Lesson 6 · 20 min
In the interactive /goal, an evaluator reading the conversation decides when it's done. In the headless loop, the test decides, run by the loop itself, with the test files protected. In this lesson you run loop-longrun.sh with a fake agent: you see it finish, you see the cheat get reverted, and you see the lock stop two loops in the same folder, without spending any quota.
Start the lesson
Goal: run loop-longrun.sh with a fake agent and watch the test decide: done, cheat reverted, stagnation, and second-copy lock.
The agent works; the loop decides. Click each step.
If what's missing only shows up in the final test (slow), every cycle counts as "no progress": put those cases in the fast test too, or raise ESTAGNACAO (a warning from the script itself).
Run it, interrupt it halfway, and run it again. One of the cycles touches a protected test.
Illustrative simulation in the browser. The real rehearsal, with a fake agent, is right below.
The loop accepts CODEX_BIN: point it at a script and no model is called. It's the cheap way to test your loop.env before spending quota.
# ensaio-loop.sh — monta um projeto de teste com 3 testes congelados e dois agentes FALSOS
E=/tmp/ensaio-loop; K=~/projetos/execucao-longa
rm -rf $E && mkdir -p $E/proj/tests && cd $E/proj
git init -q && git config user.name ensaio && git config user.email ensaio@local
cat > tests/test_conta.py <<'PY'
from conta import conta_palavras
def test_vazio(): assert conta_palavras("") == 0
def test_simples(): assert conta_palavras("um dois três") == 3
def test_espacos(): assert conta_palavras(" um dois ") == 2
PY
printf 'def conta_palavras(texto):\n raise NotImplementedError\n' > conta.py
printf '[pytest]\npythonpath = .\n' > pytest.ini
printf '__pycache__/\n' > .gitignore
git add -A && git commit -qm "base: testes congelados"
$K/tools/novo-longrun.sh $E/proj conta-palavras > /dev/null
D=$(ls -d $E/proj/longrun/*)
echo "Implemente conta_palavras em conta.py. Leia goal.md, state.md e plan.md antes. Não mexa em tests/." > $D/prompt.md
cat > $D/loop.env <<'ENV'
TESTE_RAPIDO="python3 -m pytest -q"
TESTE_FINAL="python3 -m pytest -q"
PROTEGIDOS="tests/ pytest.ini"
PERMITIDOS="conta.py"
MAX_CICLOS=4
MIN_CICLO=2
MEM=1G
ESTAGNACAO=3
ENV
git add -A && git commit -qm "longrun: pasta da execução"
printf '#!/usr/bin/env bash\ncat > /dev/null\nprintf "def conta_palavras(texto):\\n return len(texto.split())\\n" > conta.py\n' > $E/agente-bom.sh
printf '#!/usr/bin/env bash\ncat > /dev/null\nrm -f tests/test_conta.py\n' > $E/agente-trapaca.sh
printf '#!/usr/bin/env bash\ncat > /dev/null\nsleep 6\n' > $E/agente-lento.sh
chmod +x $E/agente-*.sh
echo "pronto: $D"bash ensaio-loop.sh cd /tmp/ensaio-loop/proj && D=$(ls -d $PWD/longrun/*) && L=~/projetos/execucao-longa/tools/loop-longrun.sh CODEX_BIN=/tmp/ensaio-loop/agente-trapaca.sh $L $D; echo "saída: $?" # A) trapaça CODEX_BIN=/tmp/ensaio-loop/agente-bom.sh $L $D; echo "saída: $?" # B) trabalho certo CODEX_BIN=/tmp/ensaio-loop/agente-lento.sh $L $D > /dev/null 2>&1 & # C) dois loops sleep 2; CODEX_BIN=/tmp/ensaio-loop/agente-lento.sh $L $D; echo "saída: $?"; wait
TESTE_RAPIDO="pytest -q -m 'not real'" TESTE_FINAL="pytest -q" PROTEGIDOS="tests/ pytest.ini tools/medicao/" PERMITIDOS="tools/medir-sessao.py" MAX_CICLOS=6 MIN_CICLO=20 MEM=8G ESTAGNACAO=3 AGENTE=codex MODELO=<o modelo topo da sua assinatura no Codex> # se os testes sobem servidor local, o sandbox workspace-write bloqueia até 127.0.0.1: # CODEX_ARGS="-c sandbox_workspace_write.network_access=true"
## A) trapaça 2026-10-19 03:56:46 início: rápido=0 passed · agente=codex · teto 4x2min · mem 1G 2026-10-19 03:56:46 ciclo 1: começa 2026-10-19 03:56:46 ciclo 1: agente saiu rc=0 em 0s 2026-10-19 03:56:46 ciclo 1: agente mexeu em arquivo protegido → revertido, ciclo falho 2026-10-19 03:56:46 ciclo 1: checkpoint commitado (… ciclos 2 e 3 iguais …) 2026-10-19 03:56:47 PARADO: 3 ciclos sem avanço mensurável — ver failures.md saída: 2 ## B) trabalho certo 2026-10-19 03:56:48 início: rápido=0 passed · agente=codex · teto 4x2min · mem 1G 2026-10-19 03:56:48 ciclo 1: começa 2026-10-19 03:56:48 ciclo 1: agente saiu rc=0 em 0s 2026-10-19 03:56:48 ciclo 1: rápido=3 passed (melhor antes: 0) 2026-10-19 03:56:48 ciclo 1: checkpoint commitado 2026-10-19 03:56:49 CONCLUÍDO no ciclo 1: teste final passou (3 passed) saída: 0 ## C) dois loops outro loop já roda nesta pasta saída: 3
The tests deleted by the cheat came back: the loop reverted all three cycles and wrote three "revertido pelo loop" lines in failures.md.
Without the __pycache__/ line in .gitignore, pytest itself creates new files inside tests/. The loop sees an untracked file in a protected folder and counts the GOOD agent's cycle as cheating. Caught in the same rehearsal, without the .gitignore:
2026-10-19 03:56:28 ciclo 1: agente saiu rc=0 em 0s 2026-10-19 03:56:28 ciclo 1: agente mexeu em arquivo protegido → revertido, ciclo falho 2026-10-19 03:56:29 CONCLUÍDO no ciclo 1: teste final passou (3 passed)
It finished because the final test passed, but the cycle was marked as failed and went to failures.md. With the slow agent, three cycles like this would trigger stagnation. Before you turn the loop on: git status --short tests/ after running the tests must come out empty.
The loop runs on its own; these three pieces look after it.
vigia.py looks for active runs (state.md without "concluído") and alerts once for each problem: PARADO (loop.log recorded a stop), PARADO SEM AVISO (there is a loop.log but no live loop process), and OCIOSA (nothing changed in the folder or in the sessions for N minutes).
Alert = a line in ~/.local/state/execucao-longa/alertas.log; notify-send only with --desktop.
The units in tools/systemd/ carry the absolute path from the author's machine (/home/nmaldaner/…). The sed in the practice block swaps that path for your HOME before enabling them; without it, the timer fires and the service fails silently.
grep -c nmaldaner ~/.config/systemd/user/longrun-vigia.service → 0; systemctl --user list-timers longrun-vigia.timer → shows the next run
Every cron job that calls an agent gets flock -n on a lock file and a timeout with a cap. With claude -p, add --max-turns too.
Without flock, a slow run and the next one run together in the same folder.
0 2 * * * flock -n /tmp/meu-loop.lock timeout 4h ~/projetos/execucao-longa/tools/loop-longrun.sh ~/projetos/[project]/longrun/[folder]
Copy the kit's units, swap the path for yours, and enable the timer.
mkdir -p ~/.config/systemd/user cp ~/projetos/execucao-longa/tools/systemd/longrun-vigia.* ~/.config/systemd/user/ sed -i "s#/home/nmaldaner#$HOME#g" ~/.config/systemd/user/longrun-vigia.service systemctl --user daemon-reload systemctl --user enable --now longrun-vigia.timer systemctl --user list-timers longrun-vigia.timer python3 ~/projetos/execucao-longa/tools/vigia.py --seco # one round by hand, without saving state
The three rehearsal rounds, then the watchdog.
Run ensaio-loop.sh and round A (agente-trapaca).
Done when: the last line of the log is PARADO: 3 ciclos sem avanço mensurável, saída: 2, ls tests/ still shows test_conta.py and grep -c "revertido pelo loop" $D/failures.md → 3.
Run rounds B and C.
Done when: B ends with CONCLUÍDO no ciclo 1: teste final passou (3 passed) and saída: 0; C prints outro loop já roda nesta pasta and saída: 3.
Install the watchdog timer with the path swapped. Leave the folder of your real long run (lesson 1) without "concluído" in state.md.
Done when: systemctl --user list-timers longrun-vigia.timer shows the next trigger and systemctl --user status longrun-vigia.service shows no failure in the last round.
Question 1 of 5
In the headless loop, who decides the run is over?skill
Question 2 of 5
The agent deleted a test that was failing. What does the loop do?skill
Question 3 of 5
The loop exited with code 3. What happened?
Question 4 of 5
The good agent was flagged as "touched a protected file". Most likely cause?
Question 5 of 5
What is MemoryMax per cycle for?
Pick an answer. The explanation shows up after you try.