dsh-subagent-model-config
给 DeepSeek Harness（DSH）原生 Team 编队的队员单独配模型和推理档位——在设置页点两下就配好，不改官方任何文件。
Per-teammate model & reasoning-effort configuration for DSH native agent teams — two clicks in the settings UI, zero official files modified.
中文 | English
中文
这解决什么问题
DSH 原生 Team 编队的 spawn_teammate 只有 name / description / prompt / context 四个参数，没有 model（官方讨论 #6985）——所有队员只能继承主会话的默认模型：想给审查队员配个便宜快的、给写手配个强模型的，做不到。
本插件把这个缺口补上：按队员名配规则，每个队员跑在你指定的 provider / model / 推理档位上，全部在设置页点选完成。
- 零官方改动：不改官方文件、不打补丁；包装可逆，取消勾选「启用」即回到原生行为
- 热生效：保存规则后，下一次
spawn_teammate 即生效，无需重启
- 一张规则表管全部：队员名支持
* / ? 通配、大小写不敏感，按顺序求值、第一条命中即生效；每条规则可加中文备注（纯显示，不注入）
- 待指定安全：只写队员名、模型留空的规则不注入任何东西（队员继承默认模型），UI 黄条提醒——可以先建角色、后补模型
- 取证开关：
debug 开启后每次 spawn 追加一行到 .probe.jsonl 供排查，默认关闭
快速开始
安装
方式一（推荐）：DSH 设置 → 插件市场 → 搜索 dsh-subagent-model-config → 安装并启用。
方式二（手动）：克隆本仓库，在 profile 的 package.json 里加：
{ "dsh-subagent-model-config": "link:/path/to/dsh-subagent-model-config" }
并把它加入 profile 的 dsh.bundles，重启 DSH。
配置
打开 设置 → Subagent Models，添加规则：
| 字段 | 说明 |
|---|
| 队员名匹配 | glob 模式，如 verifier*；* 任意串、? 单字符 |
| 备注 | 中文名，只用于显示 |
| provider | 留空 = 继承主会话的供应商 |
| model | 留空 = 该 provider 的默认模型；整条规则 provider/model 全空 = 不注入 |
| 推理档位 | 下拉选择，候选 = 所选模型自己声明的档位；留空 = 目标模型默认档 |
设置页全貌——左侧入口、启用开关、常用角色一键条与规则卡：
设置页：入口、开关、常用角色与规则卡
典型配置（节选）：
{
"enabled": true,
"rules": [
{ "match": "verifier*", "note": "独立验证员——便宜快的", "provider": "…", "model": "…", "reasoningEffort": "low" },
{ "match": "writer*", "note": "写手——强模型高思考", "provider": "…", "model": "…", "reasoningEffort": "high" },
{ "match": "researcher*", "note": "检索员——角色先建，模型后补" }
]
}
通用工程编队的实际用法——验证员用强模型高思考、测试编队用 flash 省额度，一张规则表各跑各的：
通用工程编队规则示例
工作原理
插件在运行时包装两个 DSH 内部接口：agentTeams.spawnTeammate（采集队员名）与 SubagentRuntime.startContinuable（注入点）。派单发生时按队员名匹配规则表，把命中的 provider / model / reasoningEffort 写进官方原生支持的 SubagentStartRequest.agentOptions——只接上编队路径没接的线，不改任何官方行为。包装装在实例与原型两层（幂等防重入），内部异常一律只记录不抛出；配置为空或关闭时，行为与未装本插件完全一致。
兼容性与测试
- 声明范围：
engines.dsh >= 0.1.7-alpha.1，Node >= 22
- 实测：DSH 0.1.7-alpha.2（首发实测）；DSH 0.2.0 桌面端（日常在用）
npm test：143 用例（17 套件）全过- 本插件钩在 DSH 内部运行时接口上，官方未来版本可能改变这些接口导致失效；遇到问题欢迎提 issue
视频预设配套：队员闸门与能力限制
这是 video 解析预设的配套约束。判定只看父会话 preset——不用视频预设的用户完全无感。
① 视频队员名硬闸门
视频解析任务用哪个模型、开不开思考，必须人工点名。名字以 video- / video_ 开头却未命中任何规则 ⇒ 抛错拒绝启动（错误码 DSH_SUBAGENT_MODEL_CONFIG_UNROUTED_VIDEO），报错实时列出当前规则表允许的命名清单。fail-closed：配置缺失、损坏或规则为空时，视频名一律拒绝。非视频名字与 enabled: false 完全不受影响。
② 队员侧能力限制（父会话 preset 为 video 时注入）
| 注入项 | 作用 |
|---|
request.persona | 队员角色卡，遮蔽部署 persona（主窗身份卡对队员不再生效） |
request.toolFilter.allow | 工具白名单（fail-closed）：watch_video / read write edit glob grep / pwsh / job_list job_output job_kill / skill / todo_write |
白名单只列「继承层」工具；团队通信工具（send_message / team_task_* 等）注册在成员自己的 scope，天然保留。standard / cordis / minimal 模式的子代理完全不受影响。
English
The problem it solves
In DSH's native agent teams, spawn_teammate takes only name / description / prompt / context — no model parameter (upstream discussion #6985). Every teammate inherits the main session's default model: you cannot give the reviewer a cheap fast model or the writer a strong one.
This plugin fills that gap: name-based rules, each teammate runs on the provider / model / reasoning effort you pick — all editable in the settings UI, effective on the next spawn.
- Zero official changes: no official files modified; the wrap is reversible — untick "enabled" and you are back to stock behavior
- Hot reload: saved rules apply to the next
spawn_teammate, no restart needed
- One rule table decides who runs on what: glob matching (
* / ?, case-insensitive), first match wins; per-rule display notes
- Safe "model pending" rules: a rule with an empty model injects nothing (teammate inherits the default); the UI flags it in amber — create roles first, fill models later
- Evidence switch: with
debug: true each spawn appends a line to .probe.jsonl; off by default
Quick start
Install: DSH Settings → Plugins → search dsh-subagent-model-config. Manual: clone this repo, link it in your profile's package.json, add it to dsh.bundles, restart DSH.
Configure: Settings → Subagent Models — add rules with a name pattern, optional provider / model / reasoning effort (dropdown candidates come from the selected model itself).
How it works
The plugin wraps two DSH internals at runtime: agentTeams.spawnTeammate (captures the teammate name) and SubagentRuntime.startContinuable (injection point). On spawn it matches the name against the rule table and writes the resolved fields into SubagentStartRequest.agentOptions — a field DSH natively supports but the team path never wired up. Wraps are installed on both instance and prototype (idempotent); internal errors are logged, never thrown; with an empty or disabled config the behavior is identical to not having the plugin.
Compatibility & tests
engines.dsh >= 0.1.7-alpha.1, Node >= 22- Tested on DSH 0.1.7-alpha.2 and the 0.2.0 desktop app (daily driver)
npm test: 143 cases, 17 suites, all passing- The plugin hooks DSH internal runtime APIs that may change in future releases — please open an issue if it breaks
Video-preset companion: gate & teammate restrictions
For the video analysis preset only — detection reads the parent session's preset, so other modes are untouched:
- Hard gate on video teammate names: a name starting with
video-/video_ that matches no rule is rejected outright (error code DSH_SUBAGENT_MODEL_CONFIG_UNROUTED_VIDEO) with a live list of allowed names. Fail-closed: missing, corrupt, or empty config rejects every video name.
- Teammate-side restrictions (injected when the parent preset is
video): request.persona (masks the deployment persona) and request.toolFilter.allow (fail-closed whitelist: watch_video, read/write/edit/glob/grep, pwsh, job_list/job_output/job_kill, skill, todo_write). Team-communication tools live on the member's own scope and are naturally preserved. Standard / cordis / minimal subagents are unaffected.
License
MIT
