ShortStackEngineer

You can author verified .loop.yaml loop-generator specs, turn vague requests into falsifiable acceptance checks, decompose full application specs into dependency-ordered vertical slices, diagnose failed or vacuous loopgen runs, and scaffold new AgentDriver backends driven to green against the conformance harness.

Developers building agentic coding feedback loops who need to write .loop.yaml specs, define a definition of done, plan build order for multi-feature apps, debug stalled loop runs, or integrate agentic coding backends such as Codex, Aider, or Cursor via the driver conformance harness.

Scaffold a new AgentDriver implementing name, preflight, and run, register it, then iterate on `loopgen verify-driver <name>` fixing one failing scenario at a time until all four conformance scenarios pass. For LM Studio, Ollama, or local OpenAI-compatible servers, use the existing opencode driver instead.

It reads the run report (terminal output or --report JSON), classifies the outcome as max-iterations, preflight-failed, error, baseline-vacuous, spec-tampered, aborted, or vacuous-success, reproduces the failing check without spending agent budget, and maps it to a concrete spec or workspace fix.

You need a target repository with real test, build, or typecheck commands that the skill can inspect, plus a developer interview defining the goal. The skill proves the resulting .loop.yaml spec is lint-clean and that its checks start RED before handing it off to loopgen.