feat: integrate CrewAI automation workspace

.env.example

@@ -2,3 +2,8 @@
 CHAT_API_URL=http://127.0.0.1:8000
 # The backend LLM request has a 60-second deadline.
 CHAT_API_TIMEOUT_MS=70000
+
+# Optional, server-only CrewAI service. Leave unset until it is available.
+# CREWAI_API_URL=http://127.0.0.1:8001
+# Flows may take longer than chat. A timeout does not cancel a running flow.
+CREWAI_API_TIMEOUT_MS=120000

.github/workflows/ci.yaml

@@ -28,8 +28,9 @@ jobs:
       - run: npm run format:check
       - run: npm run build
       - run: npm audit --audit-level=high
-      - run: npx playwright install --with-deps chromium
+      - run: npx playwright install --with-deps
       - run: npm run test:e2e
+      - run: npm run test:integration:automation
       - uses: actions/upload-artifact@v4
         if: failure()
         with:

AGENTS.md

@@ -15,6 +15,7 @@ the backend limitations and the verification matrix.
 - `src/lib/api/`: backend schemas and one browser client. Validate JSON at this boundary.
 - `src/lib/async/`: shared async-state decisions. A failed request never becomes empty or zero.
 - `src/app/api/backend/`: fixed-upstream, allowlisted proxy; no arbitrary URLs, cookies or credential forwarding.
+- `src/app/api/automation/`: separate fixed CrewAI upstream using the same transport; retain independent route/body validation.
 - LLM calls, storage, model/provider selection and secrets belong to FastAPI.
 
 ## Visual and interaction rules
@@ -33,6 +34,7 @@ the backend limitations and the verification matrix.
 - Named component exports and explicit props; use `cn`, composition and primitive variants.
 - Backend errors use FastAPI `detail`, including 422 arrays. Validate the actual response contract.
 - No automatic retry of POST: the backend has no idempotency contract.
+- Automation runs require fresh explicit acknowledgement after parameter edits and each submission. Timeout/navigation does not cancel upstream work. Never run real flows as an implementation smoke test.
 - No fake streaming, persistent-chat claims, model settings, authentication or writable behavior settings.
   Check the actual backend contract before adding a control.
 

README.md

@@ -2,6 +2,8 @@
 
 Отдельный UI для [app-poly-gitops-fastapi](https://github.com/justgithubaccount/app-poly-gitops-fastapi):
 проекты, быстрый чат, чат проекта, просмотр истории запросов и диагностика API.
+Раздел «Автоматизация» подключает отдельный
+[CrewAI backend](https://github.com/justgithubaccount/app-poly-gitops-crewai).
 Семантические дизайн-токены, общие компоненты, light/dark/system,
 адаптивная раскладка и явные состояния данных.
 
@@ -22,6 +24,35 @@ npm run dev
 ограниченный `/api/backend/...` proxy. `CHAT_API_TIMEOUT_MS` по умолчанию 70000.
 LLM-провайдер, URL модели и ключ настраиваются в FastAPI, а не в браузере.
 
+## Автоматизация / CrewAI
+
+На `/automation` доступны список flows и агентов, проверка подключения,
+ручной запуск выбранного flow или серверного варианта по умолчанию.
+Параметры: namespace, необязательное имя приложения и JSON-контекст.
+Результат содержит итоговый ответ, шаги с агентами/задачами и длительностью,
+метаданные и trace ID, если он пришёл от сервиса.
+
+Настройте **на сервере UI**, когда CrewAI запущен:
+
+```dotenv
+CREWAI_API_URL=http://127.0.0.1:8001
+CREWAI_API_TIMEOUT_MS=120000
+```
+
+URL не попадает в браузер; используется ограниченный `/api/automation/...`
+proxy. Если CrewAI не подключён, интерфейс покажет ошибку подключения,
+а проекты и чат продолжат работать независимо.
+
+Каждый запуск требует подтверждения: flow может обращаться к платной модели,
+менять инфраструктуру или создавать issues. Подтверждение сбрасывается при
+изменении параметров и после отправки. Контекст не может переопределять
+`namespace` и `app_name`; секреты остаются в CrewAI, не в форме.
+
+API синхронный: шаги приходят после завершения. Истории запусков, живого
+прогресса и отмены в нём нет. Уход со страницы или таймаут UI не отменяют
+серверное выполнение; автоматического повтора POST нет. Результат на экране
+не сохраняется после ухода/перезагрузки. Дедлайн proxy — от 1000 до 300000 мс.
+
 ## Проверки
 
 ```bash
@@ -30,6 +61,7 @@ npm run format:check
 npm run build
 npx playwright install chromium
 npm run test:e2e       # браузерные тесты с явно заданными API fixtures
+npm run test:integration:automation # настоящий proxy + отдельный loopback HTTP fixture
 ```
 
 Browser-тесты запускают production build на 3100. Бэкенд и провайдер для них
@@ -56,7 +88,7 @@ FastAPI и proxy, но подменяет только провайдера, п
 надёжная классификация таких ошибок требует изменения API.
 
 Настройки UI сейчас — тема и диагностика. API изменения модели/провайдера,
-streaming, отдельные диалоги, auth и действия CrewAI не добавлены этим setup.
+streaming, отдельные диалоги и auth не добавлены этим setup.
 `/health` проверяет доступность сервиса, а не провайдера модели.
 
 ## Проектные правила
@@ -68,3 +100,5 @@ streaming, отдельные диалоги, auth и действия CrewAI н
 
 Репозитории backend, CrewAI, Helm, GitOps и provisioning остаются отдельными
 частями стека. Этот репозиторий не поднимает кластер и не меняет их настройки.
+Публичность исходников не означает готовность публичного runtime: у текущих
+сервисов нет пользовательской авторизации. Локальный UI слушает loopback.

docs/automation-validation.md

@@ -0,0 +1,48 @@
+# Automation validation — 2026-10-29
+
+## Scope and evidence
+
+UI integration of the existing CrewAI HTTP contract. Only the UI repository
+changed; no real CrewAI flow, cluster operation or paid LLM call was executed.
+
+| Check                                                           | Result                                                               |
+| --------------------------------------------------------------- | -------------------------------------------------------------------- |
+| Clean pinned `npm ci`                                           | PASS; existing baseline 11 unit tests pass                           |
+| `npm run check`                                                 | PASS: strict TypeScript, zero-warning ESLint, 21 unit/contract tests |
+| `npm run build`                                                 | PASS: production routes including `/automation` and its proxy        |
+| `npm run format:check`                                          | PASS                                                                 |
+| `npm audit --audit-level=high`                                  | PASS: zero vulnerabilities                                           |
+| `PLAYWRIGHT_CHANNEL=chrome npm run test:e2e`                    | PASS: 19 scenarios, no retries                                       |
+| `PLAYWRIGHT_CHANNEL=chrome npm run test:integration:automation` | PASS: real Next proxy with an explicit loopback HTTP fixture         |
+
+Tests were observed failing before implementation: the new contract modules
+and the Automation workspace were missing on the baseline.
+
+## Covered behavior
+
+- Independent flows/agents/health discovery; error versus empty; stale catalog
+  remains visible after failed refresh; missing CrewAI does not break projects.
+- Selected-flow and server-default execution; namespace/app name/JSON context;
+  reserved target overrides and malformed/non-object JSON rejected.
+- Acknowledgement is mandatory, resets after input changes/submission and does
+  not automatically re-arm after failures. Pending prevents duplicate submits.
+- Returned final answer, step output, agents/tasks, duration and trace metadata.
+  No fake live steps, run persistence or cancellation claim.
+- Timeout preserves form parameters, warns that execution may continue and
+  never automatically repeats POST.
+- Real Next route forwards only validated body to the fixed destination, does
+  not forward browser cookies, and rejects cross-origin/target-override writes.
+- Desktop/mobile screenshots were opened and inspected at 1440px and 390px.
+  Long names, output and metadata fit at 320/390/1440px; navigation wraps.
+
+Local browser verification uses installed Chrome because the Browser plugin is
+not available. CI installs Playwright Chromium. Fixture browser evidence proves
+the UI contract, while the separate loopback check additionally proves the actual
+proxy hop; neither proves real CrewAI execution or cluster/LLM readiness.
+
+## Preview and runtime boundary
+
+The local preview can show the new workspace with CrewAI unconfigured. It must
+show a real connection error and disabled execution, never substitute fixture
+data. To connect an actual service, set server-only `CREWAI_API_URL` and configure
+that service's keys/tools/access independently. No cloud deployment is included.

docs/superpowers/plans/2026-10-29-automation.md

@@ -0,0 +1,74 @@
+# Automation Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Expose the existing CrewAI API through the shared UI.
+
+**Architecture:** Separate `/automation` feature and `/api/automation` proxy.
+Reuse shared transport, error validation and semantic components; preserve Chat API behavior.
+
+**Tech Stack:** Existing Next.js, React, TypeScript, Zod, SWR, Vitest and Playwright; no new dependencies.
+
+**Spec:** `docs/superpowers/specs/2026-10-29-automation-design.md`
+
+## Global Constraints
+
+- Mutation target: `justgithubaccount/app-poly-gitops-ui` only; delivery through a CI-verified PR, no merge/deployment.
+- No backend mutation, real flow run, paid provider call, browser credential forwarding or proprietary references.
+- Manual confirmed execution only; timeout never means cancellation; results are not persistent history.
+- `CREWAI_API_URL` is server-only and unset by default; deadline defaults to 120000ms with range 1000–300000ms.
+- Preserve existing Chat API behavior; no new dependencies; responsive at 320/390/1440px.
+
+## Review Focus
+
+- Context overriding visible namespace/app name must be rejected (Task 1).
+- A timeout may leave upstream work running; never auto-retry or imply cancellation (Tasks 1/2).
+- Missing/unavailable CrewAI must not disable existing chat/projects (Tasks 1/2).
+- Long flow names, outputs and metadata must not overflow mobile (Task 2).
+- Stale catalog refresh must not clear valid data or masquerade as empty (Task 2).
+
+### Task 1: Validated service boundary
+
+**Files:**
+
+- Create: `src/lib/api/automation-schemas.ts`, `automation-proxy.ts`, `upstream-proxy.ts` and boundary tests.
+- Modify: `src/lib/api/client.ts`, `proxy.ts`, `.env.example`.
+- Create: `src/app/api/automation/[...path]/route.ts`.
+
+**Interfaces:**
+
+- Produces: `automationRequest<T>(path, schema, init?)`, `flowsSchema`, `agentsSchema`, `runRequestSchema`, `runResponseSchema`, `parseRunContext(text)`.
+- Produces: `automationPath(method, segments): string | null`, `proxyAutomationRequest(request, segments, fetcher?, env?): Promise<Response>`.
+- Preserves: `apiRequest`, `backendPath`, `proxyBackendRequest` and their existing signatures/defaults.
+
+- [ ] Write failing tests for exact routes, JSON/target validation, response validation, fixed destination, no forwarded credentials, missing config, origin guard, timeout/no retry and separate service client paths.
+- [ ] Run `npm test`; expected new tests fail on missing boundary behavior.
+- [ ] Implement strict schemas, shared proxy/client transport and independent automation route/configuration.
+- [ ] Run `npm test`; expected all tests pass, including existing Chat API guards.
+- [ ] Commit the validated boundary.
+
+### Task 2: Automation workspace
+
+**Files:**
+
+- Create: `src/features/automation/{automation-workspace,automation-catalog,run-form,run-result}.tsx`.
+- Create: `src/app/automation/page.tsx`, `tests/browser/automation.spec.ts`.
+- Modify: shell/icons, existing responsive browser test, README, project instructions and foundation documentation.
+
+**Interfaces:**
+
+- Consumes: Task 1 schemas/client and shared `resolveAsyncView`, layout/data-state components.
+- Produces: `AutomationWorkspace`, independent catalog regions, validated manual form and `RunResult`.
+
+- [ ] Write failing browser tests for discovery/empty/error states, acknowledgement, named/default run, JSON validation, pending/result/error handling, preserved drafts, stale catalog, no automatic POST retry, and long content at mobile widths.
+- [ ] Run the new browser test against a clean baseline server; expected missing navigation/workspace assertions fail.
+- [ ] Implement the workspace, navigation and documentation. Use native labelled select/checkbox inside existing shared form patterns; no unneeded primitive dependency.
+- [ ] Run check/build/browser suite and formatting/audit; expected all pass without regressions.
+- [ ] Exercise the real proxy against an explicit local HTTP fixture, inspect desktop/mobile rendering and document the evidence limits.
+- [ ] Commit the workspace and verification record.
+
+### Delivery
+
+- [ ] Review the whole diff and obtain a fresh-context review; resolve important findings with regression tests.
+- [ ] Push `feat/crewai-automation`, create a PR and verify CI for the exact head.
+- [ ] Keep the user's preview available with honest missing-service state; do not merge or deploy.

docs/superpowers/specs/2026-10-29-automation-design.md

@@ -0,0 +1,57 @@
+# Automation workspace
+
+## Intent and approved scope
+
+Add the existing CrewAI API to this UI as a separate Automation workspace.
+The operator approved the existing-API proposal: flow/agent discovery, explicit
+manual execution and inspection of returned results. Backend services remain
+separate. Only `justgithubaccount/app-poly-gitops-ui` is a mutation target.
+
+## Approaches
+
+1. **Selected:** a separate `/automation` workspace and fixed-upstream proxy.
+   Reuses the application shell, semantic tokens, read states and API validation.
+2. Merge the services or introduce a gateway: unnecessary backend changes.
+3. Add an independent frontend: duplicates setup and fragments the workspace.
+
+## API and interface
+
+Verified source: `app-poly-gitops-crewai@0835d3102555e6cef4fb700c56969ecddc6f6ddd`,
+`src/auto_k8s_pilot/api.py` and `flow_runner.py`.
+
+- `GET /health`, `/flows`, `/agents`: independent regions with loading, empty,
+  error, refresh and stale states. An unavailable CrewAI must not break chat.
+- `POST /run` and `/run/{flow_name}`: select the server default or a discovered
+  flow; enter namespace, optional app name and optional JSON object context.
+- Before each run require an explicit acknowledgement that flows may call an
+  LLM, change infrastructure or create issues. Never start on load or retry POST.
+- Show pending state while waiting, then final answer, ordered step outputs,
+  agents/tasks, durations, metadata and trace ID if supplied. Results are only
+  retained in the mounted workspace. They are not a persisted run history.
+- Failures preserve the form. Timeout/disconnect means execution may continue;
+  another run requires fresh acknowledgement. No cancellation/progress promises.
+- Reject context keys `namespace` and `app_name`: the API merges context last,
+  so accepting them would silently override the visible target fields.
+- Mobile navigation must wrap and all content must fit at 320/390/1440px.
+
+## Boundary and configuration
+
+Server-only `CREWAI_API_URL` (unset by default) and
+`CREWAI_API_TIMEOUT_MS=120000`, allowed range 1000–300000ms. Browser calls only
+`/api/automation/...`. Preserve the existing Chat API configuration/defaults.
+Share transport/error handling; maintain separate route/body allowlists.
+No arbitrary URLs/paths, browser credentials, upstream redirects or automatic
+execution. Unsupported routes/methods, invalid names, non-object JSON context,
+cross-origin writes and excessive payloads are rejected before upstream calls.
+
+## Non-goals and validation
+
+No backend mutation, deployment, paid LLM call or real flow execution is part of
+implementation verification. No run persistence, live streaming, cancellation,
+editable agents/flows, writable provider settings or authentication is invented.
+Source-public is not runtime-public: the current services have no user boundary.
+
+Verify unit contracts/proxy guards and browser discovery, confirmed execution,
+result inspection, errors, no auto-retry, missing configuration, stale lists and
+responsive layout. Also exercise the real UI proxy against a loopback HTTP
+fixture; keep fixture evidence separate from real CrewAI execution.