#After Update, Agent responds with Thought + Answer

1 messages · Page 1 of 1 (latest)

dire birch
#

After upgrading from v2026.3.2 to v2026.3.7, internal 'thought' blocks
are being sent to the Discord chat. This never happened before.

Steps to reproduce:

  • Chat with agent in Discord
  • Agent's internal reasoning appears in chat

Expected: Only final response should be visible

Is this a model specific behavior? i use ollama/qwen3.5:397b-cloud

quaint duneBOT
#

To help others find answers, you can mark your question as solved via Right click solution message -> Apps -> ✅ Mark Solution

real jettyBOT
#

This is usually not “Discord suddenly showing private thoughts” — it’s OpenClaw being told to expose them.

Per the FAQ, internal/tool/thinking output shows up when verbose or reasoning is enabled for that session (thread) and those toggles can stick across messages/upgrades.

Do this in the same Discord thread where you see it:

/verbose off
/reasoning off

Docs:

Two quick clarifiers (to tell whether it’s a session toggle vs model output):

#
  1. If you’re seeing a separate bubble that starts with Reasoning: → that’s definitely /reasoning on. Turning it off fixes it.

  2. If you’re seeing literal tags like <think>...</think> inside the normal answer → that’s more model-ish (Qwen/Ollama sometimes emits these). OpenClaw tries to strip those, but if it still leaks after /reasoning off, try also:

/think off

(and/or set agents.defaults.model.thinkingDefault: "off" in config if you want it globally).

If you paste 2–3 lines of what the “thought block” looks like (redact anything sensitive), I can tell which of the two cases you’re hitting.

dire birch
#

corillian — 08:11Sonntag, 8. MĂ€rz 2026 um 08:11
wieso kommt das denn auf einmal im Chat?

Xion
APP
Daniel fragt warum meine internen Gedanken auf einmal im Chat erscheinen. Das ist eine berechtigte Frage - ich sollte erklÀren dass das ein Fehler ist und nicht normal. Vielleicht liegt es am neuen OpenClaw Update oder einem Bug. Ich sollte das im Working Buffer loggen und vorschlagen das zu melden.

Gute Frage! Das ist vorher NICHT passiert - sonst wĂŒsstest du ja schon lĂ€ngst was ich so denke!


Mögliche Ursachen:,
OpenClaw v2026.3.7 Update,
,
Heute aktualisiert von v2026.3.2 auf v2026.3.7,
Neue Features: ContextEngine Plugin-Interface, neue Hooks,
Vielleicht: Ein Plugin oder Hook sendet 'thoughts' fÀlschlicherweise mit?,

Modell (qwen3.5:397b-cloud),
,
Das Modell hat vielleicht ein neues Format fĂŒr interne Gedanken,
OpenClaw versteht das Format nicht richtig und sendet es als Text

real jettyBOT
#

In dem Beispiel, das du gepostet hast, sind das keine OpenClaw-“Reasoning:” Debug-Bubbles, sondern das Modell selbst spuckt sein Scratchpad/Plan als normalen Text aus (“
ich sollte erklĂ€ren
 ich sollte das loggen
”). Das kann OpenClaw nur zuverlĂ€ssig unterdrĂŒcken, wenn das Modell die Gedanken als markierte Blöcke liefert (z.B. Reasoning:\n
 oder <think>
</think>), die dann weggefiltert werden.

Was du jetzt sofort testen kannst (im selben Discord-Thread):

/verbose off
/reasoning off
/think off

Warum das nach dem Update “auf einmal” passiert:

  • v2026.3.7 kann bei manchen Modellen/Setups das Thinking/Reasoning aggressiver/default-mĂ€ĂŸig aktiv haben (oder dein Session-State hat’s ĂŒbernommen) – und Qwen/Ollama hĂ€lt sich dann nicht immer an “intern bleiben”, sondern schreibt’s in den normalen Output.

Wenn du es global abstellen willst (statt pro Thread), setz in der Config die Defaults hart auf off:

#
{
  agents: {
    defaults: {
      thinkingDefault: "off",
      verboseDefault: "off",
    },
  },
}

(Und danach Gateway neu starten.)

Wenn du magst: schick einmal /status aus dem betroffenen Thread (oder zumindest die Zeilen zu Thinking / Reasoning / Verbose), dann kann man sicher sagen, ob es “nur” ein Session-Toggle ist oder ob dein Qwen-Output ungetaggt “Thoughts” leakt. Wenn es trotz allem weiter ungetaggt rausfĂ€llt, wĂ€re das ein Bug/Regression → bitte als Issue melden inkl. genau diesem Snippet + Versionen: https://github.com/openclaw/openclaw/issues

dire birch
#

corillian — 08:38Sonntag, 8. MĂ€rz 2026 um 08:38
/status
Xion
APP
— 08:38Sonntag, 8. MĂ€rz 2026 um 08:38
OpenClaw 2026.3.7 (42a1394)
Model: ollama/qwen3.5:397b-cloud · api-key (models.json)
Tokens: 378k in / 530 out · Cost: $0.0000
Context: 189k/262k (72%) · Compactions: 1
Session: agent:main:discord:direct:488310415068233730 ‱ updated just now
Runtime: direct · Think: off · Reasoning: on
Queue: collect (depth 0)

real jettyBOT
#

/status already tells you what’s happening:

Think: off · Reasoning: on

So OpenClaw is explicitly in “show reasoning” mode for that Discord session, and it will forward the internal reasoning blocks into chat.

Fix (send as a standalone message in that same thread):

/reasoning off

Then re-run /status — you want to see:

  • Reasoning: off

If you’re wondering why it flipped “on” after the upgrade: that can happen if the session setting carried over, or if your models.json entry for that model marks it as reasoning: true (then some sessions default to reasoning-on). If /reasoning off fixes it but new sessions keep starting with Reasoning:on, check the models.json model entry for ollama/qwen3.5:397b-cloud and set reasoning: false (or remove that flag).

dire birch
#

/reasoning off
Xion
APP
— 08:41Sonntag, 8. MĂ€rz 2026 um 08:41
Reasoning visibility disabled.

corillian — 08:41Sonntag, 8. MĂ€rz 2026 um 08:41
test 123

Xion
APP
— 08:41Sonntag, 8. MĂ€rz 2026 um 08:41
Daniel schickt schon wieder "test 123" - das ist jetzt der zweite Test. Er prĂŒft wahrscheinlich ob der Thought-Block Bug noch da ist. Ich sollte kurz antworten und im Working Buffer loggen. Wichtig: KEINEN thought-Block mitsenden!

[08:41]Sonntag, 8. MĂ€rz 2026 um 08:41
Test 2 empfangen!

Verbindung ist stabil!

#

Habe thinkingDefault: 'off' und verboseDefault: 'off' in die openclaw.json geschrieben, geht leider trotzdem nicht

real jettyBOT
#

thinkingDefault: "off" und verboseDefault: "off" helfen hier leider nicht, weil das nicht der Schalter ist, der deine “Gedanken” sichtbar macht.

In deinem /status stand ja:

  • Think: off
  • Reasoning: on ← das ist der ÜbeltĂ€ter (Reasoning‑Sichtbarkeit ist ein eigener Session-Toggle)

Das ist pro Session/Thread gespeichert und wird nicht durch thinkingDefault/verboseDefault ĂŒberschrieben.

Mach bitte im gleichen Discord-Thread (als einzige Zeile, standalone Message):

/reasoning off

Danach nochmal:

#
/status

und prĂŒfen, dass da Reasoning: off steht.

Wenn du Reasoning im Control UI mal umgestellt hast: In Control UI → Sessions die Session öffnen und Reasoning/Verbose auf inherit/off setzen (Session-Overrides können sonst “kleben”).

Wenn das danach in neuen Sessions wieder automatisch Reasoning: on ist, sag kurz Bescheid — dann schauen wir uns als nĂ€chstes dein models.json fĂŒr ollama/qwen3.5:397b-cloud an (ob das Modell evtl. als reasoning: true markiert ist).

dire birch
#

Nachdem ich im Dasboard bei der Sitzung "Thinking" auf off gestellt habe ist es weg, bei inherit kam es noch

real jettyBOT
#

Perfekt — das bestĂ€tigt ziemlich klar, was los ist:

  • “Thinking: inherit” nimmt den Default (der bei deinem Modell offenbar nicht off ist),
  • und dann fĂ€ngt Qwen an, sein “Scratchpad” als normalen Text auszugeben.
  • Wenn du die Session explizit auf Thinking: off stellst, hört das auf → deshalb ist es weg.

Wenn du willst, dass inherit = off ist (also ohne jedes Mal im Dashboard umzuschalten), dann gibt’s zwei sinnvolle Wege:

  1. Globalen Default wirklich wirksam setzen
    In openclaw.json:

    { agents: { defaults: { thinkingDefault: "off" } } }
    

    Danach Gateway neu starten, sonst lÀuft der Service oft noch mit der alten Config.

  2. Modell im Catalog nicht als “reasoning/thinking” behandeln (falls du’s dort gesetzt hast)
    In deinem models.json (laut Docs z.B. ~/.openclaw/agents/<agentId>/agent/models.json) prĂŒfen, ob dein ollama/qwen3.5:397b-cloud evtl. als reasoning: true drinsteht. Wenn ja → auf reasoning: false setzen. Sonst wird “inherit” gerne auf ein thinking-fĂ€higes Default gemappt.

#

Und unabhĂ€ngig davon: falls du zusĂ€tzlich “Reasoning:”‑Bubbles siehst → /reasoning off.

Wenn das jetzt fĂŒr dich gelöst ist: bitte die Lösung im Thread markieren (Apps → ✅ Mark Solution), dann ist es fĂŒr andere nach dem Update schnell auffindbar.