Summary

After upgrading to 2026.2.22-2, the @openclaw/msteams plugin delivers raw reasoning/thinking tokens to the Teams channel instead of (or mixed with) the final answer when reasoning is enabled on the model. The WebChat UI shows the correct final response, but Teams receives the internal thinking stream.

This was triggered by PR #22513, which changed the default reasoningLevel to "on" for models with reasoning: true. The MS Teams plugin does not separate reasoning stream lanes from the answer stream, unlike the fixes applied to Telegram in 2026.2.21 (#20774). A parallel bug exists for Telegram (#24376) and a feature request for Slack (#19411), but no equivalent issue exists for the @openclaw/msteams plugin.

Workaround: /reasoning off restores correct behavior, but users should be able to use reasoning without it leaking into the channel output.

Steps to reproduce

1. Install @openclaw/msteams plugin and configure MS Teams channel.
2. Use a model with reasoning: true (e.g., Codex 5.3 via OpenAI OAuth).
3. Update OpenClaw to 2026.2.22-2 (or any version including PR Default reasoning to on when model has reasoning: true (fix #22456) #22513).
4. Send a message to the bot via Microsoft Teams.
5. Observe that Teams receives the reasoning/thinking tokens (e.g., internal chain-of-thought text) instead of the final answer.
6. Compare with WebChat UI — WebChat shows the correct final response while Teams gets the thinking stream.

Expected behavior

When reasoning is enabled, the MS Teams plugin should:

1. Process reasoning/thinking tokens internally (or optionally display them in a collapsed/separate format).
2. Deliver only the final answer text to the Teams channel — the same content visible in the WebChat UI.

This is consistent with the Telegram streaming fix in 2026.2.21 (#20774), which split reasoning and answer draft preview lanes to prevent cross-lane overwrites.

Actual behavior

Teams receives the raw reasoning/thinking tokens (internal chain-of-thought) instead of the final answer. The response appears as the model's internal reasoning process (e.g., "User is asking about X, I should respond with...") rather than the actual intended reply.

The WebChat UI correctly shows only the final answer, confirming the issue is in the MS Teams plugin's outbound message rendering — not in the agent/model layer.

/reasoning off restores normal behavior, confirming the reasoning stream is being routed to the channel output instead of being filtered.

OpenClaw version

2026.2.22-2

Operating system

macOS 15.7.3 (arm64), Node 22.22

Install method

npm global

Logs, screenshots, and evidence

No logs captured yet — will add gateway log snippets and Teams message screenshots if needed.

The behavior is easily reproducible: any message sent via Teams with reasoning enabled returns thinking tokens. `/reasoning off` immediately fixes the output.

Impact and severity

Affected: All MS Teams users with reasoning-capable models after upgrading to any version including #22513.
Severity: High — effectively breaks the Teams channel for reasoning-enabled models (responses are unusable).
Frequency: 100% repro when reasoning is enabled.
Consequence: Teams channel delivers internal thinking tokens instead of answers, making the bot unusable without the /reasoning off workaround.

Additional information

Model/provider: Codex 5.3 via OpenAI OAuth (ChatGPT subscription)
Integration/channel: Microsoft Teams (@openclaw/msteams plugin)

Related issues:

• [Bug]: Telegram stream + reasoning leaks toolUse intermediate status text into user-visible replies (regression) #24376 — Telegram stream + reasoning leaks toolUse intermediate status text (filed 1 day ago, same root cause in Telegram plugin)
• feat(slack): render thinking/reasoning blocks as collapsible sections #19411 — Slack feature request for collapsible reasoning blocks
• Default reasoning to on when model has reasoning: true (fix #22456) #22513 — PR that changed reasoning defaults to "on" for reasoning-capable models
• fix: split telegram reasoning and answer draft streams #20774 — Telegram streaming fix that split reasoning and answer draft preview lanes (2026.2.21)

Suggested fix direction:
The @openclaw/msteams plugin likely needs the same reasoning/answer stream lane separation that was applied to Telegram in #20774. The handleMessageEnd path (or equivalent in the Teams plugin) should filter out reasoning blocks before emitting to the Teams channel, similar to how WebChat already handles this correctly.

Temporary workaround: /reasoning off in Teams chat.