request quota exhausted: quota, wait and switch models

request quota exhausted says this request was blocked by quota; it does not say which quota. Account balance, upstream rate limits, and plugin concurrency need different recoveries.

Three quota layers

One message, three possible sources.

Account balance is total volume. It needs a top-up or cycle reset; waiting a few minutes usually does nothing.

Upstream rate limits are frequency. Too many requests in a short window; slowing down works better than waiting.

Plugin concurrency is a local cap. Some plugins or gateways limit simultaneous requests; failures cluster at peak times.

Three comparisons

Do not retry blindly; change one variable at a time.

1. Change model: another model failing points to account or global quota; only the original failing points to that channel.
2. Change channel: another provider or plugin instance working means the original is limited.
3. Change time: a minute later working suggests rate limiting; still failing suggests balance exhaustion.

Three comparisons identify the layer. Retrying without changing variables never distinguishes frequency from balance.

Record each result. Model, channel, and time; reproduction needs all three.

Recovery order: light to heavy

Sort by cost and avoid configuration changes first.

1. Wait one short window. Only for rate limits; two failures within a minute mean stop waiting.
2. Reduce frequency and concurrency. Fewer simultaneous requests, longer intervals; team and subagent tasks are most affected.
3. Switch models. Choose one with quota; watch context window differences.
4. Switch provider or channel. Official, third-party gateway, local model; verify configuration and authentication.
5. Top up or wait for reset. The only fix for exhausted balance.

Do not switch providers first. Configuration differences add variables; confirm quota first.

Open a new turn after switching. Old sessions may hold old model references.

When waiting does not help

Three cases never improve with time.

1. Balance exhausted with no reset time. Waiting creates no new quota.
2. Account limited by concurrency or total. Needs configuration or an account change.
3. Plugin rate limit hard-coded. Change the plugin configuration or implementation.

Check whether the error carries a retry time. A time suggests rate limiting; none suggests balance or configuration.

Waiting has a cost. A long task stuck on waits is better served by a prepared backup channel.

Switching models: context caveats

Switching bypasses quota but changes context.

1. Save key conclusions and unfinished items.
2. Check the target model's context window; long sessions may compact.
3. Open a new turn with necessary background rather than continuing the old session.
4. Verify tool calling and structured output compatibility.
5. Record why you switched, for comparison when returning.

Different windows mean different behavior. Moving from large to small compacts earlier content; do not assume identical results.

Capability differences affect plugins. Vision, tools, and JSON output vary; run a minimal function after switching.

Prevention: manage quota as a resource

Quota problems are visible early if you watch.

1. Monitor usage: use balance and usage plugins with thresholds.
2. Cap concurrency: set team and subagent limits.
3. Prepare a backup channel: at least one alternative model or provider.
4. Schedule off-peak: batch tasks away from peaks.
5. Record failure times: find high-risk periods and adjust schedules.

Quota is a resource, not a bug. Treat it as capacity planning rather than emergency response.

Next steps

Layer, compare, switch. For error classification, read This turn failed with transport failed. For key and discovery errors, read API key invalid and model discovery 404. For usage monitoring, read What balance plugins exist.

Sources: DeepSeek Harness docs, dsh CLI README, deepseek-harness repository