[richiemcilroy]

Shortens the Instant upload tail after Stop and removes almost every presign request while recording.

sequenceDiagram
    participant App as Desktop
    participant API as Cap API
    participant S3
    Note over App: before, at Stop
    App->>API: presign thumbnail
    App->>S3: PUT thumbnail
    App->>API: presign manifest
    App->>S3: PUT final manifest
    App->>API: recording-complete
    Note over App: after, at Stop
    par
        App->>API: presign thumbnail (while re-encoding it)
        App->>S3: PUT thumbnail
    and
        App->>S3: PUT final manifest (reused URL)
    end
    App->>API: recording-complete

Loading

• Manifest URL reuse (apps/desktop-gpui/src/upload.rs): the first batch presign already returns a URL for segments/manifest.json, which was never used. Progress and final manifests now PUT to it instead of calling /api/upload/signed before every PUT (about one presign every two seconds).
  • Kept for at most 10 minutes. The server signs these URLs for 30 minutes (createUploadTarget, expiresIn: 1800).
  • Google Drive resumable sessions are single use, so they are never reused (ReusableUploadTarget::new).
  • If the reused URL is rejected with an HTTP error (for example an expired signature), the same call signs a fresh URL. Timeouts and network errors go through the existing manifest retry backoff, so a hung network still costs one request timeout per attempt.
• Thumbnail alongside the final manifest (InstantUpload::finish_with_screenshot): the final manifest is authorized first, then the thumbnail upload and the segment finish run together. recording-complete still waits for both, and a failed thumbnail still fails the attempt, which the queue retries exactly as before.
• Thumbnail work: the JPEG re-encode runs on the blocking pool (it used to block a runtime worker) and overlaps its presign request. Studio share thumbnails get the same change.

Stop tail benchmark

upload::queue::stop_tail_bench drives a live Instant recording through the real upload queue against MockCap, with every API call and object PUT answered after a fixed delay on the virtual clock. Times are from Stop. Baseline is main.

API / object latency	Final manifest (main → PR)	recording-complete (main → PR)
120 ms / 40 ms	0.36 s → 0.12 s	0.48 s → 0.28 s
200 ms / 80 ms	0.64 s → 0.24 s	0.84 s → 0.48 s
400 ms / 150 ms	1.25 s → 0.45 s	1.65 s → 0.95 s

Recording	Requests (main → PR)	Presign calls (main → PR)
60 s	132 → 101	35 → 4
30 min	3,656 → 2,758	949 → 51

STOP_TAIL_API_MS=200 STOP_TAIL_OBJECT_MS=80 STOP_TAIL_PAIRS=30 \
  cargo test --bin cap-gpui stop_tail -- --ignored --nocapture

The Instant request budgets in upload/verify_tests.rs drop by one presign per manifest, so a return to per-manifest presigning fails them.

Confidence Score: 5/5

The PR appears safe to merge; no blocking issues were found.

Summary

Speeds up Instant uploads after Stop.

• Instant uploads reuse a signed URL for repeated manifest updates.
• The final manifest can upload while the recording thumbnail is being sent.
• Thumbnail compression and URL signing now run at the same time.

The latest change sends network errors through the existing retry backoff instead of immediately signing another URL. No new actionable issues were found.

_{Reviews (3) · Last reviewed commit: "fix(desktop-gpui): back off before re-si..." · Reviewed by Greptile}

[vercel]

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

Project	Deployment	Actions	Updated
cap-status	Ready	Preview	Oct 27, 2026 2:04am UTC

[richiemcilroy]

hey @greptileai, please re-review the PR

[richiemcilroy]

hey @greptileai, please re-review the PR

[capy-ai]

Capy found no issues.

View 2 other findings in Capy.