Developer Tool Warns Claude MAX Users Before 5-Hour Token Cap Cuts Sessions

·1 views

A developer has built a shell script called token-budget-advisor.sh that monitors token usage in Claude MAX sessions and alerts users before they hit the platform's 5-hour output token cap. Unlike the built-in status line, which only shows a percentage, the tool translates usage into absolute token counts, flagging sessions as dangerous above 800,000 tokens and critical past 1.2 million. The script integrates with a dashboard and status line, pulling data from both the ccusage command-line tool and a local cost-log.jsonl file, preferring ccusage values when available. A key technical fix addresses a double-counting bug in cost-log.jsonl, where cumulative mid-session entries under the same session ID inflate token totals two to three times over; the script resolves this by retaining only the latest entry per session. The tool is designed to fail open on errors and remain compatible with Bash 3.2, making it suitable for unattended background workflows where hitting the cap silently would otherwise stall all pending tasks.

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)

How to Verify AWS Authentication in Cypress Using STS GetCallerIdentity

Developers building Cypress test automation that interacts with AWS services can use the AWS Security Token Service (STS) and its GetCallerIdentity API to confirm successful authentication before writing further tests. The process involves installing the AWS SDK's STS client package and configuring credentials either via a local cypress.env.json file or through environment variables for CI/CD pipelines. Cypress environment variables must be mapped to Node.js process.env so the AWS SDK can automatically pick them up via its Credential Provider Chain. A Cypress task is then created to instantiate an STSClient and run GetCallerIdentityCommand, returning the authenticated account details if the connection succeeds. This foundational connectivity check ensures the automation environment is properly linked to the intended AWS account before integrating services like Lambda, S3, or DynamoDB.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Patch Any Monospace Font with Nerd Fonts Using Docker

A developer wanted to add PowerLine glyphs to Operator Mono, their preferred monospace font, which lacks these symbols by default. PowerLine glyphs enhance terminal status lines and shell prompts with version control indicators. Using the Nerd Fonts patcher tool, the developer ran a Docker container to generate a patched version of the font. The process involved mounting the original font directory as input and a new directory as output within the Docker command. The result was a custom Nerd Font variant of Operator Mono with full PowerLine and extra symbol support.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a Trading Bot's Risk Logic Outgrew Its Execution Engine

A developer building a Python-based algorithmic trading platform on Bybit documents how early architectural shortcuts created long-term structural problems. Initially, risk management logic like Stop Loss and Take Profit calculations were embedded directly inside the Trade Execution Engine for simplicity. As the platform grew to include features like Trailing Stop, Break Even, Telegram controls, and Spot/Futures trading, the execution engine gradually absorbed responsibilities far beyond its original scope. The author reflects that prototype architectures optimise for speed while production architectures must optimise for change — two fundamentally different goals. This is Part 4 of an ongoing engineering series chronicling the platform's evolution from a simple trading bot into a modular, production-ready system.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How a required schema field forced an AI pipeline to fabricate statistics every time

A developer discovered that an automated article-generation pipeline had been publishing invented statistics across multiple languages, not through a deliberate audit but while reviewing an unrelated part of the codebase. The root cause was a required output schema field that demanded a statistic and source for every article, combined with prompt instructions telling the model to estimate a plausible figure when it lacked a real one and to omit the publication year to keep content looking evergreen. A separate chart-generation block compounded the problem by independently hallucinating data while being instructed to cite official sources like the Bureau of Labor Statistics. The developer noted that adding honesty disclaimers to the prompt would not fix the issue, since a language model cannot distinguish a fabricated number from a recalled one. The fix was structural: making the statistic field optional, skipping it entirely when no verified source URL existed, and replacing unsupported figures with qualitative language instead.

0 comments Read more at DEV Community