Summary

Compaction does not reliably auto-trigger under safeguard mode, and when it does trigger, it fails due to a hardcoded ~5 minute JS timeout.
With slow local models (e.g. Ollama ~30 tok/sec), compaction never successfully completes.
As a result, compaction is effectively unusable in long-running local deployments.

Steps to reproduce

Run OpenClaw with a local Ollama model (~30 tokens/sec).
Use a large context window (e.g. 131k).
Enable compaction:

"compaction": {
  "mode": "safeguard",
  "maxHistoryShare": 0.6,
  "reserveTokensFloor": 24000
}

My openclaw.json

  "agents": {
    "defaults": {
      "timeoutSeconds": 7200,
      "contextTokens": 131072,
      "model": {
        "primary": "ollama/glm-4.7-flash"
      },
      "workspace": "/home/tsingwin/.openclaw/workspace",
      "compaction": {
        "mode": "safeguard",
        "maxHistoryShare":0.6,
        "reserveTokensFloor": 24000
      },
      "maxConcurrent": 4,
      "subagents": {
        "maxConcurrent": 8
      }
    }
  },

Expected behavior

Compaction should auto-trigger before silent trimming removes early history.
Compaction should complete successfully even for slower local models.
If timeout occurs, backend generation should be cancelled.
Compaction timeout should be configurable.
Session state should remain consistent.

Actual behavior

Compaction does not auto-trigger even when session shows:
Context: 852k / 131k (650%)
Compactions: 0
It appears that OpenClaw silently trims history instead of triggering compaction.
When compaction is manually triggered:
After ~5 minutes, JS layer times out.
Frontend/runtime returns.
Compaction is not marked successful.
However, Ollama backend continues generating.
GPU remains busy.
No cancellation is propagated to the backend model.
I have never successfully completed a compaction cycle in a real-world long session using a slow local model.

OpenClaw version

2026.2.23

Operating system

Ubuntu22.04 Arm64(Jetson Orin AGX + Ollama Glm4.7-flash)

Install method

npm global

Logs, screenshots, and evidence

/status
🦞 OpenClaw 2026.2.23 (b817600)
🧠 Model: ollama/glm-4.7-flash · 🔑 api-key ollama…-local (models.json)
🧮 Tokens: 850k in / 1.8k out · 💵 Cost: $0.0000
📚 Context: 852k/131k (650%) · 🧹 Compactions: 0
🧵 Session: agent:main:feishu:direct:ou_8554c0da0b36ebfdfd8d236726224f71 • updated just now
⚙️ Runtime: direct · Think: off
🪢 Queue: collect (depth 0)

Compaction failed: Compaction timed out • Context ?/131k

Impact and severity

No response

Additional information

No response