SShortSingh.
Back to feed

ESLint Plugin Install Counts Are Misleading Without a Proper Baseline

0
·65 views

A developer analysis published on August 12, 2026 found that raw npm install counts for ESLint plugins are largely misleading because they include the 69-package ESLint baseline that any properly declared peer dependency will pull in automatically. After subtracting that baseline, eslint-plugin-import adds 136 packages despite listing only 19 direct dependencies, revealing that direct dependency counts poorly predict actual install weight due to transitive dependencies. Installing a typical React ESLint setup with three plugins together adds only 228 packages total, far less than the sum of individual plugin counts, because npm deduplicates shared dependencies across the tree. The analysis also cautioned that a small package tree is not a sign of a healthy plugin, citing eslint-plugin-security's unusually low count as a result of missing peer dependency declarations rather than good maintenance. The author recommends developers measure the full union of plugins they actually install, subtract the ESLint baseline, and re-run measurements regularly as package versions change.

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 ·

AI Agent Costs Extend Well Beyond Token Pricing

Token prices alone do not capture what it costs to complete a useful task with an AI agent, since retries, tool calls, retrieval, orchestration, and human correction all add expense. The article recommends tracking total cost per successful task, including model usage, search or database calls, tool execution, infrastructure, and review time, while separating successful runs from failed ones. It suggests breaking down spending by workflow, step, model, and outcome, and watching for repeated tool calls, oversized context, and excessive retrieval. Cost should be weighed alongside quality, latency, and escalation rate, since a cheaper model may increase retries or review. The piece also advises setting step limits, retry policies, timeouts, cost thresholds, and fallbacks, and framing cost governance as a product decision about acceptable outcomes.

0
ProgrammingDEV Community ·

A Practical Incident Runbook for Responding to AI Workflow Failures

When an AI workflow malfunctions, responders must establish what the system attempted and which external actions may already have taken effect. The recommended approach begins with containment, such as disabling the affected trigger or tool permission and pausing queued jobs that could run duplicates, while keeping a human route open for urgent requests. Teams should log the first failure time, workflow and model or prompt changes, affected integrations, and example run IDs, then classify the problem as a poor answer, tool failure, latency spike, or unintended external change. External writes should be reconciled with the system of record before any replay, and non-idempotent actions must never be retried blindly. The runbook should name clear owners and rollback criteria, and be revisited after tool or workflow changes.

0
ProgrammingDEV Community ·

Axion Closes Username-Enumeration Timing Leak in Login Route with Dummy Hash

A developer working on Axion, a microforum platform, described a fix for a timing attack that let attackers identify valid usernames. The vulnerability arose because the server skipped expensive password hashing when a username was not found, producing faster responses than for real accounts. To address this, the login route now compares the submitted password against a dummy hash whenever no user matches. As a result, both valid and invalid usernames require the same computational work, leaving attackers no timing difference to exploit. The author invited readers to share how they protect their own login routes against timing attacks.

0
ProgrammingDEV Community ·

Testing AI Agent Workflows When Outputs Vary Between Runs

Traditional tests that demand identical outputs for identical inputs break down with AI workflows, which can phrase answers differently each time. Exact text matching can yield brittle suites or false confidence when a phrase appears even though the underlying task failed. The recommended approach is to define each workflow's user goal, allowed actions, required evidence, and unacceptable outcomes, then judge behavior against those conditions. Teams should build test sets from realistic requests, edge cases, ambiguous inputs, and known failures, keeping a stable set for regressions and a fresh set for discovering new problems. Evaluation should score multiple dimensions such as task completion, grounding, policy adherence, escalation, latency, and cost, using deterministic checks where possible and human or calibrated judge review for open-ended quality. Release thresholds for critical behaviors and production feedback loops help ensure that model or prompt changes do not quietly make a workflow less safe.