Where the tools sit in the loop

These tools aim at one picture: a development loop that an agent runs from end to end, while a person hands over the intent and keeps only two decisions: a change that cannot be undone, and turning a change on for everyone. This page draws that loop in six stages and places each tool at the stage it serves. A stage with a dashed border has no tool on this site yet.

1. Refinement

   • grilling-pane draws the intent out of a person, with the open questions asked together

2. Implementation

   • polylinedb tracks tasks, who holds each one, and project memory
   • spacequery shows which agents run where, and which worktrees sit idle

3. Verification

   • nukadoko turns the acceptance criteria written in refinement into typed checks that run
   • firstlight runs lint, typecheck, and tests in parallel
   • archstrict checks module boundaries and layers
   • stryker-agent-reporter turns surviving mutants into a gate
   • mutation-testing-skills walks an agent through mutation testing
   • depug measures what a failing test received
   • finderscope ranks where the time and memory went

4. Human judgment

   No public tool yet. A person approves the evidence and the release scope rather than the diff, and needs the knowledge to judge them: how the behavior changed, and which parts of the model it touches.

5. Staging

   No tool yet. The same scenarios should run here, and an agent should loop on what they show.

6. Production

   No tool yet. A flag should roll a change out in steps, and turn it off by itself when a declared stop condition is met.

After production, what was learned goes back into refinement, and the loop starts again.

What the picture shows

The tools so far sit in the first three stages, and most of them in verification. nukadoko sits there too: refinement writes the acceptance criteria, and nukadoko makes them a check that runs. That is where an agent’s mistakes are cheapest to catch, because a check there is a command it can run again in seconds. The right side of the loop is open: nothing here yet carries a change through staging and production, or tells an agent what happened there. That is where the next tools go.

The tools also run on each other, and how the tools fixed each other records what that found.