Specification, 99 acceptance tests, guardrails and prompts so an agent can build, in your own environment, the front desk of a small gym or studio: memberships and enrollment, classes with spots and a waitlist, check-in, workout card, animated exercises and the staff on Telegram. The system is not implemented yet: this repository is the plan, and you run the implementation.
There is no finished app here. There is the contract for what the app must do, the test battery that proves it and the guardrails that stop the agent from cheating. The builder is Claude Code or Codex, running on your machine, on your subscription. Its sibling Atende Fisioterapia follows the same mold, and atende-clinica, which came first, was built exactly this way and closed 93 of 93 tests.
Members and plans, group classes with booking, check-in and the instructor's workout card. At the center is the class with spots: booking with a cap, a waitlist with an automatic offer and an automatic no-show. Python 3 with the standard library only and SQLite, one gym per installation.
A 23-section specification (0 to 22), 99 black-box tests, 20 decisions already proposed, a map to Raio-X, prompts for /goal and for the headless loop, and the guardrails that freeze the contract. An adversarial validator went through all of it and found 10 problems, 2 of which would have stopped an honest implementer.
No-shows without notice, idle time slots and members who vanish are the leaks the plan attacks. The system hands six monthly numbers to the dashboard of Raio-X de Margem, the source of the business rules. It has no gym package: the kit uses the generic services-with-appointments model, and each mapping is an assumption.
Monthly, quarterly, annual and drop-in class plans, with price and validity. An enrollment is ativa, vencida, trancada or cancelada (active, expired, frozen, canceled). Payment is only a manual status (pago or pendente, paid or pending) set by the staff, plus an expiry notice. No Pix, card or payment processor in v1.
Members check in with a code at the front desk or in the chat (cheguei, "I'm here"), once a day. Anyone who has vanished for N days gets a win-back message, but only if they gave consent: win-back is marketing, not a contract notice. "PARAR" (stop) blocks everything automatic.
Export and anonymize a member's data. The physical assessment result (weight, height, measurements) is sensitive data: only the staff can see it, never through the chat, and it is erased together with the member. LGPD and 192 are Brazil's: LGPD is Brazil's data protection law and 192 is the Brazilian ambulance number.
This is the flow of the planned system, what the agent will build. A Python server with no dependencies, SQLite in a data folder and Docker for the VPS. Evolution and Telegram come in only through environment variables; without them the outbox is simulated and no network call goes out.
In the chat or on WhatsApp they ask about classes, book and cancel, see meu plano (my plan) and meu treino (my workout), ask explica agachamento (explain the squat), say cheguei, get FAQ answers, ask for a human and send "PARAR" when they want no more messages. A health alert word (strong pain, chest pain, shortness of breath, fainting, dizziness, injury and others) is answered with 192 and warns the instructor. The chat is in Portuguese in v1.
Notices of new and canceled bookings, and the commands /aulas [date], /fila, /responder and /encerrar, also as a direct reply to the member's message.
With a token: members, plans and enrollments, class grid, bookings, check-ins, workout cards, assessments, exercise library, FAQ, registries and the human queue.
The weekly grid has modality, instructor, room and spots. Booking respects the spot cap and the lead time; with 10 requests for the last spot, one gets 201 and the other nine get 409. When someone cancels, the first person on the waitlist receives the automatic offer.
Canceling inside the window (default 2 h before) returns the credit. After the window the spot is released and the waitlist fires, but the drop-in credit does not come back. A booking without a check-in on the day of the class becomes an automatic no-show in the task run, and the staff can correct it.
The instructor builds the card from the exercise library (sets, reps, load, rest). The member asks meu treino de hoje (my workout today) and gets the next workout in the sequence, one per check-in. The bot shows the instructor's card and never prescribes.
Each exercise is an SVG animated with CSS only, no script and no SMIL. Measured in a real render, HyperFrames has no SMIL adapter: SMIL motion came out frozen or out of phase in the MP4, and only CSS stayed faithful. This is the squat: feet fixed on the floor, knees forward, hips back.
Squat, push-up, plank, bent-over row, bench press, deadlift with a stick, lunge, hip raise, crunch, shoulder press, lat pulldown and jumping jack.
Each one has a public page /exercicios/<id> with the animated SVG, the steps, common mistakes and precautions. On WhatsApp a 4 s MP4 rendered by HyperFrames goes out, a local build step (tools/render-exercicios); without the MP4 (or without PUBLIC_URL) the page link goes out instead. The 12 SVGs are drawn by the implementer, and the motion is checked at level 4 by verificar-independente.py, which says PULADO (skipped) when the tool is missing.
To run the plan you only need Python for the tests and an agent logged in through your subscription. Docker only comes in at the final check; Evolution and Telegram are optional and yours. For the exercise videos, Node 22 or newer, HyperFrames and ffprobe.
The tests use pytest (a development tool). The application itself will have no dependencies.
# check python3 --version python3 -m pip install pytest
Through the subscription, no paid API. For the headless loop, clone execucao-longa into ~/projetos.
# check codex --version # or: claude --version
Docker does the final build of the independent check. An Evolution instance and a Telegram bot only if you want the channels; the credentials live in .env.
# check docker --version
The commands below are the repository's own. The cycle: answer the decisions, freeze the contract, run the agent until 99 passed and LIMITES OK, check with the independent verification, generate the videos and ask a physical education professional to review.
There is no ./atende yet: the repository only has the plan. Check that the suite is collected without errors and that nothing passes yet (without the implementation, the tests fail or error out, and that is expected).
git clone https://github.com/inematds/atende-academia && cd atende-academia python3 -m pytest -q --collect-only | tail -n 1 # 99 tests collected python3 -m pytest -q | tail -n 1 # 30 failed, 69 errors (0 passed)
Read docs/DECISOES-ABERTAS.md: 20 default proposals already applied in the specification and the tests. Answering "ok on everything" unlocks the run. If you change anything marked ⚠ (for example win-back only with consent, the automatic no-show or the cancellation window), edit docs/ESPECIFICACAO.md and tests/ before step 3 and check the count. For a real pilot, also replace exemplos/academia.json with your own grid, plans, instructors, FAQ and exercises.
less docs/DECISOES-ABERTAS.md python3 -m pytest -q --collect-only | tail -n 1
With a clean git status, the script records the hash of tests/, pytest.ini, the specification and the verifiers in hash-congelado.txt and makes one commit. From then on, any change to the tests is detected.
git status bash longrun/2026-10-05-academia-v1/congelar.sh
a) Headless loop with Codex (recommended). The loop.env sets 20 cycles of 30 min, 8 GB per cycle, a stop after 3 cycles with no progress, model gpt-6-astra and CODEX_ARGS="-c sandbox_workspace_write.network_access=true": without it the Codex sandbox blocks the tests' local server.
~/projetos/execucao-longa/tools/loop-longrun.sh longrun/2026-10-05-academia-v1
b) /goal in Claude Code. Open a new Claude Code session in the project folder and follow longrun/2026-10-05-academia-v1/prompt-goal-claude.md: paste the /goal condition (the output must show 99 passed and LIMITES OK) and, as the first message, the block from prompt-goal-codex.md.
claude # new session, inside atende-academia
c) Codex TUI already open. Paste the whole content of longrun/2026-10-05-academia-v1/prompt-goal-codex.md (it starts with /goal).
codex -c sandbox_workspace_write.network_access=true
In the loop, loop.log shows each cycle and progress.md has one line per checkpoint; state.md says what works, what is missing and how to resume. A test in conflict with the specification goes to failures.md: it is a human gate, and the agent must not adjust tests or the specification.
tail -f longrun/2026-10-05-academia-v1/loop.log cat longrun/2026-10-05-academia-v1/state.md
When the agent says "done", run the three commands. The last one the agent has never seen: it looks for fixture values copied into the code, brings up a gym that never appeared (another grid, other plans and windows, a holiday, another clock), opens 3 SVGs in Chromium to prove they move, renders one exercise through HyperFrames and runs docker build with a healthcheck. Without Chromium, without HyperFrames or without Docker, those stages come out as PULADO, never as OK. Then open /, /equipe and /exercicios/prancha in the browser.
python3 -m pytest -q tests/ # 99 passed bash longrun/2026-10-05-academia-v1/verificar-limites.sh # LIMITES OK python3 longrun/2026-10-05-academia-v1/verificar-independente.py # INDEPENDENTE OK
A human step, after the implementation: tools/render-exercicios (written by the agent) uses only the HyperFrames already installed, downloads nothing, and needs Node 22 or newer, Chromium and ffprobe. It produces web/exercicios/<id>.mp4 (4 s, 720×720, no audio) from each SVG. Install HyperFrames after the loop closes, because the agent must not download packages. WhatsApp fetches the video from the public URL, so the MP4 path requires PUBLIC_URL with HTTPS.
npm install hyperframes tools/render-exercicios --ids agachamento,prancha --saida web/exercicios
With the system implemented, copy the example, change the token and start the server. The member chat is at /, the staff page at /equipe (header X-Token) and each exercise at /exercicios/<id>. The VPS deploy is yours, with your own credentials; the agent writes the playbook (Docker, Evolution and Telegram webhooks, backup) in the README, section 22 of the specification.
mkdir -p dados && cp exemplos/academia.json dados/academia.json ./atende serve --porta 8080 --dados dados ./atende raiox --dados dados --mes 2026-10 --acompanhamento acompanhamento.json
The 12 SVGs are schematic figures drawn by the implementer, and the steps, common mistakes and precautions of each exercise were written as examples: a physical education professional must review everything before it reaches a real member, including the list of health alert words. The professional council rules (CREF/CONFEF, Brazil's physical education boards) were not researched: the campaign guardrail is only the safe minimum (promise of results and "before and after"), citing no article, and the owner confirms the rules before any campaign is published (decision 5). The legal basis for the physical assessment, sensitive data, is also for the owner or a lawyer to confirm (decision 6).
Everything the agent needs to build, and everything that stops it from pretending it built.
Black-box, over HTTP and the command line, in 13 files.
test_integracoes.py 16 test_exercicios_svg.py 14 test_aulas.py 12 test_alunos_planos.py 9 test_conversa.py 8 test_treino.py 6 test_cadastros.py 6 test_basico.py 6 test_docker.py 5 test_checkin.py 5 test_campanhas.py 5 test_avaliacao.py 4 test_lgpd_raiox.py 3
docs/ESPECIFICACAO.md, sections 0 to 22: execution, academia.json, members, plans and enrollments, group classes, check-in, workouts, animated explanation, physical assessment, HTTP API, chat, tasks, campaigns and the advertising guardrail, Raio-X, LGPD, Evolution, Telegram and Docker. Pix or card billing, turnstiles, a member app, personal training and several gyms on one server are explicitly out (section 16).
A frozen hash of tests/, pytest.ini, the specification and the verifiers. File scope: the agent only touches the code and its own logs. No API key, no external URL in the code, no 0.0.0.0 and no SMIL or script in the SVGs. verificar-limites.sh checks all of it and prints LIMITES OK.
Level 4, hidden from the agent: it looks for fixture values in the code, brings up a gym never seen before, checks the SVG motion and the MP4 when the tools exist, and runs docker build and a healthcheck. Only that way does "99 passed" prove general logic.
Language, channels, manual billing only, the campaign guardrail, physical assessment as sensitive data, win-back with consent, automatic no-show, late cancellation, check-in by code, the 12 exercises with CSS-only animation and Raio-X. Only the items marked ⚠ change the contract.
A validator that did not see the planning checked tests against the specification and tried to get around the guardrails. It found and fixed 10 problems: 2 blocked the run, 5 got in the way and 3 were cosmetic. The main one: the specification allowed SMIL, and the HyperFrames render has no SMIL adapter, measured in a real render. The full account is in docs/VALIDACAO.md and each fix has a line in FALHAS.md.
No-shows without notice, idle time slots, members who do not return and a front desk stuck on the calendar: docs/MAPA-RAIO-X.md links each leak of the services package to a v1 piece or to the reason it stays out. Raio-X carries no gym market figures, so no-shows, occupancy and renewal stay as assumptions to measure in your own gym. Details in the Raio-X de Margem guide.
LGPD (expiry notices, waitlist offers and canceled classes as contract performance; win-back and campaigns only with consent; physical assessment as sensitive data; export and anonymize) and 192 for emergencies. The professional council rule (CREF/CONFEF) has not been researched yet. Outside Brazil, swap in the local privacy law, professional board and emergency number.
The plan is ready and validated. The implementation does not exist yet: it is born when you run /goal or the loop, using the execucao-longa method. atende-clinica followed the same path and closed 93 of 93.
99 passed and LIMITES OK, then the independent verification. Today the system is not implemented.tools/render-exercicios once, with HyperFrames installed, and publish the files together with PUBLIC_URL.