DeepSeek Harness Plugin

YUEYUEXYS/dsh-think-ultra

Stars ★ 2 Category Models & Providers Added 2026-09-20

Pins every request to the native max-effort tier and adds a per-model/per-mode reasoning-depth panel with a 0-20 capability slider, isolated Velocity/Apex profiles, and offline reasoning toolboxes. Distributed as an obfuscated build; commercial use allowed, reverse engineering closed.

Install

# from GitHub (first run asks for allowBuilds approval — follow the hint, retry)

dsh plugin --profile web add github:YUEYUEXYS/dsh-think-ultra

Any plugin you install runs third-party code with your own permissions — it can read your files, use your credentials, and reach the network, and tool approvals don’t sandbox it. GitHub-sourced plugins also run build scripts at install time — pnpm blocks those until you allow them, so an install can stop with ERR_PNPM_GIT_DEP_PREPARE_NOT_ALLOWED or ERR_PNPM_IGNORED_BUILDS; dsh prints the exact key to add under allowBuilds in your profile’s pnpm-workspace.yaml, and the install works on the next run. Allowing a build is a trust decision: only install sources you trust, and pin a commit (github:owner/repo#sha).

README

Think Ultra

The heavy, local reasoning layer for the official DeepSeek Harness (DSH).

Think Ultra is a native plugin for DeepSeek Harness that runs a real, test-time-compute reasoning stack around the model's native maximum effort — more samples, more judges, more verification, more reconciliation — entirely on your machine, with no network calls of its own. It turns a single maximum-effort response into a layered pipeline: multi-pass tournaments, executable verification, cross-review, a metacognitive core, and model-callable toolboxes.

The floor is welded to native max. Raising a slider widens the machinery around that max call; it never changes the word in the request.


What "Ultra" actually is

The DeepSeek API exposes four reasoning efforts: off, low, high, max. There is no fifth — send reasoning_effort="ultra" and the provider rejects it. Think Ultra wraps the registered provider adapter with a proxy, adds an Ultra row to the native dropdown, and maps ultra back to max on the way out, cached per (provider, model).

Ultra is not a magic word sent to the model. It is the machinery that runs around and above a max call. The gain is meant to come from test-time compute, not adjectives in a system prompt: every heavier tier issues real, additional, independent max calls through Harness's own llm surface, then reconciles them.


What's new in v0.1.0

  • Four fully isolated surfaces. State is kept per model × mode — Flash / Pro × Velocity (迅流) / Apex (极境) — each persisted independently. Sliders, toggles, mode, authorization and directives never bleed across surfaces. Velocity is the fast single-chain base; Apex is its strict over-clocking superset.
  • Global capability control (0–20). One number drives the whole reasoning ceiling. It exists only on the Ultra tier (greyed out and frozen on Max), remembers its value across launches, and the 19→20 step is a deliberate discontinuity that arms the full chain.
  • A 46-axis reasoning kernel. The raw surface is 46 independent engine axes with zero fake controls; day-to-day they fold into 7 / 10 / 8 master rods (Flash / Pro / Vision), with every axis micro-tunable in an advanced panel.
  • Metacognition core, L1–L5. L0 linear → L2 devil's advocate → L3 multi-jury → L4 frame deconstruction → L5 recursive self-reference, with non-linear budget amplification (×1 to ×4.2), plus red, black, lab and abyss switches gated per level.
  • Seven deterministic Rust engines. std-only, zero dependencies, zero network, JSON-RPC over stdio — including an 8-wave orchestrator, a probability field and a policy compiler. The plugin transparently falls back to a built-in JS engine if the native binary is missing.
  • OC over-clocking (Apex). Multiple candidates → 2-gram clustering → self-consistency voting → cross-verification → falsification elimination → confidence ranking → optimal synthesis.
  • L1–L5 reasoning stack. L1 tournament, L1.5 executable verification in an isolated local VM, L2 cross-review, L3 multi-paradigm forge, and a recursive reasoning tree.
  • Model-callable toolboxes. 30 professional tools across five boxes plus 21 layered reasoning tools, with a three-layer fuse for wall-clock, rounds and token budget.
  • Stability and memory. 10-hour+ endurance, layered milestones, context distillation, goal anchoring, three-tier memory with semantic retrieval, and an anti-loop hard gate.
  • Super Think, code and vision. A stable graded thinking phase, a seven-step code-understanding pipeline, and a seven-step multimodal vision pipeline (Flash can read images).
  • Hardened build. The delivered client is obfuscated and self-defending. On violation it responds non-destructively — it clears in-memory markers and writes a local revocation flag; it never deletes files, opens a socket, or fingerprints hardware.

Install

Requirements: Node.js ≥ 22.6 and a working DeepSeek Harness install. The published bundle already ships a built lib/ — there is no build step.

# install into the web profile straight from GitHub
dsh plugin --profile web add https://github.com/YUEYUEXYS/dsh-think-ultra/tarball/main

# restart the profile
dsh --profile web

Or install a downloaded tarball:

dsh plugin add ./dsh-think-ultra-0.1.0.tgz

After restart, open the reasoning-tier dropdown and select Ultra to open the control panel. First launch shows the welcome screen; your configuration is persisted in the DSH profile. If the native Rust helper is absent or blocked, the plugin falls back to the JS engine automatically.


Quick start

  1. Pick a model (DeepSeek V4.1 Flash — multimodal — or V4 Pro).
  2. Set the inference tier to Ultra.
  3. Open the panel and choose Velocity or Apex, and move the global capability slider.
  4. Read the price estimator before you send on the heavy tiers.
  5. Start the conversation — the enhancements engage automatically.

The model selects and calls the toolboxes itself on Ultra sessions; describe what you want in plain language.


Compatibility

  • Fully compatible with the latest DeepSeek Harness 0.1.6-alpha.2 (the 0.16-alpha line).
  • Forward-compatible within >=0.1.2-alpha.2 <0.2.0; future 0.1.x releases keep working without changes, and only the breaking 0.2 line requires an update.
  • Supported models: DeepSeek V4.1 Flash (multimodal) and DeepSeek V4 Pro.
  • Node >=22.6.

100% offline

The plugin never phones home. Every rule, template and prompt fragment is shipped locally and offline; the Rust core is std-only with no network APIs; a static source scan shows zero network calls. The only update check is an explicit, user-triggered read of the public release feed.


Make money with Think Ultra

Think Ultra is explicitly licensed for commercial use — the deal is simple: you keep 100% of the money, you never mention us, you just never give away the plugin itself.

  • Paid delivery, freelance work and agencies — charge for the deliverable.
  • Consulting and advisory — audits, due diligence, deep research.
  • SaaS or a hosted service — run it on the backend for your own customers, with subscriptions, usage, seats, retainers, credit packs or rush fees.
  • Team and company leverage — cut turnaround on the hard tasks nobody else wants.
  • Productised offers — "deep audit in 24h", "verified code review", "hard-problem sprint" as fixed-price products.
  • Content, courses and public comparisons built on top of it.

You sell the outcome while you run it. You do not hand over the delivered files, and you do not open, change, extract or re-host the build. There is no royalty, no attribution, no audit, and no cap on your upside. See LICENSE.


A note on cost

It is supposed to be expensive. Every judge, every re-solve and every fork of the reasoning tree is a genuine extra max call, so cost climbs steeply with the tiers — the devil's advocate, double recheck, best-of-n, dual jury, tree tournament and L3 forge are the heaviest. There is no "make it stronger but use fewer tokens" switch. Match the notch to the job and let the customer pay for the fire.

The floor is native max, and every heavier notch adds real, additional independent calls — that is how the pipeline is built, not a hope. This is not a benchmarked ranking or a guaranteed head-to-head score against any named model; whether it beats a particular system on your task is something only your own tests can settle.


License

Proprietary. Think Ultra Proprietary Commercial License (TUPA-CL v1.0). Commercial use is expressly permitted inside DeepSeek Harness; reverse engineering, de-obfuscation, file inspection, technical analysis, modification, redistribution and re-hosting are prohibited. See LICENSE.


Made for DeepSeek Harness. If Think Ultra helps, a star is appreciated.

Content from the project README on GitHub ↗

Links

More in this category

View the whole category →

Community comments

Comments are public GitHub Discussions. Loading them connects to GitHub and Giscus; a GitHub account is required to post.