My OpenClaw agents have started to think like Goblins
internal assistant scratchpad / working notes are leaking into real Discord DMs, and the leaked content has degraded into compressed “goblin mode” shorthand instead of coherent English.
This is not just the OpenClaw control UI. These messages are appearing in the real user-facing Discord conversation.
While the assistant is working on a task, instead of only getting:
- the final user-facing reply
- or a clean progress update
…I sometimes get internal working scratchpad text dumped directly into my actual Discord DM.
The internal text used to look more like coherent English / progress narration.
Example from an earlier browser/avatar flow leak:
-
-
-
Still not ideal to leak, but recognizably internal narration.
Now on the newer build / current setup, the leaked text often looks like compressed goblin-mode operator shorthand, e.g.:
-
-
-
It looks like ultra-compressed internal routing notes rather than the assistant’s normal voice.
This is not just cosmetic / embarrassing.
It breaks the intended boundary between:
- user-facing assistant voice
- backstage internal working state
I do not want internal scratchpad / chain-of-work / partial progress notes dumped into the normal chat surface.
It feels like accidental “dollhouse cross-sections” of the assistant rather than the intended final interaction.
- This is happening in Discord DMs
- Not just in the OpenClaw control UI
- Initially suspected Discord message chunking / streaming
- Checked and changed:
-
- account-level Discord
-
- The weird short fragments / scratchpad leaks still happened
- So this does not appear to be just normal final message chunking
- It looks more like intermediate assistant text is being surfaced as real chat output
Relevant current state:
- source install from local checkout (
- moved onto newer version to test GPT-5.4 support
- Discord provider in config has streaming disabled globally and per-account
- issue still reproduces after that
1. Is this a known regression on dev/main?
2. Is there a config path that controls whether intermediate assistant text / internal progress notes are surfaced?
3. Is this provider/model-specific (for example Opus vs Sonnet vs Codex)?
4. Is this tied to long-running tool workflows / exec / browser automation / multi-step reasoning?
5. Is there a recommended way to ensure only final user-facing messages are delivered unless explicit progress updates are intended?
If helpful, I can also provide screenshots / exact leaked message examples.
This is not just the OpenClaw control UI. These messages are appearing in the real user-facing Discord conversation.
The symptom
While the assistant is working on a task, instead of only getting:
- the final user-facing reply
- or a clean progress update
…I sometimes get internal working scratchpad text dumped directly into my actual Discord DM.
What makes it especially weird
The internal text used to look more like coherent English / progress narration.
Example from an earlier browser/avatar flow leak:
-
the crop dialog is open-
let me click set new profile picture-
avatar done, I look like me nowStill not ideal to leak, but recognizably internal narration.
Now on the newer build / current setup, the leaked text often looks like compressed goblin-mode operator shorthand, e.g.:
-
Need do actions. maybe 1password skill none.-
Need maybe pipx? check if installed.-
Need parse env safely maybe grep specific line.It looks like ultra-compressed internal routing notes rather than the assistant’s normal voice.
Why this matters
This is not just cosmetic / embarrassing.
It breaks the intended boundary between:
- user-facing assistant voice
- backstage internal working state
I do not want internal scratchpad / chain-of-work / partial progress notes dumped into the normal chat surface.
It feels like accidental “dollhouse cross-sections” of the assistant rather than the intended final interaction.
Context / things already checked
- This is happening in Discord DMs
- Not just in the OpenClaw control UI
- Initially suspected Discord message chunking / streaming
- Checked and changed:
-
channels.discord.streaming = off- account-level Discord
streaming = off-
channels.discord.maxLinesPerMessage = 120- The weird short fragments / scratchpad leaks still happened
- So this does not appear to be just normal final message chunking
- It looks more like intermediate assistant text is being surfaced as real chat output
Config / runtime context
Relevant current state:
- source install from local checkout (
~/oclaw- moved onto newer version to test GPT-5.4 support
- Discord provider in config has streaming disabled globally and per-account
- issue still reproduces after that
What I’m trying to determine
1. Is this a known regression on dev/main?
2. Is there a config path that controls whether intermediate assistant text / internal progress notes are surfaced?
3. Is this provider/model-specific (for example Opus vs Sonnet vs Codex)?
4. Is this tied to long-running tool workflows / exec / browser automation / multi-step reasoning?
5. Is there a recommended way to ensure only final user-facing messages are delivered unless explicit progress updates are intended?
Ask
If helpful, I can also provide screenshots / exact leaked message examples.
The AI that actually does things. Emails, calendar, home automation - all from your favorite chat app. New shell, ready to help. The lobster way.
134,723Members
Resources
Was this page helpful?
