Skip to content

Goals: run goal checks (and trigger ifs) inside the bot's sandbox #355

Open

Description

@lydakis

One follow-up remains from Codex's review of #341 (Sonny goals). It was left out of that PR on purpose, by the owner's call.

Run goal checks (and trigger ifs) inside the bot's sandbox

A goal's check runs as the person and outside every sandbox, the same way a trigger's if does. #341 strips the provider keys in ~/.agent/env from the check's environment. Setting a goal also waits for an approver verdict, like triggers, and that verdict sees the check command. Even so, a hostile repo could still get a build script run unsandboxed by having an agent set cargo test as its check.

Running checks truly sandboxed would need a daemon op that runs a command under a given bot's sandbox policy, with its credentials scrubbed. That's a daemon change, so it needs discussing first. The same op would cover a trigger's if. (thread)

Done

Activity

  1. self-assigned this
    on Oct 29, 2026
  2. lydakis commented on Oct 29, 2026

    @lydakis
    OwnerAuthor

    Another follow-up from the #341 review: keep a goal's outcome until the lead hears it. When a check settles, the goal file is removed right away, and the outcome lives only in the lookout's memory until the lead's note goes out. If the lookout restarts in that window, the replay finds no goal, and the lead hears a plain completed turn with no word on how the goal ended. Fix: keep the outcome on disk (for example, in the goal file, marked ended) until its note is submitted, and have the replay read it from there. (#341 (comment))


    Generated by Claude Code

  3. changed the title [-]Goals: leftovers from #341's review[/-] [+]Goals: run goal checks (and trigger ifs) inside the bot's sandbox[/+] on Oct 31, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions