DSH Auto Preset Router
简体中文
A DeepSeek Harness profile bundle that adds an Auto Agent Preset. The first
direct user prompt is classified by the fixed
deepseek-official/deepseek-v4-flash model, then the untouched prompt runs
under one of DSH's four shipped presets with the model selected for that
session. After mounting the target preset, an isolated second V4 Flash call
selects likely necessary installed tool and Skill names from a bounded,
JSON-framed catalog; deterministic lexical ranking is the failure fallback.
Auto appends only those bounded names. It never loads a Skill body or changes
the target preset's tool prefix.
Routing policy
| Target | Selected for |
|---|
standard | Web/current information, mixed non-programming work, general assistance, full tool ecosystem, or ambiguity |
code | Any repository/source-code analysis or programming task, plus broad batch/parallel tool orchestration |
minimal | Demanding self-contained reasoning, mathematics, or algorithmic analysis with no repository or programming work |
cordis | Non-programming DSH preset/Cordis conceptual or runtime work |
The classifier uses no tools, reasoning effort off, temperature 0, and a
16-token output limit. On DSH rc.1 it receives the first prompt text plus safe
attachment metadata (name, dimensions, and media type), never image bytes.
A failed or ambiguous classification falls back to standard.
While Auto waits for the first prompt, it mounts only DSH's official filesystem
Skill catalog provider so Web and TUI can offer user-invocable /skill-name
entries. Auto itself still exposes no model-facing tools; after routing, the
selected preset owns Skill loading and prompt injection.
Install
DeepSeek Harness 0.1.1-rc.1 or a compatible later build is required.
Version 0.2.4 also supports DSH 0.1.3-alpha.2.
The first-turn check supports both legacy Session.events and the
snapshotEvents() API used by DSH 0.1.2-rc.1.
On releases that ship ptc instead of code, the classifier's code result
mounts ptc; the trace records rawOutput: "code" and finalPreset: "ptc"
without marking the rename as a fallback.
Legacy code selections also resolve to ptc when opening old sessions;
an explicitly installed code preset keeps precedence and menus are unchanged.
dsh plugin --profile web add github:yhfgyyf/dsh-auto-preset-router
dsh plugin --profile web add github:yhfgyyf/dsh-progressive-tools
dsh web
For a separately installed TUI profile that composes the official
agent-presets service, install the bundle into that profile too:
dsh plugin --profile tui add github:yhfgyyf/dsh-auto-preset-router
dsh plugin --profile tui add github:yhfgyyf/dsh-progressive-tools
dsh --profile tui
The bundle keeps the official roster service, contributes its package-owned
read-only preset through that roster, and sets the assembly default to auto.
It does not rewrite settings.yaml. An explicit user setting remains
authoritative; set this if needed:
agent-presets:
default: auto
After the first prompt is routed, DSH records auto-router/classified and then
agent-preset/selected before the first real turn. The classification event
is purely diagnostic and is persisted with the DSH ignorable envelope marker,
so a harness without this plugin can safely skip it. It contains
classifierProvider, classifierModel, rawOutput, finalPreset,
fallbackUsed, errorCode, latencyMs, capabilitySelection, toolHints,
and skillHints. The
session is then locked to the
selected preset by DSH's normal non-empty-session rule.
When at least one installed summary matches, a plugin-owned
<auto-capability-hints> message follows the untouched first user message. It
names at most five tools and three Skills. With dsh-progressive-tools, the
agent uses search_tools/describe_tools for deferred schemas; selected Skills
continue through DSH's official stable skill(name) loader.
Remove
dsh plugin --profile web remove dsh-auto-preset-router
dsh plugin --profile tui remove dsh-auto-preset-router
Existing sessions retain the preset recorded in their durable history. A local
$DSH_HOME/.agent-presets/auto directory, if present, is shadowed while this
bundle is installed and is never modified. Removing the bundle removes its
roster contribution on the next process start.
Privacy and failure behavior
The first prompt text and attachment metadata are sent to DeepSeek V4 Flash
even when the session uses a different model. Image bytes are never sent to
the classifier. The untouched multimodal prompt goes through DSH's native rc.1
attachment path after routing, so the session's selected model must genuinely
declare image input. DeepSeek's shipped Flash and Pro routes are text-only;
rc.1 adds the official deepseek-v4-flash-vision-exp image model. DSH refuses
the send and retains the draft until that model, or another image-capable
route, is selected. Prompts over 24,000 characters are reduced to their
beginning and end. Classifier failures route to standard.
Development
npm test
npm run check
npm pack --dry-run
With a configured TUI profile and DeepSeek credentials, the optional live gate
runs one fresh session for each target (and creates four short session logs):
npm run test:live:tui
The repository uses the current DSH plugin distribution contract:
package.json declares dsh.bundle.patch, and cordis.patch.yml composes the
package through dsh plugin --profile … add ….
