dsh-multi-model-orchestrator
用于在 DeepSeek Harness 中配置多模型子代理编排的 Cordis agent 插件
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:ToxicantX/dsh-multi-model-orchestrator#0a38e7159dd50e9c4fc035c97b9bcdfc11c073d3Multi Model Orchestrator 以 dsh-multi-model-orchestrator 发布，当前版本为 0.7.11。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。
在 DeepSeek Harness 中配置并运行由不同模型驱动的专业 Agent 团队。
DSH Multi-model Orchestrator 为 DeepSeek Harness Web 设置页增加了 Agent 编排功能。你可以配置最多 3 个可复用的专业 Agent，为每个 Agent 选择 DSH 中已有的模型，并设置清晰的职责。这些 Agent 是可复用的 specialist 工具；child 任务的数量由有效工作和 runtime 容量决定，而不是由 Agent 列表数量决定。
Session 使用 Multi-model orchestrator 预设后，每个已配置的 Agent 都会成为独立的子 Agent 工具，供主 Agent 按任务需要进行委派。
主 Agent 的定位是产品负责人和工程项目经理，而不是默认开发者。面对非简单工作，它先明确目标与验收标准，拆分互不重叠的范围，并在实施开始前委派开发、调查、测试和审查。每个 child 在返回前独占一个内聚任务和验收目标；主 Agent 只协调不重叠工作、等待依赖、集成返回结果并执行最终验收。真正的一步小改仍可直接完成。
将插件安装到 DSH Web profile：
dsh plugin --profile web add -w github:ToxicantX/dsh-multi-model-orchestrator
安装完成后重启 DSH Web，并刷新浏览器。Host 启动时，插件会自动预置并维护 Agent 预设，同时创建旧版 orchestrator 兼容 ID，使使用该 ID 的已有 Session 可以恢复，并从 Web 选择列表中隐藏使用官方显示名的兼容项。内容完全匹配官方旧版的无 marker preset 会被安全收编；经过自定义的用户 preset 不会被覆盖，改用不同名称时仍保持可见。
升级时，已有的超过 3 个 Agent 的配置会被完整保留。运行时先启用前 3 个，设置页继续显示完整列表，并要求在下次保存前缩减到 3 个以内。
如果通过 DSH 源码仓库运行，请将命令中的 dsh 替换为 pnpm dsh。
如果启动时提示受管预设已被修改或与现有预设冲突，并且你确认要丢弃这些本地改动，请显式修复：
dsh plugin --profile web exec dsh-orchestrator-install --force
修改 Agent 列表或模型分配后，请创建新的 Session。已经运行的 Session 会继续使用启动时的 Agent 配置。
在非简单 Session 中，主 Agent 会先确定验收标准并分配开发任务，再进入实施。它把已配置的 Agent 视为可复用的 specialist 工具，按复杂度将工作拆分为任意数量有意义、互不重叠且可独立验收的任务，并避免人为制造或过度微拆。每个任务都创建新的 child，即使它匹配已经用于其他任务的 specialist。独立任务可在 runtime 容量内并发；文件、所有权或依赖交叉的任务必须串行。存在合适 specialist 时，主 Agent 不自行承担开发，但不会为了占满 Agent 人为制造任务。每个 child 在返回前独占一个内聚任务和一个验收目标；主 Agent 可以协调其他明确不重叠的任务，但必须等待而不能并行实现相同目标。同一任务需要修正时，复用原 continuable child。specialist 返回后，主 Agent 负责检查集成边界并执行最终验收。
可靠性行为：没有独立工作剩余时，主 Agent 使用前台子调用，不反复轮询 list_agents 来消磨时间。running 表示 child 正在执行，必须等待；idle 或 ready 只表示当前没有活动轮次，不代表完成或卡死。只有 child 明确报告阻塞/死锁、依赖不可用、同一工具或执行错误连续两次，或两次状态检查都没有进展且存在未解决阻塞时，才可判定卡死。耗时、静默、不变状态、其他 child 已完成或主 Agent 准备结束，都不能单独构成卡死证据。只有用户明确取消、明确期限已经到达、有具体证据支持的真实死锁，或验证过的重复工具或执行失败才允许中断；中断前必须说明证据，只调用一次 interrupt_agent，随后用 list_agents 验证。interrupt_agent 只停止当前轮次、保留可继续 child，因此中断后显示 idle 或 ready 是预期行为，不应重复中断。主 Agent 不得在必需 child 仍运行或没有返回结果时结束；被中断或失败的 child 对应任务仍未完成，必须报告阻塞或重新规划。主 Agent 绝不能为了催要提前报告、缩短等待、重新取得控制权、开始集成或避免等待而中断，也不能仅根据运行时长或没有新消息推断死锁。send_message 只排队下一轮，不会改道当前工作。specialist 连续两次遇到相同工具或执行协议错误后，改用最简单的有效替代工具调用或报告阻塞，并避免继续发起失败或没有有效输出的重复调用。
| 配置项 | 必填 | 说明 |
|---|---|---|
| Agent ID | 是 | 用于子 Agent 工具名称和逐 Agent 运行配置的稳定身份，例如 architect 或 reviewer。 |
| 模型 | 是 | 从 DSH 模型目录中选择的 Provider 和 Model。 |
| 开发职责 | 否 | 展示给主 Agent 的简短路由摘要，用于选择 specialist；该字段不用于确定身份。 |
| 子 Agent 提示词 | 否 | 发送给对应子 Agent 的独立 persona 提示词。旧配置未填写该字段时，会回退使用开发职责。 |
| 推理等级 | 否 | 所选模型实际提供的推理等级之一；省略时使用模型默认值。 |
| 最大输出 Token | 否 | 限制 Agent 输出长度的正整数。 |
git clone https://github.com/ToxicantX/dsh-multi-model-orchestrator.git
cd dsh-multi-model-orchestrator
pnpm install
pnpm bundle
pnpm typecheck
pnpm test
发布管理遵循语义化版本（SemVer）。请在 CHANGELOG.md 记录面向用户的变更，使用 vX.Y.Z 格式创建 tag，并运行本地预检：
pnpm release:check v0.7.0
推送匹配的 vX.Y.Z tag 会触发验证并创建带自动生成说明的 GitHub Release。npm 发布不会自动执行。
将本地仓库安装到 DSH 源码 profile：
pnpm dsh plugin --profile web add -w D:/path/to/dsh-multi-model-orchestrator
DSH Multi-model Orchestrator adds an Agent orchestration page to the DSH Web settings. You can configure up to 3 reusable specialist Agents, assign an existing DSH model to each one, and give each Agent a clear responsibility. These configured Agents are reusable specialist tools; the number of child tasks is determined by meaningful work and runtime capacity, not by the roster size.
When a session uses the Multi-model orchestrator preset, every configured Agent becomes an independent subagent tool that the primary Agent can delegate work to.
The primary Agent acts as the product owner and engineering manager rather than the default developer. For non-trivial work, it defines the outcome and acceptance criteria, decomposes non-overlapping scopes, and delegates development, investigation, testing, and review before implementation begins. Each child exclusively owns one cohesive task and acceptance target until it settles; the primary coordinates only non-overlapping work, waits on dependencies, integrates returned work, and performs final acceptance. Truly small one-step changes can stay local.
Install the plugin in the DSH Web profile:
dsh plugin --profile web add -w github:ToxicantX/dsh-multi-model-orchestrator
Restart DSH Web after installation and refresh the browser. The plugin provisions and maintains its Agent preset automatically when the Host starts. It also provisions the legacy orchestrator preset ID so existing sessions created with that ID can resume while hiding official-name compatibility entries from Web selection lists. Exact official pre-marker copies are adopted safely; customized user-managed presets are never overwritten and remain visible when given a distinct name.
Existing settings with more than 3 Agents are preserved during an upgrade. The first 3 remain active, the settings page continues to show the complete roster, and the next save requires reducing it to 3 or fewer.
When running DSH from a source checkout, use pnpm dsh instead of dsh in the command.
If startup reports that the managed preset was edited or conflicts with an existing preset, and you intend to discard those local changes, repair it explicitly:
dsh plugin --profile web exec dsh-orchestrator-install --force
Create a new session after changing the Agent roster or model assignments. Sessions that are already running keep their original Agent configuration.
During a non-trivial session, the primary Agent establishes acceptance criteria and assigns development tasks before editing. It treats the configured Agents as reusable specialist tools, decomposes work by complexity into any number of meaningful, non-overlapping tasks with independent acceptance targets, and avoids artificial or overly granular splits. Each task gets a new child, even when it matches a specialist already used for another task. Independent tasks run concurrently within runtime capacity; tasks that overlap files, ownership, or dependencies run serially. A suitable specialist is not kept idle while the primary performs development, but the primary does not invent work merely to use every Agent. Each child owns one cohesive task and one acceptance target until it returns; the primary may coordinate other clearly disjoint tasks but waits instead of implementing the same outcome. For corrections to the same task, it reuses the continuable child. After specialists return, the primary reviews integration boundaries and runs the final acceptance checks.
Reliability behavior: when no independent work remains, the primary uses foreground child calls and does not repeatedly poll list_agents just to pass time. running means that a child is executing and must be allowed to continue; idle or ready means that no turn is active, not that the task is complete or deadlocked. A child is stuck only when it reports a blocker or deadlock, a required dependency is unavailable, the same tool or execution failure occurs twice, or two status checks show no progress together with an unresolved blocker. Elapsed time, silence, unchanged status, another child finishing, or the primary being ready to finish never qualifies by itself. Interrupts are limited to direct user cancellation, an explicit deadline that has arrived, a deadlock supported by concrete evidence, or verified repeated tool or execution failure. Before interrupting, the primary states the evidence, targets only the direct child, calls interrupt_agent once, and then checks list_agents; interruption stops the current turn but preserves the resumable child, so idle or ready afterward is expected. The primary never finalizes while a required child is running or has no returned result, and treats an interrupted or failed child as incomplete work requiring a blocker report or re-planned attempt. The primary never interrupts to request an early report, shorten a wait, regain control, begin integration, or avoid waiting, and never infers a deadlock from duration or lack of messages. send_message queues the next turn and does not redirect current work. After two occurrences of the same tool or execution-protocol error, a specialist switches to the simplest valid alternative tool call or reports the blocker and avoids repeated calls that fail or produce no useful output.
| Field | Required | Description |
|---|---|---|
| Agent ID | Yes | Stable identity used for the subagent tool name and per-Agent runtime settings, such as architect or reviewer. |
| Model | Yes | Provider and model selected from the DSH model catalog. |
| Development scope | No | Short routing summary shown to the primary Agent when it chooses a specialist. It does not determine Agent identity. |
| Child Agent prompt | No | Dedicated persona prompt sent to the selected child Agent. Legacy configurations without this field fall back to the development scope. |
| Reasoning effort | No | One of the exact effort levels advertised by the selected model; omission uses the model default. |
| Maximum output tokens | No | Positive integer limiting the Agent's generated output. |
git clone https://github.com/ToxicantX/dsh-multi-model-orchestrator.git
cd dsh-multi-model-orchestrator
pnpm install
pnpm bundle
pnpm typecheck
pnpm test
Release management follows Semantic Versioning (SemVer). Record user-facing changes in CHANGELOG.md, create tags as vX.Y.Z, and run the local preflight:
pnpm release:check v0.7.0
Pushing a matching vX.Y.Z tag triggers verification and a GitHub Release with generated notes. npm publishing is not automatic.
Install a local checkout into a DSH source profile:
pnpm dsh plugin --profile web add -w D:/path/to/dsh-multi-model-orchestrator
MIT