Lesson 8 · 20 min
Now put it all together in a real run: folder, level 3 goal, frozen tests, how to run it, watching without interrupting, knowing when to stop, and closing with measurement. The kit's pilot, from 10/01/2026, is a step-by-step reference. At the end, the course map and your plan with dates.
Start the lesson
Goal: carry out a real run from goal.md to a measured close, with human gates and stop rules, and leave with a dated plan.
Each step uses a lesson from the course. Click to see what to do.
Most of the work is in steps 1 and 2. In the pilot, with them done well, the run finished in the first cycle.
longrun/2026-10-01-medir-sessao/: build the medir-sessao.py from lesson 7 with a headless loop.
Goal with a complete specification of the output format, 12 fast tests and 2 with real sessions of up to 1.8 GB, frozen (hash 3ae1ba7975508cba).
loop.env: 6 cycles of 20 min, 8G of memory, stagnation 3; protected files tests/, pytest.ini, and tools/medicao/.
14:19:41 start with 0 passed. The agent exited with rc=0 in 178 s, without touching the protected files.
14:22:39 fast = 12 passed; checkpoint committed. 14:22:48 CONCLUÍDO: final test 14 passed.
2026-10-01 14:22:48 CONCLUÍDO no ciclo 1: teste final passou (14 passed)
Independent verification: test hash unchanged, no hardcoded values in the code, 14 passed reproduced, and a never-seen session (from August) checked against the old script.
Measurement: 332,791 input, 304,000 cache read, 4,296 output (cache 91.3%), 0 compactions.
With frozen tests and a precise specification, 1 cycle. The test that decides belongs to the loop, not the agent.
The compaction hook only talks to an active longrun: a run that finished without a warning is the right behavior, not a failure.
A run at cycle 4. You decide whether it continues. The stop rules are the kit's.
Choose your answer. The feedback appears right after, and the conversation continues based on your choice.
The prompt.md is what the loop sends to the agent in each cycle. Generalized from the pilot.
You are in one cycle of a long run in the repo <project path>. RESULT: <the goal.md result in one sentence>. BEFORE ACTING: read goal.md → state.md → plan.md → the last lines of progress.md and failures.md (all in longrun/<folder>/). IN THIS CYCLE: pick the next useful action, implement it, run <quick test>. If the quick ones pass, run <full test>. AT THE END OF THE CYCLE (required): - rewrite state.md (Funciona / Falta / Como retomar) and plan.md (next 3 steps); - append 1 line to progress.md: date-time, what changed, test result (N passed / N failed); - if something went wrong, 1 line in failures.md: error → attempt → result; - design decisions in decisions.md. - When the full test passes completely, write "concluído" on the first line of "## Funciona" in state.md. CONSTRAINTS: do not change <protected files>. No API, no network. Do not commit (the outer loop does that).
cd ~/projetos/<seu-projeto> head -5 longrun/<pasta>/state.md # says concluído? <teste completo> # run it yourself, don't read the log git diff --exit-code <commit-base> -- tests/ pytest.ini; echo "protegidos: $?" git diff --stat <commit-base> # only what the goal allowed? grep -rn "<valor esperado de um teste>" <pasta do código> # hardcoded value in the code? python3 ~/projetos/execucao-longa/tools/medir-sessao.py <sessão.jsonl>
A real run of your own. Start small: 1 to 3 hours.
In the folder from lessons 1 and 2, freeze the tests, pick the mode (/goal or loop-longrun.sh rehearsed with a fake agent) and let it run with the hook and the vigia on. Step in only for the stop rules.
Done when: the first line of "## Funciona" in state.md says concluído and the loop.log (in headless) ends in CONCLUÍDO no ciclo N, or the run stopped on one of the kit's rules with the reason in failures.md.
Close with the independent verification block and measure the session. Write the final line in progress.md and, if there was a failure, the smallest fix in the project's FALHAS.md.
Done when: the full test, run by you, passes; git diff --exit-code [base-commit] -- tests/ → protegidos: 0; progress.md has the final line with tokens and compactions from medir-sessao.py.
Four deliverables, in order. Click an exercise to see its "done when".
lessons 1 and 2
2 exercises
A level 3 goal.md that fails before the work and resists the cheat.
View the lessonlessons 3, 4 and 5
3 exercises
Hook enabled, a year of transcripts, and a handoff you can verify.
View the lessonlesson 6
3 exercises
loop.env rehearsed with a fake agent and vigia enabled.
View the lessonlessons 7 and 8
2 exercises
A run closed by you and measured in progress.md.
View the lessonKit installed, folder created, method in the global rules and the goal.md attacked with the cheapest cheat.
Requires day 1. Bands, pause and handoff rehearsed without spending quota; then a real handoff resumed with /prime.
Requires days 1 and 2. Done, cheat reverted and second-copy lock, before you release a real agent.
Requires everything before. The run happens in one day; the independent close and the measurement, the next.
Folder and goal, context and handoff, headless loop and a real run. Only move on to the next deliverable when the "done when" of the previous one is true.
Mandatory order: a headless loop without a level 3 goal is just a faster way to burn quota.
Pick a start date and how much time you have per week. The plan lists the 20 exercises in order.
Five questions about the whole course.
Question 1 of 5
According to the pilot, where is most of the work in a long run?skill
Question 2 of 5
The state.md says done. What's missing?skill
Question 3 of 5
Which of these is NOT a reason for the kit to stop and call the human?
Question 4 of 5
Context band 2: what does the hook ask for?
Question 5 of 5
Session idle for 2 h, 300k tokens, this session's handoff written at 45 min. When you come back, the cheapest option from there is...skill
Pick an answer. The explanation shows up after you try.