Product

A CLI and MCP server for the Product Framework — specify software as a verifiable What/How graph.

The Product Framework is an open standard for describing a software product as one connected, machine-readable graph: the What (domain model + event model — entities, commands, events, read models, UI steps typed against Abstract Interaction Objects, systems, triggers, Deciders, Projectors), the How (contracts, the screen-composition / reification model, delivery features), and the typed links between them. The graph can drive generation, gate verification, and explain itself — so "describe this system" is a query, not a stale document.

This repo is the reference tooling: a single Rust binary (product) plus an MCP server that lets an agent author and verify the graph directly. No database, no service — the graph lives as YAML/Turtle under .product/.

$ product init --demo                 # scaffold + seed the bookstore What model
$ product domain new system sys-shop --system-kind application \
      --purpose "consumer e-commerce" --target-classes gui
$ product domain validate --strict    # per-node shapes + graph-level completeness
$ product decider derive Order        # derive an aggregate's executable signature
$ product decider validate Order-decider
$ product mcp --http                  # MCP server + a live Event-Modeling web view at /

Install

# from source
cargo install --path product-cli

The binary ships with the What→How→Build Claude Code skills baked in. product init writes them into .claude/skills/ of the new repo (pass --no-skills to opt out); product skills install (re)installs them, and product skills install --global puts them in ~/.claude/skills/ for every project. Start a fresh Claude Code session to pick them up, then /product-session.

Choosing the agent CLI

product session start (and product author domain) host the What→How→Build session in an agent CLI — Claude Code or GitHub Copilot CLI. The CLI is resolved in this order:

1. the --cli claude|copilot flag, else
2. the repo's [author].cli in .product/config.toml, else
3. the global user default in $XDG_CONFIG_HOME/product/config.toml (or ~/.config/product/config.toml), else
4. claude.

# .product/config.toml — make this repo default to Copilot CLI
[author]
cli = "copilot"

Scaffold it on a new repo with product init --cli copilot, or set a personal default for every repo by putting the same [author] block in ~/.config/product/config.toml. With a default configured, product session start needs no --cli flag.

60-second tour

product init --demo                   # a worked What model to explore
product domain list                   # the captured nodes, by kind
product domain show Order             # one node and its links
product domain export                 # the graph as RDF/Turtle
product domain validate               # §3.1/§3.2 per-node conformance shapes
product domain validate --strict      # + §3.2.0/§3.2.5/§3.4/§4.5 completeness checks
product decider derive Order          # §3.3 — derive decide/evolve signature
product decider simulate Order-decider  # run its flow-derived scenarios
product guide                         # where you are + the next step

The model

• What — product domain … captures the domain + event model; product decider … (§3.3) and product projector … (§3.4) make behaviour and read models executable; product primitive … (§3.5) names irreducible algorithms.
• How — product how, product feature, product build, product seam, product preview cover the How contract, delivery features, the screen seam, and the §11/§12 design-system / content-store preview profiles.
• Everything is validated against the framework's SHACL shapes + SPARQL rules; the captured What serializes to Turtle (product domain export).

MCP + the web view

product mcp --http starts the MCP server (framework tools: product_domain_*, product_decider_*, product_projector_*, …) and serves a live web view at / that renders the active What graph across three connected views — Systems (the product → systems & journeys map, §3.0), Domain (one bounded context as an ER graph, §3.1), and Flows (a system's event-model as Event-Modeling swimlanes — triggers / commands / views over per-aggregate event streams, §3.2). A node detail panel, the What→How→Build phase stepper, dark/light theme and live SSE refresh round it out.

ddd — Decision-Driven Design governance

The workspace also ships ddd (crates ddd-core, ddd-lsp, ddd-mcp, ddd-cli), a separate tool over a separate store: a repo-local .ddd/ graph of predicates, closure claims, decisions, analyzer/linter manifests, pattern instances, seam declarations, and interception event rows (spec, umbrella PRD, formats: migrations).

cargo run -p ddd-cli -- init        # scaffold .ddd/
ddd validate                        # schema + ontology rules (CI gate)
ddd diff --sarif build.sarif        # declared vs. detected rules (CI gate)
ddd report escapes                  # diff + cadence + basis-loss report, each section stating its coverage
ddd why CA2007                      # rule -> decision -> principal -> claims (or: detected but unfiled)
ddd render                          # static self-contained HTML projection of the graph
ddd serve                           # the ddd_* MCP surface (stdio)
ddd warmup                          # pre-load the LSP hosts (Roslyn solution load)
ddd what                            # What-graph boundaries carrying no declaration

Governing the What (the framework graph)

ddd what treats the .product/ What graph as a third governed surface alongside C# and Bicep. It needs no language server: product-core already owns the What as typed data, so the adapter reads kinds and containers straight off DomainGraph and runs them through the same policy-table mechanism (ddd-core/src/surface.rs, shared with the LSP adapters).

Two kinds of row. Boundary kinds are surface whatever they connect to: a system (§3.2.5), a context mapping (§3.1), a journey crossing (§3.0.1), a quality demand (§3.6). Published kinds are surface only when a §3.2.0 Translation carries them — the View a Translation watches, the Command it issues, and the Events that View projects. Everything else is internal to its own system and never demands a declaration.

That published/internal split is the What's analogue of C# visibility, and it is load-bearing: the first table called every event and command a boundary, which the measurement in DDD-what-02 killed (0 of 39 crossed anything). See dec/ddd/what-published-qualifier.

A boundary counts as governed when a seam declaration's contract_location is what:<element-id>. --strict turns it into a CI gate; the default reports without failing, so the table can be calibrated against a real graph first.

MCP surface (M3/M4)

ddd serve exposes the ddd_* tool namespace over stdio: LSP-backed language intelligence for C# (roslyn-language-server --stdio --autoLoadProjects, the official prerelease .NET global tool) and Bicep (bicep-ls from Azure.Bicep.LangServer) — find_symbol, references, hover, diagnostics (joined to the manifests by rule id, the same join the SARIF path uses), signature, rename (computes; application funnels through the interceptor) — plus the governance tools why, graph_query, declare_seam, declare_pattern, accept_risk.

The three declare/accept tools take amend: true to revise an entry already filed. The flag is explicit in both directions — a create never silently overwrites, an amend never silently creates — and the split is by field: judgement amends (verdict_knowledge, obligation answers, rationale), while identity and LSP-derived evidence (contract_location, metadata) are carried forward untouched, so the interception rows stay machine-authored.

ddd_apply_edit runs every edit through the per-language contract-surface classifier (policy tables, PRD §9): non-surface edits apply; a surface edit applies only with a matching same-session declaration; otherwise it is rejected with a structured demand whose facts (symbol, kind, signature, visibility, reference count) are pre-filled and whose judgment fields are blank (dec/ddd/rejection-facts-prefilled). Modes: intercept: enforce | warn | off, per artifact class via intercept_by_class (config format 3); adapter.csharp.internal_is_surface flips the library-repo posture (dec/ddd/internal-not-surface). Every classified surface outcome lands as a row under .ddd/seams/events/ — the correspondence dataset. Hosts are spawned lazily, health-checked, and respawned on crash; while Roslyn loads the solution, tools return an explicit {"status": "loading"} rather than hanging. CI runs against a fixture-grade mock host (dec/ddd/fixtures-not-sdk); set DDD_LSP_E2E=1 with both tools on PATH to run the gated real-host suites.

ddd diff compares the manifests under .ddd/manifest/ against two detected sources per language and reports UNGOVERNED (detected, no manifest entry), STALE (manifest entry, absent from config and emissions), and UNCITED_SUPPRESSION (a config or in-source suppression with no risk-acceptance record):