- Home
- Plugins
- Workflow & Automation
- mstar-harness (dsh)
mstar-harness (dsh)
btspoony/mstar-harness/packages/dsh
Skill-driven harness/loop engineering workflow agent plugin.
Install
dsh plugin --profile web add github:btspoony/mstar-harnessThis plugin lives in the packages/dsh subdirectory of the repository.
README
Morning Star brings a delivery process to the AI coding tools you already use — dsh, omp, OpenCode, Cursor, Kimi Code, Devin, ZCode, and Codex. You describe what you want; a project-manager agent clarifies the request, keeps a plan, and carries the work from requirements through implementation, review, and acceptance to a pull request. Requirements, architecture, implementation, review, acceptance, and audits each go to a role that owns that part.
Why use it
- From request to pull request — describe the work once; the PM keeps the plan, brings in the right specialists, and drives everything to a pull request. If work stops partway, the plan records where it left off so the work can resume later.
- Specialists instead of one long prompt — a PM orchestrates; separate roles own requirements, architecture, implementation, QC, acceptance, audits, and ops, each with a narrow brief.
- A trail you can audit — plans, findings, review reports, and decisions are recorded under the harness directory in your repo, not just in the chat.
Install
Install through the host's native plugin manager, or this CLI's init flow where supported; Kimi and Devin are native-only. Native commands install the plugin only. Options, path layout, and host notes: INSTALL.md.
- dsh (DeepSeek Harness) — needs the
dshCLI onPATH- CLI:
npx @mstar-harness/cli init --target dsh - Native:
dsh plugin --profile web add @mstar-harness/dsh
- CLI:
- omp — needs the
ompCLI installed- CLI:
npx @mstar-harness/cli init --target omp - Native:
omp plugin install @mstar-harness/omp
- CLI:
- OpenCode
- CLI:
npx @mstar-harness/cli init --target opencode - Native:
@mstar-harness/opencodefor OpenCode 1.x,@mstar-harness/opencode-v2for OpenCode 2.x — add it to the plugin list inopencode.json, then restart OpenCode
- CLI:
- Cursor — needs
git- CLI:
npx @mstar-harness/cli init --target cursor - Native:
git clone https://github.com/btspoony/mstar-harness.git ~/.cursor/plugins/local/morning-star-harness, then Developer: Reload Window
- CLI:
- Kimi Code — no CLI target
- Native:
/plugins install https://github.com/btspoony/mstar-harness, then/plugins reload
- Native:
- Devin — native-only, no CLI target
- Native: Customize > Plugins, add
btspoony/mstar-harness, ordevin plugins install btspoony/mstar-harness
- Native: Customize > Plugins, add
- ZCode
- CLI:
npx @mstar-harness/cli init --target zcode - Native: add
https://github.com/btspoony/mstar-harnessas a ZCode plugin marketplace, then install morning-star-harness
- CLI:
- Codex — needs the
codexCLI- CLI:
npx @mstar-harness/cli init --target codex - Native:
codex plugin marketplace add btspoony/mstar-harness --ref main, thencodex plugin add morning-star-harness@mstar-repo
- CLI:
Use
Enter the PM once per session, then describe the work in your own words; the PM runs the process in that session.
| Host | Enter PM |
|---|---|
| dsh | pm skill |
| omp | /skill:pm |
| OpenCode | Project Manager agent, or /pm |
| Cursor | /pm |
| Kimi Code | auto-loads in a session, or /skill:pm |
| Devin | /morning-star-harness:pm |
| ZCode | /morning-star-harness:pm |
| Codex | /pm |
Single task (no iteration)
Give the PM a concrete request — for example, "Add rate limiting to the public API and cover it with tests." The PM clarifies the gaps, writes a plan, runs the work on a feature branch, and drives it through implement → independent QC review → acceptance → done. Confirmed problems that are not fixed in this round are captured as issues instead of being dropped. You review the result and merge.
Iteration
| Command | When |
|---|---|
/iteration-start [direction] [pause] | Phase 1: an interactive direction lock (grill-me) that produces the iteration's compass and plans, then continues automatically through execution, close, pull request, and merge-ready. pause stops after Phase 1; continue with /iteration-drive. |
/iteration-drive | Resume or advance an already locked iteration. Takes no arguments. |
/iteration-loop [direction] [scale] | The same lifecycle end to end without the interactive phase. scale (S / M / L / XL) caps how many plans the iteration takes on. |
An iteration ends only after its post-merge close: a closed phase, an open pull request, or a merged pull request is not completion by itself. Command spelling differs per host — some hosts require a plugin prefix, and Codex installs the command skills at project scope only; argument forms and per-host details are in docs/commands.md.
Audit, review & verification
| Command | When |
|---|---|
/codebase-audit [keywords] | Read-only survey that produces prioritized, ready-to-execute improvement plans; narrow it with a category focus (bug, security, perf, tech-debt, …). |
/codebase-comment-cleanup [path|subsystem] [dry-run] | Triage every code comment into delete / ADR / issue / keep-with-shape; the per-PR ratchet only reads comments a diff touches, so an untouched unclassified comment does not trip it — the whole-tree bound gate still fails a pre-existing unbounded marker. |
/amazing-pr-review [pr|branch] [quick|default|deep] | Pre-merge review of a PR or branch at three strengths, ending in one verdict — ship it, needs fixes, or blocked from the finding tally on a settled run (completed without a stop, every required domain opened, the resolved diff basis readable and complete), or the single derived incomplete on exactly three conditions — the run stopped before synthesis, a required domain seat never opened, or the resolved diff basis could not be read. A wall-clock overrun is not one of them: a fully covered run past its budget publishes its tally-derived verdict and declares the overrun as a note. |
/amazing-test-audit [scope] [quick|deep] [campaign] | Read-only audit of the test surface → plans to delete, repair, consolidate, or relocate tests. |
/amazing-e2e-check [environment] [scenarios] | Runs browser, device, or installed-deployment scenarios you explicitly asked for, in a separate workflow. |
The audit commands read and report; they do not edit your code. Two boundaries to know: /amazing-pr-review posts its findings to GitHub as a comment review when you give it a pull-request number — it never approves, requests changes, or merges — and /amazing-e2e-check runs only when you ask for it, with steps that touch a real environment requiring explicit authorization.
Workflow
The main path shows a development delivery. Every plan gets plan-level QC — three independent seats by default — and multi-task plans add per-task review. Hotfix/inline review keeps its lighter seat rules. PM Acceptance is the sole acceptance route for every scope: each required criterion maps to relevant evidence and an honest result; QC approval alone is not test proof. PM may fill a named gap with a short bounded acceptance check, not implementation, test authoring, long suites or ops. Unfixed confirmed problems remain tracked issues.
Scope notes. An iteration locks its direction with you first, then runs the per-plan cycle for every plan it takes on and closes out once before its pull request; without an iteration, the cycle runs for a single plan. A verification or report-only deliverable — an audit, for example — ends at its agreed completion, with no pull request or merge. On the development path, a plan is complete only with its evidence: plan completion, merge-ready, and merged are three different facts, and only a verified merge, followed by the post-merge close, ends the delivery. Merging is a separate authorization with one recorded source: under owner-gated you approve it or merge yourself; under reviewed-auto the recorded policy supplies it only through a settled, posted ship it report at the current PR head; under reviewed-bot only through the attached bot's newest passing review on that head — not the mere existence of a review. An iteration that never recorded a policy returns to Phase 1 instead of falling back to any of them. Branch cleanup stays a separate, explicit step.
The engine checks the mechanically checkable parts of this process — workflow state transitions, branch and worktree alignment, dispatch preconditions, plan and issue bookkeeping. These checks are advisory by default: they report problems, and a project or iteration can opt into blocking enforcement with enforcement: hard; whether a violation can actually block depends on the host. Process state (plans, workflow records, findings, review reports) lives in the harness directory, .mstar/ by default, which the scaffold's canonical .gitignore snippet ignores — the installer does not touch .gitignore.
Dashboard
A read-only page on 127.0.0.1, for this machine only, that brings project progress, open issues, and the roadmap into one view for the repository you start it from. It has Issues, Workflows, Iterations, and Roadmap; on Issues you can search and filter, open an issue's recorded history, and read the issue flow — a cumulative trend of issues captured versus retired. Nothing on the page writes; changes go through the CLI.
mstar-harness dashboard
The command prints a one-line JSON result. Open data.url (http://127.0.0.1: plus the port it actually bound) and press Ctrl-C to stop. --port and --project <projectId> are optional; omit --port and the OS chooses. The page does not push live updates.
Details: docs/runtime-reference.md.
MCP
From a coding tool that supports MCP, look up issues, the roadmap, and workflows, and run the matching CLI operations and checks, without writing a shell command. mstar-harness mcp is the stdio server in this package. It follows the same rules as the CLI — not a new permission — and there is no separate MCP package.
After install, a host usually already has its launch configuration; see INSTALL.md. Static JSON configs use the shape below (npx runs the bin on the Node on PATH, as in Prerequisites). Do not paste it over a file the host already wrote. OpenCode does not read a static file — its plugin injects the server from the config hook. dsh is the exception: the shipped row stays inert until the profile includes @deepseek-ai/dsh-mcp-client (dsh plugin --profile web add @deepseek-ai/dsh-mcp-client).
{
"mcpServers": {
"morning-star": {
"command": "npx",
"args": ["@mstar-harness/cli", "mcp"]
}
}
}
Further detail: docs/runtime-reference.md.
Roles and skills
| Agent ID | Responsibility |
|---|---|
project-manager | Routing, assignment, phase progression, evidence-backed PM Acceptance and sole Done sign-off |
product-manager | Requirements, product planning, research |
architect | Architecture and technical contracts |
fullstack-dev / fullstack-dev-2 | Backend-led implement / second parallel track |
frontend-dev | UI, interaction, frontend performance |
code-reviewer | SDD per-task review; codebase audit (audit category) |
qc-specialist / -2 / -3 | QC trio |
ops-engineer | Deploy, monitoring, infrastructure |
writing-specialist | Docs, fiction, copy, scripts |
prompt-engineer | Prompt / skill / rule work |
A project-manager runs your session, owns acceptance, and dispatches requirements/architecture, implementation, independent review, audits, operations and writing to their specialist roles. Every role follows the mstar-* skills. Acceptance procedure: mstar-roles/references/project-manager/acceptance.md.
Command line
The mstar-harness CLI covers what you do outside a session: issues, roadmap, workflow state, checks, the dashboard, and MCP.
- Ask the command itself.
--helpshows the input a command expects; for issue write verbs it lists each payload field with its type and whether it is required or only required for a particular disposition. A usage refusal tells you what it found wrong, the help route to read, and how to recover.
mstar-harness --help # command families
mstar-harness issue close --help # options and payload fields for one verb
mstar-harness schema --command issue.close # the same contract as machine-readable JSON
mstar-harness reportdrafts an offline, redacted GitHub issue report for you to review before posting.
Host wiring, the issue and roadmap command families, and the details behind these surfaces are in docs/runtime-reference.md.
Documentation
docs/runtime-reference.md— components, storage layout, roles and skills, CLI and MCP reference.INSTALL.md— install and verification per host, plus thereportcommand in detail.docs/commands.md— the eight slash commands: arguments, when to use, owning skill.CONCEPTS.md— the vocabulary the harness uses across skills and docs.- Release notes:
CHANGELOG.md/CHANGELOG_CN.md. CLI flags: themstar-use-cliskill. - Maintainers:
AGENTS.md.
License
MIT. See LICENSE.
Related plugins
webnovel-writer
lingfengqaq/webnovel-writer
loopx (dsh-loopx-plugin)
loopx-project/loopx
ouroboros (dsh-plugin)
q00/ouroboros
loopx (dsh-loopx-plugin)
huangruiteng/loopx