model-routing
Use when dispatching any subagent, teammate or reviewer, choosing a model or effort level for a piece of delegated work, escalating after an agent fails, or when a run's token or quota budget matters.
NoobyGains · 已有静态检查记录
安装与使用
运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。
npx skills add NoobyGains/godmode --skill model-routing技能说明
name: model-routing description: "Use when dispatching any subagent, teammate or reviewer, choosing a model or effort level for a piece of delegated work, escalating after an agent fails, or when a run's token or quota budget matters."
Model Routing
Every seat in a run (implementer, reviewer, investigator, lead) needs a model and an effort level. Left unnamed, a dispatch inherits the session's model, which is usually the most capable and most expensive one, and the cost discipline of the whole run silently disappears.
Core principle: give each seat the cheapest model that can do the job in few turns, name it and the effort on every dispatch, and escalate on evidence rather than by default. Cheap is measured per finished task, not per token: a cheap model that needs three times the turns, or a second attempt, costs more than a mid-tier model that finishes once.
The Prime Directive
NO DISPATCH WITHOUT A NAMED MODEL AND EFFORT
No exceptions. No workarounds. No shortcuts.
When to Use
Required:
- Every subagent, teammate, reviewer or cross-model dispatch.
- Choosing the escalation step after an agent reports BLOCKED or a fix loop stalls.
- Planning a run whose cost or quota the user cares about.
Skip:
- Work you do yourself in the current session: the lead stays on the session model.
The Tiers
This skill speaks in three tiers. Each harness reference maps them to that harness's current model names and effort settings; never hard-code a model name in a plan or brief from memory.
| Tier | Good at | Weak at |
|---|---|---|
| Cheap | Transcribing fully specified code, single-file mechanical edits, searches, summaries, classification, extracting text | Multi-step work from prose, judgment calls; tends to take many more turns |
| Mid | Implementing from a prose task, multi-file integration, per-task review, debugging with a clear reproduction | Architecture, subtle concurrency, very large diffs |
| Top | Architecture, design trade-offs, the final whole-branch review, escalations after mid-tier failure, adversarial review | Nothing on quality; spending it on bulk work wastes budget |
Effort. Where the harness has an effort or reasoning setting, set it per seat too: low for transcription and searches, medium for routine implementation and task review, high for architecture, final review and escalations. The highest settings are for problems that have already failed at high.
Seat Defaults
| Seat | Default | Move up when | Move down when |
|---|---|---|---|
| Implementer, plan pins the signature and the test | Cheap, low | the task touches several files or tests are non-trivial | (already lowest) |
| Implementer, task described in prose | Mid, medium | it needs design judgment or broad codebase understanding | it is one file with an exact spec |
| Batched same-shape edits | Cheap, low | the edits are not truly identical in shape | |
| Task reviewer | Mid, medium (the floor for reviewers) | the diff is subtle: concurrency, security, data migration | never below mid |
| Scoped re-review of a small fix | Mid, low effort (the reviewer floor holds; only the effort drops) | the fix is larger than the finding implied | never below mid |
| Final whole-branch review | Top, high | never; this is the one fresh context the run buys | |
| Fix rounds 4-5 implementer | One tier above the stuck implementer | ||
| Bulk reading, searching, summarising for the lead | Cheap, low | the summary needs judgment | |
| Investigator in a diagnosis | Mid, medium | the hypothesis space is large or adversarial | |
| Cross-model reviewer | The other family's strongest affordable model | ||
| Lead / controller | The session model |
The Protocol
-
Classify the seat using the table above. Read the task's complexity signals: number of files, whether the exact code is given, whether design judgment is needed, the risk of the change.
-
Pick tier and effort. Start from the seat default and apply the move-up and move-down columns. When torn between two tiers for implementation, choose the higher one: a failed cheap attempt plus a retry costs more than one mid-tier success.
-
Name both on the dispatch. Use the harness reference to translate the tier into the model name and effort setting that harness accepts. If the harness cannot set a model per dispatch, say so in the ledger once and note that every seat runs on the session model.
-
Keep the lead on the session model, and keep it light. The lead coordinates, rules and reads short replies. It does not read whole directories or long logs: send bulk reading to a cheap-tier agent that returns a summary or writes a file.
-
Escalate on evidence, one step at a time. The ladder:
- Same model, more context (the agent said NEEDS_CONTEXT, or the brief was missing a fact).
- Same model, smaller task (the task was too large to hold).
- One tier up, fresh agent, carrying the report file of what was tried.
- Top tier at high effort.
- A different model family (
godmode:cross-model-review), or a ruling by the lead, or the user when it is one of the four stop conditions. Never retry the same model with the same brief unchanged.
-
Choose the cheapest execution mode first. Direct work, then
godmode:inline-execution, thengodmode:delegated-executionorgodmode:parallel-execution, andgodmode:team-orchestrationlast. Every teammate is a separate full instance and token use grows with