refactor: consume shared eval-kit

docs/design/promptfoo-codex-eval-pilot.md

@@ -116,24 +116,20 @@ This design follows the docs for these reasons:
 
 The implemented boundary separates generic mechanics from the consuming suite:
 
-```text
-packages/eval-kit/      # generic CLI, schemas, Promptfoo helpers, bundled prompts
-evals/                  # technical-design cases, adapter, rubrics, tests, results
-```
-
-Generic package name:
-
 ```text
 @agentic-workflow-kit/eval-kit
+evals/                  # technical-design cases, adapter, rubrics, tests, results
 ```
 
-This package is internal infrastructure for the `technical-design` repo. It must stay private and
-must not be treated as part of the product surface shipped to users. Root `pnpm eval:*` scripts are
-the consumer command surface and call the `eval-kit` binary with `evals/eval-kit.config.json`.
+`@agentic-workflow-kit/eval-kit` is shared infrastructure consumed from the GitHub tag pinned in
+`package.json` and `pnpm-lock.yaml`. It must stay private and must not be treated as part of the
+product surface shipped to users. Root `pnpm eval:*` scripts are the consumer command surface and
+call the installed `eval-kit` binary with `evals/eval-kit.config.json`.
 
 Eval-specific material is split by ownership:
 
-- generic runner code, portable schemas, and bundled Promptfoo prompts -> `packages/eval-kit/`;
+- generic runner code, portable schemas, and bundled Promptfoo prompts -> the shared
+  `@agentic-workflow-kit/eval-kit` package;
 - `evals/README.md` and `evals/implementation-plan.md` -> consumer docs.
 - `evals/tests/**` -> `evals/tests/`.
 - `evals/schemas/**` -> `evals/schemas/`.
@@ -147,7 +143,7 @@ Eval-specific material is split by ownership:
 
 Root-level docs may still describe the evaluation strategy, because that strategy is part of how the
 product is designed and governed. Executable eval assets, fixture data, generated outputs, schemas,
-and eval-specific implementation docs belong in the internal package.
+and eval-specific implementation docs belong in `evals/`.
 
 The root package stays small. It keeps `check` as the public deterministic repo gate and exposes
 manual eval commands that call the internal kit:
@@ -176,7 +172,7 @@ Add `promptfoo` as a development dependency of the eval package and add manual s
 
 Keep `pnpm check` deterministic-only.
 
-The Promptfoo templates live in `packages/eval-kit/promptfoo/` and suite-specific variables come
+The Promptfoo templates live in the shared `@agentic-workflow-kit/eval-kit` package and suite-specific variables come
 from `evals/adapter.mjs`:
 
 - Generation suite:
@@ -196,7 +192,7 @@ from `evals/adapter.mjs`:
   - Provider: `openai:codex-app-server`.
   - Model: `gpt-5.4`.
   - Reasoning effort: medium.
-  - Output schema: `packages/eval-kit/schemas/pairwise-result.schema.json`.
+  - Output schema: `@agentic-workflow-kit/eval-kit/schemas/pairwise-result.schema.json`.
   - Test variables: case id, source facts, expected facts and boundaries, generated candidate,
     reference anchor, candidate order, original order, randomization method, and seed.
   - Assertions: JSON output, schema-valid output, evidence present, model/provider/rubric/prompt
@@ -206,7 +202,7 @@ from `evals/adapter.mjs`:
   - Provider: `openai:codex-app-server`.
   - Model: `gpt-5.4`.
   - Reasoning effort: medium.
-  - Output schema: `packages/eval-kit/schemas/pointwise-judge-result.schema.json`.
+  - Output schema: `@agentic-workflow-kit/eval-kit/schemas/pointwise-judge-result.schema.json`.
   - Test variables: case id, source facts, expected facts and boundaries, generated candidate, and
     deterministic grades if available.
   - Assertions: JSON output, schema-valid output, evidence on every covered/partial/contradicted
@@ -252,7 +248,8 @@ The final report command combines all run bundles and states:
 - Promptfoo, not shell wrappers, owns provider execution, assertions, and JSON/HTML result exports.
 - Candidate generation produces a non-empty Markdown design for `case-tiny-laundry-pickup-v1`.
 - The existing deterministic grader runs against the generated candidate.
-- The judge output validates against `packages/eval-kit/schemas/pairwise-result.schema.json`.
+- The judge output validates against
+  `@agentic-workflow-kit/eval-kit/schemas/pairwise-result.schema.json`.
 - The final report is written under ignored `evals/results/**`.
 - Generated result files are not committed unless a human reviewer explicitly promotes a redacted
   summary.

evals/README.md

@@ -2,9 +2,9 @@
 
 This directory contains the consumer-facing eval suite for the `technical-design` skills pack.
 
-It uses [`@agentic-workflow-kit/eval-kit`](../packages/eval-kit) for all generic mechanics
-(schema validation, path resolution, Promptfoo helpers, result manifests). This directory owns
-only the `technical-design` domain: DDD fixtures, expected facts and boundaries, judge rubrics,
+It uses `@agentic-workflow-kit/eval-kit` from the shared GitHub tag for all generic mechanics
+(schema validation, path resolution, Promptfoo helpers, result manifests). This directory owns only
+the `technical-design` domain: DDD fixtures, expected facts and boundaries, judge rubrics,
 lessons-ledger checks, and deterministic grading policy.
 
 The layered design-quality strategy lives in

evals/adapter.mjs

@@ -711,11 +711,11 @@ export const validateFixtures = async ({ config, manifests }) => {
   ];
 
   const expectedFactsShape = compileLocalSchema(
-    "packages/eval-kit/schemas/expected-facts.schema.json",
+    "evals/schemas/expected-facts.schema.json",
     "expected facts",
   );
   const expectedBoundariesShape = compileLocalSchema(
-    "packages/eval-kit/schemas/expected-boundaries.schema.json",
+    "evals/schemas/expected-boundaries.schema.json",
     "expected boundaries",
   );
 
@@ -910,6 +910,8 @@ export const gradeTechnicalDesignCandidate = ({
   };
 };
 
+export const gradeCandidate = gradeTechnicalDesignCandidate;
+
 export const criticalBlockerCount = (findings) =>
   countPolicyBlockers(findings, deterministicVerdictPolicy);
 

evals/eval-kit.config.json

@@ -1,5 +1,5 @@
 {
-  "$schema": "../packages/eval-kit/schemas/eval-kit.config.schema.json",
+  "$schema": "../node_modules/@agentic-workflow-kit/eval-kit/schemas/eval-kit.config.schema.json",
   "schema_version": "eval-kit.config.v1",
   "suite_id": "technical-design",
   "suite_root": ".",

evals/implementation-plan.md

@@ -275,11 +275,11 @@ Recommended OSS tools:
 
 Implemented:
 
-- Added the suite judge rubric at `evals/rubric.md`; bundled generic judge prompts live under
-  `packages/eval-kit/promptfoo/judges/`.
+- Added the suite judge rubric at `evals/rubric.md`; bundled generic judge prompts live in the
+  shared `@agentic-workflow-kit/eval-kit` package.
 - Added a pairwise comparison prompt that requires randomized candidate order to be recorded.
-- Added bundled schemas for structured judge output and pairwise results under
-  `packages/eval-kit/schemas/`.
+- Added bundled schemas for structured judge output and pairwise results in the shared
+  `@agentic-workflow-kit/eval-kit` package.
 - Added a manual Promptfoo template outside the required PR gate.
 
 Remaining:

packages/eval-kit/schemas/expected-boundaries.schema.json → evals/schemas/expected-boundaries.schema.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://json-schema.org/draft/2020-12/schema",
-  "$id": "https://agentic-workflow-kit.local/eval-kit/expected-boundaries.schema.json",
+  "$id": "https://agentic-workflow-kit.local/technical-design/evals/expected-boundaries.schema.json",
   "title": "ExpectedBoundaries",
   "type": "object",
   "required": ["case_id", "contexts"],

packages/eval-kit/schemas/expected-facts.schema.json → evals/schemas/expected-facts.schema.json

@@ -1,6 +1,6 @@
 {
   "$schema": "https://json-schema.org/draft/2020-12/schema",
-  "$id": "https://agentic-workflow-kit.local/eval-kit/expected-facts.schema.json",
+  "$id": "https://agentic-workflow-kit.local/technical-design/evals/expected-facts.schema.json",
   "title": "ExpectedFacts",
   "type": "object",
   "required": ["case_id", "facts"],

evals/tests/validate-eval-fixtures.test.mjs

@@ -55,13 +55,7 @@ const createFixtureRepo = () => {
 
   fs.symlinkSync(
     path.join(repoRoot, "node_modules"),
-    path.join(tempDir, "evals/node_modules"),
-  );
-
-  fs.mkdirSync(path.join(tempDir, "packages"), { recursive: true });
-  fs.symlinkSync(
-    path.resolve(repoRoot, "packages/eval-kit"),
-    path.join(tempDir, "packages/eval-kit"),
+    path.join(tempDir, "node_modules"),
   );
 
   return tempDir;
@@ -70,14 +64,23 @@ const createFixtureRepo = () => {
 const runValidator = (cwd) => {
   const evalKitBin = path.resolve(
     repoRoot,
-    "packages/eval-kit/bin/eval-kit.mjs",
+    "node_modules/@agentic-workflow-kit/eval-kit/bin/eval-kit.mjs",
   );
   try {
-    execFileSync(process.execPath, [evalKitBin, "validate-fixtures"], {
-      cwd: path.join(cwd, "evals"),
-      encoding: "utf8",
-      stdio: "pipe",
-    });
+    execFileSync(
+      process.execPath,
+      [
+        evalKitBin,
+        "validate-fixtures",
+        "--config",
+        path.join(cwd, "evals/eval-kit.config.json"),
+      ],
+      {
+        cwd: path.join(cwd, "evals"),
+        encoding: "utf8",
+        stdio: "pipe",
+      },
+    );
     return { ok: true, stdout: "", stderr: "" };
   } catch (error) {
     return {

package.json

@@ -33,7 +33,7 @@
     "check:links": "node scripts/check_non_eval.mjs links",
     "check:non-eval": "node scripts/check_non_eval.mjs all",
     "smoke:enforce-generator": "node scripts/smoke_generate_depcruise.mjs",
-    "check": "pnpm lint && pnpm check:non-eval && pnpm smoke:enforce-generator && pnpm --filter @agentic-workflow-kit/eval-kit test && pnpm check:deterministic",
+    "check": "pnpm lint && pnpm check:non-eval && pnpm smoke:enforce-generator && pnpm check:deterministic",
     "check:deterministic": "pnpm check:eval-static && pnpm eval:unit && pnpm eval:enforce",
     "check:eval-static": "node scripts/check_eval_static.mjs",
     "eval:unit": "vitest run evals/tests",
@@ -45,7 +45,7 @@
     "eval:manual-report": "eval-kit report --config evals/eval-kit.config.json"
   },
   "devDependencies": {
-    "@agentic-workflow-kit/eval-kit": "workspace:*",
+    "@agentic-workflow-kit/eval-kit": "github:agentic-workflow-kit/eval-kit#v0.1.0",
     "ajv": "^8.17.1",
     "dependency-cruiser": "^18.0.0",
     "prettier": "^3.6.2",