1149784810
jayhe-dsh-gamemaker
DeepSeek Harness plugin: game development role subagents (planner=minimal preset, executor=hard PTC code mode, reviewer=minimal) + bundled game-dev/game-minimal agent presets.
- Stars
- 1
- Language
- JavaScript
- Created
- Aug 15, 2026
- Updated
- Aug 15, 2026
Introduction
jayhe-dsh-gamemaker
Game development role subagents for DeepSeek Harness: a game-director pipeline built from three dedicated subagent providers plus two bundled agent presets.
What it provides
| Tool / Provider | Role | Mode |
|---|---|---|
game-planner-provider (tool game_planner) | Planner | Composed from the real minimal-family preset: fixed complete persona, no runtime context, and exactly two tools — persistent shell + str_replace_editor (bash swaps to pwsh on Windows) |
game-executor-provider (tool game_executor) | Executor | Hard PTC (Code Mode): the child declares presentAs('code') on its own scope, so the model sees only run_code + the generated SDK |
game-reviewer-provider (tool game_reviewer) | Reviewer | Same as planner: the real minimal preset, used read-only |
Bundled presets:
game-dev(游戏开发工坊) — the conductor: a mandatory pre-work interview (tech route WEB/UE/Unity/Godot, game type, and all other details), the five-phase pipeline, and explicit 【模式切换】 announcements for every mode switch.game-minimal(兼容极简模式) — identical to the shippedminimalpreset, plus one special case: on Windows the persistent bash tool is replaced by the deployment'spwshtool.
How it works
Planner/reviewer children are composed inside the child creation window via ctx.agentPresets.mount(childCtx, 'game-minimal'), which re-parents the child's scope to the preset's standing mount. The child therefore runs the real minimal composition — fixed persona, suppressed runtime context, two tools — instead of a persona + toolFilter approximation. Executor children join the parent preset, then their own scope declares presentAs('code'), making PTC mode hard: direct native calls are rejected, everything goes through run_code programs.
Because a two-tool minimal child cannot call delegation tools itself, the planner authors the dispatch manifest (docs/plan.md — module list with file boundaries, interfaces, acceptance criteria) and the conductor dispatches the executors in parallel, waits with blocking job_output(wait: true) or the system's job-completion notices (no goal-based polling), then hands merging back to the planner and review to the reviewer.
Requirements
- A dsh deployment with the
subagents,agentPresets, andcodeRuntimehost services (the stock web profile provides all three). - Node >= 18 (ESM), bash for the install scripts.
Install
# 1. add the plugin package to the web profile (or link it during development)
export PATH=/usr/local/nodejs22/bin:$PATH # pnpm may live outside PATH
dsh plugin --profile web add jayhe-dsh-gamemaker
# development: dsh plugin --profile web add link:/absolute/path/jayhe-dsh-gamemaker
# 2. link the dsh-internal seam packages from the LOCAL dsh installation
bash node_modules/jayhe-dsh-gamemaker/install/link-dsh-deps.sh
# 3. install the bundled presets into the user preset root
bash node_modules/jayhe-dsh-gamemaker/install/install-presets.sh
# 4. add the host row to the profile patch (~/.dsh/profiles/web/cordis.patch.yml)
# see install/cordis.patch.example.yml
# 5. restart dsh (systemd) and refresh the page
systemctl restart dsh
Then start a new session on the 游戏开发工坊 preset.
Dependency note
The plugin consumes @deepseek-ai/dsh-subagent, @deepseek-ai/dsh-llm, and @deepseek-ai/dsh-session. These packages ship with the dsh deployment itself and are not pinned through npm: the public registry only carries older rc versions with a different API surface. install/link-dsh-deps.sh links the deployment's own copies (auto-detected from the dsh CLI path, overridable via DSH_ROOT), which guarantees API parity with the running harness.
Configuration
The provider plugin accepts one optional field:
- id: game-roles
name: 'jayhe-dsh-gamemaker'
config:
minimalPreset: game-minimal # preset id composed for planner/reviewer children
License
MIT