OpenClaw Context Window Exceeded
Fix context length, token limit, long-context eligibility, and local backend failures without hiding the problem behind endless retries.
A turn reports context length exceeded, too many tokens, long-context access required, or a local backend crashes only on full agent prompts.
The configured window, provider entitlement, and backend’s real capacity must agree. Advertising a larger window in configuration does not give a local server more memory or a provider account long-context access.
Inspect the route and configuration
openclaw models status
openclaw config get agents.defaults.models
openclaw logs --followRecord the exact model, provider-level context defaults, per-model override, and the failing session size. A stale override can keep requesting a long-context route after the provider or model changed.
Recover safely
Use a fresh session to confirm the model works with a small prompt. Reduce the configured context to a supported standard window, compact or reset the oversized session, and remove obsolete long-context flags. For local models, lower context until the server runs without swapping or being killed.
An Anthropic 429 that says extra usage is required for long context is an entitlement and route issue, not ordinary Gateway authentication. Wait or select a standard context instead of repeatedly re-authenticating.
Expected result
- A fresh small session works.
- The configured context fits the provider or local backend.
- The normal agent prompt completes without server termination.
- Long-context requests occur only on an eligible model and account.
If the request remains under the real limit but still stalls, use Model timeout.
Source notes
Technical facts on this page were checked against primary sources.