O sintoma primário da Issue (deadlock do waitForChecks no próprio pending)
já tinha sido entregue pelo OPS61. O secundário — o rollup `checks` reportava
SUCCESS com um shard de e2e vermelho porque o Forgejo 9.0.3 reporta o
resultado do ÚLTIMO shard como o parent de matrix no `needs` (runs 604/607/610;
o `needs.<job>.results` do GitHub não existe neste runner) — foi eliminado
estruturalmente pelo OPS62 (mergeado durante a execução): ci-pr.yml e ci.yml
viraram jobs únicos sequenciais, sem matrix/shard, e o status do job `checks`
já é o veredito honesto da cascata.
O que este item entrega: (a) teste-pino no waitForChecks (cenário run 607:
rollup verde + shard vermelho → o CLI LANÇA "Checks falharam" — nunca confia
no rollup sozinho; invariante de defesa válido e que pegaria qualquer matrix
que volte); (b) docs honestas (agent-pr-workflow.mdc, AGENT-OPS.md) com a
narrativa do parent de matrix e da eliminação estrutural pelo OPS62. O script
ci-rollup-check.mjs e os edits de workflow planejados foram descartados no
gate (redundantes pós-OPS62). Unit 2392 verdes; gates locais ok.
ci-pr.yml e ci.yml deixam de ter jobs irmãos: 1 job 'checks' com steps em
série (guards → lint → format → typecheck → knip → cycles → unit → int →
build → e2e). Step vermelho interrompe o job, o run falha e o slot é
liberado — sem API de cancel (Forgejo 9.0.3 não tem), sem watchdog, sem
upgrade do Forgejo. Skips do ci-scope.mjs viram step id 'scope'; e2e em 1
processo com PLAYWRIGHT_WORKERS=4 (substitui 2 shards × 2 workers); 2
services Postgres por run com portas host derivadas do run_id (runs
concorrentes não colidem); contexto 'CI (PR) / checks' preservado (job id
checks — contrato do automerger OPS61); deploy OPS53 gated em needs checks.
Plumbing morto removido (e2eShardConfig/e2e_shards). Docs: AGENT-OPS,
TESTING.md, changelog. Impl: docs/plans/ops62-ci-fail-fast-forgejo-impl.md
(a intenção watchdog+upgrade fica como histórico do gate).
O merge POST faz um push para main; o hook pre-receive do Forgejo rejeita o
push quando o usuário do token não tem write no repo — o GITHUB_TOKEN nativo
age como o usuário sistema do Actions. Por isso nenhum auto-merge funcionou
(o CLI morria no 409; runs 'success' eram leituras de PR já mergeada
manualmente). O workflow do safety net agora injeta o mesmo PAT dos pós-merge.
O Forgejo posta o status do próprio job do safety net como pending durante o
run inteiro: o gate antigo (pending.length === 0) era insatisfazível e o CLI
nunca auto-mergeou nada — todo merge Forgejo até aqui foi manual (merged_by
do dono), inclusive a PR #52 com CI vermelho (narrativa da 'raça' corrigida).
O gate é: rollup CI (PR) / checks presente+verde (cascata inteira assentou),
ignorando o contexto do próprio workflow; mergeable false só após assentar.
- issue-done/plan-issue-ready: env FORGEJO_API_TOKEN (GITHUB_TOKEN 403 em
Issues) + fork guard (checkout de fork roda script do fork com o PAT — RCE)
- forgejo-issue-transition: falha de flip vira exit 1 (job vermelho, nunca
skip silencioso)
- waitForChecks: exige o rollup CI (PR) / checks presente+verde (raça PR #52:
snapshot com jobs não-agendados parecia verde); mergeable=false só conta
como conflito após a cascata assentar; timeouts 30→45min (CLI) e 40→55min
(workflow)
- configure-branch-protection: contexto real CI (PR) / checks* (Forgejo casa
por glob contra 'CI (PR) / checks (pull_request)'; literal 'checks' nunca
casaria), idempotente GET→POST/PATCH, verificação pós-aplicação; decisão
pura em scripts/lib/branch-protection.mjs
- docs honestas (AGENT-OPS, agent-pr-workflow.mdc, CUTOVER-MAIN-ONLY, README)
- testes: raça do rollup/rollup failure/mergeable pré-settle +
branchProtection create/update/noop + glob real
Job deploy em ci.yml (runs-on host na workstation) streama
scripts/deploy-homeserver.sh para o homeserver via SSH: build dos stages
runner+migrator com secrets do teqo-1313.env, push no registry local,
migrate pelo serviço de maintenance antes do rollout, up + smoke. Zero
secrets novos no Forgejo; HEAD guard + flock; rollback automático com
backup do compose; runbook docs/ops/teqo-1313-deploy.md; invariantes
unit do script (bash -n + estrutura).
O check falhava quando >1 PR aberto tocava src/migrations/ + payload-types.ts,
deadlockando AMBOS os PRs (9 de 13 runs de CI falhos desde 11/08). A validação
real da cadeia de migrations já roda em todo PR e na main: pnpm migrate +
test:int (incl. campaignMigrationReconciliation). Pool mantém a serialização
de spawn de issues needs:migration via countOpenSchemaPrs.
Claim extraído para scripts/lib/agent-github.mjs (claimIssue/claimBriefLines/
claimQueueEntry/claimTargetVerdict), agent-claim.mjs vira CLI fino com mesma
saída; branch validada antes do claim (sem claim órfão); hint de falha
pós-claim; guard -B contra reset de commits de sessão; diretiva launch com
prompt citado + tokenização xargs no worktree.sh; --issue sem valor/posicional
morrem com motivo (sem misclaim); unit novos (directive com issueNumber,
claimQueueEntry, claimTargetVerdict); docs/textos (AGENTS.md, /worktree, skill
worktree-next-issue, AGENT-OPS, CHANGELOG).
- work-issue reescrita como fluxo próprio do humano: sem claim/verificações
(contexto via $ARGUMENTS, 1 pergunta de fallback, validação gh issue view,
nunca claim); plano de intenção do body da Issue (contrato extractPlanPath)
- desacoplada de agent-work-issue; Passos 4-6 extraídos para
execution-pipeline.md compartilhado com deltas por ator
- model-selection volta ao cardápio único Cursor (remove model-local e o
mapeamento canônico model:->DeepSeek)
- correlatos: claim brief (metadata consultiva + footer sem stage), pool
prompt, command description, AGENT-OPS, plan-issue, rebase-on-main,
CHANGELOG-AGENTS
Safety net dual com agent:ready: no merge em main, workflow lê Related #N
e promove Issues ainda aguardando plano (canPromotePlanIssue), soft-skip
idempotente. Helper parseRelatedIssueNumbers + pins unitários.
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
Register com --plan nasce blocked (não claimável); pnpm agent:ready
promove após o plano em main. Gate da skill plan-issue endurecido.
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
Promote alone left the staged deployment on *.vercel.app only. Assign the
custom hostname via POST /v2/deployments/{id}/aliases after verifying it is
attached to the project domains list.
Disabling Git auto-deploy is correct for Actions-gated production. The
separate project toggle "Auto-assign Custom Production Domains" must stay
ON — when OFF, `vercel deploy --prod` only aliases *.vercel.app and leaves
the custom domain on an older Current (exactly what we saw after #122).
After each deploy, ensure-production-alias re-enables auto-assign if needed,
promotes the deployment, and verifies pt.jorgesolla.com.br is attached.
Also adds workflow_dispatch "Vercel promote production" for promote-without-rebuild.
checkout/setup-node/pnpm já estavam em v5/v6 (Node 24); o aviso restante
vinha de styfle/cancel-workflow-action@0.12.1 (node20) usado no fail-fast.
Documenta runtimes nativos em AGENT-OPS.md.
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
Add composite action wrapping styfle/cancel-workflow-action@0.12.1 and
invoke it with if: failure() on every parallel job in ci-pr.yml and ci.yml.
Workflow-level actions: write permission enables cancellation. Document
fail-fast behavior in docs/AGENT-OPS.md CI table.
Closes#52
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
agent-pr-workflow was correct but not alwaysApply, so Cloud agents
followed ManagePullRequest's draft default. Make the rule always-on,
document the ManagePullRequest path, and narrow the engineering-audit
"no auto-merge" exception to the audit sign-off PR only.
Remove stage/promote/Neon smoke. PR CI targets main; ci.yml full-suite
verifier deploys via vercel CLI and requeues when HEAD moves. Pool uses
POOL_GITHUB_TOKEN and POOL_BASE_REF=main. Docs/skills/rules retargeted.
Remove interview audio/transcripts/sheets from the tree, add LICENSE
(UNLICENSED), /private/ convention for local-only sensitive docs, and
public-clean CUSTOMER + session summaries for cloud agents.
agent:promote passa a escolher o SHA mais novo à frente de main com
ci-stage success e usa promote/last-green quando o head de stage está
vermelho, destravando o bootstrap manual para main.
Document native Postgres on Cloud (no db:start) and that gate:fast/push need no DB. Print rename_chat title in agent:claim brief; reinforce rename_chat + pnpm push (not raw --no-verify) in work-issue skill.
* feat(ops): auto-promote stage→main e merge em stage sem exigir head
Substitui o promote manual por workflow disparado no ci-stage green,
relaxa strict na proteção de stage (script configure:branch-protection)
e documenta o novo paradigma em AGENT-OPS e work-issue.
* style: format AGENT-OPS and configure-branch-protection for ci-pr
* feat(ops): pnpm push com ensure-repo-deps determinístico
Bootstrap install + Husky antes de todo push (pnpm push e pre-push).
Documenta fluxo canônico em work-issue, agent-pr-workflow e AGENT-OPS.
* fix(ops): pnpm push roda gate no script, sem depender do Husky
Cloud environment.json chama prepare explicitamente após install.
* test(int): isolate advisor portfolio search from shared cairu slug
Parallel int specs were stomping municipality.advisors on the fixed
cairu row; allocate per-test municipalities instead.
* feat(agent-pool): fundacao read-only do orquestrador (fase A)
- docs/plans/agent-pool-orchestrator.md: plano completo (14 secoes)
- buildClaimQueue extraido de agent-claim para agent-github (comportamento identico, pinado)
- agent-pool-eligibility: isAutonomousClaimable puro e fail-closed + buildPoolQueue
- agent-pool-state: parsePoolConfig, marcadores pool-worker, classifyPoolClaim, spawn plan, auto-stop
- agent-pool-github: leitura de repo variables (degrada nao-estrito), claims/PRs/migration-lock
- pnpm agent:pool -- status|tick --dry-run (inspecao local; live chega na fase B)
- 34 specs unit novos (992/992 verde); lint/tsc/knip/cycles/format verdes
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
* feat(agent-pool): spawn remoto de workers via Cursor API (fase B)
- agent-pool-cursor: client fino da Cloud Agents API v1 (fetch injetavel; CursorApiError com status/body)
- agent-pool-models: resolvePoolModel contra /v1/models (id/alias, sufixo -fast -> param, fallback composer-2.5 com warn)
- agent-pool-prompt: contrato do worker (claim pre-feito, pula passo 1; PR base stage + Closes #N; promote proibido)
- claim coordenado no tick: lock otimista + flip + marcador; agentId idempotente bc-<uuid>; 409 recupera; rollback em falha de spawn
- agent:pool workflow --action start|stop|pause|resume|tick|status (variables POOL_*); doctor; wrappers gh workflow run
- .github/workflows/agent-pool.yml (dispatch-only; schedule/triggers na fase C)
- 19 specs novos (1011/1011); lint/tsc/knip/cycles/format verdes
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
* feat(agent-pool): ciclo de vida do supervisor (fase C)
- derivePoolClaims com status real dos runs (getCursorRun), heal de spawn via agentId deterministico, cancel de duplicatas
- reconcilePoolClaims: falha terminal -> comentario + blocked + archive; merge concluido -> slot livre + archive
- auto-stop real: fila drenada + zero ativos -> POOL_ENABLED=false (exclusoes migration-busy nao desligam)
- job summary markdown (GITHUB_STEP_SUMMARY) — superficie remota de status
- workflow: schedule */10min + pull_request closed em stage (ativam na default branch apos promote)
- 2 specs novos (1013/1013); lint/tsc/knip/cycles/format verdes
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
* docs(agent-pool): runbook operacional e adendo do spike (fase D)
- skill agent-pool: runbook humano (pre-requisitos, start/stop/pause remoto, ciclo de vida, triage, smoke de aceite)
- AGENT-OPS: linha na tabela de comandos + secao Agent pool (canal, secrets, ativacao pos-promote)
- CHANGELOG-AGENTS: entrada curta da entrega
- engineering-audit: pause/resume do pool no protocolo do modo solitario
- plano secao 14.1: evidencias do spike 2026-07-30 (environment repo-file confirmado, id bc-<uuid>, degradacao 403, pendencias do primeiro run com CURSOR_API_KEY)
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
* fix(test): teardown flake no homeSearchUi.unit.spec (dono do CI)
O scheduler do react (prefetch do next/link) podia disparar apos o teardown
do jsdom e derrubar o job unit com 'window is not defined' (run 30587862940).
afterEach(cleanup) — padrao dos 14 specs irmaos.
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Co-authored-by: Francisco Solla <fsolla@users.noreply.github.com>
* fix(cloud): bootstrap Postgres nativo no Cursor Cloud
VMs Cloud não têm Docker (sem daemon/systemd; overlayfs falha em
docker compose). Substitui compose por ensure-postgres.sh (PG 16 via
apt, pg_ctlcluster) e corrige placeholders do .env (BLOB token no
formato vercel_blob_rw_* exigido pelo adapter).
Testado: cloud-setup.sh completa migrate + seed em teqo e teqo_test.
* fix(ci): define GH_REPO no job migration-lock
gh pr list falha sem checkout quando não há repositório git no runner
(fatal: not a git repository). GH_REPO aponta explicitamente para o repo.
---------
Co-authored-by: Cursor Agent <cursoragent@cursor.com>
Parallel ci-pr jobs no longer published the legacy `checks` context
required by stage/main branch protection, blocking auto-merge with no
link in the UI. Add a needs-based rollup that gates on all jobs green.
Root cause of the prod-build React #130 on /campanha/municipios: the
md:hidden mobile-cards tree still hydrated ~25 rows x ~5 sheet controls
(~125 eager Drawer/DrawerCloseButton instances) even on wide viewports.
Filtered ?q=... routes failed more often only because they exercise the
list with filters active — the search param itself was not the bug.
CampaignListSheetProvider hosts a single bottom Drawer for the whole
mobile section; CampaignCellEditOverlay sheet cells portal their
body/footer into it while keeping per-cell open state and autosave
scaffolds. Footers with DrawerCloseButton are only created once a sheet
actually opens. Desktop popover path is unchanged; per-cell Drawer
remains the fallback for surfaces without the provider.
Also: DrawerCloseButton (function render prop, avoids base-ui element
render in prod bundles) replaces the DrawerClose render={<Button/>}
pattern across call sites; stale e2e selectors refreshed; e2e affected
script + playwright github/list reporters for CI; use-mobile reverted
(no consumers of the gate remain).
Guardrail filed as issue #52.
- ci-pr.yml: split the serial checks job into parallel lint/format/
typecheck/knip/cycles + unit/int (vitest --changed vs PR base via
scripts/test-affected.mjs, full fallback on high-risk paths) + e2e
(manifest path->spec via scripts/e2e-affected.mjs, blocking) + build;
migration-lock gains checkout + pull-requests:read (fixes the failures
that blocked PR #50)
- ci.yml (main): full safety net — seed:minimal + int full + build + e2e
- package.json: gate:fast / gate:push; .husky/pre-push runs gate:push
- docs: "Dono do PR, dono do CI" policy in AGENT-OPS + work-issue +
engineering-standards; changelog entry; plans for OPS3-OPS7
