Category / T-CAT / SOFTWARE
Software
Software / T-2026-6372
Destructive Command Guard: An open-source safety belt for AI coding agents
The Destructive Command Guard (dcg) blocks catastrophic git and shell commands from AI coding agents, offering sub-millisecond latency and 50+ security packs.
SourceDicklesworthstone/destructive_command_guard
Software / T-2026-7019
llama.cpp b9982: The Bug That Silently Discarded Per-Request Reasoning Budgets
llama.cpp b9982 fixes a bug where per-request reasoning_budget_tokens overrides were silently discarded. A small fix with big implications for local AI tooling.
Sourceggerganov/llama.cpp b9982
Software / T-2026-6963
llama.cpp b9968 brings int8 MoE prefill to Adreno GPUs
llama.cpp b9968 ships int8 dp4 dense and MoE prefill optimizations for Adreno GPUs, co-authored by a Qualcomm engineer. On-device AI inference gets faster and cheaper.
Sourceggerganov/llama.cpp b9968
Software / T-2026-6931
llama.cpp b9957: A quiet refactor that changes how models use tools
llama.cpp b9957 ships a tools_io abstraction and drops apply_diff. The refactor points to a future where on-device agentic workflows are a first-class concern.
Sourceggerganov/llama.cpp b9957
Software / T-2026-9038
Google's Stitch Skills library turns coding agents into design tools
Google publishes a library of Agent Skills for its Stitch design platform, compatible with Claude Code, Cursor, and Gemini CLI. The plugins let coding agents convert code to…
Sourcegoogle-labs-code/stitch-skills
Software / T-2026-6899
llama.cpp b9946: Qualcomm's Hexagon backend gets a serious upgrade
llama.cpp b9946 ships a major Hexagon backend overhaul with VTCM tiling, HVX kernel refactoring, and tracing instrumentation for unary ops.
Sourceggerganov/llama.cpp b9946
Software / T-2026-4389
OpenAI ships GPT-5.6-sol support in Python SDK v2.45.0
OpenAI's latest Python SDK release adds support for GPT-5.6-sol, a model variant that signals deeper integration of chain-of-thought reasoning into the API.
Sourceopenai/openai-python v2.45.0
Software / T-2026-7219
TencentDB Agent Memory: local long-term memory with a 61% token cut
TencentCloud's open-source memory system for AI agents cuts token use by 61% and boosts pass rates by 51% using a layered, fully local architecture.
SourceTencentCloud/TencentDB-Agent-Memory
Software / T-2026-6777
llama.cpp b9908 enforces strict prompt cache RAM limits
llama.cpp b9908 fixes a long-standing bug where --cache-ram was not a hard limit, evicting old entries only after saving new ones.
Sourceggerganov/llama.cpp b9908
Software / T-2026-7082
Kyutai's Pocket TTS runs 100M-parameter speech on a CPU, no GPU required
Kyutai releases Pocket TTS, a 100M-parameter speech model that runs 6x real-time on a MacBook Air M4 CPU, with voice cloning and browser support.
Sourcekyutai-labs/pocket-tts
Software / T-2026-2120
Karakeep: The self-hosted bookmarker betting AI tagging beats the cloud graveyard
Karakeep, a self-hostable bookmark-everything app, uses LLMs for automatic tagging and summarization, carving a niche between dead cloud services and the hoarder instinct.
Sourcekarakeep-app/karakeep
Software / T-2026-6691
CodexBar turns AI coding limits into a menu bar — and exposes the industry's meter problem
CodexBar surfaces usage limits from 40+ AI coding providers in a macOS menu bar. The real story is what it reveals about a market that still can't tell you what you'll pay.
Sourcesteipete/CodexBar