Fix/inbound dedup messageid

[mosir]

📑 Description

This PR fixes repeated outbound replies observed in Telegram when heartbeat or tool-call loops occur.

Two targeted safeguards are included:

1. Ignore bot-originated inbound Telegram messages to prevent self-echo loops.
2. Suppress duplicate message tool sends within the same processing round (same channel + chat_id + content), even
   if the LLM repeatedly calls the same tool.

Additionally, this PR fixes a state-reset issue: the message tool dedup state was previously reset on every
SetContext call, which made same-round dedup ineffective.

🛠️ Type of Change

• 🐶 Bug fix (non-breaking change which fixes an issue)
• ✨ New feature (non-breaking change which adds functionality)
• 📝 Documentation update
• ♻️ Code refactoring (no functional changes, no api changes)

🤖 AI Code Generation

• 🤖 Fully AI-generated (100% AI, 0% Human)
• 🛠️ Mostly AI-generated (AI draft, Human verified/modified)
• 👆‍💻 Mostly Human-written (Human lead, AI assisted or none)

🔆 Related Issue

N/A

📎 Technical Context (Skip for Docs)

• Reference URL: N/A
• Reasoning:
  • Repeated logs showed identical tool-calls:
    • Tool call: message({"content":"HEARTBEAT_OK"}) across many iterations.
  • Telegram channel previously did not explicitly ignore bot-originated inbound messages.
  • message tool dedup state was reset too frequently (SetContext on each tool execution), so duplicate
    suppression in the same round did not hold.

Files changed

• pkg/channels/telegram/telegram.go
  • Ignore bot-originated/self messages in inbound handler.
• pkg/tools/message.go
  • Add same-round duplicate suppression guard.
  • Move round reset from SetContext to explicit BeginRound().
• pkg/agent/loop.go
  • Call BeginRound() once per inbound round before tool iterations.
• pkg/tools/message_test.go
  • Add/extend tests for duplicate suppression and round reset behavior.

Follow-up suggestion (not in this PR)

• There is a prompt-policy conflict that can still encourage repeated message tool-calls:
  • Global: ALWAYS use tools (pkg/agent/context.go)
  • Heartbeat: respond only with HEARTBEAT_OK when no action needed (pkg/heartbeat/service.go)
• Recommend a separate PR to refine heartbeat-specific prompting so HEARTBEAT_OK is treated as an internal
  completion signal instead of a tool-send objective.
• Also recommend keeping “Agent global messageID dedup” as a separate follow-up proposal if needed.

🧪 Test Environment

• Hardware: PC
• OS: Windows 11
• Model/Provider: main default model (tool-loop reproduction)
• Channels: Telegram (primary reproduction), Slack (regression check)

📳 Evidence (Optional)

Click to view Logs/Screenshots

• Reproduction logs (before fix) showed repeated:
  • LLM requested tool calls ... tools=[message]
  • Tool call: message({"content":"HEARTBEAT_OK"})
• Validation commands:
  • go test ./pkg/tools -run MessageTool ✅
  • go test ./pkg/agent ✅
  • go test ./pkg/channels/telegram ./pkg/channels/slack ./pkg/channels/discord ./pkg/channels/line ✅ (where test
    files exist)

☑️ Checklist

• My code/docs follow the style of this project.
• I have performed a self-review of my own changes.
• I have updated the documentation accordingly.

[mosir]

补充说明一下问题的原因

1. 根源

• 心跳 prompt 写了“无事可做时只回复 HEARTBEAT_OK”。
• 全局系统提示又写了“ALWAYS use tools（必须用工具）”。
• 模型在这两个约束下，会把“回复 HEARTBEAT_OK”理解成调用 message 工具发出去，而不是直接文本结束。
• 工具返回给 LLM 的是 Message sent to ...，这不等于任务目标 HEARTBEAT_OK，模型就继续再调一次 message，形成循环。

2. Telegram部分的实现缺少“忽略 bot 自己消息”的防护，会放大回声风险（WhatsApp 的实现应该也有此类问题）。