mpfaffenberger/code_puppy

Added

• 📡 Remote skills catalog from llmspec.dev with on-demand download and offline fallback
• 🧩 Tab completion for /skills commands using cached remote catalog

Changed

• 🔌 Migrate scheduler and skills commands to plugin architecture
• 🛠️ Move tool callback wrapping from decorator to pydantic-ai patch
• 🔧 Simplify model factory to always load bundled models.json
• 🎛️ Add max effort tier and 128K output tokens for Claude Opus

Fixed

• ⌨️ Slash command completions reappear after backspace
• 🔄 Update Gemini thinking bypass seed to constant string

Added

• 🧠 Support for Synthetic-Kimi-K2.5-Thinking-NVFP4 model
• 📁 Discover JSON agents and skills from project-level .code_puppy directory
• 🌐 Global custom slash commands via ~/.code-puppy/commands/

Changed

• 🔄 Replace Claude Sonnet/Opus 4.5 thinking models with Opus 4.6 variants
• ↩️ Revert Claude code tool call character length tracking fix

Removed

• 🔒 Autosave signature verification system

Changed

• 🚀 Enable DBOS backend by default

Removed

• 🗑️ Remove deprecated zai-glm-4.6 model entries

Fixed

• 🧵 Resolve race conditions in MessageBus, session cache, and circuit breakers
• 🛡️ Prevent unbounded memory growth in startup buffer and stderr capture
• 🧹 Ensure proper cleanup of temporary files and resources
• 🔄 Replace blocking operations with async equivalents
• 💾 Use atomic file writes and UTF-8 encoding for config stability
• 📢 Log dropped messages and plugin errors for better visibility

Security

• 🔐 Add HMAC-SHA256 signing to session pickle files
• 🛡️ Restrict pickle deserialization to safe types only

Changed

• 🤖 Upgrade GLM models to versions 4.7 and 5, removing support for version 4.6

Changed

• 🔗 Update Discord community link in README