Introduction

Our loop

A run covers one issue and the sub-issues under it. It claims the issue on the board, creates a git worktree for each worker, takes each change through a fixed sequence of review passes, and stops once a pull request is open. A person starts every run, and a person merges it.

tstorm

The skills describe the loop’s ten stages, from /rubber-duck to /handoff, the last of which leaves notes for the next session when work stops partway. First run lists them and walks through one. tstorm is the small app that renders the payloads, runs the gates, answers the harness hooks, reads and writes the board, and starts a session per role. The checks reference says what each check reads and what its exit code means, and the harness reference says what each agent gets.

Why does this exist?

The loop began as six directories inside a single repository. A second repository could copy them, and the two copies drifted within weeks. Installing gives both of them the same fix.

The checks are separate from the skills for a different reason. A skill is an instruction, and when an agent ignores one, nothing records that it happened. A check that exits non-zero leaves a result someone can read afterward.

Loop engineering

The loop is ReAct’s (Yao et al., ICLR 2023): the model acts, reads what happened, and acts again. What separates one coding agent from another is the middle step. Here it is a test result, an exit code from tstorm check, or a comment from a review pass that ran as its own session.

Letting the model grade itself was the cheaper option and it does not hold. Huang et al. (ICLR 2024) had models revise their own reasoning with no outside signal: accuracy did not improve, and on some tasks it fell.

Each agent gets a payload of its own because of what SWE-agent (Yang et al., NeurIPS 2024) measured. Giving a model an interface built for it, rather than the one a person uses, took SWE-bench to 12.5% pass@1 where non-interactive models had been far below that. Skills, commands and hooks are that interface here, and no two agents read them from the same place in the same format.

Fixing the stages and their order makes a run a workflow in the sense Anthropic (Schluntz and Zhang, 2024) separates from an agent, which directs its own process; the model’s discretion lives inside a stage. That post argues for composing simple patterns over adopting a framework, which is the case for a loop written as skills a person can read.

Fan-out stops at five sub-issues, each with its own worktree and its own review sequence. Cognition (Yan, 2025) is the reason for the ceiling: parallel agents make decisions none of the others can see, and the conflicts surface late, in the merge.