Skip to content
Cameron Mills

All projects

One build request in, a supervised multi-stage build out

The /build skill

Part of Life Tracker

Type one request and get a staged series of unattended agent runs (discovery, research with adversarial review, a plan, build, ship and a final review), where each step opens only once a supervisor has checked the last one's work.

Code on GitHub

How it works

How one /build request becomes a staged pipeline where each stage opens only after a check. Select a layer to open it.

Problem

Long builds done by unattended agents went wrong in repeatable ways. I wanted to ask for a build and have that handled for me. The design came from a written record of two weeks of failures. Builds aimed at the wrong target because decisions were copied into briefs and went stale. Most builds had no review at all. And a stage could start the moment a file appeared, while the previous run was still writing it.

Approach

Research and adversarial review became stages in their own right rather than paragraphs in a prompt, and each review writes findings the next stage has to answer point by point. "Done" is a claim the supervisor checks, not a fact.

Architecture

A request becomes discovery, rounds of research each followed by an adversarial review, a plan, build stages, a ship stage and a final review, all queued at once. When a request is unclear the job is sized up rather than down, because the commonest fault in the record was a job sized too small. Discovery reads the target from the repository itself, and every build happens on its own branch.

A supervisor is the only thing that can pass a stage. It checks that the output is fresh and complete, that the stage report exists, that the agent is not still waiting on something, and that a stage which finished unusually fast explains why. Build and ship stages must also record that the tests passed. A failed check sends the stage back with a note saying what was missing, and a refusal or a logged-out session pauses the whole pipeline with one notification.

What I built

Every stage brief carries the same set of guards, among them who may merge or deploy (only the ship stage) and a full test run before any merge. A self-test fails if any brief is missing one. I can add a note while a build is running, and every stage that has not yet passed must answer it. Steps only a person can take reach me as soon as the plan is agreed, not at the end.

Since 14 October the queue step also makes the build's branch and worktree and prints a seven-row preflight table, and it refuses to queue anything if a row fails. Discovery opens with a write probe, one test patch sent the whole way to a throwaway clone, and two build stages in a row that land no commits stall the pipeline. Headless runs cannot write to git, so a build in the life tracker's repository lands its code as patches that the supervisor applies to the build's own branch. Anything that reaches main goes through a Fable reviewer, not a card for me.

Results

A review found the supervisor was passing stages on the agent's word alone and not reading the test results. Each fix got its own test. When every stage in the live pipelines was re-judged under the stricter rules, all of them still passed. Builds have run through it overnight since.

The /build skill: results against their thresholds
CheckExpected or thresholdActualResult
Stages already passed, re-judged under the stricter checksAll still passAll of them, in seven pipelinesPass
Pipeline self-testExit 0234 checks, all passedPass
A review tightened the supervisor, then re-judged the live pipelines under the new rules.

What didn't work

A status check run inside a sandbox read live runs as dead, and the supervisor started second attempts in folders still in use. The fix only counts a run as dead on positive evidence, and locks each build folder. The check for an agent that is still waiting only recognises phrasings it has seen before, so the completion marker remains the real guard.

The first chief-of-staff build passed seven build stages with no code at all, because discovery could not make its worktree and the gates only looked for output files. The preflight, the probe and the empty-build checks answer that, but they apply only to pipelines made from 14 October on.

Stack

  • Python
  • Claude Code
  • launchd
  • git worktrees

How these figures are sourced

The outcome figure is agent-reported: it is taken from a test log or report written by the agents that built and ran the system. The status was read from the code repository at commit c8e90f7 on 1 Oct 2026. Every figure and claim on this page is tied to a source file or a CV line in the site's content.