feat(vercel-ai-sdk): rewrite for AI SDK v6 with full event coverage#1626
Open
MyPrototypeWhat wants to merge 8 commits into
Open
feat(vercel-ai-sdk): rewrite for AI SDK v6 with full event coverage#1626MyPrototypeWhat wants to merge 8 commits into
MyPrototypeWhat wants to merge 8 commits into
Conversation
Upgrades the Vercel AI SDK integration from v4 to v6 and re-enables it in the Dojo. This rewrite goes well beyond a pure version bump: - Streams tool input deltas: maps `tool-input-delta` → incremental `TOOL_CALL_ARGS` events so clients can pre-fill UI before the model finishes speaking (the AG-UI killer feature for tool calls). - Full text/reasoning lifecycle: emits `TEXT_MESSAGE_START/CONTENT/END` using AI SDK's part ids directly. Same for `REASONING_*`. Forwards Anthropic encrypted reasoning signatures via `REASONING_ENCRYPTED_VALUE`. - Defensive synthesis for non-streaming providers: if a model emits a final `tool-call` without prior `tool-input-start`, the handler synthesizes the START/ARGS/END trio from the parsed input. - Multi-step (`stopWhen: stepCountIs(N)`): each step gets a fresh `currentStepAssistantId` used for `parentMessageId` on tool calls. Lazy push avoids trailing empty assistant messages. - Cleanup-phase synthesis: any tool call that finishes the run without a result gets a synthesized `TOOL_CALL_RESULT` so clients don't dangle. - Proper error/abort semantics: stream-internal `error` parts emit `RUN_ERROR` but the stream continues; thrown errors emit `RUN_ERROR` then complete; `abort` parts close gracefully via `abortSignal` propagation through Observable teardown. - Replaces the 80-line hand-rolled JSON-Schema → Zod converter with the built-in `jsonSchema()` helper from `ai`. Supports any JSON Schema including `oneOf`/`anyOf`/`enum`/`pattern`. - Splits the previous single-file implementation into `tool-converter.ts`, `message-converter.ts`, `stream-handler.ts`, and `vercel-ai-sdk.ts` for readability and unit-testability. - 70 tests across 8 files using `MockLanguageModelV3` from `ai/test`, covering text, reasoning, tool streaming, tool dedup, invalid tool-calls, multi-step, abort, error part, defensive close, and cleanup synthesis. Resolves ag-ui-protocol#933.
- README Quick Start / With Tools: pass messages via `initialMessages` in constructor (RunAgentParameters does not accept `messages`). - README step counter examples: first step is `step-1` (StreamHandler increments before emitting), not `step-0`. - StreamHandler: log a console.warn for unrecognized AI SDK stream part types instead of silently skipping, surfacing future v6 part additions. - VercelAISDKAgent: catch unforeseen StreamHandler errors and surface them via subscriber.error() instead of leaking as unhandled rejections.
… & abort Previously, both the `error` and `abort` stream-part handlers fell through to the cleanup phase and emitted MESSAGES_SNAPSHOT + RUN_FINISHED after the run had already failed. AG-UI consumers treat RUN_FINISHED and RUN_ERROR as mutually exclusive terminal events, so a consumer that tore down state on RUN_ERROR was confused by the trailing RUN_FINISHED. - `error` part: emit RUN_ERROR (code: stream_error_part) and complete(). - `abort` part: emit RUN_ERROR (code: aborted) and complete() instead of silently degrading into RUN_FINISHED. Mirrors the thrown-error path. - Drop the now-unused `cancelled` flag — the for-await loop exits via `subscriber.closed` once `complete()` runs. - Rewrite the abort test with a deferred promise instead of racing a 200 ms / 50 ms setTimeout pair, removing CI flake potential. The unsubscribe-driven abort path (the only real-world abort source in this code) is unaffected: by the time the AI SDK emits the `abort` part, the subscriber is already closed and the `emit()` guard swallows the synthesized RUN_ERROR — which is exactly what the existing `vercel-ai-sdk.test.ts` "propagates abort on unsubscribe" test asserts. All 70 tests pass; tsc --noEmit clean.
… in unsubscribe test - stream-handler.ts: `onToolCall` no longer appends to `assistantMessage.toolCalls` when an entry with the same id is already present. The downstream START/ARGS/END events were already deduped via `seenToolCalls`, but the toolCalls array push was unconditional, so a misbehaving provider that emitted the same `tool-call` twice would produce a malformed assistant message in MESSAGES_SNAPSHOT. - vercel-ai-sdk.test.ts: rewrite "propagates abort on unsubscribe" with a deferred promise instead of a `setTimeout(200)` race + `setTimeout(1000)` fallback. Suite duration drops from ~1.36s to ~490ms; this single test goes from 1014ms to 21ms. Removes the last wall-clock-based timing assertion in the package. 70 tests still pass; tsc --noEmit clean.
|
@MyPrototypeWhat is attempting to deploy a commit to the CopilotKit Team on Vercel. A member of the Team first needs to authorize it. |
Author
|
Hi @NathanTarbert — picking this back up since #933 has been in Roadmap for a while. The Dojo wiring is included ( Happy to address feedback and to maintain the integration going forward, per the CONTRIBUTING.md note on community integrations. Also flagged the existing #1167 author as a courtesy. |
8 tasks
Author
|
Hi @contextablemark @tylerslaton — sorry to ping cold. This v6 rewrite (#933) has been open for ~2 weeks without a reviewer assigned. Could one of you take a look or route it to the right person? Two things:
Thanks! |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Rewrites the Vercel AI SDK integration for AI SDK v6 with full AG-UI event coverage. Resolves #933.
This is an independent implementation, not based on #1167 (which has been waiting on review for a while). See #1167 for context.
What this PR adds vs the v4 version on main
tool-input-delta→ incrementalTOOL_CALL_ARGSevents so clients can pre-fill UI as args arriveREASONING_ENCRYPTED_VALUEstopWhen: stepCountIs(N)) with per-step assistant-message rotation; tool calls link back viaparentMessageIdtool-call→ fabricatedSTART/ARGS/END) and missing tool results (cleanup-phase synthesis)RUN_ERRORis terminal for stream-internalerror/abortparts (no misleadingRUN_FINISHEDafter a failed run)jsonSchema()helper (full JSON Schema support —oneOf/anyOf/enum/pattern— for free)tool-converter.ts/message-converter.ts/stream-handler.ts/vercel-ai-sdk.tsMockLanguageModelV3fromai/testcovering text, reasoning, tool streaming, tool dedup, invalid tool-calls, multi-step, abort, error part, defensive close, and cleanup synthesisTest plan
pnpm testinintegrations/vercel-ai-sdk/typescript— 70 tests pass (~490ms)pnpm typecheck— cleanagentic_chatscenario re-enabled and wired (apps/dojo/src/agents.ts,menu.ts,scripts/generate-content-json.ts)Coordination
Notified @iamrajhans on #1167.