TypeSafe's Jev model adds probability-based safety gate to Pi coding agent shell commands

·2 views

TypeSafe released Jev on September 15, 2026, a decision-only AI model that returns probability scores between 0.0 and 1.0 for yes/no propositions rather than generating text. A developer integrated Jev into the Pi coding agent CLI as an extension called pi-jev-auto-mode, routing unfamiliar shell commands through semantic probability evaluation while handling known patterns via static rules. Testing across 18 real API fixtures revealed a clear bimodal distribution in intent coverage scores, with a natural empty gap between 0.15 and 0.77, allowing a reliable threshold to be set at 0.60. The project highlights a critical calibration risk: raising a hazard threshold even slightly can inadvertently move a dangerous command out of the rejection band and into the allowed range, meaning threshold changes must be grounded in empirical data. The extension is publicly available on npm and requires a TypeSafe API key for semantic evaluation, with the system defaulting to a fail-closed state if the evaluation engine is unreachable.

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 Safely Guard DNS Zone Deletions in Automated Migration Pipelines

Engineers running automated DNS migrations risk accidentally deleting platform-owned zones alongside customer-owned ones if safety checks rely solely on naming conventions or command-line flags. A more robust approach treats each deletion as a small state machine requiring an explicit operator approval, a strict allowlist of exact zone names, and a persisted intent record before any provider call is made. A unique operation ID should be assigned to each deletion request to prevent duplicate executions, but deduplication alone cannot protect against a correctly delivered command targeting the wrong zone. Scope validation — confirming ownership and allowlist membership — must occur before retry or deduplication logic kicks in. DNS deletions can also affect mail policy records such as DMARC, so the review scope must extend beyond web traffic to all delegated responsibilities within a zone.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

How to Use Claude Code's Permission Rules to Block Risky Terminal and Database Commands

Claude Code can execute terminal commands and database queries autonomously, raising the risk of unintended actions like premature git pushes or destructive SQL operations. Its permission system addresses this by routing every tool call into one of three outcomes — allow, ask, or deny — based on pattern rules defined in advance. These rules are stored in a project's .claude/settings.json file, which can be committed to a repository so the entire team shares the same configuration. Developers can apply targeted rules across multiple tools, such as auto-allowing safe git commands while blocking any DROP database query outright. The guide also cautions that overusing 'ask' prompts can cause developers to approve actions without reading them, undermining the system's protective value.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Benchmark of 11 MCP Servers Reveals Token Count Varies Up to 30x by Design

A technical analysis published on September 17, 2026, benchmarked 11 widely used Model Context Protocol (MCP) servers to measure how many tokens their tool definitions consume. The study found dramatic variation: Notion required approximately 19,050 tokens while Puppeteer used only around 610, a difference of over 30 times from a single server installation. Two separate measurement studies were compared, and their results for the same servers diverged by as much as 97%, due to differences in tokenizer methods, server versions, and initial configuration states. Both studies noted that token counts carry little meaning without specifying the counting method and exact server version used. The findings suggest that token budget is a fundamental design constraint for MCP tool development, not merely a performance consideration.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Capsule CRM vs HubSpot: How to Pick the Right CRM Based on Business Needs

A detailed analysis highlights that comparing Capsule CRM and HubSpot feature-by-feature misses the point, as the two tools serve fundamentally different business needs. Capsule CRM is designed for small teams of 1–20 users who need lightweight contact tracking, simple pipelines, and quick setup with minimal training. HubSpot, by contrast, suits growing companies that require marketing automation, multi-hub scaling, and complex CRM workflows for larger teams. The article recommends that buyers first define the specific operational job they need software to perform before evaluating platforms by cost, scale, or brand. Other tools like Salesforce Essentials and ZoomInfo are also mapped to distinct use cases, reinforcing a job-based framework for CRM selection.

0 comments Read more at DEV Community