---

Features

Names the conversation, and keeps up with it. The first human message gets an AI title; after that, every new message re-reads the recent ones and refreshes it, so the title follows a topic that has moved on. Turn "Refresh title as the conversation grows" off and it goes back to naming the first message once. A manual rename pins the title either way, so it is never rewritten under your hands.

Suggests titles when you rename. The AI Name button sits at the top of the conversation page. Its dialog looks and behaves like DSH's native rename window, with three extra things:

Streaming candidates	Suggestions arrive one at a time (NDJSON), each revealed by a 500 ms left-to-right sweep
One-click apply	Click a chip to fill the input; Regenerate for a fresh set
Native feel	Enter confirms, Esc closes, and Enter during IME composition never submits by mistake

Uses a model you choose. Naming follows the conversation's main model by default, or pins a cheaper one. The dropdown is built from the host's remote.session.modelCatalog(), which enumerates providers one by one and drops those that return nothing — so every entry is a model the host has actually authenticated and can really call.

Reasoning, handled for you. A title needs three or four words, but a reasoning model spends the budget thinking first — the same max_tokens the answer draws from. So four things happen automatically:

Off when it can be	The model's own switch-off level is sent when it declares one (named off on some gateways, none on others; only declared levels are ever trusted)
Lowest level when it can't	Otherwise the lowest effort the model itself declares. Some gateways silently ignore an undeclared level instead of erroring, so a hand-typed level just makes the UI lie
A bigger cap for reasoning models	Thinking that eats the budget returns an empty title; measured on one reasoning model, thinking ran 325–3172 tokens, so reasoning models get 4096 (the schema ceiling). The cap only bounds generation and returns at stop, so it does not slow anything down — the timeout is the real guard
One retry when thinking ate it	Retried only on a max-tokens cut, at most twice, each attempt getting its own full timeout; other failures repeat identically so they are not retried

The in-dialog settings panel states which of the four applies: reasoning switched off / fell back to the lowest level / not a reasoning model so no effort is sent / this model's capability could not be read. Switching model refreshes that line and the level list immediately.

Leaves the audit trail. The provider id dsh-auto-rename is written into source.provider of every session/title event, so you can always tell which titles this plugin produced.

Why it builds its own dialog instead of extending the native one

The native rename window is an internal host component with no extension point: RenameForm declares no injectable slot, the requestSessionRename it needs is passed only as a prop into sidebar menu items, and nothing in the bundle ctx.provides it. The plugin can therefore neither insert a suggestion row into it nor open it programmatically. It builds its own dialog and copies the native window's CSS rules one by one, so the two are visually indistinguishable.

Install

dsh plugin add dsh-auto-rename

From a checkout instead:

dsh plugin --profile desktop add <path-to-this-repo>

The package declares dsh.bundle.patch, so installing it also inserts dsh-auto-rename into the profile's bundle stack and disables the built-in session-title-llm row from dsh-base. That second part is required: SessionTitleService.register() accepts exactly one provider globally, so if the built-in row registers first, this plugin's registration throws.

Restart DSH afterwards. Client-side changes hot-reload; host-side changes need the restart.

Settings

Open the AI Name dialog and click the gear in its title row, left of the close button. The panel expands inside the same dialog. Edits are staged in a draft and written only when you press Save; an empty field means "use the default".

Field	Default	Notes
Refresh title as the conversation grows	on	On: every new message re-reads the recent ones and refreshes. Off: the first message is named once. Suggestions keep working either way
Naming model	Follow main model	Lists only models the host has authenticated and can currently list
Reasoning level	auto	The levels offered are the ones this model declares; auto = switch it off if possible, otherwise the lowest
Suggestions	3	Candidates fetched per dialog open (1–5)
Output limit (tokens)	256	Per-generation cap (8–4096); raised to 4096 automatically when the model thinks
Timeout (ms)	60000	Longest wait for one generated title (1000–600000). A retry gets its own full budget, so the worst case is twice this
Title width (CJK characters)	12	The only length knob (1–400); longer output is truncated

If the host has no remote assembled, or listing fails, the model row falls back to manual Provider / Model fields. They must be filled in pairs: a half-set route logs a warning at startup and errors clearly when used, rather than silently doing nothing.

How it works

It reads only what you typed. Neither path looks at assistant replies: automatic naming takes the six most recent human messages (only the first one is ever scheduled when the refresh toggle is off), suggestions take the three most recent ones, so both describe the topic in progress. Each message is clipped to maxInputBytes (default 4096) before it reaches the model, because the official shared implementation throws on oversized input.

Length is measured in display width, not characters. One CJK character costs 2 half-width units, everything else — spaces included — costs 1. The budget is Title width × 2, so the default 12 allows 12 Chinese characters or 24 Latin characters including spaces, roughly 3–4 words. One knob governs both scripts; there is no separate word count for English, because word lengths vary too much for a count to mean anything.

The official shared implementation still wants a targetWords value (prompt-only, never truncating), so it is derived from the same budget at ~6 units per word — 12 yields 4. That keeps the model's target close to the hard cut instead of writing long and being trimmed.

Truncation avoids mangling words. A cut landing mid-word falls back to the previous space, so you never get sessio. When there is no space to fall back to — a provider/model route, a URL — the title is left alone: slightly too wide beats a broken word.

Why the plugin has to truncate at all

The official caller passes Number.MAX_SAFE_INTEGER as the truncation argument of normalizeSessionTitle — that is, no truncation whatsoever. The host does enforce a maxTitleBytes ceiling (80 bytes ≈ 26 CJK characters), but it is far too loose to keep titles short. So "not too long" is this plugin's job, and the practical ceiling of the width knob sits below the host's byte limit.

Project layout

lib/config.js    Pure logic: config merge, source-message picking, title truncation.
                 Zero imports, testable on its own.
lib/index.js     Host half — registers the title provider and the HTTP bridge.
lib/client.js    Client half — header button, rename dialog, in-dialog settings panel.
cordis.patch.yml Bundle patch: insert this plugin, disable built-in session-title-llm.
locale/          Display metadata for the plugin manager (title, description).
test/            Self-checks: logic, host apply, client render, locale audit, patch.

The two halves talk over three HTTP routes, each accepting loopback origins only (127.0.0.1, ::1, ::ffff:127.0.0.1) and rejecting sec-fetch-site: cross-site:

Route	Purpose
POST /dsh-auto-rename/suggest	Stream N candidate titles as NDJSON
POST /dsh-auto-rename/config	Echo the currently effective config plus that model's reasoning capability (including whether reasoning was actually switched off), for debugging; with a provider/model pair it only probes that model and writes nothing
POST /dsh-auto-rename/settings	Write config through the host settings service, whitelist-filtered

The write route answers 503 when the host settings service is absent, rather than reporting success it did not achieve.

Development

npm run check

Six gates, all fast and all offline:

1. Syntax check of the host and config modules, plus a parse check of the client bundle.
2. Pure-logic self-test — config merge, message picking, the width budget and the byte budget.
3. Host apply() smoke against a fake context: metadata, degradation paths, provider shape, route guards, and the exact Config schema field list. The draft route on /config is probe-only and never mutates the effective config.
4. Client apply() smoke: renders the header entry and the settings panel for real, and covers all three catalog degradations. The fake context reproduces the host's restricted facade, including the fact that reading ctx.inject throws. The four reasoning states, the re-probe after switching model, and "only the host decides whether the switch-off level is off or none" are pinned down here too.
5. Locale audit: every dictionary key used in code exists in both zh and en.
6. Patch check: cordis.patch.yml disables the built-in row, inserts this one, and touches nothing else.

The smokes need @deepseek-ai/schemastery. Locally, junction it from an installed profile:

New-Item -ItemType Junction -Path node_modules\@deepseek-ai\schemastery `
  -Target "$env:USERPROFILE\.dsh\profiles\desktop\node_modules\@deepseek-ai\schemastery"

test/stubs/ holds a dsh-session-title-llm stand-in that echoes the arguments it receives, so the call wiring can be verified outside DSH without reimplementing the official provider.

Known limits

• Slot renames fail silently. The header button depends on conversation.session.header.actions, declared in DSH 0.2.1-rc.2. When a slot is not declared, the slots.inject callback never fires and nothing throws — so a host upgrade that renames it makes the button vanish without a trace. The version badge above records the host this was last verified against.
• Settings are deliberately not on the host settings page. Both available slots were tried (plugins.bundle.config and plugins.item): registration succeeds inside an instrumented stand-in, renders nothing in the real UI, and reports no error. The panel is therefore drawn inside the dialog and reads/writes over the HTTP bridge.
• Prompts are the official ones. Automatic naming and suggestions share a single LLM call path, delegated to the shared implementation in @deepseek-ai/dsh-session-title-llm, so prompts, timeouts and normalisation match the built-in provider exactly. Only the length cap is this plugin's addition.
• Suggestions are not logged. They do not write a session/title-llm-request event: that event is the one-shot audit record for automatic naming, and suggestions can be regenerated without limit, so logging them would bloat the session log.
• The 4096 floor is a measured calibration, not a theory. It was raised step by step from one and the same model repeatedly eating smaller budgets dry; a model that thinks even more verbosely needs a retest. Also, "the model declares a switch-off level" only means that level is sent to it — some gateways map it elsewhere and keep thinking anyway, and that case is covered by the retry, not by the judgement here.

License

MIT