## Summary
- **console.css + input.css:** Scope `.dark` CSS selectors to CopilotKit
container elements (`.copilotKitDevConsole
.copilotKitDebugMenuTriggerButton` and `.poweredBy` respectively). The
original selectors had bare `.dark,` as standalone entries in
comma-separated selector lists, which applied styles to *any* element
with class `.dark` instead of scoping to CopilotKit elements.
- **colors.css:** Remove broken `:root` pseudo-element from
`body[style*="color-scheme: dark"] :root` — `:root` is `<html>`, which
cannot be a descendant of `<body>`, so this selector never matched
anything.
- **E2E tests:** 4 Playwright regression tests verifying dark mode
styles don't leak into host application elements.
Prevents CopilotKit dark-mode styles from leaking into the host
application.
Closes#2920
---
*Split from #3847*
## Summary
- Widen `TanStackChatMessage.content` from `TanStackContentPart[]`
to `any[]` so messages from `convertInputToTanStackAI` are directly
passable to any TanStack AI adapter without `as any` casts
- Split `TanStackContentPart` into a proper discriminated union with
separate variants per modality
- Add `env.d.ts` with `vite/client` reference for CSS import types
- Fix `onError` callback shape in the example
## Test plan
- [x] All 17 multimodal TanStack tests pass
- [x] All 294 agent tests pass
- [ ] Verify no TS errors in react-router example IDE
🤖 Generated with [Claude Code](https://claude.com/claude-code)
- starter-langgraph-python: use next build --webpack (Turbopack default
in Next 16 can't handle serverExternalPackages)
- starter-crewai-crews: bump ag-ui-protocol>=0.1.10 to match
ag-ui-crewai 0.1.5 requirement
- shell-dojolike: add recharts dependency (used by shared frontend)
- Remove test-integration-tmp from workflow (package was deleted)
- starter-langgraph-python: disable Turbopack for Next.js build
(serverExternalPackages incompatible with Turbopack)
- starter-crewai-crews: pin crewai-tools~=0.47.1 to avoid version
conflict with crewai==0.130.0
- shell-dojolike: add missing zod dependency (required by shared
frontend modules)
Remove captured-but-never-asserted colorBefore and markerColorLight
variables. Replace silent if-guard in poweredBy test with explicit
test.skip so missing elements surface in test reports instead of
silently passing.
Verifies that .dark CSS selectors in console.css and input.css are
properly scoped to CopilotKit elements and do not leak into the host
application. Tests detect the specific rgb(69, 69, 69) color leak that
occurred when .dark was used as a standalone selector.
Templates were pinned at old @ag-ui/client versions (0.0.40–0.0.46).
@copilotkit/runtime@1.55.2 ships 0.0.52, which added private fields to
AbstractAgent. The version split caused a nominal type error on build.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Templates had stale pinned versions ranging from 1.9.3 to 1.52.1.
The CLI does a verbatim sparse checkout of these files, so users
scaffolding via `copilotkit create` were getting old versions installed.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Source route.ts files now read process.env.AGENT_URL natively (PR #3737),
so the build-time sed patches and verification greps are no longer needed.
Also removes crewai-crews docker-route-override.ts and its COPY line,
since that starter's route.ts already has AGENT_URL support.
TanStackChatMessage.content uses any[] for multimodal so
messages are directly passable to any adapter without casts.
Add env.d.ts with vite/client types for CSS imports.
- Fix convertInputToTanStackAI silently dropping multimodal
content (images, audio, video, documents) by converting
AG-UI content parts to TanStack AI ContentPart format
- Switch react-router example from Hono server to a React
Router resource route using createCopilotRuntimeHandler
- Both agents now use BuiltInAgent factory pattern
- Remove hono/react-router-hono-server deps from example
## Summary
Adds **factory mode** to `BuiltInAgent`, enabling backend-agnostic LLM
integration without introducing a new class. Users provide a factory
function that creates the LLM stream — CopilotKit handles lifecycle
events, stream-to-AG-UI conversion, error handling, and
abort/cancellation.
### What's included
- **Three backend types**: `aisdk` (Vercel AI SDK), `tanstack` (TanStack
AI), and `custom` (raw AG-UI events)
- **Stream converters**: `convertAISDKStream` and
`convertTanStackStream` extract AG-UI events from backend-native streams
- **Input helpers**: `convertInputToTanStackAI` and
`convertMessagesToVercelAISDKMessages` for message/prompt conversion
- **State management**: Automatic `STATE_SNAPSHOT`/`STATE_DELTA` event
extraction from tool results
- **Reasoning support**: Auto-close logic for providers that never emit
`reasoning-end` (e.g. `@ai-sdk/anthropic`)
- **273 tests** across 16 test files covering all backend types,
converters, state tools, multimodal, and edge cases
- **Docs**: Comprehensive factory mode guide with 16 examples covering
tools, reasoning, state, forwardedProps, structured output
- **Example app**: React Router example migrated to factory mode
### Architecture
Factory mode is a discriminated union on `BuiltInAgentConfiguration`:
```typescript
const agent = new BuiltInAgent({
type: "aisdk", // or "tanstack" | "custom"
factory: ({ input, abortSignal }) =>
streamText({ model: openai("gpt-4o"), messages: ..., abortSignal }),
});
```
No new public classes — `BuiltInAgent` handles both simple mode (model
string) and factory mode (user-owned LLM call).
## PR feedback addressed
### Critical fixes
- **C1/M3**: Fixed TOCTOU window — `abortController` now set
synchronously before Observable creation in classic `run()`, with
concurrent run guard
- **C2/M2**: Classic `tool-result` now checks both `output` and `result`
property names, with `try/catch` on `JSON.stringify`
- **C3**: Error case safely handles `undefined` `p.error` with proper
fallback chain
### Important fixes
- **I2**: `RUN_ERROR` events now include `threadId` and `runId` in
classic mode (matching factory mode)
- **I3**: State tool events guarded against `undefined` snapshot/delta
before emitting
- **I4**: `JSON.stringify` wrapped in try/catch with descriptive
fallback message
- **M1**: `assignToolsToAgents` now skips factory-mode agents (prevents
config corruption)
- **M5**: TanStack converter's dropped event types documented with clear
limitation note
### Suggestions addressed
- **S1**: `@ts-expect-error` on `clone()` middlewares documented with
rationale
- **M4**: `AgentFactoryContext.abortController` kept (needed by TanStack
AI) but JSDoc added discouraging direct `.abort()` calls
## Test plan
- [x] 273 unit tests pass across all 3 backend types
- [x] Build succeeds
- [x] React Router example compiles
- [x] Docs render correctly (Vercel preview)
## Summary
Adds a `/health` GET endpoint returning `{"status": "ok"}` to all 11
starter integration examples. This is a best practice for any deployed
service — health checks enable platform health monitoring, load balancer
probes, and container lifecycle management.
**11 starters updated:**
- adk, agent-spec, agno, crewai-crews, crewai-flows, langgraph-fastapi,
llamaindex, ms-agent-framework-dotnet, ms-agent-framework-python,
pydantic-ai, strands-python
**1 skipped:**
- langgraph-python — uses `langgraph dev` which provides its own `/ok`
endpoint
All `/health` endpoints are registered before any catch-all mount to
ensure they take priority (lesson learned from the pydantic-ai showcase
bug).
## Test plan
- [ ] Each starter builds successfully
- [ ] `curl localhost:8000/health` returns `{"status": "ok"}` on each
## Summary
- **Configurable agent URL**: All starter `route.ts` files now read
`AGENT_URL` from environment with localhost fallback, enabling
deployment to non-localhost environments without code changes. Applied
to: pydantic-ai, adk, agno, llamaindex, ms-agent-framework-python,
ms-agent-framework-dotnet, crewai-crews.
- **crewai-crews version bump**: Bumped `ag-ui-crewai` from pinned
`==0.1.3` to `>=0.1.5` to pick up the `MethodExecutionFinishedEvent`
dict state fix from ag-ui-protocol/ag-ui#1478.
- **`.env.example` files**: Added `.env.example` with `OPENAI_API_KEY`
and `AGENT_URL` (correct default port per starter) to all affected
starters.
## Starters NOT touched (already configurable or different pattern)
- strands-python (uses `STRANDS_AGENT_URL`)
- langgraph-python / langgraph-js (use `LANGGRAPH_DEPLOYMENT_URL`)
- langgraph-fastapi (has docker-route-override)
- mastra (monolith, no separate agent URL)
## Test plan
- [ ] Verify each starter builds with `next build`
- [ ] Verify `AGENT_URL` env var is respected when set
- [ ] Verify localhost fallback works when `AGENT_URL` is unset
- [ ] Verify crewai-crews agent installs with `ag-ui-crewai>=0.1.5`
The starter pinned ag-ui-crewai==0.1.3 which crashes on crew completion:
'dict' object has no attribute 'messages' in MethodExecutionFinishedEvent
handler. Fixed upstream in ag-ui-protocol/ag-ui#1478, released in 0.1.5.
Every starter's route.ts now reads process.env.AGENT_URL as the
primary agent endpoint, with framework-specific vars as fallback:
- AGENT_URL || "http://localhost:8000" (pydantic-ai, adk, agno, crewai,
ms-agent-python, ms-agent-dotnet, strands-python)
- AGENT_URL || LANGGRAPH_DEPLOYMENT_URL || "http://localhost:8123"
(langgraph-python, langgraph-js)
- AGENT_URL || "http://localhost:9000" (llamaindex)
Also adds .env.example with AGENT_URL to all starters.
Ran Shemtov