Session titles silently fail to generate on reasoning (thinking) models — no config way to disable reasoning for the auxiliary title call

[Marker689]

Problem

Session titles stop being generated entirely when the default/selected model is a reasoning (thinking) model. The auxiliary call that derives a session name fails silently — the name never appears, with no user-facing error.

Root cause

Session titles are produced by @deepseek-ai/dsh-session-title-first-prompt-llm → @deepseek-ai/dsh-session-title-llm. That provider calls the LLM with:

ctx.llm.stream({
  provider, model, messages, system,
  maxTokens: config.maxOutputTokens,   // default 64
  sessionId, purpose: "session-title", signal,
})

Two problems compound:

1. No reasoning control. The call passes no reasoningEffort and the config schema of this provider accepts no reasoning/thinking option — it only has provider/model overrides. So the title request always falls back to the model's default reasoning level, which for a reasoning model like deepseek-v4-flash is "think".

2. Reasoning is folded into the same output budget. For the underlying adapter (pi-ai at openai-completions), reasoning tokens count against maxTokens. With the shipped default maxOutputTokens: 64, the model's thinking for even a trivial title prompt consumes the whole 64-token budget, finishes with max-tokens, and emits no title text → empty title → the title provider throws and no name is ever shown.

There is no config-only way to turn reasoning off for this one auxiliary call while keeping it enabled inside real session turns.

Expected behavior

It should be possible to generate a session title on a reasoning model without it silently failing:

• Either let reasoning be controlled for auxiliary calls (e.g. a reasoning/reasoningEffort-style option on the session-title-llm provider so an operator can set it to off, keeping the main session on its own reasoning setting), or
• Keep auxiliary generation from being starved by thinking (e.g. a separate reasoning budget, or a non-trivial default maxOutputTokens for reasoning models), or
• Fall back to a deterministic title when the LLM title attempt produces nothing.

Workaround (configuration only)

Raising the generation budget on the session-title-llm row unblocks it because reasoning is counted against maxTokens:

# cordis.patch.yml
- id: session-title-llm
  config:
    targetWords: 5
    targetCjkCharacters: 10
    maxInputBytes: 4096
    maxOutputTokens: 1024
    timeoutMs: 60000

This is a workaround, not a fix — it doesn't give operators a way to choose non-thinking title generation, and a model that thinks heavily could still hit the cap.

[denial123789]

I traced this against rc.8 and documented an operator runbook: https://sandbaseai.github.io/deepseek-harness-handbook/session-title-reasoning-budget.html

One useful source-level correction to the visible symptom: session-title calls ensureFallback() before the optional LLM provider. When the auxiliary request finishes at max-tokens, automatic generation catches the failure and normally leaves the deterministic first-prompt fallback in place. The strongest proof is therefore:

1. latest session/title remains source.kind: fallback;
2. session/title-llm-request records the affected route and maxTokens: 64;
3. the Host warning contains title output reached maxOutputTokens;
4. after a bounded cap increase, a fresh Session records the new cap, finishes with stop, and appends a provider-sourced title.

The guide also separates timeout, empty-text, tool-call, and UI-projection branches, and notes that an explicit lightweight title route changes the prompt data/privacy boundary. The 1024-token patch is presented only as a bounded workaround, not a universal default or a replacement for title-specific reasoning control.

[setieroth]

+1 — confirmed on DSH Desktop 2.0.2 (macOS) with deepseek-v4-flash-vision-exp sessions (the sessions themselves run at reasoningEffort: high). The host log carries exactly the predicted symptom, repeatedly:

[W] [session-title-service] session "<id>": automatic title generation failed:
    Error: session-title-llm: title output reached maxOutputTokens

Four occurrences on Aug 24 alone (10:46 / 10:52 / 12:16 / 13:45 — four different sessions). Affected sessions keep the first-prompt-snippet fallback title in the sidebar (truncated first sentence as the "title"), which matches the ensureFallback() behavior documented above.

A reasoningEffort: "off" / minimal knob for this one auxiliary call (or a separate reasoning budget so thinking can't eat the 64-token cap) would fix it.

(AI-assisted comment; log lines verified against local host logs.)

[weijiafu14]

你的根因分析我读下来是对的，尤其是"推理 token 计进同一个 maxTokens 预算"这条——64 的默认值对一个会思考的模型来说，光 thinking 就吃光了。在官方加 reasoningEffort 选项之前，有一条纯配置的路，而且它正好利用了你已经指出的那个 provider/model 覆盖。

两步

1. 先把预算调大（最确定、最省事的一步）

maxOutputTokens: 64 是这件事的直接触发条件。把它调到几百，thinking 吃掉一部分之后仍有余量吐出标题。这一步不依赖下面第 2 步，可以单独先试。

2. 给标题调用单独指一条"不思考"的路由

llm-pi-ai 里可以为同一个端点、同一个模型声明第二条路由，只是把推理档位声明成 off：

llm-pi-ai:
  providers:
    my-provider-notitle:          # 单独一条，只给标题用
      displayName: Title helper
      api: openai-completions
      baseURL: <和主路由同一个端点>
      apiKeyEnv: <同一个凭据引用>
      models:
        - id: <同一个 model id>
          name: Title helper
          contextWindow: 131072
          reasoningEfforts:
            off:                  # 只声明 off

然后把 session-title-llm 的 provider / model 覆盖指向这一条。主会话仍然用原来那条带完整推理档位的路由，推理只在真实回合里开着——正是你"Expected behavior"里写的形状。

reasoningEfforts 是官方 profile 字段，off:（键在、值为空）这个写法我在自己的例子里用过并且进了发版回归，所以字段形状是可靠的。

必须说清的边界

• "只声明 off 是否等于该路由默认不推理"这个具体组合，我没有专门试过。 我验过的是这个字段的写法本身。如果试完标题还是空，第 1 步（调大预算）是那条更确定的。
• 这是绕过，不是修复。你要的"给辅助调用一个 reasoningEffort 选项"仍然缺，而且我认为该由官方补——用两条路由指同一个模型来表达"这次别思考"，属于配置层的变通，不是契约。
• 静默失败那部分我们更帮不上：标题生成失败没有任何用户可见的错误，这是 session-title-llm 自己的错误处理，我们不修 DSH 自家的包。你提的"至少要报错"我觉得比 reasoning 选项还重要。

利益相关：我维护 pi2dsh，上面全是 DSH 官方 llm-pi-ai 配置，不需要装任何插件。