dpm-signal
2026-05-26 · backend: claude · 277 deduped
Top stories
Numind released NuExtract3, a 4B model based on Qwen3.5-4B under Apache-2.0. It converts PDFs, screenshots, forms, tables, receipts, and other visually structured documents to Markdown or structured JSON, and replaces the earlier NuMarkdown-8B. A free HuggingFace Space is available for testing without an account.
Useful if you're building document ingestion pipelines and want a clean local extraction model — Apache-2.0 means it's production-safe without negotiating a commercial license.
release · local-models · extraction · vlm · open-weights
A fix to llama.cpp's server checkpoint creation prevents the 'forcing full prompt re-processing' behavior that fires when agentic tools like opencode or pi rewrite conversation history mid-session. Previously, a 70k-token coding context would silently reload in full after any tool that touched turn data — the concrete trigger for the 'freezes after you say thank you' experience.
This is one of the most annoying local-model agent pain points in practice; worth watching the PR for merge and picking up in your llama.cpp build.
llama.cpp · local-models · agents · bugfix
Cline's latest release wires up GPT-5.5 support on SAP AI Core, DeepSeek V4 Flash and Pro as selectable models, and Gemini 3.5 Flash across both the standard Gemini and Vertex AI providers. It also routes Poolside Laguna models through next-gen prompts and native tool calling, and fixes Vertex AI global endpoint handling for Claude.
Three frontier model additions in one release — if you're on Vertex AI with Claude, the endpoint routing fix is worth the upgrade independently of the new model support.
cline · release · agents · models
Multica is a TypeScript platform for treating coding agents as assignable teammates — task dispatch, progress tracking, and compounding skills across Claude Code, Codex, and others. It has grown from 29k to 33k stars over the past 8 days.
Notable as a self-hostable alternative to commercial agent orchestration; the skills compounding mechanism is the differentiator to evaluate against your current setup.
agents · orchestration · open-source
Pi bundles a coding agent CLI, a unified LLM API layer, TUI and web UI libraries, a Slack bot, and vLLM pod support in one TypeScript package. It gained 733 stars in 24h against a 54k base.
Worth evaluating if you want one repo covering the local agent stack rather than stitching multiple tools — the unified LLM API layer is what separates this from a plain terminal wrapper.
agents · tools · open-source · cli
An AWS ML Blog post documents three implementation paths for giving Bedrock-hosted agents a code execution environment: a self-hosted Docker sandbox on ECS for maximum control, managed AgentCore Code Interpreter, and an Anthropic SDK-compatible proxy path for teams already using the Anthropic SDK. Each approach includes architecture and sample code.
Good reference if you're building Bedrock agents and haven't settled on a sandboxing strategy — the ECS path gives you the most control, the AgentCore path the least ops.
aws · bedrock · agents · tool-use
TLDR AI's 2026-05-25 issue lists 'MCP goes stateless' as a headline item. No primary source URL appeared in today's data. If confirmed, this would remove the persistent process requirement from MCP servers and enable simpler HTTP-only hosting.
High signal if true — stateless MCP would make server deployment dramatically simpler — treat as unconfirmed until a spec update or PR surfaces.
mcp · protocol · unconfirmed
VIP updates
openclaw/openclaw
Beta.2 adds 👍/👎 iMessage tapback handling for run approvals (mirrors the WhatsApp behavior shipped in beta.1), exposes active run status and cancellation to WebUI and Discord voice callers while a consult is still running, and introduces Discord voice wake-name gating with raised profile bootstrap context budget. Multiple gateway hot-path caching improvements also landed to reduce repeated plugin file stats and manifest reads.
simonw
Datasette 1.0a30 adds a keyboard-accessible 'Jump to' menu (trigger with /) with a jump_items_sql() plugin hook so any plugin can inject searchable items. datasette-agent 0.1a4 uses this immediately to surface a 'Start a new agent chat' input directly in the jump menu on agent.datasette.io.
Anthropic
Anthropic published an open-source repository of plugins for Claude Cowork targeted at knowledge workers. The repo gained 1,448 stars in 24h and 2,891 over 8 days. Extension points and plugin categories are best understood by reading the repo directly.
GitHub risers — last 24h
affaan-m/ECC
+2052 stars
Agent harness optimization system covering skills, memory, security, and instinct layers for Claude Code, Codex, Opencode, and Cursor.
mukul975/Anthropic-Cybersecurity-Skills
+999 stars
754 structured cybersecurity skills for AI agents mapped to MITRE ATT&CK, NIST CSF 2.0, D3FEND, and MITRE ATLAS — works with Claude Code, Copilot, Codex CLI, Cursor, and 20+ platforms under Apache 2.0.
garrytan/gstack
+600 stars
Garry Tan's Claude Code skill setup: 23 opinionated tools covering CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA roles.
openai/codex
+444 stars
OpenAI's Rust-based lightweight coding agent for the terminal.
Trendshift — live top 4
stablyai/orca
3346 stars total
Orca is the next-gen IDE for working with a fleet of parallel agents. Run any coding agent with your own subscription. Available on desktop and mobile.
Lum1104/Understand-Anything
30605 stars total
Graphs that teach > graphs that impress. Turn any code into an interactive knowledge graph you can explore, search, and ask questions about. Works with Claude Code, Codex, Cursor, Copilot, Gemini CLI, and more.
esengine/DeepSeek-Reasonix
8213 stars total
DeepSeek-native AI coding agent for your terminal. Engineered around prefix-cache stability — leave it running.
GitHub risers — last 2 weeks
manaflow-ai/cmux
19432 stars total
Ghostty-based macOS terminal with vertical tabs and notifications designed for AI coding agent workflows — +2,232 stars over 8 days.
BenedictKing/ccx
1852 stars total
Single Go proxy for Claude, Codex, and Gemini APIs — +852 stars over 8 days.
langchain-ai/langgraph
32925 stars total
Stateful multi-agent orchestration framework — +625 stars over 8 days, steady baseline growth.
Curiosities
A Columbia paper proposes a 13MB adapter that lets one LLM transmit a summary of its KV state to a different model, bypassing text-based inter-agent communication. In shared-context settings it outperforms a 956MB C2C adapter; in different-context settings it is 23% more accurate and 8.5x faster than text. No shipped tool yet, but the mechanism is novel for multi-agent pipeline design.
Inductive Deductive Synthesis jointly synthesizes Lean 4 implementation and proof, learning from failed attempts to route toward promising strategies. It solved 7/7 distributed key-value store specs where Claude Code with Opus 4.6 and Codex with GPT-5.4 each managed only 2/7 — at roughly $106/spec and 6.8 hours on average. Code is referenced in the paper.
Notes
The 'MCP goes stateless' headline in TLDR AI had no primary source URL in today's data — worth tracking down the spec PR or MCP repo announcement before treating it as confirmed. The llama.cpp checkpoint PR (#22929) is unmerged as of data collection; watch the thread. rohitg00/ai-engineering-from-scratch had the highest 2w star velocity in the dataset (+10,359 over 8 days) but is course/tutorial content and was excluded per editorial bar.