Description

Session auto-title generation silently fails when using reasoning-capable LLMs (e.g., DeepSeek V4 Flash, Claude with thinking, QwQ, etc.). The resulting session titles remain NULL with no user-facing error or warning.

Root Cause

In agent/title_generator.py, the generate_title() function calls call_llm() with max_tokens=500. Reasoning models output a separate reasoning_content block before the actual response content. With only 500 tokens budgeted, these models frequently exhaust the entire budget on reasoning alone, producing finish_reason='length' and an empty content string. The function strips the empty string and returns None.

# Current upstream (line ~60):
response = call_llm(
    task="title_generation",
    messages=messages,
    max_tokens=500,        # <-- too small for reasoning models
    temperature=0.3,
    timeout=timeout,
    main_runtime=main_runtime,
)

Symptoms

• hermes sessions list shows sessions with empty titles
• SQLite query shows many rows with title IS NULL
• Agent log shows repeated INFO agent.auxiliary_client: Auxiliary title_generation: using auto … entries
• No Title generation failed warning in logs (the function returns None without raising when content is empty)

Steps to Reproduce

1. Configure a reasoning-capable model (e.g., DeepSeek V4 Flash, Claude Sonnet with thinking enabled)
2. Start a new session and send a message
3. Wait for the auto-title generation to complete
4. Check the session title via /title or hermes sessions list → it will be empty

Fix

Increase max_tokens from 500 to at least 2000 in agent/title_generator.py:

max_tokens=2000,

This provides enough headroom for the reasoning phase while still leaving tokens for the actual title output. The title itself is only 3–7 words, so 2000 tokens is generous but lightweight enough to not meaningfully impact cost or latency.

Environment

• Hermes Agent version: current main
• Model affected: DeepSeek V4 Flash (and any other model that outputs reasoning_content)
• Provider: OpenRouter / opencode-go (but provider-agnostic)

Additional Context

This is a silent failure because generate_title only logs at WARNING on exceptions, but when finish_reason='length' with empty content, no exception is raised — the code simply strips the empty string and returns None silently.

The same max_tokens issue could affect other auxiliary tasks (compression, web_extract, session_search) when used with reasoning models.