README

Code is a commodity. Harnesses get replaced. Context stays.

A time-lapsed replay of a real /wish run, host details hidden and agent launch line removed. Its figures are one labelled sample, from its own genie wish report. Full video
Give /wish one decided task. It checks that the task fits, builds it in its own worktree, runs your repository's own check, has an agent that did not write the code review the exact commit, and opens a pull request. Merging stays with you.
curl -fsSL https://raw.githubusercontent.com/automagik-si/genie/main/install.sh | bash
Claude Code runs it as /wish, Codex as $wish, and other agents by name. Docs · Skills · Discord
One /wish from a real run, shortened:
You: /wish Delete dead code: src/lib/defaults.ts has no production callers
(only a type import of AgentDefaults in src/lib/workspace.ts) and
src/__tests__/dir-ls.test.ts reimplements a removed handler and asserts
on its own copy. Move the AgentDefaults type into workspace.ts, delete
defaults.ts, defaults.test.ts and dir-ls.test.ts.
Genie: I started the wish workflow in the background on branch
wish/delete-dead-defaults. It checks that the task fits a single wish,
makes the change in its own worktree, runs the full check, gets an
independent review and opens a PR against dev.
Genie: Review: SHIP. Checks: pass. The PR is open and its remote head matches the
local one. This run merged nothing. Merging is yours.
How it works
Designs and plans stay in your repository as markdown, with the reviews that approved them, so the next agent you use reads the same context. Each step is a skill. One decided task goes straight to wish. Anything bigger takes the whole loop, where every gate is run by an agent that did not do the work:
brainstorm → design review → wish → plan review → work → implementation review → PR
- brainstorm settles an idea into a
DESIGN.md. A design review approves it before any plan is written. - wish delivers one task end to end and returns a pull request ready for you to merge. Work bigger than one task becomes a
WISH.mdwith execution groups, and a plan review approves it before anything is built. - work runs the approved groups in dependency order, each with its own scoped worker, and ends at an implementation review.
- review runs those reviews. After work, it checks the result against the plan's criteria and returns SHIP, FIX-FIRST or BLOCKED. fix repairs the blocking gaps with bounded retries and a fresh reviewer.
- council sits beside the loop. It pressure-tests a decision through independent lenses and changes nothing.
Designs, plans and verdicts land in git under .genie/brainstorms/<slug>/ and .genie/wishes/<slug>/. Task state lives in one SQLite file per repository, .genie/genie.db, and genie board shows it as a kanban. Genie is built with Genie: this repository's own designs and plans are in .genie/brainstorms and .genie/wishes.
Skills
Genie v6.261003.1 ships 18 skills. The skill catalog groups them by category, and the Skills docs show a card for each one.
All 18 skills
authoring: Write a Genie skill that passes the shipped contractbrainstorm: Settle an idea into a reviewed designcouncil: Pressure-test decisions through independent lensesdeslop: Clean up scoped prose, code or interface without losing meaningdocs: Audit docs and DX against the live product; write on requestfix: Repair review gaps with bounded retriesgenie: Route or resume work only where Genie adds valuegenie-hacks: Find grounded Genie workflow patterns and recipesmerge: Resolve a merge or rebase by intent, then re-run the gaterefine: Improve prompts for OpenAI or Claudereport: Diagnose a failure to root cause; file issues on requestresearch: Investigate against primary sources and cite every claimreview: Assess plans, code, PRs, and repository qualityskill-audit: Audit the skill catalogue for overlap, staleness, driftverify: Prove completion with fresh gate, diff, and check outputwish: Deliver one task end to end, or plan a multi-group wishwork: Execute approved wishes with evidenceworkfly: Build a saved workflow from a procedure
Workflows
更多「Agent、自動化與工作流程」外掛
dsh-agent-teams
作者 nanmicoder
DeepSeek Harness 的 Agent Teams 多智慧體協作外掛,支援多個 AI Agent 組成團隊,協同完成複雜任務,實現任務分配、並行執行、成員通訊與團隊協作。 AgentTeams plugin for DeepSeek Harness
dsh-blender-plugin
作者 sixtysevenlf
DSH × Blender 直連外掛 v3.0.6:15 個模型工具讓 AI 驅動 GUI Blender(視口 / 自定義視角 / 無頭 worker / 作業層 / 事務回滾 / 配方庫),34 family / 228 op 契約層;npm 安裝 + addon zip,GUI 優先、無頭備用;dsh-TUI 選中 chip / @ 提及 / /blender,隨包建模 skill,支援 WSL / Windows / macOS。
dsh-free-search
作者 dddmuc
為DeepSeek Harness提供的免費網路搜尋服務商:即使沒有API金鑰也可使用。Free web search provider for DeepSeek Harness:usable even without an API key.
dsh-bili-scroll
作者 c1overzz
DSH 外掛:每天替你刷一輪 B站,讀字幕挑出值得的、寫一份日報,值得的 @ 你
