Open-source tool 'merge-proof' flags PRs not validated against their final merge base

·1 views

A developer has released merge-proof, a lightweight open-source Git and GitHub Action tool designed to identify evidence gaps in pull request validation. The tool addresses a specific scenario where a PR's base branch advances after CI passes, meaning the final merged state may never have been explicitly tested. It implements two blocking checks — BASE_DRIFT_UNVERIFIED and PROTECTED_BOUNDARY — to flag cases where overlapping file changes or sensitive code paths lack verification against the actual merge target. In a study of 2,037 agent-authored PRs across 32 public repositories, 24.1% returned a NOT_PROVEN result, though the author stresses this does not indicate bugs or broken code. The tool runs entirely on local Git state, makes no external calls, and carries no telemetry, positioning it as a diagnostic aid where GitHub's native merge queue enforcement is absent or incomplete.

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)

29,000 Home Weather Stations and AI Agents Tested as Volcanic Early Warning System

Researchers used over 29,000 consumer-grade Netatmo barometric sensors across Japan, combined with Google Apps Script cloud archiving and Gemini AI agents in the Antigravity CLI environment, to detect atmospheric shockwaves from volcanic eruptions. The system was validated against two historical Japanese eruptions — the 2018 blasts at Mt. Kusatsu-Shirane and Mt. Shinmoedake. It reconstructed shockwave velocities within 98.5% of theoretical sound speeds, pinpointed unmonitored crater directions to within 1.78 degrees, and estimated explosive yields ranging from roughly 179 to 1,041 tons of TNT equivalent. The crowdsourced sensor network delivered between 2.5 and 15 minutes of advance warning with no recorded false positives, even during violent storm conditions. The findings, published on ESS Open Archive, suggest that dense consumer IoT networks grounded in physics-based AI analysis could help fill the large blind zones left by sparse government monitoring infrastructure.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Japanese TTS Bug Turned 'Shōshō' Into Gibberish Due to Unicode Whitelist Gap

A developer discovered that a text-to-speech system for a virtual inquiry desk was garbling Japanese phrases, turning '少々お待ちください' into something that sounded like 'shomo o-machi kudasai'. After ruling out the voice model, cache, and synthesis parameters, the root cause was traced to a preprocessing function that stripped characters not on an allowed list. The character '々', a common Japanese iteration mark, was excluded because Unicode classifies it under 'CJK Symbols and Punctuation' rather than 'CJK Unified Ideographs', which the whitelist covered. Further review revealed several other affected characters, including range symbols, colons used in time expressions, and rare kanji variants used in proper names. The fix required not just expanding the allowlist but converting ambiguous symbols like '〜' and ':' into Japanese-readable equivalents before synthesis.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

DeepSeek Harness: Open-Source Agent Runtime Offers Full Transparency and Plugin Control

DeepSeek Harness (dsh) is an open-source, MIT-licensed code agent framework built on Cordis, a plugin-based runtime where every component — tools, UI, sandboxing, and orchestration — is a replaceable plugin. Unlike closed tools such as Claude Code, it provides a Trajectory view that logs each agent turn with precise timing, token usage, and full tool input/output, allowing developers to pinpoint exactly where a run failed. The framework enforces granular sandbox policies at the kernel level using Seatbelt and Landlock, with three access modes ranging from read-only to full-access, and sensitive operations require explicit human approval. Structured, readable error messages replace cryptic stderr output, making permission denials actionable rather than opaque. The project can be launched via a simple CLI in web, headless, or custom-profile modes, positioning itself as a distinct category of agent tooling aimed at transparency and developer control.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

GEO Optimisation Builds on SEO Foundations, Not a Replacement for Them

Generative engine optimization (GEO), coined in a 2023 academic paper by Aggarwal et al., demonstrated visibility improvements of up to 40% within AI-generated answers using a structured benchmark called GEO-bench. However, the research also found that effective techniques vary significantly by domain, meaning no single universal GEO playbook applies across all content types. Google has confirmed that its AI Overviews and AI Mode draw from the same index and quality systems as classic Search, with no separate submission process for AI visibility. Core technical fundamentals — fast pages, clean semantic HTML, schema markup, and proper heading structure — remain essential for both traditional crawlers and AI retrieval systems alike. Experts note that most citation failures stem from technical issues such as JavaScript rendering problems or misconfigured bot rules, not content shortcomings, making a strong SEO foundation a prerequisite for any GEO effort.

0 comments Read more at DEV Community