Replace a stopped run in place
Amends ADR 0009. In the first clean agent adoption of lutzseverino/postern, a stopped run could only be ended: correcting its scope meant abandoning it, committing or discarding its changes, and adopting again, and after abandonment the agent restarted from a fresh clone, repeating every answer. On 2026-10-19 the owner asked for a recovery path that does not involve restarting; #312 records the contract.
A scope correction is still a new run with a new confirmed scope, and an
active run’s scope still never changes: scope amendment stays removed. The new
run no longer needs the old one abandoned and resolved by hand. A changed scope
answer to a run stopped for its scope, or adopt in a checkout where a run was
abandoned, replaces that run in place: its report is archived as abandonment
archives it, the unchanged selection is inspected with the new scope, and a new
run starts over the changes the old one left. The Git start blockers exempt
exactly the uncommitted state the old run’s last observation recorded, and the
last complete adoption an update replaces is read from the HEAD the old run
started from. Fixes run again because they are repeat-safe, the preserved
contextual work is reassessed under the new scope, and the new run’s change
set covers every change since that HEAD. A scope answer is Git-undoable, so
the agent may change it without asking.
Abandonment exits 0 once it succeeds, and a fix’s declined overwrite defers the whole fix instead of leaving abandonment as the only alternative: the run records it as deferred and continues, and every later run runs it again.
Consequences
Section titled “Consequences”Abandonment archives the run’s saved installation and its last observation beside its report, so a later start can verify what the run left and carry its changes. The run, state, inspection, and adoption report formats keep their keys: a deferred fix uses remote setup’s deferral record, the scope questions of a stopped run are ordinary answers-file questions, and the changes a replaced run left travel in the new run’s saved installation. The replaced run’s own evidence stays in its archived report, not in committed state, which still holds the current run only (ADR 0011).
Any uncommitted change the old run did not record still stops the agent with
DIRTY_PROJECT, under #307’s stop-and-tell rule, and selecting another CLI,
source, version, or profile does not adopt in place.
Considered options
Section titled “Considered options”- Keep correcting by adopting again from a clean checkout. Its rough edge was meant to be paid rarely, but the first clean agent adoption paid it and restarted from a fresh clone.
- Amend the active run’s scope. ADR 0009 removed that machinery: replaying fixes and revalidating outgoing scope inside one run cost every release. A replacement run reuses the ordinary inspection, start, and change-set machinery instead.
- Apply part of a declined fix. Creating missing settings while skipping overwrites needs a fix-protocol change across standards sources; deferring the whole fix needs none.