Summary
Many larger agent systems need a worker to pause at a safe boundary, wait for host/user approval, and then resume later. predict-rlm would be much easier to embed if it had explicit support for this.
Problem
Without pause/resume or checkpoint-friendly semantics, an approval workflow becomes:
- stop run
- summarize state into text
- restart later from that text summary
That loses fidelity and undermines the benefits of structured recursive execution.
Proposed direction
Support one of:
- actual pause/resume with serialized state
- checkpoint/restart from structured runtime state
- explicit host-mediated approval-required suspension points
Example use cases
- a run wants to cross a policy boundary
- host wants to inspect a proposed patch before allowing application
- long-running runs need to yield back to the embedding system
Acceptance criteria
- documented mechanism for pause/resume or checkpoints
- suspension state is distinguishable from failure
- resumed runs preserve enough context to remain useful
- event stream/trace reflects suspension and resumption
Summary
Many larger agent systems need a worker to pause at a safe boundary, wait for host/user approval, and then resume later.
predict-rlmwould be much easier to embed if it had explicit support for this.Problem
Without pause/resume or checkpoint-friendly semantics, an approval workflow becomes:
That loses fidelity and undermines the benefits of structured recursive execution.
Proposed direction
Support one of:
Example use cases
Acceptance criteria