Summary

Streaming 1:1 replies (added in PR #51808) work great for short conversational exchanges, but the final reply is dropped when an agent performs a long tool chain (3+ sequential tool calls spanning 30+ seconds).

Environment

• OpenClaw v2026.4.1
• Microsoft Teams plugin (bundled)
• Single-tenant bot registration
• DM (1:1) conversation

Steps to Reproduce

1. User sends a message that triggers multiple sequential tool calls
2. Agent begins streaming reply — initial text appears in Teams
3. Agent executes 4 sequential exec tool calls with 10-20 second delays each (~60 seconds total)
4. Agent completes tool chain and writes final reply text

Expected Behavior

• Streaming card should continue updating during tool execution (e.g., showing status updates)
• Final reply text should appear in Teams after all tool calls complete

Actual Behavior

• Streaming card shows initial pre-tool text correctly
• Card freezes during tool execution — no updates visible
• Final reply after tool chain (~60 seconds later) is silently dropped
• User only sees the partial initial text

Analysis

The Bot Framework TurnContext proxy still expires after ~30 seconds even with streaming enabled. Short replies work because the streaming card finishes before the proxy timeout. Long tool chains exceed this window.

Possible approaches:

1. Heartbeat updates during tool execution — Send periodic updateActivity calls to the streaming card during tool calls to keep the TurnContext alive (e.g., "Running command 2/4...")
2. Proactive fallback — Detect when a turn exceeds a threshold and automatically switch to proactive message delivery (using stored conversation references) for the final reply
3. Tool-aware streaming — Emit streaming card updates at tool boundaries (tool start/complete events) to maintain the connection

Workaround

We currently use a hybrid approach:

• Turn-context (streaming) for simple replies with no tools or quick tools (< 15 seconds)
• Proactive message tool sends for replies after long tool chains (30+ seconds)

This works but requires agent-side awareness of delivery rules, which is fragile and adds complexity to the system prompt.

Related

• PR msteams: implement Teams AI agent UX best practices #51808 — Original Teams SDK migration with streaming support
• The Feishu channel handles a similar pattern with streaming card transport that persists across tool calls