[fix] Bound the sidebar agent-list query and page the session list [AGE-4218] - #6525
Open
ashrafchowdury wants to merge 11 commits into
Open
[fix] Bound the sidebar agent-list query and page the session list [AGE-4218]#6525ashrafchowdury wants to merge 11 commits into
ashrafchowdury wants to merge 11 commits into
Conversation
The rail rebuilt every row object and re-rendered every row on each 15s/60s poll: toRow minted a fresh SidebarConfig per ref, the children memo depended on the whole entity sources object so any one source rebuilt all four, and no row component was memoized. Cache each row on the ref it was built from, memoize per entity, and memo the row components. A poll that changes one session's flags now re-renders one row.
maxItems for Sessions was the same number as the fetch limit, so the cap could never bite and the Show all row it guards could never render — you scrolled to row 50 and the list simply ended, with no way to reach the rest. Lift the cap instead of raising it: Sessions renders every ref it holds, and the fetched window becomes purely a page size. Show all goes with it; paging replaces it. maxItems keeps its meaning for every other entity, and still applies before grouping.
Scrolling to the bottom of the Sessions group ended the list at 50 rows with no way to reach the rest. Add a second query for sessions older than the polling head, widened one page at a time as the scroll box reaches its end. It is deliberately not polled: an interval over an accumulating list costs a request per page per tick. Liveness stays on the head, which every changed row re-enters — activity ordering means a session that just did something is newest — so a stale copy below always loses the dedupe to the fresh one above. Widening the window rather than appending a cursor page also re-reads the tail, so a row archived elsewhere leaves it. The page count resets when the filters change, and paging is suppressed mid-drag: the reorder engine caches every row's rect at dragstart.
Two latent bugs that only bite once more rows than the old cap can be loaded. The zone cap trimmed with a blind slice, so a drop in a large bucket wrote every id, discarded whatever fell past the cap, and left those rows unknown to the arrangement — where they led their bucket. One drop reordered the list. Evict the ids that are NOT on screen instead, newest first, so a visible row never loses its place. Unseen sessions led on the rationale that you had just started them, which is only true at the head of the list. Rows arriving from a later page are older than everything arranged, so place unseen rows by activity: newer leads, older trails, and a row with no activity cannot claim to be new.
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Contributor
|
Important Review skippedAuto reviews are disabled on base/target branches other than the default branch. 🗂️ Base branches to auto review (1)
Please check the settings in the CodeRabbit UI or the ⚙️ Run configurationConfiguration used: Repository YAML (base), Organization UI (inherited) Review profile: CHILL Plan: Team Run ID: You can disable this status message by setting the Use the checkbox below for a quick retry:
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
Contributor
Railway Preview Environment
|
Seeds a project with a deep workflow catalog and a pile of sessions, to reproduce and measure the #6390 sidebar stall. The cost scales on revisions per workflow, not on agent count, so the script seeds depth rather than breadth. Workflows, variants and revisions go through the HTTP API so is_agent is inferred by the real commit path; sessions are inserted directly because the streams route drives the runner and needs provider keys.
… menu opens The Sessions group is alwaysOpen, so gating the Agent facet's catalog on the group gated nothing: SessionFilterMenu renders as the group's action on every page, so every sidebar mount pulled the whole agent catalog — one revision fetch per workflow — to fill a menu nobody had opened. Gate it on the menu itself instead, and give the facet a loading row so the deferred catalog never reads as "this project has no agents".
A caller that wants one fact per workflow — is this an agent? — had to download every revision of every workflow to compute it, because the artifact list does not carry is_agent and the revision query has no way to say "just the head". Without windowing the DAO applies no LIMIT at all, so 20 apps x 30 revisions came back in full. latest_per_artifact goes on the SHARED RevisionQuery, not on the workflow subclass: the workflow service downcasts to RevisionQuery before calling the DAO, and Pydantic's extra=ignore would drop a subclass-only field there without an error. Explicit revision_refs still win over it, archived revisions are excluded as before, and the placeholder a new variant is seeded with is skipped so a multi-variant workflow resolves to its real head.
…quest Six call sites each downloaded every revision of every workflow to compute one boolean per workflow, under keys that shared nothing. The sidebar's copy also ran serially behind the artifact list and re-fetched every app's classification whenever any one app's updated_at changed. Replace them with one project-wide Map<workflowId, isAgent> behind a single query key: fetchWorkflowAgentFlags asks the server for latest_per_artifact, so one revision per workflow comes back instead of the whole history, and it parses that payload once rather than twice. Each list keeps its own query — they show different subsets — and joins against the map. Also replaces api.ts's timestamp-first "latest revision" picker with the version-first rule store.ts already uses, and fixes the last timestamp-first sort in the latest-revision-id fallback.
Two consumers in a project must share one fetch, and a workflow absent from the map must classify as a prompt — the rule the old revision-map path got for free and the new map has to state.
Sharpens the acceptance tests after checking the behaviour against the running stack: an unchanged commit is a no-op (so revisions need distinct data), a variant with no commits has no revision rows at all, and a FIRST commit lands on version "0" carrying data — which is why the placeholder guard keys off data and flags rather than the version alone.
- The older-pages query keyed the project id at index 2, so the shared placeholder helper - which reads index 1 - never matched and the tail blanked on every page. - Mobile was uncapped but had no way to page: it never wired onReachEnd, and with the cap lifted its Show all row could not render either, so it sat at one page with no affordance for the rest. - The end-of-scroll guard keyed on scrollHeight, so a page whose rows all landed in collapsed groups added no height and latched paging off for good. Throttle instead. - The tail boundary was read live off the polling head, so every poll that shifted the head refetched every loaded page. Freeze it when the first page is asked for. - Drop the four classification helpers the agent-flags map replaced, the page-size export left behind by a rename, and the tests pinning them. - Compare ref fields instead of serializing each row, and record why the new classification call is still on axios.
ashrafchowdury
force-pushed
the
fe-feat/sidebar-unbounded-sessions
branch
from
September 4, 2026 10:38
6edbca2 to
932bbeb
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Context
Two problems in the same rail.
Opening the sidebar loaded the whole revision history of every workflow in the project.
is_agentlives on a workflow revision, never on the artifact, so classifying the Agents group meant asking/workflows/revisions/queryfor every revision of every app and keeping only the newest of each. On a project with 64 workflows and 4,209 revisions that is 65 MB and 15 seconds, measured on localhost with no network latency, to produce 64 booleans. It also got worse every day a project was used, because the cost tracks commit history rather than agent count.Separately, the Sessions group stopped at 50 rows with no way past them.
maxItemswas set to the same number as the fetch limit, so therefs.length > maxItemstest that guards the "Show all" row could never be true and that row never rendered. You scrolled to row 50 and the list simply ended.Changes
Bounding the classification (#6390).
RevisionQuerygains an optionallatest_per_artifact. When set, the DAO returns one revision per artifact withDISTINCT ON (artifact_id) ... ORDER BY artifact_id, id DESC, skipping the empty placeholder a new variant is seeded with.Ordering is by
id, notversion, and that matters.versionis aStringcolumn, soORDER BY version DESCsorts"9"above"10". It is also assigned per variant (a count of earlier ids within the variant), so "max version per artifact" is not well defined.idis UUID7, and within a variant its order is exactly version order.On the same seeded project:
The frontend side collapsed six independent fetches of that data into one project-wide
agentId -> booleanmap. The sidebar, the agents page, the prompts page and app-management now share a single cache entry and join against it. The map's key also dropped its per-workflowupdated_atlist, so committing to one agent no longer re-classifies every app in the project.Paging the sessions.
maxItemsfor Sessions is now unbounded, so the group renders every row it has loaded, and a second query fetches pages older than the polling head as you reach the end of the scroll box.That tail query is deliberately not polled. A refetch interval over an accumulating list costs one request per loaded page per tick. Liveness stays on the head, and every changed row re-enters it, because activity ordering means a session that just did something is the newest one. A stale copy in the tail always loses the dedupe to the fresh copy above it. The tail boundary is frozen when the first page is requested rather than read live off the head, or every head poll would re-key the tail and refetch all of it.
Two supporting fixes the paging exposed. Rows used to be rebuilt on every 15s poll:
toRowminted a fresh object per ref, the children memo depended on the whole entity-sources object so any one source rebuilt all four, and no row component was memoized. Rows are now cached per entity and compared by field, andLeafRow,RowLabelandGroupLabelRoware memoized. A poll that changes one session now re-renders one row.The hand-arranged drag order also had two bugs that only bite once more rows than the old cap can be loaded. The zone cap trimmed with a blind
slice(0, 200), so a drop in a large bucket wrote every id, discarded whatever fell past the cap, and left those rows unknown to the arrangement, where they led their bucket. One drop reordered the list. Eviction now drops the ids that are not on screen first. Separately, unseen sessions led on the rationale that you had just started them, which is only true at the head; rows arriving from a later page are older than everything arranged, so unseen rows now place by activity.Tests
latest_per_artifactagainst the real revision semantics, including the lexicographic guard (a workflow with 10+ revisions returns"10", not"9") and the multi-variant placeholder case.@agenta/navigation95 tests,@agenta/entities1,501 tests.navigation,navigation-ui,entitiesandmobiletypecheck and lint clean.api/oss/tests/manual/workflows/seed_stress_catalog.pyto reproduce the numbers above. It seeds depth rather than breadth, because the cost scales on revisions per workflow, not on agent count.Notes for reviewers
sidebarReorderActiveAtomandentity.dragZonefrom that branch.fetchWorkflowAgentFlagsuses raw axios rather than the Fern client.latest_per_artifactships in this change, so the generated client does not carry it until the OpenAPI spec is regenerated. Flagged in the docblock to migrate with the rest of that file.activity: 7dandtype: chat. Paging removes the count ceiling, not those filters.What to QA
Needs a project with more sessions than one page.
seed_stress_catalog.pywill make one./m. Mobile shares this code and pages the same way.