What would you like to be added?

Support declarative agent definitions — define custom agents via Markdown files with YAML frontmatter, instead of hardcoding them in TypeScript. This is the pattern Claude Code shipped in 2.1.167.

How it works in Claude Code

Users place a .claude/agents/<name>.md file with frontmatter:

---
name: code-reviewer
description: "Review code for bugs and style issues"
model: sonnet
effort: high
permissionMode: plan
tools:
  - Read
  - Bash
  - Grep
disallowedTools:
  - Write
  - Edit
maxTurns: 10
skills:
  - code-review
memory: project
isolation: worktree
hooks:
  PreToolUse:
    - command: "echo reviewing..."
---

You are a code reviewer. Focus on correctness bugs...

The Markdown body becomes the agent's system prompt. The agent is then available via the Agent tool and --agent CLI flag.

Full schema (confirmed via binary reverse engineering of CC 2.1.167)

Why this matters

1. User-defined agents — users can create project-specific agents (e.g. a DB migration agent, a security reviewer) without writing code
2. Composable — agents can reference skills, hooks, and MCP servers declaratively
3. Discoverable — .qwen/agents/ directory makes agents visible and version-controllable
4. Parity — Claude Code already ships this; community expects it

Suggested implementation

• Add .qwen/agents/ directory scanning (similar to .qwen/skills/)
• Parse frontmatter with the same parser used for skills
• Register discovered agents in the existing builtin-agents.ts / subagent-manager.ts pipeline
• Validate schema with zod (like CC does)
• Support both project-level (.qwen/agents/) and user-level (~/.qwen/agents/) definitions

Prior art in this repo

• Skill frontmatter parsing already exists (skill-load.ts, skill-manager.ts)
• builtin-agents.ts has the registration path
• subagent-manager.ts handles agent lifecycle

The skill loading pipeline can be largely reused — the main new work is the schema definition and wiring agents into the Agent tool's type resolution.