Summary

OpenClaw exposes agents.defaults.compaction.model in the runtime schema/config surface, but in the installed runtime build I tested, I could not find any active code path that consumes it for session compaction.

Observed behavior

• Config includes:
  • agents.defaults.compaction.model
• Runtime schema/help includes:
  • agents.defaults.compaction.model
• But runtime search only found references in:
  • runtime-schema-B34T_6nr.js
• I could not find any execution path that reads the configured value and uses it to select a compaction model.

Evidence

1. Searching the installed runtime under:
   • /usr/local/lib/node_modules/openclaw/dist/**/*.js
2. Only hit for compaction.model was schema metadata:
   • runtime-schema-B34T_6nr.js
3. Budget-triggered compaction path appears to pass the session's active model instead:
   • compactEmbeddedPiSession({ ... model: params.followupRun.run.model, trigger: "budget", ... })
4. Memory-flush path is separate and also does not appear to read agents.defaults.compaction.model

Expected behavior

One of these should be true:

1. agents.defaults.compaction.model is actually honored for compaction runs, or
2. the config key is removed from schema/docs if it is no longer supported

Why this matters

This is misleading for operators. A config like:

"compaction": {
  "mode": "safeguard",
  "model": "openai-codex/gpt-5.3-codex-spark"
}

strongly implies compaction will use that model, but runtime behavior suggests it does not.

Impact

• makes compaction behavior hard to reason about
• can lead to false assumptions during context-budget tuning
• especially confusing when debugging large-session compaction/model limits

Suggested fix

• wire agents.defaults.compaction.model into the actual compaction executor, or
• remove/deprecate the config key and update schema/docs accordingly