mirror of
https://github.com/EveryInc/compound-engineering-plugin.git
synced 2026-11-02 23:11:29 +01:00
Trevin Chow
edd50dbf0f
A widened window is the agent's judgment about evidence it looked at. When that evidence moves the judgment may no longer hold, so the watch now hands the decision back on `review-evidence-moved` instead of sleeping the window out against a state that has since changed. settle.md carries the goal: if what moved accounts for the review being waited on, the ordinary window decides again. The engine reports the movement and concludes nothing. That is the whole difference from the earlier attempt at this, which keyed on a state value — unchanged evidence — and so re-fired every poll on a judgment the agent had already made, looping the wake. Movement cannot do that: unchanged evidence never triggers it, by construction. Regression test pins all four directions, including the two that matter: unchanged evidence stays silent however long the window has run, and the widened bound still decides when it genuinely elapses. Its comment records why the anti-loop half exists so a later restatement cannot drop it.