Add AGENTS operational notes for Spaces secrets and GPU abort debugging

AGENTS.md

@@ -0,0 +1,37 @@
+# Operational Notes (DeepSeek-OCR-Demo)
+
+## Hugging Face token handling
+- `HF_TOKEN` is **not** present by default in shell env.
+- Local HF CLI token exists at `~/.cache/huggingface/token` (used as local fallback).
+- Repo now includes:
+  - `.gitignore` (ignores `.env` and local caches)
+  - `.env.example` (`HF_TOKEN=...` placeholder only)
+  - `scripts/fetch_space_logs.sh` (Spaces log fetch helper)
+
+## Spaces-compatible secret pattern
+- Preferred: set `HF_TOKEN` in **Space Settings -> Variables and secrets**.
+- Local dev: `cp .env.example .env`, set `HF_TOKEN`, then:
+  - `set -a; . ./.env; set +a`
+
+## Space log access
+- Run logs:
+  - `./scripts/fetch_space_logs.sh ricklon/DeepSeek-OCR-2-Math run`
+- Build logs:
+  - `./scripts/fetch_space_logs.sh ricklon/DeepSeek-OCR-2-Math build`
+
+## Current findings from logs
+- Build logs looked healthy for commit `d6afca6` (no build failure seen).
+- Run logs show normal startup and large model download/initialization.
+- Observed warning:
+  - Flash Attention warning about model init on CPU before GPU move.
+  - This warning was already expected in this app flow and not by itself a crash.
+- Observed `GPU task aborted` in UI, but abort stack line was not yet captured in the streamed snippet.
+
+## Likely runtime risk (not yet conclusively proven)
+- `@spaces.GPU(duration=90)` may be too short for heavy pages now that equation refinement can trigger multiple extra inferences.
+- If abort repeats under load, first mitigation to try is increasing GPU duration for `process_image`/`process_pdf`.
+
+## Region OCR feature status
+- Region OCR UI was added with `gr.ImageEditor` behind a compatibility guard:
+  - `HAS_IMAGE_EDITOR = hasattr(gr, "ImageEditor")`
+- Space config is `sdk_version: 6.8.0`, which should support `ImageEditor`.