openclaw - 💡(How to fix) Fix [Feature]: Context pruning for non-caching providers (Xiaomi MiMo and others) [1 participants]

Summary

Enable context pruning for providers not in the cache-TTL eligible list, so token consumption doesn't grow unbounded for providers like Xiaomi MiMo.

Problem to solve

Context pruning (cache-ttl mode) is gated behind isCacheTtlEligibleProvider(), which only includes anthropic, moonshot, and zai. For Xiaomi MiMo (and any other unlisted provider), pruning is completely disabled — the context grows unbounded with every interaction, tool call, and web search. This results in massive token consumption: simple messages cost 500k+ tokens after a few tool calls because the full history is sent every time.

Proposed solution

Decouple context pruning from cache-TTL eligibility. Pruning should run for all providers — for non-caching providers it still reduces context size and token consumption, even without the caching benefit. Alternatively, add xiaomi to the eligible providers list as a first step.

Alternatives considered

No response

Impact

Affected: All users using Xiaomi MiMo (Token Plan) Severity: High — makes the provider practically unusable for long sessions or tool-heavy tasks Frequency: Always — every session grows unbounded Consequence: 500k+ tokens consumed per simple message after a few tool calls, costing significant credits

Evidence/examples

OpenClaw 2026.3.13, provider xiaomi (token-plan-ams.xiaomimimo.com), model mimo-v2-pro. Root cause found in source: buildContextPruningFactory() returns early at isCacheTtlEligibleProvider() check. Xiaomi is not in CACHE_TTL_NATIVE_PROVIDERS = ["anthropic", "moonshot", "zai"].

Additional information

No response