Emacs org-mode package

Enables users to create #+begin_ai...#+end_ai special blocks within org-mode documents that function as persistent conversation contexts. The system parses block syntax to extract configuration (model, temperature, system prompts), maintains conversation history as org-mode content, and streams responses directly into the buffer using Emacs' asynchronous request handling. The orchestration layer (org-ai.el) dispatches parsed blocks to service adapters which handle provider-specific API communication while maintaining buffer-local state for insertion positions and active requests.

Abstracts 8+ AI service providers (OpenAI, Anthropic, Google Gemini, Perplexity, DeepSeek, Azure OpenAI, local Oobabooga, Stable Diffusion) behind a single unified request interface. The org-ai-openai.el adapter module handles provider-specific API details including authentication, request formatting, response parsing, and error handling. Service selection is configured globally or per-block, and the dispatcher (org-ai-complete-block) routes requests to the appropriate adapter without requiring users to understand provider-specific APIs.

Manages API credentials for multiple AI services through Emacs' auth-source library, supporting encrypted credential storage in .authinfo.gpg or system keychains. Users configure service endpoints and credential lookup patterns, and the system retrieves credentials at request time without exposing them in configuration files. Supports per-service authentication and fallback mechanisms for multiple API keys.

Integrates with Oobabooga's text-generation-webui for running local LLMs without cloud API dependencies. The org-ai-oobabooga.el adapter communicates with the WebUI API, supporting model selection, parameter configuration, and streaming responses. Users can switch between cloud and local models using identical org-mode syntax, enabling privacy-preserving and cost-effective AI workflows for users with local GPU infrastructure.

Manages conversation history and AI responses as native org-mode content with automatic link creation and metadata tracking. Responses are inserted as org headings, lists, or code blocks depending on content type, and metadata (timestamp, model, tokens used) is stored as org properties. Supports linking between related conversations and organizing conversations hierarchically within org files.

Integrates OpenAI Whisper API for speech-to-text transcription and platform-native TTS (macOS say, espeak, greader) for text-to-speech output through the org-ai-talk.el module. Users can invoke voice input to generate prompts or voice output to hear AI responses read aloud. The system handles audio encoding/decoding, manages Whisper API communication, and coordinates with system TTS engines, enabling hands-free AI interaction workflows.

Provides image generation capabilities through two separate adapters: org-ai-openai-image.el for OpenAI DALL-E and org-ai-sd.el for local Stable Diffusion (AUTOMATIC1111 WebUI). Users specify image prompts in org-mode blocks with configuration for size, quality, and style. The system sends requests to the appropriate service, downloads/retrieves generated images, and embeds them as org-mode image links in the document. Supports both cloud-based (DALL-E) and self-hosted (Stable Diffusion) workflows.

Allows fine-grained configuration at the individual org-mode block level through special syntax headers (#+ai_model, #+ai_temperature, #+ai_system_prompt, etc.). The block parser (org-ai-block.el) extracts these headers and merges them with global configuration, creating a request-specific configuration object. This enables users to use different models, temperatures, and system prompts for different blocks without global reconfiguration, supporting experimentation and multi-purpose workflows within a single org file.

Enables AI operations across multiple files in a project through global commands that aggregate file content as context. Users can invoke commands like 'refactor this codebase' or 'summarize project documentation' which collect relevant files, construct a unified prompt with file content, and send to the AI service. The system handles file discovery, content aggregation, and response insertion back into appropriate files or a summary buffer.

Implements non-blocking asynchronous request handling using Emacs' async request library, allowing AI responses to stream into the buffer without freezing the editor. The system maintains buffer-local state (insertion position, current request ID) to coordinate streaming chunks, handles response parsing from provider APIs, and inserts text incrementally as it arrives. Supports cancellation of in-flight requests and graceful error handling without blocking user interaction.

Integrates with Emacs' yasnippet library to provide pre-built templates for common AI interaction patterns (code review, documentation generation, refactoring, etc.). Users can insert templates that expand into org-mode blocks with pre-configured prompts, system messages, and parameters. Templates support variable substitution and can be customized per-user or per-project, enabling rapid creation of well-structured AI requests without manual prompt engineering.

Allows users to define custom system prompts (#+ai_system_prompt header) that shape AI behavior for specific tasks. System prompts are merged with user prompts at request time and sent to the AI service as system-level instructions. Supports multi-line system prompts, variable substitution, and per-block customization, enabling users to create specialized AI personas or task-specific behaviors without modifying the core system.

Implements comprehensive error handling for API failures, network issues, and malformed requests with user-facing error messages and recovery options. Supports request cancellation mid-stream, timeout handling, and retry logic for transient failures. Errors are displayed in org buffer or minibuffer without disrupting editor state, and users can cancel in-flight requests or adjust configuration and retry.