Drop the NOT NULL constraint on allocation_decision.rationale to align
the DB schema with the Payload config (required: false since B134).
Hand-written idempotent migration following the consent_text_to_jsonb
precedent. Adds idempotency test in campaignMigrationReconciliation.
Closes#371
- buildClaimQueue exclui Issues ready sem frontmatter id (não viram worktree
por construção) e withoutIdReadyIssues() expõe as puladas
- agent:claim / worktree next / pnpm issue next avisam com os números em vez
de morrer — uma Issue quebrada não trava mais a fila
- pin novo em agentPoolEligibility.unit.spec.ts
Closes#864
Global (~/.config/opencode/opencode.jsonc) fica só com jina+playwright (remove github/postgres/forgejo/stalwart/office-word/jobspy) — manual como OPS89.
Repo Teqo opencode.json: penpot+postgres com permission gated (penpot*/postgres*/query deny no build, allow só nos subagentes penpot/postgres).
Novos .opencode/agent/penpot.md + postgres.md (mode subagent, docs de limite teqo vs teqo_wt*).
Re-homed: jobspy→jobs, forgejo+office-word→lei-mercados-digitais, forgejo+stalwart→infra-solla (remove postgres pin, AGENTS.md sem office-word), forgejo→iara-pwa (limpa pins postgres/stalwart).
Prosa: AGENTS-infra.md, prompt-montagem-home.md, project-status/engineering-audit sem MCP/pnpm issue.
Validação: opencode debug config nos 5 repos (Teqo 4, jobs 3, lei 6, infra 7, iara 3 MCPs), gh auth ok, gate:fast verde.
Mapa fixo commitado em scripts/lib/worktree.mjs (cheap=cheapestinference/deepseek-v4-flash, pro=opencode-go/qwen3.7-max, zen=opencode-go/ox-alpha-free, go=opencode-go/mimo-v2.5, alibaba=alibaba-token-plan/qwen3.7-max) com WORKTREE_VARIANT=max sempre; resolveWorktreeModel fail-high em múltiplas flags; opencodeLaunchDirective passa a emitir --model <map|preset> --variant max --auto; scripts/worktree.mjs valida at-most-one antes do claim e repassa modelo nos 3 propósitos (next/plan/new) com help/docblock atualizados; .agents/shell/worktree.sh, .opencode/commands/worktree.md, .agents/skills/worktree-next-issue/SKILL.md e local-database skill com contrato remapeado (--go deixa de ser no-op); tests/unit/worktree.unit.spec.ts pina as 5 flags, variant e guardrail.
Closes#859
Hoist collectUnmappedSrc helper in selectE2eSpecs so unmapped risk-area
files are computed before the high-risk early-return. When high-risk wins
and risk files lack a manifest entry, the curated result now includes
them in unmapped and appends '; curated + risk files without mapping: ...'
to reason — mode stays curated.
Pin: testAffected gains combined case (migration + unmapped risk file →
curated with unmapped in reason). Closes#846.
O primeiro deploy pós-fix faz backup do compose ainda bare e o cleanup novo
remove os aliases bare em seguida: um rollback manual restaurando esse backup
legado precisaria recriar o alias (docker tag) após o pull. Documenta a
janela de transição na seção de rollback e na linha INF13 da tabela de
falhas.
O compose do stack referenciava a tag local bare teqo-1313:<sha>, que é uma
segunda ref que nada recria se perdida (achado INF8 24/08: só restava a ref
do registry e o 'compose up -d' falhou com No such image). Agora:
- swap escreve sempre localhost:5000/teqo-1313(-migrator):<sha> (sed
idempotente aceita ambas as formas); os docker tag bare somem (sem
consumidor);
- rollback do runbook simplifica (docker pull basta, sem dance de tag);
- limpeza pós-deploy (INF3/F2) preserva TODAS as tags da imagem em uso pelo
IMAGE ID do container, imune a drift de label do compose.
Runbook docs/ops/teqo-1313-deploy.md atualizado (fluxo, rollback, tabela de
falhas com o achado). Issue de referência (tracker infra-solla):
git.solla.dev/fsolla/infra-solla#24
- opencode.json: removido provider.vercel inteiro — repo fica $schema + mcp.penpot
- docblock de OPENCODE_PRESET_MODEL e prosa de worktree-next-issue/SKILL.md apontam para ~/.config/opencode/opencode.jsonc
- Forgejo → GitHub na linha reescrita da skill (rastro pós-cutover)
- impl plan + changelog da entrega; débito OPS91 (#841) registrado com plano
The 'advisor scopes municipalities' journey casts many steps; the dossiê
leadership-list chunk can land after goto settles under 4-worker load,
tripping the leadership-link visibility assert. Adds the C106 S: settle
before the assert and scopes with .first() against transient duplicates.
The B196 'relation groups' spec asserted the '1 município encontrado'
completion footer with the 10s expect default; under 4-worker CI load the
RSC list chunk can land late ('element(s) not found'), the same
stream-latency class as the issue's three deploy-blocking assertions.
Polls the footer with the same 30s budget the sibling geometry asserts
use, scoped to .first() so a transient streaming duplicate cannot trip
strict mode.
The FAB in campaignPermissionProfile:85/:101 (carteira/tudo write present)
and the advisor checkbox in campaignMunicipalities:115 (editar page) failed
consistently under 4-worker CI load because the RSC streaming shell's
S:* copies hadn't settled when networkidle resolved, and the default
expect/action timeout wasn't sufficient on the constrained hosted runner.
A probe proved the FAB SSR-renders correctly for carteira/tudo advisors
(profile scope is honored; no seed false-negative). The fix adds the
C106 shell-settle (document.querySelectorAll('div[id^=S:]').length === 0)
before the presence assertions, matching the established idiom used by
the sibling campaign spec tests.
21/21 passed at E2E_PROD=1 PLAYWRIGHT_WORKERS=4 repeat-each=3.
gate:fast (2439 unit tests) all green.
github-api.mjs lê GITHUB_TOKEN (não GH_TOKEN); os dois workflows de flip
pós-merge (issue-done e plan-issue-ready) só definiam GH_TOKEN, então o
script falhava em autenticar ('Sem token ... defina GITHUB_TOKEN') e o
flip/promoção silenciava: o job saía success mas a Issue ficava presa
(PR #818 / #817 e PR #820 / #813). Alinha com agent-pr-ready-automerge.yml
que já passava GITHUB_TOKEN explicitamente.
