Skip to content
This repository was archived by the owner on Aug 21, 2026. It is now read-only.

CHANGELOG.md

Latest commit

 

History

History
179 lines (97 loc) · 9.69 KB

CHANGELOG.md

File metadata and controls

179 lines (97 loc) · 9.69 KB
Raw

Changelog

All notable changes to papa-ts will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[2.7.4] - 2025-12-02

Reverted

  • Provisional assistant message: Reverted the changes from v2.7.3 that emitted a provisional "assistant" message at the start of generation. The issue causing tools to block streaming was identified as a client-side fetch polyfill limitation, not a backend issue. The backend now behaves as it did in v2.7.2.

[2.7.3] - 2025-12-02

Fixed

  • Final response streaming with tools: Fixed an issue where the final response tokens would appear all at once when using tools. This was caused by client UIs not receiving an initial "Assistant" message signal after tool execution. The stream now emits a provisional assistant message when text generation begins, ensuring smooth token appending.

[2.7.2] - 2025-12-02

Fixed

  • Real-time tool streaming: Agent.streamTokens now yields empty token chunks on on_tool_start and on_tool_end events. This ensures client UIs receive immediate updates when tools are called or finish execution, rather than waiting for the final response generation.

[2.7.1] - 2025-11-30

Fixed

  • Tool call detection during streaming: Fixed an issue where tool calls were not being detected in stream chunks. The implementation now properly queries checkpoint state on chain_end events to capture tool calls and tool responses that may only be available in the checkpoint, not in individual streaming events.
  • Message format detection: Improved message normalization to handle LangChain messages with type fields (e.g., "human", "ai", "tool") in addition to role fields, ensuring correct role mapping for all message formats.
  • Duplicate message emissions: Fixed excessive message chunk emissions by only emitting when message count increases, preventing duplicate updates during streaming.

Changed

  • Message normalization fallback: Changed default role for unrecognized messages from "system" to "assistant" for more accurate fallback behavior.

[2.7.0] - 2025-11-30

Breaking

  • Removed raw event access: Completely removed includeEvents option and event-type chunks from the streaming API. Raw LangChain StreamEvent objects are no longer exposed to clients. The API now focuses exclusively on normalized data (token chunks with optional messages, and result chunks).

Changed

  • Stream event processing: The streaming implementation now extracts and normalizes messages from stream events when available, making them accessible via chunk.messages in token chunks. All LangChain internals are hidden from the public API.

[2.6.0] - 2025-11-30

Added

  • Normalized messages in stream chunks: AgentStreamChunk token chunks now include an optional messages?: ThreadMessage[] field. This allows clients to use a single, consistent message format throughout the streaming lifecycle.
  • Automatic thread title generation: Added Agent.generateTitle(threadId) method that automatically generates a concise title (max 5 words) for a thread based on its conversation history. The generated title is persisted to the thread store and can be used to provide better thread organization and navigation in user interfaces.

[2.5.1] - 2025-11-30

Added

  • Checkpoint and Runnable exports: Exported BaseCheckpointSaver, Checkpoint, CheckpointMetadata, CheckpointTuple, PendingWrite, CheckpointListOptions (from @langchain/langgraph-checkpoint) and RunnableConfig (from @langchain/core/runnables). This allows consumers to implement custom checkpoint savers and better type runnable configurations.

[2.5.0] - 2025-11-30

Added

  • LangSmith Telemetry configuration: Updated LangSmithTelemetryOptions to accept explicit apiKey, endpoint, and project parameters, allowing programmatic configuration of the LangSmith client without relying solely on environment variables.

[2.4.0] - 2025-11-29

Added

  • Thread message normalization: Introduced a canonical ThreadMessage schema and helper utilities (normalizeThreadMessages, getMessageText) that flatten LangChain constructors (chunks, tool calls, metadata) into a stable, client-friendly shape.

Changed

  • Agent now emits/hydrates normalized messages everywhere (results, history, metadata previews), ensuring consistent roles/timestamps/tool-call surfaces regardless of the underlying checkpoint format.

[2.3.0] - 2025-11-29

Breaking

  • ThreadStore metadata-only contract: ThreadSnapshot no longer contains messages. All thread stores (in-memory, FS, IndexedDB, localStorage) now persist only thread metadata (threadId, title, timestamps, custom metadata). Consumers should rely on LangGraph checkpoint savers for conversation history.
  • Thread history API: Agent.getThreadHistory() now returns a ThreadHistory object that merges thread metadata with messages loaded from the configured checkpointer. Downstream code should update type references accordingly.

Added

  • LocalStorageCheckpointSaver: New LocalStorageCheckpointSaver implementation lets browser environments persist LangGraph checkpoints and pending writes using localStorage. Exported via the public API and wired into the browser example.
  • Enhanced metadata helpers: createSnapshot now normalizes metadata/timestamps on input, simplifying upgrades from previous formats.

Changed

  • Agent persistence flow: The agent now writes only metadata (last run id, model, message preview/role) to the configured ThreadStore while leaving messages to the checkpoint saver. Fetching history automatically stitches checkpoint messages with stored metadata.
  • Thread store implementations/tests: Updated file system, IndexedDB, localStorage, and in-memory stores plus tests and examples to the metadata-only model to avoid mismatches and lost role fields.

[2.2.0] - 2025-11-29

Breaking

  • Async provider registration: ProviderRegistry.registerProvider() and useOpenAI()/useAnthropic()/useOllama()/useSapAICore() now return promises so they can await dynamic model discovery. Callers must await these methods before choosing models.

Added

  • Dynamic model discovery: Built-in providers now query their respective APIs (OpenAI, Anthropic, Ollama, SAP AI Core) to list available chat/embedding models instead of relying on hard-coded defaults. Override the discovery by passing explicit chatModels / embeddingModels if needed.

Changed

  • Provider configuration surface: Extended provider options to accept API keys, base URLs, headers, API versions, and custom fetch implementations to support discovery across environments.
  • Examples/tests: Updated sample scripts, LangGraph graph, and tests to await provider registration and showcase the new behavior.

[2.1.0] - 2025-11-29

Added

  • Agent token streaming: Expose Agent.streamTokens() so clients can display live model tokens (with optional raw events) while still receiving the final AgentResult.
  • Node streaming example: Add examples/node/streaming.ts, a runnable script that prints live tokens using the new streaming API.

[2.0.3] - 2025-11-29

Added

  • Static LangChain provider imports: Promote @langchain/openai, @langchain/anthropic, @langchain/ollama, and @sap-ai-sdk/langchain to direct, statically imported dependencies so bundlers (including Obsidian) always include them without dynamic import() fallbacks.
  • Dynamic loader removal: Drop the bespoke dynamicImport helper for simpler, more reliable provider initialization across runtimes.

[2.0.2] - 2025-11-29

Added

  • Tool helper re-export: Re-export LangChain's tool helper from the main papa-ts entrypoint so downstream consumers can define tools without importing @langchain/core directly.

[2.0.1] - 2025-11-29

Added

  • SAP AI Core embedding support: Add default embedding models and factories for SAP AI Core via @sap-ai-sdk/langchain.

Changed

  • SAP AI Core chat integration: Migrate to AzureOpenAiChatClient / AzureOpenAiEmbeddingClient usage for a more direct and flexible integration.
  • Node example: Update examples/node/tools.ts to demonstrate using SAP AI Core as the provider (with gpt-5) instead of OpenAI as the default.

[2.0.0] - 2025-11-20

Breaking Changes

  • Complete architecture rewrite: The assistant backend has been rewritten from scratch to support the new v2 architecture.
  • Provider management: Provider setup and configuration flow has changed; existing integrations will likely need configuration updates.

Added

  • Agentic assistant: New Agent abstraction and agentic assistant implementation.
  • Improved provider registry: Centralized ProviderRegistry for managing OpenAI, Anthropic, Ollama, SAP AI Core and other providers.
  • Telemetry integrations: LangSmith and Langfuse telemetry helpers for tracing and observability.
  • Better configuration handling: Enhanced model and provider configuration types, including support for new API keys and environment variables.

Changed

  • Project structure: Source code reorganized under src/agent, src/providers, src/memory, and src/telemetry.
  • Build output: Library is now built via tsup targeting both ESM and CJS with bundled type definitions.
  • Testing setup: Tests migrated to vitest with improved test utilities.

Fixed

  • Environment handling: More robust handling of environment variables for provider API keys.

[0.5.x] - 2024-xx-xx

Pre-v2 series with the original architecture and provider implementations.