self.md radar — 2026-10-13
four places to keep the operator in the loop
A credentials file, a first-run screen, a live terminal, and a tool result too large for a prompt. The machinery changes; each case still comes down to where a person can step in.
1. web-edit takes one secret-bearing step out of the chat
Simon Willison released web-edit 0.1a0 , a Python web UI for files in a folder. He was working from a phone through Codex Remote and wanted to edit a credentials file without feeding that file through the LLM on the way.
The app is meant to sit behind Tailscale. That does not turn it into a secret manager, and it leaves the operator with a web surface to secure. But it restores a direct path for a task that chat-based remote work tends to turn into a prompt exchange.
reading: web-edit 0.1a0
2. Rome’s private dashboard is local first; the first visitor still counts
Rome’s Docker quickstart
brings its dashboard up at localhost:7663 and requires an explicit --bind before it is exposed beyond the machine. It keeps state in named Docker volumes, and telemetry export stays off until an OTLP endpoint is configured.
First-run onboarding is open to whoever reaches it first. The network boundary is conservative; the ownership boundary is still a live setup task.
reading: Rome README
3. uniTerm spells out four approval settings for terminal work
uniTerm places an AI agent in the active terminal and says it can plan, run, observe, and iterate over multiple shell commands. Its documentation offers four execution settings: bypass, dangerous only, dangerous plus write, and confirm all. The agent can also stay pinned to one terminal tab or follow the active one.
The split between dangerous actions and writes avoids one blanket approval setting. Names in a README cannot prove the classification works. They do give a person something inspectable before the agent gets a shell.
reading: uniTerm README
4. OpenHuman gives a big tool result an address instead of replaying it
OpenHuman , marked early beta, describes a harness that compresses large tool results before the model reads them. For very large results, it says the model receives a searchable handle rather than the whole payload; the project links its own benchmarks and methodology alongside the claim.
The model can ask for one piece later instead of receiving the full result again. The builder still has to test whether that retrieval preserves the evidence it claims to keep.
reading: OpenHuman README
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- AI coding assistants — compare coding workbenches by review surface, permissions, cost, logs, and escape hatches
- Browser agents — where browser automation helps, where it lies, and how to keep it observable