feat(openrouter-images): add waza eval coverage

.gitignore

@@ -5,3 +5,5 @@ build/
 .env.*
 .DS_Store
 .claude/worktrees/
+.waza-results/
+.waza-cache/

.waza.yaml

@@ -0,0 +1,31 @@
+# yaml-language-server: $schema=https://raw.githubusercontent.com/microsoft/waza/main/schemas/config.schema.json
+
+paths:
+  skills: skills
+  evals: evals
+  results: .waza-results
+defaults:
+  engine: copilot-sdk
+  model: claude-sonnet-4.6
+  timeout: 300
+  parallel: false
+  workers: 4
+  verbose: false
+  sessionLog: false
+cache:
+  enabled: false
+  dir: .waza-cache
+server:
+  port: 3000
+  resultsDir: results/
+dev:
+  model: claude-sonnet-4-20250514
+  target: medium-high
+  maxIterations: 5
+tokens:
+  warningThreshold: 500
+  fallbackLimit: 1000
+graders:
+  programTimeout: 30
+storage:
+  containerName: waza-results

evals/_hooks/prepare-skill.ts

@@ -0,0 +1,55 @@
+#!/usr/bin/env bun
+
+import { spawnSync } from "node:child_process";
+import { existsSync } from "node:fs";
+import { homedir } from "node:os";
+import { dirname, join, resolve } from "node:path";
+import { fileURLToPath } from "node:url";
+
+// NOTE: Waza runs hooks without a shell, so one TypeScript command handles setup.
+
+function main(): void {
+  const skill = process.argv[2];
+  if (!skill) {
+    console.error("usage: prepare-skill.ts <skill-name>");
+    process.exit(2);
+  }
+
+  const here = dirname(fileURLToPath(import.meta.url));
+  const repoRoot = resolve(here, "..", "..");
+  const srcDir = join(repoRoot, "skills", skill);
+  if (!existsSync(srcDir)) {
+    console.error(`❌ skill source not found: ${srcDir}`);
+    process.exit(1);
+  }
+
+  const destSkillsRoot = join(homedir(), ".agents", "skills");
+  const destDir = join(destSkillsRoot, skill);
+
+  const mk = spawnSync("mkdir", ["-p", destSkillsRoot], { stdio: "inherit" });
+  if (mk.status !== 0) {
+    process.exit(mk.status ?? 1);
+  }
+
+  const sync = spawnSync("rsync", ["-a", "--delete", `${srcDir}/`, `${destDir}/`], {
+    stdio: "inherit",
+  });
+  if (sync.status !== 0) {
+    process.exit(sync.status ?? 1);
+  }
+
+  const scriptsPkg = join(destDir, "scripts", "package.json");
+  if (existsSync(scriptsPkg)) {
+    const install = spawnSync("npm", ["install", "--silent"], {
+      cwd: join(destDir, "scripts"),
+      stdio: "inherit",
+    });
+    if (install.status !== 0) {
+      process.exit(install.status ?? 1);
+    }
+  }
+
+  console.error(`✓ prepared skill "${skill}" → ${destDir}`);
+}
+
+main();

evals/openrouter-images/eval.yaml

@@ -0,0 +1,32 @@
+name: openrouter-images-eval
+description: |
+  Evaluates routing among image discovery, generation, and editing; indirect
+  triggers and anti-triggers; and migration to POST /api/v1/images with
+  data[].b64_json responses.
+skill: openrouter-images
+version: "1.0"
+config:
+  trials_per_task: 1
+  timeout_seconds: 300
+  parallel: false
+  executor: copilot-sdk
+  model: claude-opus-4.7
+metrics:
+  - name: task_completion
+    weight: 1.0
+    threshold: 0.8
+    description: Did the agent choose the correct workflow and dedicated API shape?
+
+hooks:
+  before_run:
+    - command: "bun evals/_hooks/prepare-skill.ts openrouter-images"
+
+graders:
+  - type: code
+    name: has_output
+    config:
+      assertions:
+        - "len(output) > 50"
+
+tasks:
+  - "tasks/*.yaml"

evals/openrouter-images/tasks/01-generate-basic.yaml

@@ -0,0 +1,30 @@
+id: generate-basic-001
+name: Generate Basic Image
+description: |
+  A new-image request should route to generate.ts on the dedicated Images API,
+  not edit.ts or a raw legacy endpoint.
+tags:
+  - happy-path
+  - generate
+
+inputs:
+  prompt: |
+    Generate an image of a red panda wearing sunglasses. Do not execute the
+    request because my API key is unavailable; show the exact bundled-script
+    command you would run.
+
+graders:
+  - type: code
+    name: uses_generate_script
+    config:
+      language: python
+      assertions:
+        - '"generate.ts" in output'
+        - '"edit.ts" not in output'
+        - '"red panda" in output.lower() and "sunglasses" in output.lower()'
+        - '"/api/v1/chat/completions" not in output and "/api/v1/responses" not in output'
+
+
+expected:
+  outcomes:
+    - type: task_completed

evals/openrouter-images/tasks/02-generate-with-aspect-ratio.yaml

@@ -0,0 +1,30 @@
+id: generate-aspect-ratio-001
+name: Generate With Aspect Ratio
+description: |
+  A specified landscape ratio should be passed through the bundled generate
+  script rather than only mentioned in prose.
+tags:
+  - happy-path
+  - generate
+  - aspect-ratio
+
+inputs:
+  prompt: |
+    Make a wide landscape image of a futuristic city at night, 16:9. Do not
+    execute it; show the exact bundled-script command.
+
+graders:
+  - type: code
+    name: aspect_ratio_flag
+    config:
+      language: python
+      assertions:
+        - '"generate.ts" in output'
+        - '"--aspect-ratio" in output'
+        - '"16:9" in output'
+        - '"city" in output.lower() and "night" in output.lower()'
+
+
+expected:
+  outcomes:
+    - type: task_completed

evals/openrouter-images/tasks/03-edit-image.yaml

@@ -0,0 +1,38 @@
+id: edit-image-001
+name: Edit Existing Image
+description: |
+  User wants to modify an existing image. The agent should use edit.ts,
+  pass the source path first, and avoid generate.ts and raw API calls.
+tags:
+  - happy-path
+  - edit
+
+inputs:
+  prompt: |
+    I have a file called photo.png. Edit it so the sky is purple. Do not
+    execute the request because my API key is unavailable; show the exact
+    bundled-script command you would run.
+
+graders:
+  - type: code
+    name: uses_only_edit_script
+    config:
+      language: python
+      assertions:
+        - '"edit.ts" in output'
+        - '"generate.ts" not in output'
+        - '"/api/v1/chat/completions" not in output and "/api/v1/responses" not in output'
+
+  - type: code
+    name: passes_source_then_prompt
+    config:
+      language: python
+      assertions:
+        - '"photo.png" in output'
+        - '"purple" in output.lower() and "sky" in output.lower()'
+        - 'output.find("photo.png") < output.lower().find("purple")'
+
+
+expected:
+  outcomes:
+    - type: task_completed

evals/openrouter-images/tasks/04-anti-trigger-image-theory.yaml

@@ -0,0 +1,28 @@
+id: anti-trigger-image-theory-001
+name: Anti-Trigger - Image Generation Theory
+description: |
+  A conceptual question should receive an explanation without invoking or
+  proposing image-generation scripts.
+tags:
+  - anti-trigger
+  - negative-test
+
+inputs:
+  prompt: |
+    How do diffusion models generate images from text? Explain the basic idea.
+    I am curious about the technique, not looking to generate anything.
+
+graders:
+  - type: code
+    name: no_image_scripts
+    config:
+      language: python
+      assertions:
+        - '"generate.ts" not in output'
+        - '"edit.ts" not in output'
+        - '"discover.ts" not in output'
+
+
+expected:
+  outcomes:
+    - type: task_completed

evals/openrouter-images/tasks/05-indirect-trigger-blog-hero.yaml

@@ -0,0 +1,30 @@
+id: indirect-trigger-blog-hero-001
+name: Indirect Trigger - Blog Hero Image
+description: |
+  A request for blog artwork should route to image generation even without the
+  phrase "generate an image", with an appropriate wide ratio.
+tags:
+  - happy-path
+  - indirect-trigger
+
+inputs:
+  prompt: |
+    I am writing a blog post about remote work and need a hero image for the
+    top of the page: a quiet home office with good natural light. Do not execute
+    it; show the exact bundled-script command you would run.
+
+graders:
+  - type: code
+    name: hero_generation_command
+    config:
+      language: python
+      assertions:
+        - '"generate.ts" in output'
+        - '"home office" in output.lower() or "remote work" in output.lower()'
+        - '"--aspect-ratio" in output'
+        - '("16:9" in output or "21:9" in output or "3:1" in output or "2:1" in output)'
+
+
+expected:
+  outcomes:
+    - type: task_completed

evals/openrouter-images/tasks/06-discover-model-capabilities.yaml

@@ -0,0 +1,29 @@
+id: discover-model-capabilities-001
+name: Discover Model Capabilities
+description: |
+  A model capability question should use discover.ts with the model ID before
+  recommending unsupported generation options.
+tags:
+  - discovery
+  - model-selection
+
+inputs:
+  prompt: |
+    Before I generate anything, how would you check which parameters and
+    providers google/gemini-3.1-flash-image-preview supports? Show the exact
+    bundled command, but do not execute it.
+
+graders:
+  - type: code
+    name: uses_discovery_script
+    config:
+      language: python
+      assertions:
+        - '"discover.ts" in output'
+        - '"google/gemini-3.1-flash-image-preview" in output'
+        - '"generate.ts" not in output and "edit.ts" not in output'
+
+
+expected:
+  outcomes:
+    - type: task_completed

evals/openrouter-images/tasks/07-dedicated-api-code.yaml

@@ -0,0 +1,32 @@
+id: dedicated-images-api-code-001
+name: Dedicated Images API Code
+description: |
+  A developer asking for direct API code should be taught the new dedicated
+  endpoint and response shape rather than legacy chat or responses formats.
+tags:
+  - api
+  - migration
+
+inputs:
+  prompt: |
+    Give me a minimal TypeScript fetch example for generating an image through
+    OpenRouter's new dedicated images endpoint. Include how to read the returned
+    base64 image, but do not send a real request.
+
+graders:
+  - type: code
+    name: dedicated_endpoint_shape
+    config:
+      language: python
+      assertions:
+        - '"/api/v1/images" in output'
+        - '"POST" in output.upper()'
+        - '"b64_json" in output'
+        - '"data" in output'
+        - '"/api/v1/chat/completions" not in output'
+        - '"choices" not in output or "not" in output.lower()'
+
+
+expected:
+  outcomes:
+    - type: task_completed

skills/openrouter-images/README.md

@@ -12,7 +12,7 @@ gh skill install OpenRouterTeam/skills openrouter-images
 
 Works with Claude Code, Cursor, Codex, OpenCode, Gemini CLI, Windsurf, and [many more agents](https://cli.github.com/manual/gh_skill_install). Add `--scope user` to install across every project for your current agent, or `--agent claude-code` to target a specific agent.
 
-For other install methods (Claude Code plugin marketplace, Cursor Rules, etc.) see the [root README](../../README.md#installing).
+The repository root README documents additional install methods for Claude Code, Cursor, and other agents.
 
 ## Prerequisites