Aayush Kapoor
90d · built 2026-11-06
Performance
What Aayush Kapoor shipped in the selected window, measured in ETV, and how it compares with the 90 days before it.
Effective capacity
+3.7engineers
delivers like 4.7 (4.7x pre-AI)
Output (ETV)
50.3ETV
+74.5% vs 28.8 prior
Features share
34.9%
+4.9 pp vs prior window
Fixes share
12.1%
−6.9 pp vs prior window
Work mix
34.9% Features8.6% Maintenance33.3% Tests11% Docs12.1% Fixes
165 commits over 90 days, ending 2026-11-06.
Where this dev ranks
Percentile against the global top-100 leaderboard (all-time totals).
- By commits
- Top 64 %
- By Features share
- Top 59 %
Daily performance
Daily ETV, stacked by Features, Maintenance, Tests, Docs and Fixes.
Repository spread
Where this developer's commits land. Concentrated work (top1 > 80%) vs polymath spread (top1 < 30%).
Most impactful commits
Top 10 by ETV in the last 90 days.
- 2.4ETVfeat(workflow): trace durable agent runs as a single operation (#24408) ## Background Global telemetry integrations were available inside model steps, but invisible to the isolated workflow body. Each model call produced its own root operation. A multi-step agent run therefore appeared as several disconnected traces. There was no span representing the entire run with its original prompt, final output, and tool executions. Keeping a root span in memory wouldn’t solve this because workflow steps can execute in different processes https://github.com/vercel/ai/issues/23618 ## Summary - The workflow records its lifecycle events and timestamps - A preparation step captures trace context and sampling, then passes that context into model steps. - Model and provider HTTP spans are emitted live; the enclosing root, step, and tool spans are reconstructed when the run ends or suspends. - The bridge supplies matching IDs so everything joins one tree ## End-to-End Verification verified by running the reproduction: <details> ```ts import { OpenTelemetry, OpenTelemetryIdGenerator } from '@ai-sdk/otel'; import { trace } from '@opentelemetry/api'; import { BasicTracerProvider, InMemorySpanExporter, SimpleSpanProcessor, } from '@opentelemetry/sdk-trace-base'; import { buildWorkflowTests, getWorkflowRef, setupWorkflowTests, teardownWorkflowTests, } from '@workflow/vitest'; import { registerTelemetry } from 'ai'; import { mkdir, mkdtemp, rm } from 'node:fs/promises'; import { dirname, join } from 'node:path'; import { fileURLToPath } from 'node:url'; import { start } from 'workflow/api'; import { run } from '../lib/run'; run(async () => { const exporter = new InMemorySpanExporter(); const durable = new OpenTelemetryIdGenerator(); const provider = new BasicTracerProvider({ idGenerator: durable, sampler: durable, spanProcessors: [new SimpleSpanProcessor(exporter)], }); trace.setGlobalTracerProvider(provider); registerTelemetry(new OpenTelemetry({ durable })); const cwd = dirname( fileURLToPath(import.meta.resolve('@ai-sdk/workflow/package.json')), ); const artifactsDir = join(cwd, '.workflow-vitest'); await mkdir(artifactsDir, { recursive: true }); const rootDir = await mkdtemp(join(artifactsDir, 'issue-23618-')); const options = { cwd, rootDir }; const originalCwd = process.cwd(); try { // Keep the compiler's SWC cache in the workflow package's ignored directory. process.chdir(cwd); await buildWorkflowTests(options); await setupWorkflowTests(options); // Reuse the workspace fixture: a model calls addNumbers, then returns // "The sum is 5". Calling it through start() exercises workflow isolation. const workflowRun = await start( getWorkflowRef('agentToolLoopWithGlobalTelemetry'), [2, 3], ); console.log('Result:', await workflowRun.returnValue); await provider.forceFlush(); const spans = exporter .getFinishedSpans() .filter(span => span.instrumentationScope.name === 'gen_ai'); function printTree(parentId: string | undefined, depth = 0) { for (const span of spans.filter( candidate => candidate.parentSpanContext?.spanId === parentId, )) { console.log(`${' '.repeat(depth)}${span.name}`); printTree(span.spanContext().spanId, depth + 1); } } printTree(undefined); const root = spans.find(span => span.name.startsWith('invoke_agent ')); console.log('Root prompt:', root?.attributes['gen_ai.input.messages']); console.log('Root output:', root?.attributes['gen_ai.output.messages']); console.log( 'Trace count:', new Set(spans.map(span => span.spanContext().traceId)).size, ); } finally { process.chdir(originalCwd); await teardownWorkflowTests(); await provider.shutdown(); trace.disable(); await rm(rootDir, { recursive: true, force: true, maxRetries: 5 }); } }); ``` </details> ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) ## Related Issues fixes #23618 closes #24334github.com-vercel-ai · 4d6d830c · 2026-11-02
- 2.3ETVfeat(providers): support async credential resolution (#24145) ## Background Provider credentials mostly accepted static strings, requiring custom fetch wrappers or provider recreation for asynchronous lookup and key rotation follow up to https://github.com/vercel/ai/pull/24096 ## Summary - Accept promises and synchronous/asynchronous credential callbacks, resolved per request while preserving static authentication and workflow serialization. - Make experimental OpenAI Live connection settings awaitable ## End-to-End Verification na ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [ ] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review)github.com-vercel-ai · 7716e189 · 2026-10-29
- 1.3ETVfeat(mistral): add realtime voice agent support (#23810) ## Background follow up to https://github.com/vercel/ai/pull/23802 The Mistral provider did not support managed voice conversations through the SDK’s realtime API. This builds on the core provider-owned transport support ## Summary - Add `mistral.experimental_realtime` with server-side session creation and cleanup - Handle microphone input, voice playback, transcripts, and client tool calls through a LiveKit transport contained within the Mistral provider ## End-to-End Verification verified by running the `http://localhost:3000/mistral-realtime` ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review)github.com-vercel-ai · 0912e0b1 · 2026-11-03
- 1.3ETVfeat(mcp): add compatibility for webhook based mcp events (#22155) ## Background openai recently launched support for [MCP events](https://developers.openai.com/plugins/build/mcp-events) and there is a working design schema here https://modelcontextprotocol.io/community/working-groups/triggers-events our MCP client is a great place to listen and respond to server events that are being sent. ## Summary add experimental apis and functions to extend the existing MCP client that will allow users to: - subscribe and unsubscribe to events from the server - use a specific `experimental_createMCPEventWebhook` to verify and handle the events sent to the webhook route from the mcp server - have an event store that will be a shared state between the client app and the webhook route - including callback URLs, signing secrets, cursors and expiration details ## End-to-End Verification verified by running the events example in `example/mcp` ``` pnpm server:events ``` and in a new terminal ``` pnpm client:events ``` ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review)github.com-vercel-ai · ac505482 · 2026-10-07
- 1.3ETVfix: report interrupted streams without completion signals (#23814) ## Background https://github.com/vercel/ai/issues/23449 follow up to #23778 providers could end mid-response without reporting that generation was interrupted. ## Summary - Report `interrupted` when a stream ends cleanly without its completion signal, preserving partial output - Cover Anthropic multi-message streams and Google continuations - Preserve actual error handling and OpenAI-compatible’s existing behavior ## End-to-End Verification na ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [ ] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) ## Related Issues fixes #23449github.com-vercel-ai · 43ad966e · 2026-10-21
- 1.2ETVfeat(openai): centralize model capabilities in a unified catalog (#24174) ## Background model-specific behavior was distributed across request builders and documentation tables, creating duplicated capability definitions that could drift. There was also no consistent structure for representing capabilities across language, image, speech, transcription, embedding, and realtime models. Updating a model therefore required finding and synchronizing several unrelated implementations ## Summary Introduces a shared model capability contract and consolidates openai provider's model metadata into one catalog. Runtime request handling and generated documentation now consume the same entries, while provider-specific request encoding remains separated under requestConfig ## End-to-End Verification na ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [ ] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) ## Future Work align all the providers on this specgithub.com-vercel-ai · dcd2f8d8 · 2026-10-29
- 1.2ETVfeat(ai): add support for interrupted finish reasons (#23778) ## Background https://github.com/vercel/ai/issues/23449 core has no finish reason for provider streams that end without a completion signal. historically if a stream ends without sending it's completion signal, the providers would map the finish reason as `other` and would offer no visibility into what actually happened provider's like openai-compatible would simply throw an error ## Summary - add an `interrupted` to the shared finish reasons - Preserve partial output without raising a required-tool error - Prevent tool execution, agent continuation, and automatic UI follow-ups ## End-to-End Verification na ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) ## Related Issues Need provider mapping for the providers in the original issue to use this new finish reason ## Related Issues towards #23449github.com-vercel-ai · 10026d15 · 2026-10-21
- 0.9ETVfeat(perplexity): add decisions api support (#23470) ## Background perplexity has native decisions api support https://docs.perplexity.ai/docs/decisions/quickstart this pr serves as an alternative to https://github.com/vercel/ai/pull/22185 ## Summary - `perplexity.decisionModel()` with documented model IDs - Boolean (noul), Choice, and Score with full-precision probabilities - PNG/JPEG/WebP images, including images nested in JSON - instructions is now optional or nullable for Choice, Score, and Boolean - Perplexity accepts single-level Score rubrics so the provider docs were changed ## End-to-End Verification verified by running the examples in this PR - `examples/ai-functions/src/decide/perplexity/basic.ts` - `examples/ai-functions/src/decide/perplexity/criteria-only.ts` - `examples/ai-functions/src/decide/perplexity/single-level-score.ts` ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) Co-authored-by: Hernan Alvarado <hernanvid123@gmail.com>github.com-vercel-ai · 60cc2527 · 2026-10-16
- 0.9ETVfix(openai): parse the last final answer for non-streaming structured output (#23895) ## Background OpenAI Responses can return multiple assistant messages containing JSON. Concatenating their text causes structured-output parsing to fail, even when the final answer is valid. https://github.com/vercel/ai/issues/23832 ## Summary - Select the last `final_answer` for non-streaming JSON parsing through a separate provider-supplied parser input - Support `generateText` and `generateObject`, including built-in JSON/reasoning middleware - Preserve all existing text, content, annotations, response history, and token usage. Streaming is unchanged ## End-to-End Verification verified by running this repro <details> ```ts import { createOpenAI } from '@ai-sdk/openai'; import { generateText, Output } from 'ai'; import { z } from 'zod'; const body = { id: 'resp_1', object: 'response', created_at: 1760000000, model: 'gpt-5.3-codex', status: 'completed', output: [ { type: 'message', id: 'msg_0', role: 'assistant', status: 'completed', phase: 'commentary', content: [{ type: 'output_text', text: '{"items":[]}', annotations: [] }], }, { type: 'message', id: 'msg_1', role: 'assistant', status: 'completed', phase: 'final_answer', content: [{ type: 'output_text', text: '{"items":[]}', annotations: [] }], }, ], usage: { input_tokens: 13, output_tokens: 7, input_tokens_details: { cached_tokens: 0 }, output_tokens_details: { reasoning_tokens: 2 }, }, incomplete_details: null, }; const openai = createOpenAI({ apiKey: 'test', baseURL: 'https://example.invalid/v1', fetch: async () => new Response(JSON.stringify(body), { status: 200, headers: { 'content-type': 'application/json' }, }), }); const result = await generateText({ model: openai.responses('gpt-5.3-codex'), output: Output.object({ schema: z.object({ items: z.array(z.string()) }) }), prompt: 'Return the items.', maxRetries: 0, }); console.log('Parsed output:', result.output); console.log('Full text:', result.text); console.dir({ content: result.content }, { depth: null }); console.dir({ responseMessages: result.response.messages }, { depth: null }); console.log('Token usage:', result.usage); ``` </details> ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) ## Future Work keep an eye out for the same streaming bug ## Related Issues fixes #23832github.com-vercel-ai · 54753479 · 2026-10-23
- 0.9ETVfeat(ai): introduce 'onUIMessageStepEnd' callback to persist messages (#22268) ## Background https://github.com/vercel/ai/issues/12841 before, an agent could run several steps but you could only easily save the complete chat message at the end the existing per-step callback gave you details like token usage and tool calls but not the chat message built so far ## Summary this builds on the changes of https://github.com/vercel/ai/pull/17459 + makes a fix on a timing bug in the original pr - added `onUIMessageStepEnd` which gives you that message after every completed step ## End-to-End Verification verified by running the example `examples/ai-functions/src/agent/openai/stream-ui-message-step-end.ts` ## Checklist - [x] All commits are signed (PRs with unsigned commits cannot be merged) - [x] Tests have been added / updated (for bug fixes / features) - [x] Documentation has been added / updated (for bug fixes / features) - [x] A _patch_ changeset for relevant packages has been added (for bug fixes / features - run `pnpm changeset` in the project root) - [x] I have reviewed this pull request (self-review) ## Related Issues fixes #12841 closes #12843 closes #12978 closes #17459 Co-authored-by: Lars Grammel <lars.grammel@gmail.com>github.com-vercel-ai · 3ebefff6 · 2026-10-07