feat(llm-sdk): add reasoning/thinking support via OpenAI Responses API + OpenRouter passthrough#92
Open
ankushchhabradelta4infotech-ai wants to merge 48 commits into
Conversation
…into fix/sdk-ui # Conflicts: # packages/copilot-sdk/src/chat/classes/AbstractChat.ts
# Conflicts: # pnpm-lock.yaml
…into fix/sdk-ui
…into fix/sdk-ui
…Vite 8 examples - generative-ui-demo: port of experimental generative UI to Vite+Express, recharts, dark UI with prompt suggestions sidebar - skills-demo: new demo for server-side skills system with loadSkills(), 3 skill files (code-review/concise-mode/customer-support), branching toggle via allowEdit prop Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…l activation - Replace generic skills with SaaS-focused: revenue-intelligence, customer-health, incident-runbook - Full dark dashboard UI (Bricolage Grotesque + JetBrains Mono, navy/indigo palette) - Live metrics bar, nav with module links, AI Copilot badge - Skill cards animate on load: scan line sweep → expand → capability reveal with staggered fade-in - Skill state detection via load_skill toolRenderer watching for tool call completion - Branching toggle, demo prompt injectors for demo recording - Updated server system prompt for SaaS context Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Adds @yourgpt/llm-sdk/fallback subpath export: - createFallbackChain() with priority and round-robin routing - Per-model retries with exponential/fixed backoff before fallback - FallbackExhaustedError with per-model failure breakdown - MemoryRoutingStore (default) + pluggable RoutingStore interface - onRetry / onFallback observability callbacks - Two-tier error detection (class-based for complete(), message-regex for stream()) Includes fallback-demo example and docs page. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
# Conflicts: # pnpm-lock.yaml
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…o and generative-ui-demo Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ensions with composite Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ion and UI improvements - Added support for dynamic skills registration via API, allowing skills to be registered at runtime. - Updated server to handle dynamic skills, including new endpoints for skill management. - Introduced a new skill for frontend design with detailed guidelines. - Improved UI with new font styles (DM Sans and DM Mono) and animations for skill activation. - Enhanced the skills panel to display both static and dynamic skills, improving user experience. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Added 'vaul' version 1.1.2 and 'tw-animate-css' version 1.4.0 to dependencies. - Updated 'yourgpt-server-demo' with new dependencies including 'cors', 'dotenv', 'express', and 'ws'. - Updated devDependencies for 'yourgpt-server-demo' with '@types/cors', '@types/express', '@types/ws', 'tsx', and 'typescript'. - Updated integrity checks for '@types/ws' and 'vaul' in the lockfile.
…vior - 'multi-step' (default) — one bubble per server agent iteration (OpenAI/LiteLLM style), default unchanged - 'single-turn' — all iterations collapsed into one bubble per user turn (Vercel AI SDK / Claude.ai style) Usage: <CopilotProvider streamMode="single-turn" /> Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Deleted the comprehensive documentation for the tool management branch, which included details on features, configurations, and known issues related to the tool management stack across copilot-sdk and llm-sdk.
- Replaced the existing system prompt to reflect the new focus on HR operations, including employee onboarding and performance reviews. - Removed outdated skills related to revenue intelligence, customer health, frontend design, and incident management. - Introduced new skills for employee onboarding and performance reviews with updated icons and configurations. - Adjusted loading behavior to prevent duplicate skill calls in the conversation history. - Updated initial welcome message to align with the new HR Copilot theme.
… from package files - Updated the version in package.json to 2.1.4-alpha.3. - Excluded source map files from the package distribution by adding "!dist/**/*.map" to the files array.
…ect event ordering Previously, server-side tools were executed post-loop after the adapter's for-await finished, causing message:end to arrive before action:end(result). This broke client message splitting and rendered skill cards below the assistant response instead of above. Changes: - llm-sdk/runtime: execute server tools inline in case "action:end" before message:end arrives naturally from the adapter, removing all event-ordering hacks - copilot-sdk/AbstractChat: split message turn on toolResults.size > 0 (not just text) so tool-only turns are correctly finalized at message:end - copilot-sdk/AbstractChat: skip server-side assistantWithToolCalls in done handler to prevent duplicate tool card renders in the UI - examples/playground: use workspace:* deps and add transpilePackages for Turbopack Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Introduced a new tool for retrieving the current date, including its day of the week. - Updated tool rendering logic to handle the new date tool and improve UI feedback during execution. - Adjusted the handling of assistant messages to prevent duplicate rendering of tool results in the chat interface.
…handling - Introduced a new provider configuration for "yourgpt-server" in constants and types. - Added a proxy route for yourgpt-server to handle streaming and non-streaming requests. - Updated PlaygroundPage to conditionally check for API keys based on the selected provider. - Bumped versions for copilot-sdk and llm-sdk to reflect recent changes.
…cleanup - Replace all sidebar icons with Hugeicons duotone-rounded (AiMagic, AiBook, MagicWand, Puzzle, BubbleChat, SlidersHorizontal, FileCode, ServerStack, AiChip1) - Move Generative UI to root sidebar; rewrite with two-approach structure (toolRenderers vs AI-generated iframe via useGenerativeUI/HtmlRenderer) - Move branching from chat/ to advanced/; merge message-actions into chat/ui - Collapse skills/ folder to flat skills.mdx (fixes fumadocs dropdown bug) - Move context/ pages into advanced/; move headless into customizations/ - Add 18 permanent redirects in next.config.mjs for all moved/deleted routes - Add AiMagic, AiBook, BubbleChat, FileCode, MagicWand, Puzzle, ServerStack, SlidersHorizontal icon components Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Add @yourgpt/llm-sdk/fireworks provider (OpenAI-compatible, lazy OpenAI client) - Fix tool call streaming: track by index instead of tc.id (Fireworks repeats the same tc.id across chunks, breaking multi-chunk argument accumulation) - Add examples/fireworks-demo — Vite+Express app with 26-model selector, browse link to fireworks.ai/models, tested with kimi-k2p5 + deepseek-v3p2 - Add docs: providers/fireworks.mdx and providers/openrouter.mdx (closes YourGPT#79) - Update providers/index.mdx and meta.json to include both new providers Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Introduced a new `storage.mdx` file detailing session persistence and chat history management. - Removed outdated `session.mdx` and `meta.json` files to streamline documentation. - Included examples for basic setup, browser storage, server persistence, and the `useThreadManager` hook. - Added information on data persistence, auto-eviction, and API contract for server storage.
- Removed outdated headless routes from `next.config.mjs` to streamline navigation. - Added new documentation for headless features, including `useCopilotEvent` and `useMessageMeta`, to enhance user understanding. - Introduced client-side and server-side skills documentation, detailing registration and usage. - Created new `headless` and `skills` sections in the documentation, including examples and best practices for implementation. - Deleted the `skills.mdx` file as part of the restructuring to improve clarity and organization.
- Modified the PATH variable in the pre-commit script to include the user's pnpm directory for better compatibility in non-interactive shells.
All four legacy adapters (OpenAI, Azure, Google, Ollama) were yielding action:call events but never emitting action:end, preventing the runtime from knowing when a tool call finished and triggering execution. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Add native Together AI provider using OpenAI-compatible API with streaming, tool calling, vision, and JSON mode support. Includes full docs page, Express+Vite demo app, and comprehensive test suite. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
… beta release notes Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Vercel uses npm which doesn't support pnpm's workspace: protocol, causing EUNSUPPORTEDPROTOCOL errors during install. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…tProvider - Add createTogetherAI() legacy factory (returns AIProvider for createRuntime) - Rebuild demo from Express+Vite to Next.js app using the proper copilot-sdk pattern: createRuntime + runtime.handleRequest on backend, CopilotProvider + CopilotChat on frontend - Sidebar with model selector, setup guide, and links - 11 verified models across DeepSeek, Llama, Qwen, and more Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…Runtime pattern - Add fallback chain with priority strategy + retry logic to demo API route - Add fallback toggle UI in sidebar showing chain order - Update docs: createTogetherAI + createRuntime as primary pattern, fallback chain section, cleaned up model list - Add REST test script for all models - Bump llm-sdk to 2.1.9 Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Replace the tool-call-based generative UI with a text-streaming approach (similar to Claude Artifacts pattern). The AI writes HTML wrapped in <GENUI> tags as part of its text response, which streams naturally and renders progressively in a sandboxed iframe. Key changes: SDK (copilot-sdk/experimental): - New useGenerativeUI() hook returns wrapMessage for CopilotChat - New generativeUISystemPrompt() helper for backend system prompts - New GenUIFrame component: iframe with postMessage, auto-height via ResizeObserver, copilot bridge API for interactivity - Simplified types: removed chart/table/stat/card types, HTML-only - Deleted CardRenderer, TableRenderer, StatRenderer SDK (core streaming improvements): - Progressive action:args emission in Anthropic/OpenAI adapters - tool-call-start/delta events in providers and stream-text - Runtime: proper action:args accumulation (don't null currentToolCall) - AbstractChat: handle action events when streamState is null - processChunk: parsePartialJson for streaming arg extraction - connected-chat: attach unmatched streaming executions to messages - ChatWithTools: allow action:start/args for client tools - AbstractAgentLoop: reuse streaming executions Example (generative-ui-demo): - Rewritten with sidebar, dicebear avatars, copilot logo - Uses generativeUISystemPrompt() + loadSkills() on backend - Skills system with frontend-design skill (eager) - SaaS dashboard copilot context (Acme Inc.) Docs: - Rewritten generative-ui.mdx with two approaches documented - Updated BETA-FEATURES.md Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- AbstractChat: check toolResults.size in message:end so tool-only turns reset streamState and don't attach next turn's tools to previous message - AbstractChat: in single-turn mode parent insertChainParentId from message before the streaming message, not from streamState.messageId - AbstractChat: skip role:tool messages from done.messages in single-turn (already in streamState.toolResults, re-inserting creates duplicates) - connected-chat: restrict liveExecutions to last assistant message only; prefer metadata.toolExecutions for historical messages - connected-chat: filter unmatched executions to pending/executing only so stale completed cards don't bleed into new streaming message - connected-chat: null-safe message filter to prevent render crashes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Previously request.systemPrompt || this.config.systemPrompt meant a client-sent prompt could silently override the server-configured one, breaking generativeUISystemPrompt and any server-side prompt setup. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
|
@ankushchhabradelta4infotech-ai is attempting to deploy a commit to the Delta4 Infotech Team on Vercel. A member of the Team first needs to authorize it. |
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
2 participants
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.
Description
Adds reasoning/thinking token support to the
llm-sdkfor OpenRouter-routed models. OpenAI'so1/o3/o4andgpt-5series hide chain-of-thought on the chat-completions endpoint — this PR routes those models through OpenAI's Responses API to surface reasoning summaries, and adds streamingreasoning_contentpassthrough for other OpenRouter models (e.g. DeepSeek R1, Claude with extended thinking).Changes
adapters/openai.ts— Dynamicprovidername resolved frombaseUrl(openai / openrouter / xai / azure / google); addsdisableThinkingconfig flag; adds fullstreamWithResponsesAPIpath that maps Responses API SSE events (response.reasoning_summary_text.delta,response.output_text.delta, function-call events) back to the sameStreamEventshapes the chat-completions path emitsproviders/openrouter/provider.ts—supportsThinkingflipped totrue;doStreamdetects OpenAI reasoning model IDs (openai/o1*,openai/o3*,openai/o4*,openai/gpt-5*) and delegates todoStreamResponsesAPI; addsreasoning_content/include_reasoningpassthrough for other reasoning models (DeepSeek R1 etc.); addsdisableThinkingprovider option to opt outType of Change
Testing
Checklist
Screenshots (if applicable)