Popular 30k-Star Claude Code Repo Overcounted Tokens 2.35x Due to Documented Bug

·1 views

A widely used Claude Code template repository with over 30,000 GitHub stars contained a token-counting bug that inflated usage metrics by approximately 2.35 times. The flaw stemmed from summing token usage across every JSONL record per assistant message, even though Claude Code splits a single message into multiple records — each repeating the same usage data. Anthropic's own SDK documentation explicitly warned about this behavior and provided a reference fix, meaning the bug was publicly known but never addressed in the repo. A developer identified and submitted a fix backed by synthetic corpus testing that confirmed exact over-counting, and two AI reviewers approved the pull request. As of the article's publication, the fix had been open for ten days with no response from any human maintainer.

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)

DeepSeek V4 Flash 0731 Corrupts JSON Integer Fields When Thinking Mode Is On

DeepSeek's retrained V4 Flash 0731 model has a documented defect where enabling its default thinking mode causes integer field corruption in strict JSON schema outputs, with 8 of 13 test runs returning wrong values despite schema-valid responses. The failure was observed across two independent request paths, with corrupted outputs including incorrect line-item counts and slightly wrong totals — errors silent enough to pass validation and reach production pipelines. Disabling thinking via the enable_thinking: false parameter resolved every corrupted run and reduced token usage to roughly one-seventh of the default. The corruption appears tied to the presence of thinking rather than its budget size, as even a capped 64-token budget produced the same errors. The preview build of V4 Flash remained clean across all four probed runs, while V4 Pro also showed corruption in 2 of 4 runs, suggesting the issue extends across the broader V4 thinking model family.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Model Context Protocol: The Standard Helping AI Agents Connect to External Tools

The Model Context Protocol (MCP) is a communication standard that allows AI models to discover and interact with external tools, APIs, databases, and services in a uniform way. Before MCP, developers had to build separate custom integrations for every combination of AI model and external application, resulting in significant engineering overhead. MCP solves this by defining a shared protocol through which an AI client can find available tools, understand their functions, and execute them with structured inputs and outputs. The protocol involves two main components: an AI client that requests capabilities and an MCP server that exposes those capabilities in a consistent, machine-readable format. By establishing a common language between AI systems and external software, MCP reduces integration complexity and makes tools reusable across multiple AI platforms.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

SRE Incident Response Playbook: How to Handle Exposed Secrets Effectively

Site Reliability Engineer Tiexin Guo published a detailed guide on January 27, 2026, outlining how to build and use incident response playbooks, with a focus on secret leaks. The article emphasizes that effective playbooks must define goals, roles, communication channels, and documentation standards before any incident occurs. Guo highlights that secret leaks are harder to detect than typical outages because standard metrics like CPU usage and error rates may show only minor changes. Unlike service-down incidents that trigger immediate alerts, malicious actors exploiting leaked secrets can go unnoticed for extended periods. The guide also stresses the importance of testing playbooks through simulations and training team members to ensure readiness under pressure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

OpenAI Launches GPT-Live for Real-Time, Full-Duplex Voice Conversations in ChatGPT

OpenAI has introduced GPT-Live, a new family of voice models designed to make ChatGPT conversations more natural through full-duplex audio, allowing the system to listen and speak simultaneously. The technology supports human-like conversational behaviors such as interruptions, brief acknowledgments, and mid-conversation topic changes without disrupting the exchange. GPT-Live-1 is available to paid ChatGPT subscribers, while GPT-Live-1 mini is offered to free-tier users, with API access planned for a later date. The architecture separates real-time voice interaction from heavier backend tasks — including complex reasoning and web search — which are handled by GPT-5.5 and seamlessly fed back into the ongoing conversation. At launch, GPT-Live-1 supports web search, memory, and visual widgets within ChatGPT Voice on iOS, Android, and the web.

0 comments Read more at DEV Community