dsh-generative-ideas

Roadmap ideation for DeepSeek Harness — generate, simulate, and compare distinct roadmap options via headless agent runs; pick the one that fits, export as goal.md, dispatch to an agent in the chosen workspace.

路线图构想：通过后台 agent 生成多个不同战略立场的路线图，对比后选出最合适的，导出为 goal.md 并发送到指定工作区的 agent。

Install

dsh plugin --profile web add dsh-generative-ideas

Restart the dsh web process. An Ideas entry appears in the sidebar.

The flow

1. Context — enter your focus, pick a target workspace, set horizon and constraints
2. Deep Research toggle (default ON) — the generation agent must research 12+ competitors, GitHub open-source repos, and .refs/ curated references before producing options; every thesis names what the best-in-class competitor does differently
3. Generating — the host spawns a background headless agent (30-90s); the sidebar icon pulses while running; close the panel and come back — results persist
4. Compare — 4 genuinely distinct roadmap options as bleed-row cards: name, thesis, phases with acceptance items, risks, effort, differentiator
5. Pre-flight actions (on the compare step): Reroll (fresh options, same focus), Push (force deep research ON and regenerate), Discuss (close panel, talk in chat)
6. Chosen — pick one, Export goal.md to the workspace root

Deep research doctrine

When the toggle is ON (or Push is pressed), the generation prompt mandates:

• Minimum 12 competitors studied with concrete methods captured
• GitHub open-source repos in the space — READMEs, design docs, issue trackers
• .refs/ directory curated research references
• Options grounded in specific evidence — competitor names in theses, real patterns in phases, actual risks others encountered

Architecture

src/host.js            Node half — /api/rich-ideas/{state,generate,export} routes.
                       Generation runs as a background dsh --profile headless call;
                       state persists across panel close/reopen.
src/client.bundle.js   Browser half — sidebar entry + stepper overlay panel.

License

MIT