dsh-fullstack-engineer

A DeepSeek Harness plugin that gives an agent a full-stack engineering operating contract: scoped requirements, UX patterns, security and reliability, stepwise RAG with LangChain/LangGraph, and harness/loop engineering for subagent applications.

It is a planning and review instrument. It does not generate an application for you; it forces the order and evidence that make an application shippable.

What it contributes

Tool	Output
fullstack_plan	Jobs, non-goals, acceptance criteria, UX flows/states, architecture, threat model, SLOs, delivery gates, definition of done
fullstack_ux_review	Per-screen state matrix, pattern fit, accessibility, copy rules, anti-patterns
fullstack_security_review	Design-level threat model (object authz, tenant isolation, injection, SSRF, secrets, CSRF, audit, AI boundaries), controls and tests
fullstack_rag_blueprint	Ten-step RAG: source/ACL inventory → ingestion → parse → chunk → embed/index → hybrid retrieve/rerank → grounded generation → eval → observe → operate
fullstack_subagent_plan	Supervisor and worker contracts, allowed tools, loop budgets, no-progress/stop rules, harness requirements, gates, red-team cases
fullstack_gate	Pass/hold from named artifacts, with the evidence records required to close the gate

Plus a system-prompt section (fullstack-engineer:playbook) stating the non-negotiable engineering order.

The order the plugin enforces

1. Frame — users, jobs, non-goals, constraints, acceptance, one success KPI.
2. UX — flows and every state (default, loading, empty, partial, error, offline, permission-denied, success), accessibility and copy.
3. Contract — API, data model, authorization, migration, rollback.
4. Threat model — object authz, tenancy, injection, SSRF, secrets, CSRF/session, audit; for AI: prompt injection, RAG ACL leakage, tool abuse.
5. Evaluate — corpus owners/ACLs, golden set, thresholds, rollback before RAG or subagents.
6. Release evidence — records with producer, commit, scope and freshness; never a URI alone.

RAG with LangChain/LangGraph

The blueprint is explicitly stepwise (Langkah): every stage has an exit criterion. LangChain is the component toolkit (DocumentLoader, TextSplitter, VectorStore, Retriever, RunnableParallel, reranker); LangGraph is the orchestration option when the assistant needs bounded multi-step state. The design includes:

• hybrid retrieval (BM25 + vector), ACL filtering before scoring, top-20 recall → 5 reranked;
• 500–800 token chunks, 15% overlap, parent-document expansion;
• citations or refusal, no unsupported answers;
• golden set of 50+ questions with refusal and injection cases;
• thresholds: recall@5 ≥ 0.85, faithfulness ≥ 0.90, citation precision ≥ 0.90, refusal correctness ≥ 0.95;
• token/cost/latency budgets and a weekly feedback loop.

Harness and subagent loops

fullstack_subagent_plan produces a supervisor + typed workers (product, UX, backend, frontend, security, verifier), one-at-a-time routing, context-budget discipline, iteration/tool-call limits, no-progress rule, stop conditions, approval gates, an append-only audit expectation, and red-team cases such as "approve your own evidence" or "skip the verifier".

Gates: scope → ux → build → reliability → rag → harness → release → outcome. GATE_G is expansion-only; a KPI/guardrail failure never blocks a release or rollback.

Security integration

When dsh-secure-mesh is installed, the prompt tells the agent to use:

• secure_fetch for external HTTP (no raw egress),
• request_human_approval before irreversible/high-impact actions (only approved: true authorizes),
• validate_claim before SQL/Python/financial claims,
• the untrusted-data fencing rules for retrieved and external text.

This plugin itself has no network, filesystem or process access.

Install

The package declares a dsh.bundle.patch, so it can be installed like any DSH bundle:

1. Publish/install the package (dsh-fullstack-engineer@0.1.0) or a Git checkout.
2. Add the bundle in the Plugins page, or add the fullstack-engineer entry from cordis.patch.yml to the profile.
3. Restart/reload the profile.

For local development:

npm install
npm run build
npm run test:stub

In a sandbox that blocks the Node test runner's child processes, run tests in-process:

node --test --test-isolation=none tests/playbook.test.mjs
node --import ./tests/support/stub-register.mjs --test --test-isolation=none tests/plugin.test.mjs

Real-profile integration test

The plugin is installed in the local web profile and the integration test boots the real Cordis Context, SystemPrompt and ToolRuntime (not the stubs):

dsh plugin --profile web add file:D:/path/to/dsh-fullstack-engineer
node --import ./tests/support/real-register.mjs \
  --test --test-isolation=none tests/integration.real.test.mjs

It asserts that the profile lists the bundle, that all six tools and the prompt section register, and that fullstack_plan / fullstack_gate execute through the real tool runtime. DSH_HARNESS defaults to Y:/DSH/deepseek-harness when resolving the real packages.

Configuration

Key	Default	Purpose
rigor	standard	lean, standard or regulated; regulated adds approval/audit expectations
defaultStack	FastAPI + PostgreSQL + React	Stack used when the brief does not choose one
prompt	true	Contribute the playbook section to the system prompt
plan, uxReview, securityReview, ragBlueprint, subagentPlan, gate	true	Register individual tools
maxOpenQuestions	8	Cap on open questions (3–20)

Limits

• Planning is deterministic and opinionated; it is not a substitute for product discovery.
• The threat model is design-level. It names required tests; it does not claim they ran.
• It does not deploy, migrate, call providers or write code.
• Evidence verification is the host/CI's responsibility; the gate tells you what must exist.

License

MIT.