Skip to content

Feat/aloha screenshot mode#8

Open
zhifeng-d wants to merge 8 commits into
showlab:mainfrom
hyper-instrument:feat/aloha-screenshot-mode
Open

Feat/aloha screenshot mode#8
zhifeng-d wants to merge 8 commits into
showlab:mainfrom
hyper-instrument:feat/aloha-screenshot-mode

Conversation

@zhifeng-d

Copy link
Copy Markdown

No description provided.

FingerLiu and others added 8 commits April 30, 2026 18:35
Occasionally the Windows recorder produces MP4 files that are missing
the moov atom and contain duplicate/corrupted blocks inside mdat.
These files show a black screen or refuse to open.

This commit adds:

- tools/repair_corrupted_mp4.py
  A standalone Python script that:
  1. Parses MP4 atoms to locate the mdat payload.
  2. Auto-detects duplicate/corrupted blocks via MD5 hashing.
  3. Skips corrupted boundaries and finds the first clean IDR frame.
  4. Injects a correct Baseline SPS/PPS for 2560x1440 (and fallback
     presets for 1920x1080 / 3840x2160).
  5. Converts AVCC NAL units to Annex-B and remuxes with ffmpeg.

- tools/REPAIR_GUIDE.md
  Documentation covering root-cause analysis, quick-start usage,
manual segment override, and troubleshooting tips.

Validated on a 4.5 GB, ~60 min, 2560x1440@30fps recording that was
previously unplayable on both Windows and macOS.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Add MP4 repair tool for corrupted screen recordings
Add hyperdata_blender_sphere.ipynb that loads the blender_sphere GUI-
automation dataset (screenshots, crops, action logs, trace captions),
builds Zarr arrays, creates a local HyperData dataset, and pushes it
to the HyperData Hub.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
… updates

- parser.py: add _build_meta_json() and Step 4 to run_pipeline() that
  auto-generates meta.json with task, instruments, recording, pipeline,
  and data statistics.
- hyperdata_blender_sphere.ipynb: load meta.json, push it via
  ds.metadata.set('meta_json', ...), and verify roundtrip.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- ScreenshotExtractor: source per-action frames from inputs/frames/ (written
  by `ace record export`) instead of seeking a video, reusing all crop/scale/
  marker logic from VideoScreenshotExtractor.
- parser: auto-select extractor by inputs (mp4 -> video, frames/ -> screenshot).
- process_actions: detect adjacent actions resolving to byte-identical full
  frames (frozen frames from sparse upstream capture) and warn, so degraded
  data quality is surfaced instead of silently producing mismatched crop/full.
- trace_generator: configurable per-request timeout (API_TIMEOUT_MS) and
  surface vendor error messages on failed LLM HTTP calls.

Co-authored-by: Cursor <cursoragent@cursor.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants