ArtlexYoung/dsh-super-code

中文 · English

A smarter, leaner, faster, and easier-to-use coding plugin for DeepSeek Harness.

On SWE-bench Pro hard-100, compared with the official Standard preset (same host DSH 0.2.1-rc.2, same model, both on high reasoning effort and legacy budgets):

	Standard + high	Super Code
Passed	46/100	49/100
Net time per task	20.9 min	16.4 min (-21.3%)
Tokens per task	7.81 M	5.11 M (-34.6%)
Model calls per task	92	69 (-25.5%)

Tested on version 0.3.5. Full results and limitations

Install

Supports Harness from 0.1.5-alpha.1; see compatibility for tested Web, official Desktop, and NEXT versions.

Desktop: open Plugins → Add plugin in the left sidebar, enter dsh-super-code, then click Install and Enable now (on older versions, reload or restart when prompted). Desktop includes its own runtime; Node.js is not required.

Web / CLI: requires Node.js 22 or later.

npm install -g @deepseek-ai/dsh            # skip if dsh is already installed
dsh plugin --profile web add dsh-super-code

Download the .tgz file from GitHub Releases.

Desktop: enter the full path of the package in "Package name or address", then click Install.

Web: run the following command (use the actual file name):

dsh plugin --profile web add ./dsh-super-code-0.3.10.tgz

Windows commands

Bundles GNU Bash, the required MSYS DLLs, BusyBox, and GNU grep, so Git does not need to be installed just to run Bash commands. Developer tools such as Git, Node, and Python still need separate installation. macOS and Linux keep the host environment.

Under Settings → Super Code → Preset configuration → Command environment, bundled Bash defaults on and Also apply to official presets defaults off. Enabling the latter covers Standard, PTC, and Create; Minimal retains its persistent terminal. Switches are editable only on Windows hosts, and are gray with a hover tooltip on other platforms.

In confined environments, Bash commands require DSH approval for that command. See the usage guide for configuration and compatibility limits.

Usage

1. Open Settings → Agent Presets and select Super Code (identifier super-code).
2. Create a new session and describe the task.

Model, reasoning effort, permissions, and tools follow the Harness settings and need no separate configuration. Child Agents use the parent Agent's model by default.

If the preset is not listed, check its status in Settings → Super Code, or see troubleshooting.

Select Super Code in the right sidebar. It contains three sections:

• Tasks: goal, step progress, tool-call summaries, and changed files. Read-only.
• Collaborative execution: the Agent execution tree. Click a node to view details.
• Memory: view, add, edit, and delete project memory and global preferences.

The three sections open one at a time. Tasks opens by default; click the current heading to collapse it.

Screenshots use the current UI with sample data to illustrate the plugin.

Features

Feature	Description
Batched tool calls	Reads multiple files in parallel within one model call, reducing round trips
Layered memory	Current task, project lessons, and global preferences are stored separately, each with its source
Runtime budgets	Uses the better-performing legacy compression budgets by default; adjustable in Settings
Deletion protection	Directs Agents to use recoverable trash by default; can switch to strict confirmation
Windows commands	Bundled Bash and common utilities, optionally applied to official presets
MCP connection	Allows external MCP clients to delegate tasks to DSH. Off by default
Frontend Skill	Loaded on demand for UI work to check alignment, hit areas, and browser behavior

See the usage guide.

Recent releases

• 0.3.10: Added bundled Bash on Windows and an official-preset switch, with legacy-host compatibility and tooltip fixes.
• 0.3.9: Polished Settings and the task sidebar, improved MCP conversation control, fixed PTC child-tool approvals, and verified rc.4 Web compatibility.
• 0.3.8: Added MCP connection; external clients can delegate, inspect, continue, and cancel DSH conversations. Added a Home overview in Settings.

Full changelog

Versions follow major.performance.maintenance: a change in the second number indicates improved coding results or efficiency; the third covers fixes, UI, and compatibility. Every release requires extensive regression testing, so the release cycle may be long.

Evaluation

Plugin comparison

Full SWE-bench Pro hard-100 results, sorted by tasks passed. All runs use Harness 0.2.1-rc.2, deepseek-v4.1-flash-1017, high reasoning effort, and legacy budgets. Each group reuses its environment continuously across tasks.

Configuration	Passed/valid	Mean total tokens (M)	Mean time (min)	Mean model calls	Mean tool calls	Cache rate
Super Code 0.3.5	51/100	6.435	24.751	76.660	75.840	98.927%
Official Standard	50/100	8.030	17.198	91.990	113.880	99.167%
Unmodified dsh-ops 0.4.4 · Standard	49/100	7.099	23.132	82.500	102.360	99.062%
Official Standard + billion-context 0.1.209	45/100	6.359	18.471	108.950	131.140	97.654%
Unmodified dsh-ops 0.4.4 · PTC	44/100	7.322	23.695	81.490	80.960	99.056%

M means one million tokens, including reasoning tokens. Time includes some in-run service waits. These additional results are recorded separately from the opening comparison and the historical budget experiments below. Full results and methodology

Method

• Task set: SWE-bench Pro hard-100, graded with the official grader. Tuning uses only 33 hard tasks selected from the SWE-bench dev split; hard-100 is used only to evaluate frozen candidates and investigate issues, never for tuning, to avoid overfitting to the evaluation tasks.
• Baselines: the official Standard preset and third-party plugin configurations. In the historical official-preset comparison on the old host, Standard + high passed 51 tasks and Minimal + high passed 46, with Standard using 23.5% fewer tokens per task, so Standard was selected as the baseline.
• Controlled variables: comparisons on the same host keep model, reasoning effort, and budgets fixed while comparing preset and plugin combinations. Cross-host and cross-model experiments are identified separately.
• Metrics: tasks passed, mean total tokens, time, model and tool calls, and cache rate. The plugin comparison table above includes some in-run service waits; the historical budget experiments below use net time, excluding failed requests and retry waits.

Historical process

1. Old-host baseline (DSH 0.1.5-rc.2, deepseek-v4.1-flash, high): the Super Code 0.3.0 candidate passed 56 tasks, the official Standard preset 51, and Minimal 46. Relative to Minimal, time fell by 47.6% and tokens by 44.9%.
2. Regression after the host upgrade: after upgrading to DSH 0.2.1-rc.2 and deepseek-v4.1-flash-1017, under default budgets Standard dropped to 46 and Super Code 0.3.4 to 41, below Standard. Investigation found the new host had tightened context budgets: the compression threshold fell from 800,000 to 550,464 tokens, and retained verbatim context from 160,000 to 98,560 tokens.
3. Retest with legacy budgets: both sides used legacy budgets. Standard stayed at 46 with 14.2% more tokens; Super Code 0.3.5 rose to 49 with only 0.4% more tokens and 19.4% less net time. Based on this, the plugin uses legacy budgets by default from 0.3.6.
4. Cross-model check (glm-5.3-flash, max): Super Code 0.3.4 passed 45 tasks and native ZCode 29. Super Code used less net time per task (32.9 vs 52.3 min) and more tokens (5.50 M vs 3.61 M).

Historical summary

These are earlier evaluation runs and are not pooled with the plugin comparison results above.

Comparison	Passed	Efficiency
New host, legacy budgets: Super Code 0.3.5 vs Standard + high	49 vs 46	Net time -21.3%, tokens -34.6%, model calls -25.5%
New host, default budgets: Super Code 0.3.4 vs Standard + high	41 vs 46	Tokens -25.5%, net time +13.9%
Old host: Super Code 0.3.0 candidate vs Standard + high	56 vs 51	Raw time -34.9%, tokens -27.9%
Old host: Super Code 0.3.0 candidate vs Minimal + high	56 vs 46	Time -47.6%, tokens -44.9%
GLM-5.3-flash: Super Code 0.3.4 vs native ZCode	45 vs 29	Time -37.1%, tokens +52.4%

Limitations

• Each configuration was run once. The same task can pass in one run and fail in another; run-to-run variance for a fixed configuration has not been estimated.
• The default-budget and legacy-budget runs used different plugin versions (0.3.4 → 0.3.5), so the gain cannot be attributed entirely to budgets.
• 0.3.6 – 0.3.10 have not been re-evaluated on the 100 tasks.
• The cross-model comparison is between two complete systems and cannot isolate the plugin's contribution.

Results and methodology · All result tables

Acknowledgements

This project recognizes the LINUX DO community.

Project	Reference
DeepSeek Harness (MIT)	Host runtime: Agents, sessions, tools, PTC, Skills, UI extensions
ZCode (Apache-2.0)	UI design
Agent Skills	File format of the frontend Skill
Apple HIG, WCAG 2.2	Interaction and accessibility principles in the frontend Skill
SWE-bench, SWE-bench Pro (code MIT)	Tuning and evaluation tasks
js-yaml, Zod (MIT)	YAML parsing, data validation

Detailed acknowledgements and licenses · This project is licensed under Apache-2.0.