Yarstack

Yarstack


Generating a diff is not the same as engineering a change. The hard parts are deciding what to change, keeping the design coherent, proving the behavior, and leaving the repository easier to work in.

Yarstack packages those engineering disciplines as reusable workflows for Codex and Claude Code. It helps an agent plan before coding, implement in small steps, test the contract, review concrete risks, and deliver only when the evidence supports it.

This is not a prompt that asks an agent to "be a senior engineer." It is a set of explicit workflows and engineering standards that define what the agent must inspect, what it may change, how it should verify the result, and where it must stop.

The same marketplace also publishes Yarbrain as a separate plugin. Yarbrain keeps durable agent knowledge in a user-controlled Markdown vault without mixing that state into Yarstack's engineering workflows.

What Yarstack is built for

Planning that can be implemented

A useful plan settles the decisions that change the implementation path. It names ownership, interfaces, dependencies, failure cases, compatibility needs, rollout order, and acceptance evidence. Yarstack separates architecture discussion, technical research, plan creation, and plan correction so each has one clear job.

Investigations that end in a decision

`system-investigate` owns a read-only investigation from an initial hypothesis to an evidence-backed verdict. It traces repository and runtime behavior, coordinates independent evidence gathering when useful, checks current public sources when they can change the conclusion, and recommends the smallest safe next step without starting implementation.

Correctness before completion

Code that works on the happy path is not enough. Yarstack makes normal paths, boundaries, partial failure, cleanup, cancellation, ordering, retries, stale state, and recovery part of the work when they apply. Val