8 replies

Opus 4.6 Thinking Content Leaking to Discord

The Problem

After switching to

anthropic/claude-opus-4-6

anthropic/claude-opus-4-6

as our default model, we noticed internal thinking/reasoning content rendering as visible messages in Discord — both in group channels and DMs. The model's internal monologue (e.g. "The user is asking about... Let me check...") appears as regular message text.

Root Cause (from source investigation)

Three issues compound to create the leak:

1. Opus 4.6 enables thinking by default.
Unlike previous Claude models, when the

thinking

thinking

parameter is omitted from the API request, Opus 4.6 uses adaptive thinking automatically. You must explicitly send

thinking: { type: "disabled" }

thinking: { type: "disabled" }

to turn it off. The pi-ai layer currently sets

body.thinking = undefined

body.thinking = undefined

(omits it) when no thinking budget is configured — which was effectively "off" for Opus 4.5 but means "adaptive thinking on" for 4.6.

2. The model duplicates thinking content into text blocks.
Opus 4.6 outputs reasoning in a

thinking

thinking

content block AND starts the visible

text

text

block with the same content (as a prefix, not an exact copy). OpenClaw's

repairMissingTextBlocks

repairMissingTextBlocks

src/agents/pi-embedded-runner/run/attempt.ts

src/agents/pi-embedded-runner/run/attempt.ts

has dedup logic, but it only catches exact matches — not prefix overlap. Example from our session transcript: thinking block = 149 chars, text block = 2067 chars starting with the same 149 chars.

3. Streaming bypasses dedup entirely.
Text deltas are sent to Discord in real-time via

handleMessageUpdate

handleMessageUpdate

→

onBlockReply

onBlockReply

onPartialReply

onPartialReply

. The dedup in

postProcessAssistantMessage

postProcessAssistantMessage

only runs on the final assembled message after streaming is complete — by which point the content is already in Discord.

Environment

- OpenClaw version: 2026.2.6-3
- Model: anthropic/claude-opus-4-6
- Channel: Discord (DMs and group channels)
- Thinking config: both

thinking=low

thinking=low

(explicit) and no thinking config (adaptive default) reproduce the issue
- Related GitHub issue: #6470

Steps to Reproduce

1. Set

anthropic/claude-opus-4-6

anthropic/claude-opus-4-6

as the primary model
2. Send any message via Discord
3. Observe the response — internal reasoning/thinking content appears as visible text before or mixed with the actual response

Suggested Fixes

pi-ai layer:
- For Opus 4.6+, send

thinking: { type: "disabled" }

thinking: { type: "disabled" }

instead of omitting the parameter when thinking is not configured

OpenClaw streaming path:
- Track thinking block content during streaming (

thinking_delta

thinking_delta

events) and suppress

text_delta

text_delta

content that matches/overlaps with accumulated thinking text

OpenClaw dedup:
- Extend

repairMissingTextBlocks

repairMissingTextBlocks

to handle prefix overlap, not just exact match — check if

textContent.startsWith(lastThinkingText)

textContent.startsWith(lastThinkingText)

and strip the prefix

Config workaround (current):
- Switch to

blockStreamingBreak: "message_end"

blockStreamingBreak: "message_end"

to delay Discord delivery until after dedup runs (loses streaming UX)
- Or revert to Opus 4.5 (what we did)

Workaround

We reverted to

anthropic/claude-opus-4-5

anthropic/claude-opus-4-5

as the default. Opus 4.6 is not safe for any messaging surface until the streaming dedup issue is resolved.

FotC�

Friends of the Crustacean 🦞🤝•4d ago•

8 replies

Kaelten

Opus 4.6 Thinking Content Leaking to Discord

The Problem

After switching to

anthropic/claude-opus-4-6

anthropic/claude-opus-4-6

Root Cause (from source investigation)

Three issues compound to create the leak:

1. Opus 4.6 enables thinking by default.
Unlike previous Claude models, when the

thinking

thinking

parameter is omitted from the API request, Opus 4.6 uses adaptive thinking automatically. You must explicitly send

thinking: { type: "disabled" }

thinking: { type: "disabled" }

to turn it off. The pi-ai layer currently sets

body.thinking = undefined

body.thinking = undefined

thinking

thinking

content block AND starts the visible

text

text

block with the same content (as a prefix, not an exact copy). OpenClaw's

repairMissingTextBlocks

repairMissingTextBlocks

src/agents/pi-embedded-runner/run/attempt.ts

src/agents/pi-embedded-runner/run/attempt.ts

handleMessageUpdate

handleMessageUpdate

→

onBlockReply

onBlockReply

onPartialReply

onPartialReply

. The dedup in

postProcessAssistantMessage

postProcessAssistantMessage

only runs on the final assembled message after streaming is complete — by which point the content is already in Discord.

Environment

- OpenClaw version: 2026.2.6-3
- Model: anthropic/claude-opus-4-6
- Channel: Discord (DMs and group channels)
- Thinking config: both

thinking=low

thinking=low

(explicit) and no thinking config (adaptive default) reproduce the issue
- Related GitHub issue: #6470

Steps to Reproduce

1. Set

anthropic/claude-opus-4-6

anthropic/claude-opus-4-6

as the primary model
2. Send any message via Discord
3. Observe the response — internal reasoning/thinking content appears as visible text before or mixed with the actual response

Suggested Fixes

pi-ai layer:
- For Opus 4.6+, send

thinking: { type: "disabled" }

thinking: { type: "disabled" }

instead of omitting the parameter when thinking is not configured

OpenClaw streaming path:
- Track thinking block content during streaming (

thinking_delta

thinking_delta

events) and suppress

text_delta

text_delta

content that matches/overlaps with accumulated thinking text

OpenClaw dedup:
- Extend

repairMissingTextBlocks

repairMissingTextBlocks

to handle prefix overlap, not just exact match — check if

textContent.startsWith(lastThinkingText)

textContent.startsWith(lastThinkingText)

and strip the prefix

Config workaround (current):
- Switch to

blockStreamingBreak: "message_end"

blockStreamingBreak: "message_end"

to delay Discord delivery until after dedup runs (loses streaming UX)
- Or revert to Opus 4.5 (what we did)

Workaround

We reverted to

anthropic/claude-opus-4-5

anthropic/claude-opus-4-5

as the default. Opus 4.6 is not safe for any messaging surface until the streaming dedup issue is resolved.

Opus 4.6 Thinking Content Leaking to Discord

The Problem

Root Cause (from source investigation)

Environment

Steps to Reproduce

Suggested Fixes

Workaround

Opus 4.6 Thinking Content Leaking to Discord

The Problem

Root Cause (from source investigation)

Environment

Steps to Reproduce

Suggested Fixes

Workaround

Similar Threads

Similar Threads

Similar Threads