DSH HUB
首页插件商店插件包社区排行榜资源发布指南
插件源码
返回插件目录

winghv /

winghv/dsh-acp-activity

仅 Topic 仓库

Community ACP automation server for DeepSeek Harness with committed tool-activity frames (tool_call/tool_call_update) — dsh-plugin

★ 0 Stars0 Forks0 IssuesN/A 社区评分0 已确认安装
查看 GitHub
README来源: main@7a46fc1f

@winghv/dsh-acp-activity

English | 中文

Community automation-only Agent Client Protocol server for DeepSeek Harness over JSON-RPC stdio, with committed tool-activity frames: automation clients see tool_call / tool_call_update frames alongside assistant messages, so a bash run or a file edit is visible on the wire instead of only in the session log.

This is a community alternative transport to @deepseek-ai/dsh-acp for deployments that need tool visibility. It is derived from DeepSeek Harness's packages/acp/acp and packages/examples/acp-demo (MIT), and is not an official DeepSeek AI package.

Install

npm install -g @winghv/dsh-acp-activity

Requires Node ^22.19.0 || >=24.0.0 and a Harness runtime composition (cordis.yml) providing the agent spine, model adapter, tools, and persistence — the same composition shape @deepseek-ai/dsh-acp-demo expects.

Usage

dsh-acp-activity --config ./cordis.yml

The config schema matches @deepseek-ai/dsh-acp-demo (provider, model, persona, persistenceRoot, workspaceContext, ...). In your cordis.yml, mount this package's plugin (acp-activity) in place of acp-demo; every other leaf stays the same. Stdout carries only ACP JSON-RPC frames — never attach a stdout logger.

Tool activity contract

Source Wire frame
committed tool/call one tool_call: toolCallId = session call id, title = tool name, status: 'in_progress', kind = conservative name heuristic (execute/search/edit/read/delete/move/fetch/think, else other), rawInput = reparsed model arguments JSON (dropped when invalid)
committed tool/result one tool_call_update: status = completed/failed (from the result error identity), text/image result blocks as verified content entries

Tool frames ride the same ordered output chain as assistant messages and are observational: delivery failure warns without settling or failing the owning prompt, because the model-facing outcome of a tool is still its tool result. Raw deltas, reasoning, plans, titles, usage, and tool-private meta presentation payloads stay off the wire.

Everything else — fresh-session semantics, one-shot permission negotiation, image admission, cancellation, disposal ladder, one-connection-many-sessions — matches @deepseek-ai/dsh-acp at the pinned upstream version (see peerDependencies).

Versioning and upstream tracking

The version tracks the upstream release it is ported from: 0.1.1-rc.2.1 = upstream 0.1.1-rc.2 + this package's first patch set. On every upstream bump, diff packages/acp/acp and packages/examples/acp-demo, port non-tool changes, run the ported test suite, then publish the next -N revision.

Development

npm install
npm test        # vitest — ported upstream bridge specs plus tool-activity specs
npm run build   # esbuild bundle into lib/

License

MIT — see LICENSE. Derived from deepseek-ai/deepseek-harness (MIT).

—/ 5

暂无评分

需要先验证清单

Commit 7a46fc1f0f75

社区评论

还没有评论,来写第一条。

DSH HUB

社区维护的 DSH 插件索引。不是 GitHub 或 DeepSeek AI 的官方产品。

社区资源API关于