按 Agent 单独配置 compaction，避免“一刀切”上下文策略

解决“研究型 agent 与执行型 agent 共享同一压缩策略导致上下文利用率失衡”场景：PR #14598 支持 `agents.list[].compaction` 覆盖全局默认。

GITHUBDiscovered 2026-02-12Author curtismercier

Prerequisites

You use multi-agent config (`agents.list`) with distinct workloads (research, executor, chat, etc.).
Runtime includes the per-agent compaction patch and allows config reload/restart.

Steps

Set conservative defaults in `agents.defaults.compaction` first, then override only outlier agents in `agents.list[].compaction`.
For long-context agents, raise `reserveTokensFloor` and optionally set higher `maxHistoryShare`; for short tasks, lower it.
If only overriding `memoryFlush.enabled`, keep an eye on inherited prompt/systemPrompt behavior to avoid accidental wipeout.
Run regression tests across all agents and compare compaction frequency, token use, and response quality.

Commands

openclaw gateway status

openclaw gateway restart

npm run build

# config snippet:
agents:
  defaults:
    compaction:
      mode: default
      reserveTokensFloor: 20000
  list:
    - id: researcher
      compaction:
        reserveTokensFloor: 40000
        maxHistoryShare: 0.8
    - id: executor
      compaction:
        maxHistoryShare: 0.4

Verify

Different agents exhibit expected compaction behavior (e.g., researcher retains longer context, executor compacts earlier) with no startup schema errors.

Caveats

PR discussion includes an edge-case crash warning around partial `memoryFlush` merges; validate in staging before production（需验证）.
Aggressive compaction tuning can hurt answer coherence even if token metrics improve.

Source attribution

This tip is aggregated from community/public sources and preserved with attribution.

Open original source ↗

Visit original post