dsh-coding-agents

Drive your local Codex and Claude Code from DeepSeek Harness — sessions, config and run state, from one settings page and six agent tools.

English | 中文

The plugin calls the public, stable local interfaces of Codex and Claude Code. It never parses or rewrites their private databases, and every write goes through the vendor's own configuration protocol or a backup-plus-atomic-replace.

Requirements

Item	Value
DSH	0.2.1-rc.5 (Cordis `4.0.5`, `@deepseek-ai/*` 0.2.1-rc.5)
Node.js	`^22.19.0 \|\| ^24.0.0`
Host services	`webServer`, `subprocess`, `tools`
Client services	`slots` (registers one section under Settings → Coding agents)
Vendor CLIs	Your own Codex CLI `>= 0.100.0` / Claude Code via Agent SDK `>= 0.3.0` (both optional peers)

Use your own Codex & Claude Code

The plugin never pins a vendor version. Each engine resolves independently, in order: config.codexCommand / config.claudeCommand → system PATH (codex / claude) → the CLI bundled in this package → unavailable with an actionable message. The status page reports the source (config / system / bundled) and the version read from <cli> --version.

Global install: install Codex / Claude Code normally (e.g. npm i -g @openai/codex); the plugin drives the CLI already on your PATH, including future upgrades.
Install it in this profile: dsh plugin --profile <name> add @openai/codex (or pnpm add @anthropic-ai/claude-agent-sdk) — both are optional peers, so the plugin never dictates your version.
Point at a specific build in the profile's cordis.patch.yml:

- id: dsh-coding-agents
  config:
    codexCommand: /opt/homebrew/bin/codex
    claudeCommand: /Users/me/.local/bin/claude

Without the Claude Agent SDK the plugin still loads and the Codex engine keeps working; the Claude Code row simply reports unavailable with the SDK install hint.

Install

The package ships cordis.patch.yml and a dsh.bundle.patch declaration, so it installs as a bundle:

dsh plugin --profile <profile> add dsh-coding-agents

Or add one loader entry to the profile's cordis.patch.yml:

- insert:
    - id: dsh-coding-agents
      name: 'dsh-coding-agents'

Restart the profile, then open Settings → Coding agents.

What you can do

Search the local session history of both Codex and Claude Code, and page through very long transcripts.
Start, continue or fork a native session, watch text, tool calls and run state live, and cancel a run.
Rename, tag, archive, unarchive, roll back and delete sessions. Operations that do not apply to an engine are refused explicitly instead of failing silently.
Inspect configuration values, their effective values and their origin. Codex writes through its official config protocol; Claude Code uses backup plus atomic replace.
The same capabilities are exposed to the DSH agent as six coding_agent_* tools, not only through the settings page.

Natural-language examples

"List the most recent Codex and Claude Code sessions."
"Find the history that mentions the login failure."
"Continue that Codex session and run the tests."
"Start a Claude Code task in this workspace and fix the build error."
"Read the current Codex configuration."

The plugin only drives engines that the current user has installed and logged in. Runs default to workspace-write permissions, and full access is never reachable from a web page or from a model tool: it only happens when you start the run yourself in the DSH UI.

Safety boundaries

Loopback + same-origin HTTP only. /coding-agents-api rejects any request whose TCP peer is not this machine (403), any browser request whose Origin is not exactly the http://<host> it was sent to (403), any request whose Sec-Fetch-Site is cross-site (403), every POST without content-type: application/json (415), and CORS preflight (OPTIONS, 405). A random web page therefore cannot drive your local Codex or Claude Code, even with a "simple" request that skips the CORS preflight.
No full access through HTTP. The HTTP surface accepts only read-only and workspace-write; full-access (Claude Code bypassPermissions, Codex danger-full-access) is reachable only from the DSH UI, where you confirm it yourself. The model tools mirror this ceiling: coding_agent_run refuses full-access with an actionable error.
Registered workspaces bound every cwd. A cwd is accepted only when it resolves inside a workspace registered in DSH; paths are normalised and traversal outside a registered root is refused. When the host exposes no workspace registry, requests carrying a cwd are refused instead of trusting the caller.
Configuration writes are allowlisted. HTTP may write only model and model_reasoning_effort (Codex) or permissions.defaultMode (Claude Code). Every other key — MCP servers, hooks, permission presets, arbitrary nesting — needs the DSH UI or a manual edit. Both engines back the file up before a write (Claude Code: backup + atomic replace; Codex: config.toml.backup-dsh-<timestamp> before its official config protocol write).
No vendor private database is read; only the public local interfaces.
Child processes live in the host's subprocess service and inherit a scrubbed environment (scrubbedParentEnv).
Session operations stay scoped to the workspace that owns the session.

Development

npm install
npm run typecheck   # source + client + tests
npm run build       # tsc → lib/, esbuild → lib/client.js
npm test            # vitest
npm run smoke:host  # load lib/index.js in a real cordis context with stubbed host services

lib/client.js is a window.__ModuleLoader__.load({ id, factory }) closure factory, statically verified and executed in a sandbox by scripts/verify-client.mjs. The host contract is verified by scripts/verify-host.mjs: a real @deepseek-ai/cordis context loads the module, registers the route and tools, and cleans them up on dispose.

Known limitations

Requires a local Codex / Claude Code installation and login. Without it the engine reports an explicit unavailable state.
Native executables are resolved on first use. If a package manager skipped the platform optional dependency, only that engine reports an actionable "unavailable" state (status page, coding_agent_status, run start) while the plugin itself still imports and the other engine keeps working; reinstall with optional dependencies enabled to restore it.
The bundled CLI fallback (when present) is large, so the first install depends on network speed; using your own global CLI avoids it entirely.
The HTTP API is a loopback http:// surface by design: a TLS or non-loopback DSH web deployment is refused rather than downgraded.
The plugin drives the vendor CLIs; it does not reimplement their agent loops.

dsh-coding-agents

Install

README