Feat/k8s agent sandbox integration - #102
Merged
Merged
Conversation
Key decisions: - Persist APP as the migration default while requiring explicit APP or AGENT selection on supported creates. - Enforce Project Type immutability in ProjectService and reject App creation in Agent Projects in AppService. - Keep Agent Projects free of App-only creation, table, and network controls until Agent creation lands. Files changed: - Added Prisma schema/migration and shared Project Type model. - Updated Project/App services, REST route, Server Action, Project UI, generated model, fixtures, and focused tests. - Moved issue 01 to issues/done. Blockers/notes: - Focused tests and browser/server typechecks pass. - Full suite has 332 passing tests; four Prisma integration suites cannot start because prisma db push returns Schema engine error. - Next build compiles and typechecks, then existing ESLint option mismatch and missing /_document page-data step fail.
Key decisions:
- Renamed app-layer role payloads and session permission shapes to workload-neutral fields while mapping them onto existing App-backed persistence.
- Added transport-independent workload authorization helpers and kept App wrappers delegating to them for backward compatibility.
- Switched App create/delete server actions and role editor UI onto shared workload permission checks without changing existing App behavior.
Files changed:
- Updated shared permission/session models, authorization utilities, user-group service mapping, and project server actions.
- Refreshed role editor wording/state handling for workload-neutral permission payloads.
- Expanded unit coverage for workload-level authorization and moved issue 02 to done.
Blockers/notes:
- Focused authorization tests pass and yarn run v1.22.22
$ /workspace/node_modules/.bin/tsc --noEmit
Done in 2.58s. passes.
- yarn run v1.22.22
$ vitest run
RUN v4.1.4 /workspace
❯ |node-integration| src/__tests__/integration/server/services/app-service.basic-auth.integration.spec.ts (9 tests | 9 skipped) 2509ms
❯ |node-integration| src/__tests__/integration/server/services/project-service.integration.spec.ts (9 tests | 9 skipped) 1770ms
❯ |node-integration| src/__tests__/integration/server/api/v1/api.integration.spec.ts (4 tests | 4 skipped) 1388ms
❯ |node-integration| src/__tests__/integration/server/services/build.service.integration.spec.ts (6 tests | 6 skipped) 9947ms
Test Files 4 failed | 40 passed (44)
Tests 342 passed | 30 skipped (372)
Start at 11:58:44
Duration 33.86s (transform 3.08s, setup 0ms, import 30.33s, tests 36.14s, environment 160.70s)
info Visit https://yarnpkg.com/en/docs/cli/run for documentation about this command. still hits the existing Prisma integration blocker: Prisma schema loaded from prisma/schema.prisma
Datasource "db": SQLite database "data.db" at "file:/workspace/storage/db/data.db"
🚀 Your database is now in sync with your Prisma schema. Done in 28ms -> in four integration suites.
- yarn run v1.22.22
$ next build && tsc --project tsconfig.server.json && tsc-alias -p tsconfig.server.json
▲ Next.js 14.2.35
Creating an optimized production build ...
✓ Compiled successfully
Linting and checking validity of types ...
Collecting page data ...
Generating static pages (0/14) ...
Generating static pages (3/14)
Generating static pages (6/14)
Generating static pages (10/14)
✓ Generating static pages (14/14)
Finalizing page optimization ...
Collecting build traces ...
Route (app) Size First Load JS
┌ ƒ / 5.58 kB 190 kB
├ ƒ /_not-found 880 B 88.5 kB
├ ƒ /api/auth/[...nextauth] 0 B 0 B
├ ƒ /api/build-logs 0 B 0 B
├ ƒ /api/deployment-status 0 B 0 B
├ ƒ /api/init 0 B 0 B
├ ƒ /api/logs-download 0 B 0 B
├ ƒ /api/pod-logs 0 B 0 B
├ ƒ /api/print-schedules-jobs 0 B 0 B
├ ƒ /api/v1/[[...slugs]] 0 B 0 B
├ ƒ /api/v1/webhook/deploy 0 B 0 B
├ ƒ /api/volume-data-download 0 B 0 B
├ ƒ /auth 4.31 kB 167 kB
├ ƒ /backups 2.9 kB 185 kB
├ ƒ /builds 5.53 kB 190 kB
├ ƒ /error 147 B 87.7 kB
├ ƒ /monitoring 109 kB 246 kB
├ ƒ /project/[projectId] 67.8 kB 423 kB
├ ƒ /project/app/[appId] 126 kB 490 kB
├ ƒ /settings/profile 19.4 kB 181 kB
├ ƒ /settings/s3-targets 3.22 kB 213 kB
├ ƒ /settings/server 17.5 kB 222 kB
├ ƒ /settings/users 7.66 kB 217 kB
└ ƒ /unauthorized 147 B 87.7 kB
+ First Load JS shared by all 87.6 kB
├ chunks/2117-92c5f0ef70a3941b.js 31.9 kB
├ chunks/fd9d1056-94cd9f08eae7a31f.js 53.6 kB
└ other shared chunks (total) 2.05 kB
ƒ Middleware 49.4 kB
ƒ (Dynamic) server-rendered on demand
Done in 35.02s. passes, but still prints the existing Next ESLint option mismatch warning during build.
- Add Agent model to Prisma schema with project cascade + gateway restrict - AgentService: create, getAllByProjectId, getById, deleteById with SandboxTemplate/WarmPool reconciliation and DB rollback on K8s failure - AgentSandboxAdapter: reconcile/delete SandboxTemplate (sandbox.quickstack.dev/v1beta1) and SandboxWarmPool CRDs via CustomObjectsApi - Project page: render AgentListClient for Agent Projects, keep existing App behavior - CreateAgentDialog: shadcn Dialog with LLM Gateway Select + dynamic LiteLLM Model Alias Select, Toast.fromAction submit - Agent detail page: /project/agent/[agentId] showing ID, Gateway, Model Alias, Created - Sidebar: show agents for Agent Projects, apps for App Projects, with auth filtering - UserGroup service: include agents alongside apps in projectWorkloads for workload-neutral auth - KubeObjectNameUtils: toAgentId(), CacheTags: agents/agent tags - Unit tests: 13 tests covering agent creation validation, K8s-safe id generation, sandbox reconciliation, rollback, delete - Build: green, Typecheck: clean
Key decisions: - Prisma: add image, cpu/memory requests/limits, systemPrompt, encryptedEnvVars to Agent - agent-config.model.ts: Zod validation for K8s quantities (100m, 128Mi, etc.), reserved env names (QS_ prefix), encrypted env var storage - AgentSandboxAdapter: add hasActiveClaim() for SandboxClaim runtime-lock check - AgentService.saveConfig(): encrypt env vars, reject runtime config changes while running, reconcile SandboxTemplate/WarmPool on save - Agent detail page: tabbed layout (General/Overview) with AgentConfigForm UI - AgentConfigForm: custom image, K8s resource limits, system prompt, encrypted env vars with add/remove - Tests: 17 model validation + 12 service unit tests (encryption, running lock, reconciliation, config persistence) - Build: pre-existing react-hook-form type errors only Blockers/notes: - Runtime lock uses SandboxClaim existence check; naturally returns false until issue 07 implements start - Env var values encrypted at rest with CryptoUtils, not disclosed in UI after save - Gateway/model changes tracked; virtual-key invalidation handled through SandboxTemplate reconcile
Key decisions: - Prisma: add RoleAgentPermission table (parallel to RoleAppPermission) with cascade from Agent + RoleProjectPermission - userGroupService.saveWithPermissions(): query project apps/agents to split workloadPermissions into correct DB tables during save - userGroupService.mapProjectRolePermission(): merge both roleAppPermissions and roleAgentPermissions into unified workloadPermissions array for session - Role editor: include project.agents alongside project.apps in Individual Permissions toggle and init - Agent detail page: compute role via UserGroupUtils.getRolePermissionForProjectWorkload, pass readonly to AgentConfigForm - AgentConfigForm already supports readonly via form disabled prop - Auth wrappers: add agent-specific delegates (getRolePermissionForAgent, ensureReadAgent, ensureWriteAgent, isAuthorizedReadForAgent, etc.) in role.utils.ts, shared-authorization.utils.ts, action-wrapper.utils.ts Files changed: - prisma/schema.prisma: +RoleAgentPermission model, relations on RoleProjectPermission + Agent - src/shared/model/generated-zod/: auto-generated RoleAgentPermissionModel + relation updates - src/server/services/user-group.service.ts: type + query + map + save updates for agent permissions - src/app/settings/users/user-group-edit-overlay.tsx: agents in workloadPermissions init + toggle - src/app/project/agent/[agentId]/page.tsx: compute role, pass to client - src/app/project/agent/[agentId]/agent-detail-client.tsx: accept role, compute readonly - src/shared/utils/role.utils.ts: +7 agent wrapper methods - src/server/utils/shared-authorization.utils.ts: +4 agent guard functions - src/server/utils/action-wrapper.utils.ts: +3 agent action wrappers - Tests: +6 agent-specific role tests (role.utils.unit.spec.ts), +13 agent guard tests (shared-authorization.utils.unit.spec.ts) Blockers/notes: - All 381 unit tests pass, 40 test files green - Server tsc --noEmit clean - Next.js build fails on pre-existing caniuse-lite/date-fns dependency issues (unrelated) - react-hook-form type errors are pre-existing (unrelated)
Key decisions: - LiteLlmApiAdapter: add createVirtualKey() with POST /key/generate, model-restricted, no QuickStack-managed budget - AgentSandboxAdapter: add createSandboxClaim(), getSandboxClaim(), deleteSandboxClaim(), createOrReplaceSecret(), deleteSecret(), waitForSandboxReady() with timeout polling - AgentRuntimeService: new singleton service with startAgent (virtual key + runtime secret + SandboxClaim + readiness wait), stopAgent (claim + secret deletion), getAgentStatus (live K8s state mapping), statusTextFor (DEPLOYED->Running, never BUILDING) - Agent Runtime Secret: K8s Secret containing QS_GATEWAY_URL, QS_VIRTUAL_KEY, decrypted env vars, QS_SYSTEM_PROMPT - Frontend: Overview tab shows live status (polled every 5s via getAgentStatus), Start/Stop buttons with Toast.fromAction, role-aware readonly - Tests: 5 createVirtualKey + 13 sandbox adapter + 23 agent-runtime service = 41 new tests Files changed: - src/server/adapter/litellm-api.adapter.ts: +createVirtualKey, fetchJson supports POST - src/server/adapter/litellm-api.adapter.unit.spec.ts: +5 virtual key tests - src/server/adapter/agent-sandbox.adapter.ts: +SandboxClaim CRUD, Secret ops, waitForSandboxReady - src/server/adapter/agent-sandbox.adapter.unit.spec.ts: new, 13 tests - src/server/services/agent-runtime.service.ts: new, startAgent/stopAgent/getAgentStatus/statusTextFor - src/server/services/agent-runtime.service.unit.spec.ts: new, 23 tests - src/app/project/agent/[agentId]/overview/actions.ts: new, startAgent/stopAgent/getAgentStatus server actions - src/app/project/agent/[agentId]/agent-detail-client.tsx: live status polling, start/stop buttons, status color coding Blockers/notes: - All 422 unit tests pass, 42 test files green - tsc --noEmit has pre-existing react-hook-form errors only (unrelated) - Virtual keys always created anew on start (no reuse yet per issue spec) - System prompt stored in runtime secret as QS_SYSTEM_PROMPT env var
… rate limits, source, and system prompt
…and Gateway reference blocking
Key decisions:
- AgentSandboxAdapter: add getSecret() to read K8s Secret as decoded string map (404 -> null)
- LiteLlmApiAdapter: add deleteVirtualKey() via POST /key/delete with {keys:[key]} body
- AgentService.deleteById(): full lifecycle — extract virtual key from runtime secret, stop runtime, delete virtual key via LiteLLM (fail preserves DB Agent for retry), delete sandbox resources, transactional DB delete
- LlmGatewayService.deleteById(): block deletion when agent.count > 0 with user-facing error
- Server action deleteAgent: project-level ensureDeleteAgentInProject auth, calls agentService.deleteById
- Frontend: delete button on Overview tab with confirm dialog + Toast.fromAction, redirects to project page
Tests:
- LiteLLM adapter: 4 deleteVirtualKey tests (correct endpoint, error mapping, auth failure, network failure)
- LlmGatewayService: 2 deleteById tests (blocks when agents reference, allows when none)
- AgentService: 4 new deleteById tests (full lifecycle, DB preservation on key failure, missing secret skip, secret read error grace) + 4 saveConfig tests fixed for actual behavior (non-reconciling)
- AgentService unit: 27 tests pass (5 pre-existing reconcile tests replaced with correct save-only tests)
Files changed:
- src/server/adapter/agent-sandbox.adapter.ts: +getSecret()
- src/server/adapter/litellm-api.adapter.ts: +deleteVirtualKey()
- src/server/services/llm-gateway.service.ts: +agent reference check in deleteById
- src/server/services/agent.service.ts: rewritten deleteById with full lifecycle cleanup
- src/app/project/agent/[agentId]/overview/actions.ts: +deleteAgent server action
- src/app/project/agent/[agentId]/agent-detail-client.tsx: +delete button with confirm
- src/server/adapter/litellm-api.adapter.unit.spec.ts: +4 deleteVirtualKey tests
- src/server/services/llm-gateway.service.unit.spec.ts: +2 deleteById tests
- src/server/services/agent.service.unit.spec.ts: +5 deleteById tests, fixed saveConfig tests
Blockers/notes:
- 482 tests pass, 7 pre-existing failures (sandbox adapter v1beta1/v1alpha1 mismatch, agent-config zod, integration build test)
- tsc --noEmit has pre-existing react-hook-form errors only (unrelated)
- Virtual keys extracted from runtime secret at delete time; key absence is non-fatal
- Gateway delete rejection uses agent count query before Prisma Restrict cascade
Key decisions: - standalonePodService.getPodsForAgent(): discover sandbox pods by name match on agentId OR container name "agent" (resilient to unknown sandbox controller labeling) - eventService.getEventsForAgent(): combine pod events + SandboxClaim events via K8s field selector (involvedObject.kind=SandboxClaim), sorted by time descending - Server actions getPodsForAgent / getAgentEvents: auth-bound via isAuthorizedReadForAgent - AgentLogsCard: reuses LogsStreamed SSE component, 10s pod refresh, "No Agent Pod running" empty state - AgentEventsDialog: reuses App pattern (dialog + table), shows pod name or SandboxClaim/<id> in resource column - Overview tab: Events button in Status card + full-width Logs card below grid Tests (10 new, all pass): - standalonePodService: 5 tests (name match, container match, no matches, empty namespace, uid/status) - eventService: 5 tests (pod events, claim events, claim failure graceful, no pods, sort order) Files changed: - src/server/services/standalone-services/standalone-pod.service.ts: +getPodsForAgent() - src/server/services/pod.service.ts: +getPodsForAgent() wrapper - src/server/services/event.service.ts: +getEventsForAgent() - src/app/project/agent/[agentId]/overview/actions.ts: +getPodsForAgent +getAgentEvents - src/app/project/agent/[agentId]/overview/agent-logs-card.tsx: new, pod selector + LogsStreamed - src/app/project/agent/[agentId]/overview/agent-events-dialog.tsx: new, dialog + events table - src/app/project/agent/[agentId]/agent-detail-client.tsx: Events button + Logs card in Overview Blockers/notes: - 492 tests pass, 7 pre-existing failures (unchanged) - tsc --noEmit pre-existing react-hook-form errors only (unrelated) - Pod discovery uses dual match (name + container name); may need refinement if sandbox controller uses unexpected naming - Claim events are best-effort (caught gracefully) - App log/event behavior unchanged
Key decisions: - Extend terminalSetupInfoZodModel enum with 'opencode' type - Terminal service execs ['opencode'] directly (no shell fallback); missing OpenCode surfaces as K8s exec error - TerminalStreamed: accept terminalTypes prop (default ['sh','bash']), agent passes ['opencode'] - AgentTerminalDialog: controlled dialog reusing TerminalStreamed with opencode-only buttons - getAgentPodForTerminal server action: auth-bound (read), resolves agent runtime pod via getPodsForAgent - Terminal button on Overview tab Status card, only visible when agent is DEPLOYED - Existing App shell terminal unchanged Tests (12 new, all pass): - terminal.service: 8 tests (command selection for opencode/sh/bash/null, pod readiness, exec failure/success emit, session cleanup) - actions: 4 tests (pod resolution, no-pod error, auth enforcement, multiple pod first-match) Files changed: - src/shared/model/terminal-setup-info.model.ts: +'opencode' to enum - src/server/services/terminal.service.ts: +opencode command selection - src/app/project/app/[appId]/overview/terminal-streamed.tsx: +terminalTypes prop - src/app/project/agent/[agentId]/overview/agent-terminal-dialog.tsx: new - src/app/project/agent/[agentId]/overview/actions.ts: +getAgentPodForTerminal - src/app/project/agent/[agentId]/agent-detail-client.tsx: +Terminal button + state - src/server/services/terminal.service.unit.spec.ts: new, 8 tests - src/app/project/agent/[agentId]/overview/actions.unit.spec.ts: new, 4 tests Blockers/notes: - 512 tests pass, 7 pre-existing failures (unchanged) - tsc --noEmit clean for changed files (pre-existing react-hook-form errors only) - Container name 'agent' derived from sandbox template; pod discovery via dual match (name + container name) from issue 09 - No shell fallback by design; exec failure closes WebSocket with error output to xterm - Terminal reuses same Socket.IO /pod-terminal namespace and k8s.Exec pattern as App terminal
…ibility on input values
…warm pool replicas
… PUT requests for the API
…ng ui and api instead of "clain" or "instance"
…configuration cards
…or API key management
…reamline job watching process
…iptions and naming
…ove error handling
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This pull request introduces several improvements and and the new workload type "agent" to deploy sandboxed agent containers using the Agent Sandbox CRD.