Upgrade to CopilotKit 1.77 and AG-UI 1.0

[davidmckayv]

Moves OpenBot onto CopilotKit 1.77.0 (from 1.73.1) and AG-UI 1.0. The Intelligence client ships inside @copilotkit/runtime/v2, so this is also OpenBot's side of the latest Intelligence; a managed project needs no change.

Versions

• @copilotkit/runtime, @copilotkit/react-core, @copilotkit/a2ui-renderer: 1.73.1 → 1.77.0
• @copilotkit/aimock: 1.39.0 → 1.43.0
• @ag-ui/core, @ag-ui/client: 0.0.59 → 1.0.1, matching the exact versions runtime 1.77 pins (CopilotKit moved to AG-UI 1.0 in 1.76, #7270). The server now depends on @ag-ui/core directly for @ag-ui/core/schemas.
• @ag-ui/mastra: 1.1.2 → 1.1.6 (peers >=0.0.58)

@copilotkit/channels-* 0.11.0, which runtime 1.77 depends on, still pins AG-UI 0.0.59, so a second copy remains in the lockfile. That is upstream; OpenBot's own imports all resolve to 1.0.1.

Code changes for AG-UI 1.0

• Schemas (RunAgentInputSchema, EventSchemas, MessageSchema) now come from @ag-ui/core/schemas.
• The binary content part is gone. The plan model builds 1.0 image parts directly, so the shim that converted binary parts for the Python endpoint is removed. The end-to-end test that the image reaches the endpoint as an image still passes.
• Tool message content can be content parts; the hand-off resume reads it as text.
• TOOL_CALL_RESULT has no error field in 1.0 and validation strips it, so headless tools stop sending one. The refusal remains in the content and on the tool message. Nothing read the event's field.
• A copilot.test.ts fixture built a remote agent without the standing message every registered agent has; 1.0's client no longer tolerates the resulting undefined message.

Verification

• typecheck, lint, format and build pass.
• bun run test:ci against fresh pgvector databases: this branch and origin/main (7229c68) both report 5,924 pass / 13 fail with the identical failing set (local-environment only: agent-bot startup ×3, real-provider learned catalog ×2, Mastra bridge). Before the two fixes above, the branch had two extra failures; both are covered.
• Live, against managed Intelligence, in the browser: a General Assistant chat streamed its reply; reloading restored the thread from Intelligence; a run opened Hacker News on the Bot's computer and drew a bar chart; Admin → Automatic Learning loaded; /api/capabilities resolved selfHostBanner. No AGUIError or RUN_STARTED errors in the server log or the browser console.