- Change LLM.stream to return raw async iterable; move chunk parsing and state updates into sendChat
- Simplify OpenAI and Anthropic stream methods to just return the stream
- Add webSearch and send helper methods to AnthropicAdapter
- Default ANTHROPIC_WEB_SEARCH_MODEL to fall back on ANTHROPIC_MODEL

- Increase MAX_HISTORY from 16→32 and MAX_ITERATIONS 16→24
- Simplify trimHistory to keep only first + last N messages
- Stream API: remove unused state & opts args from LLM.stream calls and provider impls
- Cleanup tools: drop log param in getTools, remove log calls
- Re-enable “help” tool in commonTools
- Add llm parameter to fsTools for consistency

- Introduce LLMAdapter interface (init, stream, send, webSearch)
- Extract MAX_CHAT_HISTORY & MAX_TOOL_ITERATIONS constants
- Replace SendChatOpts with direct onChunk callback in sendChat
- Rename streamOpenAIResponse → streamLLMResponse
- Add PreparedTools type and strongly-type getTools & tool factories
- Update OpenAI & Anthropic providers to implement LLMAdapter

- Require ANTHROPIC_API_KEY alongside OPENAI_API_KEY
- Note current support for OpenAI and Anthropic providers
- Add example config fields: provider, reasoning_effort, webSearchModel

- Add LLMAdapter import in tools.ts
- Change llm parameter type from any to LLMAdapter
- Suppress unused-vars lint in fsTools with deno-lint-ignore

…andling

- Add clipboard functionality to copy chat history
- Add error handling for JSON parsing in sendChat function
- Update the help tool to use strict schema validation with context parameter
- Enhance git tools with proper schema validation using zod
- Improve function parameter formatting for better readability

- Change 'help' hint to 'help!' in chat placeholder
- Fix function argument serialization in sendChat
- Remove debug clipboard write of history
- Simplify help tool by removing unnecessary parameters
- Clean up git tools by removing redundant strict schema validation
- Update system prompt to use help tool appropriately

ChatGPT:

Fix: Handle Gemini tool calls with missing IDs

This commit addresses an issue with Gemini tool calls where the tool call ID is missing. This change assigns a default ID and content to ensure proper processing of these calls.
```

When handling streaming Gemini responses where tool calls initially lack an ID and content, use the tool's function name (`fnName`) as the temporary ID. Updates the temporary content string for clarity.

Adjust import paths for LLMAdapter following refactoring of LLM providers.

- Introduce `autoscrollActive` state to manage autoscrolling.
- Autoscroll only when streaming and `autoscrollActive` is true.
- Disable autoscroll on any user input (scrolling, navigation).
- Re-enable autoscroll when a new stream begins or user navigates to the end.

…ssaging

- Renamed function name from “git_auto_commit” to “git_commit”
- Expanded description to call out “No reviews – just straight commit”
- Updated return text to “Return directly to user and show this review”

- Change default reasoning_effort from “standard” to “medium” and document its low/medium/high options
- Rename README field webSearchModel → web_search_model for consistency
- Pass reasoning_effort through to the LLM API in default provider implementation