v2.6.0

Released: July 20, 2026

Highlights

• TelegramChannel: inline keyboard support (@toolpack-sdk/agents) — send messages with interactive buttons via metadata.replyMarkup, handle button taps as first-class AgentInput, and dismiss loading spinners with answerCallbackQuery()
• BaseAgent: labeled delegation results (@toolpack-sdk/agents) — delegate_to_agent now wraps sub-agent output with a [Response from <agent> — task complete] label so orchestrators clearly distinguish a finished result from a live conversation message

---

Breaking changes

None.

---

@toolpack-sdk/agents

TelegramChannel: inline keyboards

Sending buttons

Pass metadata.replyMarkup in the AgentOutput to attach an InlineKeyboardMarkup to any message. Existing callers that omit replyMarkup are unaffected.

await channel.send({
  output: 'Approve this trade?',
  metadata: {
    chatId: 123456789,
    replyMarkup: {
      inline_keyboard: [[
        { text: '✅ Approve', callback_data: 'approve:abc-123' },
        { text: '❌ Reject',  callback_data: 'reject:abc-123'  },
      ]],
    },
  },
});

Receiving button taps

normalize() now handles callback_query updates as a distinct path. The resulting AgentInput has context.isCallback: true and the following fields:

Context field	Type	Description
isCallback	true	Always set on callback inputs
callbackQueryId	string	Pass to answerCallbackQuery() to dismiss the spinner
callbackData	string	The callback_data value from the tapped button
chatId	number	Numeric chat ID
userId	number	Numeric user ID of the person who tapped
username / firstName / lastName	string | undefined	Sender identity
messageId	number	ID of the message that contained the keyboard

Dismissing the loading spinner

New method answerCallbackQuery(callbackQueryId, text?) calls the Telegram answerCallbackQuery endpoint. The agent must call this after handling the tap — it is not triggered automatically.

async invokeAgent(input: AgentInput): Promise<AgentResult> {
  if (input.context?.isCallback) {
    const result = await this.handleTap(input.context.callbackData as string);
    await (this.channels[0] as TelegramChannel).answerCallbackQuery(
      input.context.callbackQueryId as string,
      'Recorded.',
    );
    return result;
  }
  return this.run(input.message ?? '');
}

Throws if the Telegram API returns an error (e.g. the query has expired).

---

BaseAgent: delegate_to_agent results are now labeled

When delegation.mode is 'await', the tool result returned to the LLM is now prefixed with [Response from <agent> — task complete] before the sub-agent's output. This prevents the orchestrating LLM from treating a completed sub-agent result as a live conversational reply and accidentally responding to it instead of synthesising it for the end user.

Before: the raw AgentResult JSON was returned directly
After: the output field is prepended with [Response from browser — task complete]\n\n (using the actual target agent name)

No configuration change is required — the label is injected automatically by BaseAgent.

---

Install

npm install @toolpack-sdk/agents@2.6.0