Summary

On Telegram, /reasoning on drops the model's thinking lane. The thinking (type:"thinking" parts) is produced and archived but is never presented as persistent messages — only the model's plain text is shown. The same model's thinking renders correctly under /reasoning stream, so the capability exists; it is specifically lost under /reasoning on.

Environment

• OpenClaw 2026.6.8 (production)
• Provider/model: deepseek/deepseek-v4-flash
• Flags: /think high, /verbose off

Steps to reproduce

1. /model deepseek/deepseek-v4-flash, then /think high, /verbose off
2. /reasoning stream → send a reasoning+tool prompt → thinking streams in the ephemeral block (works as expected).
3. /reasoning on → send the same kind of prompt → no thinking messages appear.

Expected

/reasoning on should suppress ephemeral streaming and instead persist the thinking as permanent message(s) (parity with the intended behaviour).

Actual

Under /reasoning on, no thinking is persisted. What appears are the model's type:"text" parts (its conversational narration and final answer) — which can be mistaken for thinking but are not the thinking lane. The type:"thinking" parts are confirmed in the session archive but never rendered. Under /reasoning stream the same thinking renders in the ephemeral block, so the thinking lane is lost only under on (neither streamed nor persisted).

Notes

• Reproduced with deepseek as a clean, reliable emitter; affects reasoning models generally.
• Screen recordings of both the stream (working) and on (broken) states are available and can be attached.