Deep Research: is the Witsy-driven mode broken?

[natesilva]

When using the Deep Research feature, I had been using the Multi-Step (Witsy-driven) mode. But this keeps failing with:

This model was not able to provide a research plan in the expected format. You can try again with this model or with another model.

I’d file a bug but there’s a good chance this is user error or something in my LLM setup. (can anyone else duplicate this?)

I’ve now switched to the Agentic Loop mode, which does work.

I liked the Witsy-driven mode because it gave me better feedback while it was working, with a checklist of steps that it was working on. But Agentic Loop gives good final results.

[nbonamy]

What model are you using?

[natesilva]

It doesn’t matter what model I use. I’ve tried several, including Claude models (using the Anthropic integration), and Gemini models (using the Google integration), as well as Claude/Gemini/GLM-5 via OpenRouter. All return the same message.

Here’s an example with Opus 4.6, accessed using the Anthropic API:

"Plan" section expanded:

Request Body

{
  "model": "claude-opus-4-6",
  "system": "You are a status update generator for an autonomous task execution system.\n\nGenerate a concise status update (1-2 sentences maximum) for the user based on the current progress.\n\nExamples:\n- \"Let me analyze your request and create an execution plan.\"\n- \"I've created a plan with 5 tasks. Starting execution now.\"\n- \"Working on task 2 of 5: writing blog articles.\"\n- \"Completed research phase. Moving on to content creation.\"\n\nKeep it concise, natural, and user-friendly. Do NOT include prefixes like \"Status Update:\" or technical jargon.",
  "messages": [
    {
      "role": "user",
      "content": [
        {
          "type": "text",
          "text": "I am going to create a research plan for the following topic: What are my options for smart, hardwired smoke detectors for my house?"
        }
      ]
    }
  ],
  "max_tokens": 32000,
  "stream": true
}

Response Body

event: message_start
data: {"type":"message_start","message":{"model":"claude-opus-4-6","id":"msg_REDACTED","type":"message","role":"assistant","content":[],"stop_reason":null,"stop_sequence":null,"usage":{"input_tokens":170,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"cache_creation":{"ephemeral_5m_input_tokens":0,"ephemeral_1h_input_tokens":0},"output_tokens":1,"service_tier":"standard","inference_geo":"global"}}          }

event: content_block_start
data: {"type":"content_block_start","noindex":0,"content_block":{"type":"text","text":""}}

event: ping
data: {"type": "ping"}

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":"Let"}       }

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":" me analyze"}        }

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":" your request and create a"}        }

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":" research plan for smart hard"}         }

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":"wired smoke detector"}      }

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":" options for your home"}        }

event: content_block_delta
data: {"type":"content_block_delta","noindex":0,"delta":{"type":"text_delta","text":"."}           }

event: content_block_stop
data: {"type":"content_block_stop","noindex":0   }

event: message_delta
data: {"type":"message_delta","delta":{"stop_reason":"end_turn","stop_sequence":null},"usage":{"input_tokens":170,"cache_creation_input_tokens":0,"cache_read_input_tokens":0,"output_tokens":25}    }

event: message_stop
data: {"type":"message_stop"             }

(Note: No other requests/responses show up in the debug console.)

[nbonamy-kochava]

Strange. When I tried yesterday it indeed seemed to be broken but all my tests today are working:

• Sonnet 4.6
• Opus 4.6
• GPT 5.2
• Gemini 2.5 Flash
• Groq OpenAI OSS 120B

Are you still seeing any issue?

[natesilva]

Unfortunately the same thing is still happening for me. I tried:

• Gemini 3.1 Pro Preview (via the Google integration)
• Claude Opus 4.6 (via the Anthropic integration)
• gpt-oss-120b (via OpenRouter)

[blu3knight]

Did you try looking at the Logs?

What I do to see what is going on in detail is open the logs directory in /Users//Library/Logs/Witsy in terminal. Then tail -f main.log. Able to see all interesting details there.

[natesilva]

error-redacted.log

Thanks for pointing me to that log @blu3knight.

Attached is the log. Note this is the complete log, started from before I submitted the Deep Research prompt. I know it looks like it starts in the middle of the action.

Both Opus 4.6 and GLM-5 claim that the problem is in witsy/src/renderer/services/deepresearch.ts around line 50, and again around line 159: the ... is not valid JSON.

Now that ... is in a prompt, not literal JSON. You’d think an LLM would understand that as a placeholder. But Opus and GLM seem to think that the model was parsing ... literally.

[natesilva]

Also, I’ve tried a few other things to solve this:

• use a different internet connection (cellular vs. my home fiber)
• turn off my outbound firewall, Little Snitch
• re-generate my Brave Search API token (that’s the search engine I have enabled in Settings → Plugin → Web Search)

If no one else can repro this then I may try to install Witsy in a virtual machine with a clean OS. If I have time.

[nbonamy]

Was finally able to reproduce! Bug fix coming soon!

[natesilva]

I tested the fixed version and it works. Thanks!