Repository navigation
perf(workflow): engine overhead dominates a long uncapped for_each #1162
Description
Problem
A for_each over 1,000 items with no max_parallel takes about 42 s with --driver local, even when each item only runs true. The commands take almost none of that time: the engine takes nearly all of it.
Shape (real CLI, --driver local) |
Whole run |
|---|---|
| uncapped, 1,000 items | 42 s |
max_parallel: 8, 200 items |
3.5 s |
Where the time goes
each next / report about 2 calls per item
read run file parses the whole file through the schema
reconcile settles every running item's workflow
write run file writes the whole file
Every call pays both costs, and both grow with the run. Once the settling work is cut, parsing the file is about a third of the run.
Options
| Option | Gain | Catch |
|---|---|---|
| Validate the run file more cheaply on read, for example without the per-step schema pass | up to about a third | that read is what guards against a corrupt or hand-edited file |
| Persist what is left to reconcile, and settle only those scopes | the settling cost | an edited file must still fall back to a full reconcile |
Tried and dropped in #1148: an in-memory incremental reconcile. It was correct, but it only took the run from 41 s to 33 s, and only when one process drives the whole run.
Done when
A 1,000-item uncapped run spends most of its time in its commands, measured with the real CLI.
Activity
SaulMoro commentedon Oct 20, 2026
#1175 adds work to
reportthat grows with the run. On each successful report that carrieschanges,branchStillUsedplans the branch of every unsettled step (engine.ts:1141-1149), and each plan runsplanWorkspace, which scansrun.stepsseveral times (ancestorsOf,itemsOf,itemsAfter). In an uncappedfor_eachwithworkspace: new, that is O(N²) per report. I have read this from the code, not measured it. A 1,000-itemworkspace: newrun spends most of its time ongit worktree addand commits, so I'd add it to the measurements with engine time perreportkept apart from git and tool time, and check that it stays flat from 100 to 1,000 items. A branch-to-unsettled-users index in the persisted reconcile state is one fix, if the measurement calls for it.