dsh-coding-discipline
Manifest validCoding agent preset plus an embedded coding-discipline skill for DeepSeek Harness: verify-first edits, minimal diffs, real test runs and evidence-based reports.
dsh-coding-discipline
Coding-discipline enhancement for DeepSeek Harness (dsh).
This plugin ships two pieces that work together:
- A
Codingagent preset (id: coding) — the shippedstandardpreset's plugin list with the same tools and a strengthened persona: verify-first edits, minimal diffs, real test runs, no test gaming, spec-only verification, and evidence-based reports. No new tools, no permission changes, so it is comparable with the baseline and safe to switch back. - An embedded
coding-disciplineskill — the full procedure behind the preset contract, registered at host level so any agent in the profile can discover and load it on demand. Project/user skills with the same name still win by rank.
Install
dsh plugin --profile web add dsh-coding-discipline
From GitHub instead of npm:
dsh plugin --profile web add github:JinBeiCN/dsh-coding-discipline
Install from the release tarball if neither route is available:
dsh plugin --profile web add https://github.com/JinBeiCN/dsh-coding-discipline/releases/download/v1.0.0/dsh-coding-discipline-1.0.0.tgz
Then use it:
- New sessions: choose Coding in the agent-preset selector. To make it the default for new sessions, set it in Settings → General (Agent preset).
- Any session: the
coding-disciplineskill is listed in the skill catalog; load it before coding work. Existing sessions are not changed by installing the bundle.
What the contract covers
The persona and skill state the same rules in two layers:
- read the task, acceptance criteria, tests and relevant code before editing; reproduce the failure and name the
root cause as
file:linefirst; - make the smallest change, keep public interfaces, no unrelated refactors or formatting churn;
- never edit, delete, weaken, skip or xfail tests, and never hardcode expected values;
- ground every API, import and parameter in the repository or installed dependencies before using it;
- run the repository's real verification command after the last edit and iterate until it passes;
- verify only from the repository's own spec/tests/code — do not speculate about hidden tests, graders or benchmarks, and do not read parent/sibling directories or git history for hints;
- for specs with more than ~15 numbered acceptance clauses, keep a clause → implementation → verification checklist; for performance/concurrency work, report measured evidence once instead of re-running stress loops;
- end with evidence: files changed, root cause, exact commands with observed results, unverified parts, risks.
Measured behaviour (plain statement)
The preset was evaluated against an uninjected baseline on a self-built 16-task coding benchmark with
deepseek-flash (reasoningEffort=max), 18 isolated agent runs per arm:
- solved rate was identical (15/16 vs 15/16; the single failure was one ambiguous hidden-test assertion that both arms failed in one seed and passed in another);
- injected instructions were observed in 18/18 enhanced runs and the skill was loaded in 18/18; no run tampered with tests or read outside its working directory;
- token usage and wall-clock differences were not statistically significant — run-to-run variance dominated.
So this plugin is a guardrail, not a capability booster: it changes how work is done and reported, and it matters most for weaker models, lower reasoning effort, long specs, or review-heavy delivery. It does not make an already-saturated model produce more correct code in our benchmark.
Compatibility
dsh>=0.2.1-rc.2, Node.js>=18.- Web and Desktop profiles with the agent-preset registry (the shipped
standardpreset must be present, which it is in the defaultwebprofile). Headless/ACP profiles have no preset registry and do not load the preset part; the embedded skill part is registered wherever the bundle's host row is mounted.
Size and token cost
- Persona: 325 words, charged per request only in sessions that select the
Codingpreset. - Embedded skill: 7,600 characters of body (7,923 including frontmatter), below the standard preset's 8,000-character skill budget; it is only charged when a session loads the skill.
- No tools are added or removed.
Uninstall
dsh plugin --profile web remove dsh-coding-discipline
Disabling the bundle (or removing it) leaves existing sessions untouched until they are re-created.
Development
node --input-type=module -e "import('./index.js').then(m => console.log(m.name, m.inject))"
index.js exports a plain Cordis plugin (name, inject, apply) that registers the skill from
skills/coding-discipline/SKILL.md; cordis.patch.yml inserts that host row plus the Coding preset.
No build step and no runtime dependencies.
License
MIT — see LICENSE.
Compatibility
Versions
| Latest version | Published | Size |
|---|---|---|
| 1.0.0 | — | — |
Comments
Loading…
From the same category
DeepSeek Harness plugin for Reactive Resume: bridges your resumes and job applications into a Harness session over MCP.
★ 41.7k
↓ 146/wk
MIT
Aug 24, 2026
dsh plugin --profile web add dsh-plugin-reactive-resumeby dsh-market
The plugin market inside DeepSeek Harness — browse, search, one-click install · DSH 可视化插件市场
★ 10.3k
↓ 136.6k/wk
MIT
TypeScript
Nov 2, 2026
dsh plugin --profile web add dshmarketby Tencent
Let AI agents use your real, logged-in browser without interrupting your work. CLI + extension for browser automation across any shell-capable AI agent.
★ 9.9k
↓ 7.4k/wk
MIT
TypeScript
Nov 2, 2026
dsh plugin --profile terminal add @wxg-prc-cpg/browser-skill-dsh-pluginby superdesigndev
OpenRouter for agent tools. Join community here: https://discord.gg/6mQYYfFMAn
★ 7.1k
NOASSERTION
Python
Nov 3, 2026
dsh plugin --profile web add treg-dshby yjh051108
dsh-routing-suite — injector + router-standard kit: install the runtime injector first, then the task-aware reasoning-mode router preset (measured P1-P23).
★ 7k
MIT
JavaScript
Sep 18, 2026
dsh plugin --profile web add @dsh-external/dsh-super-injectorby Q00
Agent OS: the agent gets smarter on its own. We just hold the line: Interview-gated, staged evaluation, budgeted evolution loop. MCP server, 14 runtimes: Claude Code, Codex CLI, Gemini CLI, OpenCode,
★ 6.3k
MIT
Python
Nov 1, 2026