GHSA-wvxc-jp3v-5mg5: `DELETE /api/v1/batches/delete-completed` deleted the
completed batches of EVERY api key on the instance and nulled the contents of
every file those batches referenced. Any ordinary inference key reached it —
including one with `scopes: []` — and no victim key, batch id or file id was
needed.
Two defects stacked in one endpoint:
- `deleteCompletedBatches()` carried no `api_key_id` predicate. The file
SELECT, the checkpoint DELETE and the batch DELETE were all instance-wide.
- The route only checked that SOME key was present (`!scope.apiKeyId` → 401),
never that the caller owned anything, and called the helper bare.
The helper now takes `apiKeyId` and scopes all three statements to it; the route
passes the caller's key and omits it only for session auth, so the operator's own
dashboard keeps its instance-wide cleanup and an API key clears only its own
completed batches.
None of this is a new pattern. `listBatches(apiKeyId?)` and
`countBatches(apiKeyId?)` in the same module already scope by `api_key_id`, and
`batches/[id]/route.ts` already gates per-record access with `scopeCheck` —
session auth sees everything, a key sees only its own. This one helper was the
one that never got it, which is why the fix reuses the shape instead of
inventing a second convention.
tests/unit/batch-delete-completed-ownership-wvxc.test.ts — 5 tests, 4 red before
the fix, including the two that prove the cross-tenant destruction (another
key's batch survives; another key's file content survives). It also pins the
instance-wide dashboard sweep so the fix cannot be "tightened" into breaking the
operator's own cleanup, and a source guard that the route never calls the helper
bare again.
Reported privately via GHSA-wvxc-jp3v-5mg5.
Closes GHSA-wvxc-jp3v-5mg5
Uma conversa que nunca chegou a parada limpa e não sinaliza nada é o pior estado possível de UI: indistinguível de uma que terminou. O incidente que você cita no comentário do teste — stream pesado em reasoning estourando o cap do coletor no meio, deixando a conversa presa sem sinal — é exatamente o caso que justifica o badge.
Separar `resolveTurnCompletionState` de `resolveConversationStalledState` também está certo: `tool_call_pending` é um estado legítimo em voo, não uma conversa travada.
Revalidei sobre o tip: **29/29**, typecheck:core limpo.
**Nota de integração.** O `tests/unit/responses-continuation-store.test.ts` conflitou com o #12854, que anexa a própria bateria ao mesmo arquivo. Reconstruí o arquivo como append limpo — versão do tip mais o seu bloco de 184 linhas, verificado por `esbuild` antes de rodar. Registro por que importa: na primeira tentativa eu apenas retirei os marcadores de conflito, e isso enfiou os seus testes **dentro** de um objeto literal não terminado do #12854. Compilava como erro de transform, não como conflito — só apareceu ao rodar. Resolver JSON e teste "aditivamente" sem verificar a sintaxe depois é armadilha; ficou a lição.
Validado numa worktree combinada com a onda de dashboard/monitoring desta leva sobre `release/v3.8.51`: typecheck:core limpo, check-api-typecheck OK (289), check-file-size OK após rebaseline, 130/131 nos testes focados — a falha restante é asserção de tempo de parede sob carga, verde 6/6 isolada.
Envenenar a conexão inteira por um 402 de **um** modelo é o erro clássico de granularidade em provider openai-compatible com múltiplos upstreams — derruba modelos que estavam saudáveis. Restringir ao modelo afetado é o comportamento correto, e é a mesma distinção que o guia de resiliência faz entre cooldown de conexão e lockout de modelo.
Validado numa worktree combinada com a onda de dashboard/monitoring desta leva sobre `release/v3.8.51`: typecheck:core limpo, check-api-typecheck OK (289), check-file-size OK após rebaseline, 130/131 nos testes focados — a falha restante é asserção de tempo de parede sob carga, verde 6/6 isolada.
Paginar e completar o stream em `GET /v1/models` é a correção certa para catálogo grande: um payload único que cresce com o número de providers vira timeout silencioso no cliente, não erro.
* feat(api): hydrate memoryHits from the persisted history event
`GET /api/a2a/tasks/[id]` falls back to the persisted history row once a task
leaves the in-memory TTL window, and `reconstituteHistoricalTask` hard-coded
`metadata: {}` — so the drawer's "Memory used" section vanished for any
historical task, even though `executeA2ATaskWithState` had already written a
`memory_hits` event with the hits.
The fallback now reads that event: `data_json` is parsed and, when it yields at
least one well-formed hit, exposed as `metadata.memoryHits`. The event itself is
filtered out of `events` — it is observability, not a state transition, and
without the filter it leaked into the timeline as a duplicate of the row's
current state.
Reading is defensive throughout, mirroring `DrawerMemory`'s own validation: the
payload is caller-influenced and unvalidated end to end, so `JSON.parse` runs
inside `safeJsonParse`, non-arrays are rejected, and each entry must carry `id`,
`key`, `type` and `snippet` as strings (a non-string field would be rendered as
a React child and take the drawer down). Malformed input degrades to
`metadata: {}` and a 200 — never a 500.
Refs #12639
* fix(a2a): bound the memory recall with its own deadline
collectMemoryHits() runs BEFORE the skill handler and had no deadline at all,
so a slow memory backend delayed the start of every A2A task — the HTTP
genericBackend alone defaults to a 30s timeout.
The search now races a MEMORY_RECALL_TIMEOUT_MS (1500ms) deadline. Overshooting
degrades exactly like any other recall failure: empty hits, a warn log, and the
task proceeds normally (best-effort contract unchanged, nothing propagates).
The deadline timer is cleared in a finally on BOTH paths so no handle is left
holding the event loop open, and MemoryHitsDeps.timeoutMs makes it injectable
so the tests cost milliseconds instead of 1.5s of wall clock.
Refs #12639
* fix(dashboard): carry conductor requirements and focus the repeated task
The drawer's "Repeat" for a Conductor task dropped the runner/model pinning and
left the operator staring at the finished run:
- `hubTaskSchema` now parses the hub's `requirements` (`.catch(null)` so an odd
shape never fails the whole task parse), and `ConductorTaskDetail` exposes
`cli`/`model` (`null` when the hub sends none).
- `repeatReqForConductor` carries `cli`/`model` when present and OMITS them
otherwise — the route's Zod takes both as optional strings, so a `null` would
400. The two fields are independent.
- `performAction` reads the response body once and returns it, so the repeat can
report the CANVAS id of the created task (`task_id` / `data.id` /
`result.task.id`, each with its node prefix). `OrchestrationPageClient` then
refetches and focuses it via `?node=`; History keeps its current behavior.
- `conductor-routes-auth.test.ts` covers the creation route through its `ROUTES`
array; the duplicated source assertion left `conductor-create-route.test.ts`.
Refs #12639
* chore(a2a): follow-ups changelog
Changelog fragment for the five items PR-C delivers from #12639.
The sixth item on the issue — an authenticated panel path for A2A task
creation — stays deliberately out of scope and is recorded as such in a
comment on the issue rather than silently dropped: the JSON-RPC endpoint
accepts API keys only, and widening that endpoint's auth surface to serve a
UI convenience is the operator's call, not the implementation's.
Closes#12639
Um pool `auto/*` vazio que não diz por que está vazio é a pior forma de falha: o operador vê ausência e não sabe se é config, cota ou catálogo. Registrar qual estágio removeu quantos, e carregar `dropReason` em cada entrada de `quotaHealth.providers`, transforma silêncio em diagnóstico.
Os três defeitos achados de carona valem tanto quanto a feature — em especial a cota livre recorrente sem teto sendo tratada como desconhecida em vez de segura, que é justamente o caso que mais aparece.
Revalidei após reconstruir sobre o tip: **15/15**, typecheck:core limpo, `check-api-typecheck` OK (289).
**Uma mudança minha no `catalogPaidFilter.ts`.** Mantive a sua extração — ela é mais limpa que o predicado inline — mas troquei o corpo para usar `isFreeForProvider` por id. O módulo OR-eava `providerHasFreeModels(resolved) || providerHasFreeModels(canonical)` num único `freeProvider` e depois testava `isFreeModel` em cada um. Com isso, um id cujo próprio provider não documenta tier livre passa a ler como grátis sempre que o alias irmão documenta — que é exatamente o buraco que o #12744 fechou e já está no tip. Por id preserva a garantia; nenhuma outra linha do módulo mudou.
**Dívida registrada:** o `virtualFactory.ts` foi de 1187 para 1219 somando esta onda e cruzou o teto de 1200 pela primeira vez. Congelado com os candidatos a extração nomeados na justificativa.
Validado numa worktree combinada com a onda de roteamento/free-tier desta leva sobre `release/v3.8.51`: typecheck:core limpo, check-api-typecheck OK (288), check-file-size OK após rebaseline, 84/85 nos testes focados.
Duas cópias da mesma ordem de provider com um "keep in sync" implícito é dívida que cobra juros a cada provider novo. Uma definição com re-export nos dois lados resolve a classe.
O `xao/*` ordenando depois de todo provider conhecido em vez de junto do `xai-oauth` é um sintoma concreto de que a duplicação já estava divergindo.
**Integração:** `scripts/quality/run-all-gates.mjs` conflitou com o `check:pricing-freshness` que entrou pelo #12792 na mesma onda. Aditivo — os dois gates coexistem.
Validado numa worktree combinada com a onda de roteamento/free-tier desta leva sobre `release/v3.8.51`: typecheck:core limpo, check-api-typecheck OK (288), check-file-size OK após rebaseline, 84/85 nos testes focados no runner Node.
Parar de honrar `isFree:true`, `:free` e `0/0` vindos do upstream **antes** de consultar o catálogo é a inversão certa: hoje um provider sem tier livre documentado consegue se declarar grátis e o listing diverge do roteador `auto/*`. Checar as heurísticas depois do hit de catálogo fecha a porta sem quebrar o caminho de linhas custom locais, que continuam confiáveis pelo caminho próprio.
O `isFreeModel("or", …)` com um alias que não existe é o tipo de bug que passa despercebido porque falha silenciosamente para o lado permissivo.
**Integração:** o `decideHidePaid` que o #12795 extraiu passou a usar o seu `isFreeForProvider` por id, em vez de OR-ear os dois aliases num único `freeProvider`. A forma por id é a garantia que esta PR estabelece, então ela prevaleceu.
Um 0% vermelho num install novo não é só feio: é um número contraditório, porque afirma medição onde não houve nenhuma. Ler ausência como "n/a" é a correção certa, e mantê-la display-only no caminho de combo health mantém o escopo honesto.
Registro que gostei: a PR **não** mexe no `/api/health`, e diz por quê — qualquer coisa que aquela rota devolva é pública numa instância exposta. Recusar o escopo adjacente com a razão escrita vale mais que a mudança em si.
Revalidei após mergear a base na branch: `api-health-version-source` + `combo-health-empty-snapshot` 5/5 no runner Node, `combo-health-null-quota` 1/1 no vitest, typecheck:core limpo.
**Integração:** `src/app/api/system/version/route.ts` conflitou com o `restartRunningServer` que entrou na release depois que você cortou a branch. Aditivo — ficaram os dois imports, a sua troca por `APP_CONFIG.version` e o passo de restart do outro PR.
Validado numa worktree combinada com a onda de persistência desta leva sobre `release/v3.8.51`: check-file-size e check-changelog-integrity OK, typecheck:core limpo, check-api-typecheck OK (289), 70 testes focados no runner Node e 1 no vitest, todos verdes.
"Zeros continuam zeros, latências ausentes continuam ausentes, falhas pré-coluna ganham o próprio balde" — a distinção entre ausência e zero é o miolo aqui. Um install novo mostrando 0ms como se tivesse medido é pior que mostrar nada, porque parece dado.
**Uma mudança minha na sua branch: a migration foi renumerada de 174 para 175.** A `174_server_tool_executions.sql` entrou no #12867, mergeado horas antes desta leva, então `174_call_logs_provider_stats_indexes.sql` colidia. Renomeei o arquivo e ajustei o rótulo do teste ("migration 174 creates..." → 175). Confirmei que não sobrou prefixo duplicado em `src/lib/db/migrations/` e revalidei o `call-logs-provider-stats`: 4/4.
Os dois índices compostos são a parte que paga a longo prazo — rollup por provider parando de varrer a tabela.
Validado numa worktree combinada com a onda de persistência desta leva sobre `release/v3.8.51`: check-file-size e check-changelog-integrity OK, typecheck:core limpo, check-api-typecheck OK (289, dentro da baseline), 70 testes focados no runner Node e 1 no vitest, todos verdes.
"Um checkpoint que nunca olhou o próprio resultado" é o tipo de defeito que só aparece quando o WAL fica maior que o banco. Ler a linha de retorno do pragma e diferenciar busy de sucesso é a correção; o retry `PASSIVE` um minuto depois é o que evita esperar as 6 horas do próximo tick.
Gostei da decisão de não persistir contador: streak em memória que zera no stop é o comportamento honesto para uma métrica de contenção.
Os 11 casos cobrem as formas de retorno que o pragma pode assumir — sentinela `-1`, objeto pelado, `undefined`/`null`/`[]` — que é onde esse tipo de parsing costuma quebrar em silêncio.
Validado numa worktree combinada com as 16 PRs desta leva sobre `release/v3.8.51`: typecheck:core limpo, check-file-size e check-changelog-integrity OK, complexity 2788/3218 e cognitive 1261/1437, ESLint 0 erros nos 152 arquivos alterados, 771 testes unitários focados, 49 de integração e a suíte vitest:ui completa (2149) verdes.
O ponto que sustenta a PR é o `models.dev` virar overlay de preço em vez de fonte de catálogo. Um catálogo estático que sobrevive à conta já ter listado ids mais novos é o tipo de defeito que só aparece quando o modelo novo é justamente o que se quer usar.
Nota de integração: `activeSyncedCatalog.ts` colidiu com o #12934 (união dos `customModels` do picker no catálogo de despacho). Como você extraiu o bloco original para `loadConnectionCatalog`, os dois se compõem: a união dos irmãos agy/antigravity primeiro, o `unionCustomModels` por cima. Revalidei com `custom-models-live-catalog-12597`, `live-model-catalog-reconciliation-8926`, `sync-models-degraded-cached-catalog-9683`, `models-dev-catalog-read-gate`, `discovery-class`, `reactive-model-sync` e `l1-oauth-autosync-default` juntos — 48/48 — mais typecheck:core limpo.
Sobre o `autoSync` padrão em Claude/Codex/Copilot com scheduler de 6h: passei isso pelo dono antes de mergear e a decisão foi manter como está.
Validado numa worktree combinada com as 16 PRs desta leva sobre `release/v3.8.51`: typecheck:core limpo, check-file-size e check-changelog-integrity OK, complexity 2788/3218 e cognitive 1261/1437 (ambos sob a baseline), ESLint 0 erros nos 152 arquivos alterados, 771 testes unitários focados, 49 de integração e a suíte vitest:ui completa (2149) verdes.
A decodificação dos campos aninhados 10/20/30 do `GetRemainingResets` ao vivo (último commit) é o que separa isto de um palpite sobre o formato do frame. Mostrar zero em vez de esconder a linha é a escolha certa: crédito zerado é informação, ausência de linha é ambiguidade.
Merged, with one adjustment.
Confirmed the bug end to end: `PricingTab.tsx:369` types the payload as `{ error?: string }` and feeds it to `new Error(errorPayload.error || ...)`, so the `{ message, details }` object landed in the toast as `[object Object]` — exactly what #12494 reported.
The one change I made before merging: `validation.error.message` is the fixed constant `"Invalid request"` (see `validateBody` in `src/shared/validation/helpers.ts:44`), so it would have swapped an unreadable toast for an uninformative one. The repo already has `formatValidationMessage()`, added in #10849 for precisely this case — it returns `"field: reason"` naming the first offending field. Merged with that instead, so a bad pricing value now says which field it was.
Validated on `release/v3.8.51`: `typecheck:core` clean, `check-file-size` OK. Rebased onto the release branch — the PR was cut from `main`, which is ~3695 commits behind the active branch.
Thank you for the report and the fix.
* test(catalog): pin the 2026-09-02 free-tier re-audit facts for gemini, ollama-cloud, groq, nara and mistral
* fix(catalog): re-audit gemini, ollama-cloud, groq, nara and mistral against official pages
* fix(catalog): restore the console-verified Mistral 1B pool and harden its regression test
* docs(free-tiers): move headline to the re-audited ~1.50B and refresh pool counts
* chore(free-tiers): retire stale Groq free-tier text and preset model; fix catalog header
* feat(catalog): eligibilityGate field and gatedRecurringTokens total
* docs(free-tiers): state the evidence-comment rule honestly and retire the last "14.4K RPD" Groq texts
* feat(check): docs-counts gate validates the eligibility-gated free-tier figure
* fix(docs): budget card reads computeFreeModelTotals() instead of regex-parsing the catalog
* docs(free-tiers): retire the stale Gemini onboarding quota text
* feat(radar): carry eligibilityGate through the feed schema, the merge and the summary API
* feat(dashboard): show the eligibility-gated free-tier figure apart from the headline
* docs(free-tier): refresh catalog-entry counts to 442 after base sync
* feat(catalog): ModelScope as the first eligibility-gated pool; document the gated bucket and how we count
* docs(free-tiers): restore README spacing lost in the merge and re-sync the guide counts
* docs(free-tiers): correct the unsummed-catalog comparison figure to the current catalog
* docs(free-tiers): re-sync numbers after merging release/v3.8.51 (Cerebras reclassified upstream)
* docs(free-tiers): re-sync numbers after merging PR1 (Cerebras reclassified upstream)
* fix(docs): keep the NaraRouter plans endpoint out of the API-path checker; rebaseline gateways.ts (+3)
* fix(catalog): keep eligibility-gated rows out of every headline-adjacent figure
The eligibility gate was honored by the steady headline and the pool count, but
three adjacent figures still counted gated rows: the credit reductions feeding
steadyWithRecurringCreditsTokens/firstMonthRealisticTokens, the uncappedProviders
list ("permanently free, no cap"), and the docs gate's free-forever provider set,
which was built from freeType alone.
- computeFreeModelTotals: filter !isGated in the recurring-credit, one-time-credit
and uncapped predicates; gatedProviders semantics unchanged (steady rows only).
- check-docs-counts-sync: exclude eligibility-gated rows from the FOREVER set,
which moves the live free-forever count 53 -> 52 (the base's value). README,
promise-pillars.svg and FREE-TIERS-GUIDE re-synced.
- gen-budget-card-svg: skip gated one-time credits like the totals do, and fail
loudly on `--out` without a path.
- Tests: gated one-time credit does not move firstMonthRealisticTokens; a gated
uncapped row is not in uncappedProviders; shipped gated rows carry no credit
tokens; the committed budget card is byte-identical to a fresh generation.
* test(catalog): allow eligibilityGate in the no-per-row-rating key allowlist
The allowlist landed on the base with #12318, after this branch's field was designed;
eligibilityGate says who may claim a quota, not how much a row can be trusted.
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
Co-authored-by: diegosouzapw <diegosouzapw24@gmail.com>
Validado em lote numa worktree combinada com os 10 PRs desta leva sobre o tip de `release/v3.8.51`: `typecheck:core` limpo, `check-file-size` OK e **241/242** nos 29 arquivos de teste que os PRs tocam.
A única "falha" não é falha: `tests/unit/autoCombo/strict-zero-cost-filter.test.ts` é um teste em estilo Vitest que eu incluí por engano na invocação do runner nativo do Node — ele quebra no import (`@vitest/runner`), não numa asserção. Ao investigar, descobri que esse arquivo não roda em nenhum dos dois runners hoje (o glob do `test:unit` não lista `autoCombo` e o `include` do Vitest só pega `.tsx` nessa pasta); é um problema pré-existente do repositório, sem relação com esta leva, e vou registrá-lo separadamente.
O #12636 conflitava apenas na lista de testes do `@omniroute/opencode-plugin/package.json`, de forma aditiva: o tip já tinha `models-fetcher.test.ts` (do #12607, irmão desta mesma leva) e o #12636 acrescenta `telemetry.test.ts`. Fiz a união dos dois lados (25 arquivos contra 24 de cada) em vez de escolher um, o que teria removido um arquivo da suíte do plugin em silêncio.
Obrigado, @RaviTharuma.
Validado em lote numa worktree combinada com os 10 PRs desta leva sobre o tip de `release/v3.8.51`: `typecheck:core` limpo, `check-file-size` OK e **241/242** nos 29 arquivos de teste que os PRs tocam.
A única "falha" não é falha: `tests/unit/autoCombo/strict-zero-cost-filter.test.ts` é um teste em estilo Vitest que eu incluí por engano na invocação do runner nativo do Node — ele quebra no import (`@vitest/runner`), não numa asserção. Ao investigar, descobri que esse arquivo não roda em nenhum dos dois runners hoje (o glob do `test:unit` não lista `autoCombo` e o `include` do Vitest só pega `.tsx` nessa pasta); é um problema pré-existente do repositório, sem relação com esta leva, e vou registrá-lo separadamente.
O #12636 conflitava apenas na lista de testes do `@omniroute/opencode-plugin/package.json`, de forma aditiva: o tip já tinha `models-fetcher.test.ts` (do #12607, irmão desta mesma leva) e o #12636 acrescenta `telemetry.test.ts`. Fiz a união dos dois lados (25 arquivos contra 24 de cada) em vez de escolher um, o que teria removido um arquivo da suíte do plugin em silêncio.
Obrigado, @RaviTharuma.
Validado em lote numa worktree combinada com os 6 PRs destas duas levas sobre o tip de `release/v3.8.51`: os seis boardaram **sem um único conflito**, `typecheck:core` limpo e **54/54** nos 7 arquivos de teste que trazem.
O drift de `i18n:check` (`docs/security/GUARDRAILS.md`, `STEALTH_GUIDE.md` — source-changed) foi medido também no tip puro e é idêntico: base-red pré-existente, não desta leva.
Validado em lote numa worktree combinada com os 6 PRs destas duas levas sobre o tip de `release/v3.8.51`: os seis boardaram **sem um único conflito**, `typecheck:core` limpo e **54/54** nos 7 arquivos de teste que trazem.
O drift de `i18n:check` (`docs/security/GUARDRAILS.md`, `STEALTH_GUIDE.md` — source-changed) foi medido também no tip puro e é idêntico: base-red pré-existente, não desta leva.
Validado sobre o tip de `release/v3.8.51` depois de reconciliar com o #12620, que entrou primeiro nesta mesma sessão e ataca a mesma classe de problema por outra arquitetura.
**A colisão e como foi resolvida.** O #12620 consertou o GHSA-qv45-56jc-4wmj adicionando `RAW_CREDENTIAL_PATTERNS` a `error.ts` e importando-os em `upstreamErrorPassthrough.ts`. Este PR resolve o mesmo problema quebrando `error.ts` em `errorSanitization.ts` + `errorPathRedaction.ts`. Mantive a divisão em módulos deste PR, porque ao comparar os dois vocabulários o dele já era mais amplo: o `STRONG_CREDENTIAL_TOKEN` daqui cobre `sk-`/`sk_` **com lookbehind e uma variante para a forma embutida** (que pega `sk-proj-…`), mais Slack `xox-`, AWS `AKIA`/`ASIA`, `github_pat_`/`ghp_`/`glpat-` e JWT de três segmentos.
A única forma que o #12620 carregava e este conjunto não tinha era a chave do Google (`AIza…`) — adicionada aqui, com o mesmo quantificador limitado que os irmãos usam (AGENTS.md → PII §1, já que isso roda sobre corpos upstream não confiáveis).
**A verificação não foi por inspeção.** Rodei as suítes do próprio #12620 contra esta estrutura: **48/48** em `error-sanitizer-sk-key-qv45`, `bifrost-relay-response-leak-9m72`, `search-baseurl-client-override-3f8g` e `search-baseurl-ssrf-guard` — incluindo a asserção anti-drift daquela suíte, que é o oráculo certo aqui: *para todo corpo que a camada de passthrough recusa como vazante, o sanitizador de fallback não pode devolvê-lo intacto*. Ela passa, então a propriedade de segurança dos três GHSAs sobrevive à troca de arquitetura.
Os 21 arquivos de teste deste PR: **259/259**. `typecheck:core` limpo.
Validado em worktree combinada sobre o tip de release/v3.8.51: os dois boardaram sem conflito, typecheck:core limpo, check-file-size sem violação nova (as duas restantes — codex.ts e stream.ts — são drift anterior) e 51/51 nos 5 arquivos de teste que os PRs trazem.
Validado em worktree combinada sobre o tip de release/v3.8.51: os dois boardaram sem conflito, typecheck:core limpo, check-file-size sem violação nova (as duas restantes — codex.ts e stream.ts — são drift anterior) e 51/51 nos 5 arquivos de teste que os PRs trazem.
Validado em lote numa worktree combinada com os 9 PRs desta leva sobre o tip de `release/v3.8.51` (já com a leva anterior dentro): os nove boardaram **sem um único conflito**, `typecheck:core` limpo e **80/80** nos 6 arquivos de teste que os PRs trazem.
O crescimento de arquivo próprio da leva foi rebaselinado num registro datado (`_rebaseline_2026_09_03_hartmark_batch`): `combos/page.tsx` 5012→5018 (#12355, tratar o estado degradado quando o bundling de tiktoken de um provider sem relação falha) e `open-sse/services/combo.ts` 4023→4036 (#12338, os fixes do universal-handoff). As violações restantes (`codex.ts`, `stream.ts`) foram medidas também no tip puro e são drift da base, não desta leva.
Obrigado, @hartmark.
Validado em lote numa worktree combinada com os 9 PRs desta leva sobre o tip de `release/v3.8.51` (já com a leva anterior dentro): os nove boardaram **sem um único conflito**, `typecheck:core` limpo e **80/80** nos 6 arquivos de teste que os PRs trazem.
O crescimento de arquivo próprio da leva foi rebaselinado num registro datado (`_rebaseline_2026_09_03_hartmark_batch`): `combos/page.tsx` 5012→5018 (#12355, tratar o estado degradado quando o bundling de tiktoken de um provider sem relação falha) e `open-sse/services/combo.ts` 4023→4036 (#12338, os fixes do universal-handoff). As violações restantes (`codex.ts`, `stream.ts`) foram medidas também no tip puro e são drift da base, não desta leva.
Obrigado, @hartmark.
Validado em lote numa worktree combinada com os 9 PRs desta leva sobre o tip de `release/v3.8.51`: `typecheck:core` limpo, `check:provider-consistency` OK (273 entradas REGISTRY, **356** providers canônicos), `check-docs-counts-sync` exit 0 e **300/300** nos testes que a leva toca.
O crescimento de arquivo que os PRs empilham uns sobre os outros foi rebaselinado num único registro datado (`_rebaseline_2026_09_03_houminxi_batch`), com a decomposição por arquivo: `providers/page.tsx` +18 (import CSV do #12504 + busca do #12495 no mesmo painel), `accountFallback.ts` +6 (o #12566 sobre o rebaseline que o #12590 já registrou — os dois tocam `checkFallbackError`) e `chatCore.ts` +3 (invalidez de cache de quota no 429 do #12325). As violações restantes (`codex.ts`, `stream.ts`) foram medidas também no tip puro e são drift da base, não desta leva.
* feat(api): conductor task creation route (repeat support)
* feat(a2a): record memoryHits consulted per task (observability, 2.7)
* feat(dashboard): repeat action in orchestration drawer (2.6 — repeat only)
* fix(dashboard): require every field each repeat contract needs before enabling the action
* feat(dashboard): memory-used drawer section + fase2 i18n/changelog (2.7)
* fix(dashboard): validate memoryHits shape before rendering + locale wording fixes
* fix(dashboard,a2a): stop memoryHits leaking into repeats, harden drawer guard + status clamp
Final whole-branch review fix wave for the Orchestration Canvas Fase 2 PR-C.
- a2a: `createTask` stores a COPY of `input.metadata` instead of aliasing it, so the
observability `memoryHits` written by `executeA2ATaskWithState` no longer leak into
`task.input.metadata`, into the persisted `a2a_tasks.input_json`, or into the drawer's
"Repeat" body (a repeated task was born carrying the previous run's memory snippets,
even with the `OMNIROUTE_A2A_MEMORY_HITS=0` kill-switch on).
- dashboard: `repeatReqFor` strips `memoryHits` from the a2a repeat metadata, so tasks
persisted before the copy-fix do not propagate them either.
- dashboard: the "Memory used" section now requires all four rendered fields (id, key,
type, snippet) to be strings — `{ id: "x", key: { a: 1 } }` used to throw "Objects are
not valid as a React child" and take the whole drawer down.
- dashboard: an `/a2a` action answered with a JSON-RPC error under HTTP 200 is reported as
a failure (`RPC <code>`, code only — never the upstream message) instead of a success
toast; the secured-deployment rejection keeps surfacing the sanitized `HTTP 400`.
- api: the conductor task-creation route clamps a hub status outside 400-599 to 502, so an
out-of-range status can no longer turn a hub refusal into a `RangeError`.
- dashboard: the History tab's `onActionDone` keeps the drawer mounted (and refreshes the
range) instead of closing it, so the repeat/cancel confirmation is actually visible.
- a2a: documented the recall owner-id limitation — `task.owner` is a SHA-256 key prefix
while memory rows are keyed by the DB api-key id, and no hash-to-id lookup exists today,
so recall only resolves under the keyless posture.
* refactor(dashboard): split drawer repeat helpers and test file under the size/complexity gates
---------
Co-authored-by: Markus Hartung <diegosouzapw@users.noreply.github.com>
* feat(db): a2a task history module over migration-002 tables
* feat(a2a): persist task lifecycle to a2a_tasks with 30d retention purge
* feat(api): a2a task history listing + historical detail fallback
* feat(dashboard): orchestration History tab (Airflow-grid) over persisted runs (2.2)
* fix(dashboard): assert history preset window + loading and time axis in the grid
* chore(dashboard): history i18n + changelog
* fix(dashboard): explicit history purge cascade + keep live drawer off the History tab
* docs: document OMNIROUTE_A2A_HISTORY_RETENTION_DAYS
* refactor(dashboard): split HistoryTab helpers under the complexity ratchet
---------
Co-authored-by: Markus Hartung <diegosouzapw@users.noreply.github.com>
The #11461 × #11513 merge ate the closing '],' + '},' of the maxai entry in
webSessionCredentials.ts — 11 syntax errors (TS1005/1137/1128) on the tip,
which also masked one real TS2322 the MaxAI block introduced in the models
route (providerSpecificData is unknown on the connection; cast to the exact
shape resolveMaxaiCredential already takes, zero runtime change).
API Route Typecheck gate: OK — 289 pre-existing, all baselined. typecheck:core: 0.
GET /v1/models with MODELS_CATALOG_PREFIX_MODE=canonical dropped every chat row of a provider whose registry alias is undefined (antigravity) or equal to its own id (agy, most built-ins). Each emission loop pushes alias/model only when includeAlias, and canonicalProviderId/model only when the ids differ — for a self-aliased provider both are the same string, so neither fired. #11918 fixed the class for custom nodes but not built-ins, and not the static loop. The alias row is now treated as the canonical row whenever the ids coincide, across the static, synced, custom and alias-backed loops; the canonical branch's !== alias guard is untouched, so dual and alias output cannot double up. Docs that described the omission as intended are corrected.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
The Leaderboard rendered apiKeyId.slice(0, 8)… under a column translated as "name". The route now enriches each entry with the key's display name — route-local, so the shared getTopN helper and the federation leaderboard stay id-only — and the page renders name ?? shortId with the full id in a title attribute. The lookup selects only id and name from api_keys, chunked at 200 ids, with unknown ids and blank names omitted; no key material leaves the DB layer.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
The Profile page already rendered a streak card but fed it a hard-coded useState(0) with a "streak data comes from future API" note — while streaks.ts tracked per-key streaks all along and the MCP gamification_profile tool already returned them. GET /api/gamification/level now returns streak: { current, longest } next to level: the key's own streak with apiKeyId, the operator-wide maximum otherwise, matching the aggregate mode getAggregateXp uses (#3484). No new route, no OpenAPI change, no new i18n keys; a missing or zero streak keeps the card hidden exactly as before.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
Four files embedded the U+0000 separator of a memo/group key as a raw NUL byte rather than the \\0 escape the codebase uses for the same idiom elsewhere. The runtime value is identical, but the raw byte trips the binary heuristics of git, GitHub and ripgrep: git diff --numstat reported `- -`, the introducing PRs rendered three of the files as "Binary file not shown", and rg silently skipped them in recursive mode. Rewritten as escapes, with a guard test keeping raw NUL bytes out of src/, open-sse/ and tests/.
Validated in a combined worktree with all 25 PRs of this batch boarded together: typecheck:core clean, 443/443 node-runner tests plus 14/14 vitest across every test file the batch touches, and check-changelog-integrity, check:cycles (418 files), check:provider-consistency (272 REGISTRY entries, 355 canonical providers), check:docs-counts, check:docs-sync (42 locales) and check-file-size all green.
Thanks @pacocartones.
* fix(memory): point the rerank-providers dynamic import at the real db module
#11390 landed with a dynamic import of the localDb barrel, which #12052 had
already removed from the base (and which Hard Rule #2 forbids) — the API
Route Typecheck gate reds on the tip with TS2307. getCachedProviderNodes
lives in src/lib/db/readCache.
* chore(quality): ratchet the api-typecheck baseline down (163 stale entries gone)
Regenerated with --update on a faithful npm ci environment (the .113 box)
against the current tip plus the rerank-providers import fix — the gate now
reads OK at 289 pre-existing errors, all baselined. No new entries added.
MaxAI joins as a first-class signed provider: 13 chat models discovered live from /models/get_config plus 6 image models, routed through the standard /v1 endpoints with per-request X-Authorization signing, browserless onboarding, prompted tool-calling, vision input, image generation and document RAG.
Reconciled on merge — worth reading, because the branch forked 227 commits back and 77 files conflicted. Only five carried MaxAI content; the rest was drift from the older release line and took the tip's side, taking the diff from 113 files to 37 (then 93 as counted against the current base).
- executors/index.ts: the tip has since refactored the executor map to lazy dynamic imports, so MaxAI is registered in that shape rather than the branch's static import.
- imageRegistry.ts: kept only the maxai block. The branch still carried microsoft-designer-web, which #11754 retired.
- models/route.ts: the conflicting hunk was an unrelated Vertex/Anthropic URL change, not MaxAI — tip's side.
- volcengine agent-plan/coding-plan registries: git auto-merged both sides and produced a duplicated supportsVision key, which TypeScript rejects (TS1117). Removed.
One real integration break that only the combined state shows: the MaxAI entry declared no serviceKinds, which #11392 made required a few hours ago. Provider validation threw at load time and check:provider-consistency crashed outright. Declared ["llm"] — the image kinds derive from imageRegistry, per the convention in that PR's backfill.
Every count was measured rather than taken from the branch, and each would have been wrong: reserved prefixes are 402, not the 397 the branch computed from its stale 395 base; providers are 353, not 354. PROVIDER_REFERENCE.md regenerated, the count updated across README/AGENTS.md/llm.txt and its 42 mirrors, package.json and 6 SVGs — every changed line in those files is a digit substitution and nothing else, verified by masking digits and comparing the removed and added sets (90 lines, identical). The executor-map golden snapshot was regenerated: keyCount 133 -> 134.
The branch's file-size-baseline.json predates #12411's ratchet re-tightening, so it was discarded rather than merged — taking it would have silently undone that. The three files this PR grows (proxyFetch.ts +20 for the Windows/firefox_150 TLS profile, imageGeneration.ts +12, models/route.ts +48) were entered against the current baseline under one _rebaseline annotation; no other cap moves.
Verified: typecheck:core clean, check:provider-consistency OK (269 REGISTRY entries, 353 canonical providers), check:docs-counts exit 0, check-file-size OK, check:cycles OK, and 79/79 across the MaxAI suites plus 21/21 reserved-prefix and 2/2 executor-map-golden.
Thanks @arminanton — the provider work itself is thorough; it was the 227 commits of base that needed the attention.
Claude Code (claude / cc) is correctly classified as a flat-rate subscription, so the analytics API reports $0 — accurate as billed cost, and useless as a view of what the subscription actually consumed. Neither the Costs nor the Analytics dashboard had a token-price-equivalent view.
The fix keeps both meanings rather than picking one: ordinary analytics callers keep billed-cost semantics ($0 for flat-rate), /dashboard/costs and /dashboard/analytics opt in explicitly via includeFlatRateEstimates=true, the response reports whether estimates were included so a caller cannot mistake them for vendor billing records, and the figures on /dashboard/costs are labelled as flat-rate estimates rather than presented as spend. Omitted, false and unknown values all retain the existing behaviour.
Scope note carried from the description: this is a checkpoint on #11459, not its full closure — the issue stays open.
Verified in a combined worktree with three sibling PRs of this batch: typecheck:core clean, 134/134 focused tests (4 skipped), and i18n UI coverage PASS across all 42 locales for the 43-file locale pass.
One cross-PR interaction worth recording, since it is invisible from either side: this grows CostOverviewTab.tsx from 1282 to 1318 lines, which is fine against the tip's current 2002 cap but exceeds the 1283 that #12411 (file-size ratchet re-tightening) would freeze. Neither PR fails alone. Merged first on purpose so #12411's mechanical --update recomputes against the real post-merge LOC — the cap still only goes down.
Thanks @xiaoyaner0201 — the opt-in contract plus the "were estimates included" flag is the right shape for this.
On dashboard/memory?tab=engine the Embedding Model quick-select (and the rerank selector) built their lists from a keyword heuristic over the CHAT catalog (AI_MODELS) plus OpenRouter live discovery. Providers whose embedding models are not in that catalog never appeared — mistral, gemini, nvidia nim, groq, vercel-ai-gateway and others that serve embeddings on a standard OpenAI-compatible /embeddings endpoint — and typing such a model by hand failed at runtime with "Unknown embedding provider".
The fix is one generic mechanism rather than a list of per-provider patches: deriveEmbeddingProviderForChatProvider() turns any chat-registry entry with a /chat/completions base into an OpenAI-compatible /embeddings config, with curated EMBEDDING_PROVIDERS entries always winning; the embeddings service resolves a derived config for unknown-but-configured providers instead of rejecting them; deriveRerankProviderForChatProvider() does the same for Cohere-compatible /rerank; and both memory selectors fall back to a free-text provider/model input when no static catalog exists. No provider is special-cased by name, so adding one to the chat registry now makes it embedding- and rerank-capable here automatically.
Verified on the current release tip: merged clean, typecheck:core clean, check:cycles OK across 417 files, and 35/35 across the PR's five new suites (qdrant-quick-select-catalog, memory-provider-listings, rerank-provider-listings, embedding-generic-provider-fallback, rerank-generic-provider-fallback) plus the updated hard-session-lease-bypass-inventory and embeddings-handler.
Note: the base-red disclaimer in the description referenced #9985 against release/v3.8.50 — that window is closed and the current tip carries no open base-red, so nothing was inherited here.
Thanks @rqzbeh — deriving the capability instead of enumerating providers is the version of this that stays correct as the registry grows.
* perf(compression): memory and OOM mitigations for large payload hashing and token estimation
* fix(compression): implement getMemoStats observability for result memo (#7847)
Adds the missing memo observability layer referenced by
tests/unit/compression/oom-memo-memory.test.ts and the monitoring API:
- resultMemo.ts: lifetime hit/miss counters + bounded time-ordered ring
buffer (10k entries, ~90KB) powering 1m/5m/15m/1h hit-rate windows;
getMemoStats() reports size/capacity/hits/misses/hitRate + windows.
- memoLookup() tags served results with stats.memoHit = true.
- clearMemoStore() also resets counters and the ring.
- compression/index.ts re-exports getMemoStats for the monitoring route.
- types.ts: optional memoHit field on CompressionStats.
- New GET /api/monitoring/compression route exposing the stats snapshot
(lightweight, no DB) for operators to track cache-hit efficiency.
* fix(compression): align memo contract with upstream #11727 — return caller object, reset lookup counter in clearMemoStore
* fix(compression): restore unwrapEventEnvelope in stream payload collector summaries
The OOM-mitigation commit accidentally replaced unwrapEventEnvelope(evt.data)
with asRecord(evt.data) in the summary builders and live push, breaking
translate-mode {event, data} envelope unwrapping (clientPayload type detection)
and failing 2 stream-payload-collector tests. Restored upstream semantics;
kept the jsonLength OOM optimization as the only delta in this file.
* refactor(compression): break down writeValue and writeEncodedString to pass complexity ratchets
Refactors jsonSha256 internal helpers (writeValue, writeEncodedString)
into small, single-responsibility sub-functions under the complexity
threshold (max cyclomatic 15, max cognitive 15). Preserves exact
JSON.stringify parity, circular reference guards on both arrays and
plain objects, and escape behavior (all 530 relevant tests pass).
* test(compression): make oom-memo heap assertion robust without expose-gc
The CI unit-test shard runner does not pass --expose-gc, so global.gc is
undefined and heapUsed can still momentarily hold GC-pending transients
(observed 53.4 MiB after a 3MiB body). Gate the retained-heap assertion
on forced collection being available (3 forced cycles for array buffers)
instead of skipping it silently, and keep it fully active when
--expose-gc is present.
* fix(compression): restore worker-pool offload path in runCompressionAsync
The OOM-mitigation refactor dropped the isCompressionWorkerEligible /
runCompressionInWorker dispatch at the top of runCompressionAsync, silently
removing the base's worker-thread offload for eligible large payloads.
Restore the block exactly as on release/v3.8.51, ahead of the result-memo
path, keeping the memoization and hashing improvements intact.
* docs(api): document GET /api/monitoring/compression and log route errors via pino
Add the new monitoring endpoint to docs/openapi.yaml following the
neighboring System entries, and replace the route's console.error with
the repo-standard pino logger.
* fix(skills): regenerate omni-resilience and add changelog fragment
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Co-authored-by: Andrian Balanescu <AndrianBalanescu@users.noreply.github.com>
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Free Provider Rankings sorted by the top model's Arena score, so a provider stayed first even if it failed every call — the reliability column from #11546 already showed what each one actually served, but ordering ignored it.
Adds an opt-in ?sortBy=reliability (API) and a "Most reliable first" toggle (page) sharing one comparator in freeProviderRankingsUsage.ts: measured providers first by successRate desc with ELO on ties, then unmeasured in their incoming order. The default is unchanged and locked by tests. successRate is null below MIN_USAGE_REQUESTS = 5 (existing, never zero), and ordering runs before slice(0, limit) so limit counts in the requested order. The toggle composes with the existing sortTypeFirst/groupByType grouping — a stable sort keeps reliability order within each group.
Opt-in is the right default here: the page is for discovery, including providers never called.
13 tests across three files (6 new for the comparator in isolation, plus filter and route coverage including unknown sortBy → 400 and the default path staying off call_logs).
Verified in a combined batch worktree with all 11 PRs of this batch: 174/174 focused tests, typecheck:core clean, complexity 2706/3218, cognitive-complexity 1221/1437, check:cycles and check:docs-counts green. The 42-locale i18n pass was checked with the CI gates: check-ui-keys-coverage PASS (all 42 locales at or above 65%) and check-ui-value-drift PASS against the release tip.
Thanks @maxmad64bis.
GET /api/free-tier/summary could answer from a Radar overlay built 2026-08-02 while the release ships a catalog curated 2026-08-30 (FREE_CATALOG_CURATED_AT) — totals computed from older data, still tagged catalogSource: radar-overlay. The route now refuses any overlay built before the shipped catalog and falls back to that catalog through the operator's local state.
Tightens #11550 using the generatedAt persisted by #11435.
Verified in a combined batch worktree: 174/174 focused tests across all 11 PRs of this batch (this PR's free-tier-summary-radar-overlay suite included), typecheck:core clean, check-file-size, check-changelog-integrity, check:cycles and check:docs-counts green.
Thanks @maxmad64bis.
* fix(oauth): mark empty Antigravity projectId as degraded (#11284)
The #11284 gate only fired when projectDiscoveryOutcome was set. Paste
credentials, persistOAuthConnection, and agy CLI import could persist
projectId="" as testStatus=active, so the dashboard showed Connected
while fetchAvailableModels returned 403.
Degrade on empty projectId itself. Keep the refresh token stored so
request-time bootstrap can still self-heal.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* fix(oauth): clear stale degrade fields and bind CLI imports to builtin client
persistOAuthConnection left errorCode/lastError on the row when a later
connect discovered a Cloud Code projectId. agy CLI import also kept a
leftover custom: oauthClient marker from dashboard OAuth, so the next
refresh hit the operator web client instead of the public desktop client.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* fix(oauth): null error fields on healthy create paths too
Update already cleared errorCode/lastError* when a projectId appeared.
Create payloads still omitted the keys; match the update shape so a
fresh row cannot keep a leftover degrade marker.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* test(oauth): pin healthy create/upsert nulling of degrade fields
Forge flagged create payloads omitting errorCode/lastError* when a
projectId is present. Production already writes explicit nulls; the
reader strips them via cleanNulls, so pin both the payload shape and
the upsert path that must overwrite a leftover degrade marker.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* fix(oauth): type create payload from AntigravityDegradedProjectState
The persistence helper duplicated a subset of the degrade type and
dropped warning. Align the parameter so the HTTP-only warning field
cannot drift from the exported type.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* fix(oauth): persist degrade status through a single override helper
OAuth exchange/poll-callback spread the whole degrade object, which
wrote warning into the SQLite row and left healthy updates as {}.
Centralize testStatus/errorCode/lastError* so they always win over a
spread tokenData payload.
Signed-off-by: Minxi Hou <houminxi@gmail.com>
---------
Signed-off-by: Minxi Hou <houminxi@gmail.com>
* fix(api): keep registry width and type on embedding models
* docs: changelog fragment for embedding registry fix
* test(api): cover embedding width and type merge
Exercises /v1/models rather than the registry in isolation: a synced
model colliding with an embeddingRegistry entry must keep the width the
registry states, and a synced model the registry names must be typed as
an embedding model.
Fails on catalog.ts before e7fbb62 (2 failures), passes after.
Refs #11759
Honra um `context_length` definido pelo operador em tempo de requisição no roteamento do combo (supersede #12014, que estava incluída nos mesmos commits). Boa cobertura de testes, incluindo o refactor de `resolveComboContextLimit` para módulo próprio. Validado no worktree combinado (13/13). Obrigado!
Intervalo de checagem de saúde de credencial configurável pelo operador, com boa cobertura de testes. Validado no worktree combinado (20/20).
Corrigi o import de `getCachedSettings` em `src/app/api/resilience/route.ts` e `src/lib/credentialHealth/scheduler.ts`, que apontava para `@/lib/db/settings` (path antigo antes do split para `@/lib/db/readCache`, já na tip). Resolvido também um conflito de tradução vi.json entre chaves duplicadas de outra feature (exclusive lease), sem relação com esta PR — mantida a versão já mergeada. Obrigado!
Corrige o bulk-import do Codex apagando `providerSpecificData`/`tokenExpiresAt`/duplicando `priority` de conexões existentes ao fazer upsert num match — mescla o payload importado sobre o estado existente em vez de substituir tudo, igual ao caminho de import single-file já fazia. Findings 4, 6 e 7 do #12113. Teste próprio (224 linhas). Validado no worktree combinado. Obrigado!
Restaura o log do injection-guard nas 13 rotas não-chat (embeddings, images, audio, moderations, etc.) — a correção de log duplicado anterior (#11936) silenciou completamente o único emissor de log dessas rotas, deixando tentativas de injeção sem rastro nenhum em modo warn, e sem log mesmo quando bloqueadas em modo block. Achado de segurança real (Finding 2 do #12113). Teste próprio (141 linhas). Validado no worktree combinado. Obrigado!
Deriva as modalidades do auto-combo a partir do pool de targets efetivo, com teste próprio robusto (174 linhas). Validado no worktree combinado. Obrigado!
Usa a lista de publishers v1beta1 do Model Garden para descoberta de modelos Vertex Anthropic, com teste próprio. Validado no worktree combinado. Obrigado!
Remove a reserva estática global e passa a gatear o roteamento pela ocupação real do lease exclusivo ativo, com boa cobertura de testes (5 arquivos, 54 casos, todos verdes no worktree combinado). Typecheck limpo.
Dois ajustes feitos por cima antes do merge:
1. **26 arquivos `.pyc` órfãos removidos** (`scripts/ops/__pycache__/…`, `tests/unit/ops/__pycache__/…`) — cache compilado do Python sem relação com o fix de lease, provavelmente commitado sem querer do ambiente local.
2. **Conflito em `src/app/api/keys/[id]/route.ts`**: mantida a checagem mais ampla desta PR (`instanceof ApiKeyPolicyInvariantError || código LEASE_KEY_POLICY_INVALID`), que é um superset da versão anterior — cobre o caso original e o novo caminho de erro do lease.
Obrigado pela contribuição!
