ADR 0008 — True suspend/resume of pending_approval (Q8, v2)
Date: 2026-06-11 · Status: accepted
Context
The contract froze the task-state enum at v0.1 and split Q8 in two: v1 ships the
schema (pending_approval is a state; resume_of and approval_verdict are
reserved fields), and v1 behaviour is stop-and-rerun — approving a suspended
task meant launching a brand-new task with --resume-of, which re-did all the
work and re-hit the same gate. v2 owes true suspend/resume with zero schema
change — the fields are already there.
Decision
The worker is disposable and stateless, so "resume" cannot reattach to a dead process. It means: continue the work with the approval granted. Approving a suspended task dispatches a fresh worker run that carries:
resume_of— the suspended task's id (the audit link, already in v1); andapproval_verdict— anApprovalVerdict(approved=True, actor, ts)(reserved at v0.1, populated now).
The worker honours the verdict: a granted verdict runs the orchestrator in a bounded auto mode instead of manual mode, so it proceeds past the policy gate that stopped the original run instead of re-stopping. The mode is deliberately not blanket auto-approval: commit, network, and outside-workspace actions still require explicit approval and will re-suspend — the resume is bounded, and bounded further by the Stage A sandbox and patch-as-approval.
On the supervisor side, review --approve branches on state: a completed task
means "apply the patch" (ADR 0002, unchanged); a pending_approval task means
"resume". The resume inherits the original's permissions and budget from the
audited task envelope, the original's approval row is resolved approved with a
note linking the resume it produced, and the resumed run is re-verified (ADR
0006) like any other completion.
Consequences
- Zero schema change:
resume_of+approval_verdict+ the frozen state enum carried it, exactly as Q8 predicted. No contract bump, no fixture regen. - The difference from v1 is real and proven: the same plan that stops
pending_approvalwith no verdict runs tocompletedwith one — shown both with a contract worker (a benign gated action) and end-to-end throughreview --approve. - Granular per-action approval is out of scope: a stateless re-run re-plans, so the grant is "proceed past gates (except commit/network/outside-workspace)", not "approve action id X". Recorded as the deliberate v2 semantics.
run --resume-of(v1 stop-and-rerun) still works for the cases where starting over is what you want.