From 390a2f09cf2e2c7a17c96b3122d3c469c3a72696 Mon Sep 17 00:00:00 2001 From: apple Date: Mon, 15 Jun 2026 00:24:39 +0800 Subject: [PATCH] docs: add Cursor runtime port guide MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Document path mapping (.claude → .cursor), subagent/file-handoff as Agent Teams fallback, and link from README FAQ Q3. Co-authored-by: Cursor --- CHANGELOG.md | 1 + README.md | 1 + docs/cursor-port.md | 100 ++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 102 insertions(+) create mode 100644 docs/cursor-port.md diff --git a/CHANGELOG.md b/CHANGELOG.md index 1be8060..717a810 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -5,6 +5,7 @@ ## [Unreleased] ### Added +- `docs/cursor-port.md` — Cursor 运行时路径映射与 subagent/file-handoff 降级方案 - 신규 에이전트/스킬 생성 전 중복 검토 단계 (Phase 3-0, Phase 4-0) - `references/agent-design-patterns.md` "에이전트 재사용 설계" 섹션 - `references/skill-writing-guide.md` §9 "스킬 재사용 설계" diff --git a/README.md b/README.md index b64e786..6bda686 100644 --- a/README.md +++ b/README.md @@ -295,6 +295,7 @@ Key finding: effectiveness scales with task complexity — the harder the task, **Evidence:** - Codex port: [github.com/SaehwanPark/meta-harness](https://github.com/SaehwanPark/meta-harness) - Cross-runtime scaffolder: [github.com/Gizele1/harness-init](https://github.com/Gizele1/harness-init) +- **Cursor users:** see [`docs/cursor-port.md`](docs/cursor-port.md) for path mapping and subagent/file-handoff patterns ## License diff --git a/docs/cursor-port.md b/docs/cursor-port.md new file mode 100644 index 0000000..790a858 --- /dev/null +++ b/docs/cursor-port.md @@ -0,0 +1,100 @@ +# Cursor Port Guide — Using Harness Outside Claude Code + +Harness is **Claude-Code-native** by design. This guide maps Harness outputs to **Cursor** (and similar IDEs) so you can reuse the same team architecture without Agent Teams APIs. + +> For Codex runtime, see [SaehwanPark/meta-harness](https://github.com/SaehwanPark/meta-harness). + +--- + +## Path mapping + +| Claude Code (Harness default) | Cursor equivalent | +|------------------------------|-------------------| +| `.claude/agents/{name}.md` | `.cursor/agents/{name}.md` or project rules | +| `.claude/skills/{name}/SKILL.md` | `.cursor/skills/{name}/SKILL.md` or `~/.cursor/skills-cursor/` | +| `CLAUDE.md` harness pointer | `.cursor/rules` or root `AGENTS.md` | +| `_workspace/` artifacts | Same — keep file-based handoff | + +After Harness generates files under `.claude/`, copy or symlink: + +```bash +mkdir -p .cursor/agents .cursor/skills +cp -r .claude/agents/* .cursor/agents/ +cp -r .claude/skills/* .cursor/skills/ +``` + +--- + +## Execution mode mapping + +| Harness mode | Claude Code API | Cursor approach | +|--------------|-----------------|-----------------| +| **Agent Teams** (default) | `TeamCreate`, `SendMessage`, `TaskCreate` | Use **Task subagents** in parallel; file-based handoff via `_workspace/` | +| **Subagents** | `Agent` tool | Direct Task tool with `subagent_type` — closest match | +| **Hybrid** | Phase-level team/sub switching | Phase 1: parallel Task calls → Phase 2: single integrator agent | + +Cursor does not expose `SendMessage` between subagents. **Compensate with:** + +1. **File-based protocol** (Harness Phase 5-1) — each agent writes to `_workspace/{phase}_{agent}_{artifact}.ext` +2. **Orchestrator skill** — one skill reads all artifacts and spawns the next phase +3. **Explicit dependencies** in orchestrator Phase tables + +--- + +## Trigger phrases in Cursor + +Install the harness meta-skill globally: + +```bash +cp -r skills/harness ~/.cursor/skills-cursor/harness +``` + +Then prompt: + +``` +帮我配置 harness：多源新闻情报团队 +build a harness for this project +``` + +Cursor loads skills from `SKILL.md` frontmatter `description` — same trigger rules as Claude Code. + +--- + +## Agent Teams flag + +`CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1` applies to **Claude Code only**. Cursor users can ignore it; use subagent/Task patterns instead. + +--- + +## Recommended workflow + +1. Run Harness in Claude Code **or** copy the meta-skill to Cursor and say "build a harness for …" +2. Prefer **subagent + file handoff** orchestrator template (Template B) when targeting Cursor-only teams +3. Register a minimal pointer in `AGENTS.md`: + +```markdown +## Harness: {domain} + +**Trigger:** For {domain} tasks, use `{orchestrator-skill}` skill. +**Artifacts:** `_workspace/` (preserved between runs) +``` + +--- + +## Limitations + +| Feature | Claude Code | Cursor | +|---------|-------------|--------| +| Live inter-agent messaging | Yes (`SendMessage`) | No — use files | +| Shared task board | Yes (`TaskCreate`) | Manual phase tables in orchestrator | +| Session team reconfigure | `TeamDelete` / `TeamCreate` | Re-run orchestrator phases | + +For full Agent Teams fidelity, stay on Claude Code. For day-to-day IDE work, Template B + `_workspace/` is usually sufficient. + +--- + +## See also + +- [README FAQ Q3](../README.md) — multi-runtime roadmap +- [Gizele1/harness-init](https://github.com/Gizele1/harness-init) — cross-runtime scaffolder +- [affaan-m/everything-claude-code](https://github.com/affaan-m/everything-claude-code) — L2 workflow layer on top of harnesses