Meta Footer

Appends a stats footer to every bot reply: model, thinking level, token usage, context window, cache hit rate, and compaction count.

lqqk7@lqqk7

Install

openclaw plugins install clawhub:claw-meta-footer

claw-meta-footer

An OpenClaw plugin that appends a stats footer to every bot reply, giving you at-a-glance visibility into model usage, token consumption, and session state — right inside your Telegram chat.

What It Shows

Every bot reply gets a footer like this:

text

`───────────────`
🤖 Model: `claude-sonnet-4-6`
🧠 Think: high
🔢 In: 12.3k   Out: 0.8k
📊 Context: 13.1k / 200k (6.6%)
💾 Cache: 11.9k hit (88.4%)
🔁 Compact: 2

Field	Description
Model	The model ID that generated the reply
Think	Thinking/reasoning level (`off`, `low`, `medium`, `high`, `xhigh`, `adaptive`)
In / Out	Input and output token counts for this turn
Context	Tokens currently in context vs. the model's context window limit
Cache	Cache-read token count and hit rate for this turn
Compact	Number of context compactions that have occurred in this session

Cache and Compact lines only appear when there is data to show.

Requirements

OpenClaw >= 1.0.0
Telegram channel with streaming disabled — the plugin hooks into message_sending, which is only triggered when streaming is off

Installation

bash

openclaw plugins install clawhub:claw-meta-footer

Configuration

1. Disable streaming on Telegram

In your openclaw.json, add "streaming": "off" to your Telegram channel config:

json

{
  "channels": {
    "telegram": {
      "streaming": "off"
    }
  }
}

Without this, the plugin won't fire — streamed messages bypass the message_sending hook entirely.

2. Plugin options (optional)

json

{
  "plugins": {
    "claw-meta-footer": {
      "enabled": true,
      "skipSubagent": true
    }
  }
}

Option	Type	Default	Description
`enabled`	boolean	`true`	Toggle the footer on/off
`skipSubagent`	boolean	`true`	Hide footer on subagent replies (recommended — subagents can be noisy)

How It Works

llm_output hook — captures token usage (input, output, cacheRead, cacheWrite), model ID, and provider from each LLM response, keyed by channel + chat ID
message_sending hook — before the reply is sent, retrieves the cached stats, reads thinkingLevel and compactionCount from the session file, resolves the context window size, builds the footer, and appends it to the message content

Context window sizes are resolved via a priority chain:

User-configured contextWindow in openclaw.json models
Composite provider/model lookup (mirrors OpenClaw's internal overrides)
Plain model ID lookup
Prefix matching fallback

License

MIT