feat(knowledge): index Confluence image attachments by mzxchandra · Pull Request #8949 · simstudioai/sim

mzxchandra · 2026-10-12T23:16:08Z

Summary

Confluence image attachments (PNG, JPEG, WebP, GIF) were dropped at listing, so tables, scans and screenshots attached to pages never reached knowledge-base search.
This PR indexes them: each image is read once by a Gemini vision model, which writes out what it holds (a table becomes one line per row with its headers, a scan becomes its full text, a screenshot becomes its visible text plus a description), and that text is chunked and embedded like any other document.

Per-source policy. Connectors declare an image policy on their meta (imageExtraction). Confluence is semantic; Google Drive and every other source keep today's OCR path unchanged.
Dedicated Gemini adapter. Transient server and network errors are retried within the extraction's time budget (retryWithExponentialBackoff, as OCR does); throttling and overload defer the pass through provider admission. Error bodies are reduced to Gemini's status and reason codes.
Adapter boundary. image-extraction-gemini.ts calls generateContent with a JSON-schema response, the same way the embedding adapters call their providers. Knowledge processing never goes through the workflow provider layer (which would pull the tool registry into the knowledge page). Keys resolve like embeddings: workspace BYOK, then GEMINI_API_KEY, then the GEMINI_API_KEY_1..3 pool.
Image Attachments setting, opt-in for existing connectors. A new dropdown on the Confluence connector ("Read with AI" / "Don't index"). New connectors are created with it on, through a server-side defaultSourceConfig on the connector meta that every create surface (UI, v2 API, CLI, Sim) applies. An unset value means off, so connectors created before this change never start reading (and billing) images until their owner turns the setting on. No migration. Turning it off removes indexed images on the next sync, outside the soft-delete and mass-deletion hold that guard real source deletions, for both service-account and member-credential connectors.
Decorative images. Logos, icons and images under 64 px complete with 0 chunks and show a "Nothing to index" empty state. Files under 2 KB, SVG, HEIC and AVIF are never listed.
Per-sync cap. KB_IMAGE_EXTRACTIONS_PER_SYNC (default 500) bounds new or changed images read per sync; the rest wait for the next sync, which reports partial. Held-back existing images stay indexed (they are still marked seen). A member sync shares one budget across members, the count survives a listing restart and is saved per batch, and a full resync does not re-read unchanged images.
No key, no failures. Without any Gemini key (self-hosted), images are left out of the sync instead of becoming failed documents.
Retry checkpoint. A vision checkpoint on a new shared processing-checkpoint-store (OCR and embedding checkpoints moved onto it, with keys and format unchanged) means a retried pass reuses the extraction instead of paying twice.
Billing. Each extraction is a separate recordUsage model entry (knowledge-document-vision:<docId>:<passId>), zero for BYOK. Extractions whose document then fails are not billed, matching embeddings today.
Stale embeddings. Completing a pass now always replaces the previous embeddings, so a document that re-indexes to zero chunks drops its old ones.
Runtime. sharp is added to the Trigger.dev bundle for image preparation (orientation, downscale to 2048 px, first GIF frame).

Product update (if relevant)

Teams with Confluence knowledge bases can now search what lives inside images on their pages: schedules and tables pasted as screenshots, scanned letters, UI screenshots in how-to pages, and diagrams.
It is on by default for new Confluence connectors and billed per image as a model call. Existing connectors keep images off until their owner sets Image Attachments to Read with AI.
Docs: knowledgebase/connectors.mdx (Confluence section) and the self-hosting environment variables page.

Type of Change

Testing

Unit tests for the extractor (request shape, schema validation, error mapping, token-limit cut-off, deadline reserve, undecodable images, model fallback), the Gemini adapter, the vision checkpoint, document routing (Confluence images to vision, Drive images still to OCR), billing, the per-sync cap (new and changed images, no-key hold-back, full resync), the Image Attachments opt-out removal, and the Confluence listing and hydration rules.
The connector, sync, document-processing, connector application and outbox suites pass (86 files, 1,085 tests); check:audits and lint are clean.
New integration suite confluence-images.integration.ts against real PostgreSQL, file storage and in-process indexing (synthetic connector listing behind the real Confluence meta; synthetic extraction and embeddings): the cap across syncs with held-back changed images staying indexed, full resync not re-reading unchanged images, opt-out removal and re-enable, opt-in for connectors created before the setting, zero-chunk re-indexing dropping old embeddings, and the no-key hold-back. Breaking the held-back seen stamp on purpose fails the suite. ocr-input-failures.integration.ts now seeds a Google Drive source for its image cases, since Confluence images no longer go to OCR.
End-to-end on a local stack against a synthetic Confluence space (11 attachments: a table, a scanned letter, a UI screenshot, a flowchart, a photo, a logo, an extensionless PNG, an animated GIF, a tiny icon, an SVG and a PDF).
Each of 8 search questions ranked the right image document first; the logo completed with 0 chunks and the empty state; the icon and SVG were never listed; the PDF was unchanged; the cap produced partial syncs until all images were read; a re-sync made no new charges.
This run predates the review follow-ups in the last commit (opt-out removal, cap on changed images, no-key hold-back), which are covered by unit tests.
Comparison of Gemini and OpenAI vision models on a sample of document images favored Gemini on quality and cost, so this PR is Gemini only.

Reviewers, please focus on:

lib/knowledge/connectors/sync-primitives.ts (processDocOps): images held back by the cap or a missing key are never fetched; held-back existing images are still marked seen so reconciliation keeps them.
lib/knowledge/connectors/sync-content-pass.ts (removeDisabledImages): the opt-out deletes image documents directly by media type.
lib/knowledge/documents/service.ts: the unconditional embedding replacement on completion and the separate vision usage entry.
lib/knowledge/documents/processing-checkpoint-store.ts: the refactor should be behavior-preserving for OCR and embedding checkpoints.

To verify on staging: knowledge-processing runs for image documents load sharp in the Trigger.dev worker (local runs processed in-process).

Checklist

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

Screenshots/Videos

A completed decorative image shows a "Nothing to index" empty state instead of bare table headers; the Confluence connector form gains an Image Attachments dropdown between Content Type and Filter by Label.

OCR and embedding checkpoints each carried their own copy of the framed header + body format, the outbox cleanup reservation and the cleanup handler factory. Move them into processing-checkpoint-store so a third checkpoint kind can reuse them.

Confluence images (PNG, JPEG, WebP, GIF) were dropped at listing, so tables, scans and screenshots attached to pages never reached search. - Connectors declare an image policy; Confluence reads images with a dedicated Gemini adapter that returns the image's text, every other source keeps OCR unchanged. - A new Image Attachments setting on the Confluence connector, on by default, so existing knowledge bases pick images up on their next sync. - Decorative images (logos, icons, images under 64 px) complete with no chunks; files under 2 KB and SVGs are never listed. - Each sync admits at most KB_IMAGE_EXTRACTIONS_PER_SYNC new images (default 500); the rest wait for the next sync, which reports partial. - A vision checkpoint lets a retried pass reuse the extraction instead of paying for a second call; usage is billed per image. - Completing a pass now always replaces old embeddings, so a document that re-indexes to zero chunks drops its stale ones. - sharp is added to the Trigger.dev bundle for image preparation.

A completed image with no chunks shows an empty state instead of bare table headers, and the connector and self-hosting docs cover image attachments and the two new environment variables.

Review follow-ups for Confluence image indexing. - The per-sync cap now counts changed images as well as new ones, and a member sync shares one budget across its members. Held-back images that already exist are still marked seen, so they are never reconciled away. - A full resync no longer re-reads unchanged images; their bytes are pinned to the attachment version. - Without any Gemini key, images are left out of the sync instead of becoming failed documents. - Turning Image Attachments off removes the indexed images on the next sync, outside the soft-delete and mass-deletion hold meant for source deletions. - Gemini errors: a response cut off at the token limit, a rejected key or region, and an unknown model fail permanently with an accurate message instead of retrying or blaming the image; an image whose pixel data cannot be decoded fails as invalid. - The admission wait keeps the post-call time reserve, Retry-After uses the shared parser, and a configured model must be a catalog Gemini model so its pricing is known.

vercel · 2026-10-12T23:16:14Z

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

Project	Deployment	Actions	Updated
docs	Ready	Preview	Oct 13, 2026 2:59am UTC

cubic-dev-ai

1 issue found across 38 files

Confidence score: 3/5

In apps/sim/lib/knowledge/connectors/sync-content-pass.ts, a Gemini key resolution failure can leave images unindexed while the pass reports success. Mark the checkpoint deferred when image operations are held.

Prompt for AI agents (unresolved issues)


Check if these issues are valid — if so, understand the root cause of each and fix them. When an issue isn't valid or won't be fixed in this PR, reply in its thread with the reason and then resolve the thread. If appropriate, use sub-agents to investigate and fix each issue separately.


<file name="apps/sim/lib/knowledge/connectors/sync-content-pass.ts">

<violation number="1" location="apps/sim/lib/knowledge/connectors/sync-content-pass.ts:284">
P2: When Gemini key resolution fails, `processDocOps` holds image operations without setting `cycle.imagesDeferred`, so this pass can report success while leaving images unindexed. Mark the checkpoint deferred for key-unavailable images so the sync status reflects the pending extraction.</violation>
</file>

_{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}

greptile-apps · 2026-10-12T23:24:17Z

Confidence Score: 5/5

[High impact] The PR appears safe to merge; no actionable blocking issues remain.

Summary

This PR makes Confluence image attachments searchable through Gemini, with per-connector settings, extraction limits, retry checkpoints, and separate usage charges.

Confluence connectors can list supported image attachments for indexing.
Confluence images can become searchable text through Gemini vision.
Image retries reuse saved readings, and completed passes replace old chunks.
Each sync limits image reads and removes images disabled by the owner.

Diagram

%%{init: {'theme': 'neutral'}}%%
flowchart TD
  A[Confluence attachment listing] --> B{Image reading enabled?}
  B -->|No| C[Leave images out]
  B -->|Yes| D{Key and budget available?}
  D -->|No| E[Hold back image and keep existing document seen]
  D -->|Yes| F[Download image]
  F --> G{Saved extraction available?}
  G -->|Yes| I[Chunk extracted text]
  G -->|No| H[Reserve capacity and call Gemini]
  H --> J[Save extraction checkpoint]
  J --> I
  I --> K[Replace old embeddings]
  K --> L[Record successful extraction usage]

_{Reviews (7) · Last reviewed commit: "refactor(knowledge): tighten image extra..." · Reviewed by Greptile}

…nnectors - New connectors are created with Image Attachments on, through a server-side defaultSourceConfig on the connector meta that every create surface applies; an unset value now means off, so existing connectors only read images once their owner turns the setting on. - The Gemini call retries transient server and network errors within the extraction's time budget; throttling still defers the pass. - Gemini error bodies are reduced to their status and reason codes. - Images under the byte floor complete as decorative even when the source reported no size, and the extraction model must be a generateContent Gemini model. - A member sync that held images back forces a full listing next time, the image count survives a listing restart and is saved per batch, and turning images off removes them for member-credential connectors too. - The disabled-image removal skips rows the delete would refuse. - An integration suite covers the cap, held-back images staying indexed, full resync, opt-out removal, opt-in for existing connectors, zero-chunk re-indexing and the no-key hold-back against real PostgreSQL; the OCR input-failure suite now uses a source that still sends images to OCR.

mzxchandra · 2026-10-12T23:37:37Z

@greptile

mzxchandra · 2026-10-12T23:37:38Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-10-12T23:37:44Z

@cubic-dev-ai review this PR

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

cubic-dev-ai

All reported issues were addressed across 43 files

_{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}

…es on every member run - A Gemini reply that is not valid JSON is retried within the extraction's budget and fails permanently once the attempts are spent, instead of costing a whole document retry each time. - A member sync that skips its content pass (content not due yet) still removes images once Image Attachments is turned off. - The reconciliation integration suite runs with images on, so it keeps exercising reconciliation alone.

mzxchandra · 2026-10-12T23:44:58Z

@greptile

mzxchandra · 2026-10-12T23:44:59Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-10-12T23:45:05Z

@cubic-dev-ai review this PR

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

cubic-dev-ai

All reported issues were addressed across 44 files

_{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}

…traction attempt - Member runs add the lifecycle's purges to the run's deletion counts instead of overwriting them, so image removals after Image Attachments is turned off are reported and logged as purges. - A retried extraction reports the tokens of every attempt Gemini answered, since each one was billed. - The checkpoint header check uses isRecordLike. - The create-time default is tested through the real creation path in the integration suite, replacing a mock-only unit test.

mzxchandra · 2026-10-12T23:53:02Z

@greptile

mzxchandra · 2026-10-12T23:53:03Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-10-12T23:53:09Z

@cubic-dev-ai review this PR

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

cubic-dev-ai

All reported issues were addressed across 43 files

_{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}

An unfinished disabled-image removal now keeps the member run marked as having work remaining, so the next run starts promptly instead of waiting a full sync interval with images still searchable.

mzxchandra · 2026-10-13T00:03:00Z

@greptile

mzxchandra · 2026-10-13T00:03:02Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-10-13T00:03:07Z

@cubic-dev-ai review this PR

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

cubic-dev-ai

No issues found across 43 files

Confidence score: 5/5

Automated review surfaced no issues in the provided summaries.
No files require special attention.

_{Heads up: you’ve reached your flex budget. Increase your flex budget or wait for usage to reset.

You've manually re-run cubic several times on this PR. Each manual re-review checks the full PR again and counts toward your usage quota. To preserve your usage limits, we recommend letting cubic automatically review new commits.

Turn on auto-fix | Re-trigger cubic}

…only when queued - Connectors declare image extraction once on their meta (`imageExtraction: { settingId }`), read by the shared `isImageExtractionEnabled`; drops `ConnectorConfig.imageExtractionEnabled` and the Confluence-only helper. - Turning the image setting off sets `knowledge_connector.image_removal_pending` (migration 0412); the next sync removes the indexed images and clears it, so ordinary syncs no longer scan a connector's documents for images. - Gemini specifics (key, Gemini 3 model resolution, error classification) live in the adapter. Requests send `thinkingLevel: LOW`, and any finish reason other than STOP fails without paid retries. - Image reads get their own admission budget (`vision`, KB_CONFIG_VISION_REQUESTS_PER_MINUTE, default 300) instead of sharing OCR's. - One zod schema validates the reply and generates the request's JSON schema. - The prompt is source-neutral and treats text in an image as content, never instructions; sharp loads lazily so it can only fail image documents. - Full relisting after held-back work is one named rule (`noticeRequiresFullListing`).

…ct fixtures

waleedlatif1 · 2026-10-13T02:27:31Z

@greptile

waleedlatif1 · 2026-10-13T02:27:32Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-10-13T02:27:39Z

@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 51 files

_{Heads up: you’ve reached your flex budget. Increase your flex budget or wait for usage to reset.

You've manually re-run cubic several times on this PR. Each manual re-review checks the full PR again and counts toward your usage quota. To preserve your usage limits, we recommend letting cubic automatically review new commits.

Turn on auto-fix | Re-trigger cubic}

- Every Gemini attempt takes its own admission, and 429/503 records a provider cooldown for the key; the retry delay is read from the body's RetryInfo when no Retry-After header is sent. - The image is base64-encoded once per extraction, rotated upright even when it already fits, and flattened onto white before a JPEG fallback. - Cost goes through calculateBillableModelCost; both indexing charges share one recordIndexingUsage helper. - Reconciliation waits until queued image removal finishes, and the removal walks by id instead of rescanning from the start. - The per-sync image cap is at least 1, so a sync can always finish. - canExtractImages lives with the extractor; isSemanticImage treats an undeclared policy as OCR, so callers drop their fallbacks. - Tests: unique embedding vectors, restored pool keys, waits for dispatched syncs, fake timers in finally, native noise fixtures; drops mock-only assertions.

waleedlatif1 · 2026-10-13T02:55:21Z

@greptile

waleedlatif1 · 2026-10-13T02:55:23Z

@cubic-dev-ai review this PR

cubic-dev-ai · 2026-10-13T02:55:29Z

@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

No issues found across 52 files

Confidence score: 5/5

Automated review surfaced no issues in the provided summaries.
No files require special attention.

_{Heads up: you’ve reached your flex budget. Increase your flex budget or wait for usage to reset.

You've manually re-run cubic several times on this PR. Each manual re-review checks the full PR again and counts toward your usage quota. To preserve your usage limits, we recommend letting cubic automatically review new commits.

Turn on auto-fix | Re-trigger cubic}

mzxchandra added 4 commits October 12, 2026 15:45

feat(knowledge): explain image documents with nothing to index

aa1318a

A completed image with no chunks shows an empty state instead of bare table headers, and the connector and self-hosting docs cover image attachments and the two new environment variables.

vercel Bot deployed to Preview October 12, 2026 23:20 View deployment

cubic-dev-ai Bot reviewed Oct 12, 2026

View reviewed changes

greptile-apps Bot reviewed Oct 12, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/connectors/sync-content-pass.ts

Comment thread apps/sim/lib/knowledge/connectors/sync-content-pass.ts

Comment thread apps/sim/lib/knowledge/documents/image-extraction-gemini.ts

vercel Bot deployed to Preview October 12, 2026 23:41 View deployment

cubic-dev-ai Bot reviewed Oct 12, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/documents/image-extraction.ts Outdated

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts Outdated

vercel Bot temporarily deployed to Preview October 12, 2026 23:44 Inactive

greptile-apps Bot reviewed Oct 12, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts Outdated

Comment thread apps/sim/lib/knowledge/documents/processing-checkpoint-store.ts Outdated

Comment thread apps/sim/lib/knowledge/application/connectors.test.ts Outdated

cubic-dev-ai Bot reviewed Oct 12, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/documents/image-extraction.ts Outdated

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts

greptile-apps Bot reviewed Oct 12, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/documents/image-extraction.ts Outdated

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts Outdated

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts Outdated

vercel Bot temporarily deployed to Preview October 12, 2026 23:52 Inactive

cubic-dev-ai Bot reviewed Oct 12, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts Outdated

fix(knowledge): resume a member run whose image removal hit its deadline

d01a4a8

An unfinished disabled-image removal now keeps the member run marked as having work remaining, so the next run starts promptly instead of waiting a full sync interval with images still searchable.

vercel Bot temporarily deployed to Preview October 13, 2026 00:02 Inactive

cubic-dev-ai Bot reviewed Oct 13, 2026

View reviewed changes

mzxchandra requested a review from waleedlatif1 October 13, 2026 00:28

waleedlatif1 added 2 commits October 12, 2026 19:13

Merge branch 'staging' into feat/kb-image-documents

954a8bd

vercel Bot deployed to Preview October 13, 2026 02:19 View deployment

fix(knowledge): keep op narrowing in processDocOps and clean up Proje…

89723a2

…ct fixtures

vercel Bot temporarily deployed to Preview October 13, 2026 02:27 Inactive

cubic-dev-ai Bot reviewed Oct 13, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/documents/vision-checkpoints.ts

Comment thread apps/sim/connectors/confluence/meta.ts

Comment thread apps/sim/lib/knowledge/documents/image-extraction.ts Outdated

greptile-apps Bot reviewed Oct 13, 2026

View reviewed changes

Comment thread apps/sim/lib/knowledge/connectors/member-sync-engine.ts

Comment thread apps/sim/lib/knowledge/documents/image-extraction.ts Outdated

waleedlatif1 added 2 commits October 12, 2026 19:54

Merge branch 'staging' into feat/kb-image-documents

8ecd86a

vercel Bot deployed to Preview October 13, 2026 02:59 View deployment

cubic-dev-ai Bot reviewed Oct 13, 2026

View reviewed changes

waleedlatif1 merged commit da553ac into staging Oct 13, 2026
40 checks passed

waleedlatif1 deleted the feat/kb-image-documents branch October 13, 2026 03:03

TheodoreSpeaks mentioned this pull request Oct 13, 2026

v0.9.20: project membership enforcement, db and perf improvements, realtime and auth hardening, CI and test fixes #8963

Open

waleedlatif1 mentioned this pull request Oct 13, 2026

fix(knowledge): key vision checkpoints by title and keep excluded images off while disabled #8965

Merged

6 tasks done

Conversation

mzxchandra commented Oct 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Product update (if relevant)

Type of Change

Testing

Checklist

Screenshots/Videos

Uh oh!

vercel Bot commented Oct 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

greptile-apps Bot commented Oct 12, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Confidence Score: 5/5

Summary

Diagram

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mzxchandra commented Oct 12, 2026

Uh oh!

mzxchandra commented Oct 12, 2026

Uh oh!

cubic-dev-ai Bot commented Oct 12, 2026

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

mzxchandra commented Oct 12, 2026

Uh oh!

mzxchandra commented Oct 12, 2026

Uh oh!

cubic-dev-ai Bot commented Oct 12, 2026

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

mzxchandra commented Oct 12, 2026

Uh oh!

mzxchandra commented Oct 12, 2026

Uh oh!

cubic-dev-ai Bot commented Oct 12, 2026

Uh oh!

cubic-dev-ai Bot left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

mzxchandra commented Oct 13, 2026

Uh oh!

mzxchandra commented Oct 13, 2026

Uh oh!

mzxchandra commented Oct 12, 2026 •

edited

Loading

vercel Bot commented Oct 12, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

greptile-apps Bot commented Oct 12, 2026 •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading

cubic-dev-ai Bot left a comment •

edited

Loading