dsh-reasoning-picker

A composer button for DeepSeek Harness that answers one question before you hit send: how much model do I actually need for this draft?

Click the 📊 button next to the send action. The plugin reads your draft (plus the last few turns of the conversation), asks a cheap model to judge the difficulty, and shows a recommendation — Flash or Pro, plus the reasoning effort, plus a one-line reason. Nothing changes until you click 应用 (Apply), and it never sends the message for you.

The point is to stop paying for Pro · High on "rename this variable" while still getting Pro when the task deserves it.

Install

dsh plugin add dsh-reasoning-picker

Then restart DSH (or reload the page) so the client half is loaded. The button appears in the composer toolbar, to the right of the input box.

How it works

Step	What happens
1	You type a draft and click 📊
2	The host reads the draft and the last few messages of the session
3	A small model picks the cheapest level that fits, from a ladder built from the models you actually have
4	A small panel shows Flash · Off / Pro · Low … plus a one-line reason
5	You click 应用 — only then is that provider/model/effort applied to the current session

• Never auto-sends. The draft stays in the box; you still press send.
• Never auto-switches. Evaluating alone changes nothing.
• Uses your own model list. The ladder is built from what sessionController.modelCatalog() reports, so it adapts to whatever providers you configured instead of assuming fixed tiers.
• The evaluator runs at the lowest effort the chosen model supports, so the estimate itself costs almost nothing.

Configuration

The plugin works with zero configuration. Optional keys (in the bundle config):

Key	Default	Meaning
flashModel / flashProvider	auto-detect	Force the "cheap" tier
proModel / proProvider	auto-detect	Force the "strong" tier
maxRecentMessages	6	How many recent messages are fed to the evaluator
maxRecentChars	1400	Character budget for that recent context
maxDraftChars	4000	Draft is truncated to this length
timeoutMs	15000	Evaluator timeout

Privacy & safety

• Both HTTP routes are loopback-only and reject forwarded/origin-mismatched requests.
• The draft and recent context are sent to the model you already use through DSH's own llm service — no third-party endpoint.
• No telemetry, no network calls other than the model call you configured.

License

MIT