Reading code is not running it

A pair once wrote a helper and its tests. The reviewer approved. The auditor, reading the whole, approved. The test file imported ./slugify.js for a file named slugify.ts, and the suite never loaded. Both agents had read the code. Neither had run it, and nothing in either prompt would have made them: a model reading import "./slugify.js" sees a plausible line.

/run build runs the shipped build flow, from a sentence to a working tree with the change in it, and nobody is asked anything on the way. The one thing it insists on is that your own check runs before anyone signs.

Give it a check

The check is a script of your project, .pi/checks/test.sh, and the shipped flow names that path. This page runs on a small repository, one words.js helper and its test, whose script is two lines:

#!/usr/bin/env bash
exec node --test

Without it, the run is refused before a single subagent spawns:

/run build add a slugify helper with tests

Error: run: `build` cannot run here
  flows/build.md deliver/tests.check: `.pi/checks/test.sh` is not there, from
    `/…/tut-slug`

The script is read when the run starts and run with bash after each round of work, so what runs is what was read, whatever an agent does to the file in the meantime. The command line holds no check: --model and --timeout are its only flags, and everything else, the check included, is in the flow’s file. A directory that is not a git repository is refused the same way, since each subtask works in a copy that lands back through git.

Run it

/run build add a slugify helper to slug.js, with tests --model <provider/model>

The request is the brief, as typed. Nothing asks you to confirm it and nothing asks for a commit at the end: a question in the middle of a run is a run waiting for whoever left it going. When the request needs thinking through first, /run build-attended interviews you on it, shows the specification and asks “Build this?” before building.

What runs

The widget draws the flow’s plan and fills it as it goes. Two minutes into the second of the two runs this page made, the coder had finished and the reviewer was thinking:

● build · 3 visits · 2m · ↑61k ↓6.1k
✓ locate · scout · 27s · ↑40k ↓2k
✓ plan · planner · 17s · ↑4k ↓956
● deliver · #1 of 2
  ● deliver#1
    ● deliver#1/work · 0/1 so far
      ● deliver#1/work[1]
        ● deliver#1/work[1]/pair · #1 of 3
          ● deliver#1/work[1]/pair#1
            ✓ deliver#1/work[1]/pair#1/code · coder · 1m5s · ↑17k ↓3.2k
            ● deliver#1/work[1]/pair#1/review
              ● reviewer#1  thinking…  provider/model · 12.4s
    ○ deliver#1/tests · check .pi/checks/test.sh · timeout 10m · ≤ 20m
    ○ deliver#1/audit · agent auditor (.pi/agents/auditor.md) · reads input, work, tests, diff, deliver.ledger · verd…
○ report · agent synthesiser (.pi/agents/synthesiser.md) · reads input, diff, deliver.output.last.work, deliver.outpu…
esc stops everything · ctrl+↑↓ selects · ctrl+del stops the selected one

The reviewer’s line has no tokens yet: they are read when its turn ends. The summary’s 2m is the run’s own clock, and its tokens add up the visits that have ended.

A scout maps the code. A planner splits the brief into subtasks, a typed list the flow reads rather than prose it parses; this one made one. Each subtask goes to a pair, a coder and a reviewer who remember each other for three rounds at most, the reviewer deciding through a verdict it reads from the diff, not from the coder’s summary of it.

Then your check, deliver#1/tests, and an auditor who reads the whole change against the brief, with the check’s output in front of it as evidence it cannot argue with. What the audit raises becomes the subtasks of a second round, and there is no third.

The round is over when the tests pass and the auditor approves. No approval turns a failing script into a success: the flow’s loop reads tests.output.passed && audit.output.approved, and a build whose second round ends unapproved fails.

What it leaves

The first run of that command went through in one round, under three and a half minutes on a small open-weight model served to several runs at once, and put its answer in the conversation:

Result of the build flow, asked to: add a slugify helper to slug.js, with tests.

I have added a slugify helper to slug.js that converts strings to URL-friendly slugs and created slug.test.js with a
comprehensive set of tests. Nothing is left to do.

ok · runs/2026-09-24_02-53-37

✓ build · 10 visits · 3m22s · ↑139k ↓14k
✓ locate · scout · 1m4s · ↑54k ↓3.6k
✓ plan · planner · 18s · ↑6.7k ↓1.2k
✓ deliver · 1 iteration · 1m57s · ↑76k ↓8.7k
✓ report · synthesiser · 5s · ↑2.5k ↓305

The work is in the working tree, uncommitted:

$ git status --short
?? slug.js
?? slug.test.js

Its reviewer’s first answer was LGTM in prose, without the verdict call its node asks for, and a review that ends without the call fails with schema. The node has retry: 1, so the flow sent the reviewer back once, naming the failure, and its second answer was the call. The summary counts visits, not attempts, which is why nothing on it says so.

The report is written from the diff, not from what the pairs claimed, so a file nobody asked for would be named in it. That is what the report is for: telling you what to look at. It does not replace git diff, and nothing is committed, pushed or undone on your behalf.

Interrupt it

Every fact of a run is appended to runs/<timestamp>/journal.jsonl as it happens. The same command, run again, was killed while its reviewer read; a fresh pi, in the same directory:

/run resume

run: resuming build in runs/2026-09-24_02-53-50, from deliver#1/work[1]/pair#1/review

It says what it picked up rather than asking: typing /run resume was the answer. The scout’s map, the plan and the coder’s work had ended, so they were kept; the review had not, so it ran again, in the copy the coder had left open, with a fresh reviewer that read the tree rather than a replayed conversation. Its summary counts both processes:

✓ build · 10 visits · 3m46s · ↑100k ↓22k · 2 lives (1 partial) · resumed from deliver#1/work[1]/pair#1/review
✓ locate · scout · 27s · ↑40k ↓2k
✓ plan · planner · 17s · ↑4k ↓956
✓ deliver · 1 iteration · 3m · ↑53k ↓18k
✓ report · synthesiser · 4s · ↑3k ↓285

The first life is partial because it was killed before it could write its own measurement: it is counted from the visits it ended, and the review it was in the middle of costs nothing on this bill, since nothing measured it. The fresh reviewer called verdict on its first attempt and approved, the tests passed and the auditor approved.

A run that failed by a decision of its own, a second round unapproved, would decide the same again, and /run resume refuses it with why.

What this page did not show

One subtask, one coder. It did not say where two coders write when the plan has two subtasks and the flow runs them at once, which is the whole of the next page.

Next: Two coders, one tree.