What happened
On dev (seen through PR #1632 at 458d1be73, whose recovery change does not touch this path), a task whose plan finished just before its engine restarted comes back interrupted with no way to finish it.
On 2026-09-28, in a git repo, on the engine road (plain codeaf, deepseek/deepseek-v4-flash): /task write a python script count.py that counts lines in every .txt file here, and a test for it; run the test. While the roster said Running 1, the workspace's codeaf engine daemon was killed. The surface reconnected to a new engine.
The task page then read:
interrupted · 43s · $0.0018
while its own body said done · ran 43s, and that count.py and four passing tests were written (they are on the task's branch). The page offers no enter retry (pressing Enter changes nothing). Asking the conversation to land it made the model call tasks with id: "2", resolve: "accept", which answered:
No task "2" in this project
Replication
Deterministic (no model). Not yet written. The ending is decided in reconcileSettledRun (below) for a plan root whose status is done at recovery time, so a session test can seed a checkpoint with an admitted run row plus a plan store whose root is done, open the conversation, and assert what the row says and what tasks answers for its id.
Field (real models). The steps above: any key the product accepts (OPENROUTER_API_KEY or the profile's key), a cheap model, about 2 minutes and under $0.01. What makes it fire: the plan must finish before the engine dies. Kill the daemon a few seconds after the worker reports its test passing.
Where
reconcileSettledRun in internal/session/task_run_recover.go. Its default case writes the plan finished before the restart; inspect the saved working copy before landing its changes and leaves the row interrupted.- The retry affordance in
internal/tui3/taskretry.go only offers a retry for a failed node without a run, so an interrupted plan run gets none.
- In
internal/session/tools_tasks.go, plan rows are handled for reads, but resolve and continue fall through to the ordinary task lookup, which cannot find a plan-only run row. That's where No task "2" comes from.
The fix
A task whose plan finished before a restart is either offered as finished work to land (the same resolve / accept door a finished task has), or it says what to do next and that door works. The tasks tool must never answer "No task" for a task the page is showing.
Acceptance
End to end, through the engine road: kill the engine after the plan finishes, reopen, and the task can be landed through the page or tasks resolve. The receipt names the landed commit, and the spend shown before the kill is still shown after.
What happened
On
dev(seen through PR #1632 at458d1be73, whose recovery change does not touch this path), a task whose plan finished just before its engine restarted comes back interrupted with no way to finish it.On 2026-09-28, in a git repo, on the engine road (plain
codeaf,deepseek/deepseek-v4-flash):/task write a python script count.py that counts lines in every .txt file here, and a test for it; run the test. While the roster saidRunning 1, the workspace'scodeaf enginedaemon was killed. The surface reconnected to a new engine.The task page then read:
while its own body said
done · ran 43s, and thatcount.pyand four passing tests were written (they are on the task's branch). The page offers noenter retry(pressing Enter changes nothing). Asking the conversation to land it made the model calltaskswithid: "2", resolve: "accept", which answered:Replication
Deterministic (no model). Not yet written. The ending is decided in
reconcileSettledRun(below) for a plan root whose status is done at recovery time, so a session test can seed a checkpoint with an admitted run row plus a plan store whose root is done, open the conversation, and assert what the row says and whattasksanswers for its id.Field (real models). The steps above: any key the product accepts (
OPENROUTER_API_KEYor the profile's key), a cheap model, about 2 minutes and under $0.01. What makes it fire: the plan must finish before the engine dies. Kill the daemon a few seconds after the worker reports its test passing.Where
reconcileSettledRunininternal/session/task_run_recover.go. Its default case writesthe plan finished before the restart; inspect the saved working copy before landing its changesand leaves the row interrupted.internal/tui3/taskretry.goonly offers a retry for a failed node without a run, so an interrupted plan run gets none.internal/session/tools_tasks.go, plan rows are handled for reads, butresolveandcontinuefall through to the ordinary task lookup, which cannot find a plan-only run row. That's whereNo task "2"comes from.The fix
A task whose plan finished before a restart is either offered as finished work to land (the same
resolve/acceptdoor a finished task has), or it says what to do next and that door works. Thetaskstool must never answer "No task" for a task the page is showing.Acceptance
End to end, through the engine road: kill the engine after the plan finishes, reopen, and the task can be landed through the page or
tasks resolve. The receipt names the landed commit, and the spend shown before the kill is still shown after.