feat: Implement session logging and workflow execution for Chrome tester

- Add SessionLogger class to handle logging of session details, including steps and results.
- Create main.js to orchestrate the agent's execution, including command parsing and decision-making.
- Introduce runtime.js for utility functions like sleep, JSON writing, and random ID generation.
- Develop workflow.js to manage the execution of defined workflows with error handling.
- Implement JSON parsing utilities in json.ts for handling command responses.
- Add prompt generation functions in prompt.ts for structured command interaction.
- Create safety checks in safety.ts to identify potentially unsafe commands.
- Implement truncation logic in truncate.ts for managing output length.
- Add unit tests for JSON, safety, prompt, and truncation utilities.
- Create build.rs for linking with llama-cpp-ffi common library.
- Document testing procedures in testing.md for local and CI environments.

Author: netdur

.github/workflows/ci.yml

@@ -0,0 +1,44 @@
+name: CI
+
+on:
+  push:
+  pull_request:
+
+jobs:
+  agent-cli-tests:
+    name: Agent CLI Tests
+    runs-on: ubuntu-latest
+    defaults:
+      run:
+        working-directory: agent/cli
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Node
+        uses: actions/setup-node@v4
+        with:
+          node-version: 20
+          cache: npm
+          cache-dependency-path: agent/cli/package-lock.json
+
+      - name: Install dependencies
+        run: npm ci
+
+      - name: Run tests
+        run: npm test
+
+  rust-tests:
+    name: Rust Tests
+    runs-on: ubuntu-latest
+
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v4
+
+      - name: Setup Rust
+        uses: dtolnay/rust-toolchain@stable
+
+      - name: Run cargo test --all-targets
+        run: cargo test --all-targets

Cargo.lock

@@ -2,6 +2,7 @@
 name = "hugind"
 version = "0.11.2"
 edition = "2024"
+build = "build.rs"
 
 [dependencies]
 rquickjs = { version = "0.11", features = ["loader", "futures", "classes", "macro"] }
@@ -31,7 +32,7 @@ axum = "0.7.7"
 tempfile = "3.24.0"
 zip = "2.2.0"
 semver = "1.0"
-llama-cpp = { path = "/Users/adel/Workspace/llama_cpp_rust/llama-cpp", features = ["metal", "mtmd"] }
+llama-cpp = { git = "https://github.com/netdur/llama_cpp_rust", package = "llama-cpp", features = ["metal", "mtmd"] }
 tower = { version = "0.4", features = ["util", "timeout", "limit"] }
 tower-http = { version = "0.5", features = ["trace", "cors"] }
 tracing = "0.1"

Cargo.toml

@@ -358,6 +358,9 @@ more /Users/adel/.hugind/logs/agents/ocr/20260210_122933.526.txt
 * [Stdio Bridge](docs/stdio_bridge.md)  
   NDJSON + MCP protocol reference for desktop/app integration (`hugind stdio`).
 
+* [Testing Guide](docs/testing.md)  
+  Local test commands, CI parity commands, and troubleshooting notes.
+
 * [WASM Runtime](docs/wasm_runtime.md)  
   Covers WASM execution lifecycle, hostcalls, filesystem modes, and runtime limits.
 

README.md

@@ -0,0 +1,97 @@
+# chrome_tester
+
+Autonomous web tester agent built for Hugind using the official Chrome DevTools MCP server.
+
+## What It Does
+- Discovers Chrome MCP tools dynamically via `tools.list()`.
+- Uses an LLM loop to choose the next action.
+- Acts using Chrome MCP tools (`take_snapshot`, `click`, `fill`, `wait_for`, etc.).
+- Detects blocking takeover overlays/ad-like frames (`detect_blocking_overlay`).
+- Attempts recovery (`dismiss_blocking_overlay`) by clicking close controls or hiding blockers in test mode.
+- Extracts phone numbers from page state (`extract_phone_numbers`) and prints candidates in console.
+- Logs every step under `logs/`.
+
+## Prerequisites
+- Google Chrome installed.
+- `chrome-devtools-mcp` available through npm/npx.
+
+## Run
+
+```bash
+./target/release/hugind agent run agent/chrome_tester -- \
+  --goal "Open example.com and verify the main heading is visible" \
+  --start-url "https://example.com" \
+  --max-steps 20
+```
+
+## CLI Options
+- `--goal <text>`: testing objective.
+- `--start-url <url>`: optional initial navigation target.
+- `--workflow <path.json>`: run a multi-step workflow file.
+- `--max-steps <n>`: max loop iterations.
+- `--step-delay <seconds>`: minimum delay between steps.
+
+## Workflow Mode
+
+Run:
+
+```bash
+./target/release/hugind agent run agent/chrome_tester -- \
+  --workflow agent/chrome_tester/examples/wikipedia_complex_workflow.json
+```
+
+Marrakech restaurant phone workflow:
+
+```bash
+./target/release/hugind agent run agent/chrome_tester -- \
+  --workflow agent/chrome_tester/examples/marrakech_restaurant_phone_workflow.json
+```
+
+Reddit research workflow (Google-block resistant):
+
+```bash
+./target/release/hugind agent run agent/chrome_tester -- \
+  --workflow agent/chrome_tester/examples/reddit_research_workflow.json
+```
+
+Deterministic local workflow (no external bot blocking):
+
+1. Start local static test server:
+
+```bash
+python3 -m http.server 8787 --directory agent/chrome_tester/test_site
+```
+
+2. In another terminal run:
+
+```bash
+./target/release/hugind agent run agent/chrome_tester -- \
+  --workflow agent/chrome_tester/examples/local_restaurant_phone_workflow.json
+```
+
+Workflow JSON shape:
+
+```json
+{
+  "name": "workflow_name",
+  "continueOnFailure": false,
+  "steps": [
+    {
+      "name": "step name",
+      "startUrl": "https://example.com",
+      "maxSteps": 15,
+      "goal": "What to validate",
+      "checks": ["extra check 1", "extra check 2"],
+      "instructions": "optional instructions"
+    }
+  ]
+}
+```
+
+## Notes
+- This agent intentionally uses only MCP browser tools (no shell automation).
+- Default mode is visible Chrome (no `--headless`) for easier visual debugging.
+- The manifest currently keeps `--isolated`, so it uses a separate profile/session.
+- If your environment sandboxes spawned processes heavily, run Chrome with remote debugging and configure MCP with `--browser-url`.
+- Tool names can evolve across `chrome-devtools-mcp` versions, so this agent resolves capabilities from runtime discovery instead of hardcoding one exact schema.
+- If `tools.list()` is incompatible in your MCP bridge, the agent now falls back to known Chrome tool-name variants automatically.