ABVX Agent Skills capability

Goal Loop Designer

Compile a raw agent goal into a bounded loop harness before launching a long Codex, Claude Code, MiMo, or local-LLM run.

GitHub skill Skills catalog

What it does

Critiques a raw /goal before the agent starts spending tokens.
Turns vague done criteria into an explicit rubric with evidence requirements.
Defines worker, judge, retry, budget, memory, and stop rules.
Produces portable artifacts that can be reused across Codex, Claude Code, MiMo Auto, and local OpenAI-compatible models.

Positioning

This is not a server-side agent runner. It is a preflight compiler for safer agent loops.

The source of truth lives in markoblogo/abvx-agent-skills; ABVX Lab only catalogs the capability and points to install paths.

Quickstart

Install and use

gh skill install markoblogo/abvx-agent-skills goal-loop-designer --agent codex --scope user

Use goal-loop-designer. Compile this raw goal into a bounded loop harness:

"/goal Prepare this repo for a release, fix blockers, run verification, and stop only when it is ready."

For fallback review, use MiMo Auto as a separate judge when provider access is available. Use local Ollama only after a direct response smoke test passes.