

1. Including hooks in agent frontmatter -- Plugin subagents cannot define their own hooks. Hooks belong in hooks/hooks.json at the plugin level. Remove the field entirely.
2. Alias without canonical -- Creating an alias command that points to a non-existent canonical command. Always generate the canonical first, then the alias.
3. Execution flow with unnamed skills -- Writing flows like "run validation" instead of referencing the exact skill name (validate-structure). Every step in a flow must reference a concrete skill.

Edge Cases

1. Agent with a single skill -- Valid. Still generate the full agent structure including a minimal execution flow and skills table with one row.
2. Command with no argument-hint -- Valid for commands that take no arguments (e.g., /pqa:menu). Omit argument-hint from frontmatter rather than setting it to empty string.
3. Multiple agents sharing a skill -- Valid. The skill appears in both agents' skills tables. Note this in both agents' operating principles for clarity.