---
title: 'In the News: November 12, 2026 (Morning)'
description: 'Geoffrey Huntley patches git so agents cannot force push, using one Nix overlay for laptops, CI, sandboxes and production images.'
canonical_url: 'https://darkfactory.dev/news/2026-11-12-morning'
markdown_url: 'https://darkfactory.dev/news/2026-11-12-morning.md'
collection: news
date_published: '2026-11-12T07:10:00-05:00'
date_modified: '2026-11-12T07:10:00-05:00'
---

# In the News: November 12, 2026 (Morning)


Geoffrey Huntley published a way to stop an agent from force pushing that does not rely on the agent following instructions: he patches git and ships the patched build to every environment. Margaret-Anne Storey's new paper describes a tool for recording why a codebase is the way it is, with agent skills that flag where that record has gone stale.

## 1. Huntley removes force push from git and ships the patched build to every environment

**[the answer to ai is engineering](https://ghuntley.com/answer/)** · Geoffrey Huntley, Antithesis · ghuntley.com, November 12, 2026

The post is Huntley's write-up of a talk he gave at AI DevCon New York. Its working example is a build of git that refuses to force push. He writes: "This version of git doesn't support force push." The change lives in one Nix overlay, a file that tells Nix to build git with extra patches, and Huntley says he got the patches by asking Claude for them. His [demo repository](https://github.com/ghuntley/nix-demo) applies five patches that refuse `--force`, `--force-with-lease`, `+refspec` and `--mirror` pushes, then uses the same definition in a developer shell, a virtual-machine test and a container image. Huntley says one definition generates his laptops, CI, agent sandboxes and production images, so anything that depends on git gets the patched build.

He treats security fixes the same way. When a dependency such as OpenSSL needs a patch, he patches it once and Nix rebuilds everything that depends on it. For temporary patches he asks the agent to add a note to AGENTS.md saying the patches should come out in seven days. He says the next developer or agent to pick up that machine removed them.

On agent loops, Huntley says he runs agents as root, but only on throwaway Nix machines. He tells readers never to run agents as root on production hosts or connect them to production databases. He stops a loop from accepting bad code with deterministic checks: hooks that block a push when the code does not compile, static analysis rules about which parts of the code may depend on which, and property-based tests. The last section recommends Bombadil and Hegel, two testing libraries from Antithesis, and Huntley writes about the company as "we".

The post reports no measurements. Huntley says the recent run of security incidents cost him no sleep, which is his account of his own systems.

**Why it matters:** A prompt instruction asks an agent not to force push, and a pre-push hook can be skipped with `--no-verify`, as the demo's README points out. A patched binary removes the command wherever the environment is built, including the sandbox the agent runs in. The price is running your environments on Nix, and the post shows a demo, with no account of how the approach held up across a team.

## 2. Storey's Cognitive Tours uses agent skills to flag recorded decisions that no longer match the code

**[Cognitive Tours: Making Understanding a Deliverable in AI-Assisted Development](https://arxiv.org/abs/2611.08456)** · Margaret-Anne Storey, University of Victoria · arXiv, submitted November 5, 2026

Storey describes a web-based prototype for recording why a system is built the way it is. Developers mark waypoints such as goals, user needs, design decisions, code, tests and open questions, link them, and order some of them into a tour that another person can follow. People write the descriptions. Four Claude Code skills do the upkeep: two audit existing waypoints and tours against the files they point to and the git history, and two suggest recent work worth recording. The paper says the skills never write to the map and draft text only when asked.

Storey gives one example from her own use. After an evening of AI-assisted work with more than 20 commits, she ran a skill the next day to compare the commits with the map. Its top-ranked finding was a design decision that contradicted an architecture decision record still marked as accepted.

The paper is six pages, and its arXiv listing says a shorter version is under review for the ICSE 2027 demo track. Storey calls the contribution a tool and design proposition. She says it is not yet validated that tours reduce cognitive debt, her term for a team losing its shared understanding of a system. Her evidence is her own use of the tool over six months, and the paper ends with a planned evaluation.

**Why it matters:** The audit idea transfers without the tool: have a skill read recent commits against the recorded decisions and report where the code now contradicts one. Whether writing tours by hand improves a team's understanding is the question the paper leaves open.