improvement(perf): bound trash, file-path and attachment reads

[waleedlatif1]

Summary

• Recently Deleted reads one bounded page at a time. The settings page loaded every archived workflow, folder, table, knowledge base, file, file folder and chat in full (tables also ran a per-row size subquery and a jobs lookup) and then searched and sorted in memory. It now calls a new GET /api/workspaces/[id]/recently-deleted, which returns 50 rows per page with keyset pagination. Each type's query is limited to limit + 1 rows, in page order, using its (workspace_id, deleted) partial index. The query reads one UNION ALL of those per-type queries and only the columns the list renders. Search and sort run on the server and the UI scrolls infinitely. Ties still break name then id, A→Z, in either direction.
  • Authorization runs in a new application use case: workspace read, then each type's own list capability (a withheld capability hides that type and its folders), and chats only for the viewer's own.
  • Existing per-type endpoints, the v2 API and other UI are unchanged.
• Copilot path lookups no longer list every workspace file. run_code file mounts and the file-preview target resolution now share one batched candidate query (findListedWorkspaceFiles, built on the display-name index candidate query) under the same files.list authorization. It keeps the listing's exact matching rules.
  • Directory mounts page the folder subtree in SQL and are capped at the mount limit.
• Chat attachment reads no longer load the transcript. read_attachment reads only the fileAttachments entries whose id or filename matches, through the same lateral-array pattern as the MCP server lookup.
• KeysetKey gains an optional fixed direction for ascending tiebreakers, and keysetOrderBy, so the trash reuses the shared keyset engine.
• useLoadMoreSentinel is extracted from Resource and shared.

Type of Change

• Improvement (performance)

Testing

• New integration suites run in CI:
  • lib/workspaces/__integration__/recently-deleted.integration.ts: per-type predicates, page walks at limits 1/2/3 equal the unpaged order for every sort, microsecond page boundaries, search escaping, type narrowing, cursor validation.
  • lib/mothership/chat/file-attachments.integration.ts: equivalence with the transcript scan, malformed entries, deleted messages.
  • A new case in file-names.integration.ts checks that the batched and single lookups pick exactly the file the full listing picked.
• Updated unit tests pass: read-attachment.test.ts, stream.test.ts, function-execute-file-mounts.test.ts, list-query*.test.ts, workspace-file-reference.test.ts, list-workspace-files.test.ts, folders/tables/mothership-chats hook tests.
• These checks pass locally: biome on changed files, check:api-validation:strict, check:react-query, check:boundaries, check:client-boundary, check:utils, check:comment-hygiene, check:test-patterns, check:file-names, and the permission-group enforcement audit.
• Not run locally, left to CI: full type-check, check:audits, and the integration suites. The UI has not been verified in a browser.

Checklist

• Code follows project style guidelines
• Self-reviewed my changes
• Tests added/updated and passing (new tests pass the test-audit authoring gate)
• No new warnings introduced
• I confirm that I have read and agree to the terms outlined in the Contributor License Agreement (CLA)

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
docs	Ready	Preview	Oct 12, 2026 10:02pm UTC

[cubic-dev-ai]

All reported issues were addressed across 33 files

<sub>Heads up: you’ve reached your flex budget. Increase your flex budget or wait for usage to reset.

Turn on auto-fix | Re-trigger cubic</sub>

[greptile-apps]

Confidence Score: 5/5

[Medium impact] The PR appears safe to merge; no new blocking issue was found.

Summary

This PR bounds Recently Deleted pages, narrows workspace file lookups, and reads matching chat attachments instead of whole transcripts.

The latest changes:

• Recently Deleted searches and sorts results on the server as you load more pages.
• Copilot resolves file mounts without listing every workspace file.
• Chat attachment reads fetch only entries that match the requested name.

Previous threads:

• The chat ID mismatch is fixed.
• The hook interface requirement is satisfied.
• waleedlatif1 accepted the remaining chat-page sorting cost and deferred the index migration. They reported acceptable measured latency and wanted to avoid conflicting with a sibling migration PR.

No new actionable findings were accepted. Checks were limited to source review; no tests or browser checks were run.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart LR
  UI[Recently Deleted] --> Route[Request one page]
  Route --> Access[Check workspace and resource access]
  Access --> Query[Combine bounded reads]
  Query --> Page[Rows and next cursor]
  Page --> UI
  UI --> Retry[Retry failed page]
  Retry --> Route

Loading

_{Reviews (2) · Last reviewed commit: "fix(recently-deleted): cast chat ids in ..." · Reviewed by Greptile}

[waleedlatif1]

@greptile

[waleedlatif1]

@cubic-dev-ai review this PR

[cubic-dev-ai]

@cubic-dev-ai review this PR

@waleedlatif1 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

All reported issues were addressed across 35 files

<sub>Heads up: you’ve reached your flex budget. Increase your flex budget or wait for usage to reset.

Turn on auto-fix | Re-trigger cubic</sub>