ReddRadar
Agent skill
model-routing
Select the right model for the task. Maps task cognitive tier to optimal model. Reads _b00t_ datums for available models. Prefer local/cheap for deterministic work; frontier for reasoning.
Install this agent skill to your Project
npx add-skill https://github.com/elasticdotventures/_b00t_/tree/main/plugins/next-task/skills/model-routing
SKILL.md
Model Routing
NEVER use frontier model for mechanical work. Route by cognitive tier.
Reads routing table from _b00t_/model-routing.tomllm. Falls back to hardcoded tiers.
Cognitive Tiers
| Tier | Tasks | b00t Models (in priority order) |
|---|---|---|
small (sm0l) |
grep, lint, classify, route, test pass/fail | haiku, local sm0l |
chunky (ch0nky) |
implement, refactor, debug, code review | qwen3-coder-local (RTX 3090), sonnet |
frontier |
architecture, security, novel design, planning | opus, sonnet |
Steps
- Classify the task type. # output: task_type
- Map task_type → cognitive tier. # output: tier (sm0l|ch0nky|frontier)
- Load routing config:
b00t learn model-routingvia MCP or CLI — NEVER read .tomllm directly. # output: available_models[] - Select best available model for tier (prefer local, fallback frontier). # output: selected_model
- Check resource gate:
b00t hive status— ensure RAM/GPU available. # output: resource_ok - If resource gate fails: escalate one tier up or queue. # output: model_or_queue
- Return
{model, tier, rationale}for caller to invoke.
Task → Tier Mapping
small sm0l (Haiku / local 3B):
- Running tests, checking lint output
- Classifying/routing messages
- Extracting structured data from well-defined input
- File diffing, counting, summarizing short text
- Executing known shell commands
chunky ch0nky (qwen3-coder-local → Sonnet fallback):
- Writing or refactoring code
- Debugging with stack traces
- Multi-file code review
- Translating between languages/formats
- Implementing skills from SKILL.md spec
frontier (Opus → Sonnet fallback):
- System architecture decisions
- Security threat modeling
- Novel algorithm design
- Planning complex multi-step workflows
- Evaluating ambiguous requirements
Output Contract to Executive Context
Executive context is costly. Sub-agents MUST return compressed summaries:
| Tier | Max output to executive |
|---|---|
sm0l |
PASS or FAIL: <name> <5 lines> |
ch0nky |
diff + test result (no full file dumps) |
frontier |
structured decision with rationale |
Resource Awareness
Before invoking ch0nky/frontier check hive:
b00t hive status # output: RAM free, GPU VRAM free, active profile
Anti-pattern: running vLLM (qwen3-coder, 20GB VRAM) + HuggingFace download simultaneously on 24GB.
Integration
Used by /next-task at each phase to select model.
Used by b00t-mcp agent delegation.
Load via: b00t learn model-routing (MCP preferred, CLI fallback)
🤓 NEVER read b00t/*.tomllm directly — always use b00t learn/MCP which applies guardrails, guru enrichment & tribal knowledge
Didn't find tool you were looking for?