Three follow-ups on top of PR #4837 that I had on the same branch but
didn't make it into the squash merge.
1. **packages/runtime: stamp `audio/webm` on empty-type Blobs in the
transcription handler.** Browser MediaRecorder writes the audio as
webm/opus, but the Blob's `type` field is often empty by the time it
hits the server. `isValidAudioType` lets empty / octet-stream through
for compatibility, but OpenAI Whisper then rejects the upload with
`502 Invalid file format. Supported formats: ['flac', 'm4a', 'mp3',
'mp4', 'mpeg', 'mpga', 'oga', 'ogg', 'wav', 'webm']` because it
can't pick a decoder. Reconstructing the File with an explicit
`audio/webm` type (and a `.webm` filename fallback) makes Whisper
accept the bytes that were already valid. Monorepo-wide — applies to
every integration using `/api/copilotkit-voice/transcribe`.
2. **showcase/aimock/feature-parity.json: port 12 subagents fixtures
from d5-all.json** so the three pills (cold-exposure blog, LLM
tool-calling explanation, reusable-rockets summary) work in
production. d5-all.json already has the full research → writing →
critique chain with substantive content; feature-parity only had the
single LP remote-work pill. Production aimock loads both files but
any case where feature-parity wins first-match needs the same
content. Verbatim port — no fabricated text. Net result: no more
`[sub-agent error] the writing agent...` on the demo's pills.
3. **showcase/aimock both files: scope shared-state-read-write Greet +
Plan-a-weekend fixtures with a true all-defaults systemMessage
gate.** The PR #4837 gate (`systemMessage: "tone: casual"`) only
caught tone changes — name / language / interests changes still hit
the canned fixture. Replaced with a two-element array gate (aimock
supports all-present substring matching, verified in
`/app/dist/router.js`):
- `preferences:\n- Preferred tone: casual\n` — breaks if name is
set (Name line inserts between signature and tone) or tone changes.
- `- Preferred language: English\nTailor every response` — breaks
if language changes or interests are added (Interests line
inserts between language and Tailor).
With `--provider-gemini` already wired in both local docker-compose
and Railway prod, any state change now proxies to real Gemini and
returns a personalised reply.
4. **showcase/aimock/feature-parity.json: re-remove bare 'plan' /
'steps' / 'mars' / 'dashboard' / 'report' substring catch-alls + the
bare 'alice' / 'Alice' fixtures.** These were removed in commit
`ddc2e179` on the PR #4837 branch but didn't survive the squash
merge, so they're back in main and still hijacking hitl-in-app
downgrade-#12346 ('plan'), shared-state-rw weekend pill ('plan'),
subagents 'rockets' pills, hitl-in-chat Schedule-1:1 with Alice
('alice'). Replace the alice pair with a single scoped
`Hi, my name is Alice` fixture for the showcase-assistant
introduction flow.
Local verification:
- `bin/showcase test google-adk --d5` → 38/38 green, 165s.
- Paired curl on shared-state-read-write:
- Default state → canned fixture ("Hi — I'm your shared-state co-pilot…")
- `name=alem` → real Gemini ("Hi there! …")
- `interests=[Cooking, Travel]` weekend pill → real Gemini ("Hey
there! Since you're into cooking and travel, how about a weekend
plan that combines both?")
Production deploys this PR will pick up the aimock fixture changes
(prod loads feature-parity.json from GitHub raw at boot — no image
rebuild needed for that file) plus the runtime change once the
packages/runtime build is republished.
Three layers of regression guards for the runtime reasoning-role filter and
the chained demo behavior:
1. Runtime unit test — packages/runtime/.../run-message-filtering.test.ts:
- Verifies `LangGraphAgent.run` strips `role:"reasoning"` from
`input.messages` before delegating to super.run.
- Verifies user/assistant/system/tool messages pass through in order.
- Verifies empty + missing messages arrays are tolerated.
- Verifies pre-existing forwardedProps.streamSubgraphs default + override
behavior is preserved.
- 6/6 tests pass against the runtime package's vitest config.
2. D5 harness probe — showcase/harness/.../d5-tool-rendering-reasoning-chain.ts:
- Expanded from one chained turn (flights→weather) to all three chained
pills in a single thread (stocks AAPL→MSFT, dice d20→d6, flights→weather).
- This is the canonical multi-pill regression at the harness layer:
without the runtime reasoning-role filter, the second pill would crash
before the model was called.
- Each turn asserts the per-turn delta of reasoning-block mounts (idx+1),
the minimum card count for each tool group, and unique transcript
substrings that scope to that turn.
3. Playwright e2e spec — showcase/integrations/langgraph-python/tests/e2e/
tool-rendering-reasoning-chain.spec.ts:
- Mirrors the pattern of the sibling tool-rendering-default-catchall spec
(notably its multi-pill regression at lines 162-212).
- Page-loads test verifies the 3 pills mount and no cards leak from a
prior session.
- One test per chained pill (stocks, dice, flights+weather) asserts the
full chain renders with reasoning-block + correct per-tool cards +
narration matching the aimock fixture text.
- Sequential-pills regression test clicks all 3 pills in one thread,
asserts each chain renders independently AND the reasoning-block count
increases monotonically across turns.
Agent: extend `get_stock_price` to accept optional `price_usd` and
`change_pct` arguments (mirrors the basic tool-rendering agent's signature
introduced in #4770). The aimock fixtures script the chained AAPL/MSFT
comparison by passing deterministic prices via these args; without the
wider signature, pydantic rejects the tool call and the card never mounts.
The runtime unit test is the strongest guard — it would catch any
regression on the role-filter logic without depending on the full Docker
stack. The harness probe and Playwright spec catch end-to-end regressions
in the canonical CI environment.
`@ag-ui/langgraph`'s message converter handles only user/assistant/system/tool
roles and throws `"message role is not supported."` on anything else. Agents
that stream reasoning summaries (OpenAI Responses API + `reasoning={summary:
"detailed"}`) emit AG-UI messages with `role: "reasoning"` that the AG-UI
client replays in the next turn's `input.messages`; the converter then crashes
before the model is ever called and the second pill click in a multi-turn
thread produces an `INCOMPLETE_STREAM` error.
Strip `role: "reasoning"` from `input.messages` inside CopilotKit's
LangGraphAgent.run subclass before delegating to super. This is the narrowest
fix at the runtime/AG-UI boundary — only the inbound message list is filtered,
the outbound event stream still carries reasoning summaries to the client, so
the `<ReasoningBlock>` slot continues to render on the active turn.
- Add telemetryDisabled to RuntimeInfo from COPILOTKIT_TELEMETRY_DISABLED/DO_NOT_TRACK env vars
- Mirror through AgentRegistry and expose via CopilotKitCore getter
- Guard track calls, URL param appending, and console disclosure on core.telemetryDisabled
- Move maybeShowDisclosure() to onRuntimeConnectionStatusChanged (fires after core attaches)
- Update docs to replace localStorage toggle description with env var approach
- Add telemetryDisabled test suite to get-runtime-info tests
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Move `package` from properties string to top-level `{ name }` object per
Ben's confirmed IngestPayload schema (telemetry-sink-ingest/index.ts:127-134)
- Add typed per-event helpers trackBannerViewed/trackBannerClicked/trackThreadsTabClicked
to enforce property shapes at call sites and prevent PII leakage under wrong keys
- Add trackBannerClickedOnce guard in index.ts (per-mount Set keyed by
banner_id + cta) to prevent banner_clicked inflation on repeated clicks
- Fix handleTelemetryOptOutToggle: replace ?? true fallback with
instanceof HTMLInputElement guard (wrong fallback was a privacy bug)
- Add threadsTabClicked re-selection guard (skip if already on threads tab)
- Replace getTelemetryDistinctIdForUrl() call on mount with ensureTelemetryDistinctId()
- Add inMemoryFallbackId in persistence.ts for funnel coherence when
localStorage is unavailable (same UUID returned per page load)
- Add _resetTelemetryPersistenceForTesting() for test isolation
- Remove @copilotkit/shared dep from telemetry-disclosure.ts (inline
env-var check; keeps module self-contained and testable in isolation)
- Add clearMocks: true to web-inspector vitest config (fixes spy call
history accumulating across tests)
- Expand telemetry.test.ts to 22 tests covering wire body shape, opt-out,
5 error-resilience paths, typed helpers, distinct ID lifecycle (SSR +
localStorage-throws + funnel coherence), maybeShowDisclosure, and
getTelemetryDistinctIdForUrl
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Three V1 funnel events from the inspector — oss.inspector.banner_viewed,
oss.inspector.banner_clicked, oss.inspector.threads_tab_clicked — plus a
privacy panel for opt-out, a first-run console disclosure on inspector
mount and runtime startup, and inspector content added to the canonical
/telemetry docs page on main.
Inspector POSTs directly from the browser to telemetry.copilotkit.ai/ingest
(per ticket: URL is intentionally clearly named for transparency in DevTools).
Inline fetch POST in lib/telemetry.ts — no @copilotkit/shared dep on the
inspector, no dependency on any non-main branch.
Wire body shape (conservative; needs Ben confirmation):
POST https://telemetry.copilotkit.ai/ingest
{ event, properties: { ...caller, distinct_id, package }, ts }
If the lambda expects a richer envelope, update the single JSON.stringify
in lib/telemetry.ts.
Privacy invariants:
- Opt-out toggle short-circuits before any network call (verified by test).
- Properties are scoped to event metadata only — no message content, agent
state, prompts, completions, banner markdown. Negative test pins the wire.
- Anonymous distinct ID (UUID v4 in localStorage) set on inspector load and
propagated onto banner CTA links as ?posthog_distinct_id=<uuid> so the
destination site can posthog.alias() and close the
banner_viewed → banner_clicked → signup_attributed funnel. URL param
suppressed when opted out.
- Console disclosure on first inspector mount and runtime startup. Both
link to https://docs.copilotkit.ai/telemetry.
Plan gaps addressed:
- CTA name on banner_clicked: cta:'body'|'dismiss' (click location) plus
optional cta_label read defensively. Sam: confirm dismiss treatment.
- De-anon opt-out folded into the single toggle. Docs say so explicitly.
- banner_viewed dedup: per-instance Set<string> keyed by timestamp.
- EPIC consent / pixel review: out of scope for this PR; flagged at merge.
Deferred for V1.1:
- Wire body shape (Ben).
- Event-type allowlist for oss.inspector.* (Ben — oss-path-to-production).
- posthog_distinct_id URL-param key name (Ben/Tyler/website team).
Refs https://linear.app/copilotkit/issue/OSS-96
Two CR comments addressed:
- Rename the local destructure of forwardedProps.auth.copilotkitIntelligence
from 'cki' to 'cpki' so it matches the project-wide abbreviation already
used in metadata fields (cpki_event_id, cpki_event_seq, etc).
- Replace the inline 'X-Cpki-User-Id' string literal with the existing
INTELLIGENCE_USER_ID_HEADER constant exported from intelligence-platform/client.
Applies to the runtime auto-attach in agent/index.ts and to the three
test sites in intelligence-mcp-helper.test.ts so the user-side and
runtime-side stay in sync.
Move the per-request Intelligence MCP bag from
forwardedProps.copilotkitIntelligence to
forwardedProps.auth.copilotkitIntelligence so the Intelligence-side
redaction policy strips it. The 'auth' namespace is the convention for
credentials; persistence sinks (Postgres, Redis, S3) and FE replay
paths in apps/realtime-gateway already strip everything under it.
Updates:
- Emitter (handlers/intelligence/run.ts): merge the bag into a single
forwardedProps.auth object alongside any upstream auth keys, and
only emit the auth namespace when there is something to put in it.
- Reader (agent/index.ts): read from forwardedProps.auth.copilotkitIntelligence
instead of forwardedProps.copilotkitIntelligence.
- Tests (intelligence-mcp-helper.test.ts): three fixtures rewritten
to the nested shape.
Four cases at the agent layer (BuiltInAgent reading forwardedProps):
* attaches when `copilotkitIntelligence` carries all three strings
(userId, apiKey, mcpUrl) — outbound headers carry Authorization +
X-Cpki-User-Id.
* does NOT attach when the bag is absent (no Intelligence wiring on
this run).
* does NOT attach when the bag is partial (e.g. mcpUrl missing).
* does NOT attach when the user has already configured an MCP server
pointing at the same URL — explicit user config wins, with the
user's headers and resolver hitting the wire.
Runtime side (intelligence/run.ts): when `runtime.intelligence.mcpServer`
is enabled, build a `copilotkitIntelligence` bag carrying the resolved
user-id, project apiKey, and the platform MCP URL, and pass it through
on `RunAgentInput.forwardedProps` to the agent. Skipped when the flag
is off — runs that don't go through this Intelligence path simply
don't see the bag.
Agent side (BuiltInAgent's run code): if `forwardedProps.copilotkitIntelligence`
contains all three string values AND the user's static `config.mcpServers`
doesn't already include the same URL, append a per-request
`MCPClientConfigHTTP`. Its `options.fetch` closes over apiKey + userId
and stamps `Authorization: Bearer <apiKey>` and `X-Cpki-User-Id:
<userId>` on every outbound MCP call. The custom fetch is the MCP
TypeScript SDK's documented extension point for per-request header
injection — no extra wrapper class, no separate framework concept.
The agent class is otherwise untouched: no new fields, no per-request
side channels, no typed reference to `CopilotKitIntelligence`. Other
agents that don't read `forwardedProps.copilotkitIntelligence` ignore
the bag.
Pulls in the AI SDK's stable `createMCPClient` export (rename from
`experimental_createMCPClient`) — the experimental name was deprecated;
`mcp-clients.test.ts`'s mock setup follows.
Adds `mcpServer?: boolean` to `CopilotKitIntelligenceConfig` (default
`false`). When true, the runtime emits the per-request bag the agent
needs to attach the platform's MCP server.
Internal accessors `ɵisMcpServerEnabled()` and `ɵgetApiKey()` round
out the existing `ɵgetApiUrl()`. Used by the runtime layer in the
forthcoming auto-attach commit; not part of the public surface.
Convenience re-exports so consumers wiring custom MCP clients (via
`mcpClients`) or custom transports don't need to add `@ai-sdk/mcp`
to their dependencies just to type the values.
Replaces the 501 stubs in handleGetThreadEvents and handleGetThreadState
with real delegation to CopilotKitIntelligence. Adds two new client
methods (`getThreadEvents`, `getThreadState`) on intelligence-platform/
client.ts that hit the new `/api/_inspect/threads/:id/{events,state}`
routes shipped in Intelligence PR #144 (CPK-7453). Auth flows through
the existing `resolveIntelligenceUser` path; threadId scoping happens
server-side via the API key's org/project resolution.
Wire shapes match the in-memory branch so the inspector consumes both
runtimes identically:
- events: `{ events }` (platform-internal `decodeErrorRowIds` and
`truncated` flags are stripped at the runtime boundary)
- state: `{ state }` where the platform's discriminated `ThreadStateResult`
flattens to the snapshot value for `kind: "snapshot"` and to `null` for
both `no-snapshot` and `snapshot-decode-error`
Replaces the 501 regression-protection tests with real delegation tests
that mock the platform methods, assert call args, and exercise the
no-snapshot / decode-error / throw paths.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Three new test areas covering surfaces this PR introduced or relies on:
- fetch-router: matchRoute tests for `/threads/:id/events`,
`/threads/:id/state`, and `/threads/clear` (with and without URL
encoding). Critically pins that "/threads/clear" resolves to
`threads/clear` and does NOT fall through to the more permissive
`threads/update` arm with threadId="clear" — the explicit guard in
the router exists for this reason.
- fetch-handler validation: 405 enforcement for POST/PATCH/DELETE on
the read-only `/threads/:id/events` and `/threads/:id/state`
endpoints, with `Allow: GET` header. Complementary positive case
asserts GET is NOT a 405.
- handle-run: end-to-end test that handle-run.ts:40
(`agent.agentId = agentId`) propagates the registry key onto historic
runs. Without this stamp, InMemoryAgentRunner falls back to "default"
and the agentId filter on `GET /threads?agentId=...` breaks for the
local-dev fallback.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Type quality:
- intelligence/threads.ts handleGetThreadMessages: switch on the Message
discriminant (role) and read narrowed fields directly. Removes
`as Record<string, unknown>` laundering and chained `as` casts on
toolCalls/function/arguments. AssistantMessage's toolCalls always have
`function: { name, arguments }`, so the prior fallbacks (`tc.name`,
`tc.args`) were dead branches.
- in-memory.ts getThreadState: import StateSnapshotEvent and use it
instead of `(event as { snapshot?: unknown }).snapshot`.
Comment / API alignment:
- intelligence/threads.ts handleClearThreads JSDoc no longer claims the
inspector calls this; the actual caller is the demo button.
- in-memory.ts clearThreads JSDoc updated to match.
- in-memory.ts getThreadEvents JSDoc no longer references a SQLite
runner that does not exist; just describes the compaction logic.
- web-inspector lint fix: rename unused `changed` parameter to
`_changed` per oxc no-unused-vars.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- thread-store-registry: makeStore now attaches a __testId via
intersection so callers can distinguish stubs at a glance during
debugging instead of relying on identity-by-allocation alone.
- thread-store-registry subscriber-isolation test: assert the
diagnostic content ("Subscriber onThreadStoreRegistered error") and
Error argument, not just that some error was logged.
- handle-threads identifyUser-throws test: assert
"Error identifying intelligence user" with an Error argument, since
the throw originates inside resolveIntelligenceUser which logs and
returns 500 before subscribeToThreads is reached.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Add the missing protected connect() override returning EMPTY so the
mock matches TestAgent and ThrowingAgent. Without it, a clone() that
ever exercised connect() would fall through to AbstractAgent.connect()
and may try to open a real transport in tests.
- clone() now forwards this.agentId directly instead of coercing
undefined to "". The constructor accepts string | undefined to make
this type-safe — coercion would silently turn "no agent id" into
"empty agent id", a different state per AgentConfig.agentId?: string.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
handleClearThreads is intentionally synchronous, but neither test
asserted the return type. Add `expect(response).not.toBeInstanceOf(Promise)`
on both branches so a regression that starts awaiting I/O updates the
synchronous call sites.
The handleGetThreadEvents/handleGetThreadState 501 tests asserted the
status code but not that intelligence stayed untouched. Stub spies for
both `listThreads` and a hypothetical `getThreadEvents`/`getThreadState`,
then assert neither was called — so a regression that drops the early
return and falls through to platform calls fails this test even after
the response code changes.
The handleSubscribeToThreads 500 test created an `errorSpy` to silence
console output but never asserted on it. Add `expect(errorSpy).toHaveBeenCalled()`
so a regression that quietly drops the diagnostic log is caught.
The InMemoryAgentRunner getThreadEvents test claimed the synthetic
terminal event was present but only asserted on TEXT_MESSAGE_*. Add an
explicit assertion on `RUN_ERROR` with `code: "INCOMPLETE_STREAM"` so
finalizeRunEvents' contract is locked in — a regression that stops
appending the synthetic event would leave the inspector showing an
in-progress thread forever.
Declare `onNewMessage` on `MessagePopulatingTestAgent.runAgent`'s options
type to match the runner's call site and `TestAgent` above. Without it,
a regression that starts depending on `onNewMessage` here would compile
cleanly even though the mock would silently drop the call.
handle-threads.test.ts:
- handleGetThreadMessages intelligence-path test now asserts the response
body verbatim, so a regression that swaps in a stub body is caught.
- 422-no-intelligence test issues a real DELETE request for the delete
path instead of cloning a POST request.
- handleClearThreads block carries a comment explaining why the handler
is intentionally synchronous (no I/O on either branch).
- The identifyUser-throws test now silences console.error for the
duration of the assertion, matching the pattern already used by the
subscribe-throws test.
in-memory-runner.test.ts:
- ThrowingAgent test asserts RUN_ERROR is the last emitted event and
that no RUN_FINISHED is emitted, locking in terminal-event semantics.
- getThreadEvents test asserts the full TEXT_MESSAGE triple is present
in the persisted event log, and the comment now reflects the real
finalizeRunEvents behaviour (it appends a synthetic terminal event
when the agent does not emit one, so terminal events ARE persisted).
- getThreadState multi-run test gains a cross-thread isolation
assertion: a snapshot on a different thread must not bleed into the
original thread's state.
- Bumped the inter-thread sort delay from 5ms to 20ms to absorb timer
jitter on slow CI runners.
- Removed four redundant `agent.agentId = "test-agent"` reassignments
(the constructor already sets it via super({ agentId })).
- Aligned MessagePopulatingTestAgent.runAgent with TestAgent: `onEvent`
is now required so the runner contract is exercised consistently.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
TanStack's chat() engine runs a multi-turn agent loop: after the model
returns tool calls, it tries to execute them via processToolCalls().
Frontend-only tools (like render_pie_chart) are unknown to TanStack, so
executeToolCalls() treats them as errors and buildToolResultChunks()
re-emits TOOL_CALL_END without a preceding TOOL_CALL_START. The ag-ui
verify middleware rejects this duplicate.
Fix: track a runFinished flag in convertTanStackStream and discard all
events after the first RUN_FINISHED, which marks the boundary between
the streaming pass and TanStack's internal tool execution loop.
Also adds built-in-agent to docker-compose.local.yml and local-ports.json.
Picks up ag-ui-protocol/ag-ui#1578 — `import * as jsonpatch from
"fast-json-patch"` produced an empty namespace under Node native ESM
because fast-json-patch@3.x populates exports via Object.assign, which
the CJS→ESM named-export detector cannot see. Result: every STATE_DELTA
and ACTIVITY_DELTA event threw "applyPatch is not a function", and
LangGraph generative UI streams floods the console with the failure on
each patch.
0.0.53 switches to a default import so the emitted bundle works under
both ESM and CJS consumers. Bumped @ag-ui/core and @ag-ui/encoder in
lockstep since they share the release.
## Summary
- Explicitly defaults `streamSubgraphs: true` in `LangGraphAgent.run()`
forwardedProps so subagent streaming works with `@ag-ui/langgraph`
0.0.31+, which changed the default from `true` to `undefined`
- Uses nullish coalescing (`??`) so explicit user overrides (including
`false`) are preserved
- Fixes 6 showcase integrations failing E2E probes on the `subagents`
feature
## Why
`@ag-ui/langgraph` 0.0.31 removed the `?? true` fallback on
`streamSubgraphs` in `handleStreamEvents()`. The CopilotKit runtime
never explicitly set this prop, so subgraph event forwarding became
silently disabled. This caused all subagent-dependent demos to stop
working.
A previous v1 fix (commit `21e12afca`) handled this in the now-deleted
`agui-action.ts`, but the logic was lost during the v2 migration.
## Additional context
Investigation identified a second issue: `@ag-ui/langgraph` 0.0.34
includes state snapshot fixes needed for shared-state-read/write
features, but 0.0.34 is not yet published on npm (0.0.31 is latest). The
ag-ui team needs to cut a release. Full analysis: [Notion
write-up](https://app.notion.com/p/3513aa381852812a9ecef5fbf1e71739)
## Test plan
- [x] Runtime tests pass (1412/1412)
- [x] 7-agent CR converged Round 1 (0 findings)
- [ ] Verify showcase E2E probes for subagent features turn green after
merge
Express 4 does not support the {*splat} wildcard syntax introduced in
Express 5. Replace string-based wildcard patterns with RegExp routes
that work across both major versions.
The @ag-ui/langgraph 0.0.31+ changed the default for streamSubgraphs
from true to undefined, breaking subagent streaming. This enriches the
run() input's forwardedProps with streamSubgraphs: true as a default,
while preserving any explicit user override via the nullish coalescing
operator.
Needed for turnIndex/sequenceIndex fixture matching in D5 multi-turn
conversations. Updated in both packages/runtime (devDep) and
showcase/scripts (dep).
Express 4.20+ uses path-to-regexp v8 which requires named wildcard
params. Bare * is no longer valid — use {*splat} syntax instead.
Updated production code (express.ts) and test (express-fetch-bridge).
Restores express >=4.20.0 override now that code is compatible.
Alem Tuzlak