Developer finds 9,857 tokens consumed by Claude Code config before any input

·2 views

A developer discovered that their Claude Code setup — comprising 107 skills, 38 agents, and 15 commands accumulated over several months — consumes 9,857 tokens at the start of every session before any user input is entered. The hidden cost comes from description fields: every installed component's description must remain loaded in the context window throughout the session so the model knows what tools are available. While 9,857 tokens represents roughly 5% of a 200,000-token context window, the author argues the real issue is efficiency — paying a constant overhead for components that are only triggered about 2% of the time. This baseline tax also compounds with other persistent context elements such as system prompts, MCP server schemas, and project files, quietly reducing available working memory. The developer shared a short Python script to help others audit their own Claude Code configurations and measure their per-session token overhead.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

FreeCORE Project Aims to Continue TrueNAS Core Development

A community initiative called FreeCORE has emerged to continue development of TrueNAS Core, the open-source NAS operating system. The project appears to have launched its own website at freecore.org to support ongoing maintenance and development efforts. This follows concerns within the community about the future direction of TrueNAS Core after iXsystems shifted focus toward TrueNAS SCALE. The discussion was posted on Hacker News, though it attracted minimal engagement at the time of reporting.

0 comments Read more at Hacker News

ProgrammingDEV Community ·

Security Audit Flags High-Risk Reentrancy and Access Control Flaws in Paxos Gold

A security audit of Paxos Gold (PGX), a regulated gold-backed token protocol with roughly $1.91 billion in total value locked, identified critical vulnerabilities across its smart contracts as of August 30, 2026. The most serious flaw lies in the PGXBridge withdrawal function, where an external call is made to a user-supplied address before the withdrawal nonce is updated, enabling a classic reentrancy attack that could allow double-spending of gold-backed tokens. On the access control side, the PGXController contract allows any address to nominate a pending owner without restriction, and the proxy upgrade function is guarded by a multisig key that has not been rotated since launch. Additionally, the PAUSER_ROLE is delegated to a single external contract, PGXStaking, which itself carries upgrade-path risks that could allow an attacker to seize pause privileges. Auditors rated the combined risk at 8 out of 10, warning that a coordinated exploit could break the protocol's 1:1 gold peg, freeze user funds, or enable a rug-pull scenario.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Why 50 Concurrent Users Don't Need 50 Database Connections Explained

A common misconception among backend developers is that concurrent users require an equal number of database connections, but connection pooling makes this unnecessary. A connection pool is a set of reusable connections that requests borrow and return, eliminating the overhead of creating and closing a new connection for every request. Key settings like MaxOpenConns, MaxIdleConns, and ConnMaxIdleTime control how many connections are open, kept idle, or recycled over time. In Go, setting MaxOpenConns to 5 means at most 5 simultaneous connections are allowed, but they are created on demand rather than all at startup. Understanding these pool settings is especially critical when building scalable SaaS or multi-tenant backend applications.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Meta and Alibaba Release Competing 30B Open-Weight AI Models for Consumer GPUs

In a four-day window in August 2026, Meta and Alibaba released dense ~30B parameter multimodal AI models under unrestricted Apache 2.0 licenses, marking a shift in the open-weight landscape. Meta's Muse Glimmer 30B launched on August 10, followed by Alibaba's Qwen3.8-27B on August 14, both designed to run on a single 24 GB consumer GPU such as an NVIDIA RTX 3090 or 4090. Unlike some contemporaries such as Qwen3.8-Max and Kimi K3, neither model imposes commercial revenue thresholds, making them freely usable for commercial deployment at any scale. The two models differ notably in context length and modality support — Qwen3.8-27B handles text, image, and video inputs with a 262,144-token native context window, while Muse Glimmer 30B supports text and image inputs with a 131,072-token context. Their distinct architectures and serving profiles mean each model targets different operational workflows, giving developers a meaningful choice depending on hardware and use-case requirements.

0 comments Read more at DEV Community