chore(evals): migrate to Luna and address plugin failures

.agents/skills/darrow-guide/evals/fixtures/setup.sh

@@ -29,7 +29,7 @@ for category in "$source_root"/plugins/*; do
   mkdir -p "plugins/$(basename "$category")"
   if [ -f "$category/README.md" ]; then cp "$category/README.md" "plugins/$(basename "$category")/README.md"; fi
   for plugin in "$category"/*; do
-    [ -d "$plugin" ] || continue
+    [ -d "$plugin" ] && [ -f "$plugin/README.md" ] || continue
     destination="plugins/$(basename "$category")/$(basename "$plugin")"
     mkdir -p "$destination/.claude-plugin" "$destination/.codex-plugin"
     cp "$plugin/README.md" "$destination/README.md"

.github/workflows/python-quality.yml

@@ -95,7 +95,7 @@ jobs:
       - uses: astral-sh/setup-uv@v6
         with:
           enable-cache: true
-      - run: scripts/check-python --package plugins/foundation/darrow-skill-authoring/skills/author-agent-skill/backend
+      - run: scripts/check-python --package plugins/foundation/darrow-skill-authoring/backend
 
   discovery-windows:
     name: Discovery Python ${{ matrix.python-version }} on windows-latest
@@ -119,7 +119,7 @@ jobs:
       - uses: astral-sh/setup-uv@v6
         with:
           enable-cache: true
-      - run: scripts/check-python --package plugins/capability/darrow-discovery/skills/plan-implementation/backend
+      - run: scripts/check-python --package plugins/capability/darrow-discovery/skills/work-through-decisions/backend
 
   verification-windows:
     name: Verification Python ${{ matrix.python-version }} on windows-latest
@@ -485,7 +485,7 @@ jobs:
         with:
           python-version: "3.13"
       - uses: astral-sh/setup-uv@v6
-      - run: uv run --quiet --frozen --no-dev --project plugins/foundation/darrow-skill-authoring/skills/author-agent-skill/backend python plugins/foundation/darrow-skill-authoring/skills/author-agent-skill/backend/tests/fresh_install.py
+      - run: uv run --quiet --frozen --no-dev --project plugins/foundation/darrow-skill-authoring/backend python plugins/foundation/darrow-skill-authoring/backend/tests/fresh_install.py
 
   observability-windows:
     name: Observability Langfuse Python ${{ matrix.python-version }} on windows-latest
@@ -543,7 +543,7 @@ jobs:
         with:
           python-version: "3.13"
       - uses: astral-sh/setup-uv@v6
-      - run: uv run --quiet --frozen --no-dev --project plugins/capability/darrow-discovery/skills/plan-implementation/backend python plugins/capability/darrow-discovery/skills/plan-implementation/backend/tests/fresh_install.py
+      - run: uv run --quiet --frozen --no-dev --project plugins/capability/darrow-discovery/skills/work-through-decisions/backend python plugins/capability/darrow-discovery/skills/work-through-decisions/backend/tests/fresh_install.py
 
   verification-fresh-install:
     name: Fresh install verification on ${{ matrix.os }}

AGENTS.md

@@ -37,8 +37,12 @@ between capabilities and orchestration, read [`docs/design.md`](docs/design.md).
 
 ## Skill development
 
-For skill creation, revision, or validation, follow
-[`author-agent-skill`](plugins/foundation/darrow-skill-authoring/skills/author-agent-skill/SKILL.md).
+For new skill creation, follow
+[`create-agent-skill`](plugins/foundation/darrow-skill-authoring/skills/create-agent-skill/SKILL.md).
+For a read-only skill audit, follow
+[`audit-agent-skill`](plugins/foundation/darrow-skill-authoring/skills/audit-agent-skill/SKILL.md).
+Implement audit findings or other changes to an existing skill as ordinary
+engineering work against the findings and this repository's instructions.
 Add or adjust the applicable invariant under `docs/specs/` before implementation.
 Keep contextual judgment in the skill. Put repeatable, error-prone command and
 tool-protocol mechanics behind narrow bundled scripts so the model supplies

CONTRIBUTING.md

@@ -36,8 +36,11 @@ plugin content. In particular:
 
 - Add or adjust the applicable invariant under `docs/specs/` before
   implementation.
-- For skill creation or revision, follow
-  [`author-agent-skill`](plugins/foundation/darrow-skill-authoring/skills/author-agent-skill/SKILL.md).
+- For skill creation, follow
+  [`create-agent-skill`](plugins/foundation/darrow-skill-authoring/skills/create-agent-skill/SKILL.md).
+  For a read-only skill audit, follow
+  [`audit-agent-skill`](plugins/foundation/darrow-skill-authoring/skills/audit-agent-skill/SKILL.md).
+  Implement audit findings under the repository's ordinary engineering rules.
 - Keep each plugin self-contained. Never reference files outside a plugin or
   assume a sibling plugin is installed.
 - Use imperative Conventional Commits without trailing periods, and no AI

docs/eval-development.md

@@ -16,6 +16,11 @@ constraints for producing and interpreting that evidence.
 - Prefer observable repository state, external effects, and user-visible
   outcomes over prose. When free-form text is the only public seam, assert the
   smallest semantic decision rather than parsing a complete explanation.
+- For a free-form saved document, use `semantic_artifact` to grade the artifact
+  itself. Its repository-relative `path` may contain one `*` in the filename and
+  must resolve to exactly one regular file. Keep deterministic checks for file
+  presence, status, and structural validation; `semantic_output_checks` grade
+  only the final response.
 - Give one case one concrete state and one decision. When success would require
   recognizing several conditional branches in one free-form answer, split the
   branches into separate concrete-state cases.
@@ -67,6 +72,11 @@ constraints for producing and interpreting that evidence.
 - A Codex case without the placeholder remains an implicit-discovery probe and
   requires a completed mounted-skill body read. Do not use the explicit path to
   make implicit selection pass.
+  A complete read of the byte-identical marketplace staging copy of an
+  installed plugin skill also counts: the staging path must come from that
+  eval's installed marketplace catalog, and the observed output must match
+  the installed body. Prepared target skill drafts elsewhere in the fixture
+  never count as activation.
 - Missing, repeated, malformed, or unverified observation evidence stays
   unknown. A failed compound shell command may still prove an earlier skill
   read only when the command names a mounted skill path and its output contains
@@ -237,9 +247,8 @@ evaluation that stops at the first failure. `--only <question-id>` and
 before continuing. Fixtures snapshot current public documentation, manifests,
 and relevant code while excluding inventory, hidden checks, and delivery
 conclusions from the participant repository.
-The guide driver pins semantic grading to Codex `gpt-5.6-terra` / medium;
-the shared runner's default lightweight grader remains unchanged. The route
-was calibrated against retained correct and reversed-delegation diagrams.
+The guide driver pins semantic grading to Codex `gpt-6-luna` / medium;
+the shared runner's default lightweight grader remains unchanged.
 
 ## Live-run controls
 
@@ -287,7 +296,7 @@ preserving the selected cases.
 
 Codex runs use independent defaults for each eval role:
 
-- candidate: `gpt-5.6-terra` at `medium` effort;
+- candidate: `gpt-6-luna` at `medium` effort;
 - advisory quality judge: `gpt-5.6-sol` at `low` effort;
 - gating semantic-output grader: `gpt-5.6-luna` at `low` effort.