unified-router
Intelligent task routing across Claude Code + Qwen Code — picks the right engine, model, and thinking depth for every task
SKILL.mdskills/unified-router/SKILL.md
GitHubUnified Router — Claude↔Qwen Intelligent Task Routing
Route every task to the optimal engine (Claude Code or Qwen Code), model, and thinking depth. This skill replaces ad-hoc routing decisions with a systematic framework.
Routing Decision Matrix
Engine Selection
| Signal | Route to Claude Code | Route to Qwen Code |
|---|---|---|
| Authority level | L0-L2 (strategic) | L5-L6 (worker) |
| Risk tier | T2-T3 (shared/critical) | T0-T1 (safe/local) |
| MCP servers needed | aster, playwright, stripe, k8s, terraform, slack, postgres, docker | filesystem, github, memory, context7, brave-search, tavily, sequential |
| Multi-agent orchestration | Yes | No |
| File coordination | Multiple tasks touch same files | Single-file focus |
| Thinking depth | Architecture, security, design | Review, test, docs, explore |
| Cost sensitivity | Budget available | Token-saving priority |
Model Selection (within each engine)
Claude Code models:
| Complexity | Model | When |
|---|---|---|
| Enterprise | opus | Architecture, security audit, multi-service design |
| Complex | sonnet | Multi-file implementation, debugging, integration |
| Simple | haiku | Quick lookups, simple edits, status checks |
Qwen Code models:
| Complexity | Model | When |
|---|---|---|
| Deep | qwen3.5:cloud | Complex code review, large refactors |
| Standard | qwen3:8b | Test gen, docs, simple fixes |
| Quick | qwen3:8b | Exploration, search, scaffolding |
| Creative | minimax-m2.5:cloud | Creative writing, brainstorming |
| Coding | kimi-k2.5:cloud | Strong coding tasks |
| Cloud (free) | coder-model (--qwen-oauth) | Best quality, 1K free/day |
Thinking Depth Selection
| Task Type | Thinking Budget | Model | Engine |
|---|---|---|---|
| Architecture, design, security | 8000 tokens | qwen3.5:cloud or Claude opus | Either (prefer Claude) |
| Implementation, bug fix, integration | 4000 tokens | qwen3:8b or Claude sonnet | Either |
| Review, test, docs, explore, scaffold | 1000 tokens | qwen3:8b or Claude haiku | Prefer Qwen (saves tokens) |
Dispatch Commands
Claude → Qwen (offload worker tasks)
# Standard dispatch (existing)
qwen-dispatch review "Review src/auth.py" --cwd ~/project
# With thinking (NEW)
qwen-dispatch review "Review src/auth.py" --think --cwd ~/project
# Thinking-only mode (NEW)
qwen-dispatch think "Analyze the auth flow design" --cwd ~/project
# Full auto-activation (NEW)
qwen-dispatch auto "Implement user pagination" --cwd ~/project
Qwen → Claude (escalation)
# When Qwen detects T2+ risk
qwen-escalate "Design the new microservice architecture" --cwd ~/project
# With prior thinking context
qwen-escalate "Deploy to production" --context ~/.qwen/thinking/think_abc123.json
Routing Algorithm
1. Parse task → extract domain, complexity, risk signals
2. Check MCP requirements → if needs servers Qwen lacks → Claude
3. Check risk tier → T2+ → Claude
4. Check authority level → L0-L2 → Claude, L5-L6 → Qwen
5. Check file coordination → multi-task same-file → Claude (needs orchestration)
6. Default → Qwen with --think flag (saves tokens, adds reasoning)
Integration Points
- CoreMind STAGE 4 (PLAN): Router evaluates each plan step
- CoreMind STAGE 7 (EXECUTE): Router dispatches to the selected engine
- qwen-dispatch: Updated with
think,plan,autotask types +--thinkflag - qwen-escalate: New bidirectional bridge for Qwen→Claude handoff
- auto-activate.py: Qwen's thinking engine (invoked by all new task types)
Artifact Flow
Claude thinking → ~/.claude/projects/*/memory/ (Claude's memory)
Qwen thinking → ~/.qwen/thinking/*.json (structured artifacts)
Qwen plans → ~/.qwen/plans/*.json (execution DAGs)
Qwen logs → ~/.qwen/logs/*.json (execution results)
Qwen metrics → ~/.qwen/metrics/*.jsonl (performance tracking)
Both engines can read each other's artifacts for context continuity.