QUOTAS & ACCOUNT AVAILABILITY · ANTIGRAVITY TOOLS
Antigravity Manager: Fix 429, 503 & Pool Errors
Pause repeated requests and read the complete error message. For a quota limit, check the model’s reset time. For an unavailable pool, inspect account status. For a capacity error, wait or try another model your account can use.
For version 5.0.0; beta.6 model gate covered separately · Updated
Why do I get 429 when quota looks full?
Check the reason in the failed response, then refresh quota for the account and model that served the request. A remaining quota percentage can coexist with a short-term request limit or a model-capacity shortage.
| Message or reason | Next action |
|---|---|
QUOTA_, quota reset or daily allowance | Check the affected model’s quota window and reset time. Wait for that reset, or choose an account or model with available quota. |
RATE_, requests/ | Reduce simultaneous jobs and pause duplicate client retries. Wait for the supplied delay, then try one short request. |
MODEL_ | Follow the capacity checks, even if the HTTP status is 429. |
RESOURCE_ with a vague message | Look for details and a retry delay in the same response. Begin with a pause and one small request; inspect the quota window and logs if it persists. |
In Accounts, refresh the affected account and inspect the exact model’s 5-Hour Rolling Quota and 7-Day Weekly Quota views. Check the last update and reset time. The Dashboard can combine several accounts; use the quota-reading guide to interpret the account-level figures.
If a client alias maps to another model, use the resolved model when checking quota. Find it in the request log or the X-Mapped-Model response header when present. The model-routing guide explains aliases and quota groups.
After the relevant delay, retry once with concurrent jobs paused. If that works, resume gradually. If a long conversation fails while a short request succeeds on the same model, inspect the conversation’s size, tools and reasoning settings using the request troubleshooting guide.
503 “No capacity available”
Wait for the supplied retry delay, or try another supported model if it can handle your task. This message indicates that the upstream service could not provide capacity for the requested model.
Run the comparison with one request. If the alternative works, use it while the original model is busy. If several accounts receive the same capacity response for the same model, pause that workload and try again later; adding more simultaneous attempts increases request pressure.
For a 503 with a different body, follow the named cause. Proxy service is currently disabled means you should start API Proxy. Token pool is empty needs the account checks below. Verification, eligibility or account-suspension messages belong in the account-access guide. For a timeout or refused connection, start with connection troubleshooting.
“Token pool is empty” or “No available accounts”
Open Accounts in the Manager instance serving your client and inspect each account’s status. The proxy needs an enabled account that can serve the requested model.
If your accounts are missing, check the active data directory, especially after changing a Docker volume or launching another Manager instance. Keep the existing account files while locating the correct profile.
| What you see | What to check |
|---|---|
| Disabled | Read the recorded reason. For invalid_grant, follow the reauthorization steps to replace the invalid credential. |
| Proxy Disabled | Enable proxy use for that account if you intend to include it in API requests. Preserve deliberate exclusions. |
| Protected | Compare the model’s quota with your Quota Protection reserve. Wait for quota to recover, select another eligible account, or deliberately revise that reserve. |
| Verification or eligibility message | Complete the account-specific action in the authorization guide, then refresh its state. |
| In QoS Cooldown or Probing Recovery | Check the QoS status and recovery process before sending more traffic. |
For No accounts available with quota for model, also check which models the accounts expose. In 5.0.0, this message can appear when the gateway has no matching model metadata, including before quota data has loaded. A response requiring a PRO or ULTRA subscription directs you to the account’s reported plan and model access. Refresh an enabled account’s quota and confirm the requested model appears; then review any alias or mapping.
After correcting the specific state, check that the running proxy has loaded the account and send one short request. If the pool still reports every account failed or unhealthy, find the first credential, model or upstream error that led to that state.
“All accounts limited” or “Cooldown wait Ns”
Pause the clients using this gateway until the stated wait has elapsed, then retry once. These messages mean the scheduler is waiting for a usable account to leave cooldown.
Depending on the build and response format, you may see All accounts limited. Wait 29s. or All accounts rate-limited. Cooldown wait 29s.. Keep the full message and Retry-After. If the response also gives a longer quota-reset delay, wait for the longer interval.
For a single-account gateway, waiting for that account is a valid recovery path. With several accounts, check why each was excluded if the whole pool is unavailable. If a new cooldown starts immediately, stop the repeated workload and capture its first upstream error and account-selection sequence.
What do the QoS badges mean?
QoS tracks response quality and can pause an account even while quota remains. Read the badge on the affected account:
- Suspected Throttled: abnormal response latency was detected; new session allocation is paused.
- In QoS Cooldown: the account is temporarily isolated and its session bindings have been released.
- Probing Recovery: Manager is testing whether the account can serve traffic again.
- On Probation: a recovery probe passed; Manager is checking limited traffic before full use.
Let recovery run, or use another eligible account. The QoS guide explains these controls alongside quota and account selection.
Should I use Release or Clear Rate Limit Records?
Use these controls only for a deliberate retest after you have checked the cause and paused other callers. Release resets an account’s local QoS state and clears its tracked rate-limit records. Clear Rate Limit Records clears local rate-limit records across the pool, including persisted live-limit entries.
The upstream service still evaluates the next request against its current quota and capacity. Preserve the original error before clearing anything, wait for any known reset, and send one request afterward. If the same error returns, keep the new response and return to its diagnostic branch.
502 “no quota accounts” in 5.0.1-beta.6
Check the resolved model and its account quotas. In 5.0.1-beta.6, Manager’s model gate returns this response before sending the request upstream when the model is recognized but no account has positive quota for it.
- Confirm the model ID. Use the model list exposed by your gateway. Check the client’s alias and the final target in model routing.
- Refresh account quota data. Inspect that model across the accounts you intend to use. Wait for its reset or select another model or account with available quota.
- Read any 404 separately.
model not existdirects you to the ID or model-list check. The gate also rejects retired model families; use a currently offered ID.
Immediately after startup, an empty model-data set makes this gate skip its existence and positive-quota checks. Missing IDs and retired model families are still rejected. Refresh the account data before interpreting an empty list as exhausted quota.
If this local 502 is followed by unexpected account cooldowns, record the exact build, model and before/after account states for a diagnostic report. Use the release guide to compare versions before changing builds.
Why does Manager keep choosing one account?
Check API Proxy → Account Rotation & Scheduling, then confirm that another account is eligible for the same model. Account preference, session bindings and scheduling policy all affect the choice.
- Review Fixed Account Mode. Clear an unintended preference for the comparison. In 5.0.0, the preferred account is checked for availability; some exclusion and zero-quota conditions allow fallback to another candidate.
- Check Scheduling Mode. Cache First and Balance reuse a healthy session’s account. In 5.0.0, a rate-limited binding is released immediately so another account can be selected, including in Cache First. The interface also shows Max Wait (sec); use the response’s actual cooldown time when planning your retry. Performance removes that session-binding preference; eligible-account ordering can still lead to the same account being selected.
- Review Spillover Account Cap. A configured cap narrows the candidates for new assignments; existing session bindings and fixed-account preference are handled separately. Compare it with your intended workload and available alternatives.
- Try a fresh short session. Keep the model and protocol the same and inspect the selected account in the request log. Save the old settings before changing them and restore temporary changes afterward.
When should I clear session bindings?
Use Clear Session Bindings when you intentionally want existing sessions to be assigned again on their next request. This affects all current bindings and can reduce prompt-cache reuse. Start with a fresh test session when you only need a comparison.
Follow the account-rotation guide for normal scheduling setup. If selection changes but every attempt receives the same capacity message, return to the capacity checks.
Confirm recovery with one short request
After the indicated wait or account repair, run the minimal text-request test from the same environment as your client. Keep the model and protocol fixed for the first comparison and pause other jobs using the gateway.
- A complete answer arrives: check the resolved model and selected account, then gradually resume the original workload.
- The short request works, the original task fails: compare concurrency, conversation size, tools and thinking fields. Follow the request and streaming checks.
- The same failure returns: record the first error after the wait. If you try a second model, save both results so the failing model remains clear.
Quota refresh updates the displayed account data. Use the completed model answer to decide when to resume the workload.
Report a limit or pool failure that persists
Include the first upstream error and the result of one request after the wait. These show what triggered the cooldown and whether the same cause remains.
Add the Manager build and client version, protocol, requested and resolved model, time and time zone, status and error body, retry delay, pool count, exclusion reasons and account-selection sequence. Replace account identities with consistent labels such as Account A and Account B.
Remove API keys, OAuth tokens, emails and private prompts, including identifiers in response headers. Keep account JSON and raw traffic captures private.