Tolerating unrecognized refresh errors everywhere meant a dead token on a
provider without a handleErrors override was never flagged: the refresh
that follows a rejected post went through the same code and stayed silent.
The refresh service now takes a tolerateUnknownErrors flag, off by default.
The refreshToken activity turns it on only when the calling workflow is
refreshTokenWorkflow, read from the activity context, so post workflows of
every version (including the still-running v1.0.1 runs that share this
activity) keep disconnecting after a rejected post.
Also add 'disconnect' to the optional handleErrors interface type, which the
base class returns since this branch was opened.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Add POST /public/v1/clipping, GET /public/v1/clipping and GET /public/v1/clipping/:id to the public API, calling the existing ClippingService.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Turns a YouTube video into captioned vertical clips that land in the media
library and as draft posts. Adds the Clipping / ClippingClip models, the
clipping_minutes credit type, clippingWorkflow + per-clip child workflows,
the /clipping REST routes, the MCP clipping tools and the ui://postiz/clipping
status widget.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
MCP has no way to send a file from the user's device, so add an MCP Apps
(SEP-1865) widget: uploadWidgetTool renders ui://postiz/upload, the widget
gets a short-lived ticket through the app-only uploadWidgetTicketTool and
posts the file to /media-widget/upload, then reports the media back to the
model. uploadWidgetStatusTool is the host-independent fallback.
The widget tools are mcpOnly, so the in-app agent keeps its current toolset.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Mastra core 1.67 / mcp 1.18 / memory 1.30 / pg 1.25 / cli 1.30,
CopilotKit 1.72 and @ag-ui/mastra 1.1.4.
CopilotKit 1.72 adaptations: /copilot/agent uses the node-http handler,
explicit CORS on both copilot routes, integrations read from the
single-route envelope (forwardedProps), root type imports, interrupt via
useCopilotChatInternal, multimodal message content, useSingleEndpoint on
the providers. Pin @types/express v5 and declare runtime-client-gql.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The CustomerModal rendered `@mantine/core`'s `<Autocomplete>`, which triggered a
React 19 console error ("Accessing element.ref was removed in React 19. ref is now
a regular prop.") whenever the modal opened (e.g. adding a channel to a new group).
The root cause is `@mantine/core@5.10.5` (Mantine v5, built for React 17/18) running
under React 19, where the library internally accesses the now-removed `element.ref`.
Replaced the single Autocomplete usage with a native `<input>` + `<datalist>`,
styled to match the app's inputs. Free-text entry (needed to create a new
customer/group) and existing-name suggestions are preserved. Scoped entirely to
customer.modal.tsx; no dependency or behavior changes elsewhere.
The alternative was to upgrade @mantine/core to a React 19-compatible major (v8),
which was clearly unfavorable: Mantine is used in only two files, the project's
convention is to prefer native components over npm UI libraries, and a v5->v8 bump
is a large breaking change carrying far more risk than removing one component.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
RFC 6749 搂2.3.1 requires the token endpoint to accept HTTP Basic client
authentication for clients that hold a secret. Until now only
client_secret_post worked, so a client sending the id and secret in the
Authorization header failed validation before the code was checked.
- extractBasicCredentials() next to extractBearerToken() decodes the header
- the token controller merges header credentials with the body
- registration mirrors back client_secret_basic when a client asks for it
- both discovery documents advertise client_secret_basic
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The workflows pass the platform's post id to sendWebhooks, so the
DB-id lookup matched nothing and every webhook body was [].
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JkiuuMQcG1QRkbagfSkBtb
refreshChannel redirected to the url returned by
/integrations/social/:id?refresh=..., which for custom-fields providers
(Bluesky etc.) is the random state string rather than an OAuth page, so
the click landed on a 404. Open the CustomVariables modal the channel
menu already uses for those providers instead.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PyDRJpL27E7WCK5UGsUope
showMediaBox() rendered MediaBox through ShowMediaBoxModal as a plain
in-flow div mounted near the top of the layout tree, with no overlay, no
positioning and no z-index. Every other caller opens MediaBox through the
modal manager instead, which is why the picker only misbehaved here.
Since BotPicture is itself a fixed, full-viewport modal at z-index 200,
the picker was laid out at the top of the document underneath it, and the
modal manager force-applies overflow:hidden to body/html while a modal is
open, so it could not even be scrolled to. Slack channels were the only
ones affected in practice, as Slack is the only provider implementing
changeProfilePicture, and the bot picture could not be changed at all.
Route showMediaBox through modals.openModal with the same options
MultiMediaComponent already uses, so the picker gets its own stack entry
above the dialog that opened it. Its confirm button calls
modals.closeCurrent(), which now closes the picker alone and leaves the
Bot Picture dialog open with the chosen image applied; previously it ran
outside any CurrentModalContext and closed nothing. MediaBox itself is
unchanged, and the callback is narrowed to the picker's first selection
to match the single-item shape the caller expects.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
After the managed-signer migration, clicking Farcaster on the login and
signup pages opened a new tab to the approval link, which only redirects
to the farcaster:// scheme and stays blank on desktop, while the QR was
squeezed into the narrow provider button slot.
The approval view now opens straight in a modal on the login and signup
pages and replaces the intermediate button step in the add/reconnect
channel modal. No tab opens; a Copy Farcaster link button replaces the
Open in Farcaster link. The auth layout mounts the shared modal manager
and the toaster, so errors on those pages are visible. A response
without an approval link (for example the rate limit) now shows the
failure toaster instead of leaving the modal empty, and the signer
creation limit is raised from 10 to 30 per hour per address.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8gvBNgPqqayC9zGPB5R5a
GET /auth/farcaster/signer is unauthenticated and calls Neynar on every
request. Apply the same per-address guard as the create route, with a
ceiling of 1000 per hour so the modal's 2s poll over a 10 minute window
still fits several approvals.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8gvBNgPqqayC9zGPB5R5a
POST /auth/farcaster/signer is unauthenticated and creates a signer at
Neynar on every call. The global throttler guard only covers the public
posts API, so add a route-level guard keyed by the forwarded client
address and cap the route at 10 requests per hour per address.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8gvBNgPqqayC9zGPB5R5a
A thrown fetch inside the async generator ended the for-await loop and
left the modal spinning with no timeout. Catch it and keep polling, as
the Moltbook status poll does; the 10 minute cap still applies.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8gvBNgPqqayC9zGPB5R5a
Neynar retired Sign In With Neynar, so the hosted popup used for Farcaster
login and channel connect no longer completes. Replace it with Neynar managed
signers: the backend creates a signer, signs the EIP-712 signed key request
with the app custody mnemonic (viem), registers it with Neynar and renders the
approval deep link as a QR (qrcode) since the link is a bare 302 to the
farcaster:// scheme. The frontend polls the new public /auth/farcaster/signer
routes and builds the same base64 blob the popup produced, so authenticate(),
the auth provider, the stored token format and existing users/channels are
unchanged. @neynar/react is removed.
Also thread the oauth_state nonce through the Farcaster login button: since
the Apple login change, checkExists rejects logins without a matching state
cookie, which broke Farcaster login independently of the SIWN retirement.
New env vars: NEYNAR_APP_FID, NEYNAR_APP_MNEMONIC, NEYNAR_SPONSOR_SIGNERS.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01R8gvBNgPqqayC9zGPB5R5a
Restores a 3 minute heartbeatTimeout on postSocialPending, finalizePost and
postComment and retries at the workflow level when the heartbeat timeout
fires with no heartbeat ever received, which means the worker never ran the
activity. Comment proxy drops SDK retries; exhausted retries now mark the
post unconfirmed instead of leaving it in QUEUE.