delegating

name	delegating
description	Proactively delegates tasks to cost-optimized agents before working in main session. MUST suggest for: tests, typos, formatting, docs, refactors, code reviews, feature implementation, debugging, commit validation. Skips main session token usage by routing to GLM (simple/deterministic), Gemini (reasoning/analysis), Qwen (quality/patterns), or multi-agent orchestration (review, feature dev, bug hunt). Never suggest for: architecture decisions, security-critical code, unknown-cause bugs, performance optimization.
allowed-tools	Bash
disable-model-invocation	true

Delegating Tasks

Delegate tasks to cost-optimized models (CCS) or multi-agent orchestration workflows (Gemini/Qwen).

When to Suggest

Task Pattern → Backend Mapping (auto-selection logic):

Task Pattern	Backend	Cost	Reason
typo|test|doc|format	CCS (GLM)	LOW	Simple deterministic
think|analyze|reason	CCS (Gemini)	MEDIUM	Requires reasoning
review.*(code|security)	Orchestration	HIGH	Multi-agent code review
implement.*feature	Orchestration	HIGH	Full development workflow
validate.*commit	Orchestration	MEDIUM	Security+Quality validation
debug|bug.*unknown	Orchestration	HIGH	Root cause investigation

Never Suggest For:

• Architecture decisions requiring human judgment
• Security-critical without review
• Performance optimization (needs profiling first)

---

Interactive Menu

Step 1: Delegation Choice

Use AskUserQuestion with:

• question: "This task can be delegated. How would you like to proceed?"
• header: "Execution"
• options:
  • "Delegate (Recommended)" — Execute via optimal backend. Saves main session tokens and uses cost-efficient models.
  • "Work in main session" — Execute in current Claude session. Better for tasks requiring discussion or complex context.

If user selects "Delegate" → Continue to Step 2 If user selects "Work in main session" → Execute task normally (don't delegate)

Step 2: Backend Selection

Use AskUserQuestion with:

• question: "Which backend should handle this task?"
• header: "Backend"
• options:
  • "Auto-select (Recommended)" — Analyzes task keywords and selects optimal backend/profile automatically.
  • "GLM - Cost-optimized" — Fast model for tests, typos, formatting [LOW COST]
  • "Gemini - Reasoning" — Analysis, thinking, architecture tasks [MEDIUM COST]
  • "Qwen - Quality" — Code quality, pattern detection [MEDIUM COST]
  • "Multi-Agent Orchestration" — Direct Gemini/Qwen collaboration for complex tasks (review, feature dev, debugging) [HIGH COST]

---

Auto-Selection Logic

Configuration-Driven: All pattern matching is defined in config.yaml, not hardcoded.

Configuration Structure

The skill reads config.yaml to determine:

1. Available backends (CCS profiles + Orchestration workflows)
2. Pattern mappings (task keywords → backend selection)
3. Priority order (Orchestration workflows checked before CCS)
4. Default fallback (when no pattern matches)

---

Orchestration Workflow Selection (Autonomous)

When backend: 'orchestration' is selected, Claude autonomously chooses the appropriate workflow and orchestrates between gemini and qwen CLI tools.

Selection Process

1. Load config - Read workflow definitions from config.yaml
2. Match patterns - Determine which orchestration pattern (collaborative, handshake, troubleshoot) applies.
3. Execute turn protocol - Use CLI commands sequentially:
   • gemini -p "..."
   • qwen "..."
   • gemini -r latest -p "..." (to refine)

Turn Protocols

Workflow	Protocol
handshake	1 turn: Agent A (Gemini) proposes -> Agent B (Qwen) validates.
collaborative	3 turns: Gemini designs -> Qwen critiques -> Gemini refines.
troubleshoot	4 turns: Gemini hypothesis -> Qwen verification -> Gemini root cause -> Final synthesis.

---

Execution Flow

For Direct Invocation (/delegation [task] or /delegate [task])

1. Parse override flag (if present: --glm, --gemini, --orchestrate, etc.)
2. Auto-select backend using keyword-based logic.
3. Route to appropriate backend (see sections below).
4. Report results: Backend, Workflow (if Orchestration), Cost indicator, Duration.

---

CCS / GLM Execution

CCS wraps the claude CLI and only works with GLM. It requires a PTY for output and is blocked by the CLAUDECODE nested-session guard — always run via tmux:

tmux new-session -d -s ccs_task "env -u CLAUDECODE ccs glm -p '{task}' > /tmp/ccs_out.txt 2>&1"
sleep 30  # or poll until session exits
cat /tmp/ccs_out.txt

If the task requires file modifications, add --dangerously-skip-permissions:

tmux new-session -d -s ccs_task "env -u CLAUDECODE ccs glm --dangerously-skip-permissions -p '{task}' > /tmp/ccs_out.txt 2>&1"

---

Gemini CLI Execution

Gemini is a direct CLI — no tmux workaround needed:

gemini -p "task description"

If the task requires file modifications, add -y to auto-approve all tool calls:

gemini -y -p "task description"

---

Qwen CLI Execution

Qwen is a direct CLI — no tmux workaround needed:

qwen "task description"

If the task requires file modifications, add -y to auto-approve all tool calls:

qwen -y "task description"

---

Orchestration Execution

Multi-turn sequences use Gemini and Qwen direct CLIs. Add -y to all turns when modifications are expected:

gemini -y -p "..."            # Turn 1
qwen -y "..."                 # Turn 2
gemini -y -r latest -p "..."  # Turn 3 (refine)

---

Examples

Auto-Selection Examples

CCS Simple:

• /delegate add unit tests for UserService → CCS (GLM)
• /delegate think about the best database schema → CCS (Gemini)

Orchestration Workflows:

• /delegate review this code for security issues → Orchestration (parallel-review)
• /delegate implement OAuth authentication feature → Orchestration (feature-design)
• /delegate debug crash on startup → Orchestration (bug-hunt)

---

Notes

Version 7.0.0 - Direct Orchestration

• Independent of unitAI: Now uses direct gemini and qwen CLI calls.
• Unified backends: CCS (cost-optimized) + Direct Orchestration.
• Config-driven: All behavior defined in config.yaml.
• Reference: Uses patterns from orchestrating-agents.