- Fix ChatBubble to handle non-string content with String() wrapper - Fix API route to use generateText for non-streaming requests - Add @ai-sdk/openai-compatible for non-OpenAI providers (DeepSeek, etc.) - Use Chat Completions API instead of Responses API for compatible providers - Update ChatBubble tests and fix component exports to kebab-case - Remove stale PascalCase ChatBubble.tsx file
9.0 KiB
Agent Compilation: YAML Source → Final Agent
For the LLM running this workflow: This document explains what the compiler adds. When building agents, focus on the YAML structure defined here—do NOT add things the compiler handles automatically.
Example reference: Compare
{workflow_path}/data/reference/module-examples/architect.agent.yaml(source, 32 lines) witharchitect.md(compiled, 69 lines) to see what the compiler adds.
Quick Overview
You write: YAML source file (agent-name.agent.yaml)
Compiler produces: Markdown with XML (agent-name.md) for LLM consumption
The compiler transforms your clean YAML into a fully functional agent by adding:
- Frontmatter (name, description)
- XML activation block with numbered steps
- Menu handlers (workflow, exec, action)
- Auto-injected menu items (MH, CH, PM, DA)
- Rules section
What YOU Provide (YAML Source)
Your YAML contains ONLY these sections:
agent:
metadata:
id: "_bmad/..."
name: "Persona Name"
title: "Agent Title"
icon: "🔧"
module: "stand-alone" or "bmm" or "cis" or "bmgd"
persona:
role: "First-person role description"
identity: "Background and specializations"
communication_style: "How the agent speaks"
principles:
- "Core belief or methodology"
critical_actions: # Optional - for Expert agents only
- "Load ./sidecar/memories.md"
- "Load ./sidecar/instructions.md"
- "ONLY access ./sidecar/"
prompts: # Optional - for Simple/Expert agents
- id: prompt-name
content: |
<instructions>Prompt content</instructions>
menu: # Your custom items only
- trigger: XX or fuzzy match on command-name
workflow: "path/to/workflow.yaml" # OR
exec: "path/to/file.md" # OR
action: "#prompt-id"
description: "[XX] Command description"
What COMPILER Adds (DO NOT Include)
1. Frontmatter
---
name: "architect"
description: "Architect"
---
DO NOT add frontmatter to your YAML.
2. XML Activation Block
<activation critical="MANDATORY">
<step n="1">Load persona from this current agent file</step>
<step n="2">Load config to get {user_name}, {communication_language}</step>
<step n="3">Remember: user's name is {user_name}</step>
<!-- YOUR critical_actions inserted here as steps 4, 5, etc. -->
<step n="N">ALWAYS communicate in {communication_language}</step>
<step n="N+1">Show greeting + numbered menu</step>
<step n="N+2">STOP and WAIT for user input</step>
<step n="N+3">Input resolution rules</step>
<menu-handlers>...</menu-handlers>
<rules>...</rules>
</activation>
DO NOT create activation sections—the compiler builds them.
3. Auto-Injected Menu Items
Every agent gets these 4 items automatically. DO NOT add them to your YAML:
| Code | Trigger | Description |
|---|---|---|
| MH | menu or help | Redisplay Menu Help |
| CH | chat | Chat with the Agent about anything |
| PM | party-mode | Start Party Mode |
| DA | exit, leave, goodbye, dismiss agent | Dismiss Agent |
4. Menu Handlers
<handler type="workflow">
When menu item has: workflow="path/to/workflow.yaml"
→ Load workflow.xml and execute with workflow-config parameter
</handler>
<handler type="exec">
When menu item has: exec="path/to/file.md"
→ Load and execute the file at that path
</handler>
DO NOT add handlers—the compiler detects and generates them.
Before/After Example: Architect Agent
Source: architect.agent.yaml (32 lines - YOU WRITE)
agent:
metadata:
id: "_bmad/bmm/agents/architect.md"
name: Winston
title: Architect
icon: 🏗️
module: bmm
persona:
role: System Architect + Technical Design Leader
identity: Senior architect with expertise in distributed systems...
communication_style: "Speaks in calm, pragmatic tones..."
principles: |
- User journeys drive technical decisions...
menu:
- trigger: WS or fuzzy match on workflow-status
workflow: "{project-root}/_bmad/bmm/workflows/workflow-status/workflow.yaml"
description: "[WS] Get workflow status..."
- trigger: CA or fuzzy match on create-architecture
exec: "{project-root}/_bmad/bmm/workflows/3-solutioning/create-architecture/workflow.md"
description: "[CA] Create an Architecture Document"
- trigger: IR or fuzzy match on implementation-readiness
exec: "{project-root}/_bmad/bmm/workflows/3-solutioning/check-implementation-readiness/workflow.md"
description: "[IR] Implementation Readiness Review"
Compiled: architect.md (69 lines - COMPILER PRODUCES)
---
name: "architect"
description: "Architect"
---
You must fully embody this agent's persona...
```xml
<agent id="architect.agent.yaml" name="Winston" title="Architect" icon="🏗️">
<activation critical="MANDATORY">
<step n="1">Load persona from this current agent file (already in context)</step>
<step n="2">🚨 IMMEDIATE ACTION REQUIRED - BEFORE ANY OUTPUT...</step>
<step n="3">Remember: user's name is {user_name}</step>
<step n="4">Show greeting using {user_name} from config...</step>
<step n="5">STOP and WAIT for user input...</step>
<step n="6">On user input: Number → execute menu item[n]...</step>
<step n="7">When executing a menu item: Check menu-handlers section...</step>
<menu-handlers>
<handlers>
<handler type="workflow">...</handler>
<handler type="exec">...</handler>
</handlers>
</menu-handlers>
<rules>
<r>ALWAYS communicate in {communication_language}</r>
<r>Stay in character until exit selected</r>
<r>Display Menu items as the item dictates...</r>
<r>Load files ONLY when executing menu items...</r>
</rules>
</activation>
<persona>
<role>System Architect + Technical Design Leader</role>
<identity>Senior architect with expertise...</identity>
<communication_style>Speaks in calm, pragmatic tones...</communication_style>
<principles>- User journeys drive technical decisions...</principles>
</persona>
<menu>
<item cmd="MH or fuzzy match on menu or help">[MH] Redisplay Menu Help</item>
<item cmd="CH or fuzzy match on chat">[CH] Chat with the Agent about anything</item>
<item cmd="WS...">[WS] Get workflow status...</item> ← YOUR CUSTOM ITEMS
<item cmd="CA...">[CA] Create an Architecture Document</item>
<item cmd="IR...">[IR] Implementation Readiness Review</item>
<item cmd="PM...">[PM] Start Party Mode</item>
<item cmd="DA...">[DA] Dismiss Agent</item>
</menu>
</agent>
Key additions by compiler: Frontmatter, activation block, handlers, rules, MH/CH/PM/DA menu items.
DO NOT DO Checklist
When building agent YAML, DO NOT:
- Add frontmatter (
---name/description---) to YAML - Create activation blocks or XML sections
- Add MH (menu/help) menu item
- Add CH (chat) menu item
- Add PM (party-mode) menu item
- Add DA (dismiss/exit) menu item
- Add menu handlers (workflow/exec logic)
- Add rules section
- Duplicate any auto-injected content
DO:
- Define metadata (id, name, title, icon, module)
- Define persona (role, identity, communication_style, principles)
- Define critical_actions (Expert agents only)
- Define prompts with IDs (Simple/Expert agents only)
- Define menu with your custom items only
- Use proper trigger format:
XX or fuzzy match on command-name - Use proper description format:
[XX] Description text
Expert Agent: critical_actions
For Expert agents with sidecars, your critical_actions become activation steps:
critical_actions:
- "Load COMPLETE file ./agent-sidecar/memories.md"
- "Load COMPLETE file ./agent-sidecar/instructions.md"
- "ONLY read/write files in ./agent-sidecar/"
The compiler injects these as steps 4, 5, 6 in the activation block:
<step n="4">Load COMPLETE file ./agent-sidecar/memories.md</step>
<step n="5">Load COMPLETE file ./agent-sidecar/instructions.md</step>
<step n="6">ONLY read/write files in ./agent-sidecar/</step>
<step n="7">ALWAYS communicate in {communication_language}</step>
Division of Responsibilities
| Aspect | YOU Provide (YAML) | COMPILER Adds |
|---|---|---|
| Agent identity | metadata + persona | Wrapped in XML |
| Memory/actions | critical_actions | Inserted as activation steps |
| Prompts | prompts with IDs | Referenced by menu actions |
| Menu items | Your custom commands only | + MH, CH, PM, DA (auto) |
| Activation | — | Full XML block with handlers |
| Rules | — | Standardized rules section |
| Frontmatter | — | name/description header |
Quick Reference for LLM
- Focus on: Clean YAML structure, persona definition, custom menu items
- Ignore: What happens after compilation—that's the compiler's job
- Remember: Every agent gets MH, CH, PM, DA automatically—don't add them
- Expert agents: Use
critical_actionsfor sidecar file loading - Module agents: Use
workflow:orexec:references, not inline actions