greenrun-cli 0.1.4 → 0.1.6

package/dist/commands/init.js

@@ -138,10 +138,60 @@ function installClaudeMd() {
         console.log('  Created CLAUDE.md with Greenrun instructions');
     }
 }
+function installSettings() {
+    const settingsDir = join(process.cwd(), '.claude');
+    mkdirSync(settingsDir, { recursive: true });
+    const settingsPath = join(settingsDir, 'settings.local.json');
+    let existing = {};
+    if (existsSync(settingsPath)) {
+        try {
+            existing = JSON.parse(readFileSync(settingsPath, 'utf-8'));
+        }
+        catch {
+            // overwrite invalid JSON
+        }
+    }
+    const greenrunTools = [
+        'mcp__greenrun__list_projects',
+        'mcp__greenrun__get_project',
+        'mcp__greenrun__create_project',
+        'mcp__greenrun__list_pages',
+        'mcp__greenrun__create_page',
+        'mcp__greenrun__list_tests',
+        'mcp__greenrun__get_test',
+        'mcp__greenrun__create_test',
+        'mcp__greenrun__update_test',
+        'mcp__greenrun__start_run',
+        'mcp__greenrun__complete_run',
+        'mcp__greenrun__get_run',
+        'mcp__greenrun__list_runs',
+        'mcp__greenrun__sweep',
+    ];
+    const browserTools = [
+        'mcp__claude-in-chrome__tabs_context_mcp',
+        'mcp__claude-in-chrome__tabs_create_mcp',
+        'mcp__claude-in-chrome__navigate',
+        'mcp__claude-in-chrome__computer',
+        'mcp__claude-in-chrome__read_page',
+        'mcp__claude-in-chrome__find',
+        'mcp__claude-in-chrome__form_input',
+        'mcp__claude-in-chrome__javascript_tool',
+        'mcp__claude-in-chrome__get_page_text',
+        'mcp__claude-in-chrome__read_console_messages',
+        'mcp__claude-in-chrome__read_network_requests',
+    ];
+    const requiredTools = [...greenrunTools, ...browserTools];
+    existing.permissions = existing.permissions || {};
+    const currentAllow = existing.permissions.allow || [];
+    const merged = [...new Set([...currentAllow, ...requiredTools])];
+    existing.permissions.allow = merged;
+    writeFileSync(settingsPath, JSON.stringify(existing, null, 2) + '\n');
+    console.log('  Updated .claude/settings.local.json with tool permissions');
+}
 function installCommands() {
     const commandsDir = join(process.cwd(), '.claude', 'commands');
     mkdirSync(commandsDir, { recursive: true });
-    const commands = ['greenrun.md', 'greenrun-sweep.md'];
+    const commands = ['greenrun.md', 'greenrun-sweep.md', 'procedures.md'];
     for (const cmd of commands) {
         const src = join(TEMPLATES_DIR, 'commands', cmd);
         if (!existsSync(src)) {
@@ -156,6 +206,7 @@ function installCommands() {
 export function runUpdate() {
     console.log('\nGreenrun - Updating templates\n');
     installCommands();
+    installSettings();
     installClaudeMd();
     console.log('\nDone! Templates updated to latest version.\n');
 }
@@ -247,6 +298,7 @@ export async function runInit(args) {
     if (opts.commands) {
         installCommands();
     }
+    installSettings();
     console.log(`
 Done! Restart Claude Code to connect.
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "greenrun-cli",
-  "version": "0.1.4",
+  "version": "0.1.6",
   "description": "CLI and MCP server for Greenrun - browser test management for Claude Code",
   "type": "module",
   "main": "dist/server.js",

package/templates/commands/greenrun-sweep.md

@@ -36,41 +36,4 @@ Present the affected tests:
 
 ### 6. Offer to run
 
-Ask the user if they want to run the affected tests. If yes, execute them **in parallel** using the same approach as the `/greenrun` command:
-
-Use the project's `concurrency` setting (default: 5) to determine batch size. Split affected tests into batches and launch each batch simultaneously using the **Task tool** with `run_in_background: true`.
-
-For each test in a batch, launch a background agent with `max_turns: 30` and `model: "sonnet"`. Use this prompt:
-
-```
-You are executing a single Greenrun browser test. You have access to browser automation tools and Greenrun MCP tools.
-
-**Test: {test_name}** (ID: {test_id})
-
-Step 1: Call `get_test` with test_id "{test_id}" to get full instructions.
-Step 2: Call `start_run` with test_id "{test_id}" to begin - save the returned `run_id`.
-Step 3: Execute the test instructions using browser automation:
-   - Call `tabs_context_mcp` then create a new browser tab for this test
-   - Follow each instruction step exactly as written
-   - The instructions will tell you where to navigate and what to do
-   - Only take a screenshot when you need to verify a visual assertion — not for every navigation or click
-   - When reading page content, prefer `find` or `read_page` with `filter: "interactive"` over full DOM reads
-   - NEVER trigger JavaScript alerts, confirms, or prompts — they block the browser extension entirely. Before clicking delete buttons or other destructive actions, use `javascript_tool` to override: `window.alert = () => {}; window.confirm = () => true; window.prompt = () => null;`
-   - If browser tools stop responding (no result or timeout), assume a dialog is blocking — report the error and stop. Do not keep retrying.
-   - If you get stuck or a step fails, record the failure and move on — do not retry more than once
-Step 4: Call `complete_run` with:
-   - run_id: the run ID from step 2
-   - status: "passed" if all checks succeeded, "failed" if any check failed, "error" if execution was blocked
-   - result: a brief summary of what happened (include the failure reason if failed/error)
-Step 5: Close the browser tab you created to clean up.
-
-Return a single line summary: {test_name} | {status} | {result_summary}
-```
-
-Wait for each batch to complete before launching the next. After all tests finish, present a summary table:
-
-| Test | Pages | Tags | Status | Result |
-|------|-------|------|--------|--------|
-| Test name | Affected page URLs | tag1, tag2 | passed/failed/error | Brief summary |
-
-Include the total count: "X passed, Y failed, Z errors out of N tests"
+Ask the user if they want to run the affected tests. If yes, read `.claude/commands/procedures.md` for the agent prompt template and execution procedures. Follow those procedures to pre-fetch test details, launch agents in batches, collect results, and summarize.

package/templates/commands/greenrun.md

@@ -23,58 +23,6 @@ If no argument is given, run all active tests.
 
 If there are no matching active tests, tell the user and stop.
 
-### 3. Execute tests in parallel
+### 3. Execute tests
 
-Split the test list into batches of size `concurrency` (from the project settings).
-
-For each batch, launch all tests simultaneously using the **Task tool** with `run_in_background: true`. Each background agent receives a prompt containing everything it needs to execute one test independently:
-
-```
-For each test in the current batch, call the Task tool with:
-- subagent_type: "general-purpose"
-- run_in_background: true
-- max_turns: 30
-- model: "sonnet"
-- prompt: (see below)
-```
-
-The prompt for each background agent should be:
-
-```
-You are executing a single Greenrun browser test. You have access to browser automation tools and Greenrun MCP tools.
-
-**Test: {test_name}** (ID: {test_id})
-
-Step 1: Call `get_test` with test_id "{test_id}" to get full instructions.
-Step 2: Call `start_run` with test_id "{test_id}" to begin - save the returned `run_id`.
-Step 3: Execute the test instructions using browser automation:
-   - Call `tabs_context_mcp` then create a new browser tab for this test
-   - Follow each instruction step exactly as written
-   - The instructions will tell you where to navigate and what to do
-   - Only take a screenshot when you need to verify a visual assertion — not for every navigation or click
-   - When reading page content, prefer `find` or `read_page` with `filter: "interactive"` over full DOM reads
-   - NEVER trigger JavaScript alerts, confirms, or prompts — they block the browser extension entirely. Before clicking delete buttons or other destructive actions, use `javascript_tool` to override: `window.alert = () => {}; window.confirm = () => true; window.prompt = () => null;`
-   - If browser tools stop responding (no result or timeout), assume a dialog is blocking — report the error and stop. Do not keep retrying.
-   - If you get stuck or a step fails, record the failure and move on — do not retry more than once
-Step 4: Call `complete_run` with:
-   - run_id: the run ID from step 2
-   - status: "passed" if all checks succeeded, "failed" if any check failed, "error" if execution was blocked
-   - result: a brief summary of what happened (include the failure reason if failed/error)
-Step 5: Close the browser tab you created to clean up.
-
-Return a single line summary: {test_name} | {status} | {result_summary}
-```
-
-After launching all agents in a batch, wait for them all to complete (use `TaskOutput` to collect results) before launching the next batch.
-
-### 4. Summarize results
-
-After all batches complete, collect results from all background agents and present a summary table:
-
-| Test | Pages | Tags | Status | Result |
-|------|-------|------|--------|--------|
-| Test name | /login, /dashboard | smoke, auth | passed/failed/error | Brief summary |
-
-Include the total count: "X passed, Y failed, Z errors out of N tests"
-
-If any tests failed, highlight what went wrong and suggest next steps.
+Read `.claude/commands/procedures.md` for the agent prompt template and execution procedures. Follow those procedures to pre-fetch test details, launch agents in batches, collect results, and summarize.

package/templates/commands/procedures.md

@@ -0,0 +1,64 @@
+Shared procedures for executing Greenrun browser tests in parallel. Referenced by `/greenrun` and `/greenrun-sweep`.
+
+## Pre-fetch
+
+Before launching agents, call `get_test` for ALL tests **in parallel** to get full instructions. Then call `start_run` for ALL tests **in parallel** to get run IDs.
+
+## Launch agents
+
+Split tests into batches of size `concurrency` (from project settings, default: 5).
+
+For each batch, launch all tests simultaneously using the **Task tool** with `run_in_background: true`:
+
+```
+For each test in the current batch, call the Task tool with:
+- subagent_type: "general-purpose"
+- run_in_background: true
+- max_turns: 25
+- model: "haiku"
+- prompt: (see agent prompt below)
+```
+
+### Agent prompt
+
+```
+Execute a Greenrun browser test. Run ID: {run_id}
+
+**Test: {test_name}**
+
+## Instructions
+{paste the full test instructions from get_test here}
+
+## Setup
+1. Call `tabs_context_mcp`, then `tabs_create_mcp` to create YOUR tab. Use ONLY this tabId — other tabs belong to parallel tests.
+2. Navigate to the first URL. Run `javascript_tool`: `window.location.pathname`. If it returns `/login`, call `complete_run` with status "error", result "Not authenticated", then `window.close()` and stop.
+
+## Execution rules
+- Verify assertions with `screenshot` after actions that change the page. Do NOT use `read_page` for verification.
+- Use `find` to locate elements, then `ref` parameter on `computer` tool or `form_input` to interact.
+- Navigate with absolute URLs via `navigate` — don't click nav links.
+- Before destructive buttons: `window.alert = () => {}; window.confirm = () => true; window.prompt = () => null;`
+- On failure or timeout, retry ONCE then move on. Max 35 tool calls total.
+
+## Finish
+Call `complete_run` with run_id "{run_id}", status ("passed"/"failed"/"error"), and a brief result summary.
+Then run `javascript_tool`: `window.close()`.
+
+Return: {test_name} | {status} | {result_summary}
+```
+
+## Collect results
+
+After launching all agents in a batch, wait for them all to complete (use `TaskOutput`) before launching the next batch.
+
+## Summarize
+
+After all batches complete, present a summary table:
+
+| Test | Pages | Tags | Status | Result |
+|------|-------|------|--------|--------|
+| Test name | /login, /dashboard | smoke, auth | passed/failed/error | Brief summary |
+
+Include the total count: "X passed, Y failed, Z errors out of N tests"
+
+If any tests failed, highlight what went wrong and suggest next steps.