Add harness-level multi-agent tests and controls #51

Open
opened 2026-10-24 23:26:55 +00:00 by TheComputerGuy · 0 comments

Problem

Chock implements bounded subagent trees, but the roadmap notes that full trees of real provider-backed chock run sessions have not been exercised end to end. Users also need a clear view of delegated budgets, progress, failures, and cancellation.

Proposed outcome

Build provider-backed and deterministic-fixture integration suites for real agent trees, then expose tree status and controls in the CLI, daemon, and editor. Preserve budget partitioning and inherited policy restrictions across every branch.

Acceptance criteria

  • Tests cover depth, width, concurrent children, reviewers, cancellation, rate limits, and restart.
  • Parent and child token/cost budgets reconcile exactly.
  • One branch can be inspected or canceled without corrupting siblings.
  • Every child gets a terminal completion record, including crashes.
  • Policy and self-restriction inheritance are verified end to end.
### Problem Chock implements bounded subagent trees, but the roadmap notes that full trees of real provider-backed `chock run` sessions have not been exercised end to end. Users also need a clear view of delegated budgets, progress, failures, and cancellation. ### Proposed outcome Build provider-backed and deterministic-fixture integration suites for real agent trees, then expose tree status and controls in the CLI, daemon, and editor. Preserve budget partitioning and inherited policy restrictions across every branch. ### Acceptance criteria - Tests cover depth, width, concurrent children, reviewers, cancellation, rate limits, and restart. - Parent and child token/cost budgets reconcile exactly. - One branch can be inspected or canceled without corrupting siblings. - Every child gets a terminal completion record, including crashes. - Policy and self-restriction inheritance are verified end to end.
Sign in to join this conversation.
No description provided.