Skip to content

Course-correction template orders every agent to stop and wait, even on Warning — costs an extra turn per correction #42

Description

@batcave-official

What happens

build_correction_injection uses one hardcoded template for every verdict, and it ends with:

Ask for permission to take that next step, then stop and wait for the user's answer.

So a Warning injects exactly the same stop order as OffTrack. Across all corrections delivered on our machines (~375, roughly half of them warning), every delivery was advisory at the harness level (exitCode: 0, never a blocking exit) — the stop comes purely from the template text. The practical effect: each correction forces an extra assistant turn (report + wait for the user), which the operator pays for, and on false positives it stalls an authorized session outright.

What we'd like

Either make the correction tail verdict-aware (advisory "take into account and continue" for warning, stop order reserved for off_track), or make the tail configurable. Related gap: the judge prompt lists both "warning" and "off_track" as verdict values but gives no criterion distinguishing them, so in practice they are chosen arbitrarily and behave identically.

What we run locally

We carry a local patch that makes the template advisory for every verdict (mention the flagged step in the next user-facing report, keep working, stop only before a destructive/hard-to-reverse step). Happy to send a PR in whichever shape you prefer — verdict-split, fully advisory, or config-keyed.

Tested against v0.2.2.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions