agentleFS
mikeparcewskimikeparcewski/wicked-garden/skills/domain/SKILL.md

wicked-garden-domain

Skill9 starsChanged 7 months agoMITView on GitHub

The FOUNDATIONAL domain-model capability: extract a codebase's domain — testable business rules (with confidence + provenance), entities, requirements — as a schema-conformant model on the estate graph. The workers annotate the store; wicked-core reads it and builds the requirements graph, fail-closed below the configured minimum share of real requirements (80% by default). Steers three workers.

  • Installs packages

The file

We split this skill into the rules it gives an agent

3 rules. About 0 are common, 2 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈2

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    If none of these is available, or Python 3 is missing, skip the script-backed step, say so, and follow the manual alternative where one is …

    line 47
  2. Very specific:

    Corroboration: Every requirement in the built graph carries corroboration: validated when two or more independent sources support it …

    line 196
  3. Distinctive:

    Before extracting, check for wicked-understanding repo playbooks (the "how to work in THIS repo" layer).

    line 256

≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

wicked-garden-domain

The domain-model capability — a foundational substrate, not a step inside one archetype. It produces the shared domain model (business rules + entities + requirements on the estate graph); the modernize archetype derives from it, and build / migrate / review / specify / explore consume the same model. Understanding a codebase's domain is upstream of building, porting, reviewing, or governing it.

The contract in one line: the only thing that crosses repo lines is a document validating against the domain-model schema 1.1.0 (1.0.0 documents stay valid; vendored at vendor/domain-model.schema.json, owner: wicked-core/crates/wicked-governance/schemas/) plus an estate SymbolId string. Garden STEERS (annotates the estate store + cross-checks the built document), core BUILDS (wicked-core domain-graph reads the store, builds the requirements graph, and coverage-gates fail-closed below the configured minimum), estate GROUNDS (owns SymbolId identity + the graph), crew GOVERNS (drives the run). The peer CLIs are shelled via scripts/domain/_clients.py, or mocked behind fixtures when absent.

Runtime

Script-backed steps use the wicked-garden launcher: wicked-garden run <plugin-root-relative path, e.g. scripts/…> [args]. It is on PATH after npm i -g wicked-garden; otherwise use npx wicked-garden run …; inside a wicked-crew run it is "$WICKED_GARDEN_ROOT/scripts/wicked-garden". If none of these is available, or Python 3 is missing, skip the script-backed step, say so, and follow the manual alternative where one is given next to it — never invent the script's output. Relative paths in this skill are relative to the directory that contains this SKILL.md.

The four-way seam

crew GOVERNS  →  garden STEERS  →  core BUILDS  →  estate GROUNDS
(drives the run) (this skill +     (domain-graph +  (SymbolId + graph
                  3 workers)        coverage gate)   + Louvain clusters)

Garden never imports core or estate code — it shells their CLIs (argv lists, no shell string) or mocks them. It annotates the estate store and references SymbolId strings; core reads the store and builds + coverage-gates the requirements graph; estate is the sole writer of graph structure.

Routing — the three workers

AskWorkerProduces
Mine business rules from the estate → domain-model docdomain-extractor (wicked-garden-domain-extractor)business_rules[] with confidence + provenance; estate requirement annotations
Group clusters into domains → invoke core's domain-graph builddomain-modeler (wicked-garden-domain-modeler)an advisory domains{} grouping by estate community; the requirements graph saved into the estate store by wicked-core domain-graph (core groups its domains by package directory; --out writes a JSON export)
Threat-model the extracted model before builddomain-coverage (wicked-garden-domain-coverage)pre-build threat list / RISK-flag reasons

Hand-off — open the worker's skill by name (wicked-garden-domain-extractor, wicked-garden-domain-modeler, wicked-garden-domain-coverage) and carry out its SKILL.md with the ask as the argument; on Claude Code this is the Skill tool, on any other seat open the named skill from your catalog and run it inline, keeping its output separate from yours, then continue here. Run order for a full extraction: extractor → coverage → modeler (coverage is the pre-build check; the modeler's build comes last, as in crew's domain-extraction preset). Each is independently invocable.

The document this skill emits

A domain-model@1.0.0 document (see vendor/domain-model.schema.json). Top-level { metadata, domains }; every domain has requirements + entities; every requirement carries the full field set and ≥1 business rule; every business rule carries a numeric confidence ∈ [0,1] and a provenance{source, ref, source_kinds}. Full field map + the hard invariants: refs/domain-model-emit.md.

Reference-not-copy: legacy_components[] and provenance.ref hold estate SymbolId references, never a copy of the symbol's code, file, or line. estate resolves them live.

The working core slice (what ships in PHASE-1)

Garden is skills-only, but the deterministic assembly + invariant enforcement is real code you can run and test today:

# Emit a conformant domain-model doc from mocked estate + rule inputs
wicked-garden run scripts/domain/emit_domain_model.py --fixture > /tmp/domain-model.json

# Validate any doc against the vendored schema + hard invariants
wicked-garden run scripts/domain/validate_domain_model.py /tmp/domain-model.json
  • emit_domain_model.py — the extractor's deterministic core: takes a mocked estate cluster + a list of extracted rules and assembles a schema-conformant document, enforcing the hard invariants at build time (drop needs a reason, ≥1 rule per requirement, numeric confidence).
  • validate_domain_model.py — a stdlib-only draft-07 subset validation against the vendored schema (no third-party dependency — the repo is stdlib+pytest only), plus the extra invariants the schema can't express (numeric confidence type, SymbolId-shaped references).
  • scripts/domain/_clients.py — the selection seam: estate_client(db) / core_client() return CLI-backed clients when wicked-estate / wicked-core resolve, else the mocks. Shells the peers (argv lists, no shell string).
  • scripts/domain/_mocks.py — the disjoint fixtures: a fake estate client (canned clusters + resolve/annotate) and a fake brain client (the hermetic doc-assembly lane). These mock the peers; no other-product code is imported.

Honest scope — implemented vs stubbed

Implemented + tested: the vendored + pinned schema; the deterministic document assembler; the validator (schema + hard invariants); the CLI-backed clients (_clients.py) that shell the real wicked-estate + wicked-core, with argv-construction tests + an opt-in real-CLI lane; the fixture mocks; a conformant fixture document; the pytest suite that validates emitter output and the fixture against the schema.

Wired to live CLIs (_clients.py): the estate surface (resolve, annotate --replace, semantics, clusters --json --summary, annotations) → CliEstateClient; the domain-graph build (wicked-core domain-graph — reads the store, builds the requirements graph, coverage-gates fail-closed below the configured minimum) → CliCoreClient. When the peers are absent, the selection seam falls back to _mocks for hermetic testing.

The model's one job: stating each rule. The harness calls a bounded rule model (WICKED_RULE_MODEL_BIN, default claude -p); in a crew run's file mode the seat states the rules itself when no model is available. Everything around that call (enumeration, validation, the writes, coverage) is deterministic code. A real build needs an INDEXED store whose real requirements reach the configured minimum; a crew domain-extraction run maps it there page by page.

Coverage minimum, extraction order, corroboration (operator ruling 2026-10-31)

  • Minimum, not 100%, and only real requirements count. The graph is built once real requirements (behaviours with a business rule at or above the confidence threshold, 0.75 by default, or a validated requirement) reach the configured minimum share of the behaviours in product code: crew's requirementsGraph.minCoveragePct (Settings, default 80), which the daemon exports as WICKED_REQUIREMENTS_MIN_COVERAGE_PCT. A placeholder (a risk flag, or a rule under the threshold) is shown, never counted. Product code is decided by file path: test, spec and fixture files and data, manifest and document files are left out, though a test module inside a product file still counts, and an extracted rules-engine rule outside test and fixture paths counts whatever its file extension. The harness checks each rule's shape (a statement, a numeric confidence, a provenance) and that its source is readable product code before writing it; whether a statement is grounded in that source is for the page's review and the coverage evaluator to judge. wicked-core coverage --min-coverage-pct and wicked-core domain-graph both read it. Below it they say "23% real requirements, minimum 80% (1362 of 5786 behaviours; 4397 placeholders, 27 unmapped)" and fail closed; the mapping done so far stays in the estate store, so the next run continues from it. The built graph carries metadata.coverage (real_pct, min_pct, behavior_bearing, real, placeholders, unmapped, not_product). A repository's own minimum (wicked-core requirements-target --set N, kept in its store) comes before the daemon-wide one.
  • A generation keeps going. A crew domain-extraction run maps one page (worklist, 500 behaviours by default, WICKED_DOMAIN_PAGE_SIZE) at a time; after each page the host step extract-next (scripts/domain/next_page.py) prints one progress line ("Mapping wicked-crew · real requirements 3% → 12% → 19% · minimum 80% · page 4 · 2,174 placeholders · about 2 h left") and the run loops back to extract-prepare until the minimum is reached. Each page frames the unmapped behaviours first, then the placeholders earlier pages left, and skips a unit the seat already answered in this run: an answered placeholder waits for a later run, while a unit the seat left unanswered can be framed again. The run asks the operator instead of mapping on when the budget per grant (WICKED_DOMAIN_MAX_PAGES 20, WICKED_DOMAIN_MAX_MINUTES 240, crew Settings) is spent, when two pages added no real requirement, when a page of 20 or more units resolved under 1 in 20, or when nothing is left to map. It offers map more (while units are left), lower the minimum for this repository to the real share so far and build the graph (once that share is at least 1%), or stop (what was mapped is kept). With nothing left to map and a share under 1%, nothing can be lowered, so the run moves on to the coverage step, which reports the share (after a low-yield last page it first asks, with one choice: go on to the coverage step). A minimum the units left to ask cannot reach gets no time estimate.
  • Hub-first, fan-out order. The extraction harness maps the most central symbols first (estate's PageRank, wicked-estate rank), then their neighbours and onward by graph distance, so the load-bearing behaviour is covered first. The order is deterministic: a later run re-derives it and resumes at the frontier.
  • Across the seat boundary. In a wicked-crew domain-extraction run the extractor seat runs in a container with no wicked-core, no wicked-estate and a read-only store. Two host steps bracket it: extract-prepare writes the hub-first worklist (.wicked/domain/worklist.json, scripts/domain/prepare_worklist.py), the seat states the rules into .wicked/domain/rules.json (the extractor's file mode), and extract-apply (scripts/domain/apply_rules.py) checks that file against its schema and the store, refuses anything outside it or about a file the repository does not track, judges and writes the rest with the harness's own validation (authored extract-loop), and writes coverage-report.json.
  • Corroboration. Every requirement in the built graph carries corroboration: validated when two or more independent sources support it (callers, tests, docs, agreeing extractor seats), else unconfirmed with a why code and sentence (unindexed_file, dead_code, config_only, single_caller, test_only, doc_only, extractors_only). Core computes it; nothing in garden asserts it.

Estate id-scheme migration (2026-08) — when the pipeline must be re-run

Estate id-scheme "2" nests method/field SymbolIds under their enclosing type (<module>/Repo#save(). instead of <module>/save().). The first wicked-estate index per repo under a new-scheme binary hits the id_scheme meta mismatch and performs a forced full re-extract that re-mints every definition id — and the pipeline's outputs (business_rule/risk annotations + node_semantics written by scripts/domain/extract_loop.py) stay keyed to the OLD ids. Method/field-keyed rows orphan; module-level (top-level function / file) rows survive. Annotations are documented loss, not re-keyed.

Re-runs are required for every repo that carries domain annotations, after its first index under the new binary — in exactly this order:

  1. wicked-estate index <repo> with the new binary (the loud full re-extract).
  2. Re-run the extraction loop: wicked-garden run scripts/domain/extract_loop.py --db "<store>" … (it re-seeds from wicked-core coverage's unaccounted_nodes, so orphaned nodes simply reappear as unaccounted; the loop's pre-flight also warns when it detects orphaned annotations before writing).
  3. Rebuild the domain graph: wicked-core domain-graph (coverage-gates fail-closed below the configured minimum against the re-annotated store).

Previously emitted requirements_graph.json / domain-model documents embedding old-scheme SymbolIds are stale until step 3 completes.

Quantify the damage first (operators) — how many annotation rows would orphan (method/field-keyed) vs survive (still-live node), against any store:

-- rows whose node no longer exists (orphaned) vs rows still keyed to a live node
SELECT (n.symbol IS NULL) AS orphaned, count(*) AS rows
FROM annotations a LEFT JOIN nodes n ON n.symbol = a.node_sym
GROUP BY orphaned;

-- pre-migration estimate on an OLD-scheme store: rows keyed to a definition
-- nested inside a type (its id re-mints → orphans) vs module-level rows (id
-- unchanged → survive). Old flat ids don't encode nesting, so containment is
-- derived from byte spans: a Type-kinded node in the same file encloses it.
SELECT EXISTS (
         SELECT 1 FROM nodes t
         WHERE t.file = n.file AND t.symbol != n.symbol
           AND t.kind IN ('"class"','"struct"','"enum"','"trait"','"interface"',
                          '"module"','"namespace"','"type_alias"','"type"')
           AND json_extract(t.data,'$.location.span.start_byte')
               <= json_extract(n.data,'$.location.span.start_byte')
           AND json_extract(t.data,'$.location.span.end_byte')
               >= json_extract(n.data,'$.location.span.end_byte')
       ) AS would_orphan, count(*) AS rows
FROM annotations a JOIN nodes n ON n.symbol = a.node_sym
GROUP BY would_orphan;

Ground in the repo's method first

Before extracting, check for wicked-understanding repo playbooks (the "how to work in THIS repo" layer). A modernization lives or dies on repo-specific wiring. If present, load the matching playbook.

Depth

  • refs/domain-model-emit.md — the full field map, the seven hard invariants, and the SymbolId reference rule.
  • refs/extraction-flow.md — how the extractor drives estate (resolve → annotate → semantics) and core (domain-graph), the _clients.py selection seam, and the mock lane.
  • vendor/README.md — the vendored-schema pin + drift discipline.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.