openclaw - 💡(How to fix) Fix lossless-claw contextThreshold does not adapt to actual model context window after fallback

Bug Description

When an agent session falls back from a small-context model (e.g. glm-5.1, 200K) to a large-context model (e.g. deepseek-v4-pro, 1M), the lossless-claw plugin continues to use the original model's contextWindow to calculate the compression threshold, rather than dynamically adapting to the currently active model.

Impact

• Premature compression: With contextThreshold: 0.75, a 200K model triggers LCM compression at 150K tokens
• Context loss: Historical conversation is summarized/lost before the session falls back to a larger model
• User-visible interruption: The combination of premature compression + fallback chain causes conversations to appear "cut off" around 190K tokens
• Cascading failure: After compression on the small model still fails, the fallback chain tries multiple small-window models (all fail) before reaching a large-window model, by which time context is already lost

Reproduction

1. Configure an agent without its own model field, inheriting global default zai/glm-5.1 (contextWindow: 200K)
2. Include deepseek/deepseek-v4-pro (contextWindow: 1M) in the fallback chain
3. Have a long conversation that grows past 150K tokens
4. Observe: LCM compression triggers (based on 200K × 0.75 = 150K threshold), then model fails, falls back through chain to deepseek
5. Result: conversation history is compressed/lost despite now having a 1M context window

Expected Behavior

lossless-claw should dynamically read the currently active model's contextWindow to calculate the threshold, not the originally configured primary model's window. When a session is running on a 1M context model, the threshold should be 750K, not 150K.

Environment

• OpenClaw: 2026.5.20
• lossless-claw config: contextThreshold: 0.75
• Observed on agent tax-invoice-analyst session