wicked-garden 12.37.1 → 12.38.0

package/.claude-plugin/components.json

@@ -1,9 +1,9 @@
 {
   "_comment": "Issue #328: Component declarations for wicked-garden plugin. Claude Code plugin.json does not yet support inline component declarations beyond path overrides. This manifest documents all components for tooling, validation, and future schema support. When plugin.json gains a 'components' field, merge this data into plugin.json. The plugin is skills-only: former commands/ became actions of consolidated per-domain skills, and former agents/ became standalone context:fork worker skills (enumerated under fork_skills). Regenerate derived blocks with scripts/ci/sync_components.py.",
-  "domains": ["agentic", "crew", "data", "engineering", "jam", "mem", "persona", "platform", "product", "qe", "search", "smaht"],
+  "domains": ["agentic", "crew", "data", "engineering", "jam", "mem", "persona", "platform", "product", "qe", "search"],
   "summary": {
-    "skills": 144,
-    "fork_skills": 73,
+    "skills": 133,
+    "fork_skills": 69,
     "hooks": 13,
     "specialists": 7
   },
@@ -15,9 +15,6 @@
     "archetype": ["archetype"],
     "classify": ["classify"],
     "core": ["core"],
-    "crew-implementer": ["crew-implementer"],
-    "crew-researcher": ["crew-researcher"],
-    "crew-reviewer": ["crew-reviewer"],
     "data": ["data"],
     "data-engineer": ["data-engineer"],
     "deliberate": ["deliberate"],
@@ -28,12 +25,10 @@
     "draft": ["draft"],
     "engineering": ["architecture", "backend", "debugging", "docs", "engineering", "frontend", "integration", "large-scale-migration", "patch", "system-design", "unit-test-quality"],
     "engineering-api-documentarian": ["engineering-api-documentarian"],
-    "engineering-conformance-reviewer": ["engineering-conformance-reviewer"],
     "engineering-migration-engineer": ["engineering-migration-engineer"],
     "engineering-solution-architect": ["engineering-solution-architect"],
     "governed-worker": ["governed-worker"],
     "ground": ["ground"],
-    "integration-discovery": ["integration-discovery"],
     "jam": ["jam"],
     "jam-brainstorm-facilitator": ["jam-brainstorm-facilitator"],
     "jam-council": ["jam-council"],
@@ -41,7 +36,6 @@
     "mem-capture": ["mem-capture"],
     "mem-ingest": ["mem-ingest"],
     "persona": ["persona"],
-    "persona-agent": ["persona-agent"],
     "platform": ["audit", "compliance", "errors", "gate-benchmark-rebaseline", "gh-cli", "github-actions", "gitlab-ci", "glab-cli", "health", "observability", "peer-health", "platform", "prereq-doctor"],
     "platform-compliance-officer": ["platform-compliance-officer"],
     "platform-privacy-expert": ["platform-privacy-expert"],
@@ -95,15 +89,13 @@
     "qe-ui-component-test-engineer": ["qe-ui-component-test-engineer"],
     "qe-visual-regression-engineer": ["qe-visual-regression-engineer"],
     "repo-learn": ["repo-learn"],
-    "runtime-exec": ["runtime-exec"],
     "search": ["codebase-narrator", "search"],
-    "smaht": ["discovery", "intent", "propose-skills", "smaht"],
     "swarm": ["swarm"],
     "wickedizer": ["wickedizer"],
     "workflow": ["workflow"],
     "worktrees": ["worktrees"]
   },
-  "fork_skills": ["wicked-garden-agentic-architect", "wicked-garden-agentic-performance-analyst", "wicked-garden-agentic-safety-reviewer", "wicked-garden-archetype", "wicked-garden-classify", "wicked-garden-data-engineer", "wicked-garden-domain-coverage", "wicked-garden-domain-extractor", "wicked-garden-domain-modeler", "wicked-garden-engineering-api-documentarian", "wicked-garden-engineering-conformance-reviewer", "wicked-garden-engineering-migration-engineer", "wicked-garden-engineering-solution-architect", "wicked-garden-ground", "wicked-garden-integration-discovery", "wicked-garden-jam-brainstorm-facilitator", "wicked-garden-jam-council", "wicked-garden-mem-capture", "wicked-garden-mem-ingest", "wicked-garden-persona-agent", "wicked-garden-platform-compliance-officer", "wicked-garden-platform-privacy-expert", "wicked-garden-platform-security-engineer", "wicked-garden-product-a11y-expert", "wicked-garden-product-requirements-analyst", "wicked-garden-product-ui-reviewer", "wicked-garden-product-ux-designer", "wicked-garden-product-value-strategist", "wicked-garden-qe-a11y-test-engineer", "wicked-garden-qe-acceptance-test-executor", "wicked-garden-qe-acceptance-test-reviewer", "wicked-garden-qe-acceptance-test-writer", "wicked-garden-qe-ai-feature-test-engineer", "wicked-garden-qe-chaos-test-engineer", "wicked-garden-qe-code-analyzer", "wicked-garden-qe-compliance-test-engineer", "wicked-garden-qe-contract-testing-engineer", "wicked-garden-qe-coverage-archaeologist", "wicked-garden-qe-data-quality-tester", "wicked-garden-qe-e2e-orchestrator", "wicked-garden-qe-exploratory-tester", "wicked-garden-qe-flaky-test-hunter", "wicked-garden-qe-fuzz-property-engineer", "wicked-garden-qe-iac-test-engineer", "wicked-garden-qe-incident-to-scenario-synthesizer", "wicked-garden-qe-integration-test-engineer", "wicked-garden-qe-load-performance-engineer", "wicked-garden-qe-localization-test-engineer", "wicked-garden-qe-mutation-test-engineer", "wicked-garden-qe-observability-test-engineer", "wicked-garden-qe-production-quality-engineer", "wicked-garden-qe-release-readiness-engineer", "wicked-garden-qe-requirements-quality-analyst", "wicked-garden-qe-risk-assessor", "wicked-garden-qe-scenario-executor", "wicked-garden-qe-security-test-engineer", "wicked-garden-qe-semantic-reviewer", "wicked-garden-qe-snapshot-hygiene-auditor", "wicked-garden-qe-test-automation-engineer", "wicked-garden-qe-test-code-quality-auditor", "wicked-garden-qe-test-data-manager", "wicked-garden-qe-test-designer", "wicked-garden-qe-test-impact-analyzer", "wicked-garden-qe-test-oracle", "wicked-garden-qe-test-strategist", "wicked-garden-qe-testability-reviewer", "wicked-garden-qe-ui-component-test-engineer", "wicked-garden-qe-visual-regression-engineer", "wicked-garden-smaht", "wicked-garden-swarm", "wicked-garden-wickedizer", "wicked-garden-workflow", "wicked-garden-worktrees"],
+  "fork_skills": ["wicked-garden-agentic-architect", "wicked-garden-agentic-performance-analyst", "wicked-garden-agentic-safety-reviewer", "wicked-garden-archetype", "wicked-garden-classify", "wicked-garden-data-engineer", "wicked-garden-domain-coverage", "wicked-garden-domain-extractor", "wicked-garden-domain-modeler", "wicked-garden-engineering-api-documentarian", "wicked-garden-engineering-migration-engineer", "wicked-garden-engineering-solution-architect", "wicked-garden-ground", "wicked-garden-jam-brainstorm-facilitator", "wicked-garden-jam-council", "wicked-garden-mem-capture", "wicked-garden-mem-ingest", "wicked-garden-platform-compliance-officer", "wicked-garden-platform-privacy-expert", "wicked-garden-platform-security-engineer", "wicked-garden-product-a11y-expert", "wicked-garden-product-requirements-analyst", "wicked-garden-product-ui-reviewer", "wicked-garden-product-ux-designer", "wicked-garden-product-value-strategist", "wicked-garden-qe-a11y-test-engineer", "wicked-garden-qe-acceptance-test-executor", "wicked-garden-qe-acceptance-test-reviewer", "wicked-garden-qe-acceptance-test-writer", "wicked-garden-qe-ai-feature-test-engineer", "wicked-garden-qe-chaos-test-engineer", "wicked-garden-qe-code-analyzer", "wicked-garden-qe-compliance-test-engineer", "wicked-garden-qe-contract-testing-engineer", "wicked-garden-qe-coverage-archaeologist", "wicked-garden-qe-data-quality-tester", "wicked-garden-qe-e2e-orchestrator", "wicked-garden-qe-exploratory-tester", "wicked-garden-qe-flaky-test-hunter", "wicked-garden-qe-fuzz-property-engineer", "wicked-garden-qe-iac-test-engineer", "wicked-garden-qe-incident-to-scenario-synthesizer", "wicked-garden-qe-integration-test-engineer", "wicked-garden-qe-load-performance-engineer", "wicked-garden-qe-localization-test-engineer", "wicked-garden-qe-mutation-test-engineer", "wicked-garden-qe-observability-test-engineer", "wicked-garden-qe-production-quality-engineer", "wicked-garden-qe-release-readiness-engineer", "wicked-garden-qe-requirements-quality-analyst", "wicked-garden-qe-risk-assessor", "wicked-garden-qe-scenario-executor", "wicked-garden-qe-security-test-engineer", "wicked-garden-qe-semantic-reviewer", "wicked-garden-qe-snapshot-hygiene-auditor", "wicked-garden-qe-test-automation-engineer", "wicked-garden-qe-test-code-quality-auditor", "wicked-garden-qe-test-data-manager", "wicked-garden-qe-test-designer", "wicked-garden-qe-test-impact-analyzer", "wicked-garden-qe-test-oracle", "wicked-garden-qe-test-strategist", "wicked-garden-qe-testability-reviewer", "wicked-garden-qe-ui-component-test-engineer", "wicked-garden-qe-visual-regression-engineer", "wicked-garden-swarm", "wicked-garden-wickedizer", "wicked-garden-workflow", "wicked-garden-worktrees"],
   "hooks": ["SessionStart", "UserPromptSubmit", "PreToolUse", "PostToolUse", "PostToolUseFailure", "TaskCompleted", "SubagentStart", "SubagentStop", "PermissionRequest", "Notification", "PreCompact", "Stop", "SessionEnd"],
   "specialists": [
     {

package/.claude-plugin/marketplace.json

@@ -1 +1 @@
-{"name": "wicked-garden", "owner": {"name": "Mike Parcewski"}, "plugins": [{"name": "wicked-garden", "source": "./", "description": "Your coding agent already plans and swarms \u2014 Wicked Garden is the curated toolkit for what it can't: proof-not-claims gates, code links grep can't see, deterministic refactor, cross-session memory, real multi-model second opinions. Don't fight the harness; fill its gaps. The gate needs wicked-vault (a self-contained infra peer, installed directly) and re-derives through garden's built-in engine; estate/bus are opt-in layers (wicked-testing and wicked-brain are retired \u2014 QE ships in-catalog as the qe domain, memory lives in wicked-estate).", "version": "12.37.1"}], "version": "12.37.1"}
+{"name": "wicked-garden", "owner": {"name": "Mike Parcewski"}, "plugins": [{"name": "wicked-garden", "source": "./", "description": "Your coding agent already plans and swarms \u2014 Wicked Garden is the curated toolkit for what it can't: proof-not-claims gates, code links grep can't see, deterministic refactor, cross-session memory, real multi-model second opinions. Don't fight the harness; fill its gaps. The gate needs wicked-vault (a self-contained infra peer, installed directly) and re-derives through garden's built-in engine; estate/bus are opt-in layers (wicked-testing and wicked-brain are retired \u2014 QE ships in-catalog as the qe domain, memory lives in wicked-estate).", "version": "12.38.0"}], "version": "12.38.0"}

package/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "wicked-garden",
-  "version": "12.37.1",
+  "version": "12.38.0",
   "description": "Your coding agent's harness already plans and swarms; Wicked Garden is the curated toolkit for what it can't do alone. Proof, not claims \u2014 gates re-derive 'done' through the internal loom engine (fail-closed, independently attested via wicked-vault), so a self-asserted pass can't lie its way green. Relationships grep can't see \u2014 the wicked-estate code graph + injected bus/dispatch/capability/archetype edges power blast-radius & lineage. Plus deterministic multi-file refactor (wicked-patch), cross-session memory + knowledge (the mem domain over wicked-estate), real multi-model second opinions (jam:council), portable expertise as on-demand skill-refs, and evidence-gated testing (the in-catalog qe domain). It reads the shape of your work to apply the right rigor \u2014 steering, not a pipeline \u2014 then gets out of the way. The compiler stamps a self-contained evidence gate into any repo that runs with no wicked-garden installed. Don't fight the harness; fill its gaps. The evidence gate requires wicked-vault, a self-contained infra peer wicked-garden installs directly; wicked-bus is an opt-in layer \u2014 install what you'll use, the toolkit works without the rest.",
   "wicked_vault_version": "^0.5.0",
   "wicked_ledger_version": "^0.4.0",

package/.claude-plugin/specialist.json

@@ -1,6 +1,6 @@
 {
   "plugin": "wicked-garden",
-  "_comment": "Specialist roles listed here are user-facing disciplines surfaced by crew for phase routing. Internal infrastructure domains (crew, mem, persona, smaht, search, jam) are NOT listed as specialists — they are invoked directly via their consolidated domain skills, not routed through the specialist dispatch system. Specialist roles resolve to context:fork worker skills (skills/<domain>-<role>/SKILL.md) via scripts/crew/specialist_resolver.py.",
+  "_comment": "Specialist roles listed here are user-facing disciplines surfaced by crew for phase routing. Internal infrastructure domains (crew, mem, persona, search, jam) are NOT listed as specialists — they are invoked directly via their consolidated domain skills, not routed through the specialist dispatch system. Specialist roles resolve to context:fork worker skills (skills/<domain>-<role>/SKILL.md) via scripts/crew/specialist_resolver.py.",
   "specialists": [
     {
       "name": "engineering",

package/CHANGELOG.md

@@ -2,6 +2,26 @@
 
 ## [Unreleased]
 
+## [12.38.0] — 2026-09-15
+
+### Changed
+- **Wave-2 catalog batch B16 (DES-L6 B14–B17; third of the 40 `qe-*` workers): 10 qe specialist workers (`qe-mutation-test-engineer`, `qe-observability-test-engineer`, `qe-production-quality-engineer`, `qe-release-readiness-engineer`, `qe-requirements-quality-analyst`, `qe-risk-assessor`, `qe-scenario-executor`, `qe-security-test-engineer`, `qe-semantic-reviewer`, `qe-snapshot-hygiene-auditor`) go cross-CLI (fixall L6-B16; D-21 — BC-37 / BC-39).** Each drops its Claude-only frontmatter for the closed key set with `metadata.role: worker` and loses the fork sentence; `qe-security-test-engineer` (1777 → 1015) and `qe-snapshot-hygiene-auditor` (1070 → 621) descriptions are trimmed under 1024 (example blocks dropped, security's two NOT/advisor lists condensed). `qe-scenario-executor`'s "Use the Read tool to read the scenario file" reads "Open the scenario file with your file reader". The `qe-mutation-test-engineer` `VERDICT={PASS|FAIL} REVIEWER=… RUN_ID=…` output footer becomes the D-9 grammar on separate lines and leaves `verdict_spelling.legacy_footer_files` (2 → 1, leaving only `qe-visual-regression-engineer` for B17 — the verdict-spelling test fixture target, whose `legacy_footer_files` entry and `_LEGACY_FOOTER` fixture rows move together in B17). `tests/cross_cli_baseline.json` shrinks by exactly this batch (13 pairs across 10 files → 18; 0 stale; nothing added); `components.json` unchanged. BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives the 10 workers in `skillsSnapshotHanded` with 0 `seatConstraint`, and a mutation run writes `VERDICT: PASS|FAIL` on its own line.
+- **Wave-2 catalog batch B17 (DES-L6 B14–B17; the LAST 10 of the 40 `qe-*` workers): `qe-test-automation-engineer`, `qe-test-code-quality-auditor`, `qe-test-data-manager`, `qe-test-designer`, `qe-test-impact-analyzer`, `qe-test-oracle`, `qe-test-strategist`, `qe-testability-reviewer`, `qe-ui-component-test-engineer`, `qe-visual-regression-engineer` go cross-CLI — completing all 40 qe workers (fixall L6-B17; D-21 — BC-37 / BC-39).** Each drops its Claude-only frontmatter for the closed key set with `metadata.role: worker` and loses the fork sentence; `qe-test-code-quality-auditor`'s over-long description (1117 → 615) drops its `<example>` block. `qe-visual-regression-engineer`'s `VERDICT={PASS|FAIL} REVIEWER=… RUN_ID=…` output footer becomes the D-9 grammar on separate lines (`VERDICT: PASS|FAIL` / `REVIEWER:` / `RUN_ID:`) and leaves `verdict_spelling.legacy_footer_files` — which is now **EMPTY** (all 40 qe footers normalized across B14–B17). Per the B14 carry-forward, the verdict-spelling test fixture moves WITH the list: `tests/test_skill_portability.py`'s three `_LEGACY_FOOTER` rows keyed to `qe-visual-regression-engineer` and the corpus quiet entry are removed (no legacy file survives to point at), `_LEGACY_FOOTER` becomes a generic example, and a legacy-footer-shaped line is now a verdict-spelling hit in EVERY file. `tests/cross_cli_baseline.json` shrinks by exactly this batch (11 pairs across 10 files → **7**; 0 stale; nothing added) — the remaining 7 entries are non-L6 files (`mem`, `search`, `search/codebase-narrator`, `repo-learn`, `wickedizer`, cleared by their owners / B18). `components.json` unchanged. BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives the 10 workers in `skillsSnapshotHanded` with 0 `seatConstraint`, and a visual-regression run writes `VERDICT: PASS|FAIL` on its own line. With B17, every `qe-*` worker skill is cross-CLI.
+- **Wave-2 catalog batch B18 (DES-L6 — the final L6 garden lint/guard/cleanup batch): N1 CI-hard-fail when PyYAML is unimportable in a CI context, the `#1157` dispatch-lint bare-identifier gap, the `#1156` capability-router retirement, the retired `crew-*` trio stub removal, and five engineering unpaired-fence refs (fixall L6-B18; D-21 — BC-38).** `scripts/ci/validate.py` gains `ci_yaml_requirement_error()` — in a CI context (`CI` / `GITHUB_ACTIONS`) it now FAILS when PyYAML is unimportable, so the frontmatter guard can never silently fall back to the weaker stdlib approximation if a future edit drops `pip install pyyaml` from a workflow; locally the graceful stdlib fallback stays (N1, the 12.37.1 / 12.37.2 guard-durability follow-up). The `claude-dispatch` / `claude-tool-call` lint (`tests/portability_rules.garden.json`) now also catches a bare CAPITALISED first positional argument followed by more args (`Agent(Explore, prompt=…)`, `Read(FilePath, encoding)`) — the `#1157` lookahead gap; a lone `X(Cap)` prose noun (`recall replaces Grep/Glob/Agent(Explore)`) and the plurals `Task(s)` / `Agent(s)` stay quiet (adversarial rows added; 0 new corpus hits — guard hardening). The Claude-plugin capability-router is retired (`#1156`): `scripts/_capability_resolver.py` (322 lines) and `scripts/_capability_registry.py` (334 lines) are deleted, `hooks/scripts/bootstrap.py` drops SessionStart step 3b and `_resolve_capabilities()`, `scripts/_session.py` drops the unread `resolved_capabilities` field, `.claude/commands/wg-check.md` drops §4b + §12 + the summary-table row, `scripts/ci/find_orphan_agents.py` drops the `_capability_registry` root-allowlist clause, and the ADR 0001 injected-edge table notes the runtime router is retired (estate's `agent → tool` edge reads the `tool-capabilities:` frontmatter directly — no registry import). The resolver failed open and NOTHING read its result, so nothing observable changes; this supersedes the loom-extraction spec's "stays in garden" disposition (which meant "don't migrate to loom", not "never retire"). The three retired `crew-implementer` / `crew-researcher` / `crew-reviewer` module stubs (folded into `wicked-garden-governed-worker` §Creator/Neutral/Evaluator in an earlier batch) are deleted; `governed-worker`'s History block marks the three names `<!-- not-a-skill -->`; `components.json` regenerates (skills 140 → 137). Five `skills/engineering/**/refs/*.md` (architecture-template-design, debugging/process, component-template-structure, interface-template-maintenance, interface-template-structure) promote their outer three-backtick `markdown` wrapper (opener + closer) to a four-backtick fence so a three-backtick inner block nests per CommonMark (delisted from `FENCE_PAIRING_BASELINE`). Archetype role ruling: `skills/archetype/SKILL.md` stays `metadata.role: worker` (system-reminder-triggered, not user-invocable) — no change. `tests/cross_cli_baseline.json` is UNCHANGED (its remaining 7 entries are non-L6 `mem` / `search` / `search/codebase-narrator` / `repo-learn` / `wickedizer` files, cleared by their owners' PRs). Deferred with reasons: the colon-form (`wicked-garden:x:y`) token (54 corpus hits across three semantic categories — slash-invocation, bare skill-name cross-refs, and bus `source_agent` JSON payloads — needs a spec + shrink baseline), `detect_state.py` and description-historical treatment (need specs), four ambiguous-boundary fence refs (`examples-saas-trading`, `event-schemas-best-practices`, `scenario-format`, `qe-incident-to-scenario-synthesizer` — swallowed-sections / pre-wrapper classes, a focused fence follow-up), the crew fixture v3 re-vendor (no v3 upstream), and the two retirements (mem-capture / mem-ingest ride L4's mem PR; smaht is a product PR needing the user's call).
+- **The `smaht` domain is retired whole (fixall L6; BC-77 — APPROVED by the user 2026-09-15).** The pull-model context-assembly domain — its skills (`smaht` router + `briefing`/`state`/`events-import` refs, and the `discovery`/`intent`/`propose-skills` sub-skills) and the `scripts/smaht/**` package (adapters, `context_package`, `knowledge_graph`, `propose_skills`, `cheatsheet_store`) — is deleted; garden drops from 9 domains to 8. **Routing is unchanged:** the session-intent primitive that drives archetype routing (auto-detection in `hooks/scripts/prompt_submit.py::_detect_intent` + the sticky `state.intent`) lives in the hook and `_session.py`, never in smaht — the `smaht-intent` skill was only the manual OVERRIDE writer. Retiring it removes the override path (`state.intent_explicit` and the `<wg intent="X" t=N />` label are dropped; auto-detected intent was already invisible, so the observable directive stream is unchanged) while the UserPromptSubmit hook still classifies every prompt to an archetype. No production code imported `scripts/smaht` (verified); consumers are rewired: the `smaht` entry leaves the domain lists (`_agents.py`, `_validate_registry.py`, `specialist_resolver.py`, `find_orphan_agents.py`, `components.json`, `_relevance_bulkadd.py`), the `reset.py` domain map, and the routing tables in `skills/core/SKILL.md` + `docs/domains.md`; the live `wicked.garden.fact.extracted` bus event (now produced by `scripts/mem/session_fact_extractor.py`) is relabelled off the `smaht` domain; `notification.py`'s dead context-pressure flag writes (whose reader — the v2 orchestrator — was deleted in #428) are removed. The six GENERAL onboarding/install scenarios mis-filed under `scenarios/smaht/` (fresh-install, plugin-degradation, mcp-integration, the two onboarding guards, wizard-environment-scan — 0 smaht refs; they exercise core plugin bootstrap) are RELOCATED to `scenarios/core/` (preserving coverage and re-rooting `wicked-garden-wickedizer`); the 3 genuinely smaht-specific scenarios are deleted. `tests/hooks/test_prompt_submit_intent.py` is the gate: intent auto-detects, no override field/label survives, and archetype routing classifies with smaht gone. `find_orphan_agents` stays 0 orphaned; `validate.py` / registry / sync pass. Cross-lane follow-ups (NOT in this PR): `scenarios/search/lifecycle-scoring.md` (L4-owned) references the deleted `smaht.adapters.domain_adapter` — scenarios are not CI-executed, so it rides L4's cleanup; and `site/src/data/garden.ts` still lists the `smaht` domain (rides the site lane). The `_context_backend` stats folder keeps its legacy `smaht` name (renaming would orphan on-disk stats).
+- **The estate shim scopes capture proposals to the run's studio project (BC-79 reader half; #1172).** `scripts/_estate_client.py` reads `WICKED_RUN_PROJECT` (stamped on the worker env by wicked-core's WRAPPED carrier since core-ts 0.7.29) into `facets.project` before `propose(...)` sends a capture, so an approved capture lands at the run's project scope instead of the state-home default; absent or blank — an older core that does not stamp it, or a non-run invocation — leaves `facets` unchanged, preserving prior behavior. `_scope_facets_to_run_project` never mutates the caller's dict; `tests/test_estate_client.py` covers the set / override / absent / blank paths.
+
+## [12.37.2] — 2026-09-15
+
+### Changed
+- **Guard upgrade (fast-follow to the 12.37.1 frontmatter-drop incident): the SKILL.md frontmatter guard now runs the real `yaml.safe_load` in CI, not a stdlib approximation (fixall L6-guard-upgrade).** The hotfix's stdlib structural check caught the three observed shapes but a reviewer's adversarial battery found 8 cases the publish's parser REJECTS while the stdlib guard PASSED — most seriously the same frontmatter-drop class under a NESTED `metadata:` / `mandates:` key (the block every wave-2 skill carries), plus tab-indentation, mis-indented nested maps, unterminated quotes, colon-in-unquoted-value, and a col-0 sequence item mixed into the mapping. `scripts/ci/validate.py::frontmatter_yaml_error` now calls `yaml.safe_load` on every frontmatter block — the exact mirror of the skills publish's parser — so nothing the publish would reject goes green in CI, and it still flags the SILENT subclass a plain parse accepts (a bare-token list value absorbed into a `|`/`>` block scalar); a stdlib structural check remains only as a graceful fallback if PyYAML is unimportable. `pip install pyyaml` is added to `.github/workflows/validate.yml`, `test.yml` (alongside pytest) and `release.yml`'s pre-publish `test_cmd` (PyYAML is a declared garden dependency; the legs previously ran bare, which is why the first guard had to be stdlib). `tests/test_validate_frontmatter.py` gains the 8 divergence rows (each now FAILS the guard), a guard-mirrors-`yaml.safe_load` check across the battery, and the live-tree clean assertion.
+- **Hotfix (frontmatter-drop class; ship-blocker for 12.37.1 → cut as 12.37.2): three translated skills had a dropped `tool-capabilities:` key that left its value line(s) ORPHANED in the frontmatter (fixall L6-hotfix).** `skills/data-engineer/SKILL.md` (batch B8) left `  - data-query` after the key was removed, which made the frontmatter INVALID YAML (`expected <block end>, but found '<block sequence start>'`) — `yaml.safe_load` refused it, blocking the ENTIRE skills publish (rig can't republish → `baseSkillPolicy=require` refuses every launch). Two more of the same class were absorbed into a `|` block description (valid YAML, corrupt content, so `yaml.safe_load` did not catch them): `skills/agentic-safety-reviewer/SKILL.md` (B8, `  - security-scanning`) and `skills/platform-security-engineer/SKILL.md` (B6, `  - security-scanning` + `  - version-control`). All orphaned value lines are deleted (the `tool-capabilities` key was intentionally dropped in those batches; its values go with it). **Systemic guard:** `scripts/ci/validate.py` now runs a strict STRUCTURAL frontmatter check on every SKILL.md (stdlib only — garden CI runs the validate/test legs with a bare interpreter, no PyYAML, so it cannot import the publish's `yaml.safe_load`; the check reproduces the one failure mode that matters) and FAILS on an orphaned value line under a scalar key (the syntax break that stopped 12.37.1's publish) AND on a bare-token list value absorbed into a `|` block scalar (the silent subclass); `tests/test_validate_frontmatter.py` carries the adversarial rows (an orphan after a quoted scalar, an orphan absorbed into a `|` block, a nested-map orphan, and missing frontmatter all fail; clean frontmatter, a `|` block, `mandates:` lists, and legitimate prose bullets pass) plus a live-tree assertion. Root cause: the cross-CLI batches' frontmatter rewriter dropped a key's line without its indented value lines, and garden CI read frontmatter with a lenient regex, not a strict YAML parse — so it slipped through fifteen batches. A full repo sweep of all 140 SKILL.md confirms these three were the only affected files.
+- **Wave-2 catalog batch B9 (DES-L6 B11): the `jam` and `persona` domains — two routers, `jam-brainstorm-facilitator` / `jam-council`, three refs — are cross-CLI, and the `persona-agent` worker is RETIRED into `persona/refs/as.md` (fixall L6-B9; D-21 — BC-37 / BC-39).** The routers carry the closed key set with `metadata.role: router` (`jam` phases `clarify,design`; `persona` phases `*`): `jam`'s `brainstorm` and `council` sub-actions hand off to the two worker skills through Hand-off paragraphs carrying the exact asks (the routing table says "Hand-off →", the Workers section drops "forked", council seats are "separate-worker seats"); `persona`'s Source Priority names "the harness's plugin cache" instead of a `.claude/plugins` path. The two jam workers are `role: worker` (Claude hint keys, fork sentences and the pre-v12.25 `Task(subagent_type=…)` compat tails dropped — reached by name through a Hand-off; "dispatched by" → "handed off by" in their descriptions); the facilitator's `TaskCreate(` / `TaskUpdate(` session tracking reads as "track the session in the harness's task list where it has one, else your working notes" with the subject, metadata and appends named, its three `Skill(skill="wicked-garden-mem", …)` recall/store calls are Hand-off paragraphs and its Step 1b names the `wicked-garden-search` skill; the council's "subagent tier" is "the separate-worker tier" (its seat-independence Hand-off was already cross-CLI) and its outcome store is a Hand-off paragraph. `persona-agent` (a 122-line fork worker whose whole job was to run the profile the `as` action assembled) is retired: `persona/refs/as.md` Step 5 "Embody the persona" carries the profile as a ```markdown template and a new "Embodying the persona" section carries the worker's behavioral guidelines, archetype patterns, response format and task-type adaptations verbatim (harness-neutral tool wording); `skills/persona-agent/` is deleted, `components.json` regenerated (`fork_skills` 73 → 72, skills 144 → 143), `docs/domains.md` says so. `jam/refs/revisit.md`'s two mem calls are Hand-off paragraphs; `council-verdict.md` says "worker returns". `tests/test_skill_naming.py` loses the ">= 3 skills keep a `subagent_type` compat key" floor — the batches drop the key on purpose (BC-37), so the drop is not silent; the well-formedness test still runs over the two remaining declarers (`mem-capture`, `mem-ingest`) until B10 retires them. Kept on purpose: the `{topic-slug}` / `{persona}` template placeholders, description trigger phrases → the B18 colon-form token. `tests/cross_cli_baseline.json` shrinks by exactly this batch (20 pairs across 7 files → 119; 0 stale; nothing added — the deleted `persona-agent/SKILL.md` entry goes with the file). BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives both routers and the two jam workers in `skillsSnapshotHanded` with 0 `seatConstraint`, a jam `brainstorm` unit on that seat hands off to `wicked-garden-jam-brainstorm-facilitator` by name, and a persona `as` unit embodies the persona inline with no `Skill(` in its output.
+- **Wave-2 catalog batch B10 (DES-L6 B12): the fold/retire batch — `integration-discovery` RETIRED, `runtime-exec` folded into `core/refs/runtime.md`, `engineering-conformance-reviewer` folded into `engineering/refs/conformance.md`, and `smaht` (+ `discovery` / `intent` / `propose-skills` / `refs/briefing.md`), `mem-capture` and `mem-ingest` are cross-CLI (fixall L6-B10; D-21 — BC-37 / BC-39).** `wicked-garden-integration-discovery` (a `user-invocable: false`, `status: experimental` capability router nothing dispatched — the `integration-discovery` strings in `_domain_store.py` / `_session.py` / `domain_adapter.py` name the routing mechanism, not the skill) is deleted with its four refs. `wicked-garden-runtime-exec`'s body moves verbatim into `skills/core/refs/runtime.md` minus its frontmatter, H1 and the shared launcher paragraph `core/SKILL.md` already carries, and the router's Runtime section points at it; `wicked-garden-engineering-conformance-reviewer`'s body moves verbatim into `skills/engineering/refs/conformance.md`, a pointer lands at the end of the router's Knowledge modules section, `scripts/platform/guard_pipeline.py` stops telling the operator to invoke a skill that no longer exists, and `docs/outgov-per-turn.md` names the ref. The `smaht` router carries the closed key set with `metadata.role: router` (phases `*`; description 1042 → under 1024 chars — the alias line drops and `wicked-garden:search` reads as the `wicked-garden-search` skill), loses its fork sentence and the dispatch-fallback sentence, and its Quick Reference's two `Skill(skill="wicked-garden-mem", …)` calls become one Hand-off paragraph (the symbol-lookup comment now says "through the estate read-only shim", matching the grounding doctrine that landed in 12.37.1); `discovery` / `intent` / `propose-skills` are `role: module` (discovery's "Fork-worker dispatches (`Skill(skill=…)`, or the legacy `subagent_type=…` compat form)" detection bullets describe Hand-off paragraphs and worker names, propose-skills reads "the harness's session transcripts"), and `refs/briefing.md`'s fallback memory call is a Hand-off paragraph. `mem-capture` and `mem-ingest` get the closed key set with `metadata.role: worker` — bodies untouched. **Scope deviation from DES B12, disclosed:** the DES says those two workers and `smaht` RETIRE "— `mem/SKILL.md` untouched", but both clauses cannot hold — the L4-owned `skills/mem/SKILL.md` dispatches `wicked-garden-mem-ingest` and `wicked-garden-mem-capture` by name at four sites (and `mem/refs/scopes.md` at one), so deleting them dangles five identifiers in files this lane does not own and `unresolved-skill-name` fails loud on exactly that (B9's guard); `wicked-garden-smaht-intent` is a live user-override path in `hooks/scripts/prompt_submit.py` and `scripts/_session.py`, `core/SKILL.md` lists smaht in two routing tables, and `docs/domains.md` counts it among the 9 domains. Both are translated here and their retirement is owed as a follow-up that carries the L4 router edit (mem) and the hook/docs changes (smaht) in the same PR. `tests/cross_cli_baseline.json` shrinks by exactly this batch (26 pairs across 10 files → 93; 0 stale; nothing added — the deleted files' entries go with them); `.claude-plugin/components.json` regenerated (`fork_skills` 72 → 69 — `engineering-conformance-reviewer` and `integration-discovery` leave with their deleted files and `smaht` leaves the set because it is now `metadata.role: router` instead of a `context: fork`-inferred worker; `runtime-exec` was already a module); `scripts/wg/_relevance_bulkadd.py` drops the two rows for the deleted dirs and the `portability_rules.garden.json` corpus entry pinned to the retired `runtime-exec` file is re-pointed at its new home, so no corpus row names a deleted file. BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives `wicked-garden-smaht`, `wicked-garden-mem-capture` and `wicked-garden-mem-ingest` in `skillsSnapshotHanded` with 0 `seatConstraint` and none of the three retired names, and a smaht `briefing` unit on that seat hands off to `wicked-garden-mem` by name with no `Skill(` in its output.
+- **Wave-2 catalog batch B11 (DES-L6 B13): the `qe` router and its refs go cross-CLI (fixall L6-B11; D-21 — BC-37 / BC-39; C7).** The router carries the closed key set with `metadata.role: router` (phases `*`; description trimmed 1339 → 1012 chars), its "Fork workers (dispatch with the Skill tool)" section is "Workers", and the dispatch-guard names the hand-off instead of "the Skill call". Nine dispatch fences across `accept` (writer / executor / reviewer), `author`, `execute`, `insight`, `plan`, `review` become Hand-off paragraphs that keep each worker's full arg brief verbatim in a ```markdown block; the "Every id … invoke it with the Skill tool" intros read "reach it by name through a Hand-off", the shared "Dispatch uses the Skill tool on Claude Code …" fallback sentences are dropped, the `Skill(skill="wicked-garden-mem")` / `-search` recalls in `accept` / `campaign` are Hand-off paragraphs or plain skill references, and the "one `Skill(...)` call each" dispatch prose in `execute` / `plan` / `campaign` reads "one hand-off". Two arg-body tool nouns ("Read the scenario via the Read tool", "Use the Read tool to open") say "with your file reader"; the reviewer-isolation enforcement tables in `accept` / `campaign-grading` are keyed by capability ("a host that enforces `allowed-tools`") instead of naming "Claude Code"; and `integration.md`'s retained-for-compat `subagent_type` wire-field note is marked historical. **C7 (DES B13):** the router's preflight is scoped to writing actions (author / execute / intake / accept) — read-only actions (plan / insight / review) never force setup — and `plan.md`'s preflight reads "config is optional for planning; never run setup from plan". Every `.wicked-testing/` dual-read line (router :36/:68/:69, `accept` :103/:104/:124, `campaign` :250) is marked `<!-- historical -->` (the functional bash lines via a trailing `# <!-- historical -->` comment so the illustrative command still parses). No lint rule or regex changed. `tests/cross_cli_baseline.json` shrinks by exactly this batch (30 pairs across 10 files → 63; 0 stale; nothing added); `components.json` unchanged (the qe router was already a router; the 40 `qe-*` workers are B14–B17). BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives `wicked-garden-qe` in `skillsSnapshotHanded` with 0 `seatConstraint`, and a qe `accept` unit on that seat hands off to `wicked-garden-qe-acceptance-test-{writer,executor,reviewer}` by name (no `Skill(` in its output).
+- **Wave-2 catalog batch B14 (DES-L6 B14; first of the 40 `qe-*` workers): 10 qe specialist workers (`qe-a11y-test-engineer`, `qe-acceptance-test-{executor,reviewer,writer}`, `qe-ai-feature-test-engineer`, `qe-chaos-test-engineer`, `qe-code-analyzer`, `qe-compliance-test-engineer`, `qe-contract-testing-engineer`, `qe-coverage-archaeologist`) go cross-CLI (fixall L6-B14; D-21 — BC-37 / BC-39).** Each drops its Claude-only frontmatter (`context: fork`, `model`, `effort`, `max-turns`, `allowed-tools`, `phase_relevance`, `archetype_relevance`) for the closed key set with `metadata.role: worker`, and the "This skill is designed to run as an isolated worker …" fork sentence is removed. Four over-long descriptions are trimmed under 1024 (a11y 1160 → 772, ai-feature 1593 → 991, chaos 1046 → 508, compliance 1289 → 807 — the illustrative `<example>` blocks removed, ai-feature's NOT/Cross-ref lists condensed). The three acceptance workers' dispatch is translated: `qe-acceptance-test-executor` and `-writer`'s `Skill(skill="wicked-garden-mem", …)` recalls are Hand-off paragraphs, the executor's "Use Bash tool" / "Use Read tool" evidence rows say "your shell" / "your file reader", and the reviewer's enforcement note is keyed by capability ("a host that enforces `allowed-tools`") instead of naming "Claude Code". The `qe-chaos-test-engineer` and `qe-compliance-test-engineer` `VERDICT={PASS|FAIL} REVIEWER=… RUN_ID=…` output footers become the D-9 grammar on separate lines (`VERDICT: PASS|FAIL` / `REVIEWER: …` / `RUN_ID: …`) and leave `verdict_spelling.legacy_footer_files` (7 → 5); the verdict-spelling test fixture + corpus entry re-point from chaos to `qe-visual-regression-engineer` (a survivor through B17). B14 also converted the LAST `Skill(skill="wicked-garden-…")` dispatch call-form in the catalog, so `test_reference_extraction_is_not_vacuous`'s Skill-ref non-vacuity assert inverts to the zero-invariant (a reappearing `Skill(` ref is now a regression; the launcher-count guards keep the test non-vacuous) — the B9 review's N2 item, retired by the batch that made it come due. `tests/cross_cli_baseline.json` shrinks by exactly this batch (21 pairs across 10 files → 42; 0 stale; nothing added); `components.json` unchanged (all 10 were already `context: fork`-inferred workers). BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives the 10 workers in `skillsSnapshotHanded` with 0 `seatConstraint`; a qe `accept` unit on that seat hands off to `wicked-garden-qe-acceptance-test-{writer,executor,reviewer}` by name (no `Skill(` in its output) and a chaos/compliance run writes `VERDICT: PASS|FAIL` on its own line.
+- **Wave-2 catalog batch B15 (DES-L6 B14–B17; second of the 40 `qe-*` workers): 10 qe specialist workers (`qe-data-quality-tester`, `qe-e2e-orchestrator`, `qe-exploratory-tester`, `qe-flaky-test-hunter`, `qe-fuzz-property-engineer`, `qe-iac-test-engineer`, `qe-incident-to-scenario-synthesizer`, `qe-integration-test-engineer`, `qe-load-performance-engineer`, `qe-localization-test-engineer`) go cross-CLI (fixall L6-B15; D-21 — BC-37 / BC-39).** Each drops its Claude-only frontmatter for the closed key set with `metadata.role: worker` and loses the "This skill is designed to run as an isolated worker …" fork sentence; `qe-incident-to-scenario-synthesizer`'s over-long description (1165 → 643) drops its illustrative `<example>` block. The `qe-flaky-test-hunter` and `qe-iac-test-engineer` `VERDICT={PASS|FAIL} REVIEWER=… RUN_ID=…` output footers become the D-9 grammar on separate lines (`VERDICT: PASS|FAIL` / `REVIEWER:` / `RUN_ID:`), and `qe-incident-to-scenario-synthesizer`'s concrete `VERDICT=PASS …` footer becomes `VERDICT: PASS` + separate lines; all three leave `verdict_spelling.legacy_footer_files` (5 → 2, leaving `qe-mutation-test-engineer` (B16) and `qe-visual-regression-engineer` (B17, the verdict-spelling test fixture target)). No dispatch/prose in these 10 (a full ungrouped sweep confirmed it); no test-fixture move (the fixture points at visual-regression, which stays legacy). `tests/cross_cli_baseline.json` shrinks by exactly this batch (11 pairs across 10 files → 31; 0 stale; nothing added); `components.json` unchanged. BM-style check for the wave-2 P2b/BM re-run: a non-claude seat receives the 10 workers in `skillsSnapshotHanded` with 0 `seatConstraint`, and a flaky/iac/incident run writes `VERDICT: …` on its own line.
+
 ## [12.37.1] — 2026-09-14
 
 ### Changed
@@ -908,6 +928,8 @@ through `v9.2.18`. Major themes across that span:
 
 For each of those releases, see the corresponding tag in git history.
 
-[Unreleased]: https://github.com/mikeparcewski/wicked-garden/compare/v12.37.1...HEAD
+[Unreleased]: https://github.com/mikeparcewski/wicked-garden/compare/v12.38.0...HEAD
+[12.38.0]: https://github.com/mikeparcewski/wicked-garden/compare/v12.37.2...v12.38.0
+[12.37.2]: https://github.com/mikeparcewski/wicked-garden/compare/v12.37.1...v12.37.2
 [12.37.1]: https://github.com/mikeparcewski/wicked-garden/compare/v12.37.0...v12.37.1
 [12.37.0]: https://github.com/mikeparcewski/wicked-garden/compare/v12.36.0...v12.37.0

package/docs/adr/0001-code-relationship-graph-engine.md

@@ -30,7 +30,7 @@ reference, so neither `grep` nor a static call-graph can see them:
 |---|---|
 | event producer → consumer | `emit_event("wicked.x")` + `_bus_consumers.json` `event_filter` |
 | command/Task → agent | `subagent_type:` frontmatter |
-| agent → tool | `tool-capabilities` + `_capability_registry` |
+| agent → tool | `tool-capabilities` frontmatter (the `_capability_registry` runtime router was retired 2026-09, #1156) |
 | hook event → script | `hooks.json` |
 | archetype → playbook/gate | `archetypes.json` + `refs/` |
 

package/docs/domains.md

@@ -6,7 +6,7 @@ Work is shaped by the **archetypes** (triage, explore, specify, decide, ship, re
 
 The plugin is **skills-only** (v12.25). Each domain is one consolidated, user-invokable skill — `skills/{domain}/SKILL.md`, frontmatter name `wicked-garden-{domain}` — that routes to its **actions** (the former `/wicked-garden:{domain}:{command}` surfaces) and loads rubrics on demand from `skills/{domain}/refs/` and nested knowledge modules. The former dispatch agents are **fork worker skills** — `skills/{domain}-{role}/SKILL.md` with `context: fork`, frontmatter name `wicked-garden-{domain}-{role}` — reached by `Skill()`/`Task()` dispatch, never invoked as entry points. Invoke a domain action as `/wicked-garden-{domain} {action} [args]` (or just describe what you want — the skill descriptions carry the triggers).
 
-There are **9 domains**: engineering, platform, product, data, jam, search, agentic, persona, smaht.
+There are **8 domains**: engineering, platform, product, data, jam, search, agentic, persona.
 
 > **v12 cleanup (ADR 0002).** Most former domain commands were *rubric-wrappers* — a checklist the agent already applies. These were **collapsed**: the rubric moved to an on-demand `skills/{domain}/refs/{name}.md` and the action now loads it and works **inline** (no `Task` dispatch hop). Dispatch is kept only where it earns it — real parallelism (multiple lenses at once), a real external tool, or an independent gate. The dispatch-only agents that nothing reaches anymore were removed; workers still referenced by a surviving action, skill, scenario, or the specialist registry live on as fork skills. Capability is preserved; only the token-burning hop is gone.
 
@@ -159,22 +159,7 @@ They do not overlap; pick by target:
 | `define` | Create or update a custom persona |
 | `list` | List all available personas (Methodology vs Generic tiers) |
 
-**Fork worker**: `wicked-garden-persona-agent` (the `as` action builds its prompt from the registry at runtime and dispatches to it — kept). `define` runs inline from `refs/define.md`; the no-op `persona:submit` stub was deleted.
-
-## smaht — Context Assembly
-
-`wicked-garden-smaht` — on-demand context assembly over the wicked-estate knowledge layer and the search index. A pull-model skill — archetypes and subagents call it when they need a briefing, rather than pushing context onto every prompt.
-
-| Action | What It Does |
-|--------|-------------|
-| `briefing` | What happened since the last session — recent events and updates |
-| `state` | Snapshot and report current session state |
-| `events-import` | Import existing domain JSON records into the event log |
-| `intent` | Set or inspect the active session intent (nested `intent` sub-skill) |
-
-> `briefing` and `events-import` are real event-store tools and stay; `state` runs inline from `refs/state.md` (its dead v6 half dropped). Nested sub-skills: `discovery`, `intent`, `propose-skills`.
-
----
+**Worker**: none — the former `wicked-garden-persona-agent` worker is retired into `refs/as.md` (wave-2 B9): the `as` action builds the profile from the registry at runtime and embodies it inline. `define` runs inline from `refs/define.md`; the no-op `persona:submit` stub was deleted.
 
 ## How archetypes invoke domains
 

package/docs/getting-started.md

@@ -77,7 +77,7 @@ All ten are available: `triage`, `explore`, `specify`, `decide`, `ship`, `review
 
 ### 3. Use a Domain Skill Directly
 
-Archetypes invoke the domain skills (engineering, platform, product, data, jam, search, agentic, persona, smaht) under the hood, but you can call them directly when you want a single focused pass. Each domain is one skill with routed actions:
+Archetypes invoke the domain skills (engineering, platform, product, data, jam, search, agentic, persona) under the hood, but you can call them directly when you want a single focused pass. Each domain is one skill with routed actions:
 
 ```bash
 /wicked-garden-engineering review                          # senior-perspective code review
@@ -193,7 +193,7 @@ To list everything:
 
 When you submit a prompt, a `UserPromptSubmit` hook runs the archetype detector and emits a steering reminder for the matched work-shape(s). The selected archetype then drives its own phase shape — there is no universal pipeline forcing every kind of work through the same gates.
 
-A context assembly layer called **smaht** enriches prompts with relevant context — recent memory, active work, native tasks, and code intelligence. Every `TaskCreate` / `TaskUpdate` carries a structured metadata envelope (`chain_id`, `event_type`, `source_agent`, `phase`, `archetype`) validated by a PreToolUse hook, and a SubagentStart hook injects the matching procedure bundle — R1–R6 bulletproof standards for coding tasks, the Gate Finding Protocol for review findings, and per-role procedures otherwise.
+Every `TaskCreate` / `TaskUpdate` carries a structured metadata envelope (`chain_id`, `event_type`, `source_agent`, `phase`, `archetype`) validated by a PreToolUse hook, and a SubagentStart hook injects the matching procedure bundle — R1–R6 bulletproof standards for coding tasks, the Gate Finding Protocol for review findings, and per-role procedures otherwise.
 
 Gates re-derive every claim through wicked-loom — which re-runs the verifier via wicked-vault — and fail closed if loom is unavailable: "done" is *re-derived from evidence*, never self-asserted. Hard gates (incident mitigate, migrate cutover, review final-verdict) require explicit human approval; discrete gates auto-pass only when their produces contract is met.
 

package/docs/outgov-per-turn.md

@@ -136,7 +136,7 @@ Hermetic tests — no live estate needed, fixture rules dirs only:
 - `hooks/scripts/stop.py` — `_check_outgov_compliance` (per-turn advisory)
 - `scripts/platform/guard_pipeline.py` — Check 6 (`check_outgov_pattern`),
   `compute_rules_content_hash`, `write_rules_provenance`, the `hash` CLI
-- `skills/engineering-conformance-reviewer/SKILL.md` — the agent-half that
+- `skills/engineering/refs/conformance.md` — the agent-half rubric that
   evaluates Check 6's surfaced rules semantically
 - wicked-core `crates/wicked-governance` — the graph side: `rules ingest`,
   `rules fanout` (the manifest + store split), `rules.recall`'s Rust twin

package/hooks/scripts/bootstrap.py

@@ -2,8 +2,7 @@
 """
 SessionStart hook — wicked-garden unified bootstrap.
 
-Consolidates: crew session_start, mem session_start,
-search session_start, smaht session_start.
+Consolidates: crew session_start, mem session_start, search session_start.
 
 Flow:
 1. Read config from ~/.something-wicked/wicked-garden/config.json
@@ -305,22 +304,6 @@ def _load_agents():
         return 0, {}
 
 
-def _resolve_capabilities(agents, config=None):
-    """Run capability resolution for pre-loaded agents.
-
-    Returns dict[agent_name, tool_list] or empty dict on any error.
-    Fails open — never blocks bootstrap.
-    """
-    try:
-        from _capability_resolver import resolve_all_agents, discover_mcp_servers
-
-        mcp_servers = discover_mcp_servers()
-        return resolve_all_agents(agents, config, mcp_servers)
-    except Exception as e:
-        print(f"[wicked-garden] capability resolution error: {e}", file=sys.stderr)
-        return {}
-
-
 def _find_active_crew_project(workspace: str = ""):
     """Return (project_data, project_name) for most recently updated active crew project
     scoped to the given workspace.
@@ -1306,11 +1289,6 @@ def main():
         if state is not None:
             state.update(agents_loaded=agents_loaded)
 
-        # 3b. Resolve tool-capabilities for agents that declare them
-        resolutions = _resolve_capabilities(agents_dict, config)
-        if resolutions and state is not None:
-            state.update(resolved_capabilities=resolutions)
-
         # 4. Load last crew project for this workspace
         workspace = os.environ.get("CLAUDE_PROJECT_NAME") or Path.cwd().name
         project_data, project_name = _find_active_crew_project(workspace)
@@ -1455,11 +1433,8 @@ def main():
                 is_active = active_phase not in ("", "complete", "done", "archived")
                 state.update(active_project_id=project_name if is_active else None)
 
-                # v9.2.4: also write active_chain_id so smaht's chain-aware
-                # event scoring (CLAUDE.md "Chain-aware smaht scoring") actually
-                # works. Field declared in v9.2.3, producer was missing — events
-                # got the 0.1 baseline instead of the documented 0.8+ chain
-                # boost. Format mirrors the chain_id convention in CLAUDE.md
+                # v9.2.4: also write active_chain_id so chain-aware event
+                # scoring works. Format mirrors the chain_id convention in CLAUDE.md
                 # "Native Tasks as Dual-Purpose Event Queue":
                 #   {slug}.root            for the root chain
                 #   {slug}.{phase}         per active phase

package/hooks/scripts/notification.py

@@ -2,12 +2,11 @@
 """
 Notification hook — wicked-garden context limit adaptation.
 
-Issue #343: Detect context limit warnings and adjust smaht behavior.
+Issue #343: Detect context limit warnings and warn Claude.
 
 When Claude Code emits notifications about approaching context limits,
-this hook adjusts session state to:
-  - Force smaht to use the HOT path (skip expensive adapters)
-  - Reduce context assembly scope
+this hook returns a systemMessage nudging Claude to:
+  - Minimize context — keep only what the current step needs
   - Encourage delegation over inline execution
   - Log the context pressure event for observability
 
@@ -80,32 +79,14 @@ def _is_context_limit_notification(payload: dict) -> bool:
 # ---------------------------------------------------------------------------
 
 def _handle_context_limit(payload: dict) -> str:
-    """Adapt smaht behavior when context limits are approaching.
+    """Warn Claude when context limits are approaching.
 
-    Sets session state flags that smaht's orchestrator checks:
-    - context_pressure: True — forces HOT path (skip expensive adapters)
-    - context_pressure_since: ISO timestamp
-    - force_delegation: True — encourages delegation over inline work
-
-    Returns a systemMessage with guidance for Claude.
+    Returns a systemMessage with guidance for Claude (prefer delegation,
+    keep responses concise, save critical context before compaction).
     """
     _log("notification", "warn", "context.limit_approaching",
          detail={"payload_type": payload.get("type", ""), "message": (payload.get("message", "") or "")[:100]})
 
-    # Update session state
-    try:
-        from _session import SessionState
-        from datetime import datetime, timezone
-
-        state = SessionState.load()
-        state.update(
-            context_pressure=True,
-            context_pressure_since=datetime.now(timezone.utc).isoformat().replace("+00:00", "Z"),
-            force_delegation=True,
-        )
-    except Exception:
-        pass
-
     # Name the memory surface (estate memory.capture). Fail-open to the
     # same wording.
     try:
@@ -117,8 +98,8 @@ def _handle_context_limit(payload: dict) -> str:
     return json.dumps({
         "systemMessage": (
             "[Context Pressure] Context limit is approaching. Adapting behavior:\n"
-            "1. Smaht context assembly is now in HOT-path-only mode (minimal context injection)\n"
-            "2. Prefer delegation to subagents via Task() over inline execution\n"
+            "1. Minimize context — keep only what the current step needs\n"
+            "2. Prefer delegation to subagents over inline execution\n"
             "3. Keep responses concise — avoid large code dumps\n"
             "4. If in a crew project, consider completing the current phase before starting new work\n"
             f"5. Use {_mem_target} to save critical context before compaction"

package/hooks/scripts/prompt_submit.py

@@ -607,7 +607,7 @@ def _check_onboarding_gate(prompt: str) -> str | None:
 # ---------------------------------------------------------------------------
 # The intent variable names what five overlapping classifiers were collectively
 # detecting. Auto-detected on turn 1-2 from existing complexity/risk signals,
-# sticky for the session, overridable via the wicked-garden-smaht-intent skill.
+# sticky for the session (auto-detected from the first turn's prompt).
 # Hooks gate directive emission on intent. Design: brainstorms/v10-session-01-intent-and-hook-gating.md.
 
 _INTENT_VALUES = ("simple-edit", "feature", "rigor", "research")
@@ -637,15 +637,13 @@ def _detect_intent(prompt: str, complexity: float, is_risky: bool, state) -> str
     """Auto-detect session intent from prompt + complexity/risk signals.
 
     Called once per session on turn 1-2. Result is sticky — subsequent turns
-    read state.intent directly. User overrides via the wicked-garden-smaht-intent
-    skill, which sets state.intent_explicit=True.
+    read state.intent directly.
 
     Detection priority:
       1. Rigor invocations (wicked-garden-crew* skills, /wg-issue, /wg-test) → rigor
       2. Research signals (explain how, why does, walk me through) → research
       3. Risk OR complexity ≥ threshold OR technical verb → feature
-      4. Default → simple-edit (conservative; user can raise it via the
-         wicked-garden-smaht-intent skill)
+      4. Default → simple-edit (conservative)
     """
     if not prompt or not prompt.strip():
         return "simple-edit"
@@ -679,7 +677,7 @@ def _ensure_intent_set(prompt: str, state, complexity: float, is_risky: bool) ->
     ``state.intent`` for a turn. On turn ≤ _AUTO_DETECT_TURN_LIMIT and
     intent is None → detect via ``_detect_intent``, persist via
     ``state.update``, return. On later turns or when intent is already
-    set → return existing value. Never overwrites an explicit override.
+    set → return existing value.
 
     Returns the effective intent value (always a string from _INTENT_VALUES).
     """
@@ -696,30 +694,26 @@ def _ensure_intent_set(prompt: str, state, complexity: float, is_risky: bool) ->
     detected = _detect_intent(prompt, complexity, is_risky, state)
     try:
         if state:
-            state.update(intent=detected, intent_explicit=False)
+            state.update(intent=detected)
     except Exception:
         pass  # fail open — return value still usable for this turn
     return detected
 
 
-def _build_intent_directive(intent: str, turn_count: int, explicit: bool, state=None) -> str:
+def _build_intent_directive(intent: str, turn_count: int, state=None) -> str:
     """Build the system-reminder directive for a given intent.
 
     Output rules (brainstorm decision):
-      simple-edit + auto-detect → empty string (silence is the signal)
-      simple-edit + explicit    → bare label only ("user said simple-edit")
-      feature                   → synthesis directive (+ label if explicit)
-      research                  → synthesis directive (+ label if explicit)
-      rigor                     → synthesis directive + chain context (+ label if explicit)
-
-    The label-only-on-explicit rule prevents confirmation-bias drift —
-    auto-detected intent stays invisible to the model so it doesn't
+      simple-edit → empty string (silence is the signal)
+      feature     → synthesis directive
+      research    → synthesis directive
+      rigor       → synthesis directive + chain context
+
+    Auto-detected intent stays invisible to the model so it doesn't
     second-guess the framework's silent classification.
     """
-    label = f'<wg intent="{intent}" t={turn_count} />' if explicit else ""
-
     if intent == "simple-edit":
-        return label  # empty when auto-detected
+        return ""  # silence is the signal
 
     # synthesis directive shared by feature / research / rigor.
     # Fail-open to a matching inline copy if the backend module cannot be
@@ -748,10 +742,7 @@ def _build_intent_directive(intent: str, turn_count: int, explicit: bool, state=
         )
     directive_lines.append("Only after grounding is complete, answer the original prompt.")
 
-    directive = "\n".join(directive_lines)
-    if label:
-        return f"{label}\n{directive}"
-    return directive
+    return "\n".join(directive_lines)
 
 
 # ---------------------------------------------------------------------------
@@ -793,7 +784,7 @@ def _build_wip_recovery_block(session_id: str, project: str) -> str:
     # (current_task, decisions, file_scope, active_constraints, open_questions)
     # are gone. If nothing is in-progress, emit nothing.
     if not in_progress_tasks:
-        _log("smaht", "debug", "wip_recovery.empty")
+        _log("prompt", "debug", "wip_recovery.empty")
         return ""
 
     parts = ["## [Recovery] Work in progress\n"]
@@ -808,7 +799,7 @@ def _build_wip_recovery_block(session_id: str, project: str) -> str:
     if len(block) > 1000:
         block = block[:997] + "..."
 
-    _log("smaht", "debug", "wip_recovery.done", detail={"chars": len(block)})
+    _log("prompt", "debug", "wip_recovery.done", detail={"chars": len(block)})
     return block
 
 
@@ -818,7 +809,7 @@ def _build_wip_recovery_block(session_id: str, project: str) -> str:
 
 def main():
     _t0 = time.monotonic()
-    _log("smaht", "debug", "hook.start")
+    _log("prompt", "debug", "hook.start")
 
     try:
         raw = sys.stdin.read()
@@ -938,7 +929,6 @@ def main():
         _log("prompt", "debug", "intent.resolved",
              detail={
                  "intent": intent,
-                 "explicit": bool(getattr(state, "intent_explicit", False)),
                  "complexity": round(complexity, 2),
                  "risk": is_risky,
                  "turn": turn_count,
@@ -950,12 +940,10 @@ def main():
         # block runs from the onboarding-pending branch where no intent was
         # computed) drives directive emission. Auto-detected simple-edit
         # turns produce an empty directive; feature/research/rigor produce
-        # the synthesis directive. Explicit overrides additionally echo a
-        # bare label so the model knows the user steered the framework.
+        # the synthesis directive.
         _intent = getattr(state, "intent", None) if state else None
-        _intent_explicit = bool(getattr(state, "intent_explicit", False)) if state else False
         intent_directive = _build_intent_directive(
-            _intent or "simple-edit", turn_count, _intent_explicit, state
+            _intent or "simple-edit", turn_count, state
         )
 
         # WIP recovery: surface native in-progress tasks (v5 ticket rail gone)
@@ -1024,8 +1012,8 @@ def main():
 
         merged_context = f"<system-reminder>\n{chr(10).join(all_parts)}\n</system-reminder>"
 
-        _log("smaht", "debug", "prompt.intent_directive",
-             detail={"intent": _intent, "explicit": _intent_explicit, "turn": turn_count,
+        _log("prompt", "debug", "prompt.intent_directive",
+             detail={"intent": _intent, "turn": turn_count,
                      "directive_bytes": len(intent_directive.encode("utf-8"))})
 
         output = {
@@ -1035,7 +1023,7 @@ def main():
             },
             "continue": True,
         }
-        _log("smaht", "debug", "hook.end", ms=int((time.monotonic() - _t0) * 1000))
+        _log("prompt", "debug", "hook.end", ms=int((time.monotonic() - _t0) * 1000))
         print(json.dumps(output))
 
     except Exception as e:

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "wicked-garden",
-  "version": "12.37.1",
+  "version": "12.38.0",
   "type": "module",
   "description": "Curated toolkit for what AI coding agents can't do alone: evidence-gated work, codegraph relationships, deterministic multi-file refactor, multi-model second opinions, and cross-session memory. Installs as a Claude Code plugin.",
   "keywords": [

package/scripts/_agents.py

@@ -367,7 +367,7 @@ def _parse_simple_yaml(text: str) -> dict[str, Any]:
 
 # Known plugin domains (mirrors .claude-plugin/components.json "domains")
 _DOMAINS = ("agentic", "crew", "data", "engineering", "jam", "mem", "persona",
-            "platform", "product", "qe", "search", "smaht")
+            "platform", "product", "qe", "search")
 
 
 def _infer_domain(path: Path) -> str:

package/scripts/_bus.py

@@ -290,9 +290,9 @@ BUS_EVENT_MAP: Dict[str, Dict[str, str]] = {
         "subdomain": "crew.yolo",
         "description": "Yolo auto-approval revoked due to scope-increase mutation (audit + observability)",
     },
-    # Smaht domain — fact_extractor.py → brain auto-memorize subscriber
+    # Garden fact stream — scripts/mem/session_fact_extractor.py → auto-memorize subscriber
     "wicked.garden.fact.extracted": {
-        "domain": "smaht",
+        "domain": "wicked-garden",
         "subdomain": "facts",
         "description": "Structured fact extracted from conversation (consumed by the garden-run auto-memorize drain -> estate memory)",
     },
@@ -799,7 +799,7 @@ def _resolve_bus_db_path() -> Optional[str]:
 def tail_events(limit: int = 10) -> List[Dict[str, Any]]:
     """Return the most recent ``limit`` bus events, newest first. Read-only.
 
-    This is a *display* helper (e.g. the wicked-garden-smaht skill's state action). It does
+    This is a *display* helper (e.g. a session-state inspection). It does
     NOT register a subscriber, advance a cursor, or mutate the bus in any way:
     the wicked-bus ``replay`` command resets a cursor rather than returning
     rows, and ``list`` returns subscribers, so neither yields event rows

package/scripts/_domain_store.py

@@ -21,7 +21,7 @@ Usage:
 
     # For truly ephemeral files (caches, temp session state) only:
     from _domain_store import get_local_path
-    cache_dir = get_local_path("wicked-smaht", "cache", "context7")
+    cache_dir = get_local_path("wicked-crew", "cache", "sessions")
 
 hook_mode=True:
     When True, skips integration-discovery entirely and operates local-only.
@@ -71,7 +71,7 @@ DOMAIN_MCP_PATTERNS: dict[str, list[str]] = {
     ],
     "wicked-crew": ["jira", "linear", "github", "rally", "ado", "azure-devops", "project"],
     "wicked-garden:search": ["elasticsearch", "algolia", "typesense", "search"],
-    # wicked-smaht, wicked-qe, wicked-patch: local-only (no MCP patterns)
+    # wicked-qe, wicked-patch: local-only (no MCP patterns)
 }