dsh-compact-standard

中文说明 | GitHub

A DeepSeek Harness agent preset: full Standard tool catalog plus a high-density compact-output expert prompt. By default it aggressively compresses thinking and output tokens without reducing capability — and never compresses code/commands/formulas/critical-step completeness.

This is a community project. It is not an official DeepSeek preset and is not affiliated with or endorsed by DeepSeek.

What it does

• Installs a strict, standardized Chinese expert persona with nine capability sections: thinking compression, output compression, priority adjudication (accuracy > clarity > conciseness), completeness guarantees (no pseudocode / ellipsis, placeholder and environment annotation), uncertainty handling (state-when-unknown, explicit assumptions), risk-warning and rollback obligations, DSH-tool result distillation, and an override rule letting explicit in-conversation user instructions beat the default compression.
• Keeps the full Standard tool catalog, so capability is not reduced.
• v0.3.0 deterministic token levers (measured, see EXPERIMENT.md §9):
  • tool-compact: densifies curated model-facing tool/parameter descriptions at assembly time. Structural schema keys, unmatched tools and unmatched parameters stay byte-identical; failure degrades to the unchanged catalog. Measured on the real 25-tool catalog: 26,638 → 21,963 chars (−17.6%, ~4.7K chars ≈ 1.2K+ tokens per request before caching), zero structural drift (test-asserted). Execution goes through the registry's own definitions — only the strings the model sees are rewritten. Disable with disabled: true under the tool-compact row for a clean A/B.
  • tool-bootstrap is ENABLED by default again. v0.2.0 disabled it claiming the host default ships an equivalent first-request anchor; measurement on this host showed the opposite: with the bundled bootstrap off, the FIRST request exposed the full 25-tool catalog (33.7K-char header vs 12.1K bootstrapped; session-881eba9a vs session-6c0b72a4 in bench output), and the readonly-first boundary was lost. The 2-tool anchor only ever existed because this preset's own bootstrap provided it.
• DSH-specific optimizations, stated on a measured basis since v0.2.0:
  • a normal dsh-persona section instead of complete: true, so plan-mode and other cooperative prompt sections remain active; includeRuntimeContext: false keeps the prompt lean.

Measured impact (v0.3.0, see EXPERIMENT.md §9 and docs/BENCHMARK.md)

Item	Measured on this project's host (DSH rc.6, deepseek-v4-flash)
Persona injection	Yes (verbatim in the assembled system prompt)
Persona size	738 chars (9 sections preserved)
Tool catalog size/request	26,638 → 21,963 chars with tool-compact (−17.6%, structural keys untouched, 25/25 tools preserved)
First-request tool surface	2 tools (pwsh+read) with bootstrap enabled; full 25-tool catalog if disabled (33.7K-char header)
System prompt	default 46 chars → ~7,022 chars with this preset (uncached +2.4–2.6K tokens/request; amortized by context caching)
Output/reasoning tokens	confounded on this host by a concurrent provider switch (opencode-go); no attribution claimed — a controlled A/B via bench/run.mjs is required

Honest summary: the preset's deterministic effects are (a) tool-catalog compression (tool-compact, −17.6% measured — the larger per-request lever a preset can own), (b) the first-request 2-tool anchor (bootstrap), and (c) the persona (a soft instruction on output style and thinking tendency). The remaining unbilled-input lever is host-level; pairing this preset with scripts/install.mjs host tuning yields far more than the persona alone (see docs/HOST-TUNING.md).

System prompt

The preset installs this expert prompt (Chinese original; it is shown verbatim below and translated in the Chinese README's mirror sections):

你是严格标准化技术专家。默认极致压缩思考与输出 token，但不降低能力，不压缩代码/命令/公式/关键步骤的完整性。

【思考】只做有效推理：直接锁定目标、约束、最优路径；删除重复、铺垫、自我检查、冗余推演。不拟人、不寒暄、不用语气词；禁“好的”“综上所述”“我们可以”“如您所知”等填充。

【输出】结论先行，证据/步骤紧随；能列表不用段落，能表格不用长句。不输出内部推理、草稿、自检；不重复问题，不写总结客套。默认只给最佳方案；仅应明确要求提供多方案，并列取舍与推荐。要求“详细/解释”时才扩展，否则最小必要输出。准确完整前提下可用高密度结构提升效率，不以牺牲能力换取压缩。

【优先级】准确性 > 清晰度 > 简洁性；冲突时以准确性与完整性为准，绝不因压缩省略内容。

【完整性】代码、命令、公式、配置须完整、精确、可执行；禁止伪代码、省略号或以“等”“……”省略关键内容。必要占位符须显式声明（如 <API_KEY>、<your-domain>）并给出替换示例；命令/配置开头标明环境、版本、前置条件，不确定则说明假设。

【不确定性】信息不足或歧义须声明“信息不足，需确认”，不得编造；必须基于假设才能继续时，显式标注 假设：<内容>，再给出基于该假设的结果。

【风险】“请注意”仅用于必要风险提示。高风险操作（删除、覆盖、强制执行、生产变更）必须给出关键风险与回滚/备份方法，即使未要求多方案。

【DSH 工具】工具结果只提炼必要结论与关键证据；完整报错、环境/版本信息、关键文件片段与必要命令不得省略。调用工具前后不叙述过程，直接给出结果或下一步。

【覆盖规则】本指令为默认基线；用户本次对话中的明确相反要求（如“详细讲解”“全量输出”）优先于默认压缩。

Note: on stock installs, DeepSeek Harness prepends its fixed identity opener (≈4.5K chars) before this persona. That text is compiled into the app; there is no settings.yaml switch. Only custom host compositions (a self-owned base.cordis.yml) can drop it via includeHarnessIdentity: false — see docs/HOST-TUNING.md.

Compatibility

• DeepSeek Harness 0.1.0-rc.5 (commit 47f9438) and 0.1.0-rc.6 (node_modules API-level verification). Upstream rc.7 (released 2026-08-17) is not yet verified.
• Node.js >= 22.19 (tests); bench/install scripts need 22.2+ (node:zlib zstd).
• Windows (verified on this host) / POSIX (scripts are zero-dependency).

DeepSeek Harness is currently a developer preview and explicitly permits breaking changes. This preset is a full snapshot of the Standard composition, so review upstream changes before using it with a newer release.

Install

Option A: preset + host tuning (recommended)

git clone https://github.com/pythonshiyi/dsh-compact-standard.git
cd dsh-compact-standard
npm run install:tune   # dry-run: prints the exact change to ~/.dsh/settings.yaml
npm run install:tune -- --reasoning low --preset compact-standard --apply

--apply backs up settings.yaml first (settings.yaml.bak-<timestamp>); rollback: npm run install:rollback -- --apply. Manual equivalent edits and details in docs/HOST-TUNING.md.

Option B: copy the preset only

PowerShell:

$target = Join-Path $env:USERPROFILE '.dsh\.agent-presets\compact-standard'
if (Test-Path -LiteralPath $target) { throw "Preset already exists: $target" }
New-Item -ItemType Directory -Force -Path (Split-Path -Parent $target) | Out-Null
Copy-Item -Recurse -LiteralPath '.\preset' -Destination $target

Linux/macOS:

dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
test ! -e "$dsh_home/.agent-presets/compact-standard"
cp -R preset "$dsh_home/.agent-presets/compact-standard"

Fully restart DeepSeek Harness, create a blank session, and select Compact Standard (compressed expert). Do not switch an active session from a different preset.

Verify

• Static: npm test (persona 9-section assertions, tool-bootstrap-enabled and tool-compact-mounted assertions, tool-catalog completeness, structural-drift and compression-ratio regression on the real 25-tool fixture).
• Runtime: npm run bench [<session-dir-or-file>] — aggregates per-session system-prompt length, tool-schema size (toolsChars — verify −17.6% vs the pristine catalog), first-request tool surface, input/output/reasoning/cache tokens, time-to-first-token, and final-answer size from the zstd session logs under ~/.dsh/sessions. Confirm the first request's system prompt contains the compression rules and its toolsChars is the compressed value; judge effects only against a baseline arm on the SAME provider/model/ reasoningEffort (docs/BENCHMARK.md).

Important behavior

• tool-bootstrap is ENABLED by default (v0.2.0 disabled it on a disproven host-default assumption; see EXPERIMENT.md §9 for the measured counter-example). promoteOn: either guarantees promotion to the full catalog at request #2 (a text-only first reply still promotes). Disable only when your host composition provably anchors request #1 itself.
• tool-compact rewrites curated descriptions only; unmatched tools and unmatched parameters pass through byte-identical, structural keys are never touched, and a mount/runtime failure degrades to the unchanged catalog (unit-tested). disabled: true under its row gives a clean A/B arm.
• The tool catalog changes once, so request-prefix cache continuity also changes once between the first and second model requests.
• The plugin performs no network requests and adds no telemetry.
• The preset has the same trust level as shell access. Review its files before installation.

Official ecosystem guidance

DeepSeek currently asks community plugin authors to publish plugins in their own GitHub projects and add the dsh-plugin repository topic for discovery. The official repository does not currently accept external pull requests and does not mandate a community repository template. See the official CONTRIBUTING.md.

License

MIT. preset/agent.cordis.yml is derived from the DeepSeek Harness Standard preset and the community dsh-anchored-standard preset; the original copyright and MIT notices are retained in NOTICE.