You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Tracked in ROADMAP.md — the narrative there carries the full design rationale + the per-bullet checkboxes. This issue exists for board-level Status / Area / Track / Milestone tracking.
A fixture OpenAPI document served from a local HTTP server, swapped underneath a running watch so the poller saw a real change: /cranes added, /doomed removed, and listTugs keeping its name while its response shape gained two properties. That last one is the case worth having — a saved request survives syntactically and fails at runtime.
All three delta kinds surfaced, in the sidebar, from one poll:
where
what it showed
banner
+1 method, −1 method, ~1 changed, naming the removed Fleet / GET /doomed
service header chip
+1 −1 ~1, with a per-kind tooltip
method rows
+ on listCranes, ~ on listTugs
Then reverted the schema and got the mirror image — /cranes reported removed, goingAway marked added, listTugs marked changed again. Console clean throughout.
Two behaviours worth recording, because neither is obvious from the code
A quiet poll does not clear the delta. It persists until dismissed. That is the right way round: clearing on the next poll would make a change visible for one interval and then vanish, so you would have to be looking at the right moment to ever see it.
Dismissing clears all three surfaces at once — banner, chip and row marks — and leaves the watch running.
#551 stopped building method rows for collapsed services, so a row-level +/~ only exists once a group is expanded. The per-service chip is built for every service regardless, which is what keeps a change findable in a collapsed tree. Checked deliberately rather than assumed, because the row marks would otherwise have been silently unreachable for anyone with a large catalogue.
What this issue does not cover
The freeze investigated earlier in the v2.4 cycle was triggered by this poller but was never a property of it: the mechanisms found were unbounded observer accumulation (#550), an unbounded SSE reconnect loop (#552), and the render-path costs in #551. The first two are fixed and closed; #551 has one item still open (coalescing beyond the live-stream panels). The freeze itself was never reproduced under a profiler — that residual doubt belongs to those issues, not to this one, but it should not disappear just because this closes.
Covered by 24 unit tests in schema-watch.test.mjs; full JS suite green at 446.
Activity
chore(planning): v2.0 → v1.9 milestone + draft-release-notes workflow…
feat(workbench): schema watch reports what changed, not how many (#48)
fix(workbench): a failed schema-watch poll must not report a delta (#48)
fix(workbench): stop schema watch from saturating the main thread (#48)
fix(workbench): guard imperative wiring with a property, not an attri…
thomas-stegemann commentedon Aug 3, 2026
Verified end to end and closing.
What was checked
A fixture OpenAPI document served from a local HTTP server, swapped underneath a running watch so the poller saw a real change:
/cranesadded,/doomedremoved, andlistTugskeeping its name while its response shape gained two properties. That last one is the case worth having — a saved request survives syntactically and fails at runtime.All three delta kinds surfaced, in the sidebar, from one poll:
+1 method, −1 method, ~1 changed, naming the removedFleet / GET /doomed+1 −1 ~1, with a per-kind tooltip+onlistCranes,~onlistTugsThen reverted the schema and got the mirror image —
/cranesreported removed,goingAwaymarked added,listTugsmarked changed again. Console clean throughout.Two behaviours worth recording, because neither is obvious from the code
A quiet poll does not clear the delta. It persists until dismissed. That is the right way round: clearing on the next poll would make a change visible for one interval and then vanish, so you would have to be looking at the right moment to ever see it.
Dismissing clears all three surfaces at once — banner, chip and row marks — and leaves the watch running.
Interaction with #551
#551 stopped building method rows for collapsed services, so a row-level
+/~only exists once a group is expanded. The per-service chip is built for every service regardless, which is what keeps a change findable in a collapsed tree. Checked deliberately rather than assumed, because the row marks would otherwise have been silently unreachable for anyone with a large catalogue.What this issue does not cover
The freeze investigated earlier in the v2.4 cycle was triggered by this poller but was never a property of it: the mechanisms found were unbounded observer accumulation (#550), an unbounded SSE reconnect loop (#552), and the render-path costs in #551. The first two are fixed and closed; #551 has one item still open (coalescing beyond the live-stream panels). The freeze itself was never reproduced under a profiler — that residual doubt belongs to those issues, not to this one, but it should not disappear just because this closes.
Covered by 24 unit tests in
schema-watch.test.mjs; full JS suite green at 446.feat(workbench): schema-watch diff view — what changed since you last…