✨ Add MCP App groundwork with Claude/Codex parity

.agents/skills/check-changed-folders/SKILL.md

@@ -1,11 +1,11 @@
 ---
 name: check-changed-folders
-description: Run typecheck, tests, and builds for Rudder's core and plugin workspaces on the current branch versus main, verify the centralized agent-instruction layout, and verify agent attribution. Use when asked to run "/check", to validate a branch before commit/PR, or whenever a user asks to run checks before publishing.
+description: Run typecheck, tests, and builds for Rudder's core and plugin workspaces on the current branch versus main, verify Claude/Codex provider parity, verify the centralized agent-instruction layout, and verify agent attribution. Use when asked to run "/check", to validate a branch before commit/PR, whenever a user asks to run checks before publishing, or whenever changes affect provider-facing plugin surfaces such as manifests, hooks, skills, MCP servers, apps, or UI.
 ---
 
 # Check Changed Folders
 
-Identify what changed on the branch, verify the centralized agent-instruction layout and agent attribution, run the package checks, and report pass/fail status with actionable failure output.
+Identify what changed on the branch, verify provider parity, the centralized agent-instruction layout, and agent attribution, run the package checks, and report pass/fail status with actionable failure output.
 The root npm scripts validate the core package and the Rudder plugin workspace.
 Checks are therefore repo-wide rather than per-package.
 
@@ -20,7 +20,7 @@ git diff --name-only
 git diff --name-only --cached
 ```
 
-   If anything under `src/`, `bin/`, `test/`, or a build/config file (`package.json`, `tsconfig*.json`, `.github/`) changed, the package checks below apply.
+   If anything under `src/`, `bin/`, `test/`, `ui/`, `.claude-plugin/`, `.codex-plugin/`, `hooks/`, `skills/`, or a build/config file (`package.json`, `tsconfig*.json`, `.mcp.json`, `.github/`) changed, the package checks below apply.
 
 2. Verify the centralized agent-instruction layout before running checks:
 
@@ -33,39 +33,51 @@ git diff --name-only --cached
   - `.codex/skills` -> `../.agents/skills`
 - If any link is missing or resolves outside `.agents/`, mark the check as failed and report the broken path.
 
-3. Verify agent attribution.
+3. Verify provider parity.
+
+- Treat Claude and Codex as supported providers for the same Rudder plugin package.
+- When a change adds, removes, or modifies a provider-facing component—manifest metadata, skills, hooks, MCP servers, apps, UI, or packaged artifacts—inspect both provider manifests and runtime paths.
+- Require equivalent discovery, launch behavior, packaged resources, and user-facing capability across providers.
+  Provider-specific schema or environment-variable differences are allowed, but they must remain explicit implementation details rather than functional gaps.
+- Add or update a regression assertion in `test/plugin-package.test.ts` or the relevant runtime test that exercises both provider paths.
+  If provider-facing behavior changes without parity coverage, fail this check.
+- Permit a provider-specific exception only when the user explicitly requests it or the provider cannot support the capability.
+  Report the reason, the affected provider, and the test that protects the supported behavior.
+
+4. Verify agent attribution.
    If a coding agent wrote code included in the
 branch, inspect `git log origin/main..HEAD` and require every such agent to appear as a commit author or in a `Co-authored-by:` trailer.
 Missing agent attribution on committed work fails the check.
 If the agent-written work is still uncommitted, report attribution as pending and name the trailer that must be added when committing.
 Human-only changes are not subject to this check.
 
-4. Install dependencies if `node_modules/` does not exist: run `npm install`.
+5. Install dependencies if `node_modules/` does not exist: run `npm install`.
 
-5. Run the package checks (the same set `prepublishOnly` runs, so green means publishable):
+6. Run the package checks (the same set `prepublishOnly` runs, so green means publishable):
 
 ```bash
 npm run typecheck
 npm test
 npm run build
 ```
 
-6. Surface and address open PR comments.
+7. Surface and address open PR comments.
    If the current branch has an open GitHub PR, always invoke the `address-pr-comments` skill before finishing.
    That skill fetches open review comments (Greptile, human reviewers) for the PR, dedupes them, and fixes/declines/defers each one.
    Only if `gh` is unavailable or there is no PR for the current branch, treat this step as `skipped`.
    If any comment is acted on, re-run the checks before reporting.
 
-7. Report concise results:
+8. Report concise results:
 
 - State whether typecheck, tests, and build passed, failed, or were skipped.
-- Include dedicated results for agent-instruction layout, agent attribution, and PR comments (`passed`, `failed`, `pending`, or `skipped`) and why.
+- Include dedicated results for provider parity, agent-instruction layout, agent attribution, and PR comments (`passed`, `failed`, `pending`, or `skipped`) and why.
 - For failures, include the key error output and which command failed.
-- Distinguish real failures (broken agent links, missing agent attribution, type errors, test failures, unaddressed P0/P1 PR comments) from environment issues (missing CLI tools, no PR).
+- Distinguish real failures (provider drift, missing parity coverage, broken agent links, missing agent attribution, type errors, test failures, unaddressed P0/P1 PR comments) from environment issues (missing CLI tools, no PR).
 
 ## Notes
 
 - Default comparison branch is `origin/main` (not local `main`, which may be stale).
+- Provider parity means equivalent behavior and coverage, not necessarily byte-identical provider configuration.
 - The `.agents/` layout is a required gate, not an optional reminder.
 - Agent attribution is required only when an agent contributed code.
 - If nothing relevant changed, state that no checks were required.

.agents/skills/check-changed-folders/agents/openai.yaml

@@ -1,4 +1,4 @@
 interface:
   display_name: "Check"
-  short_description: "Run checks and verify agent attribution and the shared workflow layout."
-  default_prompt: "Use $check-changed-folders to verify agent attribution and the .agents-based workflow layout, then run typecheck, tests, and build."
+  short_description: "Validate provider parity, layout, attribution, and builds."
+  default_prompt: "Use $check-changed-folders to verify Claude/Codex provider parity, attribution, and the shared workflow layout, then run the repository checks."

.claude-mcp.json

@@ -0,0 +1,13 @@
+{
+  "mcpServers": {
+    "rudder": {
+      "command": "node",
+      "args": [
+        "--input-type=module",
+        "-e",
+        "import { resolve } from 'node:path'; import { pathToFileURL } from 'node:url'; const root = process.env.PLUGIN_ROOT || process.env.CLAUDE_PLUGIN_ROOT || process.cwd(); await import(pathToFileURL(resolve(root, 'dist/rudder-mcp-server.mjs')).href)"
+      ],
+      "cwd": "."
+    }
+  }
+}

.claude-plugin/plugin.json

@@ -17,5 +17,6 @@
     "coverage"
   ],
   "skills": "./skills/",
+  "mcpServers": "./.claude-mcp.json",
   "hooks": "./hooks/hooks.json"
 }

.codex-plugin/plugin.json

@@ -18,13 +18,15 @@
     "coverage"
   ],
   "skills": "./skills/",
+  "mcpServers": "./.mcp.json",
   "interface": {
     "displayName": "Rudder",
     "shortDescription": "Tests from intent and changes",
     "longDescription": "Rudder combines the intent captured in your coding session with the current worktree, then guides your existing coding agent through a fresh-slate unit-test workflow. It discovers the repository's native test and coverage tooling, generates focused tests for changed production behavior, and asks concrete follow-up questions until the configured coverage target is met.",
     "developerName": "RudderCode",
     "category": "Productivity",
     "capabilities": [
+      "Interactive",
       "Write"
     ],
     "websiteURL": "https://github.com/RudderCode/Rudder",

.mcp.json

@@ -0,0 +1,11 @@
+{
+  "rudder": {
+    "command": "node",
+    "args": [
+      "--input-type=module",
+      "-e",
+      "import { resolve } from 'node:path'; import { pathToFileURL } from 'node:url'; const root = process.env.PLUGIN_ROOT || process.env.CLAUDE_PLUGIN_ROOT || process.cwd(); await import(pathToFileURL(resolve(root, 'dist/rudder-mcp-server.mjs')).href)"
+    ],
+    "cwd": "."
+  }
+}

almanac/architecture/automation/contributor-automation.md

@@ -1,6 +1,6 @@
 ---
 title: "Contributor Automation"
-summary: "Contributor automation connects centralized agent skills, local check flows, PR-comment remediation, CI validation, and Danger-based agent guards into one guarded branch workflow."
+summary: "Contributor automation connects centralized agent skills, provider-parity checks, local check flows, PR-comment remediation, CI validation, and Danger-based agent guards into one guarded branch workflow."
 topics: [architecture, automation, contributor-workflow, validation]
 sources:
   - id: agents-readme
@@ -29,14 +29,16 @@ sources:
     path: dangerfile.ts
 ---
 
-Rudder's contributor automation is a set of local and CI gates for a repository that currently has one root plugin package and centralized agent workflows. `.agents/skills/` is the only reusable-workflow source, with `.claude/skills` and `.codex/skills` as compatibility symlinks [@agents-readme]. The `check-changed-folders` skill validates branches, verifies layout and attribution, runs local package checks, and delegates PR-comment remediation when a PR exists [@check-skill]. The `prepare-package-release` skill synchronizes package and plugin versions, ingests the complete range since the previous release, Gardens the whole CodeAlmanac wiki, and validates the prepared release [@release-skill]. GitHub Actions repeats package validation on branch pushes, while the Danger workflow enforces protected paths and inline agent guards for agent-authored pull requests [@test-workflow] [@danger-workflow] [@dangerfile].
+Rudder's contributor automation is a set of local and CI gates for a repository that currently has one root plugin package and centralized agent workflows. `.agents/skills/` is the only reusable-workflow source, with `.claude/skills` and `.codex/skills` as compatibility symlinks [@agents-readme]. The `check-changed-folders` skill validates branches, verifies provider parity, layout, and attribution, runs local package checks, and delegates PR-comment remediation when a PR exists [@check-skill]. The `prepare-package-release` skill synchronizes package and plugin versions, ingests the complete range since the previous release, Gardens the whole CodeAlmanac wiki, and validates the prepared release [@release-skill]. GitHub Actions repeats package validation on branch pushes, while the Danger workflow enforces protected paths and inline agent guards for agent-authored pull requests [@test-workflow] [@danger-workflow] [@dangerfile].
 
 ## Local Check Surface
 
-The check surface is centralized in `.agents/skills/check-changed-folders/SKILL.md` [@check-skill]. It starts by fetching `origin/main`, collecting changed files from the merge-base diff plus unstaged and staged local changes, and treating Rudder as repo-wide rather than per-package [@check-skill]. When files under `src/`, `bin/`, `test/`, package or TypeScript configuration, or `.github/` change, the package checks apply [@check-skill].
+The check surface is centralized in `.agents/skills/check-changed-folders/SKILL.md` [@check-skill]. It starts by fetching `origin/main`, collecting changed files from the merge-base diff plus unstaged and staged local changes, and treating Rudder as repo-wide rather than per-package [@check-skill]. When files under `src/`, `bin/`, `test/`, `ui/`, `.claude-plugin/`, `.codex-plugin/`, `hooks/`, `skills/`, package or TypeScript configuration, `.mcp.json`, or `.github/` change, the package checks apply [@check-skill].
 
 The centralized layout itself is a hard gate. The check skill requires `AGENTS.md` as canonical guidance, `.agents/skills/` as the only reusable workflow source, `skills/<skill-name>/agents/openai.yaml` metadata for shared skills, `.claude/skills` and `.codex/skills` symlinks to `.agents/skills`, and no `.claude/commands` aliases [@agents-readme] [@check-skill]. The [Run Checks](../../guides/contributor/run-checks) guide turns this architecture into the step-by-step contributor procedure.
 
+Provider parity is a hard local gate for provider-facing changes. The check skill treats Claude and Codex as supported providers for the same package, requires equivalent discovery, launch behavior, packaged resources, and user-facing capability, and requires regression coverage in `test/plugin-package.test.ts` or the relevant runtime test for both provider paths [@check-skill]. Provider-specific schema or environment-variable differences are allowed only when they do not become functional gaps; exceptions need an explicit user request or an unsupported provider capability, plus a reported reason and protective test for the supported path [@check-skill].
+
 Before package checks, the local flow also checks that each coding agent represented in committed work is listed as a commit author or `Co-authored-by` trailer. Agent-written uncommitted work is reported as pending attribution until it is committed; human-only changes are outside this gate [@check-skill].
 
 ## Package Checks