Prompt Logger

An OpenClaw plugin that records every complete LLM request (system prompt, conversation history, tool calls, responses, token usage) to disk as structured JSON files.

Built for debugging — never wonder what OpenClaw actually sent to the model again.

How it works

The plugin uses Plugin Hooks (model_call_started, llm_output, model_call_ended, agent_end) to capture everything the Gateway sends to the LLM provider.

Output structure

<workspace>/prompts/
├── 2026-05-09.jsonl          # Lightweight index (search with jq/grep)
├── 2026-05-10.jsonl
└── raw/
    ├── 2026-05-09-104212-prompt-a1b2c3d4.json   # Full request
    └── 2026-05-10-091500-prompt-e5f6g7h8.json

JSONL index (quick search)

Each line contains key metadata:

{
  "ts": "2026-05-09T02:42:12.608Z",
  "session": "agent:main:...",
  "model": "custom-provider/model-name",
  "inputTokens": 24072,
  "outputTokens": 70,
  "success": true,
  "rawFile": "raw/2026-05-09-104212-prompt-xxx.json"
}

Raw files (full details)

Each raw file contains the complete agent messages array (system prompt, user messages, assistant replies, tool calls/results) plus LLM usage and timing info.

Installation

openclaw plugin install prompt-logger

Requires OpenClaw Gateway v2026.5.0 or later.

Configuration

Add to openclaw.json:

{
  "plugins": {
    "entries": {
      "prompt-logger": {
        "enabled": true,
        // Required: this plugin uses llm_output/agent_end hooks which
        // need explicit conversation access opt-in (security gate)
        "hooks": {
          "allowConversationAccess": true
        },
        "config": {
          "retentionDays": 30,   // Auto-delete raw files older than N days
          "logInput": true,       // Record full prompt messages
          "logOutput": true       // Record LLM responses
        }
      }
    }
  }
}

⚠️ Important: Without hooks.allowConversationAccess: true, the plugin will load but silently produce no output. This is a Gateway security measure — non-bundled code plugins that access raw conversation content must opt in explicitly.

Query examples

# Find all failed requests today
jq 'select(.success == false)' prompts/2026-05-09.jsonl

# Find requests for a specific model
jq 'select(.model | test("deepseek"))' prompts/2026-05-09.jsonl

# Count total tokens used today
cat prompts/2026-05-09.jsonl | jq -s 'map(.inputTokens + .outputTokens) | add'

# View a specific full prompt
less prompts/raw/2026-05-09-104212-prompt-xxx.json

Disk usage

Each raw file is 100-500 KB depending on conversation length. The JSONL index is negligible (~200 bytes/entry). With retentionDays: 30, old files are cleaned automatically.

Publishing

1. Create an account on clawhub.ai
2. Install the clawhub CLI: npm i -g clawhub
3. Run clawhub login
4. Publish:

clawhub package publish prompt-logger --dry-run
clawhub package publish prompt-logger

License

MIT