feat: add tools.playwright.mode (mcp/cli) and enable cli mode in smoke-copilot

.github/workflows/smoke-copilot.md

@@ -44,6 +44,7 @@ tools:
     trusted-users:
       - pelikhan
   playwright:
+    mode: cli
   web-fetch:
   cli-proxy: true
 runtimes:
@@ -142,7 +143,7 @@ Talk like a caveman in all your responses and outputs. Use short, broken sentenc
 
 This workflow uses `cli-proxy: true`. The following MCP servers are **NOT available as MCP tools** — they are mounted exclusively as **shell CLI commands** (see `<mcp-clis>` section above). You **must** use them via the `bash` tool:
 
-- **`playwright`** — use `playwright <tool> [--param value...]` in bash (e.g. `playwright browser_navigate --url ...`)
+- **`playwright`** — installed as `@playwright/cli`, use `playwright-cli <command>` in bash (e.g. `playwright-cli open https://github.com`, `playwright-cli screenshot`)
 - **`serena`** — use `serena <tool> [--param value...]` in bash (e.g. `serena activate_project --path ...`)
 - **`agenticworkflows`** — use `agenticworkflows <tool> [--param value...]` in bash
 - **`safeoutputs`** — use `safeoutputs <tool> [--param value...]` in bash (e.g. `safeoutputs add_comment --body "..."`)
@@ -161,7 +162,7 @@ These are **not** MCP protocol tools — they are bash executables. Call them wi
 3. **Serena CLI Testing**: 
    - Use bash to run `serena activate_project --path ${{ github.workspace }}` to initialize the workspace and verify it succeeds (do NOT use bash to run go commands - use the serena CLI only)
    - After initialization, use bash to run `serena find_symbol --name_path <symbol>` to search for symbols and verify that at least 3 symbols are found in the results
-4. **Playwright CLI Testing**: Use bash to run `playwright browser_navigate --url https://github.com` to navigate to <https://github.com>, then `playwright browser_snapshot` to capture the page and verify the title contains "GitHub" (do NOT try to install playwright - use the `playwright` CLI command via bash only)
+4. **Playwright CLI Testing**: Use bash to run `playwright-cli open https://github.com` to navigate to <https://github.com>, then `playwright-cli screenshot` to take a screenshot and verify that the output indicates a successful navigation to "GitHub" (do NOT try to install playwright - use the `playwright-cli` command via bash only)
 5. **Web Fetch Testing**: Use the web-fetch tool to fetch https://github.com and verify the response contains "GitHub" (do NOT use bash or playwright for this test - use the web-fetch tool directly)
 6. **File Writing Testing**: Create a test file `/tmp/gh-aw/agent/smoke-test-copilot-${{ github.run_id }}.txt` with content "Smoke test passed for Copilot at $(date)" (create the directory if it doesn't exist)
 7. **Bash Tool Testing**: Execute bash commands to verify file creation was successful (use `cat` to read the file back)

pkg/constants/version_constants.go

@@ -91,6 +91,10 @@ const MCPGIntegrityReactionsMinVersion Version = "v0.2.18"
 // DefaultPlaywrightMCPVersion is the default version of the @playwright/mcp package
 const DefaultPlaywrightMCPVersion Version = "0.0.73"
 
+// DefaultPlaywrightCLIVersion is the default version of the @playwright/cli package
+// Used when tools.playwright.mode is "cli" to install the CLI tool instead of the MCP server.
+const DefaultPlaywrightCLIVersion Version = "0.1.11"
+
 // DefaultPlaywrightBrowserVersion is the default version of the Playwright browser Docker image
 const DefaultPlaywrightBrowserVersion Version = "v1.59.1"
 

pkg/parser/schemas/main_workflow_schema.json

@@ -4054,6 +4054,11 @@
                   "items": {
                     "type": "string"
                   }
+                },
+                "mode": {
+                  "type": "string",
+                  "description": "Integration mode: 'mcp' (default) runs a Docker-based MCP server; 'cli' installs @playwright/cli via npm for token-efficient CLI invocations.",
+                  "enum": ["mcp", "cli"]
                 }
               },
               "additionalProperties": false

pkg/workflow/compiler_yaml_main_job.go

@@ -333,6 +333,15 @@ func (c *Compiler) generateEngineInstallAndPreAgentSteps(yaml *strings.Builder,
 		}
 	}
 
+	// Add Playwright CLI install steps when playwright is configured in CLI mode.
+	// These run after Node.js is available (set up by the engine install steps above).
+	for _, step := range generatePlaywrightCLIInstallSteps(data) {
+		for _, line := range step {
+			yaml.WriteString(line)
+			yaml.WriteByte('\n')
+		}
+	}
+
 	// GH_AW_SAFE_OUTPUTS is now set at job level, no setup step needed
 
 	// Mint the GitHub MCP App token directly in the agent job.

pkg/workflow/docker.go

@@ -34,11 +34,14 @@ func collectDockerImages(tools map[string]any, workflowData *WorkflowData, actio
 	}
 
 	// Check for Playwright tool (uses Docker image - no version tag, only one image)
+	// Only in MCP mode; CLI mode installs @playwright/cli via npm instead.
 	if _, hasPlaywright := tools["playwright"]; hasPlaywright {
-		image := "mcr.microsoft.com/playwright/mcp"
-		if !imageSet[image] {
-			images = append(images, image)
-			imageSet[image] = true
+		if !isPlaywrightCLIMode(tools) {
+			image := "mcr.microsoft.com/playwright/mcp"
+			if !imageSet[image] {
+				images = append(images, image)
+				imageSet[image] = true
+			}
 		}
 	}
 

pkg/workflow/mcp_cli_mount.go

@@ -62,7 +62,13 @@ func getMCPCLIServerNames(data *WorkflowData) []string {
 			// Only include tools that have MCP servers (skip bash, web-fetch, web-search, edit, cache-memory, etc.)
 			// Note: "github" is excluded — it is handled differently and should not be CLI-mounted.
 			switch toolName {
-			case "playwright", "qmd":
+			case "playwright":
+				// In CLI mode, playwright is installed as @playwright/cli via npm and is NOT
+				// an MCP server, so it must not appear in the CLI-mounted servers list.
+				if !isPlaywrightCLIMode(data.Tools) {
+					servers = append(servers, toolName)
+				}
+			case "qmd":
 				servers = append(servers, toolName)
 			case "agentic-workflows":
 				// The gateway and manifest use "agenticworkflows" (no hyphen) as the server ID.

pkg/workflow/mcp_setup_generator.go

@@ -140,6 +140,11 @@ func collectMCPTools(workflowData *WorkflowData) []string {
 			continue
 		}
 		if toolName == "github" || toolName == "playwright" || toolName == "cache-memory" || toolName == "agentic-workflows" {
+			// Playwright in CLI mode is not an MCP server; skip it here.
+			if toolName == "playwright" && isPlaywrightCLIMode(workflowData.Tools) {
+				mcpSetupGeneratorLog.Print("Skipping playwright MCP registration: tools.playwright.mode is cli")
+				continue
+			}
 			mcpTools = append(mcpTools, toolName)
 			continue
 		}

pkg/workflow/playwright_cli.go

@@ -0,0 +1,92 @@
+package workflow
+
+// Package workflow provides support for Playwright CLI mode.
+//
+// # Playwright CLI Mode
+//
+// When tools.playwright.mode is set to "cli", the compiler installs the
+// @playwright/cli npm package instead of launching the Docker-based MCP server.
+// This is a token-efficient alternative for coding agents that prefer CLI-based
+// workflows over MCP: CLI invocations avoid loading large tool schemas and verbose
+// accessibility trees into the model context.
+//
+// See https://github.com/microsoft/playwright-cli for details.
+//
+// In CLI mode:
+//   - The mcr.microsoft.com/playwright/mcp Docker image is NOT pulled.
+//   - playwright is NOT registered as an MCP server in the gateway config.
+//   - @playwright/cli is installed via npm (global) before the agent runs.
+//   - playwright-cli install --skills installs agent skill files so the coding
+//     agent can discover and use the available playwright-cli commands.
+//   - The agent uses `playwright-cli <command>` directly via bash.
+//
+// Example workflow frontmatter:
+//
+//	tools:
+//	  playwright:
+//	    mode: cli
+
+import (
+	"github.com/github/gh-aw/pkg/constants"
+	"github.com/github/gh-aw/pkg/logger"
+)
+
+var playwrightCLILog = logger.New("workflow:playwright_cli")
+
+// isPlaywrightCLIMode returns true when the playwright tool in the given tools map
+// is configured with mode: cli.
+func isPlaywrightCLIMode(tools map[string]any) bool {
+	playwrightTool, ok := tools["playwright"]
+	if !ok || playwrightTool == false {
+		return false
+	}
+	config := parsePlaywrightTool(playwrightTool)
+	return config != nil && config.IsCLIMode()
+}
+
+// generatePlaywrightCLIInstallSteps returns npm install steps for @playwright/cli
+// when playwright is configured in CLI mode. Returns nil if playwright is in MCP mode.
+//
+// Node.js setup is intentionally omitted here because all supported engines
+// (copilot, claude, codex, gemini) include a Node.js setup step in their own
+// installation steps, which run before this function is called.
+func generatePlaywrightCLIInstallSteps(workflowData *WorkflowData) []GitHubActionStep {
+	if !isPlaywrightCLIMode(workflowData.Tools) {
+		return nil
+	}
+
+	playwrightCLILog.Print("Generating @playwright/cli install steps (CLI mode)")
+
+	version := string(constants.DefaultPlaywrightCLIVersion)
+	// Use version override from playwright config if provided
+	if playwrightTool, ok := workflowData.Tools["playwright"]; ok {
+		config := parsePlaywrightTool(playwrightTool)
+		if config != nil && config.Version != "" {
+			version = config.Version
+			playwrightCLILog.Printf("Using playwright CLI version from config: %s", version)
+		}
+	}
+
+	// Install @playwright/cli globally.
+	// runInstallScripts is true because playwright's postinstall script may set up
+	// browser binaries needed at runtime.
+	steps := GenerateNpmInstallStepsWithScope(
+		"@playwright/cli",
+		version,
+		"Install Playwright CLI",
+		"playwright-cli",
+		false, // Node.js is already set up by the engine install steps
+		true,  // Global install so playwright-cli is on PATH
+		true,  // Allow install scripts for browser setup
+	)
+
+	// Install playwright-cli skills so the coding agent can discover available commands.
+	installSkillsStep := GitHubActionStep{
+		"      - name: Install Playwright CLI skills",
+		"        run: playwright-cli install --skills",
+	}
+	steps = append(steps, installSkillsStep)
+
+	playwrightCLILog.Printf("Generated %d Playwright CLI install steps", len(steps))
+	return steps
+}

pkg/workflow/testdata/TestWasmGolden_CompileFixtures/basic-copilot.golden

@@ -53,8 +53,8 @@ jobs:
           GH_AW_INFO_ENGINE_ID: "copilot"
           GH_AW_INFO_ENGINE_NAME: "GitHub Copilot CLI"
           GH_AW_INFO_MODEL: ${{ vars.GH_AW_MODEL_AGENT_COPILOT || 'claude-sonnet-4.6' }}
-          GH_AW_INFO_VERSION: "1.0.39"
-          GH_AW_INFO_AGENT_VERSION: "1.0.39"
+          GH_AW_INFO_VERSION: "1.0.40"
+          GH_AW_INFO_AGENT_VERSION: "1.0.40"
           GH_AW_INFO_WORKFLOW_NAME: "basic-copilot-test"
           GH_AW_INFO_EXPERIMENTAL: "false"
           GH_AW_INFO_SUPPORTS_TOOLS_ALLOWLIST: "true"
@@ -310,7 +310,7 @@ jobs:
             const { main } = require('${{ runner.temp }}/gh-aw/actions/checkout_pr_branch.cjs');
             await main();
       - name: Install GitHub Copilot CLI
-        run: bash "${RUNNER_TEMP}/gh-aw/actions/install_copilot_cli.sh" 1.0.39
+        run: bash "${RUNNER_TEMP}/gh-aw/actions/install_copilot_cli.sh" 1.0.40
         env:
           GH_HOST: github.com
       - name: Install AWF binary