self.md radar — 2026-09-27
four ways to make an agent session leave a receipt
A check that cannot run should not wear a green badge. A whiteboard can now point at a live coding session; a menu-bar app can find its orphan processes; an MCP selector can keep specialist servers out of the room until they are needed. These are small control surfaces, but each is trying to make the tail of an agent run legible.
1. keel gives an unrun check its own exit code
keel
is a macOS harness for AI-assisted delivery. Its gates return pass, fail, or blocked; the last is exit code 3, for a check that could not run. Missing credentials, a stale index, or an absent tool are not quietly promoted into success, and they are not filed as an agent mistake either.
The project makes another useful admission on its own front page: G2/store-drift has not failed in 26 runs across two repositories. That can mean the check is working on a quiet system, or it can mean nobody has yet found the condition it was meant to catch. keel labels the ambiguity instead of borrowing certainty from a green dashboard. Its broader performance evidence is still young: two repositories, one operator.
reading: keel project site
2. drawgent turns a diagram note into an agent request
drawgent
connects a local Excalidraw canvas to Claude Code, Codex, or opencode. A note beginning AGENT: can refer to nearby shapes; the tool sends that context to the session and resolves the note to DONE after the work is handled. The editor and API bind to 127.0.0.1, and the scene lives in .drawgent/scene.json inside the workspace.
The attachment mechanics are not flattened into one sales sentence. opencode can receive messages in the running session; Codex uses codex queue --thread; Claude Code gets a fork carrying the conversation because its terminal session has no public injection path. A canvas is useful here precisely because it has to say which kind of connection it is offering.
reading: drawgent documentation
3. Leftovers names the processes after the agent leaves
Leftovers 1.0
is a newly released macOS menu-bar app for processes and agent sessions left running after work ends. Its project describes normal app quitting rather than force-quitting, a separate confirmation before ending an agent session and its child processes, plus duplicate MCP-server detection. It also offers --report, --json, and --clean modes.
That is a modest claim, which is good. Process attribution is messy, and “clean” can become an expensive verb fast. The useful move is making the leftover state inspectable before it becomes a fan spinning at 2am or another server quietly holding a port.
reading: Leftovers repository and release
4. withmcp keeps extra servers out of the default room
withmcp selects Model Context Protocol servers when launching Claude, Codex, or Pi. It supports profile groups, per-directory rules, and native-server collision checks; authentication stays inside the chosen harness. A one-off server can be present in the configuration without being present in every session.
The README also puts one sharp limitation in writing: Codex receives server settings as command-line arguments, so expanded environment values can be visible in the process list. Choosing a tool set is not the same job as securing it. But it is better than pretending every discovered connector belongs in the default context forever.
reading: withmcp repository
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
- Best MCP servers — connect files, browsers, memory, search, and workflow tools without turning the stack into soup