R-LEI2536/dsh-more-agent-presets

中文文档

Multiple selectable Agent Presets for DeepSeek Harness.

The presets are adapted for DeepSeek Harness (DSH).

Benefits

This plugin is particularly useful for:

Non-DeepSeek or Older Models:

• Improves performance for models that haven't been specifically adapted for DSH
• Enhances older models that lack DSH-specific adaptations

Personal Preference:

• Some users may prefer more interactive, discussion-based coding assistance over independent work

Why It Helps:

• DSH's default prompts are optimized for DeepSeek models
• Non-DeepSeek or older models may not perform optimally with default prompts
• These presets provide alternative interaction patterns suitable for different models

Available Presets

Qwen Code Coding Mode (qwencode-coding-agent)

A professional coding assistant emphasizing code standards and project conventions, using iterative workflows and CLI-friendly interaction style.

IFlow Coding Mode (iflow-coding-agent)

An interactive CLI agent with dynamic environment awareness, automatic Git context injection, and structured task workflows. Features auto-detected platform information, security-first permissions handling, and CLI-optimized communication style.

IFlow Creation Mode (iflow-cre-agent)

An enhanced mode for creating custom Agent presets. Includes all standard capabilities plus runtime inspection, plugin experimentation, and preset authoring guidance.

Features:

• DSH context information (Web UI URL, source root)
• Git repository awareness
• Custom prompt sections
• Skills for Cordis plugin development and composition editing

Current Limitation: ⚠️ Cordis Tool (@deepseek-ai/dsh-tool-cordis) is currently unavailable due to upstream issues. Dynamic Cordis plugin creation will be available after upstream fixes.

Pair Coding Mode (pair-coding-agent)

A coding assistant that works alongside the user as a pair programmer: aligns direction before each move, discusses before non-trivial edits, and waits for confirmation rather than driving the task to completion on its own.

Codex Coding Mode (codex-coding-agent)

A port of the open-source Codex CLI agent prompt: it states what it is about to do before each tool call, reads the codebase before editing, and drives the task to completion on its own, following Codex's planning, validation, and final-answer formatting rules. Tool names are mapped to this harness's tools (apply_patch → edit/write, update_plan → todo_write, request_user_input → ask_user_question), and Codex's plan-mode protocol is preserved through exit_plan_mode. The three Codex input[] blocks — <permissions instructions>, <collaboration_mode>, and <environment_context> — are injected as user-role messages at the head of every admitted step via three DSH agent/pre-step plugins (codex-permissions.mjs, codex-collab-mode.mjs, codex-environment.mjs), with the live sandbox / approval policy and cwd / platform / shell / date resolved from the DSH runtime. Unlike the other presets here, this one keeps Codex's autonomous posture rather than a discussion-first one.

General Chat Mode (chat-agent)

A pure-conversation assistant: it mounts no tool row, masks every inherited tool through a preset-plane ctx.tools.restrict({ allow: [] }), and refuses execution of anything registered in the agent's own scope through a ctx.tools.guard(...) — so it cannot act: no reading or writing files, no commands, no web search, no spawning a teammate. Two boundaries are deliberate and documented (presets/chat-agent/no-tools.mjs has the mechanism with source pointers). First, a preset-plane restriction covers the inherited surface only — the global layer plus every ancestor scope — while an agent's own registrations are exempt; a profile that mounts such a plugin, e.g. @deepseek-ai/dsh-experimental-agent-team-profile (whose nine Team tools go into agent.ctx for every root agent), therefore still lists them in that session's tool catalog, and the guard is what makes their call return a refusal instead of running. Second, that refusal is not the unknown tool a filtered tool produces. The system prompt switches with the current model (Default / Codex GPT-5 · GPT-6 / Qwen / DeepSeek-R1 / Gemini 3.7 · 3.8 texts; the verbatim table lives in presets/shared/model-personas.json), and the picked text is the session's entire prompt. Automatic (context pressure / overflow) compaction and the human /compact command stay on, because a chat session is long by nature. Use it when you want a conversation partner rather than an agent that acts on your machine.

General Chat Mode with Delegation (chat-subagent)

A chat assistant whose whole capability is one tool that acts: subagent — plus the read-only list_subagent_models the model-selection opt-in registers, which only reports the routes the session may name. The top-level session's tool catalog holds exactly those two names: a depth-0 catalog mask — ctx.tools.restrict({ allow: [] }) on the Agent's own scope, installed by presets/chat-subagent/delegator.mjs — hides the preset's inherited rows, while a restriction never filters the agent's own layer, where the opt-in registers both delegation tools. Hiding is not enforcement: one standing guard — ctx.tools.guard(...), installed once on the preset's standing plane by the same plugin — refuses every call outside the two names. The mask sits on the Agent scope, so a delegated child never inherits it (the child's scope parent is the preset's standing scope); the guard is read from the calling agent's chain, so the same one install covers the root session and every delegated child, and the delegation depth decides which names pass. The mask deliberately installs one macrotask after agent/created (and after a switch), and lifts the same way on leave: restrict() notifies tools/change where guard() does not, and the official delegation plugin installs into a live Agent on exactly that event, so an agent-scope restriction installed synchronously from agent/created gave one scope two installs, the second threw tool "list_subagent_models" is already registered in this scope, its install() aborted before mount(), and the session was left with no delegation tool at all (1.7.3). The session therefore cannot act directly; it calls subagent once and answers the user from the report. (The one residual is the own-plane Team-tool boundary noted below.) The single call is foreground (enableRunInBackground: false removes the background parameter from the schema entirely) on the spawn provider, with maxDepth: 1 stating the recursion budget.

The delegated subagent is what actually works, and it works read-only. A delegated child joins its parent's preset, so it inherits the same rows — and the same guard, whose depth branch is what bounds it: the child may call the declaration's subagentTools list — by default read, glob, grep, read_image, list_directory, todo_write, job_list — and nothing else. There is no bash, write, or edit in that list: the subagent investigates and reports, and cannot run commands, run tests, or change files (TODO.md tracks the writable follow-up). Two of those names are deployment-dependent — read_image exists only while an attachment store is mounted, and list_directory comes from a third-party profile plugin — and nothing probes for them any more: the guard judges by the declared list, so a name no row registered is simply absent from the catalog. The mask is on the parent Agent's scope, never on the preset plane, for exactly this reason: a preset-plane restriction would be on the child's chain too (it joins the same preset) and a restriction can only narrow, so the child could not undo it.

The system prompt switches with the current model. It is the shared table (presets/shared/model-personas.json, read through presets/shared/model-persona.mjs, which chat-agent mounts as its own prompt) plus a delegation note (presets/chat-subagent/delegation-note.md, read verbatim when the row mounts), and together they are the session's entire prompt. Automatic (pressure / overflow) compaction, the human /compact command, and tool-result pruning stay on: the child greps repositories, so its tool results need pruning even though the top-level session produces none.

Changing the subagent's tool list means overriding the declaration by its Loader row id in a profile patch. config is replaced wholesale, so every row must be restated; the list is the guard's allow-list for the child, so a name no row registers is simply not in the catalog:

# ~/.dsh/profiles/<profile>/cordis.patch.yml (abridged: restate every row)
- id: preset-chat-subagent
  name: '@deepseek-ai/dsh-agent-preset'
  config:
    id: chat-subagent
    order: 16
    name: "通用聊天模式 · 可派活"
    description: "顶层目录只有 subagent 与 list_subagent_models（可为子 agent 指定模型）：把任务交给只读子 agent 去查、去读，自己专注对话。"
    plugins:
      - id: delegator
        name: 'dsh-more-agent-presets/presets/chat-subagent/delegator.mjs'
        config:
          presetId: chat-subagent
          subagentTools: [read, glob, grep, todo_write]
      # …every remaining row of the shipped declaration, unchanged…

Widening that list does not rewrite the delegation note: the note is the file presets/chat-subagent/delegation-note.md, read verbatim when the row mounts, and it describes the shipped read-only default — revise the file when the child's abilities change, or override it for one profile with the delegator row's delegationNote. Nothing reads the tool list out of the note: it says in prose what the subagent can and cannot do.

One boundary is deliberate and stays: a profile that mounts the agent-team bundle still lists its nine Team tools for the top-level session, because an agent's own registrations are exempt from any restriction — the depth-0 mask cannot hide them, and the guard refuses their call instead of hiding them. The preset's own read/write/todo_*/job_* rows, by contrast, are hidden from that catalog by the mask and refused by the guard if a call is attempted anyway. And job_list is inert for the subagent — it cannot start a job — but it is part of the default list so that the writable follow-up needs no second edit.

Subagent Model Selection (qwencode / iflow / iflow-cre / pair / codex)

The five coding presets and chat-subagent opt into the Host's subagent model-selection setting with modelSelectionSettings: true on their delegation row — the same one-key opt-in the shipped standard/ptc/cordis presets carry. The top-level Agent then gets two things:

• list_subagent_models reports the routes this session may use. No arguments list the authorized providers, provider lists that provider's advertised models, and provider + model adds that model's reasoning efforts.
• subagent gains the optional provider, model, and reasoning_effort parameters. Omit all three and the child still inherits the parent session's model — the official default, and exactly what these presets did before the opt-in.

The authorized list is the Host's @deepseek-ai/dsh-tool-subagent/model-selection-settings row, i.e. the "Allow Agent to select a model for subagents" panel in the Web UI. An explicit route outside that list fails with a message listing the available ones. The list is sampled when a session is created, so it only affects new sessions. With the setting off (or its list empty) the behaviour falls back to inheritance without an error.

That row must exist in the profile. The shipped dsh-base + dsh-web-app stack mounts it by default, and so do the profiles in this repository; without it a preset that opts in fails to mount as a whole, with tool-subagent: \modelSelectionSettings` requires @deepseek-ai/dsh-tool-subagent/model-selection-settings in the Host scope`.

chat-subagent opts in too, so a chat session can name the child's model like the coding presets do. Its top-level catalog is not empty — the two delegation names are the ONLY entries in it — and exactly one of them acts, which is why its invariant reads "exactly one tool that ACTS". The flag also moves the delegation tool from the preset's standing plane into each agent's own layer: that own layer is precisely what a restriction never filters, which is why the depth-0 mask can be allow: [] without stripping the delegation itself (see the preset section above). chat-agent mounts no tool row at all, so it has nothing to opt in.

Which model a child actually ran on is a terminal line. Every delegation prints one, from presets/shared/subagent-route-log.mjs — mounted by these five presets — or from chat-subagent's own delegator:

[I] pair-coding-agent/subagent-route-log: subagent 3f1c… route opencode-chat/deepseek-v4.1-flash
[I] chat-subagent/delegator: subagent 3f1c… route opencode-go/deepseek-v4-flash

The line names the route the child was created with, which is the route the call asked for or the one it inherited from the parent session; unknown appears only when the harness stated neither half. Top-level sessions print nothing.

Two upstream warnings can appear around a short-lived child and are cosmetic: agent-registry … agent/disposed listener threw: TypeError: Cannot read properties of undefined (reading 'catch') (the tool chains .catch onto a dispose() that may return undefined — the per-Agent install path the opt-in activates) and session-projection-cache … write for "…" failed (cache stays stale) (a fail-soft race when a fast child's handle closes before a throttled checkpoint). TODO.md records the sources; the child's cached session record is in fact complete.

Design Philosophy

These presets differ from the default DSH prompt in their approach to user interaction and planning:

Interaction Style:

• Default DSH: Works independently with minimal user interaction
• Qwen/IFlow/Pair: Actively discusses with users, maintaining communication throughout

Plan Mode:

• Default DSH: Static approval process — AI produces a complete plan document, then waits for user approval
• Qwen/IFlow/Pair: Dynamic collaboration — AI iterates with the user through multiple rounds, refining the plan step by step

Known Limitations

Preset display text does not follow the Web UI locale. The name and description for every preset shipped by this plugin are read from each preset's declaration and rendered verbatim by the Web UI, regardless of which UI language is selected. Only the presets shipped with DeepSeek Harness itself (standard, ptc, minimal, cordis) are localized through the harness's i18n system; community-shipped presets, including every preset in this plugin, are not. Switching the Web UI from Chinese to English will leave these presets' display text in Chinese.

This is a limitation of how the harness consumes preset metadata, not of this plugin. As of this writing the harness exposes no mechanism for plugins to register localized strings for their own presets.

Requirements

• DeepSeek Harness 0.2.0-rc.1 or newer. This release is a declarative preset bundle (the pre-0.1.7 .agent-presets directory installer no longer exists). A host on 0.1.7-rc.x skips the whole bundle through the peer compatibility gate, so update the host first.
• The subagent model-selection opt-in needs the Host's @deepseek-ai/dsh-tool-subagent/model-selection-settings row. The shipped dsh-base + dsh-web-app stack mounts it by default. Without it, a preset that opts in — the five coding presets and chat-subagent — fails to mount as a whole.

Install

dsh plugin --profile web add github:R-LEI2536/dsh-more-agent-presets

Restart the Web profile, then select the preset when creating a session.

The plugin is a declarative preset bundle: installing it adds seven @deepseek-ai/dsh-agent-preset declaration rows to the profile through dsh.bundle.patch (one per preset, in presets/<id>.patch.yml). The declaration rows are registered with @deepseek-ai/dsh-agent-preset-registry and appear in the preset picker automatically. The plugin writes nothing to the user directory — no files are copied anywhere.

Behavior Notes

Declarative presets (DSH 0.1.7+):

• Each preset is a @deepseek-ai/dsh-agent-preset row declared in a bundle patch; config.plugins holds the child plugin rows (tools, prompt sections, groups with isolate realms, and the plugin's inline .mjs plugins referenced as package subpath exports under dsh-more-agent-presets/presets/<id>/, with assets shared by more than one preset under dsh-more-agent-presets/presets/shared/).
• Preset id values (chat-agent, chat-subagent, codex-coding-agent, iflow-coding-agent, iflow-cre-agent, pair-coding-agent, qwencode-coding-agent) are unchanged from previous releases, so existing session selections keep working. chat-subagent is the one row this bundle adds, and it is a new preset rather than a change to chat-agent.
• The registry retains each declaration's composition for live sessions; editing a declaration affects subsequently created Agents.

Multi-file patch, the bundle's own row, and the root entry artifact:

• dsh.bundle.patch is an ordered list (cordis.patch.yml plus one presets/<id>.patch.yml per preset) — the array form DSH core supports. A host composes all eight layers and, because this is a bundle, activation happens on the next start, not on hot mount.
• cordis.patch.yml declares the bundle's own loader row (id/name: dsh-more-agent-presets), and index.mjs is the no-op plugin behind it. That row is what makes the bundle visible in the running composition, and it is the shape every bundle uses for itself (dshmarket inserts dsh-market / dshmarket; @deepseek-ai/dsh-web-app inserts web-runtime / @deepseek-ai/dsh-web-app). Without it, dshmarket's liveness check (liveIncludes) has nothing to match — the seven preset rows are named @deepseek-ai/dsh-agent-preset — and the plugin manager reports "已安装，重启后生效" on every page load, restart or not, while the presets keep working.
• Some third-party tooling only understands the string form of dsh.bundle.patch, and judges an installed build by its declared root entry artifact. index.mjs — declared through main and exports["."] and shipped via files — exists so that check sees a real artifact and so the row above has a plugin behind it.
• Do not remove the self row, index.mjs, main, or exports["."]. Without them the plugin manager reports the plugin as broken ("已安装，校验未通过 / the declared entry artifact is missing"), refuses to install or update it ("updated build has no loadable entry"), or shows it as permanently pending a restart — even though DSH itself loads the bundle correctly.

Migration from ≤1.4.x:

• Versions ≤1.4.4 copied preset directories into $DSH_HOME/.agent-presets (normally ~/.dsh/.agent-presets) with a .dsh-preset-owner.json marker. DSH 0.1.7 no longer reads that directory, and this plugin no longer writes it.
• After upgrading, the old copied directories are inert leftovers. Remove them manually if you want a clean slate:

  rm -rf ~/.dsh/.agent-presets
  

  (Only your own ~/.dsh/.agent-presets; nothing from another plugin or hand-authored presets should live there in the new mechanism, since presets are profile patches now.)

Remove

dsh plugin --profile web remove dsh-more-agent-presets

Removing the bundle removes its patch layer, so the seven preset declarations disappear from the profile. The preset plugins assets installed under the profile's node_modules are removed with the package. No ~/.dsh/.agent-presets cleanup is needed on 0.1.7+ (see Behavior Notes if you are migrating from ≤1.4.x).

License

MIT. Some preset compositions are derived from other open source projects (see NOTICE.md).

Contributing

Contributions are welcome! Feel free to submit issues and pull requests.