build-project-model

name	build-project-model
description	Bridge skill — author a valid ProjectModel JSON (per normalized_schema) from greenfield inputs (parsed customer docs + brainstorm report + user-story list + risk register + clarification answers), then validate it so /morkit:init can render docs/srs.md and friends with no hand-authored JSON. The missing brainstorm→init link for the greenfield pipeline.
category	documentation
keywords	["project-model","bridge","greenfield","normalized-schema","srs","provenance","doc-ingest","brse"]
argument-hint	--workspace morkit/output/greenfield/<proj> [--language JP|EN|VN]
metadata	{"author":"morkit-greenfield","version":"1.0.0"}

Build ProjectModel (Bridge) 🌉

The keystone of /morkit:greenfield. Turns the BA/BrSE artifacts into a single project-model.json that the existing docs-hero init pipeline consumes. After this skill, customer docs → project-model.json → init → docs/srs.md works end-to-end.

Conventions (workspace layout, provenance, classification) are defined once in ../greenfield-orchestrator/references/greenfield-conventions.md. Schema authoring rules: references/schema-cheatsheet.md. Do not restate them — follow them.

Inputs (read what exists, skip what doesn't)

All under the run workspace morkit/output/greenfield/<proj>/:

File	Produced by	Use
inputs/*	G0 intake	raw customer docs (parsed below)
brainstorm-report.md	G1 /morkit:brainstorming	scope, goals, actors
user-story-list.md	G2 generate-user-stories	FR / UseCase candidates
gap-analysis.md	G3 gap-risk-analysis	gaps → open_questions
risk-register.md	G3 gap-risk-analysis	rows → constraints_risks.risks
clarification-log.md	G4 clarification-loop	answers → fill <TBD>, FR detail

Step 1 — Doc-ingest (reuse parse_inputs.py, no new parser)

PY="${HOME}/.claude/plugins/data/docs-hero/.venv/bin/python3"
ORCH="${CLAUDE_PLUGIN_ROOT:?}/skills/docs-hero-orchestrator/scripts"
WS="morkit/output/greenfield/<proj>"

"$PY" "$ORCH/parse_inputs.py" --inputs-dir "$WS/inputs" \
  --language "$LANG" --project-name "<proj>" --output "$WS/raw-bundle.json"

parse_inputs.py does structural extraction only (text/tables) and never crashes on empty inputs (returns warnings). Then you (LLM) read raw-bundle.json and, per source, write a short fact list — never dump full text. Emit source-manifest.json per the schema: each source → classified_as (entity types, see conventions §5) → facts. For large docs, summarize each source in a subagent so context stays small.

Step 2 — Author project-model.json

Using raw-bundle.json + source-manifest.json + the markdown artifacts above, author $WS/project-model.json following the cheatsheet. Non-negotiable rules:

• Only meta.project_name is required. Start minimal, add only justified entities.
• No fiction. Every seeded entity carries source (SourceRef) and/or external_sources: [path], and doc_status: "Draft". A fact with no source becomes an open_questions entry or a <TBD: reason> string — never invented.
• Skip code-derived entities (screen items, db tables/columns, api endpoints, naming/lint, modules) — greenfield has no code; init backfills later.
• Canonical/drift (conventions §6): external doc is primary for WHAT/WHY; on a conflict between two docs, record an open_questions/risk row rather than guess.
• Map user-stories to entities by format: brse → functional_requirements, agile → business_flow.use_cases (+ derived FR). Keep ids stable from G2.

Step 3 — Validate (hard gate, loop until clean)

VAL="${CLAUDE_PLUGIN_ROOT:?}/skills/build-project-model/scripts/validate_project_model.py"
"$PY" "$VAL" --project-model "$WS/project-model.json"

• Exit 0 → valid; proceed to G6 (init).
• Exit 1 → the script prints exact field.path: message lines. Re-author only the offending fields and re-run. Cap at ~5 iterations; if still failing, surface the remaining errors to the user (don't loop forever).
• Exit 2 → file missing / malformed JSON → fix the JSON and re-run.

Common fixes: id regex (FR-001, UC-001, ENT-001, DATA-001, INT-001, CONS-001, ASM-001); priority/status enum typos; provenance draft placed in doc_status, not status.

Step 4 — Hand off to init

The validated file is consumed unchanged by the orchestrator:

"$PY" "$ORCH/dispatch_coordinator.py" init \
  --project-model "$WS/project-model.json" \
  --language "$LANG" --outputs srs --docs-dir "$PWD/docs"

(The greenfield-orchestrator G6/G7 stages call this for you with the resolved --outputs.)

Tests

tests/test_validate_project_model.py — valid minimal model, missing-required error path, provenance extras preserved, bad id rejected, the status-vs-doc_status pitfall. tests/fixtures/seeded-greenfield.json is a known-good bridge output.