mikeparcewski/wicked-garden / skills/domain-modeler
wicked-garden-domain-modeler
VerifiedDomain-graph worker for the modernize archetype. Proposes the repository's BUSINESS domains (3 to 9 areas named for what the code does for its users, each with a one-line purpose and the code that shows it) and assigns each rule one with a confidence. Groups estate's Louvain communities as advisory provenance, invokes wicked-core's domain-graph build (reads the annotated store, recomputes coverage, fail-closed below the configured minimum), then validates core's output against the vendored schema. Use when: dispatched by wicked-garden-domain after rule extraction; when a domain-extraction worklist asks for the repository's business domains; "what are this repo's business areas", "group these into domains", "build the requirements graph", "translate clusters into a domain model". NOT for mining the rules themselves (that is domain-extractor) or threat-modeling (that is domain-coverage).
Domain Modeler
You turn a flat set of extracted rules into a cluster-keyed domain model and invoke core's domain-graph build to produce the requirements graph. Your grouping is derived from estate's Louvain communities, not hand-partitioned.
Contract you emit against
Same vendored domain-model schema as the extractor (1.1.0; 1.0.0 documents
stay valid). Your domains{}
grouping — each domain groups the requirements whose components fall in one estate
community, carries an advisory cluster_id (the Louvain index), and declares its
entities{} — is advisory provenance. The authoritative requirements graph is
built by core from the annotated store, and core keys its domains by package
directory (the parent directory of each behaviour's file; it sets no cluster_id),
because dense modern code collapses into one community. That key is not the business
domain: each requirement also carries the business domain its rule picked
(business_domain), and the model lists them (metadata.business_domains), from the
"Business domains" catalog below. You invoke core, you
don't reimplement it.
Business domains (propose, then assign)
A requirement's directory is not its domain: backend/integrations/oauth is where the code
lives, "Credential management" is what it does for the product's users. Package-directory
domains fit a product owner's view on about half of the rows measured, so the repository gets a
small catalog of business areas, and every rule picks one. The directory is evidence, never the
name.
Propose (once per repository: a domain-extraction worklist says
business_domains.propose: true; a stored catalog is never replaced, so pick from
business_domains.catalog when it is there):
- Read the code for what the product does for its users, not how it is laid out: entry points
(
main, CLI commands and their verbs, workers and schedulers), HTTP or RPC routes and their handlers, the nouns in public APIs and data models, and what the README or docs say the product is for. Cover every entry-point group the code shows, including scheduled and housekeeping jobs (cron/scheduler jobs, maintenance, notices). - Propose 3 to 9 domains. Each is
{"id": "<lowercase-slug>", "name": "<2 to 5 words>", "purpose": "<one line: what this area does for its users>", "evidence": [...]}. Name the area a product owner would recognise ("Credential management", "Workflow scheduling", "Usage billing"), never a directory, package, module or class name (the host refuses a name that is a directory name in this repository: the worklist'sbusiness_domains.directory_names). Prefer fewer, broader areas over many thin ones; one area per user-facing capability. - Ground every domain in code.
evidencelists 1 to 8 items{"kind": "entry-point" | "route" | "cli-verb" | "api-noun" | "directory", "file": "<repo-relative path>", "symbol": "<optional name in that file>", "text": "<optional: the route or verb, as it appears in the file>"}. Every item other than a directory names a product-code file (not a test, data or document file) and asymbolortextin it, and every domain names at least one symbol. The host checks each file is tracked by git, each symbol is in the code graph at that file and each text occurs in the file. An item the code does not show is dropped, and a domain left with no symbol the graph holds is dropped. The catalog is refused only when fewer than 3 domains remain. The seat's--check-rulesnames a bad file or text first. The page's review then judges whether that code is really the area's.
Assign each rule one domain with a confidence in [0,1]: the area the rule serves, judged from what the rule does for the product's users, with the file's directory as one piece of evidence. High (0.8 or more) when the rule plainly serves that area; lower when two areas compete. A multi-purpose module's symbols can land in different domains. When no area fits, leave the rule without a domain rather than forcing one.
In a domain-extraction run all of this goes into the rules file the extractor writes
(.wicked/domain/rules.json): the catalog as its top-level business_domains list (only when
proposing), each rule entry's "domain": "<id>" and "domain_confidence": <0..1>, and, for the
worklist's assign list (requirements an earlier page resolved before there was a catalog), a
top-level "assignments": [{"symbol_id": "<verbatim>", "domain": "<id>", "confidence": <0..1>}].
The host's apply step stores the catalog beside the code graph and writes each pick on its code
node; wicked-core domain-graph reads it into each requirement's business_domain. Requirement
keys (<directory>::REQ-NNN) do not change, so existing links, overrides and citations hold.
The loop
- Read clusters from estate — the full-membership feed
clusters --json --summary→[{id, size, members:[symbol_id], …}], viaestate_client(db).read_clusters()(CLI) or_mocks.EstateClient(hermetic). On the real path this is informational (core builds the authoritative domains itself when it builds the graph); the id is a volatile positional index, so key durable grouping on a hash of the sorted member set, not the raw id (record the raw id only as advisorycluster_id). - Assign each requirement to a domain by the community its
legacy_components[]SymbolIds belong to (advisory grouping / provenance for the coverage evaluator). A requirement whose components span communities goes to the dominant one; note the split. - Invoke core's domain-graph build.
core_client().domain_graph(db, out)shellswicked-core domain-graph --db <db> --out requirements_graph.json, which reads the annotated store, recomputes front-half coverage, and saves the requirements graph into the estate store (domain, requirement and business-rule nodes,SatisfiedByedges to code — the source of truth crew reads);--outalso writes it as therequirements_graph.jsonexport this step validates (a capability plan, not a code skeleton). It fails closed below the configured minimum (writes nothing, non-zero exit, and says "N% real requirements, minimum M% (…)") — real requirements (behaviours with a business rule at or above the confidence threshold, or a validated requirement; placeholders never count) must reach at least that share of the behaviours in product code (the repository's own minimum, elseWICKED_REQUIREMENTS_MIN_COVERAGE_PCT, which crew exports from itsrequirementsGraph.minCoveragePct, else 80). There is no doc handed in and no--overlay: the store is the input. Whencore_client()isNone(no peer), fall back to the hermetic lane —emit_domain_model.pyassembles the doc and_mocks.BrainClientstands in. - Validate core's output. Cross-check the built
requirements_graph.jsonwithscripts/domain/validate_domain_model.py(schema + hard invariants) before returning.
Invariants you must not break
domainsis required; every domain hasrequirementsandentities.cluster_idis advisory provenance only — never authoritative. estate is the sole owner of graph structure; you reference it.- No silent drop: a requirement dropped from a domain needs
disposition:"drop"and adisposition_reason, or it counts against coverage. metadata.migration_mode ∈ {structural, functional}—functional(capability-grouped) is the default;structuralis 1:1.
What is wired vs stubbed
The estate clusters read and core's domain-graph build are wired to the
live CLIs via scripts/domain/_clients.py (estate_client / core_client),
falling back to _mocks.py when the peers are absent. The grouping heuristic
(community → domain) and the schema validation are real. A real build
additionally needs an INDEXED store whose real requirements reach the configured
minimum; a crew domain-extraction run maps it there, not this worker. Full field map:
the wicked-garden-domain skill's refs/domain-model-emit.md.
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 472k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |