Conversation title generator emits refusal text as the title #34

Observation

From conv_30f049cdb75d464f on ws_mat, the title metadata event for an otherwise-successful conversation ended up as:

"I appreciate your request, but I need to clarify that I'm Claude, an AI assistant made by Anthropic. I don't have the"

(truncated — character limit)

This is clearly a refusal response from whichever model was asked to generate the title, not a title. The user's actual request ("Create a deal titled 'Smoke test' at $10k in qualified stage...") completed successfully with proper tool calls; only the title derivation misfired.

Cause

The title-generator prompt is getting something it interprets as "what are you?" rather than "summarize this conversation in a short title." One likely culprit: the prompt includes system text or tool-schema fragments that the model reads as a direct identity question. Another possibility: the transcript passed to the title generator starts with a boundary token that reads like a fresh conversation opener.

Impact

Low on correctness. The conversation itself works. But:

• Title UI renders an AI refusal as the conversation's identity
• Listing views look broken
• If the title is used anywhere in search/indexing, those tokens pollute retrieval

Suggested

1. Review the title-generator system prompt: ensure it's unambiguously "produce a short title for the following conversation" with clear bounding of the transcript content.
2. Validate the output: if the title starts with "I appreciate" / "I cannot" / "I need to clarify" / similar refusal patterns, treat as a bad title and either retry with a simpler prompt or fall back to "Untitled" + first user message excerpt.
3. Truncate at sentence boundary, not character count. Current truncation mid-word ("I don't have the") is a clear symptom of no post-processing.

Source

conv_30f049cdb75d464f on ws_mat — the metadata.title event at line 36 of the JSONL log.