workflow evals with one scenario

mise.toml

@@ -46,15 +46,11 @@ sources = ["test/**", "skills/**"]
 # ── Eval tasks ────────────────────────────────────────────────────────
 
 [tasks.eval]
-description = "Run code-fix evals for all configured models (local, no upload)"
+description = "Run workflow evals"
 run = "npm --prefix packages/evals run eval"
-sources = ["packages/evals/src/**", "skills/**/references/**"]
+sources = ["packages/evals/src/**", "packages/evals/evals/**"]
-
-[tasks."eval:model"]
-description = "Run code-fix eval for a single model (local, no upload)"
-run = "EVAL_MODEL={{arg(name='model')}} npm --prefix packages/evals run eval"
 
 [tasks."eval:upload"]
-description = "Run code-fix evals for all models and upload to Braintrust"
+description = "Run workflow evals and upload to Braintrust"
 run = "npm --prefix packages/evals run eval:upload"
-sources = ["packages/evals/src/**", "skills/**/references/**"]
+sources = ["packages/evals/src/**", "packages/evals/evals/**"]

packages/evals/.env.example

@@ -1,5 +1,4 @@
-BRAINTRUST_API_KEY=
-BRAINTRUST_PROJECT_ID=
 ANTHROPIC_API_KEY=
-# Provider API keys for eval models are configured in the Braintrust dashboard
+# Required for Braintrust upload (BRAINTRUST_UPLOAD=true)
-# under Settings → AI providers (not needed locally).
+# BRAINTRUST_API_KEY=
+# BRAINTRUST_PROJECT_ID=

packages/evals/AGENTS.md

@@ -1,171 +1,108 @@
 # Evals — Agent Guide
 
-This package evaluates whether LLMs correctly apply Supabase best practices
+This package evaluates whether AI agents correctly implement Supabase tasks
-using skill documentation as context. It uses
+when using skill documentation. Modeled after
-[Braintrust](https://www.braintrust.dev/) for eval orchestration and the
+[Vercel's next-evals-oss](https://github.com/vercel-labs/next-evals-oss): each
-[Vercel AI SDK](https://sdk.vercel.ai/) for LLM calls.
+eval is a self-contained project with a task prompt, the agent works on it, and
+hidden tests check the result. Binary pass/fail.
 
 ## Architecture
 
-Two-step **LLM-as-judge** pattern powered by Braintrust's `Eval()`:
-
-1. The **eval model** receives a prompt with skill context and produces a code
-   fix. All eval model calls go through the **Braintrust AI proxy** — a single
-   OpenAI-compatible endpoint that routes to any provider (Anthropic, OpenAI,
-   Google, etc.).
-2. Five independent **judge scorers** (`claude-opus-4-6` via direct Anthropic
-   API) evaluate the fix via structured output (Zod schemas via AI SDK's
-   `Output.object()`).
-
-The eval runs once per model in the model matrix, creating a separate Braintrust
-experiment per model for side-by-side comparison.
-
-Key files:
-
 ```
-src/
+1. Create temp dir with project skeleton (PROMPT.md, supabase/ dir)
-  code-fix.eval.ts        # Braintrust Eval() entry point (loops over models)
+2. Symlink supabase skill into workspace (or skip for baseline)
-  dataset.ts              # Maps extracted test cases to EvalCase format
+3. Run: claude -p "prompt" --cwd /tmp/eval-xxx
-  scorer.ts               # Five AI SDK-based scorers (quality, safety, minimality)
+4. Agent reads skill, creates migrations/code in the workspace
-  models.ts               # Braintrust proxy + direct Anthropic provider
+5. Copy hidden EVAL.ts into workspace, run vitest
-  models.config.ts        # Model matrix (add/remove models here)
+6. Capture pass/fail
-  dataset/
-    types.ts              # CodeFixTestCase interface
-    extract.ts            # Auto-extracts test cases from skill references
-  prompts/
-    code-fix.ts           # System + user prompts for the eval model
 ```
 
-## How It Works
+The agent is **Claude Code** invoked via `claude -p` (print mode). It operates
+on a real filesystem in a temp directory and can read/write files freely.
 
-**Test cases are auto-extracted** from `skills/*/references/*.md`. The extractor
+## Eval Structure
-(`dataset/extract.ts`) finds consecutive `**Incorrect:**` / `**Correct:**` code
-block pairs under `##` sections. Each pair becomes one test case.
 
-Five independent scorers evaluate each fix (0–1 scale):
+Each eval lives in `evals/{scenario-name}/`:
 
-- **Correctness** — does the fix address the core issue?
+```
-- **Completeness** — does the fix include all necessary changes?
+evals/auth-rls-new-project/
-- **Best Practice** — does the fix follow Supabase conventions?
+  PROMPT.md          # Task description (visible to agent)
-- **Regression Safety** — does the fix avoid introducing new problems (broken
+  EVAL.ts            # Vitest assertions (hidden from agent during run)
-  functionality, removed security measures, new vulnerabilities)?
+  package.json       # Minimal project manifest
-- **Minimality** — is the fix tightly scoped to the identified issue without
+  supabase/
-  unnecessary rewrites or over-engineering?
+    config.toml      # Pre-initialized supabase config
-
+    migrations/      # Empty — agent creates files here
-Each model in the matrix generates a separate Braintrust experiment. The
-dashboard supports side-by-side comparison of experiments.
-
-## Adding Test Cases
-
-No code changes needed. Add paired Incorrect/Correct blocks to any skill
-reference file. The extractor picks them up automatically.
-
-Required format in a reference `.md` file:
-
-```markdown
-## Section Title
-
-Explanation of the issue.
-
-**Incorrect:**
-
-\```sql
--- bad code
-\```
-
-**Correct:**
-
-\```sql
--- good code
-\```
 ```
 
-Rules:
+**EVAL.ts** is never copied to the workspace until after the agent finishes.
-
+This prevents the agent from "teaching to the test."
-- Pairs must be consecutive — an Incorrect block immediately followed by a
-  Correct block
-- Labels are matched case-insensitively. Bad labels: `Incorrect`, `Wrong`, `Bad`.
-  Good labels: `Correct`, `Good`, `Usage`, `Implementation`, `Example`,
-  `Recommended`
-- The optional parenthetical in the label becomes the `description` field:
-  `**Incorrect (missing RLS):**`
-- Files prefixed with `_` (like `_sections.md`, `_template.md`) are skipped
-- Each pair gets an ID like `supabase/db-rls-mandatory#0` (skill/filename#index)
-
-## Adding/Removing Models
-
-Edit the `EVAL_MODELS` array in `src/models.config.ts`:
-
-```typescript
-export const EVAL_MODELS: EvalModelConfig[] = [
-  { id: "claude-sonnet-4-5-20250929", label: "Claude Sonnet 4.5", provider: "anthropic", ci: true },
-  { id: "gpt-5.3", label: "GPT 5.3", provider: "openai", ci: true },
-  // Add new models here
-];
-```
-
-Provider API keys must be configured in the Braintrust dashboard under
-Settings → AI providers.
 
 ## Running Evals
 
 ```bash
-# Run all models locally (no Braintrust upload)
+# Run all scenarios with Claude Sonnet 4.5 (default)
 mise run eval
 
-# Run a single model
+# Run a specific scenario
-mise run eval:model model=claude-sonnet-4-5-20250929
+EVAL_SCENARIO=auth-rls-new-project mise run eval
 
-# Run and upload to Braintrust dashboard
+# Override model
-mise run eval:upload
+EVAL_MODEL=claude-opus-4-6 mise run eval
+
+# Run with baseline comparison (with-skill vs without-skill)
+EVAL_BASELINE=true mise run eval
 ```
 
 Or directly:
 
 ```bash
 cd packages/evals
+npx tsx src/runner.ts
 
-# Local run (all models)
+# Single scenario with baseline
-npx braintrust eval --no-send-logs src/code-fix.eval.ts
+EVAL_SCENARIO=auth-rls-new-project EVAL_BASELINE=true npx tsx src/runner.ts
-
-# Single model
-EVAL_MODEL=claude-sonnet-4-5-20250929 npx braintrust eval --no-send-logs src/code-fix.eval.ts
-
-# Filter to one test case (across all models)
-npx braintrust eval --no-send-logs src/code-fix.eval.ts --filter 'input.testCase.id=db-migrations-idempotent'
 ```
 
+## Baseline Comparison
+
+Set `EVAL_BASELINE=true` to run each scenario twice:
+
+- **With skill**: The supabase skill is symlinked into the workspace. Claude
+  Code discovers it and uses reference files for guidance.
+- **Baseline**: No skill available. The agent relies on innate knowledge.
+
+Compare pass rates to measure how much the skill improves agent output.
+
+## Adding Scenarios
+
+1. Create `evals/{scenario-name}/` with `PROMPT.md`, `EVAL.ts`, `package.json`
+2. Add any starter files the agent should see (e.g., `supabase/config.toml`)
+3. Write vitest assertions in `EVAL.ts` that check the agent's output files
+4. Document the scenario in `scenarios/SCENARIOS.md`
+
 ## Environment
 
-API keys are loaded by mise from `packages/evals/.env` (configured in root
-`mise.toml`). Copy `.env.example` to `.env` and fill in the keys.
-
 ```
-BRAINTRUST_API_KEY=...         # Required: proxy routing + dashboard upload
+ANTHROPIC_API_KEY=sk-ant-...    # Required: Claude Code authentication
-BRAINTRUST_PROJECT_ID=...      # Required: Braintrust project identifier
+EVAL_MODEL=...                  # Optional: override model (default: claude-sonnet-4-5-20250929)
-ANTHROPIC_API_KEY=sk-ant-...   # Required: judge model (Claude Opus 4.6)
+EVAL_SCENARIO=...               # Optional: run single scenario
+EVAL_BASELINE=true              # Optional: run baseline comparison
+BRAINTRUST_UPLOAD=true          # Optional: upload results to Braintrust
 ```
 
-Optional overrides:
+## Key Files
 
 ```
-EVAL_MODEL=claude-sonnet-4-5-20250929  # Run only this model (skips matrix)
+src/
-EVAL_JUDGE_MODEL=claude-opus-4-6       # Judge model for scorers
+  runner.ts              # Main orchestrator
+  types.ts               # Core interfaces
+  runner/
+    scaffold.ts          # Creates temp workspace from eval template
+    agent.ts             # Invokes claude -p as subprocess
+    test.ts              # Runs vitest EVAL.ts against workspace
+    results.ts           # Collects results and prints summary
+evals/
+  auth-rls-new-project/  # Scenario 1
+scenarios/
+  SCENARIOS.md           # Scenario descriptions
 ```
-
-## Modifying Prompts
-
-- `src/prompts/code-fix.ts` — what the eval model sees
-- `src/scorer.ts` — judge prompts for each scorer dimension
-
-Temperature settings:
-
-- Eval model: `0.2` (in `code-fix.eval.ts`)
-- Judge model: `0.1` (in `scorer.ts`)
-
-## Modifying Scoring
-
-Each scorer in `src/scorer.ts` is independent. To add a new dimension:
-
-1. Create a new `EvalScorer` function in `scorer.ts`
-2. Add it to the `scores` array in `code-fix.eval.ts`

packages/evals/README.md

@@ -1,46 +1,51 @@
 # Evals
 
-LLM evaluation system for Supabase agent skills, powered by [Braintrust](https://www.braintrust.dev/). Tests whether models can correctly apply Supabase best practices using skill documentation as context.
+Agent evaluation system for Supabase skills. Tests whether AI agents (starting
+with Claude Code) correctly implement Supabase tasks when given access to skill
+documentation.
 
 ## How It Works
 
-Each eval follows a two-step **LLM-as-judge** pattern orchestrated by Braintrust's `Eval()`:
+Each eval is a self-contained project directory with a task prompt. The agent
+works on it autonomously, then hidden vitest assertions check the result.
+Binary pass/fail.
 
-1. **Generate** — The eval model (e.g. Sonnet 4.5) receives a prompt with skill context and produces a code fix.
+```
-2. **Judge** — Three independent scorers using a stronger model (Opus 4.6 by default) evaluate the fix via the Vercel AI SDK with structured output.
+1. Create temp workspace from eval template
-
+2. Agent (claude -p) reads prompt and creates files
-Test cases are extracted automatically from skill reference files (`skills/*/references/*.md`). Each file contains paired **Incorrect** / **Correct** code blocks — the model receives the bad code and must produce the fix.
+3. Hidden EVAL.ts runs vitest assertions against the output
-
+4. Pass/fail
-**Scoring dimensions (each 0–1):**
+```
-
-| Scorer | Description |
-|--------|-------------|
-| Correctness | Does the fix address the core issue? |
-| Completeness | Does it include all necessary changes? |
-| Best Practice | Does it follow Supabase best practices? |
 
 ## Usage
 
 ```bash
-# Run locally (no Braintrust upload)
+# Run all scenarios
 mise run eval
 
-# Run and upload to Braintrust dashboard
+# Run a specific scenario
-mise run eval:upload
+EVAL_SCENARIO=auth-rls-new-project mise run eval
+
+# Run with baseline comparison (with-skill vs without-skill)
+EVAL_BASELINE=true mise run eval
+
+# Override model
+EVAL_MODEL=claude-opus-4-6 mise run eval
 ```
 
 ### Environment Variables
 
-API keys are loaded via mise from `packages/evals/.env` (see root `mise.toml`).
-
 ```
-ANTHROPIC_API_KEY         Required: eval model + judge model
+ANTHROPIC_API_KEY         Required: Claude Code authentication
-BRAINTRUST_API_KEY        Required for Braintrust dashboard upload
+EVAL_MODEL                Override model (default: claude-sonnet-4-5-20250929)
-BRAINTRUST_PROJECT_ID     Required for Braintrust dashboard upload
+EVAL_SCENARIO             Run single scenario by name
-EVAL_MODEL                Override default eval model (claude-sonnet-4-5-20250929)
+EVAL_BASELINE=true        Run baseline comparison (no skill)
-EVAL_JUDGE_MODEL          Override default judge model (claude-opus-4-6)
 ```
 
-## Adding Test Cases
+## Adding Scenarios
 
-Add paired Incorrect/Correct code blocks to any skill reference file. The extractor picks them up automatically on the next run.
+1. Create `evals/{name}/` with `PROMPT.md`, `EVAL.ts`, and starter files
+2. Write vitest assertions in `EVAL.ts`
+3. Document in `scenarios/SCENARIOS.md`
+
+See [AGENTS.md](AGENTS.md) for full details.

packages/evals/evals/auth-rls-new-project/EVAL.ts

@@ -0,0 +1,107 @@
+import { existsSync, readdirSync, readFileSync } from "node:fs";
+import { join } from "node:path";
+import { expect, test } from "vitest";
+
+const supabaseDir = join(process.cwd(), "supabase");
+const migrationsDir = join(supabaseDir, "migrations");
+
+/** Find the first .sql migration file (agent may name it differently). */
+function findMigrationFile(): string | null {
+	if (!existsSync(migrationsDir)) return null;
+	const files = readdirSync(migrationsDir).filter((f) => f.endsWith(".sql"));
+	return files.length > 0 ? join(migrationsDir, files[0]) : null;
+}
+
+function getMigrationSQL(): string {
+	const file = findMigrationFile();
+	if (!file) throw new Error("No migration file found in supabase/migrations/");
+	return readFileSync(file, "utf-8");
+}
+
+test("supabase project initialized (config.toml exists)", () => {
+	expect(existsSync(join(supabaseDir, "config.toml"))).toBe(true);
+});
+
+test("migration file exists in supabase/migrations/", () => {
+	expect(findMigrationFile()).not.toBeNull();
+});
+
+test("creates tasks table", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	expect(sql).toMatch(/create\s+table/);
+	expect(sql).toMatch(/tasks/);
+});
+
+test("enables RLS on tasks table", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	expect(sql).toMatch(/alter\s+table.*tasks.*enable\s+row\s+level\s+security/);
+});
+
+test("has foreign key to auth.users", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	expect(sql).toMatch(/references\s+auth\.users/);
+});
+
+test("uses ON DELETE CASCADE for auth FK", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	expect(sql).toMatch(/on\s+delete\s+cascade/);
+});
+
+test("uses (select auth.uid()) not bare auth.uid() in policies", () => {
+	const sql = getMigrationSQL();
+	const policyBlocks = sql.match(/create\s+policy[\s\S]*?;/gi) ?? [];
+	for (const policy of policyBlocks) {
+		if (policy.includes("auth.uid()")) {
+			// The subselect form: (select auth.uid())
+			expect(policy).toMatch(/\(\s*select\s+auth\.uid\(\)\s*\)/i);
+		}
+	}
+});
+
+test("policies use TO authenticated", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	const policyBlocks = sql.match(/create\s+policy[\s\S]*?;/gi) ?? [];
+	expect(policyBlocks.length).toBeGreaterThan(0);
+	for (const policy of policyBlocks) {
+		expect(policy).toMatch(/to\s+authenticated/);
+	}
+});
+
+test("uses timestamptz not plain timestamp for time columns", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	// Match "timestamp" that is NOT followed by "tz" or "with time zone"
+	const hasPlainTimestamp = /\btimestamp\b(?!\s*tz)(?!\s+with\s+time\s+zone)/;
+	// Only fail if the migration defines time columns with plain timestamp
+	if (
+		sql.includes("created_at") ||
+		sql.includes("updated_at") ||
+		sql.includes("due_date")
+	) {
+		expect(sql).not.toMatch(hasPlainTimestamp);
+	}
+});
+
+test("creates index on user_id column", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	expect(sql).toMatch(/create\s+index/);
+	expect(sql).toMatch(/user_id/);
+});
+
+test("migration is idempotent (uses IF NOT EXISTS)", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	expect(sql).toMatch(/if\s+not\s+exists/);
+});
+
+test("overall quality: demonstrates Supabase best practices", () => {
+	const sql = getMigrationSQL().toLowerCase();
+	// A high-quality migration should contain most of these patterns
+	const signals = [
+		/enable\s+row\s+level\s+security/,
+		/\(select\s+auth\.uid\(\)\)/,
+		/to\s+authenticated/,
+		/on\s+delete\s+cascade/,
+		/create\s+index/,
+	];
+	const matches = signals.filter((r) => r.test(sql));
+	expect(matches.length).toBeGreaterThanOrEqual(4);
+});

packages/evals/evals/auth-rls-new-project/PROMPT.md

@@ -0,0 +1,16 @@
+I'm starting a new Supabase project from scratch for a task management app. Users should sign up with email/password, and each user should only see their own tasks.
+
+Set up the project:
+
+1. Initialize the Supabase project with the CLI (`npx supabase init`)
+2. Start the local Supabase stack (`npx supabase start`)
+3. Create a SQL migration for a tasks table with columns: title (text), description (text), status (text), and due_date
+
+The migration must:
+
+- Create the tasks table with proper column types
+- Link tasks to authenticated users
+- Enable Row Level Security
+- Create policies so users can only CRUD their own tasks
+- Add appropriate indexes
+- Be idempotent (safe to run multiple times)

packages/evals/evals/auth-rls-new-project/package.json

@@ -0,0 +1,5 @@
+{
+	"name": "auth-rls-new-project",
+	"private": true,
+	"type": "module"
+}

packages/evals/package.json

@@ -4,21 +4,18 @@
 	"type": "module",
 	"author": "Supabase",
 	"license": "MIT",
-	"description": "LLM evaluation system for Supabase agent skills",
+	"description": "Agent evaluation system for Supabase skills",
 	"scripts": {
-		"eval": "braintrust eval --no-send-logs src/code-fix.eval.ts",
+		"eval": "tsx src/runner.ts",
-		"eval:upload": "braintrust eval src/code-fix.eval.ts"
+		"eval:upload": "BRAINTRUST_UPLOAD=true tsx src/runner.ts"
 	},
 	"dependencies": {
-		"@ai-sdk/anthropic": "^3.0.44",
+		"braintrust": "^1.0.2"
-		"@ai-sdk/openai": "^3.0.29",
-		"ai": "^6.0.86",
-		"braintrust": "^1.0.2",
-		"zod": "^3.23.0"
 	},
 	"devDependencies": {
 		"@types/node": "^20.10.0",
 		"tsx": "^4.7.0",
-		"typescript": "^5.3.0"
+		"typescript": "^5.3.0",
+		"vitest": "^3.1.0"
 	}
 }

packages/evals/scenarios/SCENARIOS.md

@@ -0,0 +1,51 @@
+# Supabase Skills Eval Scenarios
+
+## Scenario 1: auth-rls-new-project
+
+**Description:** Set up a new Supabase project from scratch and add
+authentication with RLS. The agent must initialize the project with the CLI,
+start the local Supabase stack, then create a tasks table with proper security
+(RLS policies, auth FK, indexes) in a single idempotent migration.
+
+**Setup:** The workspace starts empty (no `supabase/` directory). The agent is
+expected to run `npx supabase init` and `npx supabase start` before creating
+the migration.
+
+**Expected skill files read:**
+
+- `SKILL.md` (skill body with reference file index)
+- `references/dev-getting-started.md`
+- `references/db-rls-mandatory.md`
+- `references/db-rls-policy-types.md`
+- `references/db-rls-common-mistakes.md`
+- `references/db-schema-auth-fk.md`
+- `references/db-schema-timestamps.md`
+- `references/db-migrations-idempotent.md`
+
+**Expected result:**
+
+The agent initializes a Supabase project and creates a migration file that:
+
+- Creates tasks table with `timestamptz` columns
+- Has `user_id` FK to `auth.users(id)` with `ON DELETE CASCADE`
+- Enables RLS (`ALTER TABLE tasks ENABLE ROW LEVEL SECURITY`)
+- Creates per-operation policies using `(select auth.uid())` with `TO authenticated`
+- Creates index on `user_id`
+- Uses `IF NOT EXISTS` for idempotency
+
+**Scorer:** Binary pass/fail (12 vitest assertions)
+
+| Test | What it checks |
+| --- | --- |
+| supabase project initialized | `supabase/config.toml` exists after agent runs |
+| migration file exists | Agent created a `.sql` file in `supabase/migrations/` |
+| creates tasks table | SQL contains `CREATE TABLE ... tasks` |
+| enables RLS | `ALTER TABLE tasks ENABLE ROW LEVEL SECURITY` |
+| FK to auth.users | `REFERENCES auth.users` |
+| ON DELETE CASCADE | Cascade delete on auth FK |
+| (select auth.uid()) | Subselect form in policies (performance) |
+| TO authenticated | Policies scoped to authenticated role |
+| timestamptz | No plain `timestamp` for time columns |
+| index on user_id | `CREATE INDEX` on the FK column |
+| IF NOT EXISTS | Idempotent migration |
+| overall quality | At least 4/5 best-practice signals present |

packages/evals/src/code-fix.eval.ts

@@ -1,82 +0,0 @@
-import assert from "node:assert";
-import { generateText } from "ai";
-import { Eval } from "braintrust";
-import { dataset } from "./dataset.js";
-import type { EvalModelConfig } from "./models.config.js";
-import { EVAL_MODELS } from "./models.config.js";
-import { getProxyModel } from "./models.js";
-import {
-	buildCodeFixPrompt,
-	buildCodeFixSystemPrompt,
-} from "./prompts/code-fix.js";
-import {
-	bestPracticeScorer,
-	completenessScorer,
-	correctnessScorer,
-	minimalityScorer,
-	regressionSafetyScorer,
-} from "./scorer.js";
-
-assert(process.env.BRAINTRUST_API_KEY, "BRAINTRUST_API_KEY is not set");
-assert(process.env.ANTHROPIC_API_KEY, "ANTHROPIC_API_KEY is not set");
-
-/**
- * Resolve which models to run based on environment:
- * - EVAL_MODEL set → single model only (local dev / cost control)
- * - CI without EVAL_ALL_MODELS → ci:true models only
- * - Otherwise → all models
- */
-function getModelsToRun(): EvalModelConfig[] {
-	const singleModel = process.env.EVAL_MODEL;
-	if (singleModel) {
-		const found = EVAL_MODELS.find((m) => m.id === singleModel);
-		return [
-			found ?? {
-				id: singleModel,
-				label: singleModel,
-				provider: "unknown",
-				ci: false,
-			},
-		];
-	}
-
-	if (process.env.CI && !process.env.EVAL_ALL_MODELS) {
-		return EVAL_MODELS.filter((m) => m.ci);
-	}
-
-	return EVAL_MODELS;
-}
-
-const models = getModelsToRun();
-
-for (const modelConfig of models) {
-	Eval("CodeFix", {
-		experimentName: modelConfig.id,
-		projectId: process.env.BRAINTRUST_PROJECT_ID,
-		trialCount: process.env.CI ? 3 : 1,
-		metadata: {
-			model: modelConfig.id,
-			label: modelConfig.label,
-			provider: modelConfig.provider,
-		},
-		data: () => dataset(),
-		task: async (input) => {
-			const model = getProxyModel(modelConfig.id);
-			const response = await generateText({
-				model,
-				system: buildCodeFixSystemPrompt(),
-				prompt: buildCodeFixPrompt(input.testCase),
-				temperature: 0.2,
-				maxRetries: 2,
-			});
-			return { llmOutput: response.text };
-		},
-		scores: [
-			correctnessScorer,
-			completenessScorer,
-			bestPracticeScorer,
-			regressionSafetyScorer,
-			minimalityScorer,
-		],
-	});
-}

packages/evals/src/dataset.ts

@@ -1,51 +0,0 @@
-import type { EvalCase } from "braintrust";
-import { extractCodeFixDataset } from "./dataset/extract.js";
-import type { CodeFixTestCase } from "./dataset/types.js";
-
-export type Input = { testCase: CodeFixTestCase };
-
-export type Expected = {
-	correctCode: string;
-	correctLanguage?: string;
-};
-
-export type Metadata = {
-	name: string;
-	skillName: string;
-	section: string;
-	referenceFile: string;
-	tags: string[];
-};
-
-export type Output = { llmOutput: string };
-
-/**
- * Extract the feature category from a reference filename.
- * e.g. "db-migrations-idempotent.md" → "db"
- *      "auth-core-sessions.md" → "auth"
- */
-function featureCategory(filename: string): string {
-	return filename.replace(/\.md$/, "").split("-")[0];
-}
-
-export function dataset(): EvalCase<Input, Expected, Metadata>[] {
-	return extractCodeFixDataset().map((tc) => ({
-		id: tc.id,
-		input: { testCase: tc },
-		tags: [
-			featureCategory(tc.referenceFilename),
-			tc.referenceFilename.replace(/\.md$/, ""),
-		],
-		expected: {
-			correctCode: tc.goodExample.code,
-			correctLanguage: tc.goodExample.language,
-		},
-		metadata: {
-			name: tc.title,
-			skillName: tc.skillName,
-			section: tc.section,
-			referenceFile: tc.referenceFilename,
-			tags: tc.tags,
-		},
-	}));
-}

packages/evals/src/dataset/extract.ts

@@ -1,277 +0,0 @@
-import { existsSync, readdirSync, readFileSync } from "node:fs";
-import { basename, join, resolve } from "node:path";
-import type { CodeFixTestCase } from "./types.js";
-
-function findSkillsRoot(): string {
-	let dir = process.cwd();
-	for (let i = 0; i < 10; i++) {
-		const candidate = join(dir, "skills");
-		if (existsSync(candidate)) return candidate;
-		const parent = resolve(dir, "..");
-		if (parent === dir) break;
-		dir = parent;
-	}
-	throw new Error(
-		"Could not find skills/ directory. Run from the repository root or a subdirectory.",
-	);
-}
-
-const SKILLS_ROOT = findSkillsRoot();
-
-// --- Duplicated from skills-build/src/parser.ts for isolation ---
-
-interface CodeExample {
-	label: string;
-	description?: string;
-	code: string;
-	language?: string;
-}
-
-function parseFrontmatter(content: string): {
-	frontmatter: Record<string, string>;
-	body: string;
-} {
-	const frontmatter: Record<string, string> = {};
-
-	if (!content.startsWith("---")) {
-		return { frontmatter, body: content };
-	}
-
-	const endIndex = content.indexOf("---", 3);
-	if (endIndex === -1) {
-		return { frontmatter, body: content };
-	}
-
-	const frontmatterContent = content.slice(3, endIndex).trim();
-	const body = content.slice(endIndex + 3).trim();
-
-	for (const line of frontmatterContent.split("\n")) {
-		const colonIndex = line.indexOf(":");
-		if (colonIndex === -1) continue;
-
-		const key = line.slice(0, colonIndex).trim();
-		let value = line.slice(colonIndex + 1).trim();
-
-		if (
-			(value.startsWith('"') && value.endsWith('"')) ||
-			(value.startsWith("'") && value.endsWith("'"))
-		) {
-			value = value.slice(1, -1);
-		}
-
-		frontmatter[key] = value;
-	}
-
-	return { frontmatter, body };
-}
-
-function extractTitle(body: string): string | null {
-	const match = body.match(/^##\s+(.+)$/m);
-	return match ? match[1].trim() : null;
-}
-
-interface Section {
-	title: string;
-	explanation: string;
-	examples: CodeExample[];
-}
-
-function extractSections(body: string): Section[] {
-	const sections: Section[] = [];
-	const lines = body.split("\n");
-
-	let currentTitle = "";
-	let explanationLines: string[] = [];
-	let currentExamples: CodeExample[] = [];
-	let currentLabel = "";
-	let currentDescription = "";
-	let inCodeBlock = false;
-	let codeBlockLang = "";
-	let codeBlockContent: string[] = [];
-	let collectingExplanation = false;
-
-	function flushExample() {
-		if (currentLabel && codeBlockContent.length > 0) {
-			currentExamples.push({
-				label: currentLabel,
-				description: currentDescription || undefined,
-				code: codeBlockContent.join("\n"),
-				language: codeBlockLang || undefined,
-			});
-		}
-		currentLabel = "";
-		currentDescription = "";
-		codeBlockContent = [];
-		codeBlockLang = "";
-	}
-
-	function flushSection() {
-		if (currentTitle && currentExamples.length > 0) {
-			sections.push({
-				title: currentTitle,
-				explanation: explanationLines.join("\n").trim(),
-				examples: currentExamples,
-			});
-		}
-		currentExamples = [];
-		explanationLines = [];
-	}
-
-	for (const line of lines) {
-		if (line.startsWith("## ") && !inCodeBlock) {
-			flushExample();
-			flushSection();
-			currentTitle = line.replace(/^##\s+/, "").trim();
-			collectingExplanation = true;
-			continue;
-		}
-
-		const labelMatch = line.match(
-			/^\*\*([^*]+?)(?:\s*\(([^)]+)\))?\s*:\*\*\s*$/,
-		);
-		if (labelMatch && !inCodeBlock) {
-			collectingExplanation = false;
-			flushExample();
-			currentLabel = labelMatch[1].trim();
-			currentDescription = labelMatch[2]?.trim() || "";
-			continue;
-		}
-
-		if (line.startsWith("```") && !inCodeBlock) {
-			collectingExplanation = false;
-			inCodeBlock = true;
-			codeBlockLang = line.slice(3).trim();
-			continue;
-		}
-
-		if (line.startsWith("```") && inCodeBlock) {
-			inCodeBlock = false;
-			continue;
-		}
-
-		if (inCodeBlock) {
-			codeBlockContent.push(line);
-		} else if (collectingExplanation) {
-			explanationLines.push(line);
-		}
-	}
-
-	flushExample();
-	flushSection();
-
-	return sections;
-}
-
-// --- Duplicated from skills-build/src/validate.ts ---
-
-function isBadExample(label: string): boolean {
-	const lower = label.toLowerCase();
-	return (
-		lower.includes("incorrect") ||
-		lower.includes("wrong") ||
-		lower.includes("bad")
-	);
-}
-
-function isGoodExample(label: string): boolean {
-	const lower = label.toLowerCase();
-	return (
-		lower.includes("correct") ||
-		lower.includes("good") ||
-		lower.includes("usage") ||
-		lower.includes("implementation") ||
-		lower.includes("example") ||
-		lower.includes("recommended")
-	);
-}
-
-// --- Extraction logic ---
-
-function pairExamples(
-	examples: CodeExample[],
-): Array<{ bad: CodeExample; good: CodeExample }> {
-	const pairs: Array<{ bad: CodeExample; good: CodeExample }> = [];
-
-	for (let i = 0; i < examples.length - 1; i++) {
-		if (
-			isBadExample(examples[i].label) &&
-			isGoodExample(examples[i + 1].label)
-		) {
-			pairs.push({ bad: examples[i], good: examples[i + 1] });
-		}
-	}
-
-	return pairs;
-}
-
-function discoverSkillNames(): string[] {
-	if (!existsSync(SKILLS_ROOT)) return [];
-
-	return readdirSync(SKILLS_ROOT, { withFileTypes: true })
-		.filter((d) => d.isDirectory())
-		.filter((d) => existsSync(join(SKILLS_ROOT, d.name, "SKILL.md")))
-		.map((d) => d.name);
-}
-
-function getMarkdownFiles(dir: string): string[] {
-	if (!existsSync(dir)) return [];
-
-	return readdirSync(dir)
-		.filter((f) => f.endsWith(".md") && !f.startsWith("_"))
-		.map((f) => join(dir, f));
-}
-
-export function extractCodeFixDataset(skillName?: string): CodeFixTestCase[] {
-	const skills = skillName ? [skillName] : discoverSkillNames();
-	const testCases: CodeFixTestCase[] = [];
-
-	for (const skill of skills) {
-		const referencesDir = join(SKILLS_ROOT, skill, "references");
-		const files = getMarkdownFiles(referencesDir);
-
-		for (const filePath of files) {
-			const content = readFileSync(filePath, "utf-8");
-			const { frontmatter, body } = parseFrontmatter(content);
-			const fileTitle =
-				frontmatter.title || extractTitle(body) || basename(filePath, ".md");
-			const tags = frontmatter.tags?.split(",").map((t) => t.trim()) || [];
-			const section = basename(filePath, ".md").split("-")[0];
-
-			const sections = extractSections(body);
-			let pairIndex = 0;
-
-			for (const sec of sections) {
-				const pairs = pairExamples(sec.examples);
-
-				for (const { bad, good } of pairs) {
-					testCases.push({
-						id: `${skill}/${basename(filePath, ".md")}#${pairIndex}`,
-						skillName: skill,
-						referenceFile: filePath,
-						referenceFilename: basename(filePath),
-						title: sec.title || fileTitle,
-						explanation: sec.explanation,
-						section,
-						tags,
-						pairIndex,
-						badExample: {
-							label: bad.label,
-							description: bad.description,
-							code: bad.code,
-							language: bad.language,
-						},
-						goodExample: {
-							label: good.label,
-							description: good.description,
-							code: good.code,
-							language: good.language,
-						},
-					});
-					pairIndex++;
-				}
-			}
-		}
-	}
-
-	return testCases;
-}

packages/evals/src/dataset/types.ts

@@ -1,24 +0,0 @@
-export interface CodeFixTestCase {
-	/** Unique ID, e.g. "supabase/db-rls-mandatory#0" */
-	id: string;
-	skillName: string;
-	referenceFile: string;
-	referenceFilename: string;
-	title: string;
-	explanation: string;
-	section: string;
-	tags: string[];
-	pairIndex: number;
-	badExample: {
-		label: string;
-		description?: string;
-		code: string;
-		language?: string;
-	};
-	goodExample: {
-		label: string;
-		description?: string;
-		code: string;
-		language?: string;
-	};
-}

packages/evals/src/models.config.ts

@@ -1,47 +0,0 @@
-export interface EvalModelConfig {
-	/** Model ID passed to the Braintrust proxy */
-	id: string;
-	/** Human-readable label for dashboards */
-	label: string;
-	/** Provider name for display/grouping */
-	provider: string;
-	/** Whether to include in CI runs by default */
-	ci: boolean;
-}
-
-/**
- * Models to evaluate. Add/remove entries to change the eval matrix.
- * Set `ci: false` to exclude expensive models from automated CI runs.
- */
-export const EVAL_MODELS: EvalModelConfig[] = [
-	{
-		id: "claude-sonnet-4-5-20250929",
-		label: "Claude Sonnet 4.5",
-		provider: "anthropic",
-		ci: true,
-	},
-	{
-		id: "gpt-5.3",
-		label: "GPT 5.3",
-		provider: "openai",
-		ci: true,
-	},
-	{
-		id: "gpt-5.2",
-		label: "GPT 5.2",
-		provider: "openai",
-		ci: true,
-	},
-	{
-		id: "gemini-3-pro",
-		label: "Gemini 3.0 Pro",
-		provider: "google",
-		ci: true,
-	},
-	{
-		id: "claude-opus-4-6",
-		label: "Claude Opus 4.6",
-		provider: "anthropic",
-		ci: false,
-	},
-];

packages/evals/src/models.ts

@@ -1,52 +0,0 @@
-import type { AnthropicProvider } from "@ai-sdk/anthropic";
-import { anthropic } from "@ai-sdk/anthropic";
-import { createOpenAI } from "@ai-sdk/openai";
-import type { LanguageModel } from "ai";
-
-/** Model ID accepted by the Anthropic provider (string literal union + string). */
-export type AnthropicModelId = Parameters<AnthropicProvider["chat"]>[0];
-
-/**
- * Braintrust AI proxy — routes to any provider (Anthropic, OpenAI, Google)
- * via a single OpenAI-compatible endpoint.
- *
- * Provider API keys are configured in the Braintrust dashboard at
- * project or org level. The x-bt-parent header scopes the request to
- * the project so project-level keys are resolved.
- */
-const braintrustProxy = createOpenAI({
-	baseURL: "https://api.braintrust.dev/v1/proxy",
-	apiKey: process.env.BRAINTRUST_API_KEY ?? "",
-	headers: process.env.BRAINTRUST_PROJECT_ID
-		? { "x-bt-parent": `project_id:${process.env.BRAINTRUST_PROJECT_ID}` }
-		: undefined,
-});
-
-/**
- * Get a model for the eval task. Claude models use the Anthropic SDK
- * directly (via ANTHROPIC_API_KEY). All other models route through the
- * Braintrust proxy (keys configured at the org level in Braintrust).
- */
-export function getProxyModel(modelId: string): LanguageModel {
-	if (modelId.startsWith("claude")) {
-		return anthropic(modelId as AnthropicModelId);
-	}
-	return braintrustProxy(modelId);
-}
-
-/**
- * Get a model using direct provider SDKs. Used for the judge model which
- * is always Claude and uses ANTHROPIC_API_KEY directly (no proxy).
- */
-export function getModel(modelId: string): LanguageModel {
-	if (modelId.startsWith("claude")) {
-		return anthropic(modelId as AnthropicModelId);
-	}
-
-	return getProxyModel(modelId);
-}
-
-export function getJudgeModel(): LanguageModel {
-	const judgeModelId = process.env.EVAL_JUDGE_MODEL || "claude-opus-4-6";
-	return getModel(judgeModelId);
-}

packages/evals/src/prompts/code-fix.ts

@@ -1,33 +0,0 @@
-import type { CodeFixTestCase } from "../dataset/types.js";
-
-export function buildCodeFixSystemPrompt(): string {
-	return `You are a senior Supabase developer and database architect. You fix code to follow Supabase best practices including:
-- Row Level Security (RLS) policies
-- Proper authentication patterns
-- Safe migration workflows
-- Correct SDK usage patterns
-- Edge Function best practices
-- Connection pooling configuration
-- Security-first defaults
-
-When fixing code, ensure the fix is complete, production-ready, and follows the latest Supabase conventions. Return only the corrected code inside a single code block.`;
-}
-
-export function buildCodeFixPrompt(testCase: CodeFixTestCase): string {
-	const langHint = testCase.badExample.language
-		? ` (${testCase.badExample.language})`
-		: "";
-
-	return `The following code has a problem related to: ${testCase.title}
-
-Context: ${testCase.explanation}
-
-Here is the problematic code${langHint}:
-
-\`\`\`${testCase.badExample.language || ""}
-${testCase.badExample.code}
-\`\`\`
-${testCase.badExample.description ? `\nIssue hint: ${testCase.badExample.description}` : ""}
-
-Fix this code to follow Supabase best practices. Return ONLY the corrected code inside a single code block. Do not include any explanation outside the code block.`;
-}

packages/evals/src/runner.ts

@@ -0,0 +1,189 @@
+import { existsSync, readdirSync, readFileSync } from "node:fs";
+import { join, resolve } from "node:path";
+import { runAgent } from "./runner/agent.js";
+import { uploadToBraintrust } from "./runner/braintrust.js";
+import { preflight } from "./runner/preflight.js";
+import { listModifiedFiles, printSummary } from "./runner/results.js";
+import { createWorkspace } from "./runner/scaffold.js";
+import { runTests } from "./runner/test.js";
+import type { EvalRunResult, EvalScenario } from "./types.js";
+
+// ---------------------------------------------------------------------------
+// Configuration from environment
+// ---------------------------------------------------------------------------
+
+const DEFAULT_MODEL = "claude-sonnet-4-5-20250929";
+const AGENT_TIMEOUT = 30 * 60 * 1000; // 30 minutes
+
+const model = process.env.EVAL_MODEL ?? DEFAULT_MODEL;
+const scenarioFilter = process.env.EVAL_SCENARIO;
+const runBaseline = process.env.EVAL_BASELINE === "true";
+
+// ---------------------------------------------------------------------------
+// Discover scenarios
+// ---------------------------------------------------------------------------
+
+function findEvalsDir(): string {
+	let dir = process.cwd();
+	for (let i = 0; i < 10; i++) {
+		const candidate = join(dir, "packages", "evals", "evals");
+		if (existsSync(candidate)) return candidate;
+		const parent = resolve(dir, "..");
+		if (parent === dir) break;
+		dir = parent;
+	}
+	throw new Error("Could not find packages/evals/evals/ directory");
+}
+
+function discoverScenarios(): EvalScenario[] {
+	const evalsDir = findEvalsDir();
+	const dirs = readdirSync(evalsDir, { withFileTypes: true }).filter(
+		(d) => d.isDirectory() && existsSync(join(evalsDir, d.name, "PROMPT.md")),
+	);
+
+	return dirs.map((d) => ({
+		id: d.name,
+		name: d.name,
+		tags: [],
+	}));
+}
+
+// ---------------------------------------------------------------------------
+// Run a single eval
+// ---------------------------------------------------------------------------
+
+async function runEval(
+	scenario: EvalScenario,
+	skillEnabled: boolean,
+): Promise<EvalRunResult> {
+	const evalsDir = findEvalsDir();
+	const evalDir = join(evalsDir, scenario.id);
+
+	console.log(
+		`\n--- ${scenario.id} (${skillEnabled ? "with-skill" : "baseline"}) ---`,
+	);
+
+	// 1. Create isolated workspace
+	const { workspacePath, cleanup } = createWorkspace({
+		evalDir,
+		skillEnabled,
+	});
+	console.log(`  Workspace: ${workspacePath}`);
+
+	try {
+		// 2. Read the prompt
+		const prompt = readFileSync(join(evalDir, "PROMPT.md"), "utf-8").trim();
+
+		// 3. Run the agent
+		console.log(`  Running agent (${model})...`);
+		const agentResult = await runAgent({
+			cwd: workspacePath,
+			prompt,
+			model,
+			timeout: AGENT_TIMEOUT,
+			skillEnabled,
+		});
+		console.log(
+			`  Agent finished in ${(agentResult.duration / 1000).toFixed(1)}s`,
+		);
+
+		// 4. Run the hidden tests
+		const evalFilePath = existsSync(join(evalDir, "EVAL.tsx"))
+			? join(evalDir, "EVAL.tsx")
+			: join(evalDir, "EVAL.ts");
+
+		console.log("  Running tests...");
+		const testResult = await runTests({
+			workspacePath,
+			evalFilePath,
+		});
+		console.log(
+			`  Tests: ${testResult.passedCount}/${testResult.totalCount} passed`,
+		);
+
+		// 5. Collect modified files
+		const filesModified = listModifiedFiles(workspacePath, evalDir);
+
+		return {
+			scenario: scenario.id,
+			agent: "claude-code",
+			model,
+			skillEnabled,
+			status: testResult.passed ? "passed" : "failed",
+			duration: agentResult.duration,
+			testOutput: testResult.output,
+			agentOutput: agentResult.output,
+			testsPassed: testResult.passedCount,
+			testsTotal: testResult.totalCount,
+			filesModified,
+		};
+	} catch (error) {
+		const err = error as Error;
+		return {
+			scenario: scenario.id,
+			agent: "claude-code",
+			model,
+			skillEnabled,
+			status: "error",
+			duration: 0,
+			testOutput: "",
+			agentOutput: "",
+			testsPassed: 0,
+			testsTotal: 0,
+			filesModified: [],
+			error: err.message,
+		};
+	} finally {
+		cleanup();
+	}
+}
+
+// ---------------------------------------------------------------------------
+// Main
+// ---------------------------------------------------------------------------
+
+async function main() {
+	preflight();
+
+	console.log("Supabase Skills Evals");
+	console.log(`Model: ${model}`);
+	console.log(`Baseline: ${runBaseline}`);
+
+	let scenarios = discoverScenarios();
+
+	if (scenarioFilter) {
+		scenarios = scenarios.filter((s) => s.id === scenarioFilter);
+		if (scenarios.length === 0) {
+			console.error(`Scenario not found: ${scenarioFilter}`);
+			process.exit(1);
+		}
+	}
+
+	console.log(`Scenarios: ${scenarios.map((s) => s.id).join(", ")}`);
+
+	const results: EvalRunResult[] = [];
+
+	for (const scenario of scenarios) {
+		// Run with skill enabled
+		const withSkill = await runEval(scenario, true);
+		results.push(withSkill);
+
+		// Optionally run baseline (no skill)
+		if (runBaseline) {
+			const baseline = await runEval(scenario, false);
+			results.push(baseline);
+		}
+	}
+
+	printSummary(results);
+
+	if (process.env.BRAINTRUST_UPLOAD === "true") {
+		console.log("\nUploading to Braintrust...");
+		await uploadToBraintrust(results);
+	}
+}
+
+main().catch((err) => {
+	console.error("Fatal error:", err);
+	process.exit(1);
+});

packages/evals/src/runner/agent.ts

@@ -0,0 +1,82 @@
+import { spawn } from "node:child_process";
+
+export interface AgentRunResult {
+	output: string;
+	duration: number;
+}
+
+/**
+ * Invoke Claude Code in print mode as a subprocess.
+ *
+ * The agent operates in the workspace directory and can read/write files.
+ * When the skill is installed (symlinked into workspace), Claude Code
+ * discovers it automatically and uses it for guidance.
+ */
+export async function runAgent(opts: {
+	cwd: string;
+	prompt: string;
+	model: string;
+	timeout: number;
+	skillEnabled: boolean;
+}): Promise<AgentRunResult> {
+	const start = Date.now();
+
+	const args = [
+		"-p", // Print mode (non-interactive)
+		"--output-format",
+		"text",
+		"--model",
+		opts.model,
+		"--no-session-persistence",
+		"--dangerously-skip-permissions",
+		"--tools",
+		"Edit,Write,Bash,Read,Glob,Grep",
+	];
+
+	// Disable skills for baseline runs so the agent relies on innate knowledge
+	if (!opts.skillEnabled) {
+		args.push("--disable-slash-commands");
+	}
+
+	const env = { ...process.env };
+	// Remove all Claude-related env vars to avoid nested-session detection
+	for (const key of Object.keys(env)) {
+		if (key === "CLAUDECODE" || key.startsWith("CLAUDE_")) {
+			delete env[key];
+		}
+	}
+
+	return new Promise<AgentRunResult>((resolve) => {
+		const child = spawn("claude", args, {
+			cwd: opts.cwd,
+			env,
+			stdio: ["pipe", "pipe", "pipe"],
+		});
+
+		// Pipe prompt via stdin and close — this is the standard way to
+		// pass multi-line prompts to `claude -p`.
+		child.stdin.write(opts.prompt);
+		child.stdin.end();
+
+		let stdout = "";
+		let stderr = "";
+		child.stdout.on("data", (d: Buffer) => {
+			stdout += d.toString();
+		});
+		child.stderr.on("data", (d: Buffer) => {
+			stderr += d.toString();
+		});
+
+		const timer = setTimeout(() => {
+			child.kill();
+		}, opts.timeout);
+
+		child.on("close", () => {
+			clearTimeout(timer);
+			resolve({
+				output: stdout || stderr,
+				duration: Date.now() - start,
+			});
+		});
+	});
+}

packages/evals/src/runner/braintrust.ts

@@ -0,0 +1,53 @@
+import assert from "node:assert";
+import { init } from "braintrust";
+import type { EvalRunResult } from "../types.js";
+
+/**
+ * Upload eval results to Braintrust as an experiment.
+ *
+ * Each EvalRunResult becomes a row in the experiment with:
+ * - input: scenario name + config
+ * - output: agent output summary
+ * - scores: pass (0 or 1)
+ * - metadata: model, skill toggle, duration, files modified
+ */
+export async function uploadToBraintrust(
+	results: EvalRunResult[],
+): Promise<void> {
+	assert(process.env.BRAINTRUST_API_KEY, "BRAINTRUST_API_KEY is not set");
+	assert(process.env.BRAINTRUST_PROJECT_ID, "BRAINTRUST_PROJECT_ID is not set");
+
+	const experiment = await init({
+		projectId: process.env.BRAINTRUST_PROJECT_ID,
+	});
+
+	for (const r of results) {
+		experiment.log({
+			input: {
+				scenario: r.scenario,
+				skillEnabled: r.skillEnabled,
+			},
+			output: {
+				status: r.status,
+				filesModified: r.filesModified,
+				testOutput: r.testOutput,
+			},
+			scores: {
+				pass: r.status === "passed" ? 1 : 0,
+			},
+			metadata: {
+				agent: r.agent,
+				model: r.model,
+				skillEnabled: r.skillEnabled,
+				duration: r.duration,
+				testsPassed: r.testsPassed,
+				testsTotal: r.testsTotal,
+				...(r.error ? { error: r.error } : {}),
+			},
+		});
+	}
+
+	const summary = await experiment.summarize();
+	console.log(`\nBraintrust experiment: ${summary.experimentUrl}`);
+	await experiment.close();
+}

packages/evals/src/runner/preflight.ts

@@ -0,0 +1,43 @@
+import { execFileSync } from "node:child_process";
+
+/**
+ * Verify the host environment has everything needed before spending
+ * API credits on an eval run.
+ *
+ * Checks: Node >= 20, Docker running, claude CLI available.
+ */
+export function preflight(): void {
+	const errors: string[] = [];
+
+	// Node.js >= 20
+	const [major] = process.versions.node.split(".").map(Number);
+	if (major < 20) {
+		errors.push(`Node.js >= 20 required (found ${process.versions.node})`);
+	}
+
+	// Docker daemon running
+	try {
+		execFileSync("docker", ["info"], { stdio: "ignore", timeout: 10_000 });
+	} catch {
+		errors.push("Docker is not running (required by supabase CLI)");
+	}
+
+	// Claude CLI available
+	try {
+		execFileSync("claude", ["--version"], {
+			stdio: "ignore",
+			timeout: 10_000,
+		});
+	} catch {
+		errors.push("claude CLI not found on PATH");
+	}
+
+	if (errors.length > 0) {
+		console.error("Preflight checks failed:\n");
+		for (const e of errors) {
+			console.error(`  - ${e}`);
+		}
+		console.error("");
+		process.exit(1);
+	}
+}

packages/evals/src/runner/results.ts

@@ -0,0 +1,68 @@
+import { mkdirSync, readdirSync, statSync, writeFileSync } from "node:fs";
+import { join, resolve } from "node:path";
+import type { EvalRunResult } from "../types.js";
+
+/**
+ * List files created or modified by the agent in the workspace.
+ * Compares against the original eval directory to find new files.
+ */
+export function listModifiedFiles(
+	workspacePath: string,
+	originalEvalDir: string,
+): string[] {
+	const modified: string[] = [];
+
+	function walk(dir: string, prefix: string) {
+		const entries = readdirSync(dir, { withFileTypes: true });
+		for (const entry of entries) {
+			if (
+				entry.name === "node_modules" ||
+				entry.name === "skills" ||
+				entry.name === "EVAL.ts" ||
+				entry.name === "EVAL.tsx"
+			)
+				continue;
+
+			const relPath = prefix ? `${prefix}/${entry.name}` : entry.name;
+			const fullPath = join(dir, entry.name);
+
+			if (entry.isDirectory()) {
+				walk(fullPath, relPath);
+			} else {
+				// Check if file is new (not in original eval dir)
+				const originalPath = join(originalEvalDir, relPath);
+				try {
+					statSync(originalPath);
+				} catch {
+					// File doesn't exist in original — it was created by the agent
+					modified.push(relPath);
+				}
+			}
+		}
+	}
+
+	walk(workspacePath, "");
+	return modified;
+}
+
+/** Print a summary table of eval results. */
+export function printSummary(results: EvalRunResult[]): void {
+	console.log("\n=== Eval Results ===\n");
+
+	for (const r of results) {
+		const icon = r.status === "passed" ? "PASS" : "FAIL";
+		const skill = r.skillEnabled ? "with-skill" : "baseline";
+		console.log(
+			`[${icon}] ${r.scenario} | ${r.model} | ${skill} | ${(r.duration / 1000).toFixed(1)}s`,
+		);
+		if (r.filesModified.length > 0) {
+			console.log(`       Files: ${r.filesModified.join(", ")}`);
+		}
+		if (r.status === "error" && r.error) {
+			console.log(`       Error: ${r.error}`);
+		}
+	}
+
+	const passed = results.filter((r) => r.status === "passed").length;
+	console.log(`\nTotal: ${passed}/${results.length} passed`);
+}

packages/evals/src/runner/scaffold.ts

@@ -0,0 +1,65 @@
+import {
+	cpSync,
+	existsSync,
+	mkdtempSync,
+	readdirSync,
+	rmSync,
+	symlinkSync,
+} from "node:fs";
+import { tmpdir } from "node:os";
+import { join, resolve } from "node:path";
+
+/** Walk up from cwd to find the repository root (contains skills/ and packages/). */
+function findRepoRoot(): string {
+	let dir = process.cwd();
+	for (let i = 0; i < 10; i++) {
+		if (existsSync(join(dir, "skills")) && existsSync(join(dir, "packages"))) {
+			return dir;
+		}
+		const parent = resolve(dir, "..");
+		if (parent === dir) break;
+		dir = parent;
+	}
+	throw new Error("Could not find repository root (skills/ + packages/)");
+}
+
+/**
+ * Create an isolated workspace for an eval run.
+ *
+ * 1. Copy the eval directory to a temp folder (excluding EVAL.ts)
+ * 2. Optionally symlink the supabase skill so Claude Code can discover it
+ *
+ * Returns the path to the workspace and a cleanup function.
+ */
+export function createWorkspace(opts: {
+	evalDir: string;
+	skillEnabled: boolean;
+}): { workspacePath: string; cleanup: () => void } {
+	const repoRoot = findRepoRoot();
+	const workspacePath = mkdtempSync(join(tmpdir(), "supabase-eval-"));
+
+	// Copy eval directory, excluding EVAL.ts (hidden from agent)
+	const entries = readdirSync(opts.evalDir, { withFileTypes: true });
+	for (const entry of entries) {
+		if (entry.name === "EVAL.ts" || entry.name === "EVAL.tsx") continue;
+		const src = join(opts.evalDir, entry.name);
+		const dest = join(workspacePath, entry.name);
+		cpSync(src, dest, { recursive: true });
+	}
+
+	// Make the skill available to the agent by symlinking the skills dir
+	if (opts.skillEnabled) {
+		const skillsDir = join(repoRoot, "skills");
+		if (existsSync(skillsDir)) {
+			const destSkills = join(workspacePath, "skills");
+			symlinkSync(skillsDir, destSkills);
+		}
+	}
+
+	return {
+		workspacePath,
+		cleanup: () => {
+			rmSync(workspacePath, { recursive: true, force: true });
+		},
+	};
+}

packages/evals/src/runner/test.ts

@@ -0,0 +1,97 @@
+import { execFile } from "node:child_process";
+import { copyFileSync, existsSync, writeFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+import { fileURLToPath } from "node:url";
+import { promisify } from "node:util";
+
+const __filename = fileURLToPath(import.meta.url);
+const __dirname = dirname(__filename);
+
+const exec = promisify(execFile);
+
+export interface TestResult {
+	passed: boolean;
+	output: string;
+	/** Number of tests that passed */
+	passedCount: number;
+	/** Total number of tests */
+	totalCount: number;
+}
+
+/**
+ * Run the hidden EVAL.ts tests against the agent's workspace.
+ *
+ * 1. Copy EVAL.ts into the workspace (agent is done, safe to expose)
+ * 2. Run vitest against it
+ * 3. Parse the output for pass/fail
+ */
+export async function runTests(opts: {
+	workspacePath: string;
+	evalFilePath: string;
+}): Promise<TestResult> {
+	// Copy the hidden test file into the workspace
+	const evalFileName = opts.evalFilePath.endsWith(".tsx")
+		? "EVAL.tsx"
+		: "EVAL.ts";
+	const destPath = join(opts.workspacePath, evalFileName);
+	copyFileSync(opts.evalFilePath, destPath);
+
+	// Write a minimal vitest config that overrides the default include pattern
+	// so EVAL.ts (without .test. or .spec.) is picked up.
+	const vitestConfigPath = join(opts.workspacePath, "vitest.config.mjs");
+	if (!existsSync(vitestConfigPath)) {
+		writeFileSync(
+			vitestConfigPath,
+			`export default { test: { include: ["EVAL.{ts,tsx}"] } };\n`,
+		);
+	}
+
+	// Use the vitest binary from the evals package (always available)
+	const evalsVitest = join(
+		__dirname,
+		"..",
+		"..",
+		"node_modules",
+		".bin",
+		"vitest",
+	);
+	const vitestBin = join(opts.workspacePath, "node_modules", ".bin", "vitest");
+	const cmd = existsSync(vitestBin) ? vitestBin : evalsVitest;
+	const args = ["run", evalFileName, "--reporter=verbose", "--no-color"];
+
+	try {
+		const { stdout, stderr } = await exec(cmd, args, {
+			cwd: opts.workspacePath,
+			timeout: 60_000,
+			env: { ...process.env },
+			maxBuffer: 5 * 1024 * 1024,
+		});
+
+		const output = `${stdout}\n${stderr}`;
+		return parseTestOutput(output);
+	} catch (error) {
+		const err = error as Error & { stdout?: string; stderr?: string };
+		const output = `${err.stdout ?? ""}\n${err.stderr ?? ""}`;
+		return parseTestOutput(output);
+	}
+}
+
+function parseTestOutput(output: string): TestResult {
+	// Parse vitest output for pass/fail counts
+	// Format: "Tests  N passed (M)" or "Tests  N failed | M passed (T)"
+	const testsLine = output.match(
+		/Tests\s+(?:(\d+)\s+failed\s+\|\s+)?(\d+)\s+passed\s+\((\d+)\)/,
+	);
+
+	let passedCount = 0;
+	let totalCount = 0;
+
+	if (testsLine) {
+		passedCount = Number.parseInt(testsLine[2], 10);
+		totalCount = Number.parseInt(testsLine[3], 10);
+	}
+
+	const passed = totalCount > 0 && passedCount === totalCount;
+
+	return { passed, output, passedCount, totalCount };
+}

packages/evals/src/scorer.ts

@@ -1,187 +0,0 @@
-import { generateText, Output } from "ai";
-import type { EvalScorer } from "braintrust";
-import { z } from "zod";
-import type { CodeFixTestCase } from "./dataset/types.js";
-import type { Expected, Input, Output as TaskOutput } from "./dataset.js";
-import { getModel } from "./models.js";
-
-const judgeModelId = process.env.EVAL_JUDGE_MODEL || "claude-opus-4-6";
-
-const scoreSchema = z.object({
-	score: z
-		.number()
-		.describe("Score from 0 to 1 (0 = bad, 0.5 = partial, 1 = good)"),
-	reasoning: z.string().describe("Brief reasoning for the score"),
-});
-
-const SYSTEM_PROMPT =
-	"You are a precise, consistent evaluator of Supabase code fixes. You assess whether LLM-generated code correctly addresses Supabase anti-patterns by comparing against reference solutions. You are fair: functionally equivalent solutions that differ in style or approach from the reference still receive high scores. You are strict: partial fixes, missing security measures, or incorrect patterns receive low scores. Always provide specific evidence for your scoring.";
-
-function buildContext(tc: CodeFixTestCase, llmOutput: string): string {
-	return `## Reference Information
-
-**Topic:** ${tc.title}
-**Explanation:** ${tc.explanation}
-
-## Original Incorrect Code
-
-\`\`\`${tc.badExample.language || ""}
-${tc.badExample.code}
-\`\`\`
-
-## Reference Correct Code (ground truth)
-
-\`\`\`${tc.goodExample.language || ""}
-${tc.goodExample.code}
-\`\`\`
-
-## LLM's Attempted Fix
-
-${llmOutput}`;
-}
-
-async function judge(
-	prompt: string,
-): Promise<{ score: number; reasoning: string }> {
-	const model = getModel(judgeModelId);
-	const { output } = await generateText({
-		model,
-		system: SYSTEM_PROMPT,
-		prompt,
-		output: Output.object({ schema: scoreSchema }),
-		temperature: 0.1,
-		maxRetries: 2,
-	});
-	if (!output) throw new Error("Judge returned no structured output");
-	return output;
-}
-
-export const correctnessScorer: EvalScorer<
-	Input,
-	TaskOutput,
-	Expected
-> = async ({ input, output }) => {
-	const context = buildContext(input.testCase, output.llmOutput);
-	const result = await judge(`${context}
-
-## Task
-
-Evaluate **correctness**: Does the LLM's fix address the core issue identified in the incorrect code?
-
-The fix does not need to be character-identical to the reference, but it must solve the same problem. Functionally equivalent or improved solutions should score well.
-
-Score 1 if the fix fully addresses the core issue, 0.5 if it partially addresses it, 0 if it fails to address the core issue or introduces new problems.`);
-
-	return {
-		name: "Correctness",
-		score: result.score,
-		metadata: { reasoning: result.reasoning },
-	};
-};
-
-export const completenessScorer: EvalScorer<
-	Input,
-	TaskOutput,
-	Expected
-> = async ({ input, output }) => {
-	const context = buildContext(input.testCase, output.llmOutput);
-	const result = await judge(`${context}
-
-## Task
-
-Evaluate **completeness**: Does the LLM's fix include ALL necessary changes shown in the reference?
-
-Check for missing RLS enablement, missing policy clauses, missing columns, incomplete migrations, or any partial fixes. The fix should be production-ready.
-
-Score 1 if all necessary changes are present, 0.5 if most changes are present but some are missing, 0 if significant changes are missing.`);
-
-	return {
-		name: "Completeness",
-		score: result.score,
-		metadata: { reasoning: result.reasoning },
-	};
-};
-
-export const bestPracticeScorer: EvalScorer<
-	Input,
-	TaskOutput,
-	Expected
-> = async ({ input, output }) => {
-	const context = buildContext(input.testCase, output.llmOutput);
-	const result = await judge(`${context}
-
-## Task
-
-Evaluate **best practices**: Does the LLM's fix follow Supabase best practices as demonstrated in the reference?
-
-Consider: RLS patterns, auth.users references, migration conventions, connection pooling, edge function patterns, SDK usage, and security-first defaults. Alternative correct approaches that achieve the same security/correctness goal are acceptable.
-
-Score 1 if the fix follows best practices, 0.5 if it mostly follows best practices with minor deviations, 0 if it uses anti-patterns or ignores conventions.`);
-
-	return {
-		name: "Best Practice",
-		score: result.score,
-		metadata: { reasoning: result.reasoning },
-	};
-};
-
-export const regressionSafetyScorer: EvalScorer<
-	Input,
-	TaskOutput,
-	Expected
-> = async ({ input, output }) => {
-	const context = buildContext(input.testCase, output.llmOutput);
-	const result = await judge(`${context}
-
-## Task
-
-Evaluate **regression safety**: Does the LLM's fix avoid introducing new problems?
-
-Carefully check whether the fix:
-- Breaks existing functionality that was working in the original code
-- Removes security measures (RLS policies, auth checks, input validation) that were already present
-- Changes function signatures, return types, or column names in ways that would break callers
-- Introduces SQL injection, XSS, or other security vulnerabilities not present in the original
-- Drops data, removes columns, or alters schemas destructively without necessity
-- Changes behavior beyond the scope of the identified issue
-
-The fix should repair the identified problem WITHOUT creating new ones. A fix that solves the original issue but breaks something else is dangerous in production.
-
-Score 1 if the fix introduces no new problems. Score 0.5 if the fix introduces minor issues (e.g., slightly different naming that could confuse but not break). Score 0 if the fix introduces a new bug, security vulnerability, or breaking change.`);
-
-	return {
-		name: "Regression Safety",
-		score: result.score,
-		metadata: { reasoning: result.reasoning },
-	};
-};
-
-export const minimalityScorer: EvalScorer<
-	Input,
-	TaskOutput,
-	Expected
-> = async ({ input, output }) => {
-	const context = buildContext(input.testCase, output.llmOutput);
-	const result = await judge(`${context}
-
-## Task
-
-Evaluate **minimality**: Does the LLM's fix make only the changes necessary to address the identified issue?
-
-Check whether the fix:
-- Rewrites or restructures code beyond what is needed to fix the problem
-- Adds features, abstractions, or utilities not present in the reference solution
-- Changes formatting, variable names, or style in unrelated parts of the code
-- Adds excessive comments, logging, or error handling not required by the fix
-- Over-engineers the solution (e.g., adding configuration options, generalization, or layers of abstraction when a simple targeted fix suffices)
-
-Compare the scope of changes in the LLM's fix against the reference. The reference represents the ideal minimal fix. The LLM's fix should be similarly focused.
-
-Score 1 if the fix is tightly scoped to the identified issue (similar scope to the reference). Score 0.5 if the fix includes some unnecessary changes but the core fix is present. Score 0 if the fix significantly over-reaches — rewriting large portions of code, adding unrelated features, or restructuring beyond what is needed.`);
-
-	return {
-		name: "Minimality",
-		score: result.score,
-		metadata: { reasoning: result.reasoning },
-	};
-};

packages/evals/src/types.ts

@@ -0,0 +1,35 @@
+export interface EvalScenario {
+	/** Directory name under evals/ */
+	id: string;
+	/** Human-readable name */
+	name: string;
+	/** Tags for filtering */
+	tags: string[];
+}
+
+export interface AgentConfig {
+	/** Agent identifier */
+	agent: "claude-code";
+	/** Model to use */
+	model: string;
+	/** Whether the supabase skill is available */
+	skillEnabled: boolean;
+}
+
+export interface EvalRunResult {
+	scenario: string;
+	agent: string;
+	model: string;
+	skillEnabled: boolean;
+	status: "passed" | "failed" | "error";
+	duration: number;
+	testOutput: string;
+	agentOutput: string;
+	/** Number of vitest tests that passed */
+	testsPassed: number;
+	/** Total number of vitest tests */
+	testsTotal: number;
+	/** Files the agent created or modified in the workspace */
+	filesModified: string[];
+	error?: string;
+}

packages/evals/tsconfig.json

@@ -12,5 +12,5 @@
 		"resolveJsonModule": true
 	},
 	"include": ["src/**/*"],
-	"exclude": ["node_modules", "dist"]
+	"exclude": ["node_modules", "dist", "evals"]
 }