Skip to content

Self-hosted workflows for coding-agent delivery. Orchestrate CodeBuddy, Codex and Claude Code with scripts and human approval gates. Includes Local workflows for coding assistants.

feat(local): add Hermes and DSH integrations (#158)

Chen Zhonge0ce2de54a95
diff --git a/README.md b/README.md
index d2045d2f..d4ec28d9 100644
--- a/README.md
+++ b/README.md
@@ -181,6 +181,8 @@ Then run `/start-devflow "<your request>"` in Claude Code.
| --- | --- |
| Claude Code · CodeBuddy · Codex · Cursor | Classic (default) and Portable editions |
| Pi | Portable edition |
+| DeepSeek Harness | Portable edition |
+| Hermes Agent | Portable edition |
See the [local guide](docs/local/usage.md) for each host's entry point and [editions](docs/local/editions.md) for the differences. Source lives in [`local/`](local/). Local runs the built-in workflows today; exporting a platform SOP for standalone local execution is planned.
diff --git a/docs/local/architecture.md b/docs/local/architecture.md
index 05ab392f..8cd04172 100644
--- a/docs/local/architecture.md
+++ b/docs/local/architecture.md
@@ -11,8 +11,8 @@
## Portable Edition
`skills/` 提供独立的新 DevFlow 状态机、模板、规则、测试和 adapter,
-支持 CodeBuddy、Codex、Cursor、Claude Code 与 Pi(Pi 只有声明型 adapter,
-isolated 模式依赖公开扩展 `npm:pi-subagents`)。脚本只依赖 Python 3.8+
+支持 CodeBuddy、Codex、Cursor、Claude Code、Pi、DeepSeek Harness 和 Hermes Agent(Pi 只有声明型 adapter,
+isolated 模式依赖公开扩展 `npm:pi-subagents`;DSH 使用内置 `subagent` 工具;Hermes 使用 `delegate_task`)。脚本只依赖 Python 3.8+
标准库;具体 Agent 调度能力由宿主提供。
`skills/manifest.json` 是跨宿主 Skill 集合清单。`skills/` 是这一层的唯一
diff --git a/docs/local/compatibility.md b/docs/local/compatibility.md
index 86c76275..5377a87f 100644
--- a/docs/local/compatibility.md
+++ b/docs/local/compatibility.md
@@ -11,6 +11,8 @@
| Cursor | `.cursor/` 完整包 | `.cursor/skills` + 两个执行器 |
| Claude Code | `.claude/` 完整包 | `.claude/skills` + 两个执行器 |
| Pi | 无(Portable only) | `.pi/skills` + `subagent` adapter(需 `npm:pi-subagents`) |
+| DeepSeek Harness | 无(Portable only) | `.dsh/skills` + `subagent` adapter(DSH base 默认提供) |
+| Hermes Agent | 无(Portable only) | `.hermes/skills` + `delegate_task` adapter;首次使用需 `hermes skills trust .`;可选 `hermes-plugin/` 提供 `/loopforge` 控制命令 |
Classic 的 Cursor/Claude 目录是生成产物;Portable 的事实只保存在 `skills/`。
两者只共享显式行为合同和测试,不共享宿主运行文件。
@@ -19,7 +21,7 @@ Classic 的 Cursor/Claude 目录是生成产物;Portable 的事实只保存在
1. **结构门禁**:校验 JSON/YAML、Skill frontmatter、manifest 引用和链接。
2. **漂移门禁**:`build-classic-hosts.py --check` 逐字比较 Classic 生成包。
-3. **安装门禁**:统一 CLI 覆盖 5 宿主 × 2 edition(pi 仅 Portable),并验证重复安装、更新、卸载、
+3. **安装门禁**:统一 CLI 覆盖 7 宿主 × 2 edition(Pi、DSH 和 Hermes 仅 Portable),并验证重复安装、更新、卸载、
mixed-edition 拒绝和用户文件保护。
4. **流程回归**:标准库单测和 CLI E2E 验证状态机、门禁、产物与恢复协议。
5. **真实宿主 canary**:发布前或宿主大版本升级后,在隔离 fixture 仓库用真实
@@ -27,6 +29,10 @@ Classic 的 Cursor/Claude 目录是生成产物;Portable 的事实只保存在
版本、调用轨迹、最终状态和 diff。此层需要各宿主凭据和运行环境,不应伪装成
无凭据的普通单测。
+Hermes 与 DSH 不绑定私有版本号:当前稳定版及后续版本必须提供清单中列出的公开能力。
+能力缺失时,Hermes 不得伪造 isolated 执行器,DSH/Hermes 均须记录原因并按适配器合同降级到
+`single-context`;这条能力策略就是版本边界。
+
前四层在普通 PR CI 中运行。第五层建议使用受保护的手动或定时流水线;任何宿主
变更应先更新 manifest 的核对信息,再运行 canary。宿主 canary 失败时不能只更新
快照,应先判断是字段协议变化、工具缺失、权限变化还是模型行为回归。
diff --git a/docs/local/editions.md b/docs/local/editions.md
index 8e456e0d..0dfeaf3b 100644
--- a/docs/local/editions.md
+++ b/docs/local/editions.md
@@ -6,7 +6,7 @@ LoopForge Local 内置两套工作流,源码位于 `local/`。为避免同一
| Edition | 仓库事实源 | 入口 | 维护方式 |
| --- | --- | --- | --- |
-| Portable | `skills/` | CodeBuddy/Cursor/Claude 使用 `/devflow`,Codex 使用 `$devflow`,Pi 使用 `/skill:devflow` | 修改 canonical Skill 和 adapter manifest |
+| Portable | `skills/` | CodeBuddy/Cursor/Claude/DeepSeek Harness/Hermes 使用 `/devflow`,Codex 使用 `$devflow`,Pi 使用 `/skill:devflow` | 修改 canonical Skill 和 adapter manifest |
| Classic(默认) | `.codebuddy/`、`.codex/`、`.cursor/`、`.claude/` | CodeBuddy/Cursor/Claude 使用 `/start-devflow`,Codex 使用 `$devflow-codex` | 以 `.codebuddy` 为参考行为;Cursor/Claude 由生成器产生 |
表中路径均相对于 `local/`。安装器将对应资产安装到目标项目的宿主原生目录,
@@ -27,3 +27,13 @@ loopforge skills install claude
Portable 的 Pi 适配是声明型的:`skills install pi` 只安装 `.pi/skills/`,isolated
(medium/large)模式还需要公开扩展 `pi install npm:pi-subagents` 提供 `subagent` 工具。
+
+DeepSeek Harness 使用 `skills install dsh` 安装到 `.dsh/skills/`。DSH base profile 默认提供
+`subagent` 工具;如果当前 profile 没有该工具,插件会记录具体原因并将 medium/large 流程降级为
+`single-context`,不会伪造 isolated 执行器。
+
+Hermes Agent 使用 `skills install hermes` 安装到项目的 `.hermes/skills/`。Hermes 首次发现项目技能时
+要求显式信任仓库:在项目根目录运行 `hermes skills trust .`。medium/large 阶段使用 Hermes 的
+`delegate_task` 创建独立子 Agent;当前会话缺少该工具时必须记录原因并降级为 `single-context`。
+如需独立查看状态、产物和审批,可选安装 `local/hermes-plugin/`;插件只调用项目内的
+`workflow_state.py`,不会覆盖 Hermes profile 配置。
diff --git a/docs/local/usage.md b/docs/local/usage.md
index b60ad726..cd894ba9 100644
--- a/docs/local/usage.md
+++ b/docs/local/usage.md
@@ -14,6 +14,8 @@ Requires Node.js 20+ and Python 3.8+. The commands below install the default Cla
| Codex | `npx -y loopforge-cli@latest install codex` | `$devflow-codex "<request>"` |
| Cursor | `npx -y loopforge-cli@latest install cursor` | `/start-devflow "<request>"` |
| Claude Code | `npx -y loopforge-cli@latest install claude` | `/start-devflow "<request>"` |
+| DeepSeek Harness | `npx -y loopforge-cli@latest skills install dsh` | `/devflow "<request>"` |
+| Hermes Agent | `npx -y loopforge-cli@latest skills install hermes` then `hermes skills trust .` | `/devflow "<request>"` |
For example, with Claude Code:
@@ -32,7 +34,7 @@ Check any installation with `npx -y loopforge-cli@latest status <agent>`. If ins
npx -y loopforge-cli@latest doctor
```
-Pi is supported through the [Portable Edition](#portable-edition).
+Pi, DeepSeek Harness and Hermes Agent are supported through the [Portable Edition](#portable-edition). DSH has an optional [plugin](../../local/dsh-plugin/README.md) for task controls, status, artifacts and approval gates. Hermes uses its project-local Agent Skills discovery and `delegate_task` tool; it does not modify `~/.hermes/config.yaml`. Hermes users can optionally install the native [Hermes plugin](../../local/hermes-plugin/README.md) for the same `/loopforge status`, `/loopforge result` and `/loopforge approve` controls.
## How it works
@@ -64,6 +66,8 @@ Your request
| Cursor | ✓ | ✓ |
| Claude Code | ✓ | ✓ |
| Pi | — | ✓ |
+| DeepSeek Harness | — | ✓ |
+| Hermes Agent | — | ✓ |
LoopForge enables only one edition per coding agent in each project. See [Edition selection](editions.md) for the differences and directory layout.
@@ -136,6 +140,8 @@ The Classic Edition from the quick start is the right choice for most users. If
| Cursor | `npx -y loopforge-cli@latest skills install cursor` | `/devflow "<request>"` |
| Claude Code | `npx -y loopforge-cli@latest skills install claude` | `/devflow "<request>"` |
| Pi | `npx -y loopforge-cli@latest skills install pi`, then `pi install npm:pi-subagents` | `/skill:devflow "<request>"` |
+| DeepSeek Harness | `npx -y loopforge-cli@latest skills install dsh` | `/devflow "<request>"` |
+| Hermes Agent | `npx -y loopforge-cli@latest skills install hermes`, then `hermes skills trust .` | `/devflow "<request>"` (optional plugin: `/loopforge`) |
See [Edition selection](editions.md) for the complete boundaries and directory layout.
@@ -147,7 +153,7 @@ For ongoing use, install the CLI globally:
npm install --global loopforge-cli
```
-Then maintain an installation with the same agent names as above (`codebuddy`, `codex`, `cursor`, `claude`, and `pi` for the Portable Edition):
+Then maintain an installation with the same agent names as above (`codebuddy`, `codex`, `cursor`, `claude`, `pi`, `dsh`, and `hermes` for the Portable Edition):
```bash
# Classic Edition
diff --git a/docs/local/usage.zh-CN.md b/docs/local/usage.zh-CN.md
index f3eb0741..9aadf874 100644
--- a/docs/local/usage.zh-CN.md
+++ b/docs/local/usage.zh-CN.md
@@ -14,6 +14,8 @@ LoopForge Local 在已有的编程助手中运行本地开发工作流,无需
| Codex | `npx -y loopforge-cli@latest install codex` | `$devflow-codex "<需求>"` |
| Cursor | `npx -y loopforge-cli@latest install cursor` | `/start-devflow "<需求>"` |
| Claude Code | `npx -y loopforge-cli@latest install claude` | `/start-devflow "<需求>"` |
+| DeepSeek Harness | `npx -y loopforge-cli@latest skills install dsh` | `/devflow "<需求>"` |
+| Hermes Agent | `npx -y loopforge-cli@latest skills install hermes`,然后 `hermes skills trust .` | `/devflow "<需求>"` |
以 Claude Code 为例:
@@ -32,7 +34,7 @@ npx -y loopforge-cli@latest status claude
npx -y loopforge-cli@latest doctor
```
-Pi 通过 [Portable Edition](#portable-edition) 支持。
+Pi、DeepSeek Harness 和 Hermes Agent 通过 [Portable Edition](#portable-edition) 支持。DSH 还可以安装[可选插件](../../local/dsh-plugin/README.md),提供任务控制、状态、产物和人工审批入口。Hermes 使用项目级 `.hermes/skills/` 和 `delegate_task`,不会修改 `~/.hermes/config.yaml`;需要独立命令时可安装[可选 Hermes 插件](../../local/hermes-plugin/README.md),提供 `/loopforge status`、`/loopforge result` 和 `/loopforge approve`。
## 工作方式
@@ -63,6 +65,8 @@ LoopForge 会根据任务复杂度选择路径:小改动走精简的单角色
| Cursor | ✓ | ✓ |
| Claude Code | ✓ | ✓ |
| Pi | — | ✓ |
+| DeepSeek Harness | — | ✓ |
+| Hermes Agent | — | ✓ |
同一项目中,每个编码 Agent 只能启用一种 Edition。两者的区别和目录说明见 [Edition selection](editions.md)。
@@ -135,6 +139,8 @@ LoopForge 尤其适合这些任务:
| Cursor | `npx -y loopforge-cli@latest skills install cursor` | `/devflow "<需求>"` |
| Claude Code | `npx -y loopforge-cli@latest skills install claude` | `/devflow "<需求>"` |
| Pi | `npx -y loopforge-cli@latest skills install pi`,然后 `pi install npm:pi-subagents` | `/skill:devflow "<需求>"` |
+| DeepSeek Harness | `npx -y loopforge-cli@latest skills install dsh` | `/devflow "<需求>"` |
+| Hermes Agent | `npx -y loopforge-cli@latest skills install hermes`,然后 `hermes skills trust .` | `/devflow "<需求>"`(可选插件:`/loopforge`) |
完整边界和目录说明见 [Classic / Portable 说明](editions.md)。
@@ -146,7 +152,7 @@ LoopForge 尤其适合这些任务:
npm install --global loopforge-cli
```
-之后用上面的宿主名(`codebuddy`、`codex`、`cursor`、`claude`,Portable Edition 另有 `pi`)维护安装:
+之后用上面的宿主名(`codebuddy`、`codex`、`cursor`、`claude`,Portable Edition 另有 `pi`、`dsh` 和 `hermes`)维护安装:
```bash
# Classic Edition
diff --git a/local/README.md b/local/README.md
index 302406c2..752163e0 100644
--- a/local/README.md
+++ b/local/README.md
@@ -1,6 +1,6 @@
# LoopForge Local
-LoopForge 的本地工作流组件,在 CodeBuddy、Codex、Cursor、Claude Code 或 Pi 中使用,无需启动平台。
+LoopForge 的本地工作流组件,在 CodeBuddy、Codex、Cursor、Claude Code、Pi、DeepSeek Harness 或 Hermes Agent 中使用,无需启动平台。
Local workflows for LoopForge, running inside an existing coding assistant without the platform server.
@@ -19,7 +19,7 @@ npx -y loopforge-cli@latest install claude
npx -y loopforge-cli@latest status claude
```
-Start with `/start-devflow "<request>"` in Claude Code. Replace `claude` with `codebuddy`, `codex`, or `cursor` when installing for another host; the guides list each host's entry point. Pi uses the Portable edition.
+Start with `/start-devflow "<request>"` in Claude Code. Replace `claude` with `codebuddy`, `codex`, or `cursor` when installing for another host; use `skills install pi`, `skills install dsh`, or `skills install hermes` for Portable hosts. DSH uses `/devflow "<request>"` after installing the Skill, and the optional DSH plugin adds `/loopforge` task controls. Hermes loads the project skill after `hermes skills trust .`; its optional native plugin adds the same `/loopforge` controls.
在 Claude Code 中通过 `/start-devflow "<需求>"` 启动。其他宿主的安装命令、入口及 Portable 用法见统一文档。
diff --git a/local/dsh-plugin/LICENSE b/local/dsh-plugin/LICENSE
new file mode 100644
index 00000000..54a9df3a
--- /dev/null
+++ b/local/dsh-plugin/LICENSE
@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Tencent LoopForge Contributors
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.
diff --git a/local/dsh-plugin/README.md b/local/dsh-plugin/README.md
new file mode 100644
index 00000000..0896e108
--- /dev/null
+++ b/local/dsh-plugin/README.md
@@ -0,0 +1,51 @@
+# LoopForge DSH 插件
+
+这个插件是一个适用于 DSH 0.2.1 及后续 bundle-compatible 版本的 Host bundle,把
+LoopForge Local 接入 DeepSeek Harness(DSH)Profile。它复用项目中
+`.dsh/skills/devflow` 的 Portable 工作流,不复制工作流规则或产物模板。
+
+## 本地安装
+
+先在目标项目安装 DSH Portable Skill:
+
+```bash
+loopforge skills install dsh
+```
+
+再从 LoopForge 源码目录把 bundle 加入目标 DSH profile(`web` 也可以换成你自己的
+profile 名称):
+
+```bash
+dsh plugin --profile web add -w /path/to/LoopForge/local/dsh-plugin
+```
+
+安装器会把 `@tencent/loopforge-dsh` 写入 profile 的 `dsh.profile.bundles`,并由
+`cordis.patch.yml` 插入插件行;旧版 `dsh-plugin.json` 清单不会被当前 DSH Loader
+读取。安装后可用下面的只读检查确认 bundle 已组合:
+
+```bash
+dsh --profile web --dump-config
+dsh --profile web --dump-config-schema
+```
+
+输出应包含 `@tencent/loopforge-dsh`、`loopforge-dsh`,且 schema 的
+`diagnostics` 为空。
+
+重启 DSH 后可用:
+
+- `/devflow <需求>`:直接按 Skill 执行完整工作流。
+- `/loopforge start <需求>`:初始化一个 medium 任务并显示状态文件位置。
+- `/loopforge status`:查看最近任务、当前阶段和待审批门禁。
+- `/loopforge result`:查看任务状态和产物目录。
+- `/loopforge approve [阶段]`:批准当前人工门禁;REQUIREMENT 会强制带用户确认。
+- `loopforge_status`:供 DSH Agent 读取结构化状态。
+
+阶段 Agent 使用 DSH 的 `subagent` 工具。没有该工具时,工作流必须按协议记录
+`single-context` 降级原因,不能伪造隔离执行器。
+
+## 卸载
+
+```bash
+dsh plugin --profile web remove loopforge-dsh
+loopforge skills uninstall dsh
+```
diff --git a/local/dsh-plugin/cordis.patch.yml b/local/dsh-plugin/cordis.patch.yml
new file mode 100644
index 00000000..4e4a04dd
--- /dev/null
+++ b/local/dsh-plugin/cordis.patch.yml
@@ -0,0 +1,3 @@
+- insert:
+ - id: loopforge-dsh
+ name: '@tencent/loopforge-dsh'
diff --git a/local/dsh-plugin/index.js b/local/dsh-plugin/index.js
new file mode 100644
index 00000000..eb70e45e
--- /dev/null
+++ b/local/dsh-plugin/index.js
@@ -0,0 +1,228 @@
+import { spawn } from 'node:child_process'
+import { existsSync, readdirSync, readFileSync, statSync } from 'node:fs'
+import { join, resolve } from 'node:path'
+
+export const name = 'loopforge-dsh'
+export const inject = ['commands', 'tools']
+
+const SKILL_ROOT = '.dsh/skills/devflow'
+const WORKFLOW = 'scripts/workflow_state.py'
+const SAFE_SLUG = /^[A-Za-z0-9][A-Za-z0-9_-]{0,127}$/u
+const APPROVAL_STAGES = new Set(['REQUIREMENT', 'DESIGN', 'REVIEW', 'TEST'])
+const MAX_RESULT_FILES = 40
+
+function cwdOf(owner) {
+ const agent = owner?.agent
+ const session = agent?.session
+ const service = agent?.ctx?.workingDirectory
+ ?? agent?.ctx?.get?.('workingDirectory')
+ ?? owner?.ctx?.workingDirectory
+ ?? owner?.ctx?.get?.('workingDirectory')
+ const candidate = typeof service?.get === 'function' && session
+ ? service.get(session)
+ : session?.header?.cwd
+ ?? process.env.DSH_PROJECT_ROOT
+ ?? process.cwd()
+ return resolve(candidate)
+}
+
+function run(argv, cwd, signal) {
+ return new Promise((resolvePromise, reject) => {
+ if (signal?.aborted) {
+ reject(signal.reason instanceof Error ? signal.reason : new Error('LoopForge command aborted'))
+ return
+ }
+ let settled = false
+ const child = spawn(argv[0], argv.slice(1), { cwd, stdio: ['ignore', 'pipe', 'pipe'] })
+ let stdout = ''
+ let stderr = ''
+ child.stdout.on('data', chunk => { stdout += chunk })
+ child.stderr.on('data', chunk => { stderr += chunk })
+ const abort = () => {
+ if (!child.killed) child.kill('SIGTERM')
+ }
+ signal?.addEventListener('abort', abort, { once: true })
+ child.on('error', error => {
+ if (settled) return
+ settled = true
+ signal?.removeEventListener('abort', abort)
+ reject(error)
+ })
+ child.on('close', code => {
+ if (settled) return
+ settled = true
+ signal?.removeEventListener('abort', abort)
+ if (code === 0) resolvePromise(stdout.trim())
+ else if (signal?.aborted) {
+ reject(signal.reason instanceof Error ? signal.reason : new Error('LoopForge command aborted'))
+ } else reject(new Error(stderr.trim() || `LoopForge command exited with ${code}`))
+ })
+ })
+}
+
+function scriptPath(cwd) {
+ const path = join(cwd, SKILL_ROOT, WORKFLOW)
+ if (!existsSync(path)) {
+ throw new Error('LoopForge Skill is not installed in this project; run `loopforge skills install dsh` first.')
+ }
+ return path
+}
+
+function stateFiles(cwd) {
+ const root = join(cwd, 'artifacts')
+ if (!existsSync(root)) return []
+ const files = []
+ for (const entry of readdirSync(root)) {
+ const path = join(root, entry, 'workflow-state.json')
+ if (existsSync(path) && statSync(path).isFile()) files.push(path)
+ }
+ return files.sort((a, b) => statSync(b).mtimeMs - statSync(a).mtimeMs)
+}
+
+function readState(cwd, slug) {
+ if (slug !== undefined && !SAFE_SLUG.test(slug)) {
+ throw new Error('Task slug must contain only letters, digits, hyphens, and underscores.')
+ }
+ const paths = stateFiles(cwd)
+ const path = slug ? join(cwd, 'artifacts', slug, 'workflow-state.json') : paths[0]
+ if (!path || !existsSync(path)) return undefined
+ return { path, state: JSON.parse(readFileSync(path, 'utf8')) }
+}
+
+function renderState(record) {
+ if (!record) return 'No LoopForge task found in this project.'
+ const { state, path } = record
+ const completed = Object.entries(state.stages ?? {})
+ .filter(([, stage]) => stage?.status === 'completed')
+ .map(([stage]) => stage)
+ return [
+ `Task: ${state.task_slug ?? 'unknown'}`,
+ `Status: ${state.status ?? 'unknown'}`,
+ `Size: ${state.size_class ?? 'unknown'}`,
+ `Artifacts: ${path}`,
+ completed.length ? `Completed stages: ${completed.join(', ')}` : '',
+ state.awaiting_approval ? `Awaiting approval: ${state.awaiting_approval}` : '',
+ state.next_stage ? `Next stage: ${state.next_stage}` : '',
+ ].filter(Boolean).join('\n')
+}
+
+function resultFiles(cwd, slug) {
+ const root = join(cwd, 'artifacts', slug)
+ if (!existsSync(root)) return []
+ const files = []
+ const visit = directory => {
+ for (const entry of readdirSync(directory, { withFileTypes: true })) {
+ const path = join(directory, entry.name)
+ if (entry.isDirectory()) visit(path)
+ else if (entry.isFile()) files.push(path.slice(root.length + 1))
+ }
+ }
+ visit(root)
+ return files.sort().slice(0, MAX_RESULT_FILES)
+}
+
+function parseInput(rawInput) {
+ const input = rawInput.trim()
+ const match = /^(start|status|approve|result|help)(?:\s+([\s\S]*))?$/u.exec(input)
+ return match ? { command: match[1], value: match[2]?.trim() ?? '' } : { command: 'help', value: '' }
+}
+
+async function handle(invocation, ctx) {
+ const cwd = cwdOf(invocation)
+ const { command, value } = parseInput(invocation.rawInput)
+ if (command === 'help') {
+ return { kind: 'success', text: 'Usage: /loopforge start <request> | status [slug] | result [slug] | approve [stage] | help' }
+ }
+ if (command === 'status') return { kind: 'success', text: renderState(readState(cwd, value || undefined)) }
+ if (command === 'result') {
+ const record = readState(cwd, value || undefined)
+ if (!record) return { kind: 'success', text: 'No LoopForge task found in this project.' }
+ const files = resultFiles(cwd, record.state.task_slug)
+ const listing = files.length
+ ? `\n\nArtifact files:\n${files.map(file => `- ${file}`).join('\n')}`
+ : '\n\nNo artifact files have been written yet.'
+ return { kind: 'success', text: `${renderState(record)}${listing}` }
+ }
+ const script = scriptPath(cwd)
+ if (command === 'approve') {
+ const record = readState(cwd)
+ if (!record) return { kind: 'error', text: 'No LoopForge task is waiting for approval.' }
+ const stage = value || record.state.awaiting_approval
+ if (!stage) return { kind: 'error', text: 'No approval stage is pending.' }
+ if (!APPROVAL_STAGES.has(stage)) {
+ return { kind: 'error', text: `Unknown approval stage: ${stage}. Use REQUIREMENT, DESIGN, REVIEW, or TEST.` }
+ }
+ const args = ['python3', script, 'approve', '--state', record.path, '--stage', stage]
+ if (stage === 'REQUIREMENT') args.push('--user-confirmed')
+ await run(args, cwd, invocation.signal)
+ return { kind: 'success', text: `Approved ${stage}. Resume the DSH workflow to continue.` }
+ }
+ if (!value) return { kind: 'error', text: 'A request is required: /loopforge start <request>' }
+ const slug = value.toLowerCase().replace(/[^a-z0-9]+/g, '-').replace(/^-|-$/g, '').slice(0, 48) || `dsh-${Date.now()}`
+ const statePath = join(cwd, 'artifacts', slug, 'workflow-state.json')
+ const hasSubagent = typeof ctx?.tools?.get === 'function'
+ && ctx.tools.get('subagent', invocation.agent) !== undefined
+ const args = [
+ 'python3', script, 'init', '--project-root', cwd, '--slug', slug,
+ '--size', 'medium', '--mode', 'auto', '--coordinator-id', 'dsh-main',
+ ]
+ if (hasSubagent) {
+ args.push('--execution-mode', 'isolated', '--host-adapter', 'dsh')
+ } else {
+ args.push(
+ '--execution-mode', 'single-context',
+ '--fallback-reason', 'DSH subagent tool is unavailable in this session',
+ )
+ }
+ const output = await run(args, cwd, invocation.signal)
+ const execution = hasSubagent
+ ? 'isolated stages via the DSH subagent tool'
+ : 'single-context fallback because the DSH subagent tool is unavailable'
+ return { kind: 'success', text: `LoopForge task initialized: ${slug}\nRequest: ${value}\nState: ${statePath}\n${output}\nExecution: ${execution}. Follow the devflow Skill to resume the workflow.` }
+}
+
+export function apply(ctx) {
+ ctx.commands.register({
+ name: 'loopforge',
+ description: 'Start, inspect, and approve LoopForge Local workflows',
+ input: { hint: '[start <request>|status [slug]|result [slug]|approve [stage]|help]' },
+ handler: invocation => handle(invocation, ctx),
+ })
+ ctx.tools.register({
+ name: 'loopforge_status',
+ description: 'Read the current LoopForge Local workflow state and pending human approval for the current DSH project. Read-only.',
+ parameters: {
+ type: 'object',
+ properties: {
+ slug: { type: 'string', description: 'Optional task slug; omit to inspect the most recently updated task.' },
+ },
+ additionalProperties: false,
+ },
+ output: {
+ schema: {
+ type: 'object',
+ additionalProperties: false,
+ properties: {
+ found: { type: 'boolean' },
+ status: { type: 'string' },
+ task_slug: { type: 'string' },
+ awaiting_approval: { type: 'string' },
+ state_path: { type: 'string' },
+ },
+ required: ['found'],
+ },
+ render: (_args, value) => [{ type: 'text', text: JSON.stringify(value) }],
+ },
+ execute: async (args, exec) => {
+ const record = readState(cwdOf(exec), args?.slug)
+ if (!record) return { found: false }
+ return {
+ found: true,
+ status: record.state.status,
+ task_slug: record.state.task_slug,
+ awaiting_approval: record.state.awaiting_approval ?? '',
+ state_path: record.path,
+ }
+ },
+ })
+}
diff --git a/local/dsh-plugin/package.json b/local/dsh-plugin/package.json
new file mode 100644
index 00000000..e785e40c
--- /dev/null
+++ b/local/dsh-plugin/package.json
@@ -0,0 +1,18 @@
+{
+ "name": "@tencent/loopforge-dsh",
+ "version": "0.1.0",
+ "description": "LoopForge Local integration for DeepSeek Harness",
+ "type": "module",
+ "main": "./index.js",
+ "exports": {
+ ".": "./index.js",
+ "./cordis.patch.yml": "./cordis.patch.yml"
+ },
+ "files": ["index.js", "cordis.patch.yml", "README.md", "LICENSE"],
+ "dsh": {
+ "bundle": {
+ "patch": "./cordis.patch.yml"
+ }
+ },
+ "license": "MIT"
+}
diff --git a/local/hermes-plugin/README.md b/local/hermes-plugin/README.md
new file mode 100644
index 00000000..7841f1ab
--- /dev/null
+++ b/local/hermes-plugin/README.md
@@ -0,0 +1,57 @@
+# LoopForge Hermes plugin
+
+This native Hermes Agent plugin adds `/loopforge` controls for a project that
+already has the LoopForge Hermes Skill installed:
+
+```text
+/loopforge start <request>
+/loopforge status [slug]
+/loopforge result [slug]
+/loopforge approve [REQUIREMENT|DESIGN|REVIEW|TEST]
+```
+
+The plugin calls the project's `.hermes/skills/devflow/scripts/workflow_state.py`
+contract. It stores state under `artifacts/<slug>/` and does not read or modify
+`~/.hermes/config.yaml`, profile credentials, or the profile skill directory.
+
+## Local installation
+
+From a LoopForge checkout, validate the plugin first:
+
+```bash
+hermes plugins doctor ./local/hermes-plugin
+```
+
+For a local Hermes profile, copy the directory to the profile's plugin folder
+and restart Hermes. The copy is intentionally explicit so an existing plugin
+directory is never overwritten by LoopForge:
+
+```bash
+plugin_dir="${HERMES_HOME:-$HOME/.hermes}/plugins/loopforge"
+if [ -e "$plugin_dir" ]; then
+ echo "Refusing to overwrite existing Hermes plugin: $plugin_dir" >&2
+ exit 1
+fi
+mkdir -p "$(dirname "$plugin_dir")"
+cp -a ./local/hermes-plugin "$plugin_dir"
+```
+
+When this repository is published, install the same directory through Hermes'
+normal Git plugin flow (`hermes plugins install <owner>/<repo>`), then enable it
+with `hermes plugins enable loopforge`.
+
+To update a published installation, use `hermes plugins update loopforge`. To
+disable it without deleting files, use `hermes plugins disable loopforge`; to
+remove it, disable it first and then run `hermes plugins remove loopforge`.
+For a local copied installation, remove only the exact directory you copied
+after disabling the plugin.
+
+Run Hermes from the project root after installing the Skill and trust it:
+
+```bash
+loopforge skills install hermes
+hermes skills trust .
+hermes
+```
+
+The plugin is optional. `/devflow` remains the canonical workflow entrypoint.
diff --git a/local/hermes-plugin/__init__.py b/local/hermes-plugin/__init__.py
new file mode 100644
index 00000000..69c8c680
--- /dev/null
+++ b/local/hermes-plugin/__init__.py
@@ -0,0 +1,213 @@
+"""Hermes Agent command plugin for LoopForge Local.
+
+The plugin deliberately keeps its state in the current project. It calls the
+same ``workflow_state.py`` contract as the ``devflow`` Skill and never reads or
+writes Hermes' profile configuration.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+import subprocess
+import sys
+from pathlib import Path
+from typing import List, Optional, Tuple
+
+
+_SAFE_SLUG = re.compile(r"^[A-Za-z0-9][A-Za-z0-9_.-]{0,127}$")
+_APPROVAL_STAGES = {"REQUIREMENT", "DESIGN", "REVIEW", "TEST"}
+_MAX_RESULT_FILES = 40
+
+
+def _project_root() -> Path:
+ """Resolve the Hermes session project without touching Hermes config.
+
+ Hermes normally starts from the active project directory. The explicit
+ environment override is useful for gateway hosts that keep a different
+ process working directory and makes the behavior testable.
+ """
+
+ override = os.environ.get("LOOPFORGE_PROJECT_ROOT")
+ if override:
+ return Path(override).resolve()
+ try:
+ # Hermes documents this as the public way for plugins to access the
+ # invoking session's working directory. The gateway process cwd is
+ # the Hermes home and is therefore not a safe fallback.
+ from agent.runtime_cwd import resolve_agent_cwd
+
+ return Path(resolve_agent_cwd()).resolve()
+ except (ImportError, RuntimeError, TypeError):
+ return Path(os.getcwd()).resolve()
+
+
+def _workflow_script(root: Path) -> Path:
+ path = root / ".hermes" / "skills" / "devflow" / "scripts" / "workflow_state.py"
+ if not path.is_file():
+ raise RuntimeError(
+ "LoopForge Skill is not installed in this project; run "
+ "`loopforge skills install hermes` first."
+ )
+ return path
+
+
+def _run(root: Path, script: Path, *args: str) -> str:
+ result = subprocess.run(
+ [sys.executable, str(script), *args],
+ cwd=str(root),
+ capture_output=True,
+ text=True,
+ check=False,
+ )
+ output = (result.stdout or result.stderr).strip()
+ if result.returncode:
+ raise RuntimeError(output or f"workflow_state.py exited with {result.returncode}")
+ return output
+
+
+def _state_files(root: Path) -> List[Path]:
+ artifacts = root / "artifacts"
+ if not artifacts.is_dir():
+ return []
+ return sorted(
+ (
+ path
+ for path in artifacts.glob("*/workflow-state.json")
+ if path.is_file()
+ ),
+ key=lambda path: path.stat().st_mtime_ns,
+ reverse=True,
+ )
+
+
+def _record(root: Path, slug: Optional[str] = None) -> Optional[Tuple[Path, dict]]:
+ if slug is not None and not _SAFE_SLUG.fullmatch(slug):
+ raise RuntimeError("Task slug may contain only letters, digits, dots, hyphens, and underscores.")
+ path = root / "artifacts" / slug / "workflow-state.json" if slug else None
+ if path is None:
+ files = _state_files(root)
+ path = files[0] if files else None
+ if path is None or not path.is_file():
+ return None
+ try:
+ return path, json.loads(path.read_text(encoding="utf-8"))
+ except (OSError, json.JSONDecodeError) as exc:
+ raise RuntimeError(f"Cannot read LoopForge state: {path}: {exc}") from exc
+
+
+def _render(record: Optional[Tuple[Path, dict]]) -> str:
+ if record is None:
+ return "No LoopForge task found in this project."
+ path, state = record
+ completed = [
+ name
+ for name, stage in (state.get("stages") or {}).items()
+ if stage.get("status") == "completed"
+ ]
+ lines = [
+ f"Task: {state.get('task_slug', 'unknown')}",
+ f"Status: {state.get('status', 'unknown')}",
+ f"Size: {state.get('size_class', 'unknown')}",
+ f"State: {path}",
+ ]
+ if completed:
+ lines.append("Completed stages: " + ", ".join(completed))
+ if state.get("awaiting_approval"):
+ lines.append(f"Awaiting approval: {state['awaiting_approval']}")
+ if state.get("next_stage"):
+ lines.append(f"Next stage: {state['next_stage']}")
+ return "\n".join(lines)
+
+
+def _result_files(root: Path, slug: str) -> List[str]:
+ artifact_root = root / "artifacts" / slug
+ if not artifact_root.is_dir():
+ return []
+ return sorted(
+ str(path.relative_to(artifact_root))
+ for path in artifact_root.rglob("*")
+ if path.is_file()
+ )[:_MAX_RESULT_FILES]
+
+
+def _parse(raw: str) -> Tuple[str, str]:
+ value = raw.strip()
+ match = re.match(r"^(start|status|result|approve|help)(?:\s+([\s\S]*))?$", value)
+ return (match.group(1), (match.group(2) or "").strip()) if match else ("help", "")
+
+
+def _handle(raw: str) -> str:
+ command, value = _parse(raw)
+ root = _project_root()
+ if command == "help":
+ return (
+ "Usage: /loopforge start <request> | status [slug] | result [slug] | "
+ "approve [stage] | help"
+ )
+ if command == "status":
+ return _render(_record(root, value or None))
+ record = _record(root, value or None) if command == "result" else _record(root)
+ if command == "result":
+ if record is None:
+ return "No LoopForge task found in this project."
+ _path, state = record
+ files = _result_files(root, state.get("task_slug", ""))
+ listing = "\n\nArtifact files:\n" + "\n".join(f"- {item}" for item in files)
+ return _render(record) + (listing if files else "\n\nNo artifact files have been written yet.")
+ script = _workflow_script(root)
+ if command == "approve":
+ if record is None:
+ return "No LoopForge task is waiting for approval."
+ state_path, state = record
+ stage = value.upper() if value else state.get("awaiting_approval")
+ if not stage:
+ return "No approval stage is pending."
+ if stage not in _APPROVAL_STAGES:
+ return "Unknown approval stage. Use REQUIREMENT, DESIGN, REVIEW, or TEST."
+ args = ["approve", "--state", str(state_path), "--stage", stage]
+ if stage == "REQUIREMENT":
+ args.append("--user-confirmed")
+ _run(root, script, *args)
+ return f"Approved {stage}. Resume the Hermes workflow to continue."
+ if not value:
+ return "A request is required: /loopforge start <request>"
+ slug = re.sub(r"[^a-z0-9]+", "-", value.lower()).strip("-")[:48] or "hermes-task"
+ state_path = root / "artifacts" / slug / "workflow-state.json"
+ output = _run(
+ root,
+ script,
+ "init",
+ "--project-root",
+ str(root),
+ "--slug",
+ slug,
+ "--size",
+ "medium",
+ "--mode",
+ "auto",
+ "--execution-mode",
+ "isolated",
+ "--host-adapter",
+ "hermes",
+ "--coordinator-id",
+ "hermes-main",
+ )
+ return (
+ f"LoopForge task initialized: {slug}\nRequest: {value}\nState: {state_path}\n"
+ f"{output}\nExecution mode: the Hermes adapter will use delegate_task when available "
+ "and record a single-context fallback with its reason when it is not. "
+ "Continue with /devflow or resume the task in Hermes."
+ )
+
+
+def register(ctx) -> None:
+ """Register the command on Hermes' documented PluginContext surface."""
+
+ ctx.register_command(
+ "loopforge",
+ _handle,
+ description="Start, inspect, and approve LoopForge Local workflows",
+ args_hint="[start <request>|status [slug]|result [slug]|approve [stage]|help]",
+ )
diff --git a/local/hermes-plugin/plugin.yaml b/local/hermes-plugin/plugin.yaml
new file mode 100644
index 00000000..d2d885e1
--- /dev/null
+++ b/local/hermes-plugin/plugin.yaml
@@ -0,0 +1,9 @@
+name: loopforge
+version: 0.1.0
+description: "LoopForge Local workflow controls for Hermes Agent"
+author: Tencent LoopForge
+kind: standalone
+platforms:
+ - linux
+ - macos
+ - windows
diff --git a/local/package.json b/local/package.json
index ecaf361e..306f18d5 100644
--- a/local/package.json
+++ b/local/package.json
@@ -1,7 +1,7 @@
{
"name": "loopforge-cli",
"version": "0.2.3",
- "description": "LoopForge Local development workflows for CodeBuddy, Codex, Cursor, Claude Code and Pi (Portable)",
+ "description": "LoopForge Local development workflows for CodeBuddy, Codex, Cursor, Claude Code, Pi, DeepSeek Harness and Hermes Agent",
"keywords": [
"ai-agents",
"multi-agent",
@@ -9,7 +9,9 @@
"codex",
"claude-code",
"cursor",
- "codebuddy"
+ "codebuddy",
+ "deepseek-harness",
+ "hermes-agent"
],
"repository": {
"type": "git",
@@ -33,6 +35,8 @@
".codex/",
".cursor/",
".claude/",
+ "hermes-plugin/",
+ "dsh-plugin/",
"README.md",
"LICENSE",
"THIRD_PARTY_NOTICES.md",
diff --git a/local/skills/README.md b/local/skills/README.md
index b269534c..e36312ab 100644
--- a/local/skills/README.md
+++ b/local/skills/README.md
@@ -5,7 +5,7 @@
- `devflow/`:任务分级、需求澄清、方案、计划、实现、审查、测试、知识沉淀和总结。
- `devflow-clarify-requirements/`:需求阶段使用的澄清 Skill,由主协调 Agent 在当前会话中调用。
-脚本只依赖 Python 3.8+ 标准库。当前支持 CodeBuddy、Codex、Cursor 和 Claude Code。
+脚本只依赖 Python 3.8+ 标准库。当前支持 CodeBuddy、Codex、Cursor、Claude Code、Pi、DeepSeek Harness 和 Hermes Agent。
## 流程
@@ -34,8 +34,16 @@ loopforge skills install codebuddy
loopforge skills install codex
loopforge skills install cursor
loopforge skills install claude
+loopforge skills install hermes
```
+Hermes Agent 使用项目级 `.hermes/skills/`,安装后在项目根目录执行
+`hermes skills trust .`,然后使用 `/devflow <软件交付需求>`。Hermes 的 medium/large
+阶段通过公开的 `delegate_task` 工具创建独立子 Agent;缺少该工具时必须记录原因并降级为
+`single-context`。安装器不会改动 `~/.hermes/config.yaml` 或用户 profile。需要独立的
+状态、结果和审批命令时,可按 [`hermes-plugin/README.md`](../hermes-plugin/README.md)
+安装可选原生插件,使用 `/loopforge status|result|approve`;`/devflow` 仍是工作流入口。
+
`loopforge skills status/update/uninstall <host>` 和 `loopforge doctor` 使用
`.devflow/install-state.json` 保护用户文件。下面的 adapter 命令只保留为开发和
故障排查入口,不再是面向使用者的主安装方式。
diff --git a/local/skills/devflow/adapters/dsh/adapter.md b/local/skills/devflow/adapters/dsh/adapter.md
new file mode 100644
index 00000000..96fdf4c5
--- /dev/null
+++ b/local/skills/devflow/adapters/dsh/adapter.md
@@ -0,0 +1,27 @@
+# DeepSeek Harness 适配器
+
+支持当前稳定版以及后续仍提供公开 `subagent` 工具的 DSH 版本;不绑定私有版本号。
+没有该工具的 profile 仍可运行,但 medium/large 必须记录原因并降级为
+`single-context`。
+
+DSH 通过 `subagent` 工具创建独立阶段执行器。每个阶段必须使用新的子 Agent,
+并将工具返回的子 Agent 名称作为 `executor_id` 登记到 `workflow_state.py`。
+
+## 能力映射
+
+1. `spawn(role, prompt) -> executor_id`:调用 `subagent`,使用完整阶段提示;返回的
+ 子 Agent 名称就是 `executor_id`。
+2. `wait(executor_id)`:结束当前轮次,等待 DSH 自动投递子 Agent 完成通知。
+3. `close(executor_id)`:DSH 子 Agent 完成后由宿主回收,不需要额外调用。
+4. `spawn_helper(role, bounded_prompt) -> executor_id`:同样使用 `subagent`,只读范围
+ 由提示明确限制。
+
+DSH 没有 LoopForge 专用的 send/close 生命周期 API。不要伪造 executor ID,也不要把
+普通当前上下文工作标记为 isolated;如果当前会话没有 `subagent` 工具,必须用具体
+原因降级为 `single-context` 并记录 `fallback_reason`。
+
+## 使用入口
+
+安装 Portable Skill 后,在 DSH 中使用 `/devflow <需求>`,或通过 LoopForge DSH 插件的
+`/loopforge start <需求>` 初始化任务。任务状态和产物保存在项目的
+`artifacts/<task_slug>/` 下;需要人工确认时,插件会显示待批准阶段和批准命令。
diff --git a/local/skills/devflow/adapters/dsh/manifest.json b/local/skills/devflow/adapters/dsh/manifest.json
new file mode 100644
index 00000000..692fc5c5
--- /dev/null
+++ b/local/skills/devflow/adapters/dsh/manifest.json
@@ -0,0 +1,14 @@
+{
+ "version": "1.0",
+ "id": "dsh",
+ "compatibility": {
+ "verified_on": "2026-11-09",
+ "version_policy": "Current stable DSH releases exposing the subagent tool; profiles without it use the recorded single-context fallback",
+ "documentation": "https://github.com/deepseek-ai/deepseek-harness"
+ },
+ "dispatch_tool": "subagent",
+ "topology": "spawn",
+ "stage_executor": "subagent",
+ "research_executor": "subagent",
+ "required_tools": ["subagent"]
+}
diff --git a/local/skills/devflow/adapters/hermes/adapter.md b/local/skills/devflow/adapters/hermes/adapter.md
new file mode 100644
index 00000000..c792d8fb
--- /dev/null
+++ b/local/skills/devflow/adapters/hermes/adapter.md
@@ -0,0 +1,40 @@
+# Hermes Agent 适配器
+
+Hermes Agent 通过项目内的 Agent Skills 目录加载 LoopForge:安装器写入
+`.hermes/skills/`,不会修改 Hermes 用户目录或 `config.yaml`。首次在该项目运行
+Hermes 时,需要按宿主的信任提示执行 `hermes skills trust`,之后 `/devflow` 才会
+出现在会话中。
+
+## 版本和能力边界
+
+适配器支持当前稳定版以及后续仍提供项目级 `.hermes/skills/`、`hermes skills trust` 和
+`delegate_task` 的 Hermes Agent 版本,不绑定私有版本号。Hermes 的插件和 skill API
+仍在演进;LoopForge 只依赖这三个公开能力,不读取 Hermes 私有状态文件。若当前会话没有
+`delegate_task`,medium/large 工作流必须记录具体原因并降级为
+`single-context`,不能把当前上下文伪造为 isolated。
+
+## 能力映射
+
+1. `spawn(role, prompt) -> executor_id`:调用
+ `delegate_task(goal=<完整阶段提示>, context=<项目路径和交接信息>)`,将返回值中的
+ `delegation_id`(或宿主提供的等价句柄)登记为 `executor_id`。
+2. `wait(executor_id)`:等待 Hermes 的 delegation 完成通知;不得轮询或猜测句柄。
+3. `close(executor_id)`:Hermes 会在 delegation 完成后回收子 Agent,不要求额外关闭。
+4. `spawn_helper(role, bounded_prompt) -> executor_id`:同样调用
+ `delegate_task`,并在提示中限定只读范围;用返回的 delegation 句柄登记 helper。
+
+REQUIREMENT 和 SUMMARY 仍由协调者执行。每个 DESIGN、IMPLEMENT、REVIEW、TEST
+阶段必须新建独立 delegation;子 Agent 只能通过完整阶段提示和允许的上游产物获得上下文。
+
+## 安装和入口
+
+在项目根目录执行:
+
+```bash
+loopforge skills install hermes
+hermes skills trust .
+```
+
+重启或重新打开 Hermes 会话后使用 `/devflow <需求>`。任务状态和产物保存在
+`artifacts/<task_slug>/`,与其他 Portable 宿主相同;卸载使用
+`loopforge skills uninstall hermes`,只删除 LoopForge 自己未修改的托管文件。
diff --git a/local/skills/devflow/adapters/hermes/manifest.json b/local/skills/devflow/adapters/hermes/manifest.json
new file mode 100644
index 00000000..a74ec490
--- /dev/null
+++ b/local/skills/devflow/adapters/hermes/manifest.json
@@ -0,0 +1,16 @@
+{
+ "version": "1.0",
+ "id": "hermes",
+ "compatibility": {
+ "verified_on": "2026-11-09",
+ "product": "Nous Research Hermes Agent",
+ "version_policy": "Current stable Hermes releases exposing project-local skills, hermes skills trust, and delegate_task; no private-version pin",
+ "capabilities": ["project-local-skills", "hermes-skills-trust", "delegate_task"],
+ "documentation": "https://github.com/NousResearch/hermes-agent"
+ },
+ "dispatch_tool": "delegate_task",
+ "topology": "spawn",
+ "stage_executor": "dynamic",
+ "research_executor": "dynamic",
+ "required_tools": ["delegate_task"]
+}
diff --git a/local/skills/devflow/references/runtime-core.md b/local/skills/devflow/references/runtime-core.md
index 75906f99..5cba1052 100644
--- a/local/skills/devflow/references/runtime-core.md
+++ b/local/skills/devflow/references/runtime-core.md
@@ -18,6 +18,7 @@
- CodeBuddy 提供 `task|team_create|send_message|team_delete|use_skill`:加载 [adapters/codebuddy/adapter.md](../adapters/codebuddy/adapter.md),使用 Team member 拓扑;不得把普通 task subagent 当成阶段执行者。
- Codex 提供 `spawn_agent|send_message|wait_agent`:加载 [adapters/codex/adapter.md](../adapters/codex/adapter.md)。
+- Hermes Agent 提供 `delegate_task`:加载 [adapters/hermes/adapter.md](../adapters/hermes/adapter.md),使用 delegation spawn 拓扑;只把宿主返回的 `delegation_id`(或等价句柄)登记为 executor ID。
- 没有受支持的生命周期 API:先报告限制并记录 `fallback_reason`,之后才能使用 `single-context`。
一个 isolated 流程只能选择一个适配器。除非适配器明确提供桥接,否则不要混用项目原生流程与本状态机。适配器只实现 `spawn/send/wait/close`,不得复制或改写核心状态、模板和规则语义。
diff --git a/local/skills/devflow/scripts/workflow_state.py b/local/skills/devflow/scripts/workflow_state.py
index 8f1cc80f..92424309 100755
--- a/local/skills/devflow/scripts/workflow_state.py
+++ b/local/skills/devflow/scripts/workflow_state.py
@@ -1090,11 +1090,19 @@ def command_resume(args: argparse.Namespace) -> int:
"executor_role": STAGE_ROLES[next_stage],
"executor_type": (
"coordinator" if STAGE_EXECUTION[next_stage] == "coordinator"
- else expected_executor_type(data)
+ else (
+ "single-context"
+ if data.get("execution_mode") == "single-context"
+ else expected_executor_type(data)
+ )
),
"executor_topology": (
"coordinator" if STAGE_EXECUTION[next_stage] == "coordinator"
- else TOPOLOGIES.get(data.get("host_adapter"))
+ else (
+ "single-context"
+ if data.get("execution_mode") == "single-context"
+ else TOPOLOGIES.get(data.get("host_adapter"))
+ )
),
"team_name": (
data.get("team_name") if STAGE_EXECUTION[next_stage] != "coordinator"
@@ -1105,6 +1113,12 @@ def command_resume(args: argparse.Namespace) -> int:
action.update(action="prepare", **common)
elif STAGE_EXECUTION[next_stage] == "coordinator":
action.update(action="start-coordinator", **common)
+ elif data.get("execution_mode") == "single-context":
+ action.update(
+ action="start-single-context",
+ message="宿主没有独立 Agent 生命周期 API;在当前上下文执行本阶段并记录降级原因",
+ **common,
+ )
elif not stage.get("executor_id"):
action.update(
action="spawn-and-assign",
diff --git a/local/src/devflow_cli/cli.py b/local/src/devflow_cli/cli.py
index 574d967d..741edcac 100644
--- a/local/src/devflow_cli/cli.py
+++ b/local/src/devflow_cli/cli.py
@@ -146,6 +146,8 @@ def main(argv=None) -> int:
)
if args.host == "pi" and args.edition == "portable":
print("Note: Pi isolated mode also needs the public extension: pi install npm:pi-subagents")
+ if args.host == "hermes" and args.edition == "portable":
+ print("Note: Hermes must trust this project before loading .hermes/skills: hermes skills trust .")
return 0
if args.command == "skills":
if args.skills_command in {"install", "update"}:
@@ -168,6 +170,8 @@ def main(argv=None) -> int:
)
if args.host == "pi":
print("Note: Pi isolated mode also needs the public extension: pi install npm:pi-subagents")
+ if args.host == "hermes":
+ print("Note: Hermes must trust this project before loading .hermes/skills: hermes skills trust .")
return 0
if args.skills_command == "status":
rows = [
diff --git a/local/src/devflow_cli/core.py b/local/src/devflow_cli/core.py
index 13a3a984..8f7a08b8 100644
--- a/local/src/devflow_cli/core.py
+++ b/local/src/devflow_cli/core.py
@@ -157,6 +157,25 @@ def materialize_portable(host: str, stage: Path) -> None:
for name in manifest["skill_sets"]["portable"]:
copy_skill(skills_root / name, target / name)
return
+ if host == "dsh":
+ target = stage / ".dsh/skills"
+ target.mkdir(parents=True, exist_ok=True)
+ manifest = json.loads((skills_root / "manifest.json").read_text(encoding="utf-8"))
+ copy_skills_manifest(skills_root, target)
+ for name in manifest["skill_sets"]["portable"]:
+ copy_skill(skills_root / name, target / name)
+ return
+ if host == "hermes":
+ # Hermes discovers project skills under .hermes/skills after the user
+ # explicitly trusts the repository. Keep installation project-local;
+ # never mutate ~/.hermes/config.yaml or the user's profile skills.
+ target = stage / ".hermes/skills"
+ target.mkdir(parents=True, exist_ok=True)
+ manifest = json.loads((skills_root / "manifest.json").read_text(encoding="utf-8"))
+ copy_skills_manifest(skills_root, target)
+ for name in manifest["skill_sets"]["portable"]:
+ copy_skill(skills_root / name, target / name)
+ return
raise DevFlowError(f"不支持的宿主: {host}")
diff --git a/local/src/devflow_cli/editions.py b/local/src/devflow_cli/editions.py
index bcc3ead8..2e3d2d75 100644
--- a/local/src/devflow_cli/editions.py
+++ b/local/src/devflow_cli/editions.py
@@ -4,26 +4,30 @@ from typing import Dict, Tuple
DEFAULT_EDITION = "classic"
-HOSTS: Tuple[str, ...] = ("codebuddy", "codex", "cursor", "claude", "pi")
+HOSTS: Tuple[str, ...] = ("codebuddy", "codex", "cursor", "claude", "pi", "dsh", "hermes")
HOST_LABELS = {
"codebuddy": "CodeBuddy",
"codex": "Codex",
"cursor": "Cursor",
"claude": "Claude Code",
"pi": "Pi",
+ "dsh": "DeepSeek Harness",
+ "hermes": "Hermes Agent",
}
EDITION_SPECS: Dict[str, dict] = {
"portable": {
"description": "Agent Skills based portable workflow",
"source_roots": ("skills/",),
- "hosts": ("codebuddy", "codex", "cursor", "claude", "pi"),
+ "hosts": ("codebuddy", "codex", "cursor", "claude", "pi", "dsh", "hermes"),
"entrypoints": {
"codebuddy": "/devflow",
"codex": "$devflow",
"cursor": "/devflow",
"claude": "/devflow",
"pi": "/skill:devflow",
+ "dsh": "/devflow",
+ "hermes": "/devflow",
},
},
"classic": {
diff --git a/local/src/devflow_cli/targets/dsh.py b/local/src/devflow_cli/targets/dsh.py
new file mode 100644
index 00000000..49ff78e9
--- /dev/null
+++ b/local/src/devflow_cli/targets/dsh.py
@@ -0,0 +1,12 @@
+HOST = "dsh"
+
+
+def describe() -> str:
+ return "DeepSeek Harness project workflow target"
+
+
+def materialize(edition, stage) -> None:
+ from devflow_cli.core import materialize_portable
+ if edition != "portable":
+ raise ValueError("DeepSeek Harness 只支持 Portable edition")
+ materialize_portable(HOST, stage)
diff --git a/local/src/devflow_cli/targets/hermes.py b/local/src/devflow_cli/targets/hermes.py
new file mode 100644
index 00000000..8377bbfc
--- /dev/null
+++ b/local/src/devflow_cli/targets/hermes.py
@@ -0,0 +1,12 @@
+HOST = "hermes"
+
+
+def describe() -> str:
+ return "Hermes Agent project workflow target"
+
+
+def materialize(edition, stage) -> None:
+ from devflow_cli.core import materialize_portable
+ if edition != "portable":
+ raise ValueError("Hermes Agent 只支持 Portable edition")
+ materialize_portable(HOST, stage)
diff --git a/local/tests/e2e/fake_agent.py b/local/tests/e2e/fake_agent.py
index 13946000..a24f09b4 100644
--- a/local/tests/e2e/fake_agent.py
+++ b/local/tests/e2e/fake_agent.py
@@ -34,6 +34,8 @@ HOST_SKILL_DIRS = {
"cursor": ".cursor/skills",
"claude": ".claude/skills",
"pi": ".pi/skills",
+ "dsh": ".dsh/skills",
+ "hermes": ".hermes/skills",
}
FIXTURES = Path(__file__).resolve().parent / "fixtures"
EXIT_COMPLETED, EXIT_CONTRACT, EXIT_WAITING, EXIT_BLOCKED, EXIT_INTERRUPTED = 0, 2, 3, 4, 75
@@ -175,11 +177,13 @@ class FakeAgent:
self.log("spawned", stage=action["stage"], executor_id=executor_id,
executor_type=action["executor_type"],
dispatch_tool=action["dispatch_tool"])
- elif kind in ("start-assigned", "start-coordinator"):
+ elif kind in ("start-assigned", "start-coordinator", "start-single-context"):
self.run("start", "--state", self.state_path(), "--stage", action["stage"])
if self.args.crash_in == action["stage"]:
self.log("crash", stage=action["stage"])
return EXIT_INTERRUPTED
+ if kind == "start-single-context":
+ self.log("fallback", stage=action["stage"], reason=self.state().get("fallback_reason"))
self.execute_stage(action["stage"])
elif kind == "continue-executor":
self.execute_stage(action["stage"])
diff --git a/local/tests/e2e/test_e2e_workflow_modes.py b/local/tests/e2e/test_e2e_workflow_modes.py
index e97abc65..6dec87c3 100644
--- a/local/tests/e2e/test_e2e_workflow_modes.py
+++ b/local/tests/e2e/test_e2e_workflow_modes.py
@@ -62,6 +62,39 @@ class WorkflowModesE2E(unittest.TestCase):
self.assertEqual(["REQUIREMENT"], [e["stage"] for e in events(project, "approved")])
self.assert_code_fixed(project)
+ def test_dsh_medium_workflow_uses_the_subagent_adapter(self):
+ project = self.install("dsh")
+ h.fake_agent("dsh", project, check_rc=0)
+ state = h.workflow_state(project)
+ self.assertEqual("completed", state["status"])
+ self.assertEqual("dsh", state["host_adapter"])
+ self.assertEqual("isolated", state["execution_mode"])
+ self.assertTrue(events(project, "spawned"))
+
+ def test_dsh_can_record_single_context_fallback(self):
+ project = self.install("dsh")
+ state_file = h.run_workflow(
+ project, "init", "--project-root", project, "--slug", "dsh-fallback",
+ "--size", "medium", "--mode", "auto", "--execution-mode", "single-context",
+ "--fallback-reason", "subagent tool unavailable",
+ ).stdout.strip()
+ result = h.fake_agent("dsh", project, "--slug", "dsh-fallback", "--resume", check_rc=0)
+ self.assertEqual(0, result.returncode)
+ state = h.workflow_state(project, "dsh-fallback")
+ self.assertEqual("single-context", state["execution_mode"])
+ self.assertEqual("subagent tool unavailable", state["fallback_reason"])
+ self.assertTrue(events(project, "fallback"))
+
+ def test_hermes_medium_workflow_uses_delegate_task_adapter(self):
+ project = self.install("hermes")
+ h.fake_agent("hermes", project, check_rc=0)
+ state = h.workflow_state(project)
+ self.assertEqual("completed", state["status"])
+ self.assertEqual("hermes", state["host_adapter"])
+ self.assertEqual("isolated", state["execution_mode"])
+ self.assertTrue(events(project, "spawned"))
+ self.assertTrue((project / ".hermes/skills/devflow/adapters/hermes/manifest.json").is_file())
+
def test_large_task_runs_the_knowledge_stage_when_it_has_value(self):
project = self.install("codex")
h.fake_agent("codex", project, "--size", "large", "--knowledge", "run", check_rc=0)
diff --git a/local/tests/test_hermes_plugin.py b/local/tests/test_hermes_plugin.py
new file mode 100644
index 00000000..5c623b80
--- /dev/null
+++ b/local/tests/test_hermes_plugin.py
@@ -0,0 +1,68 @@
+import importlib.util
+import json
+import os
+import sys
+import tempfile
+import unittest
+import shutil
+from pathlib import Path
+
+
+ROOT = Path(__file__).resolve().parents[1]
+PLUGIN = ROOT / "hermes-plugin" / "__init__.py"
+
+
+def load_plugin():
+ spec = importlib.util.spec_from_file_location("loopforge_hermes_plugin", PLUGIN)
+ module = importlib.util.module_from_spec(spec)
+ spec.loader.exec_module(module)
+ return module
+
+
+class HermesPluginTests(unittest.TestCase):
+ def test_register_uses_public_slash_command_api(self):
+ module = load_plugin()
+ calls = []
+
+ class Context:
+ def register_command(self, name, handler, description="", args_hint=""):
+ calls.append((name, handler, description, args_hint))
+
+ module.register(Context())
+ self.assertEqual("loopforge", calls[0][0])
+ self.assertIn("LoopForge", calls[0][2])
+ self.assertEqual("[start <request>|status [slug]|result [slug]|approve [stage]|help]", calls[0][3])
+
+ def test_commands_use_project_state_and_leave_profile_unmodified(self):
+ module = load_plugin()
+ with tempfile.TemporaryDirectory() as temporary:
+ project = Path(temporary)
+ skill_root = project / ".hermes/skills/devflow"
+ source = ROOT / "skills"
+ shutil.copytree(source, skill_root.parent)
+ profile = project / ".hermes/config.yaml"
+ profile.write_text("user_setting: keep\n", encoding="utf-8")
+ old_root = os.environ.get("LOOPFORGE_PROJECT_ROOT")
+ os.environ["LOOPFORGE_PROJECT_ROOT"] = str(project)
+ try:
+ started = module._handle("start add tests")
+ self.assertIn("LoopForge task initialized", started)
+ status = module._handle("status")
+ self.assertIn("Task: add-tests", status)
+ result = module._handle("result")
+ self.assertIn("- workflow-state.json", result)
+ finally:
+ if old_root is None:
+ os.environ.pop("LOOPFORGE_PROJECT_ROOT", None)
+ else:
+ os.environ["LOOPFORGE_PROJECT_ROOT"] = old_root
+ self.assertEqual("user_setting: keep\n", profile.read_text(encoding="utf-8"))
+ state = json.loads(
+ (project / "artifacts/add-tests/workflow-state.json").read_text(encoding="utf-8")
+ )
+ self.assertEqual("hermes", state["host_adapter"])
+ self.assertEqual("isolated", state["execution_mode"])
+
+
+if __name__ == "__main__":
+ unittest.main()
diff --git a/local/tests/test_installer.py b/local/tests/test_installer.py
index 5dd10c73..138daa98 100644
--- a/local/tests/test_installer.py
+++ b/local/tests/test_installer.py
@@ -174,6 +174,28 @@ class InstallerTests(unittest.TestCase):
with self.assertRaises(DevFlowError):
apply_install(project, "pi", "classic")
+ def test_dsh_installs_portable_skills_and_rejects_classic(self):
+ with tempfile.TemporaryDirectory() as td:
+ project = Path(td)
+ installed, _unchanged = apply_install(project, "dsh", "portable")
+ self.assertGreater(installed, 0)
+ self.assertTrue((project / ".dsh/skills/devflow/SKILL.md").is_file())
+ self.assertTrue((project / ".dsh/skills/manifest.json").is_file())
+ self.assertTrue((project / ".dsh/skills/devflow/adapters/dsh/manifest.json").is_file())
+ with self.assertRaises(DevFlowError):
+ apply_install(project, "dsh", "classic")
+
+ def test_hermes_installs_project_skills_and_rejects_classic(self):
+ with tempfile.TemporaryDirectory() as td:
+ project = Path(td)
+ installed, _unchanged = apply_install(project, "hermes", "portable")
+ self.assertGreater(installed, 0)
+ self.assertTrue((project / ".hermes/skills/devflow/SKILL.md").is_file())
+ self.assertTrue((project / ".hermes/skills/manifest.json").is_file())
+ self.assertTrue((project / ".hermes/skills/devflow/adapters/hermes/manifest.json").is_file())
+ with self.assertRaises(DevFlowError):
+ apply_install(project, "hermes", "classic")
+
def test_editions_and_plan_make_the_split_visible_without_writing(self):
output = StringIO()
with redirect_stdout(output):