Troubleshooting¶

Common issues and their resolutions for the Curaway platform.

Git Sync Issues¶

"Already up to date" but code is stale¶

Symptom: git pull says up to date, but local code doesn't match what's deployed.

Fix: Use the gsync alias:

git fetch origin && git reset --hard origin/main

Deployment Issues¶

Cloud Run health check timeout¶

Symptom: Cloud Run deploy fails — the new revision never becomes ready and traffic isn't routed to it.

Cause: The startup probe points at /health, which is too slow (11 DB queries + 3 HTTP calls) and exceeds the probe window.

Fix: Point the Cloud Run startup probe at the lightweight /ready endpoint (the container listens on port 8000):

gcloud run services update curaway-backend \
  --project=curaway-dev --region=asia-south1 \
  --startup-probe=httpGet.path=/ready,httpGet.port=8000,timeoutSeconds=10,periodSeconds=10,failureThreshold=3

The canonical probe config lives in service.yaml in the curaway-ai/curaway-gcp-infra repo (applied by the Cloud Deploy pipeline) — prefer fixing it there so a pipeline deploy doesn't overwrite a manual gcloud override.

API Issues¶

Cloudflare proxy returning HTML instead of JSON¶

Symptom: API calls return HTML challenge pages instead of JSON.

Cause: Cloudflare proxy mode (orange cloud) intercepts requests.

Fix: Use DNS-only mode (grey cloud) for all Cloudflare records. See ADR-0008.

File Upload Issues¶

Agent "can't see attachment"¶

Root causes (in order):

OCR not completed: Async OCR queued but not finished when chat runs. Fix: synchronous PyMuPDF inline. See ADR-0010.
has_issues status: Document validator flagged issues but status not counted as analyzed. Fix: count has_issues as analyzed.
Scanned PDF: PyMuPDF returns 0 chars. Fix: falls back to Unstructured.io or Claude Vision.
Toxic history: Old "can't see file" messages poison LLM context. Fix: filter toxic messages.

First message with attachment — agent asks for records it already has¶

Cause: Orchestrator routed to procedure identification before processing attachments. The procedure ID handler always appended "do you have records?" without checking.

Fix (Session 27B): Orchestrator now detects attachments during procedure identification and processes them inline, replacing the records request with document analysis.

EHR showing duplicate conditions (e.g. 53 instead of 7)¶

Cause: Clinical Context Agent ran multiple times on the same document (QStash retries, re-analysis). Each run created duplicate FHIR Condition resources. EHR builder appended them all without dedup.

Fix (Session 27B): Two layers — EHR builder deduplicates by ICD-10 code (or name) at construction time. EHR API endpoint also deduplicates at response time for existing dirty snapshots.

Document matching wrong count¶

Fix: Embedding-based matching (Session 25B). Re-seed: python -m app.seed_embeddings

Frontend Issues¶

React error #310¶

Cause: useState after early return. Move ALL hooks before any conditional returns.

Cause: ReferenceError from variable in wrong scope. Check browser console.

Flagsmith warning flood¶

Cause: Checks for nonexistent flags. Fixed with 60s cache in feature_flags.py.

Database Issues¶

EHR showing 50% with nothing populated¶

Cause: len(comorbidities) >= 0 always true. Fixed to > 0.

Alembic migration conflicts¶

alembic current        # Check state
alembic stamp head     # Force stamp if stuck
alembic upgrade head   # Re-run

Agent Issues¶

Duplicate assistant messages¶

Cause: QStash callback raced sync chat path. Fixed: removed async message insertion.

Quick Diagnostics¶

curl -s https://api.curaway.ai/ready                          # API health
python -c "from app.services.graph_service import get_graph_stats; print(get_graph_stats())"  # Neo4j
pytest tests/ -v                                                    # All tests