novahiz 0.2.2 → 0.2.5

package/install/bootstrap.mjs

@@ -63,7 +63,6 @@ function detectShell() {
 
 function generateOpenCodeJson(configDir, NovahizHome) {
   const skillsDir = join(configDir, "skills");
-  const bundledDir = join(NovahizHome, "bundled-skills");
   const agentsSkillsDir = join(homedir(), ".config", ".agents", "skills");
 
   const config = {
@@ -90,18 +89,13 @@ function generateOpenCodeJson(configDir, NovahizHome) {
         command: ["npx", "@playwright/mcp@latest", "--browser=msedge"],
         enabled: true,
       },
-      expo: {
-        type: "remote",
-        url: "https://mcp.expo.dev/mcp",
-        enabled: true,
-      },
     },
     skills: {
       paths: [skillsDir],
     },
     plugin: [
-      "@mohak34/opencode-notifier@0.2.8",
-      "@tarquinen/opencode-dcp@latest",
+      "@mohak34/opencode-notifier@0.3.0",
+      "@tarquinen/opencode-dcp@3.2.0",
       join(NovahizHome, "adapters", "opencode", "novahiz.ts"),
     ],
     compaction: {
@@ -112,11 +106,6 @@ function generateOpenCodeJson(configDir, NovahizHome) {
     shell: detectShell(),
   };
 
-  // Add bundled-skills if it exists
-  if (existsSync(bundledDir)) {
-    config.skills.paths.push(bundledDir);
-  }
-
   // Add .agents/skills if it exists
   if (existsSync(agentsSkillsDir)) {
     config.skills.paths.push(agentsSkillsDir);
@@ -151,8 +140,8 @@ async function main() {
   // 3. Install opencode globally if not present
   if (!which("opencode")) {
     log("Installing opencode globally...");
-    if (!run("npm", ["install", "-g", "opencode"])) {
-      error("Failed to install opencode. Try: npm install -g opencode");
+    if (!run("npm", ["install", "-g", "opencode-ai"])) {
+      error("Failed to install opencode. Try: npm install -g opencode-ai");
       process.exit(1);
     }
     log("opencode installed");
@@ -179,8 +168,9 @@ async function main() {
   log("Installing MCP servers...");
   const mcpServers = [
     { pkg: "@upstash/context7-mcp", bin: "context7-mcp" },
-    { pkg: "@anthropic-ai/narsil-mcp", bin: "narsil-mcp" },
-    { pkg: "@anthropic-ai/mcp-cron", bin: "mcp-cron" },
+    { pkg: "narsil-mcp", bin: "narsil-mcp" },
+    { pkg: "mcp-cron", bin: "mcp-cron" },
+    { pkg: "security-mcp", bin: "security-mcp" },
   ];
 
   for (const server of mcpServers) {
@@ -200,8 +190,8 @@ async function main() {
   log("");
   log("Installing opencode plugins...");
   const plugins = [
-    "@mohak34/opencode-notifier@0.2.8",
-    "@tarquinen/opencode-dcp@latest",
+    "@mohak34/opencode-notifier@0.3.0",
+    "@tarquinen/opencode-dcp@3.2.0",
   ];
 
   for (const plugin of plugins) {

package/install/install.mjs

@@ -29,7 +29,6 @@ const CORE_ITEMS = [
   "mcp",
   "adapters",
   "skills",
-  "bundled-skills",
   "docs",
   "package.json",
   "tsconfig.json",
@@ -42,10 +41,12 @@ const CORE_ITEMS = [
 function defaultConfig(skillsDir) {
   return {
     dbPath: "novahiz.sqlite",
-    skillRoots: ["./skills", "./bundled-skills"],
+    skillRoots: ["./skills"],
     gate: {
       enabled: true,
       mode: "block",
+      // Kept for schema compatibility only — the kill-switch name is hardcoded
+      // to NOVAHIZ_GATE in the CLI, MCP gate, and plugin (see src/spec.ts).
       envEscape: "NOVAHIZ_GATE",
       tools: ["edit", "write", "patch", "apply_patch", "bash", "shell"]
     },
@@ -85,16 +86,20 @@ async function main() {
 
   // Check for opencode and auto-install if missing
   if (!which("opencode")) {
-    note("opencode not detected. Global installation...");
-    const installResult = spawnSync("npm", ["install", "-g", "opencode"], {
-      encoding: "utf8",
-      stdio: "inherit"
-    });
-    if (installResult.status !== 0) {
-      process.stderr.write("Failed to install opencode. Try: npm install -g opencode\n");
-      process.exit(1);
+    if (dryRun) {
+      note("opencode not detected. Would install globally (npm install -g opencode-ai).");
+    } else {
+      note("opencode not detected. Global installation...");
+      const installResult = spawnSync("npm", ["install", "-g", "opencode-ai"], {
+        encoding: "utf8",
+        stdio: "inherit"
+      });
+      if (installResult.status !== 0) {
+        process.stderr.write("Failed to install opencode. Try: npm install -g opencode-ai\n");
+        process.exit(1);
+      }
+      note("opencode installed successfully.");
     }
-    note("opencode installed successfully.");
   } else {
     note("opencode detected.");
   }
@@ -195,19 +200,6 @@ async function main() {
     } else {
       note(`No skills folder found at ${skillsSource}`);
     }
-
-    // Copy bundled-skills
-    const bundledSource = existsSync(join(home, "bundled-skills")) ? join(home, "bundled-skills") : join(root, "bundled-skills");
-    if (existsSync(bundledSource)) {
-      const bundledTarget = join(home, "bundled-skills");
-      note(`Installing bundled-skills in ${bundledTarget}`);
-      if (!dryRun) {
-        const result = copyInto(bundledSource, bundledTarget, true);
-        created.push(...result.created);
-        backups.push(...result.backups);
-        note(`  ${result.total} files copied`);
-      }
-    }
   }
 
   const pluginSource = join(home, "adapters", "opencode", "novahiz.ts");
@@ -319,8 +311,9 @@ async function main() {
   // Install MCP servers globally
   const mcpServers = [
     { pkg: "@upstash/context7-mcp", bin: "context7-mcp", name: "context7" },
-    { pkg: "@anthropic-ai/narsil-mcp", bin: "narsil-mcp", name: "narsil" },
-    { pkg: "@anthropic-ai/mcp-cron", bin: "mcp-cron", name: "cron" },
+    { pkg: "narsil-mcp", bin: "narsil-mcp", name: "narsil" },
+    { pkg: "mcp-cron", bin: "mcp-cron", name: "cron" },
+    { pkg: "security-mcp", bin: "security-mcp", name: "security" },
   ];
 
   if (!dryRun) {
@@ -349,8 +342,8 @@ async function main() {
 
   // Install opencode plugins globally
   const plugins = [
-    "@mohak34/opencode-notifier@0.2.8",
-    "@tarquinen/opencode-dcp@latest",
+    "@mohak34/opencode-notifier@0.3.0",
+    "@tarquinen/opencode-dcp@3.2.0",
   ];
 
   if (!dryRun) {
@@ -374,7 +367,6 @@ async function main() {
     const configPath = join(configDir, "opencode.jsonc");
     if (!existsSync(configPath)) {
       note(`\nCreating ${configPath}`);
-      const bundledDir = join(home, "bundled-skills");
       const agentsSkillsDir = join(homedir(), ".config", ".agents", "skills");
       
       const openCodeConfig = {
@@ -400,19 +392,14 @@ async function main() {
             "type": "local",
             "command": ["npx", "@playwright/mcp@latest", "--browser=msedge"],
             "enabled": true
-          },
-          "expo": {
-            "type": "remote",
-            "url": "https://mcp.expo.dev/mcp",
-            "enabled": true
           }
         },
         "skills": {
           "paths": [skillsDir]
         },
         "plugin": [
-          "@mohak34/opencode-notifier@0.2.8",
-          "@tarquinen/opencode-dcp@latest",
+          "@mohak34/opencode-notifier@0.3.0",
+          "@tarquinen/opencode-dcp@3.2.0",
           join(home, "adapters", "opencode", "novahiz.ts")
         ],
         "compaction": {
@@ -423,11 +410,6 @@ async function main() {
         "shell": process.platform === "win32" ? "pwsh" : "bash"
       };
 
-      // Add bundled-skills if it exists
-      if (existsSync(bundledDir)) {
-        openCodeConfig.skills.paths.push(bundledDir);
-      }
-
       // Add .agents/skills if it exists
       if (existsSync(agentsSkillsDir)) {
         openCodeConfig.skills.paths.push(agentsSkillsDir);

package/mcp/novahiz-tools/index.mjs

@@ -11,7 +11,7 @@ import { rankSkills } from "../../src/relevance.ts";
 import { openDb } from "../../src/db.ts";
 import { enabledProviders } from "../../src/providers.ts";
 import { checkDependencies } from "../../src/deps.ts";
-import { activeTask, addTodos, amendTodo, blockTodo, buildWorkPackets, completeTodo, createTask, dropTodo, getTask, getTodo, insertTodo, ledgerSummary, listTodos, recordTodoDone, reorderTodos, resume, reviewDue, reviewTask, revisionSignals, startTodo } from "../../src/ledger.ts";
+import { activeTask, addTodos, amendTodo, blockTodo, buildWorkPackets, completeTodo, createTask, dropTask, dropTodo, getTask, getTodo, insertTodo, ledgerSummary, listTodos, recordTodoDone, reorderTodos, resume, reviewDue, reviewTask, revisionSignals, startTodo } from "../../src/ledger.ts";
 import { DEFAULT_LIMIT_CHARS, DEFAULT_LIMIT_LINES, ensureMemoryRoot, getSlot, listSlots, memoryRoot, parseSlotInput, rebuildIndex, writeEntry } from "../../src/memory.ts";
 
 const SUPPORTED_PROTOCOLS = ["2024-11-05", "2025-06-18"];
@@ -63,7 +63,8 @@ const TOOLS = [
         filePath: { type: "string", description: "Alias of file. Some harnesses rename the parameter when they surface the tool." },
         tool: { type: "string", description: "edit, write or patch." },
         content: { type: "string", description: "The edited content, used for content-aware rules." },
-        categories: { type: "array", items: { type: "string" } },
+        prompt: { type: "string", description: "Optional prompt used to auto-classify when categories is omitted or empty." },
+        categories: { type: "array", items: { type: "string" }, description: "Category ids. When omitted or empty, inferred from prompt, content, or file path." },
         loaded: { type: "array", items: { type: "string" } }
       }
     }
@@ -118,7 +119,7 @@ const TOOLS = [
           description: "What to do with the ledger."
         },
         title: { type: "string", description: "Task title for action new." },
-        id: { type: "string", description: "Task id (new) or todo id (start, done, block)." },
+        id: { type: "string", description: "Task id (new) or todo/task id (start, done, block, drop)." },
         task: { type: "string", description: "Task id. Defaults to the active task." },
         session: { type: "string", description: "Session id used to scope the active task." },
         label: { type: "string", description: "Todo label for action todo." },
@@ -329,12 +330,34 @@ function callTool(name, args) {
     if (args?.tool !== undefined && typeof args.tool !== "string") {
       throw new Error("Invalid params: tool must be a string");
     }
+    if (args?.prompt !== undefined && typeof args.prompt !== "string") {
+      throw new Error("Invalid params: prompt must be a string");
+    }
+    if (typeof args?.prompt === "string" && args.prompt.length > MAX_PROMPT_LEN) {
+      throw new Error(`Invalid params: prompt exceeds ${MAX_PROMPT_LEN} characters`);
+    }
+    // Auto-classify when categories is omitted or empty: seed from prompt,
+    // fall back to content, then file path. Explicit categories always win.
+    let categories = args?.categories ? args.categories.map(String) : [];
+    if (categories.length === 0) {
+      const seed =
+        (typeof args?.prompt === "string" && args.prompt.length > 0 ? args.prompt : "") ||
+        (typeof args?.content === "string" && args.content.length > 0 ? args.content : "") ||
+        file;
+      try {
+        categories = classify(spec, seed).categories.map((entry) => entry.id);
+      } catch {
+        // Fail closed on classify errors: keep empty categories so
+        // evaluateGate still runs path/content rules instead of crashing.
+        categories = [];
+      }
+    }
     const index = loadInstalledSkills(spec);
     const result = evaluateGate({
       tool: String(args?.tool ?? "edit"),
       filePath: file,
       content: typeof args?.content === "string" ? args.content : "",
-      categories: args?.categories ? args.categories.map(String) : [],
+      categories,
       loadedSkills: args?.loaded ? args.loaded.map(String) : [],
       installedSkills: index.skills,
       installedIndexAvailable: index.available,
@@ -472,8 +495,14 @@ function callTool(name, args) {
         const position = raw === undefined || raw === "" ? "end" : /^\d+$/.test(String(raw)) ? Number(raw) : String(raw);
         return toolResult(insertTodo(db, taskId, normalizeTodo(args), position));
       }
-      if (action === "drop") return toolResult(dropTodo(db, String(args?.id ?? ""), args?.reason ? String(args.reason) : ""));
-      if (action === "reorder") {
+        if (action === "drop") {
+          const dropId = String(args?.id ?? "");
+          const reason = args?.reason ? String(args.reason) : "";
+          if (getTask(db, dropId)) return toolResult({ task: dropTask(db, dropId, reason) });
+          if (getTodo(db, dropId)) return toolResult({ todo: dropTodo(db, dropId, reason) });
+          throw new Error(`unknown id: ${dropId} (expected a task id or todo id)`);
+        }
+        if (action === "reorder") {
         const taskId = args?.task ? String(args.task) : activeTask(db, session)?.id;
         if (!taskId) return toolResult("no active task", true);
         const order = Array.isArray(args?.order) ? args.order.map(String) : [];

package/novahiz.config.example.json

@@ -2,11 +2,7 @@
   "dbPath": "novahiz.sqlite",
   "skillRoots": [
     "./skills",
-    "~/.config/opencode/skills",
-    "~/.agents/skills",
-    "~/.config/.agents/skills",
-    "~/.config/marketingskills/skills",
-    "~/.config/humanizer"
+    "~/.config/opencode/skills"
   ],
   "gate": {
     "enabled": true,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "novahiz",
-  "version": "0.2.2",
+  "version": "0.2.5",
   "description": "Enforce AI agent skills before every file edit: prompt classifier, skill catalog, and pre-edit gate.",
   "type": "module",
   "license": "Apache-2.0",
@@ -27,11 +27,11 @@
     "access": "public"
   },
   "bin": {
-    "novahiz": "./bin/novahiz.mjs",
-    "novahiz-bootstrap": "./install/bootstrap.mjs",
-    "novahiz-install": "./install/install.mjs",
-    "novahiz-uninstall": "./install/uninstall.mjs",
-    "novahiz-mcp": "./mcp/novahiz-tools/index.mjs"
+    "novahiz": "bin/novahiz.mjs",
+    "novahiz-bootstrap": "install/bootstrap.mjs",
+    "novahiz-install": "install/install.mjs",
+    "novahiz-uninstall": "install/uninstall.mjs",
+    "novahiz-mcp": "mcp/novahiz-tools/index.mjs"
   },
   "engines": {
     "node": ">=22.18.0"
@@ -59,7 +59,6 @@
     "install",
     "mcp",
     "skills",
-    "bundled-skills",
     "docs",
     "README.md",
     "NOTICE.md",

package/skills/apple-ui-audit/SKILL.md

@@ -0,0 +1,95 @@
+---
+name: "apple-ui-audit"
+description: "Audit iOS/macOS/watchOS/visionOS interfaces against Apple HIG and Liquid Glass (iOS 26): navigation hierarchy, materials, controls, accessibility settings, screenshots vs binary. Use when reviewing an Apple-platform mockup or build for HIG compliance, contrast, tap targets, reduced transparency, or App Store UI review readiness."
+license: "Apache-2.0"
+metadata:
+  author: Novahiz
+  organization: Novahiz
+  version: "1.0.0"
+  date: September 2026
+---
+
+# apple-ui-audit
+
+Structured review of an Apple-platform interface against the Human Interface Guidelines and current Liquid Glass guidance (iOS 26 / 2025-2026). The report lists findings with severity and a fix direction.
+
+## Scope
+
+SwiftUI, UIKit, AppKit screens; mockups intended for App Store submission; cross-platform (iOS/iPadOS/macOS/tvOS/watchOS/visionOS) consistency checks. Web UI that only *looks* iOS-like is out of scope (use `ui-slop-remover` / `ui-craft-rules`).
+
+## 1. Navigation and hierarchy
+
+- Content layer vs navigation layer clearly separated. Tab bars, sidebars, toolbars sit in the functional layer above content.
+- One obvious primary path per screen; no competing equal-weight actions.
+- Standard back / dismiss patterns; no custom chrome that fights system gestures.
+- Search and navigation work the same way across iPhone, iPad, Mac where the HIG expects it.
+- Sidebar / inspector safe areas: content peeks correctly; nothing clipped under Liquid Glass.
+
+## 2. Liquid Glass (when targeting latest OS)
+
+Primary references: Apple "Adopting Liquid Glass", HIG Materials.
+
+| Check | Fail signal |
+|---|---|
+| Use system components first | Custom opaque bars fighting the system material |
+| Glass on controls/navigation, not content layer | Decorative glass cards in body content |
+| Sparingly on custom controls | Every cell frosted |
+| Regular vs clear variant | Clear over bright text without dimming |
+| Reduce Transparency path | Broken layout when the setting strips blur |
+| Scroll edge effect | Unreadable controls over scrolling content |
+| Color on controls | Hard-coded brand color with no light/dark/increase-contrast variants |
+
+Rebuild with latest SDK, run on latest OS, then judge. App Review 2.3 rejects marketing shots whose glass the binary does not ship.
+
+## 3. Controls and touch
+
+- Minimum tap targets per HIG (44×44 pt default; follow current platform tables).
+- Standard control sizes: do not hard-code metrics that the system now updates.
+- Toolbars: no empty spacer hacks; hide the whole item, not an empty shell.
+- Tab bar auto-minimize: intentional opt-in, tested both scroll directions.
+- Focus engine / keyboard / remote (tvOS) paths for every actionable control.
+
+## 4. Color, type, and accessibility
+
+- System colors or custom colors with light, dark, and increased-contrast variants.
+- Dynamic Type: no fixed-height labels that clip at accessibility sizes.
+- Test matrix: Reduce Transparency, Reduce Motion, Increase Contrast, Bold Text, larger text, VoiceOver, Switch Control.
+- Liquid Glass fallback under those settings must remain legible and navigable.
+- Contrast on frosted surfaces: measure. Blur alone does not guarantee a ratio.
+- State must remain readable without color.
+
+## 5. Motion and materials
+
+- Prefer system transitions; custom morphs respect Reduce Motion.
+- `GlassEffectContainer` (where applicable) for combined custom glass performance.
+- Profile scroll hitching and modal latency on representative devices.
+
+## 6. Store readiness (UI-adjacent)
+
+- Screenshots match the submitted build (current chrome, not last year's).
+- Privacy manifest present when required; AI features disclosed where guidelines require.
+- No mock-only UI in the listing.
+
+## Report format
+
+```
+## Apple UI audit
+Verdict: PASS | PASS_WITH_FIXES | FAIL
+
+### Blockers
+- [HIGH] screen:path : finding
+  HIG / doc: ...
+  Fix: ...
+
+### Should fix
+- ...
+
+### Notes / accepted risk
+- ...
+```
+
+Each finding cites a screen path and a guideline source (HIG section or Apple doc title). No finding without a location.
+
+## Sources
+
+Apple Human Interface Guidelines; Adopting Liquid Glass; Materials (Liquid Glass + standard materials); App Review guidelines (2.3 screenshots, privacy). Original Novahiz checklist synthesis.

package/skills/apple-ui-audit/references/screen-walkthrough.md

@@ -0,0 +1,29 @@
+# Screen walkthrough template
+
+Copy into the audit report. One block per screen (home, detail, search, modal, settings, empty, error).
+
+## Screen: <name>
+
+| Check | Pass? | Evidence |
+|---|---|---|
+| Nav vs content layers separated | | |
+| Primary action obvious | | |
+| Liquid Glass only on chrome/controls | | |
+| Reduce Transparency still legible | | |
+| Reduce Motion: no essential motion | | |
+| Tap targets ≥ platform minimum | | |
+| Dynamic Type no clip | | |
+| VoiceOver labels on controls | | |
+| Light / dark / increase contrast | | |
+| Empty + error states designed | | |
+| Screenshot matches current build | | |
+
+Findings:
+
+- [severity] what / where / fix
+
+Matrix (run at least one device per class):
+
+- Current OS plus one older supported OS
+- Light and dark
+- Accessibility settings toggled on as listed above

package/skills/browser-session/SKILL.md

@@ -0,0 +1,61 @@
+---
+name: "browser-session"
+description: "Drive a persistent Microsoft Edge browser session for research and verification: navigate, snapshot, click, fill, console and network inspection, keep login state across runs. Use when automating a web page, checking a UI in the real browser, or reusing an authenticated session."
+license: "Apache-2.0"
+metadata:
+  author: Novahiz
+  organization: Novahiz
+  version: "1.0.0"
+  date: September 2026
+---
+
+# browser-session
+
+All browser work goes through the Playwright MCP tools (`playwright_browser_*`). The profile is persistent at `C:\Users\hiz\.opencode\playwright-profile` so cookies and logins survive restarts.
+
+## Hard rules
+
+1. **Never launch Chrome, Chromium, or any manual browser process.** Use only the MCP tools, which run Microsoft Edge (`--browser=msedge`). If Edge is missing, fall back to another non-Chromium engine via the same MCP server.
+2. Never disable or redirect `--user-data-dir`. Never purge the profile folder without the user's explicit consent.
+3. Never open an ephemeral context when the task needs saved login state.
+4. If the profile looks corrupted, stop and propose a backup before any reset.
+
+## Session loop
+
+1. **Snapshot first.** Prefer `playwright_browser_snapshot` over screenshots when you need to act. The accessibility tree gives stable element refs for clicks and fills.
+2. **Navigate** with a full URL (`playwright_browser_navigate`). Wait for the content you care about with `playwright_browser_wait_for` (text visible or gone) rather than fixed sleeps.
+3. **Act** via ref from the snapshot: `click`, `type`, `fill_form`, `select_option`, `press_key`. One logical action, then re-snapshot before the next if the DOM may have changed.
+4. **Read** with `find` (search the snapshot for text/regex) before capturing a whole new tree when you only need one node.
+5. **Inspect when debugging:** `console_messages` (level `error` first), `network_requests` / `network_request` for failed calls and payloads.
+6. **Screenshot** only to show the user something visual (`take_screenshot`, save under a clear filename). Screenshots are not for finding elements.
+
+## Authenticated flows
+
+The persistent profile keeps sessions. On login pages:
+
+- Check whether you are already signed in (snapshot the account menu or user name) before typing credentials.
+- Never hard-code passwords in prompts or scripts. Ask the user to complete credential entry in the open browser when possible.
+- After login, re-navigate to the deep link you actually need; do not assume the redirect landed where you wanted.
+
+## Rate limits and failures
+
+- Repeated identical navigations can hit HTTP 429. Back off (seconds, then minutes), vary the path only if the target allows it, and tell the user rather than hammering.
+- On navigation error, capture console + last network response once, then stop and report. Do not retry in a tight loop.
+- Dialogs: `handle_dialog` explicitly (accept or dismiss with prompt text). Ignored dialogs block the page.
+
+## When the task needs a dev server
+
+Start the app server in an **external terminal** before any browser interaction. Point the browser at the real origin. Do not claim a UI works from code reading alone; load it, snapshot it, act on it.
+
+## Checklist
+
+- [ ] MCP tools only; Edge; persistent profile untouched.
+- [ ] Snapshot taken before each interaction sequence.
+- [ ] Waits are content-based, not fixed sleeps.
+- [ ] Console and network checked if anything failed.
+- [ ] Credentials never written into the repo or chat.
+- [ ] Result reported from observed browser state, not assumed.
+
+## Sources
+
+Novahiz global Playwright rules (`instructions.md`, persistent profile + Edge-only policy). Playwright MCP tool set (snapshot, navigate, click, type, console, network, dialogs). Original synthesis for this skill body.

package/skills/browser-session/references/session-run-sheet.md

@@ -0,0 +1,30 @@
+# Browser session run sheet
+
+Task: __________________  URL / app: __________________  Date: __________________
+
+## Preconditions
+
+- [ ] Dev server up in external terminal (if local app)
+- [ ] Using `playwright_browser_*` only (no chrome/chromium launch)
+- [ ] Persistent profile path unchanged
+
+## Steps
+
+| # | Intent | Tool / ref | Expected observation | OK? |
+|---|---|---|---|---|
+| 1 | | snapshot / navigate | | |
+| 2 | | find / click | | |
+| 3 | | fill / type | | |
+| 4 | | wait_for text | | |
+
+## Failures
+
+| Symptom | Console | Network (status) | Action taken |
+|---|---|---|---|
+| | | | |
+
+## Verdict
+
+- [ ] Goal observed in browser (state described from snapshot/screenshot)
+- [ ] Login state kept for next run
+- [ ] No rate-limit / no profile reset

package/skills/browser-session/scripts/browser_session_preflight.mjs

@@ -0,0 +1,74 @@
+#!/usr/bin/env node
+/**
+ * browser-session: preflight check that MCP browser rules are respected in a run log.
+ * Input: JSON array of steps [{tool, url?, note?}...]
+ * Flags violations: manual chrome/chromium launch, user-data-dir override, ephemeral context.
+ *
+ * Usage: node scripts/browser_session_preflight.mjs <run.json>
+ * Exit 1 on HIGH violation.
+ */
+
+import { readFileSync, existsSync } from "node:fs";
+
+const path = process.argv[2];
+if (!path) {
+  console.error('usage: node scripts/browser_session_preflight.mjs <run.json>');
+  process.exit(2);
+}
+if (!existsSync(path)) {
+  console.error(`missing: ${path}`);
+  process.exit(1);
+}
+
+let steps;
+try {
+  steps = JSON.parse(readFileSync(path, "utf8"));
+} catch (e) {
+  console.error(`HIGH\tinvalid-json\t${e.message}`);
+  process.exit(1);
+}
+if (!Array.isArray(steps)) {
+  console.error("HIGH\tnot-array\troot must be an array of steps");
+  process.exit(1);
+}
+
+const findings = [];
+const push = (level, id, msg, i) => findings.push({ level, id, msg, i });
+
+steps.forEach((step, i) => {
+  const tool = String(step.tool || "");
+  const blob = JSON.stringify(step).toLowerCase();
+
+  if (/chrome|chromium/.test(blob) && !/msedge|edge/.test(blob)) {
+    push("HIGH", "forbidden-browser", `step ${i}: chrome/chromium reference`, i);
+  }
+  if (/user-data-dir/.test(blob) && !/playwright-profile/.test(blob)) {
+    push("HIGH", "profile-override", `step ${i}: user-data-dir override or foreign profile`, i);
+  }
+  if (/newcontext|ephemeral|incognito/.test(blob)) {
+    push("HIGH", "ephemeral-context", `step ${i}: ephemeral context would drop login state`, i);
+  }
+  if (/rmdir|rimraf|remove.*profile|purge/.test(blob)) {
+    push("HIGH", "profile-purge", `step ${i}: profile purge requires explicit user consent`, i);
+  }
+  if (tool && !tool.startsWith("playwright_browser_")) {
+    push("MED", "non-mcp-tool", `step ${i}: tool ${tool} is outside playwright_browser_*`, i);
+  }
+  if (tool === "playwright_browser_type" || tool === "playwright_browser_fill_form") {
+    if (blob.includes("password") && blob.includes("value")) {
+      push("MED", "credential-in-log", `step ${i}: looks like a stored credential in the step payload`, i);
+    }
+  }
+  if (tool === "playwright_browser_navigate" && !step.url) {
+    push("HIGH", "navigate-no-url", `step ${i}: navigate without url`, i);
+  }
+});
+
+for (const level of ["HIGH", "MED", "LOW"]) {
+  for (const f of findings.filter((x) => x.level === level)) {
+    console.log(`${level}\t${f.id}\t${f.msg}`);
+  }
+}
+const high = findings.filter((f) => f.level === "HIGH").length;
+console.error(`browser_session_preflight: ${steps.length} steps, ${findings.length} findings (${high} HIGH)`);
+process.exit(high > 0 ? 1 : 0);

package/skills/browser-session/scripts/sample-run.json

@@ -0,0 +1,5 @@
+[
+  { "tool": "playwright_browser_navigate", "url": "http://localhost:3000" },
+  { "tool": "playwright_browser_snapshot" },
+  { "tool": "playwright_browser_click", "target": "ref=e1" }
+]