Fix: terminal Muapi errors are retried for 30 minutes instead of failing fast

[alnavone]

Fix: terminal Muapi errors are retried for 30 minutes instead of failing fast

A generation that has already failed server-side — out of credits, rate limited, bad request — is indistinguishable from one still running, for up to half an hour. The UI shows a spinner the whole time and never surfaces an error.

Found while debugging a video generation that appeared to hang for 15+ minutes. The job had actually been rejected with a 400 almost immediately.

The bug

pollForResult throws on a fatal HTTP status and on an explicit failed job status. Both throws are inside the try block, so the loop's own catch swallows them and keeps polling:

if (!response.ok) {
    const errText = await response.text();
    if (response.status >= 500) continue;
    throw new Error(`Poll Failed: ${response.status} - ...`);   // thrown inside try…
}
const data = await response.json();
const status = data.status?.toLowerCase();
if (status === 'completed' || ...) return data;
if (status === 'failed' || status === 'error') {
    throw new Error(`Generation failed: ${data.error || 'Unknown error'}`);   // …and this one too
}
} catch (error) {
    if (attempt === maxAttempts) throw error;                    // …caught right here
    console.warn('[Muapi] Poll attempt failed, retrying...', error.message);
}

The if (response.status >= 500) continue guard shows the intent was to treat 5xx as transient and anything else as fatal. The control flow never did that.

Impact scales with maxAttempts:

Caller	maxAttempts	Interval	Time spent retrying a dead job
generateImage	60	2s	2 minutes
generateVideo	900	2s	30 minutes

A 402 (out of credits), 429 (rate limited), 401, or 400 gets retried 900 times. So does a job the server has explicitly reported as failed.

The fix

Tag terminal errors and rethrow them from the catch:

function fatal(message) {
    const error = new Error(message);
    error.isFatal = true;
    return error;
}

} catch (error) {
    if (error.isFatal || attempt === maxAttempts) throw error;
    console.warn('[Muapi] Poll attempt failed, retrying...', error.message);
}

5xx responses and network errors stay retryable — those genuinely are transient.

Applied to all three polling loops, since the same pattern is duplicated:

• src/lib/muapi.js — pollForResult (Electron renderer)
• packages/studio/src/muapi.js — pollForResult (web studio)
• packages/studio/src/muapi.js — pollWorkflowResult

Tests

tests/muapiPolling.test.js, matching the existing node:test suite. It extracts the real pollForResult from src/lib/muapi.js and runs it against a stubbed fetch, so it tests the shipped code rather than a copy of the logic. Extraction is brace-matched, not offset-based, so it survives edits to the surrounding file.

• a 402 stops after exactly one request
• a failed job status stops after exactly one request
• 503 is retried, then succeeds
• pending keeps polling until the job completes
• network errors remain retryable and use every attempt

node --test "tests/**/*.test.js"
# tests 22 | pass 22 | fail 0

17 pre-existing tests still pass.

Note

This is independent of the Windows CUDA / sd.cpp packaging fixes I opened separately — different subsystem, different failure mode. The two branches touch no common files and can be reviewed and merged in either order.