Why Your Claude Code Subagents Fail: Description, Name Conflicts, and Bad Tool Entries

·1 views

Claude Code allows developers to create custom subagents using Markdown files with YAML frontmatter, stored in either a project-level or global agents directory. The routing system works entirely through the description field — Claude reads each subagent's description and decides when to delegate based on how well it matches a given task. Vague, title-like descriptions are the most common reason subagents go unused, while detailed descriptions that specify when to delegate work far more reliably. Two additional failure points include name collisions that silently drop a file and invalid tool entries that prevent the subagent from launching altogether. Optional fields such as model, maxTurns, memory, and isolation offer further control over how each subagent behaves and executes.

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 Intel SGX and Gramine Can Shield Health Data During Cloud Inference

Confidential Computing using Trusted Execution Environments (TEEs) offers a hardware-based approach to protecting sensitive health data processed in the cloud. Intel SGX creates an isolated memory region called an enclave, where data remains encrypted even if the host operating system or root administrator is compromised. A practical pipeline can be built using Intel SGX, the Gramine library OS, and C++, allowing health inference models to run on encrypted inputs without exposing data to the underlying cloud infrastructure. Gramine acts as a bridge between standard Linux binaries and SGX hardware, enabling containerized deployment via Docker without rewriting application code. The approach ensures that both model weights and patient data, such as heart rate readings, stay protected throughout the entire inference lifecycle.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Key Tips for Running Stable ML Inference as a Background Process on macOS

Developers running machine learning inference services as background processes on macOS face several platform-specific pitfalls that differ from Linux environments. Unlike Linux, macOS lacks GNU coreutils tools such as setsid and timeout by default, requiring alternatives like nohup combined with disown to keep processes running after a terminal is closed. Log processing tools like grep and tr can crash on macOS when encountering binary or garbled multibyte characters in inference logs, a problem resolved by setting LC_ALL=C to handle text as raw bytes. Because ML models take variable time to load, using a fixed sleep delay before sending requests is unreliable; instead, polling a health endpoint until it returns a 200 status ensures the service is truly ready. These tips were compiled from real-world experience running a Seed-VC voice conversion service built on FastAPI and uvicorn on macOS.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Handle Notification Taps in Expo Apps Across All Launch States

When a user taps a push notification, the app's entry point varies depending on whether it was terminated, backgrounded, or active at the time. Expo provides two separate mechanisms to cover these cases: getLastNotificationResponseAsync() for cold launches and addNotificationResponseReceivedListener() for background or foreground taps. Developers should avoid triggering navigation before authentication checks or the Navigator is fully ready, as this can cause routing conflicts. To prevent duplicate listeners during screen remounts or Fast Refresh, the listener subscription should be cleaned up with remove(). For security, externally provided URLs should be validated against allowed hosts and converted to internal routes before being passed to the router.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google, Bing, and Brave Agree on Top Search Result Only 29% of the Time

A comparative analysis of 15 searches run across Google, Bing, and Brave found that all three engines agreed on the number-one organic result only 29% of the time. On average, Google and Bing shared just three of their top-ten results, while Google and Brave overlapped on about five. Each engine showed a distinct content preference: Google heavily featured its own properties like YouTube and Reddit, Bing favored traditional publishers such as Forbes and PCMag, and Brave fell somewhere in between. Notably, Reddit, YouTube, and Wikipedia appeared in none of Bing's top-ten results, despite featuring prominently on Google. The findings highlight that tracking search rankings on a single engine gives an incomplete picture of overall search visibility.

0 comments Read more at DEV Community