The triage runtime connected its MCP clients with Promise.all, so a single
unreachable/misconfigured server (bad key, sidecar down, hang) rejected the
whole run and surfaced as a fatal "⚠️ Agent error: Failed to connect to
MCP server" — the bot was dead for every request, even ones needing no MCP.
Connect each server independently (Promise.allSettled) with an 8s timeout,
drop the ones that fail, and let the agent run with whatever's left (web search,
rendering, thread reading, and any MCP that did connect). A per-turn system note
lists any down sources so the model only tells the user a source is unreachable
if they actually ask for it, and never invents data. Connections are retried
each turn, so a transient outage self-heals.
Repairs TypeScript check-types across the monorepo and adds a CI gate so
regressions are caught going forward:
- core: bundler module resolution and strict-mode fixes
- sdk-js: bundler module resolution; keep codegen, formatter, packaging working
- react-core: fixes across components, hooks, and tests
- react-native: restore catch binding referenced by TypeError cause
- runtime: repair check-types and bound AI SDK schema inference
- web-inspector: nodenext import extensions, export Anchor
- remaining packages and node example: assorted check-types repairs
- deps: add missing type-only devDependencies
- license context driven from /info licenseStatus
- ci: run check-types in the static quality workflow
Squashed from 12 commits for a single, easily-revertable change.
## What does this PR do?
Routes the Gemini canvas stack analyzer through its registered `end`
node instead of directly to LangGraph's `END` sentinel.
This keeps the workflow wiring consistent with
`workflow.set_finish_point("end")` and ensures the cleanup/final state
emission in `end_node` is reachable.
## Related PRs and Issues
- Fixes#5605
## Tests
- `python3.12 -m py_compile examples/canvas/gemini/agent/stack_agent.py`
- `python3.12 - <<'PY' ... PY` (AST check that
`workflow.add_edge("analyze", "end")` exists and
`workflow.add_edge("analyze", END)` does not)
- `. /tmp/oss-pr-pipeline/langgraph-venv/bin/activate && python - <<'PY'
... PY` (LangGraph topology reproduction asserts `analyze -> end ->
__end__` and that `end_node` runs)
- `. /tmp/oss-pr-pipeline/langgraph-venv/bin/activate && python - <<'PY'
... PY` (imports `stack_agent.py` with CopilotKit/Gemini stubs and
asserts compiled `stack_analysis_graph` edges include `analyze -> end`
and not `analyze -> __end__`)
- `git diff --check`
## Checklist
- [x] I have read the [Contribution
Guide](https://github.com/copilotkit/copilotkit/blob/master/CONTRIBUTING.md)
- [x] If the PR changes or adds functionality, I have updated the
relevant documentation (not applicable: example graph wiring bug fix)
- [x] "Allow edits by maintainers" is checked (lets us help iterate on
your PR directly — faster turnaround for everyone)
Add per-feature demos to examples/slack that narrate per-platform degradation
explicitly rather than failing silently:
- emoji triage — 🐛/🔥/✅ reactions file/escalate/ack via the agent
- /preview — ephemeral draft issue (native only-you on Slack, DM fallback on
Discord/Telegram)
- /file-issue — modal form (Slack rich, Discord text-only, Telegram
conversational fallback)
Also updates the Slack frontend guide (slack.mdx) with the capability matrix.
The aisdk + tanstack agents expose a needsApproval bookFlight tool; /interrupts renders interrupts in-chat via useInterrupt(renderInChat) with a reusable InterruptCard and single/multiple suggestion pills. Requires OPENAI_API_KEY.
Adds a durable persistence layer for @copilotkit/bot, replacing the
in-memory-only ActionStore with a pluggable StateStore.
- StateStore interface (kv/list/lock/dedup/queue) with a shared
conformance suite; MemoryStore default plus @copilotkit/bot-store-redis
and @copilotkit/bot-store-postgres backends.
- createBot({ store }): typed per-thread state via Standard Schema,
action snapshots persisted through the store, per-conversation turn
lock (onLockConflict drop|force), and inbound-event dedup keyed on a
stable eventId. ActionStore is kept as a deprecated alias.
- Cross-platform transcripts (bot.transcripts + identity resolver) with
age-bounded retention (prune on append + filter on read), and
runAgent({ transcript: true }) to auto-inject history and capture the
reply.
- createBot({ components }) re-registers components so durable actions
re-fire after a restart; restart-durability demo in examples/slack.
- Dedup is marked seen only after the turn lock is acquired, so a turn
dropped on lock-conflict does not burn its eventId (no lost retries).
- Release lockstep: bot-store-redis/postgres version with bot + bot-ui.
The build script enumerated bot-slack/bot-discord/runtime, so it silently
omitted bot-telegram and bot-whatsapp — deploying with TELEGRAM_*/WHATSAPP_*
secrets would fail at runtime because those adapters' dist/ was never built
(start runs via tsx against the workspace packages' compiled output).
Use the nx project glob '@copilotkit/bot*' (+ runtime) so every bot adapter,
including any added later, is built without editing this script.
## What
Brings `@copilotkit/bot-slack` up to the **current** Slack native
streaming API surface (`chat.startStream` / `appendStream` /
`stopStream`, GA Oct 2025; structured chunks + AI feedback elements) and
removes the type-erasure workarounds. Result of an audit
cross-referencing the live `@slack/web-api@7.16.0` /
`@slack/types@2.21.1` types and Vercel's `vercel/chat` Slack adapter.
## Changes
- **No more `as unknown as Parameters<…>` casts** — every
streaming/post/update call uses the SDK's typed args.
- **One streamed message per turn** — dropped the per-message
continuation splitting (Slack documents only a 12k-per-append limit, no
cumulative cap; matches `vercel/chat`), keeping ≤12k per-append
chunking.
- **Native `task_update` tool-progress chunks** (`task_display_mode:
"timeline"`) interleaved into the streamed reply, replacing the separate
`🔧` status messages — with automatic degradation back to
`🔧` rows where structured chunks aren't supported.
- **Opt-in AI feedback buttons** via `slack({ feedback })` — a typed
`context_actions` + `feedback_buttons` row attached at `stopStream`;
clicks are routed adapter-locally (bypassing the engine's interaction
dispatch). No handler ⇒ no buttons.
- **Recipient scoping** — `recipient_user_id` / `recipient_team_id` only
for channel targets.
- **Cadence** — native flush floor lowered to ~600ms (appendStream
Tier-4); legacy `chat.update` stays 800ms.
### Engine (`@copilotkit/bot`)
One small, backward-compatible addition: optional
`RunRenderer.finish?()`, called after `runAgentLoop` resolves, so a
turn-scoped renderer can finalize its single stream. No-op for existing
adapters.
## Verification
- `bot-slack`: type-check (both tsconfigs) clean, **208 tests pass**,
oxfmt + oxlint clean.
- `bot`: type-check clean, **33 tests pass** (incl. a new `finish()`
test).
- Reviewed for correctness (stream lifecycle, finish/interrupt
interaction, delta tracking, degradation) — no high-confidence bugs;
feedback-on-interrupt and missing-ref-logging were tightened.
### Not verifiable without a live workspace (flagged in-code)
- That a >12k reply truly streams into one message (the
no-cumulative-cap assumption).
- `startStream` with no initial content.
## What
Removes the live-consumed local Intelligence overlay and the dead
references the deletion would leave behind.
- Deletes `examples/integrations/_intelligence/` (`docker-compose.yml`,
`.env.intelligence`, `README.md`) — the overlay the currently-shipped
CLI clones at runtime.
- Strips the now-dangling `# see
examples/integrations/_intelligence/.env.intelligence for the seed
value` pointer from 8 integration `.env.example` files (adk, agno,
langgraph-fastapi, langgraph-js, langgraph-python,
ms-agent-framework-{dotnet,python}, strands-python).
Closes ENT-834.
## ⚠️ DO NOT MERGE until launch
The **currently-shipped** CLI clones
`CopilotKit/CopilotKit@main:examples/integrations/_intelligence` at
runtime via `fetchIntelligenceOverlay` (hardcoded to `main`). Deleting
this dir from `main` **immediately breaks** the shipped CLI's
threads-framework `init` — the overlay fetch 404s.
**Merge gate (verify at merge time):**
- [ ] The managed-only CLI build has **dropped
`fetchIntelligenceOverlay`** (Intelligence-repo removal ticket)
- [ ] That managed-only CLI has been **released**
- [ ] Merge in lockstep with launch
## Scope
- ✅ **Included:** the `_intelligence/` overlay dir + the 8 dead
`.env.example` pointers it leaves behind.
- ✅ **Already done elsewhere:**
`examples/integrations/langgraph-python-threads/` (the ticket's "maybe"
scope) was already removed via ENT-800 — it is no longer on
`origin/main`.
- ⏭️ **Deferred to a launch-coordinated follow-up:** the local-stack
`INTELLIGENCE_API_URL`/`GATEWAY_WS_URL` defaults (`localhost:4201` /
`ws://localhost:4401`) baked into ~18 `copilotkit:intelligence` route
blocks + `agentcore/docker/docker-compose.yml`, and the matching
local-dev block in each `.env.example`. These depend on the managed
CLI's hosted env contract (Intelligence repo) and shouldn't be guessed
at now. The `copilotkit license` locked-state copy is owned by ENT-804.
## Not affected
The `docker-compose.test.yml` + `docker/Dockerfile.{agent,app}` files
across the integrations are the **e2e/CI test harness** for the example
apps, unrelated to the Intelligence pivot. The Threads feature, the
activation-gated `copilotkit:intelligence` block, and the themed
threads-drawer UI are the product and work against hosted Intelligence —
only the local-stack scaffolding is being removed.
## Notes
Branched off fresh `origin/main` (`c540734143`). No in-repo code
references the overlay dir (the `_intelligence` matches under
`packages/` are an unrelated private field on the agent registry).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
The build script only built bot-slack, bot-discord, and runtime, so
@copilotkit/bot-telegram (plus bot and bot-ui) were never compiled — the
Railway deploy then crashed on `import { telegram } from @copilotkit/bot-telegram`.
Build every workspace package the app imports.
Resolve event-renderer.ts onRunFinishedEvent: keep the native turn stream open
(finalized in finish()) AND retain the legacy per-message stream drain from main
(#5573) as a no-op-in-native safety net. app/index.ts (telegram adapter from
#5520 + showToolStatus:false) and create-bot.test.ts auto-merged.
Bump @tanstack/ai-openai 0.14.4 → 0.15.2 (pins @tanstack/openai-base 0.8.7,
TanStack/ai#790) and @tanstack/ai → 0.32.0. 0.8.7 emits strict:false for tool
schemas outside OpenAI's strict subset in the provider-path function-tool
converter, so MCP tools (e.g. Notion's API-post-search) no longer 400 — no
local patch needed. Verified end-to-end against the live Slack bot.
- runtime.ts: pass the bridge's forwarded client tools (convertInputToTanStackAI's
tools) into chat() alongside web_search + MCP, so generative-UI cards and the
confirm_write HITL gate work; drop the prompt line that made the model narrate
charts with a trailing "Charting …" sentence (it landed after the image).
- render-chart / render-diagram: post the caption as a header BEFORE the image
(a file upload's message lands a beat after postFile resolves, so caption-first
keeps a stable caption → image order) and drop the false "rendered above" wording.
Depends on the @copilotkit/runtime factory tool-lifecycle fix (#5572) and the
bot-slack HITL/render-order fix (#5573).
The example's runtime agent needed OpenAI's hosted `web_search` tool, but
BuiltInAgent's classic `tools` only accepts handler-based `ToolDefinition[]`
(needs `execute`) — it can't carry a provider/hosted tool. So switch the
agent to BuiltInAgent **factory mode** (`type: "tanstack"`) and drive it with
TanStack AI's `chat()`:
- `openaiText(model)` adapter (OpenAI Responses API; gpt-5.5 default)
- `webSearchTool({ type: "web_search" })` provider tool (`@tanstack/ai-openai/tools`)
- Linear/Notion MCP via `@tanstack/ai-mcp` `createMCPClient` (HTTP + bearer),
created per-run; `chat()` discovers their tools and closes the connections
- `convertInputToTanStackAI(ctx.input)` bridges AG-UI input → `chat()`;
BuiltInAgent converts `chat()`'s stream back to AG-UI events
OpenAI-only now (web search is OpenAI-specific); AGENT_MODEL accepts a bare
OpenAI id or an "openai/<id>" form. Adds @tanstack/ai, @tanstack/ai-openai,
@tanstack/ai-mcp to the example.
## What
1. Add `suppressHydrationWarning` to `<body>` across **all 14
integration demo templates**
(`examples/integrations/*/src/app/layout.tsx`).
2. Fix a pre-existing **double-escaped Windows path** bug in the parity
manifest's `packageJsonOverrides`.
## Why (hydration)
**Mike Ryan hit a hydration error on first load of a fresh
`langgraph-python` init — caused by his Grammarly browser extension.**
Grammarly (and similar extensions) inject attributes onto `<body>`
*before* React hydrates:
```
data-new-gr-c-s-check-loaded="9.98.0"
data-gr-ext-installed=""
```
Those attributes are in the client DOM but absent from the server HTML,
so Next.js reports:
> A tree hydrated but some attributes of the server rendered HTML didn't
match the client properties.
It's a **false positive** — the app works, and end users (without dev
extensions) never see it — but it's a red console error on the first
load of our flagship eval/showcase templates, which is a poor first
impression.
## Fix (hydration)
`suppressHydrationWarning` on `<body>` is the React/Next.js-recommended
escape hatch for this. It is **scoped and one level deep**: it only
relaxes the check for `<body>`'s *own* attributes/text — **everything
rendered inside `<body>` (the whole app) is still fully
hydration-checked** — and `<body>`'s only attribute here is a static
`className`, so none of our own markup is masked. An inline comment
documents this so a future maintainer who adds dynamic `<body>`
attributes knows the check is relaxed.
`agent-spec` already had `suppressHydrationWarning` on `<html>`; the
Grammarly attributes land on `<body>`, so it needed the body-level
relaxation too (the `<html>` one is a level up and doesn't cover
`<body>`'s attributes).
## Commits
1. `b1fa482a7` — north-star (`langgraph-python`) + parity instances
(`langgraph-js`, `langgraph-fastapi`, `strands-python`) via `pnpm
parity:sync`.
2. `9f9c415d9` — the non-parity templates (not tracked by
`_parity/manifest.json`): `adk`, `agno`, `crewai-crews`, `crewai-flows`,
`llamaindex`, `mastra`, `ms-agent-framework-dotnet`,
`ms-agent-framework-python`, `pydantic-ai`, `agent-spec`. *(The repo's
`oxfmt` pre-commit hook also collapsed some multiline `<CopilotKit …>`
JSX in these files — standard auto-format on touched files; the only
semantic change is the suppression.)*
3. `7d60e49de` — parity manifest path-escaping fix (see below).
## The manifest bug (commit 3)
While syncing I found the `langgraph-js` and `strands-python`
`packageJsonOverrides` double-escaped the Windows `.bat` fallback,
producing `scripts\\run-agent.bat` (two backslashes) instead of
`scripts\run-agent.bat`:
- `langgraph-js/package.json` had already been synced with the broken
value.
- `strands-python/package.json` was still correct — and `parity:sync`
would have **corrupted** it on the next run (which is what surfaced
this).
Fixed the three overrides and re-ran `parity:sync`, which corrects
`langgraph-js/package.json` and leaves `strands-python`'s correct value
intact.
## Test plan
- [x] `pnpm parity:verify` → 0 errors
- [x] lefthook pre-commit green on all 3 commits (lint + `packages/**`
tests + commitlint)
- [x] All 14 templates confirmed to have body-level
`suppressHydrationWarning`
- [ ] Reviewer with Grammarly installed: run/`init` a template and
confirm no hydration error on first load
The langgraph-js and strands-python packageJsonOverrides in
_parity/manifest.json double-escaped the Windows .bat fallback, so the
synced value became `scripts\\run-agent.bat` (two backslashes) instead of
the intended `scripts\run-agent.bat`. langgraph-js's package.json had
already been synced with the broken value; strands-python's was still
correct (and parity:sync would have corrupted it on the next run).
Fix the three overrides and re-run parity:sync, which corrects
langgraph-js/package.json and leaves strands-python's correct value intact.
Extend the same `<body suppressHydrationWarning>` fix to the integration
templates that are not tracked by examples/integrations/_parity/manifest.json,
so they don't surface a Grammarly-style hydration mismatch on first load:
adk, agno, crewai-crews, crewai-flows, llamaindex, mastra,
ms-agent-framework-dotnet, ms-agent-framework-python, pydantic-ai, agent-spec.
agent-spec already had suppressHydrationWarning on <html>; the Grammarly
attributes land on <body>, so it needs the body-level relaxation too (the
<html> one is one level up and does not cover <body>'s attributes).
The opengenui route used the nonexistent model id "openai/gpt-5.2", causing
every chat turn to error. Switch to "openai/gpt-4o", the verified-real id used
by the sibling showcase/shell copilotkit route.
The example's API routes import @copilotkit/runtime/v2 (a server-only package),
but next.config.ts lacked serverExternalPackages, so Next.js attempted to bundle
it. Add serverExternalPackages: ["@copilotkit/runtime"] (base package name covers
the /v2 subpath), mirroring showcase/shell/next.config.ts.
Browser extensions such as Grammarly inject attributes onto <body>
(data-gr-ext-installed, data-new-gr-c-s-check-loaded) before React
hydrates, which surfaces as a hydration mismatch error on first load of
the generated Next.js app.
Add suppressHydrationWarning to <body> in the langgraph-python north-star
and propagate to the parity instances (langgraph-js, langgraph-fastapi,
strands-python) via parity:sync. This only relaxes the check for <body>'s
own attributes (one level deep); everything rendered inside <body> is
still fully hydration-checked, and <body>'s className is static so none of
our own markup is masked.
## Summary
- Update locked Threads drawer copy in integration examples to use `npx
copilotkit@latest license`
- Keep the command consistent across the example variants that render
the licensed feature panel
## Validation
- Pre-commit hooks ran lint/check package hooks successfully
- Verified integration examples no longer render the stale `copilotkit
license` command
1.61.0 is now published. Regenerates all 20 starter package-lock.json files
so @copilotkit/* resolves to 1.61.0 and the transitive
@copilotkit/license-verifier moves 0.4.2 -> 0.5.0 (shipped by runtime@1.61.0).
ENT-939
Prep for the 1.61.0 release. Pins every @copilotkit/* dependency in the
examples/integrations starter projects to exact 1.61.0 (not yet published).
ENT-939
`showToolStatus` only gated the legacy `🔧` rows — the native
`task_update` chunks and the pane "is using `tool`…" composer status
ignored it, so there was no single switch to hide tool-call progress.
Promote `showToolStatus` to the master toggle: when `false`, tool progress
is suppressed on ALL surfaces (native chunks, legacy rows, pane status);
tools still run, only the display is hidden. When `true`, the surface is
still chosen by target (native chunks / legacy rows / pane status, the
latter further gated by the pane's own `toolStatus`).
Flip it off in the slack example (`showToolStatus: false`).
The deleted examples/integrations/_intelligence/ dir was referenced by a
"see .../.env.intelligence for the seed value" comment on the
INTELLIGENCE_API_KEY line in 8 integration .env.example files. Removing the
overlay leaves those pointers dangling, so strip them in the same PR.
Localhost INTELLIGENCE_API_URL/GATEWAY_WS_URL defaults are intentionally left
in place — the hosted-only rewrite of those is a launch-coordinated follow-up
(depends on the managed CLI env contract).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Removes examples/integrations/_intelligence/ — the docker-compose +
.env.intelligence + README overlay that the currently-shipped CLI clones
at runtime from CopilotKit@main via fetchIntelligenceOverlay.
DRAFT — do NOT merge until the managed-only CLI build has dropped
fetchIntelligenceOverlay (Intelligence-repo removal) AND been released.
Deleting this dir from main before then breaks the shipped CLI's
threads-framework init (overlay fetch 404s). Merge in lockstep with launch.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The example pinned @ai-sdk/mcp to 1.0.21 (protocolVersion incompat, see
88a2d82) via its own pnpm.overrides. That only took effect when the example
was installed in isolation; as a workspace member pnpm ignores package-level
overrides, so the pin was silently dropped — packages/runtime's `^1.0.21`
could drift to a newer, incompatible 1.x on the next lockfile regen.
Move the override to the root package.json's pnpm.overrides (runtime is the
only consumer, so this enforces exactly 1.0.21 with no wider impact) and
remove the now-dead override from the example (also silences the pnpm warning
that surfaced once the example became a workspace member).
`examples/slack/pnpm-lock.yaml` only existed for the old isolated deploy
(root dir `/examples/slack`, `pnpm install --ignore-workspace
--frozen-lockfile` resolving the @copilotkit/* deps from npm). Now that the
example is a workspace member built from source (`workspace:*`, root-dir
`/`), pnpm uses the single root `pnpm-lock.yaml`; the per-example lock is
never consulted and was left stale — it still pins the published `~0.0.2`
versions, which contradicts the `workspace:*` package.json and would break
any `--frozen-lockfile` install.
The example declared its sibling @copilotkit/* packages as npm version
ranges, so the Railway service (which builds examples/slack in isolation)
resolved them from the registry — forcing a "publish first, then bump the
example" dance on every PR, with a broken deploy window in between.
Switch those deps to the workspace:* protocol (the example is private, so
it never affects publishing) so the example always builds from in-repo
source, and add a graph-aware `build` script that compiles the workspace
libs it imports (and their deps) via Nx. README documents the Railway
settings (root dir / build / start / watch paths) and the copy-out caveat.
Result: a packages/** change redeploys the example with the new code
immediately, and npm publishing becomes an independent manual step.
Alem Tuzlak