- ThreadStoreRegistry.register() now fires onThreadStoreUnregistered
before overwriting an existing store so subscribers (web inspector)
don't stay subscribed to stale stores on replacement; test extended
to verify both events fire in order
- Wire handleClearThreads to POST /threads/clear route; add RouteInfo
variant, router pattern, and fetch-handler case so the inspector can
actually call it
- Add handleGetThreadMessages tool-call mapping test using properly
typed Message objects (role as const, type as const) — exercises the
real mapping code path without mocking getThreadMessages
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Excludes showcase/shell-docs and showcase/shell-dojo demo-content.json from
the size check in lefthook.yml — these data files were added by main but
weren't in the exemption list, causing the pre-commit hook to reject them.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- middleware SKILL: beforeRequestMiddleware runs AFTER hooks.onRequest (not before); cite fetch-handler.ts:136-147 for exact order
- agent-runners/in-memory ref: delete globalThis[Symbol] leaves module-captured reference intact; use in-place Map.clear() on data.stores + data.historicRunsBackup; flag missing official reset helper as follow-up
- intelligence-mode SKILL: identifyUser does NOT forward thrown Response (resolveIntelligenceUser converts to generic 500); gate auth rejection in hooks.onRequest which does forward Responses
- built-in-agent/factory-modes ref: custom factory example now checks abortSignal.aborted at entry and between yields; note that agent.abortRun() only flips the flag — generator must consult it
- setup-endpoint Express delegate: stream SSE via Readable.fromWeb().pipe(res) instead of buffering via arrayBuffer(); applies to both the recommended pattern and the CRITICAL "Correct" example. Add Node 18.17+ note.
- built-in-agent frontmatter: maxSteps defaults to undefined (AI SDK stops after one generation), not 1.
- built-in-agent examples: replace bare anthropic("claude-sonnet-4") / anthropicText("claude-sonnet-4") with the dated id "claude-sonnet-4-5-20250929" (AnthropicMessagesModelId in @ai-sdk/anthropic@3.0.49 does not accept the bare form).
- 0-to-working-chat Next.js: align catch-all with migration-playbook — use [[...slug]] so the bare /api/copilotkit basePath also routes.
- migration-playbook v1->v2 parameters: v1 entries default to required: true; only map to .optional() when required: false was explicit. Prevents silently flipping the contract.
- switching-agents Recipe 2 (Tabs): hold activeAgent in a ref so the onAgentsChanged subscribe effect only re-binds on copilotkit change, not on every tab click.
- a2ui-rendering custom-catalog example: use the real createCatalog(definitions, renderers) signature with Zod schemas; remove the bogus extractSchema<T>() generic (types erase at runtime).
- spa-without-runtime clipboard handler: correct the comment — returning a structured error does NOT trigger onError with tool_handler_failed; only thrown handlers do.
- agent-runners: agent_thread_locked fires only on Intelligence 409, not SSE;
SSE surfaces "Thread already running" as a plain 500 body
- agent-runners: SqliteAgentRunner install-hint is unreachable — the import of
@copilotkit/sqlite-runner fails at module load if better-sqlite3 is missing
- agent-runners: soften "CopilotIntelligenceRuntimeOptions explicitly excludes
runner" — the field is simply not declared; excess-property checks flag it
- agent-runners/sqlite: mirror real schema (CREATE INDEX IF NOT EXISTS)
- middleware: consistent afterRequestMiddleware signature across Wrong/Correct
examples; document the full { runtime, response, path, messages?, threadId?,
runId? } shape
- server-side-tools: Factory Mode example now uses AI SDK + the real converter
(convertToolDefinitionsToVercelAITools); TanStack caveat called out since
defineTool output is not a TanStack tool and there is no built-in converter
- setup-endpoint: Cloudflare Workers example hoists runtime/handler to module
scope behind a lazy getter; openaiText signature corrected to createOpenaiChat
when passing an explicit API key (openaiText config omits apiKey)
- transcription: JSON payload mode works in both multi-route and single-endpoint
— dispatch is by Content-Type, not endpoint mode
- wiring-external-agents/mastra: getLocalAgents requires resourceId; remote
Mastra section now points at HttpAgent instead of misshaping getRemoteAgents
- package.json: add per-condition types to every exports entry so bundler and
moduleResolution: "bundler"/"node16" both pick up the matching .d.mts/.d.cts
intent edit-package-json replaced "files": ["skills"] instead of appending, excluding dist/ from runtime and react-core published packages. publint caught it (pkg.exports.[].import/require -> file not published). Restoring dist + skills.
a2ui-renderer was unaffected (had existing "files": ["dist"] which got appended correctly).
- Keep deletion of sdk-python/copilotkit/langgraph_agent.py (deprecated LangGraphAgent
removed in this PR; main's unrelated bug fixes are superseded by our removal)
- Resolve poetry.lock conflict by taking main's ag_ui_langgraph 0.0.33
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Review-loop fixes on the runtime surface:
- sse-response.ts: wrap DebugEventBus.broadcast in try/catch inside the
observer's next — a buggy subscriber used to propagate an exception
that RxJS routed to `error`, killing the SSE stream for an unrelated
reason. Logged via the existing logError helper instead.
- sse-response.ts: non-abort write failures no longer fall through
silently. Previously streamClosed stayed false and the next event
re-attempted a broken writer. Now logs + sets streamClosed.
- sse-connect-agent-id.test.ts: new unit test that feeds a synthetic
observable into handleSseConnect and asserts the emitted envelope
carries the route-resolved agentId ('weather-agent') rather than
the literal 'connect'. Regression guard for the fix in 7f3a8c27c
that the /run-based integration tests didn't cover.
- debug-events.suite.ts: replaced a version of this assertion that
couldn't reliably drive /connect in the integration runtime with a
pointer to the unit test above.
- fetch-router.ts: clarify the cpk-debug-events comment — the prefix
is what guards against agent-name collision, not the ordering of
router branches.
- debug-stream.ts (extension client): include the underlying JSON parse
error message in the emitted error so a truncated SSE frame can be
distinguished from a runtime/protocol version mismatch.
Per PR review suggestion (14) — a cpk- prefix makes collision with a
user-named agent essentially impossible. The earlier change only
documented the route as reserved; renaming it removes the hazard.
Rename scope:
- URL path segment (/debug-events → /cpk-debug-events) in the router,
handler switch arms, method-not-allowed check, and integration suite.
- RouteInfo union discriminator ('debug-events' → 'cpk-debug-events')
in hooks.ts and its two consumers in fetch-handler.ts.
- VS Code DebugStream client URL.
- Docs page (event-inspector.mdx) — both the Callout and the Steps copy.
Left unchanged (internal names that describe the feature but aren't the
public route):
- packages/runtime/src/v2/runtime/handlers/handle-debug-events.ts
- packages/runtime/src/v2/runtime/core/debug-event-bus.ts
- packages/shared/src/debug-event-envelope.ts (and the DebugEventEnvelope
type)
- 'debug-event' (singular) message type in the VS Code webview bridge —
that's the extension host ↔ webview protocol, not the HTTP route.
Critical
- scripts/hooks/check-binaries.sh: restore showcase data-file exclusions
(demo-content / search-index / starter-content >1 MB) that the inline
refactor dropped; add 'set -eu' so silent shell failures don't hide
policy violations.
- packages/vscode-extension/src/extension/preview-panel.ts: drop 'blob:'
from CSP script-src — it lets arbitrary-string JS execute via Blob URL
and defeats most of CSP's XSS protection.
- packages/vscode-extension/package.json: set private: true. The extension
ships as a .vsix via vsce, not npm, and workspace:* devDependencies
would break an accidental 'npm publish'.
Important
- runtime: forward the real agentId from handleConnectAgent into
handleSseConnect / createSseEventResponse so DebugEventBus envelopes
on /connect carry the actual agent name instead of the literal
'connect'. Updates handle-connect.ts, sse/connect.ts.
- hooks/panel.ts CSP: narrow connect-src from 'https:' to just the
Tailwind CDN. The preview path never drives a real CopilotKit runtime
— all hook calls route through the stub — so there's no legitimate
https: fetch to allow from inside bundled user code.
- extension/utils.ts getNonce(): switch to crypto.randomBytes. Math.random()
is not acceptable for a value that gates inline-script execution.
- .github/workflows/vscode-extension.yml:
* Build step uses 'nx run copilotkit-vscode-extension:build' instead
of 'pnpm run build' (targeted build with Nx caching, not full
monorepo rebuild).
* Added explicit 'Type check' step (tsc --noEmit).
* Added Lint step gated with continue-on-error until the Nx target
exists, so a missing target doesn't break the pipeline.
* Publish job now queries the Marketplace for the current published
version and skips 'vsce publish' when the local package.json
version matches — stops every docs/CI-only push to main from
failing on duplicate-version errors.
- hooks/hook-scanner.ts: bound the synchronous walk at 20 000 files so a
pathologically large workspace can't freeze the extension host; flag
kept in module-level constant with a rationale comment.
- sse-response.ts: document that debugEventBus.broadcast intentionally
runs before the stream-closed gate so debug subscribers see trailing
events even after the SSE client disconnects.
- inspector-panel.ts: subscribe to DebugStream lazily on show() rather
than in the constructor; unsubscribe on panel dispose. Avoids firing
the event callback on every envelope when no panel is open.
Suggestions
- fetch-router.ts: document 'debug-events' as a reserved route so it
can't be shadowed by an agent literally named 'debug-events'.
- activate.ts findValuePosition: add optional startOffset parameter,
document the first-occurrence limitation + the follow-up path for
per-fixture precision.
- activate-hook-explorer.ts isInsideWorkspace: fix JSDoc to reflect the
code (root itself is excluded).
Tests the /debug-events SSE endpoint across Express, Hono, Node, and
Fetch Direct adapters. Verifies SSE response format, event flow through
the DebugEventBus, envelope structure (timestamp, agentId, threadId,
runId, event), full event sequence, HTTP method validation (405 for
POST), and NODE_ENV=production guard (404).
Key implementation detail: reader.cancel() must NOT be awaited on
tee'd ReadableStreams (created by response.clone() in the fetch
handler) because Node.js blocks until the other tee branch is also
consumed/cancelled.
## Summary
- Fixes#2986
- `reqOrRequest instanceof Request` fails when `@hono/node-server`
polyfills the Request class with a different prototype
- Replaced with a duck-type check (`isRequestLike`) that verifies `url`,
`method`, and `headers` properties exist
- The `!res` guard ensures IncomingMessage objects (which always come
with a ServerResponse) are still routed correctly
## Test plan
- [x] Added `request-duck-type.test.ts` covering native Request,
polyfilled Request, null/undefined, and missing properties
- [x] All 1225 existing runtime tests pass
- [x] Build passes
## Summary
- LangChain adapter now always uses `randomId()` for message stream IDs
instead of relying on `lc_kwargs.id` (which GoogleGenerativeAI sets to
`"0"` for all messages)
Closes#2929
## Test plan
- [x] `@copilotkit/runtime` tests pass
🤖 Generated with [Claude Code](https://claude.com/claude-code)
## Summary
- Default `max_tokens` was hardcoded to 1024, causing truncated
responses for most use cases
- Increased default from 1024 to 4096 while still respecting
user-provided `maxTokens`
Fixes#2354
## Test plan
- [x] Added test verifying default max_tokens is 4096 when not specified
- [x] Added test verifying user-provided maxTokens is passed through
- [x] All existing anthropic adapter tests pass
## Summary
- Providers like `@ai-sdk/openai-compatible` emit sequential IDs
(`txt-0`, `reasoning-0`, `msg-0`) that are treated as unique but collide
across requests
- The existing check only caught the literal `"0"` — expanded to match
the `^(txt|reasoning|msg)-0$` pattern
- Both `text-start` and `reasoning-start` event handlers now use
`randomUUID()` for these non-unique IDs
## Test plan
- [x] Added `provider-id-collision.test.ts` with 4 tests covering txt-0,
reasoning-0, msg-0, and legitimate ID preservation
- [x] All 1226 runtime tests pass (including existing
config-tools-execution tests)
Fixes#3410, #3623
## Summary
- `onAfterRequest` middleware callback was called with `{}` instead of
actual hook parameters
- Now forwards `threadId`, `runId`, `inputMessages`, `outputMessages`,
and `url` from the v2 runtime's `hookParams`
- The `onBeforeRequest` handler at line 566 already did this correctly;
this fix makes `onAfterRequest` consistent
## Test plan
- [x] Added `on-after-request.test.ts` verifying the callback receives
threadId and runId (not empty object)
- [x] All 1226 runtime tests pass
Fixes#2124
## Summary
Fixes#2450
Token trimming in the Anthropic adapter could remove assistant messages
containing `tool_use` blocks while keeping the corresponding user
messages with `tool_result` blocks. Anthropic's API rejects orphaned
`tool_result` messages. This adds a post-processing step that removes
orphaned `tool_result` blocks after trimming.
## Test plan
- [ ] Verify token trimming preserves tool_use/tool_result pairing
- [ ] Verify orphaned tool_result blocks are removed
- [ ] Verify messages with mixed content (tool_result + text) retain
non-orphaned content
## Summary
Fixes#2405
Both OpenAI and Anthropic adapters used a hardcoded token limit for
message trimming. This adds an optional `maxInputTokens` constructor
parameter to both adapter classes, allowing consumers to override the
default context window limit.
## Test plan
- [ ] Verify default behavior unchanged when `maxInputTokens` not
provided
- [ ] Verify custom `maxInputTokens` is passed through to
`limitMessagesToTokenCount`
## Summary
Fixes#2504
When Anthropic returns a `tool_use` block for a tool not registered in
the current action set, the adapter would crash trying to process it.
This adds a check against known action names and skips unknown tool_use
blocks gracefully.
## Test plan
- [ ] Verify known tools are still processed normally
- [ ] Verify unknown tool_use blocks are skipped without crashing
## Summary
- langchain-mcp-adapters sends tool call result content as an array of
`{type:"text", text:string}` objects
- Extracts and joins text parts so downstream consumers always receive a
plain string
Closes#2922
Closes#1936
`extractParametersFromSchema` now preserves enum values as structured
data on string parameters, recursively extracts nested object
attributes, and converts object arrays to `object[]` type with
attributes. This ensures complex MCP tool schemas survive the
`Parameter[]` -> Zod conversion without losing nested structure.
Split from #3838.
The compatibility layer correctly handles v6 (same API surface as v5),
but the version range ^5.0.0 excluded v6.x. Changed to >=5.0.0 so
users on the latest openai SDK don't get peer dependency warnings.
Mike Ryan