SShortSingh.
Back to feed

Claude Code's allowed-tools grants per-turn permission, not a persistent sandbox

0
·1 views

Claude Code's allowed-tools field in skill frontmatter only pre-approves listed tools for the single turn that invokes the skill, not for the entire session. Once the user sends a follow-up message, the permission grant clears and Claude will prompt again, even if the same skill is active. The field also does not restrict other tools — it merely skips prompts for the listed ones, while all other tools remain accessible under normal permission rules. To block a tool from a skill entirely, developers should use the separate disallowed-tools field instead. For permissions that need to persist across a session or project, the recommended approach is to configure allow or deny rules in the project's permission settings.

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)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

Developer builds zero-dependency Go clipboard tool for terminal-based text sharing

A developer has released cpynet, a single-file Go binary designed to simplify text sharing between remote servers and local machines entirely from the terminal. The tool requires no signup, no database, and no external dependencies, operating purely through curl commands over standard HTTP. Snippets are stored in memory only and auto-delete either after the first read or when a configurable TTL expires. For sensitive data, the tool supports AES-256-GCM client-side encryption, ensuring the server never processes plaintext. The project was motivated by the friction developers face when using traditional pastebins or clipboard utilities in headless or restricted server environments.

0
ProgrammingDEV Community ·

ZWISERFIT's LAO System Logged 5 AI Agent Failures in 24 Hours, Prevented All Repeats

On August 8, 2026, ZWISERFIT's AI orchestration framework LAO completed a 24-hour autonomous cycle during which five distinct failures were detected across three agents named Shuyu, Luna, and Hermes. Rather than filing bug reports, the system automatically repaired each failure and encoded 114 persistent rules called 'anchors' to structurally block the same error class from recurring. The failures included an agent pushing unrequested platform integrations and another executing the wrong port despite understanding the correct pattern. No founder intervention was required at any point during the repair process, and none of the five errors repeated. ZWISERFIT claims this structural gate approach outperforms prompt-tuning, which it says reduces errors by only 1–2%, and has made the LAO library publicly available via GitHub and PyPI.

0
ProgrammingDEV Community ·

Free Technical Documentation Template Offers Five-Page Structure With Built-In Validation

Developer and writer Ninad Pathak has released a reusable technical documentation template designed to give contributors a ready-made starting structure instead of building documentation from scratch. The template includes five focused pages covering an index, getting started, a task guide, configuration reference, and troubleshooting. It is packaged with MkDocs configuration, a local validator script, and a GitHub Actions deployment workflow to form a complete publishing pipeline. Each page is assigned a specific reader job, such as completing first setup or recovering from a known failure, with defined evidence requirements before publishing. The template is available for download and is intended to be adapted by replacing placeholder content with product-specific details.

0
ProgrammingDEV Community ·

AI Coding Tools Boost Speed but Risk Eroding Core Developer Skills and Ownership

AI coding assistants like GitHub Copilot and Cursor have become standard in modern software development, promising faster output and less repetitive work. However, developers increasingly risk shifting from writing code to simply reviewing AI-generated output, weakening their deep understanding of the systems they build. A key concern is the 'fluency heuristic,' where code that looks correct is accepted without rigorous verification, potentially hiding critical flaws that surface only under real-world conditions. Cognitive offloading to AI tools may also degrade algorithmic intuition and debugging skills over time, much as GPS has diminished natural navigation ability. Experts argue that preserving developer agency — the ability to understand, question, and own technical decisions — is not optional but essential for building secure and maintainable software.

Claude Code's allowed-tools grants per-turn permission, not a persistent sandbox · ShortSingh