Product Delivery Workflow

English | 中文

A product delivery workflow plugin for DeepSeek Harness: a full product-to-release pipeline with loop closure, built for high-quality autonomous completion of long, hard tasks.
Kick it off before you sleep; wake up to the final report.
100% AI-native — no human in the loop: autonomous research, architecture, implementation, review, testing, and release audit, end to end.

What this plugin provides

Installing this bundle adds one workflow skill + nine role-agent skills to every session of the target dsh profile (via the model-facing skill tool and ctx.skills). No other configuration is needed.

Skill	Role
product-delivery-workflow	The orchestrator: 10 stage gates from brief to release, artifact-driven, with review loops and a required final report.
product-agent	Product manager: evidence-backed decisions (BUILD / REFINE / REJECT), PRDs with acceptance criteria and metrics.
architect-agent	Senior architect: OpenSpec proposals, component boundaries, interfaces, contracts, and dependency-ordered task graphs.
senior-dev-agent	Senior developer: bounded implementation from contracts, deterministic tests, explicit error handling.
agent-team	Multi-agent team/swarm: bounded subagents, parallel workflows, disjoint ownership, independent review.
perf-reviewer-agent	Read-only performance reviewer against PRD budgets.
test-agent	Automated testing and manual platform/device verification against contracts and acceptance criteria.
memory-curator	Safe cross-session memory candidates from approved evidence only.
evolution-agent	Measurable workflow improvements as evidence-backed proposals.
release-audit-agent	Read-only pre-commit / pre-GitHub release audit (secrets, privacy, hygiene).

The workflow in one picture

Brief → Research → Product Decision → PRD → OpenSpec Architecture
      → Parallel Implementation (one bounded agent per task)
      → Independent Review Loop (correctness / architecture / performance)
      → Tests & Acceptance → Memory & Evolution → Release Audit
      → Required Final Report ──┐
                                └── findings route back to workers → loop until PASS

The loop is the point: every review finding routes back to the responsible worker and repeats until all reviewers pass (5 cycles max, then it asks for human direction). Nothing is claimed unless its artifact or verification result exists.

Why run it overnight

• Loop closure — review → fix → re-review continues without human babysitting; it stops only when gates pass or a hard blocker needs you.
• Artifact persistence — everything lands under .opencode/workflow/<slug>/ (01-research.md → 10-*), so you can audit the whole run after sleeping.
• Bounded delegation — parallel waves with disjoint file ownership and explicit contracts; workers never commit, push, or redesign contracts.
• Safety rails — the release audit runs before any git commit / PR / push, and memory curation never stores secrets or transcripts.
• Self-improving — evolution-agent + memory-curator turn each run's evidence into proposals and durable memory for the next one.

Install

Requires the dsh CLI (install DeepSeek Harness).

# from GitHub (recommended)
dsh plugin --profile web add github:wellorbetter/dsh-product-delivery-workflow

# or from a local checkout
dsh plugin --profile web add ./dsh-product-delivery-workflow

Restart the profile (dsh web / dsh --profile <name>) after installing. The plugin is a bundle: it declares dsh.bundle.patch, so dsh plugin reconciles it into the profile's dsh.profile.bundles layer list automatically. It ships plain JavaScript (lib/index.js), so a GitHub install needs no build step or allowBuilds permission.

Use

In any session of the profile, ask for the workflow — the orchestrator loads the role skills itself:

Run the product delivery workflow on <brief>. Produce the full pipeline and the final report. You may work autonomously; loop until the gates pass.

Or invoke a single role for a bounded task:

Use product-agent to review this PRD against the delivery results and report gaps.

For overnight runs, pair it with the goal/headless mode: start a session with the brief, let it loop, and check the .opencode/workflow/<slug>/ artifacts and final report in the morning.

Development

pnpm install
pnpm verify     # discovers + loads all 10 packaged skills through the provider

License

MIT