Objective: The environment decides done, not the model. Board items carry machine-readable acceptance (commands, required/forbidden files, endpoint assertions, review-required). Before risky ops, auto-checkpoint (commit/stash + diff + task/db state). On failure, restore the pre-change checkpoint and retry a clean task with a structured failure bundle — not by stuffing the old conversation back into context.
Fits the north star: amux owns verification and recovery. Generalizes the existing board done/verified gate into a reusable verification engine (tests, linters, browser assertions, deploy health, custom scripts). The model can improve indefinitely while success criteria stay external and deterministic. Loop: agent reports done → amux runs gates → failures become evidence → same/different model receives only the failure bundle → task can't advance until gates pass or a human waives.
Scope
- Verification engine runners (shell/test/lint/browser/deploy-health/custom).
- Checkpoint create/restore (git + working-tree diff + task/db snapshot + pending messages + runtime metadata).
retry --from last-good wired to the failure bundle.
Depends on: event sourcing; task envelopes.
Avoid: asking the model whether it finished.
Part of #46 (amux harness roadmap).
Objective: The environment decides done, not the model. Board items carry machine-readable acceptance (commands, required/forbidden files, endpoint assertions, review-required). Before risky ops, auto-checkpoint (commit/stash + diff + task/db state). On failure, restore the pre-change checkpoint and retry a clean task with a structured failure bundle — not by stuffing the old conversation back into context.
Fits the north star: amux owns verification and recovery. Generalizes the existing board done/verified gate into a reusable verification engine (tests, linters, browser assertions, deploy health, custom scripts). The model can improve indefinitely while success criteria stay external and deterministic. Loop: agent reports done → amux runs gates → failures become evidence → same/different model receives only the failure bundle → task can't advance until gates pass or a human waives.
Scope
retry --from last-goodwired to the failure bundle.Depends on: event sourcing; task envelopes.
Avoid: asking the model whether it finished.
Part of #46 (amux harness roadmap).