Inspiration
Extraction-shooter streams often make combat skill easy to see, but another part of the experience is harder for viewers to follow: how the streamer’s assets rise or fall after each run. I wanted to make that progression visible without reading game memory or intercepting network traffic.
What it does
DeltaForceHUD is a Windows-first, local OCR HUD for Delta Force streamers. It captures the user’s own screen through OBS Virtual Camera, uses OpenCV and Tesseract Japanese OCR to read the on-screen asset value, and displays live gain/loss through an OBS browser-source overlay. It also provides a browser setup wizard, lobby-state detection, session history, a trend graph, CSV export, cut-based segment tracking, and camera-freeze recovery.
Demo flow
- Start the local FastAPI server and open the browser setup wizard.
- Select a capture device and draw the OCR region.
- Confirm that the lobby value is read from the game screen.
- Show the live OBS gain/loss overlay.
- Open session history, press “Cut Here,” change the balance, and show the net change from that marker.
What was completed during Build Week
DeltaForceHUD existed before the event. The OCR pipeline, setup wizard, OBS overlay, history view, and the cut UI predated Build Week. At the official event baseline, however, the cut UI called backend endpoints that were missing from main.py, so the workflow was incomplete.
During Build Week, Codex running on GPT-5.6 inspected the frontend call sites and the current backend, restored and completed POST /cut and GET /current_segment, added self-contained marker persistence in markers.jsonl, and aligned failure responses with the frontend’s res.ok behavior. The implementation is commit 85ef29b; judge documentation was added in a later event-period documentation commit.
Codex also built a supporting Windows video-production harness and used a dedicated OBS/FFmpeg workflow to validate and export the demo. The harness is documented separately; the final demo was rendered through the dedicated workflow rather than a canonical harness project.
How Codex and GPT-5.6 were used
This was a human-directed, multi-model development workflow. Claude assisted with structured specifications and review. Codex on GPT-5.6 performed the repository inspection, implementation, diff production, syntax check, and endpoint verification. A follow-up review challenged the first assumption about the frontend error contract: the UI branches on the HTTP status through res.ok, not only on the JSON body. Codex then corrected the failure paths and verified:
- missing balance → HTTP 400 with
ok: false - successful marker save → HTTP 200 with
ok: true - forced marker-write failure → HTTP 500 with
ok: false - browser error alert → displayed on the existing
/logUI
Codex handled repository diagnosis, implementation, endpoint verification, OBS recording control, footage validation, synchronization, and export.
The human developer handled the product goal, in-game actions, footage judgment, narration alignment, final approval, and publication.
Primary Codex Session ID: 019f74dc-5f11-79e3-86e1-a290fa97cc42
Impact
DeltaForceHUD gives streamers a visible progression story that does not depend on high-level combat highlights. Viewers can understand whether a run or a segment was profitable, while the tool remains local and does not access game memory or modify network traffic.
Challenges
- Combat scenes produced OCR noise, so lobby-state validation had to use meaningful on-screen words rather than a raw character-count threshold.
- OBS Virtual Camera could freeze, requiring frame-change detection and reconnect handling.
- The pre-existing cut UI and current backend had drifted apart, so the event-period work required tracing the actual call contract instead of assuming that the visible feature was complete.
- Windows, Tesseract, Japanese OCR data, capture resolution, and crop configuration all affect reproducibility.
What we learned
Repository history and visible UI are not enough to prove a feature works. The decisive step was tracing the frontend call sites to the event-baseline backend, testing both success and failure paths, and preserving evidence that maps the Codex session to the resulting commit.
I started using AI tools less than three months ago and do not work as a full-time software engineer. With a human-directed Codex workflow, I was able to inspect an existing codebase, complete and verify a missing backend path, diagnose the OBS capture pipeline, and produce the final demo within the Build Week window.
What’s next
- A standalone Windows package with Tesseract included
- Additional game and screen-layout profiles
- More viewer-facing overlay options
Disclaimer
DeltaForceHUD is an unofficial fan-made tool. It is not affiliated with or endorsed by Team Jade, TiMi Studio Group, Level Infinite, OBS Project, OpenAI, or Anthropic. Product names and trademarks belong to their respective owners.
Log in or sign up for Devpost to join the conversation.