self.md radar — 2026-09-22
four places an agent’s reach gets specified
today’s useful documents are unusually literal. one tool writes its plans into .git; one operating system separates three kinds of AI refusal; one vendor makes a graph the memory layer; one runtime admits Python to the production roster.
1. Foremerge lets agents declare the job before they touch the code
Foremerge is a local-first coordination layer for coding agents working in separate Git worktrees. Its claim is narrower than “multi-agent collaboration”: agents publish targets, dependencies, provisional changes, decisions, validation, and provenance before the code changes land.
The example is annoyingly plausible. One agent migrates callers to StripePaymentService; another adds PayPal support to the old PaymentService. Different files, clean merge, useless PayPal branch. Git sees text. It does not see that the second agent has built on a road the first one has already removed.
Foremerge stores its shared record in the repository’s .git directory and says its deterministic conflict detector and SQLite store are already implemented. The README also says shared multi-machine mode and published benchmarks do not yet exist. For now, its coordination record stops at one machine and one repository.
reading: Foremerge repository
2. Apple gives Apple Intelligence three separate doors to close
Apple’s Mac guide separates Writing Tools, image creation, and Intelligence Extensions inside Screen Time. The route is ordinary: enable Content & Privacy Restrictions, open Intelligence & Siri, then disable the capability in question.
The extension switch is the useful one. Apple names third-party providers such as ChatGPT there, instead of treating them as an inseparable part of the operating-system bundle. A household can leave some features on and refuse the external-provider path. That is a better question than whether “AI” is enabled.
The guide says nothing about enterprise management, retention, or what happens to data. It documents a consumer control surface with separate feature and provider-extension switches.
reading: Apple’s Screen Time guide
3. V7 sells a graph where the agent can point back to the file
In an OpenAI profile of V7 , the company describes V7 Go as a layer that reads connected repositories, extracts entities and relationships, and attaches cited evidence to the resulting Context Graph. Its agents can search that graph through MCP rather than repeatedly trawling a long prompt for the same fund name or policy clause.
V7 says an incoming document is connected to an existing company, fund, person, or other ontology record, with the original evidence retained. The graph then becomes part of the decision about which context arrives with a task.
OpenAI reports 89% accuracy for GPT-6 Astra on V7’s hardest graph-query tests, while V7 claims 99.9% accuracy for some workflows. Those are vendor-case-study numbers. The documented product move is a source trail inside the context layer.
reading: OpenAI: How V7 gives AI agents institutional memory
4. Cloudflare puts Python on the Workers production list
Cloudflare has made Python Workers generally available
. The announcement calls Python a first-class language in the Workers runtime, and its queue example still crosses a visible bridge: to_js turns a Python dictionary into something self.env.QUEUE.send can use.
That small conversion is more useful than a language-war headline. Python joins the runtime, but the surrounding platform objects have not quietly become Python objects. Teams get another supported entry point and still need to know where the bridge is.
reading: Cloudflare: Python Workers are now generally available