I've been running OpenClaw for a few days now. It's powerful out of the box, but the default configuration left me with a few problems: Feishu group messages silently dropped, cron jobs failing to deliver, and sessions getting blown up by massive tool outputs. Here's what I did to fix it.

The Setup

Three agents, each bound to a different Feishu account:

• 二毛 (main) — my primary assistant
• 三毛 (sanmao) — food planner, industry research
• 四毛 (simao) — DevOps and frontend tasks

All running on MiniMax M2.7 via OAuth, with Ollama as a local fallback.

Fixing Feishu Channel Issues

Group messages silently dropped

The groupPolicy was set to allowlist but no groupAllowFrom was configured. Every group message was being discarded without any error.

Fix: set groupPolicy to "open" at both the top-level and per-account:

"channels": {
  "feishu": {
    "groupPolicy": "open",
    "accounts": {
      "main": { "groupPolicy": "open" },
      "sanmao": { "groupPolicy": "open" },
      "simao": { "groupPolicy": "open" }
    }
  }
}

Cron job delivery failures

Cron jobs require an explicit delivery.to field for Feishu. Without it, you get:

Delivering to Feishu requires target

Fix via CLI:

openclaw cron edit <job-id> \
  --to "user:ou_XXXXXXXXXXXXXXXXXXXXXXXXXX" \
  --channel feishu \
  --announce \
  --best-effort-deliver

Context Management: Preventing Session Blowups

This was my biggest pain point. Web research, large file reads, and recursive commands would dump tens of thousands of tokens into the session, making it unusable.

Layer 1: Context Pruning

Automatically trims old tool outputs based on time and size:

"agents": {
  "defaults": {
    "contextPruning": {
      "mode": "cache-ttl",
      "ttl": "30m",
      "softTrim": {
        "maxChars": 8000,
        "headChars": 2000,
        "tailChars": 1000
      },
      "hardClear": {
        "enabled": true,
        "placeholder": "[内容已截断]"
      }
    }
  }
}

• ttl: "30m" — tool results older than 30 minutes get pruned
• softTrim — any single result over 8000 chars gets truncated, keeping 2000 chars from the head and 1000 from the tail
• hardClear — if pruning still can't bring context under control, old results get replaced with a placeholder

Layer 2: Compaction

When the context window is nearly full, OpenClaw automatically compresses the conversation history:

"compaction": {
  "mode": "safeguard",
  "maxHistoryShare": 0.5,
  "reserveTokens": 8000,
  "keepRecentTokens": 4000
}

• safeguard mode is more aggressive about protecting the session
• History never exceeds 50% of the context window
• 8000 tokens reserved for the model's response
• 4000 most recent tokens kept verbatim

Layer 3: Subagent Delegation

For tasks known to produce massive output, I added rules to each agent's TOOLS.md telling them to spawn subagents instead of executing directly. A subagent runs in its own isolated session — the huge output stays there, and the parent only gets back a summary.

Rules written into workspace/TOOLS.md:

• Must spawn subagent: web research, large file reads (>200 lines), recursive commands, multi-step analysis
• Do it directly: single-line commands, short file reads (<50 lines), simple config changes

This is a prompt-level rule, not a config-level enforcement. But it works well in practice.

Memory Search with Ollama

OpenClaw supports semantic memory search — querying past conversations and notes by meaning rather than keyword matching. It needs an embedding provider.

Why Ollama

The default local option requires node-llama-cpp, which takes forever to install via npm. I already had Ollama running locally, and OpenClaw has native support for it.

Setup

Pull an embedding model:

ollama pull qwen3-embedding:8b

Configure in openclaw.json:

"agents": {
  "defaults": {
    "memorySearch": {
      "enabled": true,
      "provider": "ollama",
      "model": "qwen3-embedding:8b"
    }
  }
}

OpenClaw's Ollama embedding provider automatically reads the base URL from models.providers.ollama.baseUrl (already configured as http://127.0.0.1:11434).

Verify

openclaw memory status --deep
openclaw memory index --force

You should see Embeddings: ready and Vector: ready for each agent.

Quick Reference: Key Config Paths

Lessons Learned

• openclaw doctor --fix catches and fixes a lot of common misconfigurations automatically
• Context pruning is the single most impactful setting for keeping sessions healthy
• Subagent delegation is better than tool restrictions for handling large outputs — isolation beats limitation
• Ollama as a local embedding provider works well if you have the RAM (qwen3-embedding:8b needs ~5GB)
• Cron jobs need explicit delivery.to — the error message is clear but easy to miss