Developer Builds Token Sentinel Script to Prevent Claude Code Rate-Limit Surprises

·1 views

A developer has shared a shell script called token-budget-advisor.sh designed to monitor token usage within Claude Code's 5-hour output-token blocks before the cap is reached. The problem arose when automated background jobs silently consumed the entire token budget overnight, leaving the developer with no work completed by morning. The script uses a dual-source approach, treating the ccusage command-line tool as the primary data source and a local cost-log.jsonl file as a fallback, ensuring the guard remains active even if one source fails. Two alert thresholds are defined — a warning at 800,000 output tokens and a critical level at 1.2 million — giving users time to shift heavy workloads to the next block. The tool also tracks daily session counts and weekly costs, emitting a single-line status suitable for dashboards or daily briefings.

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)

Retaining Splink Match Scores as Graph Edge Properties Improves Graph RAG Accuracy

Most entity resolution pipelines discard match probabilities after applying a binary match/non-match threshold, losing valuable confidence information before it reaches downstream systems. A developer building er-api, a multilingual entity resolution service for Korean, Japanese, Chinese, and English corporate data, chose instead to preserve Splink match probabilities as edge properties on SAME_AS relationships in the knowledge graph. This allows different retrieval queries to apply different confidence thresholds — for example, legal compliance queries can require 95% confidence while exploratory searches may accept 70%. During graph traversal for Graph RAG, the system propagates the worst-case confidence score along each path, so language model prompts can explicitly flag low-confidence identity claims. The approach also enables automatic approval of high-confidence matches while routing uncertain ones to a human review queue.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer Builds Offline AI Agent Memory Using SQLite Instead of Vector Embeddings

A data engineer building the LoreConvo tool chose SQLite with its FTS5 full-text search extension over cloud-based vector embeddings for storing AI agent session memory. The decision was driven by concerns over network latency, per-call costs, and the lack of transparency and reproducibility that embedding-based storage introduces. Each conversation session is saved as a row in a local SQLite file, capturing transcripts, summaries, tool usage, and tags, with no user action required. Because the file resides entirely on the user's machine, it works offline and can be easily backed up, migrated, or version-controlled. The developer noted that while a hybrid approach combining full-text search with embeddings still makes sense in some scenarios, SQLite alone delivered comparable recall quality without the overhead of managed embedding services.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

LangGraph vs LangChain in 2026: A Practical Guide for Complex Pipelines

LangChain and LangGraph are both open-source Python libraries for building AI-powered workflows, but they differ significantly in how they handle control flow and state management. LangChain structures workflows as linear chains of steps, making it simple and effective when logic flows in a straight line without branching or looping. LangGraph extends this by modeling workflows as directed graphs, enabling conditional routing, parallel execution, shared state, and built-in retry policies across nodes. In a real-world 19-node finance pipeline moving from raw market data to a daily profit-and-loss statement, LangGraph's graph abstraction reduced boilerplate and made complex branching logic — such as routing trades to compliance review — easier to define and audit. The practical guidance is to use LangChain for straightforward pipelines and LangGraph when workflows require dynamic branching, loops, or shared state across many interdependent steps.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How Hidden Browser Tabs Caused an AI QA Agent to Falsely Pass Blank Canvas Tests

A developer using an AI-powered QA agent built on Claude and Chrome MCP discovered that the tool was falsely approving broken features, including a completely blank canvas, by marking them as fully functional. The root cause was browser throttling in hidden background tabs, where Chrome aggressively limits requestAnimationFrame to save power, resulting in zero rendered frames despite clean JavaScript execution. Measurements confirmed that rAF fired zero times in hidden tabs, while setInterval ran at roughly one-eighth its expected rate and setTimeout drifted significantly. The developer found that AI agents were evaluating code-path health — such as event wiring and absence of JS errors — rather than actual visual or behavioral output. The fix involved opening tabs in a visible state before screenshotting, requiring explicit behavior verification for any dynamically rendered features, and mandating that unverified dynamic features never be silently graded as passing.

0 comments Read more at DEV Community