Adopt generated registries and repo-root Gemini/Kiro emission

.aipm/generated-root.json

@@ -0,0 +1,17 @@
+{
+  "version": 1,
+  "paths": [
+    ".kiro/agents/experimenter.json",
+    ".kiro/agents/test-subject.json",
+    "GEMINI.md",
+    "POWER.md",
+    "agents/experimenter.md",
+    "agents/test-subject.md",
+    "commands/evaluate.toml",
+    "gemini-extension.json",
+    "hooks/hooks.json",
+    "mcp.json",
+    "skills/evaluate-skill/SKILL.md",
+    "steering/evaluation-workflow.md"
+  ]
+}

.claude-plugin/marketplace.json

@@ -4,15 +4,19 @@
     "name": "AI Plugin Marketplace Template"
   },
   "metadata": {
-    "version": "0.0.1",
     "description": "Universal AI Plugin Marketplace — author once, distribute to all platforms"
   },
   "plugins": [
     {
       "name": "skill-evaluator",
       "source": "./plugins/skill-evaluator",
       "description": "Evaluate AI skills across model tiers with blind testing and refinement recommendations",
-      "tags": ["evaluation", "testing", "skills", "model-tiers"]
+      "tags": [
+        "evaluation",
+        "testing",
+        "skills",
+        "model-tiers"
+      ]
     }
   ]
 }

.cursor-plugin/marketplace.json

@@ -4,15 +4,19 @@
     "name": "AI Plugin Marketplace Template"
   },
   "metadata": {
-    "version": "0.0.1",
     "description": "Universal AI Plugin Marketplace — author once, distribute to all platforms"
   },
   "plugins": [
     {
       "name": "skill-evaluator",
       "source": "./plugins/skill-evaluator",
       "description": "Evaluate AI skills across model tiers with blind testing and refinement recommendations",
-      "tags": ["evaluation", "testing", "skills", "model-tiers"]
+      "tags": [
+        "evaluation",
+        "testing",
+        "skills",
+        "model-tiers"
+      ]
     }
   ]
 }

.kiro/agents/experimenter.json

@@ -0,0 +1,21 @@
+{
+  "name": "experimenter",
+  "description": "Orchestrates blind skill evaluation across model tiers",
+  "prompt": "# Experimenter Agent\n\nYou are the experimenter in a blind skill evaluation. Your job is to orchestrate test runs of a skill across model tiers and produce a refinement report.\n\n## Principles\n\n- **Blind testing**: Never reveal expected outcomes to test subjects\n- **Structured protocol**: Define pass/fail criteria BEFORE running tests\n- **Systematic comparison**: Evaluate each tier independently before comparing across tiers\n- **Actionable output**: Every identified failure must include a specific recommendation\n\n## Workflow\n\n1. Receive the skill content and test cases from the evaluate-skill skill\n2. For each model tier (opus, sonnet, haiku):\n   a. For each test case, spawn a test-subject agent at the appropriate tier\n   b. Provide only the skill content and the input — never the expected outcome\n   c. Collect and store the output\n3. Compare outputs against expected outcomes\n4. Generate a structured refinement report\n\n## Report Format\n\n```\n# Skill Evaluation Report\n\n## Summary\n- Skill: [name]\n- Clarity Floor: [lowest passing tier]\n- Overall Pass Rate: [X/Y]\n\n## Per-Tier Results\n### Opus\n| Test Case | Pass/Fail | Notes |\n|-----------|-----------|-------|\n| ...       | ...       | ...   |\n\n### Sonnet\n...\n\n### Haiku\n...\n\n## Failure Analysis\n### [Test Case N at Tier X]\n- **Symptom**: [what went wrong]\n- **Root Cause**: [why the lower-tier agent failed]\n- **Recommendation**: [specific improvement to the skill]\n\n## Recommendations\n1. [Ordered list of improvements, highest impact first]\n```",
+  "mcpServers": {},
+  "tools": [
+    "delegate",
+    "read",
+    "write",
+    "glob",
+    "grep",
+    "shell"
+  ],
+  "toolAliases": {},
+  "allowedTools": [],
+  "resources": [],
+  "hooks": {},
+  "toolsSettings": {},
+  "includeMcpJson": true,
+  "model": null
+}

.kiro/agents/test-subject.json

@@ -0,0 +1,20 @@
+{
+  "name": "test-subject",
+  "description": "Blind agent that executes a skill and produces output without knowledge of expected outcomes",
+  "prompt": "# Test Subject Agent\n\n<!-- The model tier above (`sonnet`) is a default. The experimenter agent typically overrides this when spawning test subjects at different tiers (opus → sonnet → haiku). -->\n\nYou are a test subject in a blind skill evaluation. You will receive a skill and an input. Execute the skill to the best of your ability and produce your output.\n\n## Rules\n\n- You do NOT know what the expected outcome is — just do your best\n- Follow the skill instructions exactly as written\n- If the skill instructions are ambiguous, make your best interpretation and note the ambiguity\n- Do not ask clarifying questions — work with what you have\n- Produce your output in a clear, structured format\n\n## Process\n\n1. Read the skill content provided to you\n2. Read the input provided to you\n3. Execute the skill's instructions against the input\n4. Produce your output\n\n## Output Format\n\nProduce your result in this format:\n\n```\n## Output\n\n[Your skill execution output here]\n\n## Notes\n\n- [Any ambiguities encountered]\n- [Assumptions made]\n- [Difficulties faced]\n```",
+  "mcpServers": {},
+  "tools": [
+    "read",
+    "write",
+    "shell",
+    "glob",
+    "grep"
+  ],
+  "toolAliases": {},
+  "allowedTools": [],
+  "resources": [],
+  "hooks": {},
+  "toolsSettings": {},
+  "includeMcpJson": true,
+  "model": null
+}

CONTRIBUTING.md

@@ -9,17 +9,19 @@ cloning, run:
 pnpm install
 ```
 
-This requires `@ai-plugin-marketplace/cli` and `@ai-plugin-marketplace/core` to
-be **published to npm**. A fresh `pnpm install` regenerates `pnpm-lock.yaml`
-against the published versions.
+`@ai-plugin-marketplace/cli` and `@ai-plugin-marketplace/core` are published to
+npm, so a fresh `pnpm install` resolves them and writes `pnpm-lock.yaml` against
+the published versions.
 
 ## Workflow
 
 1. Add or edit a plugin under `plugins/<name>/`. Use `aipm scaffold <name>` to
    create a new one or `aipm add-target <plugin> <target>` to expand an
    existing plugin's support envelope.
-2. Run `aipm build` to regenerate toolkit-owned artifacts (hook JSON and the
-   `dist/` bundles). Commit both your authored sources and the regenerated
+2. Run `aipm build` to regenerate toolkit-owned artifacts: hook JSON, the
+   `dist/` bundles, the marketplace registries (generated from
+   `aipm.workspace.ts`), and — for a single-plugin marketplace — the repo-root
+   Gemini/Kiro artifacts. Commit both your authored sources and the regenerated
    output — CI enforces that the tree is clean after a build (freshness).
 3. Run `aipm validate` and resolve any hard findings.
 

GEMINI.md

@@ -0,0 +1,26 @@
+# Skill Evaluator
+
+Evaluate AI skills across model tiers with blind testing and refinement recommendations.
+
+## Overview
+
+This extension tests AI skills across model tiers (opus → sonnet → haiku) using blind sub-agent testing. It identifies where skill clarity degrades and provides actionable refinement recommendations.
+
+## Usage
+
+Use the `/evaluate` command to start an evaluation:
+
+```
+/evaluate <skill-path> <test-cases-path>
+```
+
+- `skill-path`: Path to the SKILL.md file to evaluate
+- `test-cases-path`: Path to a JSON file with test cases (array of `{ input, expectedOutcome }`)
+
+## How It Works
+
+1. The experimenter agent reads the skill and test cases
+2. Separate blind test-subject agents are spawned per test case at the highest tier (opus)
+3. The process repeats at lower tiers (sonnet, then haiku)
+4. Results are compared against expected outcomes per tier
+5. A refinement report identifies where clarity breaks down and recommends improvements

POWER.md

@@ -0,0 +1,28 @@
+---
+schemaVersion: 0.1.0
+name: skill-evaluator
+description: Evaluate AI skills across model tiers with blind testing and refinement recommendations
+version: 0.0.1
+---
+
+# Skill Evaluator
+
+This power evaluates AI skills across model tiers (opus → sonnet → haiku) using blind sub-agent testing.
+
+## Capabilities
+
+- **Blind evaluation**: Test subjects execute skills without knowledge of expected outcomes
+- **Multi-tier testing**: Run skills at progressively lower model tiers to find clarity breakpoints
+- **Refinement recommendations**: Generate specific suggestions to improve skill clarity
+
+## Workflow
+
+1. Provide a skill path and test cases (input/expected-outcome pairs)
+2. The experimenter agent orchestrates blind test runs across model tiers
+3. Results are compared against expected outcomes
+4. A refinement report is generated with specific recommendations
+
+## Related Files
+
+- `steering/evaluation-workflow.md` — Evaluation workflow steering
+- `mcp.json` — MCP server configuration (empty — this power is skill/agent-based)

README.md

@@ -22,8 +22,8 @@ then:
 pnpm install
 ```
 
-> **Note:** `pnpm install` requires the `@ai-plugin-marketplace/*` packages to be
-> published to npm. See [CONTRIBUTING.md](CONTRIBUTING.md).
+The `@ai-plugin-marketplace/*` packages are published to npm, so a fresh
+`pnpm install` resolves them directly. See [CONTRIBUTING.md](CONTRIBUTING.md).
 
 ## Authoring plugins
 
@@ -43,11 +43,20 @@ repo-root marketplace registries.
 aipm build
 ```
 
-`aipm build` regenerates every toolkit-owned artifact — per-plugin hook JSON
-(`hooks/claude.json`, `hooks/hooks.json`) and the standalone bundles under
-`dist/gemini/` and `dist/kiro/`. Both authored sources and generated outputs are
-committed, so plugins stay browsable on GitHub and pull-request diffs stay
-honest.
+`aipm build` regenerates every toolkit-owned artifact:
+
+- per-plugin hook JSON (`hooks/claude.json`, `hooks/hooks.json`);
+- the standalone bundles under `dist/gemini/` and `dist/kiro/`;
+- the marketplace registries (`.claude-plugin/marketplace.json`,
+  `.cursor-plugin/marketplace.json`) — generated from `aipm.workspace.ts` plus
+  each plugin's `aipm.config.ts` (see [Marketplace metadata](#marketplace-metadata-and-generated-registries));
+- for a single-plugin marketplace, the **repo-root Gemini extension and Kiro
+  power** (`gemini-extension.json`, `GEMINI.md`, `POWER.md`, `commands/`,
+  `skills/`, `steering/`, `.kiro/`, …) so the repo installs natively into hosts
+  that have no marketplace concept.
+
+Both authored sources and generated outputs are committed, so plugins stay
+browsable on GitHub and pull-request diffs stay honest.
 
 ### Validate
 
@@ -57,8 +66,8 @@ aipm validate
 
 Validation checks the support envelope, every target manifest's schema,
 cross-target name consistency, MCP server-key sync, marketplace registration,
-and freshness (that the committed generated artifacts match what `aipm build`
-would produce).
+and freshness — that the committed generated artifacts (registries, bundles,
+and the repo-root Gemini/Kiro emission) match what `aipm build` would produce.
 
 `package.json` exposes these as `pnpm build`, `pnpm check`, and
 `pnpm scaffold` if you prefer the npm-script entry points.
@@ -81,24 +90,76 @@ plugin that carries files for a target outside its envelope. To expand a
 plugin's envelope, run `aipm add-target <plugin> <target>` to scaffold the
 skeleton files for a new target, then fill in the manifest fields.
 
+## Marketplace metadata and generated registries
+
+Repo-level marketplace metadata lives in `aipm.workspace.ts` at the repo root:
+
+```ts
+import { defineWorkspace } from '@ai-plugin-marketplace/core';
+
+export default defineWorkspace({
+  marketplace: {
+    name: 'ai-plugin-marketplace',
+    owner: { name: 'AI Plugin Marketplace Template' },
+    description: 'Universal AI Plugin Marketplace — author once, distribute to all platforms',
+  },
+});
+```
+
+The **presence of this file opts the repo into generated registries.** Instead
+of hand-maintaining `.claude-plugin/marketplace.json` and
+`.cursor-plugin/marketplace.json`, `aipm build` generates them from the
+workspace metadata plus each plugin's `aipm.config.ts` (its `description` and
+`keywords` become the registry entry's `description` and `tags`). The generated
+registries are committed and freshness-checked like every other artifact.
+
+### Gemini and Kiro: the single-plugin marketplace
+
+Gemini CLI and Kiro have **no marketplace concept** — a repo is installed as one
+extension (Gemini) or one power (Kiro) from its root. So when a marketplace
+exposes **exactly one plugin**, `aipm build` additionally emits that plugin's
+Gemini/Kiro artifacts at the **repo root** (`gemini-extension.json`, `GEMINI.md`,
+`POWER.md`, `commands/`, `skills/`, `steering/`, `.kiro/`), letting the repo
+install natively into those hosts. A sidecar at `.aipm/generated-root.json`
+records exactly which root paths the toolkit owns.
+
+If you add a **second plugin**, the repo can no longer be a single Gemini/Kiro
+artifact: `aipm validate` reports a `single-artifact-host` finding, and you keep
+full Claude/Cursor marketplace support (which can host many plugins) while
+choosing one plugin to expose to Gemini/Kiro. Because the repo root is the
+single Gemini/Kiro artifact, the plugin does **not** carry its own `LICENSE` or
+`README.md` — the repo-root `LICENSE`/`README.md` are canonical and are never
+overwritten by generation.
+
 ## Repository structure
 
 ```
 .
+├── aipm.workspace.ts        # Marketplace metadata (opts into generated registries)
 ├── .claude-plugin/
-│   └── marketplace.json     # Claude Code marketplace registry
+│   └── marketplace.json     # Claude Code registry (GENERATED)
 ├── .cursor-plugin/
-│   └── marketplace.json     # Cursor marketplace registry
+│   └── marketplace.json     # Cursor registry (GENERATED)
 ├── plugins/
 │   └── <plugin-name>/       # One directory per plugin (authored sources)
-│       ├── aipm.config.ts   # Support envelope + plugin version
+│       ├── aipm.config.ts   # Support envelope + version + description/keywords
 │       └── ...              # Manifests, agents, skills, commands, hooks, rules
 ├── dist/                    # Generated standalone bundles (committed)
 │   ├── gemini/
 │   └── kiro/
+│
+│   # Repo-root Gemini/Kiro artifacts — emitted for a single-plugin marketplace
+│   # (GENERATED; tracked in .aipm/generated-root.json):
+├── gemini-extension.json
+├── GEMINI.md
+├── POWER.md
+├── commands/  skills/  steering/  agents/  hooks/  mcp.json  .kiro/
+├── .aipm/
+│   └── generated-root.json  # Records which repo-root paths the toolkit owns
+│
 ├── package.json             # Depends on @ai-plugin-marketplace/cli + core
-├── LICENSE
-└── README.md
+├── LICENSE                  # Canonical — also serves the Gemini/Kiro artifact
+└── README.md                # Canonical — also serves the Gemini/Kiro artifact
 ```
 
 ## Upgrading the toolkit
@@ -119,9 +180,18 @@ when.
 ## Example plugin: skill-evaluator
 
 The included `skill-evaluator` plugin demonstrates the full multi-platform
-pattern end-to-end — it evaluates AI skills across model tiers (opus → sonnet →
-haiku) using blind sub-agent testing. See
-[plugins/skill-evaluator/README.md](plugins/skill-evaluator/README.md).
+pattern end-to-end. Given a skill (`SKILL.md`) and a set of test cases, it:
+
+1. runs the skill with blind test-subject agents at different model tiers
+   (opus → sonnet → haiku);
+2. compares outputs against expected outcomes;
+3. identifies where skill clarity degrades at lower tiers;
+4. generates actionable refinement recommendations.
+
+Invoke it in any host with `/evaluate path/to/SKILL.md path/to/test-cases.json`,
+where each test case is `{ "input": "...", "expectedOutcome": "..." }`. The skill
+definition lives at
+[plugins/skill-evaluator/skills/evaluate-skill/SKILL.md](plugins/skill-evaluator/skills/evaluate-skill/SKILL.md).
 
 ## License