Usage, retries and recovery

Owners and Admins must configure connection limits before use. Editors spend through those shared limits. Daily counters reset at 00:00 UTC; reservations belong to the day the call was claimed, including a response received after midnight.

What the counters mean

• Estimated upper bound reserves conservative request/output capacity. It is not a price quotation.
• Reported usage comes from a provider usage field when available.
• Reserved / unreported includes in-flight requests and consumption that cannot be established. It continues to count against local limits.
• Outcome unknown is a call whose result may have been lost. Its reservation is retained, even after a worker restart.

Google Translate/DeepL limits use characters; OpenAI/Gemini use tokens. Native placeholder protection and model instructions contribute to the request budget. The model output cap bounds response generation. Localisync cannot know the provider’s complete balance, bill or requests made elsewhere.

A preview can exceed the configured per-job limit; Generate rejects it. Narrow the selection or ask an administrator to update limits and review a new preview. Limits and permits are checked atomically across worker instances.

Retry deliberately

A provider request gets a persistent attempt record before it is sent. A timeout, lost response or worker crash cannot safely prove that no charge occurred. Localisync marks the attempt Outcome unknown and makes no automatic paid retry.

At most two delayed retries are made for a definite temporary rejection that the adapter considers safe to retry. Permanent authentication, quota, model, language and validation failures do not loop indefinitely. Retrying an uncertain or failed paid request requires a new explicit approval and may incur another charge.

Changing source content requires a new preview. Replacing a key or changing a connection stops queued work until settings are reviewed again. A disabled/removed connection cannot make future calls. Cancelling a job cannot refund an already sent request.

Technical limits and retention

Boundary	Limit
Job scope	10,000 keys / 50,000 target cells / 100 MiB temporary content
Provider batch	At most 20 cells, one target language; lower provider byte limits apply
Provider timeout	60 seconds
Concurrent calls	2 total, 1 per workspace and 1 per connection
Lists	25 default, 100 maximum
Preview and proposals	24 hours; expired jobs cannot generate or Apply
Content-free job summary	30 days

The content budget includes retained suggestions as well as source previews. A response that would exceed the budget is not stored; reported or uncertain provider use still counts. Use a smaller scope rather than retrying the same full job. Inline edits also respect the budget.

Applied translation history and audit events follow their existing retention. Preview cleanup does not remove applied translations. Temporary cleanup does not race an active attempt. No partial Apply is made when a limit or validation check fails.

If the vault, queue or database is unavailable, follow the operator guide. Provider response bodies, source/proposal text and API keys are excluded from logs and traces; share only the safe request/job reference when asking an operator for help.