Skip to content

n23eos/jev-skills / src/jev_skills/skills/jev-model-router

jev-model-router

Verified

Suggest one host-provided coding model for a bounded task when several available models differ in cost and capability.

Model routing

Use this skill when the user wants help choosing a model for a specific coding task. Jev does not replace the model powering this conversation or switch an existing agent session.

  1. List only models actually available in the destination host. Map up to four useful choices to tiny, everyday, large, and hardest using verified host capabilities and cost information. Use exact host model IDs, order candidates cheapest/smallest first, and describe task-relevant capability differences. Do not hardcode or invent model IDs, availability, prices, or size assignments.
  2. Prepare a reusable local profile for the user: {"agent":"codex","models":[{"id":"actual-host-model-id","description":"verified capability for this task","size":"everyday"}]}. Use claude for Claude Code. Support one to four models; never require the user to repeatedly assemble JSON. Remove credentials and private notes from the task. Use the four size labels above only when justified.
  3. For a one-time authorized network trial, run jev-skills route-model --profile PROFILE.json --request-file - --live, supplying the bounded task on stdin. Without --live, it is an offline dry-run. --automatic works only if the model workflow was enabled and project instructions request it. Never use network modes for private or contextual follow-up requests.
  4. Treat route: recommendation and selected as advice, then verify the chosen host model is still available and appropriate. Delegate only a scoped, authorized task through a mechanism the host actually supports. If delegation occurs, have the helper's final report state the actual model used, not merely the suggested model. For fallback, selected: null, or uncertainty, choose with normal judgment or ask for a material missing constraint. Never claim that running this command changed the current model.

For an explicitly authorized helper execution, add --execute to route-model. This starts a separate authenticated host CLI in a temporary directory with conservative tool restrictions, not a repository-editing agent. Pass all necessary non-sensitive text in the request. Return the helper answer to the user only after reviewing it; apply suggested code through the parent agent's normal permissions. Unsupported hosts or errors fall back to normal work. Do not recursively call Jev from a helper (JEV_SKILLS_CHILD=1).

End a delegated answer with the reported model if available. Otherwise say Requested model: ID; serving model not reported by host. Never turn a requested ID into a verified claim. Helper usage is separate from Jev selection cost. Confidence is not a correctness guarantee, permission, or measured savings.

slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall