Add MCP harness regression guard

.github/workflows/verify.yml

@@ -0,0 +1,33 @@
+name: Verify
+
+on:
+  pull_request:
+    branches: [master]
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  verify:
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout
+        uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
+
+      - name: Setup Pnpm
+        run: |
+          npm install -g corepack@latest --force
+          corepack enable
+
+      - name: Setup Node.js
+        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e # v6.4.0
+        with:
+          node-version: 24
+          cache: pnpm
+
+      - name: Install Dependencies
+        run: pnpm install --frozen-lockfile
+
+      - name: Verify
+        run: pnpm verify

AGENTS.md

@@ -12,23 +12,28 @@
   - `packages/cli`
   - `packages/mcp`
 
-`error2fix` is a post-failure terminal diagnosis CLI. Users do not wrap commands with `e2f`. They run `e2f init` once, let normal shell commands fail, and then run `e2f` to inspect the latest captured failure context.
+`error2fix` is a post-failure frontend failure diagnosis toolkit for CLI users and AI coding agents. Users do not wrap commands with `e2f`. They run `e2f init` once, let normal shell commands fail, and then use the CLI or MCP tools to inspect compact failure context.
 
 ## Key Commands
 
 - Install dependencies: `pnpm install`
 - Build: `pnpm build`
 - Run tests: `pnpm test`
 - Typecheck: `pnpm typecheck`
+- Verify: `pnpm verify`
 - Format: `pnpm format`
 - Biome check: `pnpm check`
 - Biome lint: `pnpm lint`
+- MCP harness regression check: `pnpm harness:mcp`
+- Update MCP harness baseline: `pnpm harness:mcp:update`
 
 ## Repository Structure
 
 - `packages/core/src/`: shared diagnosis logic, parsing, storage, prompt state, and eval logic
 - `packages/cli/src/`: CLI commands, shell integration, and terminal formatting
-- `packages/mcp/src/`: MCP server scaffold and future MCP integrations
+- `packages/mcp/src/`: frontend-focused MCP tools for agent-facing failure diagnosis
+- `harness/`: MCP benchmark harness, regression baseline, and verification scripts
+- `benchmarks/`: frontend failure log cases, expected signals, and generated reports
 - `test/`: unit and white-box integration tests
 - `e2e/`: black-box CLI tests
 
@@ -43,12 +48,13 @@
 
 ## Testing Expectations
 
-Before finishing non-trivial changes, run the relevant checks:
+Before finishing non-trivial changes, agents MUST run:
 
-- `pnpm check`
-- `pnpm lint`
-- `pnpm typecheck`
-- `pnpm test`
+- `pnpm verify`
+
+`pnpm verify` includes Biome checks, linting, typecheck, tests, and the MCP harness regression check.
+
+If a diagnosis strategy or MCP behavior intentionally changes the benchmark baseline, agents MUST explain why, run `pnpm harness:mcp:update`, then run `pnpm verify`.
 
 For CLI behavior changes, prefer adding or updating tests in `e2e/` when the behavior is user-visible.
 

package.json

@@ -25,7 +25,8 @@
     "release": "pnpm build && changeset publish",
     "test": "vitest run",
     "test:watch": "vitest",
-    "typecheck": "tsc --noEmit"
+    "typecheck": "tsc --noEmit",
+    "verify": "pnpm check && pnpm lint && pnpm typecheck && pnpm test && pnpm harness:mcp"
   },
   "lint-staged": {
     "*.{js,mjs,cjs,ts,json,md,yaml,yml}": "biome check --write --no-errors-on-unmatched"