If I get distracted, an agent yells at me
August experiment: when I start a session, a task initiation gate reads the prompt and compares it to my day plan. If the task isn’t on the critical path, it challenges me:

After the challenge, I can type #a to capture and close the pane, or give a justification for why the task is kosher. If the agent agrees with the justification, it waves it through. If not, it blocks again:

The gate takes 0.1–5 seconds to run.1 Decisions are logged, of course, so we can iterate on the criteria. I’m early with this, but I think it’s a keeper.
Footnotes
Implemented as a
UserPromptSubmithook. Deterministic “green light” conditions wave some prompts through ~instantly. Those that need LLM judgement go to a fast model (currently: Luna) so the decision is made in 1–5 seconds. ↩
