assistant-burst-summary-hook

assistant-burst-summary-hook

Cross-harness hooks for Ethan-style Telegram voice workflows.

After a voice note, if the assistant sends 3+ Telegram messages and then goes quiet for 5 minutes, this package sends one Telegram blockquote summary of that burst.

It includes:

• a shared deterministic burst-summary state machine,
• an OpenClaw plugin using message_received / message_sending / message_sent,
• a Claude Code plugin/hook pack using UserPromptSubmit / PostToolUse / Stop,
• install notes for Agent Bridge rollouts,
• synthetic tests proving threshold, quiet-window, voice-boundary, and no-loop behavior.

Default behavior

• threshold: 3 assistant messages.
• inactivityMs: 300000 (5 minutes).
• summary format: one Telegram MarkdownV2 blockquote.
• summary messages do not count toward the next burst.
• after a summary, another summary requires another threshold unsummarized assistant messages.

Install: OpenClaw

From a local checkout:

openclaw plugins install /Users/ethansk/Projects/assistant-burst-summary-hook --link
openclaw plugins enable assistant-burst-summary-hook

Config lives under plugins.entries.assistant-burst-summary-hook. Minimal example:

{
  plugins: {
    entries: {
      "assistant-burst-summary-hook": {
        enabled: true,
        threshold: 3,
        inactivityMs: 300000,
        telegramBotTokenEnv: "TELEGRAM_BOT_TOKEN",
        telegramParseMode: "MarkdownV2"
      }
    }
  }
}

The OpenClaw adapter resolves the Telegram bot token from, in order:

1. plugin config telegramBotToken,
2. env var named by telegramBotTokenEnv, default TELEGRAM_BOT_TOKEN,
3. active OpenClaw Telegram channel config.

It uses the outgoing OpenClaw message target as the Telegram chat id, so a separate telegramChatId is usually not needed for OpenClaw.

Install: Claude Code

Claude Code hooks need a chat id because command hooks do not have OpenClaw's channel delivery context.

claude plugin marketplace add /Users/ethansk/Projects/assistant-burst-summary-hook --scope user
claude plugin install assistant-burst-summary-hook@assistant-burst-summary-hook --scope user

Create ~/.config/assistant-burst-summary-hook/config.json:

{
  "threshold": 3,
  "inactivityMs": 300000,
  "telegramBotTokenEnv": "TELEGRAM_BOT_TOKEN",
  "telegramChatId": "6164541473",
  "telegramParseMode": "MarkdownV2"
}

Or point Claude hooks at another file:

export ASSISTANT_BURST_SUMMARY_CONFIG=/path/to/config.json

Config reference

See examples/config.json for a full example.

Common fields:

• enabled: boolean, default true.
• threshold: number of unsummarized assistant messages required, default 3.
• inactivityMs: quiet window after newest assistant message, default 300000.
• maxMessagesInSummary: cap on messages included in one summary, default 8.
• stateDir: durable JSON state directory.
• telegramBotTokenEnv: env var containing the bot token.
• telegramBotToken: direct token override; avoid committing this.
• telegramChatId: required for Claude Code unless PostToolUse extracts a chat id.
• telegramParseMode: MarkdownV2, HTML, or none.
• voiceBoundaryPatterns: extra strings that identify a voice-note transcript.
• dryRun: evaluate/send without calling Telegram.

Validation

npm test
node bin/assistant-burst-summary-hook.js simulate --json
node --input-type=module -e "await import('./src/openclaw/index.js')"
claude plugin validate .

Security and privacy

• No secrets are committed.
• Tests use synthetic transcripts only.
• State files contain compact assistant-message excerpts, not raw session JSONL.
• The Telegram Bot API call only sends the generated summary to the configured/observed chat target.
• Public repo contents are safe to publish; machine-local config files are intentionally excluded.

Agent Bridge rollout

See docs/agent-bridge-rollout.md.