Skip to content
self.md

■ SIGNALS // RADAR SIGNAL

the browser has two different kinds of yes

Browser consent switches, skill-package inspection, and deterministic code-review mechanics.

Published Read 2 min
■ [!] ON THIS PAGE ▼

self.md radar — 2026-09-19

the browser has two different kinds of yes

A browser tab, a repository hook and a code-review comment are small surfaces with large consequences: an agent gets a real permission, a hidden route to execution, or a finding somebody has to believe.

1. BrowserSkill separates borrowing a tab from asking for help

Tencent’s BrowserSkill now exposes two browser-level switches: one for confirming that an agent may borrow a tab, another for allowing it to ask a human for help. The README spells out the awkward cases. Turn confirmation off and pending borrow requests are released. Turn help off and pending help requests finish as disabled. A completed borrow stays completed.

The older --unattended and --no-confirm inputs remain for compatibility, but the extension’s saved settings are the authority. Borrowing a logged-in tab and asking the person to intervene now have separate controls.

reading: BrowserSkill release notes

2. SkillSpector follows a skill into its side doors

NVIDIA’s SkillSpector 2.11.0 now inspects hooks/hooks.json, .claude/settings.json and .claude/settings.local.json as part of a skill scan. It also adds opt-in traversal for referenced skills and bounded inspection of hidden or nested ZIP, DOCX, XLSX and PPTX content.

The nested material is inspected without extraction or execution, and excluded from LLM prompts. Runtime behaviour, images, encrypted content and non-English material remain outside its reach. A clean report says less than “safe to install,” but it covers more than the visible SKILL.md.

reading: SkillSpector 2.11.0 release notes

3. OpenCodeReview moves repeatable review work out of the model

Alibaba’s OpenCodeReview puts file selection, related-file bundling, rule matching, comment placement and reflection into deterministic components; the LLM agent reads context and forms the review. Its public benchmark covers 50 repositories, 200 pull requests, 10 languages and 1,505 annotated issues.

The project reports higher precision and about one ninth of the tokens against general-purpose agents using the same model, alongside lower recall. The declared trade is fewer false alarms for the reviewer and an admitted chance that some defects remain in the code. The benchmark is the project’s own, not a general result, but its stated choices are exposed rather than hidden inside a prompt.

reading: OpenCodeReview repository and benchmark

voice-identity