dsh-acp-agentv0.3.0
goodjin/dsh-acp-agent
dsh plugin --profile web add github:goodjin/dsh-acp-agent#70efcc17fd691e299b6f8009a7827df993ddddfe
介绍
DSH 插件:把 Codex / Claude Code / Grok / Cursor / Qoder / ZCode(GLM) / MiniMax / DeepSeek 接进 DeepSeek Harness,内置各家额度胶囊与账户余额显示
标签
概述
@goodjin2/dsh-acp-agent
A DeepSeek Harness plugin that connects external coding agents to DSH's native session interface. Supports GLM (ZCode), MiniMax, Codex, Claude, Grok, Cursor, and Qoder.
Developing or debugging this plugin? The host contract (preset declaration rows, client rows, quota routes, the ACP agent framework) is defined by the official DeepSeek Harness documentation. A complete offline Chinese mirror lives at ../docs/official/README.md (with a task-oriented navigation table); the online version is at .
What it does
Host side (loads in the DSH process):
- ACP agent framework — transports (stdio, WebSocket), per-agent dialects, frame projection, terminal & file sub-protocols
- GLM (ZCode) model client — reads the local ZCode coding-plan key, calls GLM Coding Plan, streams responses
- MiniMax model client — reads/refreshes MiniMax Code credentials, calls managed gateway
- Preset dispatch — registers external agents as DSH presets, takes over sessions, routes turns, handles model switching
- Quota endpoints — host-side GLM/MiniMax/Codex/Claude/Grok quota routes (credentials never leave the host process)
- Roster filter — hides marked presets from the new-session picker
- Mermaid asset — serves the Mermaid rendering runtime on demand
Browser side (./client entry, auto-loaded by DSH Web via the dsh.client declaration):
- Quota pills — the GLM/MiniMax/Codex/Claude/Grok pills under the composer, polling the host routes
- Model visibility panel — settings section to hide providers and models from the pickers
- Model label — the composer's model chip
- Switch feedback — handover notice and switch progress bar when the conversation moves to a new session
- Turn recovery — the expand button for interrupted no-text turns; resend for requests that never entered the transcript
- ACP process view — data projection for external agent turns (
acpProcessSegmenton assistant steps). Theassistant-stepshadowing wrapper was removed on both sides (2026-11-13): the native AssistantNodeView declares child seats (conversation.chat.reasoning.body,conversation.chat.message.images), and a shadowing wrap cannot legally render it — renderSlot checks the wrapping entry's children table and the native seats cannot be re-declared. ACP segment rendering will return via a legal seat (e.g. registering intoconversation.chat.reasoning.body).
MiniMax login: refresh, and what happens when it can't
The MiniMax client borrows the local MiniMax Code login
(~/.minimax/auth/<buildEnv>/<region>/mcode-public/auth.json) and refreshes it through
the same official OAuth core — same credential store, same cross-process lock — so a
refresh here never invalidates the desktop app or the CLI.
When the refresh token itself is dead, retrying is pointless, so the plugin now says where to authorize instead of failing quietly:
resolveCredentialraisesMiniMaxLoginRequiredError(reason: 'missing' | 'expired');- the quota route (
GET /acp-agent/minimax/quota) answers{ ok: false, needsLogin: true, login: { url, userCode?, expiresInSec? } }; - a model request that cannot authenticate puts the authorization link into the error it streams into the conversation, so the link is clickable where the failure shows up.
The link comes from the same device-authorization flow the desktop app uses
(POST /oauth2/device/code → verification_uri_complete). The plugin starts one in the
background and waits at most 3 s for the prompt, because while MiniMax Code is running it
holds the auth lock and our login has to queue: the first response may carry just the
account origin, and the next poll (30 s later) the complete link with the user code. Once
the user authorizes, the official core writes the new credential back to the shared
auth.json and the next call simply works. Tokens never enter these payloads, log lines
or error strings.
Where the GLM (ZCode) key comes from
The GLM client reads the coding-plan key from the ZCode desktop app's own files, and it reads them per request — never from a cache, because the app rewrites those files when the user signs in again or switches plans, and a cached key is a key that no longer works.
Two places hold it, the newer one first:
~/.zcode/v2/config.json— the app's provider configuration. The coding-plan provider entry carries the key in plain text underoptions.apiKey, right next to theoptions.baseURLthat entry serves. The entry is picked by the gateway it serves, not by its name: the same file also carries the start-plan entry, whose value is a short-lived JWT for a different gateway.~/.zcode/v2/credentials.json— the encrypted credential file, entryaccount-provider:coding-plan:…:api-key. Older apps wrote the key there. The current app (3.17.2, checked 2026-11-06) writes that entry no more; the key lives only in the provider configuration.
ZCODE_DATA_BASE_DIR moves both files; ZCODE_ACP_CONFIG_FILE and
ZCODE_ACP_CREDENTIALS_FILE point at them one by one. When neither place holds a key, the
error names both files and the gateway — never the key itself.
When a row doesn't mount: the self-report seam
A module that cannot be imported takes its whole row down with it, and it cannot report that
itself. So this package ships one leaf module (./health, row acp-agent-health) that
imports nothing from the package — it still loads and speaks while everything else is broken.
After boot it prints one line naming what is missing and why:
acp-agent health: 12/13 行已挂载;未挂载 1 行 — acp-agent-minimaxi-quota(MiniMax 套餐额度接口):SyntaxError: TypeScript parameter property is not supported in strip-only mode
and serves the same ledger as JSON at GET /acp-agent/health:
{ ok, mounted, total, disabled, rows: [{ id, what, status, detail }] }. status is one of
mounted, failed (module never imported, or apply threw), pending (waiting on injected
services), absent (the composed row never reached the loader tree) or disabled (turned off
on purpose — not a failure). No tokens or credentials are in that payload.
The same line is appended, timestamped, to $DSH_HOME/logs/acp-agent-health.log — one line
per boot. That file is the point of the exercise: the desktop app hands the host's stderr to
nobody, so after an incident this is what you can read without rerunning anything.
Why it exists: on 2026-11-03 a constructor parameter property in src/minimaxi/credential.ts
made every row importing it fail under DSH's strip-only TypeScript loader. The host half and
the browser half vanished together with nothing visible in any log; the cause was only found
by probing whether a quota route existed. Two gates now stand in front of that: the audit row
above, and tests/entry-loadability.spec.ts, which imports every host entry with a real
node (no transform) before a change can ship.
0.3.0 — what moved out
Generic conversation-area UI moved to @goodjin2/dsh-chat-area
in 2026-11: the answer card, the conversation file cards, the Mermaid renderer and
its host route, and the short session-ID bar. This package keeps the ACP side and
no longer exports ./mermaid-asset — if your profile still lists
@goodjin2/dsh-acp-agent/mermaid-asset, replace it with
@goodjin2/dsh-chat-area/mermaid-asset (or drop it; chat-area's own bundle patch
declares it).
Browser layering: neither package wraps assistant-step anymore (2026-11-13).
The native renderer declares child seats, and a shadowing wrap cannot legally
re-declare them (already declared) nor render without them
(SlotOwnershipError on the native's first renderSlot call). The old layering
(chat-area -50, this package -40 fallback) is gone with the wraps.
Install
From GitHub Packages:
# Set up auth (one-time)
echo "//npm.pkg.github.com/:_authToken=YOUR_GITHUB_TOKEN" >> ~/.npmrc
echo "@goodjin:registry=https://npm.pkg.github.com" >> ~/.npmrc
dsh plugin --profile web add @goodjin2/dsh-acp-agent
Or add to your profile's package.json:
{
"dependencies": { "@goodjin2/dsh-acp-agent": "^0.1.0" }
}
Then add the agent presets you want to your profile's cordis.patch.yml. This repo ships its own cordis.patch.yml with the seven agent preset rows, the browser-half roster row (acp-agent-client, a bare package-name row — DSH only treats rows named exactly that way as client entries), the host-side roster-filter / zcode-quota / minimaxi-quota / codex-quota / claude-quota mounts (moved here from the threadharbor umbrella in 2026-11), and the health self-report row. See the installation guide for details.
Requirements
- DSH ≥ 0.2.1-rc.2
- Node ≥ 22.19 (or ≥ 24)
- For GLM: ZCode desktop app installed and logged in
- For MiniMax: MiniMax Code installed and logged in
Publish
gh auth refresh -h github.com -s write:packages
NODE_AUTH_TOKEN=$(gh auth token) npm publish --registry https://npm.pkg.github.com
License
MIT
Qoder 国际版
新增独立模型预置(DSH 执行工具),在设置的「Qoder 国际版账号」中浏览器授权。国内版 ACP 预置保留。模型清单登录后由宿主从 Qoder 国际版拉取,选择器以那份为准;真实授权回合尚待验证。详见 拆解、协议与验证边界。
Version anchor tags
This repo participates in the container's DSH version-correspondence mechanism: every
plugin repo carries tags named dsh-<dsh版本> (e.g. dsh-0.2.2-rc.1), each an annotated
tag recording the dsh version it was verified against, the check result, and the date.
The tag is the immutable "known-good coordinate" of that dsh version — a snapshot of
main taken the moment the whole-plugin row check went green on every installed host.
- Why tags instead of releases: the desktop app reads this checkout by path and the web install pulls
repo#ref— a git tag is already the furthest installable form, so no npm publish or GitHub Release is needed. - Rollback:
scripts/pin-to-dsh.sh 0.2.2-rc.1(from the container) checks out the tag in every repo at once;scripts/check-host-rows.mjs --reportshows which repos are on the anchor for the running dsh version. - Do not move or delete a
dsh-*tag; it is a rollback coordinate. A follow-up fix on the same dsh version adds a-p2(then-p3) suffixed tag instead. - New tags are cut after each dsh upgrade, as step 4 of the upgrade runbook — see the container's
AGENTS.mdfor the full flow.
星级评价
基本信息
插件信息来自公开 GitHub 仓库与 SkillHub,版权归原作者所有;安装前请自行审查仓库代码。