Summary

On OpenClaw 2026.4.5, assistant messages tagged internally as phase: "commentary" are leaking to Telegram as user-visible messages.

This should never happen. Commentary/tool-chatter should remain internal, while only the final user-facing reply should be delivered.

Impact

High trust / privacy issue.

In a direct Telegram chat, raw internal working notes were surfaced to the user, for example:

• We need answer yes there is risk because script deletes any gateway orphan...
• Need adjust config to use openai api models maybe gpt-4.1/gpt-4o-mini?...

These were not final answers. They were internal drafting/commentary text.

Environment

• OpenClaw: 2026.4.5 (3e72c03)
• Surface: Telegram direct chat
• Channel config had channels.telegram.streaming: partial at the time of the observed leaks
• Provider/model path when the leaks occurred:
  • provider: openai-codex
  • model: gpt-5.4
• The leaked session artifacts explicitly mark these messages as phase: "commentary"

Evidence from session logs

From the session JSONL, leaked assistant entries are stored like this:

• timestamp: 2026-04-06T12:50:17.088Z
• assistant text begins:
  • We need answer yes there is risk because script deletes any gateway orphan, including internal dreaming cron if not in supabase...
• metadata includes:
  • textSignature.phase = "commentary"

Another example:

• timestamp: 2026-04-07T08:47:33.195Z
• assistant text begins:
  • Need adjust config to use openai api models maybe gpt-4.1/gpt-4o-mini?...
• metadata includes:
  • textSignature.phase = "commentary"

These same commentary texts were observed on Telegram as user-visible assistant messages.

Expected behavior

• Commentary-phase content should never be sent to the channel surface.
• Only final user-facing assistant output should be delivered.
• Tool-call planning / scratch text / commentary should remain internal even when preview streaming is enabled.

Actual behavior

Commentary-phase assistant text was delivered to Telegram as visible chat messages.

Notes / likely area to inspect

This may be specific to the OpenAI Codex / Responses path or to how commentary-phase chunks are handled in Telegram delivery/stream finalization.

The key issue is not merely “reasoning was exposed” — the log evidence shows these are explicitly tagged commentary-phase messages.

Suggested debugging directions

• Verify Telegram delivery excludes any message/content block tagged phase: commentary
• Inspect OpenAI Codex / Responses provider path for commentary vs final separation
• Inspect streaming/finalization path on Telegram for accidental commentary flush or wrong phase selection
• Add an invariant/assertion that commentary-phase text can never be routed to channel delivery