dsh-autopilot

Governed plan-execute-audit for DeepSeek Harness: mechanical gates, owner-only egress, and evidence-bound closeout as an out-of-tree plugin.

The project name is dsh-autopilot. The npm package is dsh-goal-autopilot. Unscoped dsh-autopilot on npm belongs to a separate, independent project; install this plugin as dsh-goal-autopilot.

This is a developer preview. Experimental. Latest is 0.2.0, which adapts the plugin to dsh 0.2.0 (CLI and the Desktop app's bundled dsh command). On WSL2 + dsh 0.2.0-rc.2, install, mount, and a delegated run through needs-fix resume are measured with a local mock model. Desktop is not run here, and Windows is not verified. Not a production release.

dsh-autopilot turns a natural-language goal into a gated run: plan, audit, execute, audit, then close only when the evidence is bound. Design lineage is the Claude Code goal-autopilot-harness skill, rebuilt on dsh plugin seams so the gates are executable instead of prompt-only. Deep design: DESIGN.md.

flowchart LR
  Goal --> Plan --> Audit1[Audit] --> Execute --> Audit2[Audit] --> Closeout

Capabilities

• Plan gate — standard runs stay read-only until an independent plan audit passes.
• Independent audits — one-shot auditors, structured verdicts, no verdict means no gate flip.
• Delegated executor — a continuable child implements; the same child resumes on needs-fix.
• Owner-only egress — git / gh / publish writes need a manifest and owner authority.
• Evidence-bound closeout — completed needs both gates, required roles, and one bearer per acceptance criterion.
• Usage + stop reminder — standard runs must declare how a change was operated; a bounded 3x nudge fires if a turn ends mid-execution.

Quick Start

Requires dsh 0.2.0 (or 0.1.2-rc.1 / 0.1.5-rc.1) and Node 22+. On dsh Desktop, install dsh-goal-autopilot from the plugin manager or with the bundled dsh command.

dsh plugin --profile <name> add dsh-goal-autopilot

Pin the current release:

dsh plugin --profile <name> add [email protected]

Then ask, in natural language:

Use dsh-autopilot to implement <goal> and verify the result.

You do not need the internal autopilot_* tool names to start. Operators who want the full sequence can read docs/reference.md.

How it works

A standard run is event-sourced (planning → review → executing → review → close). The engine is the only writer. Audits are independent subagents. Closeout refuses completed until the gates and evidence line up.

That is the shape. The invariants, honest limits, and host traces live in DESIGN.md.

Status

Topic	Current fact
Maturity	Experimental developer preview (not a production release)
npm	[email protected]
0.2.0	Adapts to dsh 0.2.0: peer range, executor resume, executor recognition, reminder source kind, PTC card events
Hosts	dsh 0.2.0 line, 0.1.5-rc.1, 0.1.2-rc.1 (0.1.7 and 0.2.1 are outside the range)
WSL2 + 0.2.0-rc.2	Install, mount, skill, plan audit → delegated executor → needs-fix resume measured with a mock model
Full NL write → execution audit → closeout	Not verified with a real model
Desktop (macOS / Windows)	Not run — not verified
Windows	Not verified — not claimed as supported
Security	Quality gates + fail-closed egress; not complete security

Host matrix, Desktop notes, the 0.2.0 fixes, and the sandbox residual: docs/compatibility.md.

Documentation

Doc	For
DESIGN.md	Architecture, state machine, invariants, verification
docs/compatibility.md	DSH versions, Desktop, WSL/Windows, 0.2.0 fixes
docs/security.md	Approval, egress, sandbox, non-goals
docs/reference.md	Tool protocol, run state, usage classes
CHANGELOG.md	Release history

Develop

From a local checkout:

pnpm install
pnpm run build
dsh plugin --profile <name> add .

pnpm run check   # typecheck
pnpm run test    # vitest
pnpm run build   # emit lib/

Client-half typecheck needs a dsh tree at DSH_SRC (default ~/dsh). CI runs the host half only. See docs/compatibility.md.

License

MIT