translation agent loop detected retrying failed crowdin upload

Fix the translation agent loop retrying a failed Crowdin upload

TL;DR

The agent retries forever because it never distinguishes retryable from fatal errors, so cap the retries and classify the error. A 401 retried 200 times is a loop, not resilience. Retry only transient errors, at most a few times, then stop and report.

The error

Translation agent log: retrying Crowdin upload, attempt 147
the API returns 401 every time, the agent never stops

Fix it

Step 1: Stop the runaway agent

pkill -f translate-agent || true

Expected: The loop stops.

Step 2: Check what the API actually returns

curl -s -o /dev/null -w "%{http_code}" -H "your auth header https://api.crowdin.com/api/v2/user

Expected: You see the real status, likely 401.

Step 3: Add a retry cap and error classification

node -e "console.log('max 3 retries, only on 429/5xx/network errors, everything else fails immediately')"

Expected: Fatal errors stop the loop on the first try.

Step 4: Fix the underlying 401, then rerun

node -e "console.log('refresh the credential, then run the agent -- it should fail fast or succeed, never loop')"

Expected: The agent completes or errors cleanly.

When to use this

• A translation agent loops retrying a failed upload
• Retry counts climb into the hundreds

When NOT to use this

• Retries are bounded and eventually succeed, that is working as designed
• The loop is in extraction, different fix

Tool and version compatibility

• Any agent with upload retry logic
• Crowdin API v2 error codes

Variant phrasings

loop on 429 without honoring retry-after

Bounded retries plus honoring retry-after. The cap alone still hammers the API.

loop across restarts, not within one run

The checkpoint says incomplete so it restarts forever. Fix the root error, then clear the poisoned state.

Why it happens

Unbounded retry with no error classification treats a permanent 401 like a transient blip. Each attempt fails identically, the agent tries again, and the loop runs until someone kills it, burning API quota the whole time.

Edge cases

• Alert on final failure, a capped retry that always fails is still a bug
• Exponential backoff without a cap is still unbounded, cap the count too
• Log the error body on final failure, the classification depends on it

Provenance

Resolved from the public thread: https://vectle.com/posts/pst_V0YLHBDkvs64oqDMxWFDEA