dsh-plugin-cc

English | 简体中文

A Claude Code plugin marketplace that bridges to the DeepSeek Harness (dsh) agent: code review, adversarial critique, task delegation, background runs, and multi-turn resumable dsh sessions — all from Claude Code slash commands.

Built against a source checkout of DeepSeek Harness 0.1.0-rc.5 (developer preview; the CLI ships as @deepseek-ai/dsh on npm, but the SDK server this plugin needs is outside that package's dependency closure). The exact DSH behaviors this plugin depends on are pinned in docs/dsh-compat.md; re-verify that table when upgrading dsh.

Quick start

Plugin commands require Node >= 20 and a DEEPSEEK_API_KEY. The one-command source build also requires git, Node >= 22.19, and pnpm (or corepack enable) because those are DeepSeek Harness build requirements.

# 1. Install the plugin
/plugin marketplace add cpj-dev/dsh-plugin-cc
/plugin install dsh@deepseek-dsh

# 2. One command, one time: /dsh:setup does everything
#    (the CLI is on npm as @deepseek-ai/dsh; setup still clones a pinned
#     source checkout because the cc profile's SDK server is not in the
#     CLI dependency tree. First run takes a few minutes.)
/dsh:setup

# 3. In any git repository:
/dsh:check          # readiness probe
/dsh:review         # read-only review of your local changes

Have your own deepseek-harness checkout (any location, any version)? /dsh:setup --harness <path> uses it instead of cloning. Have a built dsh already? DSH_BINARY selects that executable; plain /dsh:setup may still clone the verified source so it can install the separately published SDK server (@deepseek-ai/dsh-sdk-jsonrpc-server, outside the CLI dependency closure) from that pinned checkout. Uninstalling: remove the plugin, the plugin data directory (setup's clone lives there), and ~/.dsh/profiles/cc.

Commands

Command	What it does	Needs setup?
/dsh:check	Readiness probe (dsh, source checkout, credentials, profile, broker)	no
/dsh:setup	Install/link dsh from a source checkout (--harness <path>) and create the multi-turn cc profile	—
/dsh:review [focus]	Read-only code review of local changes	no
/dsh:critique [focus]	Structured adversarial design critique	no
/dsh:run <task>	Run a task (read-only by default; --write, --session, --resume, --model, --effort, --background)	only for --session/--resume
/dsh:delegate <task>	Background delegation via the dsh-delegate subagent	no
/dsh:import	Transfer this conversation into a resumable dsh session	yes
/dsh:runs [id]	List runs / one run's status	no
/dsh:show [id]	Stored result of a finished run	no
/dsh:stop [id] / --broker	Kill a run's process tree / the shared broker	no

Full command semantics: docs/commands.md. Installation and runtime failures are covered in docs/troubleshooting.md.

Project scope — what each part is and how to use it

.claude-plugin/marketplace.json   Marketplace manifest. Users point /plugin marketplace add here.
plugins/dsh/                      The single plugin this marketplace ships.
  .claude-plugin/plugin.json      Plugin identity; bump `version` on every release.
  commands/*.md                   Slash-command surfaces. Each file = one /dsh:* command; the body
                                  tells Claude exactly which bridge invocation to run and how to
                                  present the output. Edit these to change UX wording, never logic.
  agents/dsh-delegate.md          Background-delegation subagent. Claude invokes it for substantial
                                  tasks; it drives the bridge and reports the run id/result.
  skills/                         Internal skills (progressive disclosure for Claude itself):
    dsh-delegate-runtime/         the bridge call contract — read by the agent before first use
    dsh-run-output/               presentation rules for bridge results
  hooks/hooks.json                SessionStart/SessionEnd wiring. Start exports DSH_CC_SESSION_ID /
                                  DSH_CC_TRANSCRIPT_PATH via CLAUDE_ENV_FILE; End cancels this
                                  session's runs. No edits needed unless Claude Code's hook API moves.
  scripts/dsh-bridge.mjs          The only entry point commands call. One subcommand per capability;
                                  stdout is user-facing. Run `node scripts/dsh-bridge.mjs` for usage.
  scripts/dsh-broker.mjs          Per-workspace daemon owning one live `dsh --profile cc` SDK runtime
                                  — the only way DSH sessions can span multiple turns. Started on
                                  demand by the bridge; never start it manually.
  scripts/session-lifecycle-hook.mjs  Hook implementation behind hooks.json.
  scripts/lib/                    Implementation layers; dsh.mjs is the only file that knows how to
                                  invoke DeepSeek Harness, broker-client.mjs the only one that talks
                                  to the broker. Everything else is plumbing (state, jobs, git, render).
  prompts/*.md                    Prompt templates ({{VAR}} interpolation) for critique and import.
  schemas/review-output.schema.json  JSON contract embedded into critique prompts.
docs/                             The documentation set (see reading order below).
tests/                            node:test suite with a fake `dsh` fixture; `npm test`.

Documentation — progressive disclosure

Read in this order; stop at the layer you need.

1. This README — what exists and how to invoke it.
2. docs/commands.md — full user-facing command semantics and flags.
3. docs/troubleshooting.md — setup, credentials, profile, broker, and timeout recovery.
4. docs/architecture.md — the two drive paths (one-shot headless vs broker) and why each design decision was forced by a DSH fact.
5. docs/dsh-compat.md — the pinned DSH behavior table; the first thing to re-verify on a dsh upgrade.
6. Deep dives: docs/broker.md, docs/state-and-jobs.md.
7. Contributing: docs/development.md, docs/testing.md.

docs/README.md is the maintained index with per-document ownership rules. Simplified Chinese user documentation starts at docs/zh-CN/README.md.

Known limitations (v1)

• No mid-run approvals: permissions are decided before launch (--write or not). DSH's interactive approval seam needs a UI the plugin doesn't have.
• Fresh one-shot runs are not resumable; only broker-backed runs (--session, --resume, /dsh:import) record session ids, and those sessions live only as long as the broker process.
• Stop = kill: the DSH SDK wire has no per-turn cancel; stopping a mid-turn broker run discards the broker's in-memory sessions.
• /dsh:import is a weak import (compressed text digest), not a native history replay.
• POSIX only (unix sockets, pgrep); Windows is out of scope for v1.

License

MIT — see LICENSE. Design provenance is credited in NOTICE.

Community and support

• Read CONTRIBUTING.md before opening a pull request.
• Use SUPPORT.md for support boundaries and help channels.
• Report vulnerabilities through the private process in SECURITY.md.
• Participation is governed by CODE_OF_CONDUCT.md.