Summary

When using ollama/kimi-k2.6:cloud (and likely kimi-k2.5:cloud) with Think: off, the model's inline reasoning text leaks into the visible chat output. The gateway correctly sends think: false (native Ollama) and thinking: { type: "disabled" } (Moonshot wrapper) on outgoing requests, but the model still emits reasoning text inline — separated from the actual response by a boundary delimiter. The Ollama provider has no response-level stripper for this inline reasoning, unlike the opencode-go provider which has stripOpencodeGoKimiReasoningPayload.

Environment

• OpenClaw version: 2026.5.22 (a374c3a)
• Provider: ollama
• Model: ollama/kimi-k2.6:cloud (also observed with kimi-k2.5:cloud)
• Runtime: Think: off
• OS: Windows 10.0.26200 (x64)
• Ollama base URL: http://192.168.1.72:11434

Reproduction

Config (relevant excerpt)

{
  models: {
    providers: {
      ollama: {
        baseUrl: "http://192.168.1.72:11434",
        api: "ollama",
        models: [
          {
            id: "kimi-k2.6:cloud",
            name: "kimi-k2.6:cloud",
            reasoning: false,
            params: {
              num_ctx: 262144
            }
          }
        ]
      }
    }
  },
  agents: {
    defaults: {
      model: { primary: "ollama/kimi-k2.6:cloud" }
    }
  }
}

Steps

1. Set primary model to ollama/kimi-k2.6:cloud
2. Ensure thinking / Think is off
3. Send any message that triggers the agent
4. Observe the assistant response in chat

Actual behavior

The visible response contains the model's internal reasoning monologue, followed by a boundary delimiter, then the actual response. Example from session history:

"The user is asking what projects we've worked on so far. Based on my memory files...

...Let me provide a clear summary. ️ So far we've got **3 active projects** tracked:"

The text before ️ is raw reasoning that should be internal only.

Expected behavior

Only the text after the reasoning delimiter should be visible to the user. The reasoning block should be stripped, discarded, or stored as internal metadata — never rendered as chat content.

Root cause analysis

1. Request side is correct — createConfiguredOllamaCompatStreamWrapper applies both:

   • createOllamaThinkingWrapper(..., false) → sets think: false on native Ollama payload
   • createMoonshotThinkingWrapper(..., "disabled") → sets thinking: { type: "disabled" }

2. Model ignores the disable signal — kimi-k2.6 still outputs reasoning inline, likely because the Ollama API passthrough doesn't propagate the disable parameter correctly to the underlying model, or the model inherently emits reasoning regardless.

3. Missing response stripper — The opencode-go provider has stripOpencodeGoKimiReasoningPayload which:

   • Deletes reasoning, reasoning_details, reasoning_content, reasoning_text fields
   • Filters out type: "thinking" / type: "reasoning" content parts from messages
   • Replaces stripped content with [assistant reasoning omitted]

   The ollama provider has no equivalent response sanitizer for Kimi models.

Cross-references

• opencode-go/kimi-k2.6 (and other openai-completions models): reasoning field leaks through passthrough replay policy, rejected as "Extra inputs are not permitted" on multi-turn #81988 — opencode-go/kimi-k2.6: reasoning field leaks through passthrough replay policy (same family, different provider)
• [Bug]: OpenCode Go / Kimi K2.6 sends unsupported reasoning_details in replayed messages #83812 — opencode-go/kimi-k2.6 sends unsupported reasoning_details in replayed messages (request-side fix for opencode-go)
• [Bug]: Reasoning/thinking content leaking to Discord regardless of model #6470 — Discord: reasoning content posted as regular messages (general reasoning leak class)
• how can i disable thinking mode? ollama/ollama#10456 — Ollama-level discussion on disabling thinking mode

Suggested fix

Add a Kimi-specific response sanitizer in the Ollama provider, analogous to what exists for opencode-go:

Option A (provider-level): In createConfiguredOllamaCompatStreamWrapper, when isOllamaCloudKimiModelRef(modelId) is true, wrap the stream with a response interceptor that strips inline reasoning text from assistant message content before it reaches the user.

Option B (gateway-level): Add a general stripInlineReasoningFromAssistantText utility in the message processing pipeline that recognizes the ️ (or equivalent) delimiter and splits/omits the reasoning portion.

Option C (model registry): Mark kimi-k2.6:cloud and kimi-k2.5:cloud under the Ollama provider as reasoning: true with a reasoningOutputMode: "inline" so the gateway knows to apply stripping regardless of what the model claims.

Workarounds

• Switch to a non-reasoning model (e.g., llama3.2:3b, gemma4)
• Modify the model's Ollama Modelfile to inject a system prompt forbidding reasoning output

Impact

• Severity: Medium-High — leaks internal decision-making and planning to user-visible chat
• Affected channels: All (webchat, Discord, Telegram, etc.)
• Frequency: Every multi-turn assistant message with ollama/kimi-k2.6:cloud