dsh-anchored-standard
具备任务感知能力的两阶段 DeepSeek Harness 预设：首次请求使用 spec/react/weak 锚点（bash+read±edit/write），首次工具调用后使用完整的 Standard 工具目录。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add github:Jungod1121/dsh-anchored-standard#560530f0e1af6dba3528ac8962f5e5e63951b1bbAnchored Standard 以 dsh-anchored-standard 发布，当前版本为 1.2.1。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
点击下方 Badge 复制 Markdown，粘贴到 README 即可。
验证 package.json 声明的 GitHub 仓库，即可管理这个公开页面。认领后，Hub 会优先安排当前版本的安全扫描，并在通过后公开展示结果。
认领这个 Plugin →Anchor the first request. Unlock the full stack.
A two-phase agent preset for DeepSeek Harness: bootstrap DeepSeek V4 Pro with the Minimal-aligned prompt and two tools, then expose the complete Standard catalog after the first durable tool call or reply.
V4 Pro's ability ceiling is high — but which reasoning strategy it uses is
decided by what the first API request shows it. The community evaluation
at xiaobright/modeltest measured:
| Preset | First request sees | Full task sees | Project2 V4.1b (max) |
|---|---|---|---|
| Standard | 25 tools + long prompt | 25 tools | 91 / 92 |
| Minimal | 1-line prompt + 2 tools | 2 tools only | 99 / 96 — but too narrow for real work |
| Anchored Standard | 1-line prompt + 2 tools | all 25 tools | 98 / 99 |
The gain comes from anchoring the first-turn trajectory — not from keeping the tool surface small forever. This preset keeps Minimal's brain and gives it Standard's hands.
sequenceDiagram
participant U as User
participant A as Agent loop
participant M as DeepSeek V4 Pro
U->>A: first message (new session)
A->>M: request #1<br/>task-matched anchor persona<br/>tools: bash + read (± edit/write)
Note over M: reasoning chain anchors on the task-matched scaffold
M->>A: first tool/call
Note over A: promotion signal (durable session event)
A->>M: request #2<br/>same persona · all 25 Standard tools
Note over M: full-capability work continues on the anchored trajectory
The session's first user message picks the anchor:
| Task | Persona | First-request tools |
|---|---|---|
| spec — fix / maintain / debug | Minimal's exact prompt | bash + read + edit |
| react — build / create from zero | hands-on doer prompt | bash + read + write |
| weak — ambiguous | model self-picks (Pro: classify instruction) | bash + read |
glob/grep stay out of every bootstrap catalog — measured trajectory
boundary for V4 Pro; edit/write are anchor-safe.
On request #1 the persona is the ONLY prompt section and runtime contexts are cleared — the cleanest possible opening.
After the session records its first durable tool/call, every later
request sees the full Standard catalog; the chosen persona stays constant
and the remaining sections (plan-mode etc.) return. Nothing is injected
after turn one (measured: post-anchor guidance hurts Pro).
Mode and promotion state live in durable session events, so refresh and resume preserve them.
dsh plugin --profile web add github:Jungod1121/dsh-anchored-standard
Restart DeepSeek Harness. The bundle copies the preset into
$DSH_HOME/.agent-presets/anchored-standard/ (idempotent — local edits are
never overwritten), then create a blank session and select
Anchored Standard (experimental). To make it the default preset for new
sessions:
# $DSH_HOME/settings.yaml
agent-presets:
default: anchored-standard
dsh_home="${DSH_HOME:-$HOME/.dsh}"
mkdir -p "$dsh_home/.agent-presets"
cp -R preset "$dsh_home/.agent-presets/anchored-standard"
Export the session JSONL and inspect request/header events — the first
header contains only bash/read, every later header the full catalog:
seq 18 -> 2 tools ['bash', 'read'] # request #1
seq 137 -> 25 tools ['ask_user_question', 'bash', ...] # request #2 (promoted)
| You get | Trade-off | |
|---|---|---|
| vs Minimal | 23 more tools: glob/grep/edit/write/web_search/subagents/workflows/skills/goals/todo… | one extra reasoning round for the bootstrap |
| vs Standard | task-matched trajectory anchors (98/99 vs 91/92 in the reference eval) + a first turn that is cleaner than Standard's | the first-request catalog is decided by keyword classification — ambiguous tasks fall back to the weak anchor and let the model pick |
0.1.0-rc.6 at the wire level; the classifier and anchor tables are
unit-tested.xiaobright/modeltest Project2 V4.1b — n=2
on one frozen task. They are reproducible evidence for that task, not a
universal guarantee across models or workloads. The spec/react/weak anchor
design incorporates measured findings from
yjh051108/dsh-router-standard (P1-P24) and xiaobright's trajectory
boundary probes.reasoningEffort: max.dsh plugin --profile web remove dsh-anchored-standard
rm -rf "${DSH_HOME:-$HOME/.dsh}/.agent-presets/anchored-standard"
Developed and verified against DeepSeek Harness 0.1.0-rc.6. The harness is a
developer preview with breaking changes — review upstream changes before
upgrading.
This is a community project: not an official DeepSeek preset, not affiliated
with or endorsed by DeepSeek. Design inspired by
xiaobright/dsh-anchored-standard
(see NOTICE). MIT — LICENSE.