Problem

Four compaction parameters use absolute token counts, not relative percentages of the context window:

Meanwhile, contextThreshold (0.75) and maxHistoryShare (0.8) correctly use ratios that scale with the context window.

Impact

When switching a session from a 1M-token model (e.g. deepseek-v4-pro) to a 200K-token model (e.g. zai/glm-5.1), softThresholdTokens=400000 exceeds the entire 200K context window. This causes immediate memory flush on every turn, preventing the context from ever building up — effectively rendering LCM unusable.

Every model change with a different context window requires manually editing the JSON config and restarting the gateway. This is error-prone and creates a silent footgun.

Proposed Fix

Allow percentage strings for these parameters (similar to how contextThreshold works):

{
  "softThresholdTokens": "40%",
  "reserveTokens": "4%",
  "keepRecentTokens": "2%",
  "reserveTokensFloor": "3%"
}

Resolved at runtime against the active model's context window. Backward-compatible: integer values keep the current behavior.

Workaround (current)

Manual JSON edit + gateway restart per model switch. Not sustainable.

Environment

• openclaw 2026.5.20 (e510042)
• macOS, Node 24