Description

Fresh OpenClaw installations receive compaction.mode: "safeguard" by default (via applyCompactionDefaults in src/config/defaults.ts), but this mode silently fails when contexts reach ~180k tokens, producing "Summary unavailable due to context limits" instead of actual AI-generated summaries.

This causes conversations to lose context without any warning to the user.

Evidence

1. Code default is safeguard:

// src/config/defaults.ts, line ~460
export function applyCompactionDefaults(cfg: OpenClawConfig): OpenClawConfig {
  // ...
  return {
    ...cfg,
    agents: {
      ...cfg.agents,
      defaults: {
        ...defaults,
        compaction: {
          ...compaction,
          mode: "safeguard",  // ← This is applied to all new installations
        },
      },
    },
  };
}

2. Documentation says otherwise:

The docs at docs.clawdbot.com state:

agents.defaults.compaction.mode selects the compaction summarization strategy. Defaults to default; set safeguard to enable chunked summarization for very long histories.

3. Session logs show failures:

{
  "type": "compaction",
  "summary": "Summary unavailable due to context limits. Older messages were truncated.",
  "deletedCount": 847
}

This appears repeatedly in session JSONL files when using the default config.

4. Wizard doesn't configure this:

The setup wizard (src/commands/configure.wizard.ts) has no section for compaction settings. Users cannot opt out of safeguard mode without manually editing their config.

Impact

• All new users get broken compaction by default
• Silent failure — no warning that summarization failed
• Context loss — AI loses conversation history without proper summary
• Users may not realize the issue for weeks/months

Expected Behavior

Either:

1. Change the default from "safeguard" to "default" (which works reliably)
2. Fix safeguard mode to actually produce summaries on large contexts
3. Add a warning when safeguard mode can't summarize
4. Add compaction settings to the setup wizard
5. Update documentation to match actual code behavior

Workaround

Users can manually set in ~/.openclaw/openclaw.json:

{
  "agents": {
    "defaults": {
      "compaction": {
        "mode": "default",
        "reserveTokensFloor": 40000
      }
    }
  }
}

Increasing reserveTokensFloor from the default 20,000 to 40,000 triggers compaction earlier, giving more headroom for the summarization API call.

Environment

• OpenClaw version: 2026.1.29 (latest)
• Model: Claude Opus 4.5 (200k context)
• OS: Linux

Roadmap Alignment

This issue aligns with the current priority listed in CONTRIBUTING.md:

Performance: Optimizing token usage and compaction logic.

Happy to help with a PR if maintainers point me in the right direction.

Related

• src/config/defaults.ts — applyCompactionDefaults()
• src/agents/pi-settings.ts — DEFAULT_PI_COMPACTION_RESERVE_TOKENS_FLOOR
• src/commands/configure.wizard.ts — no compaction section
• Documentation at docs.clawdbot.com/gateway/configuration

I'm writing a blog post about OpenClaw memory optimization and discovered this while investigating why my bot kept "forgetting" context. Happy to help test any fix.

I can take a crack at a PR if you'd like — the fix seems straightforward (change default in applyCompactionDefaults or update docs). Let me know which direction you prefer.