Fix Playwright cleanup crash, CDP port conflicts, localhost IPv6, and URL null-string pollution

[andrii-mazurchuk]

Summary

Four operational bugs found while running ApplyPilot on Ubuntu 26.04 with multiple concurrent instances.

1. Enrich stage crashes on exit when Chrome launch fails (detail.py)

Problem: When Playwright fails to launch Chrome mid-batch (e.g. because the system Chrome binary is unavailable), the cleanup path of with sync_playwright() as p: raises 'PlaywrightContextManager' object has no attribute '_playwright', propagating an unhandled exception that kills the entire enrich stage with exit 1.

Fix: Wrap the with sync_playwright() as p: block in a broad except Exception at the batch level. Any jobs that hadn't been processed yet are written to the DB as browser_unavailable so they can be retried, and the stage exits cleanly.

---

2. CDP endpoint uses localhost which resolves to IPv6 on Ubuntu 26.04 (launcher.py)

Problem: http://localhost:{port} in _make_mcp_config() resolves to ::1 (IPv6) on Ubuntu 26.04, but Chrome's remote debugging port only binds to 127.0.0.1 (IPv4). Result: ECONNREFUSED ::1:9222 on every apply attempt.

Fix: Replace localhost with 127.0.0.1 in the CDP endpoint URL.

---

3. Concurrent apply runs on multiple instances share port 9222 (chrome.py)

Problem: All instances hard-code BASE_CDP_PORT = 9222. When two instances run apply concurrently, one instance's _kill_on_port(9222) kills the other instance's Chrome mid-session.

Fix: Derive a stable per-instance base port from the MD5 hash of the instance directory name, mapping each instance into a unique 10-port window in the range 9200–9399. Deterministic: same instance always gets the same port.

def _instance_base_port() -> int:
    instance_dir = os.environ.get("APPLYPILOT_DIR", "")
    name = Path(instance_dir).name if instance_dir else "default"
    offset = int(hashlib.md5(name.encode()).hexdigest()[:2], 16) % 20 * 10
    return 9200 + offset

---

4. LLM returns "None" string for missing application_url (detail.py, launcher.py, prompt.py)

Problem: When no apply URL is found, the LLM returns the string "None" rather than JSON null. Python's or operator treats non-empty strings as truthy, so application_url = "None" passes all guards and gets stored in the DB and shown to the apply agent as a valid URL (URL: None), causing the agent to refuse to proceed.

Fix:

• detail.py: sanitize LLM output before writing — treat "None", "null", "N/A", "" as NULL
• launcher.py: guard apply_url construction with explicit != "None" check
• prompt.py: same guard when building the URL line shown to the apply agent

---

5. Site-specific password support (prompt.py)

Problem: Different job sites may require different account credentials but the profile only supports one global password.

Fix: Check personal.get('site_passwords', {}).get(job_site, personal.get('password')) so per-site passwords in profile.json take precedence over the default.

Test environment

• Ubuntu 26.04, system Chrome (Playwright bundled Chromium not supported on this kernel)
• Multiple concurrent instances via APPLYPILOT_DIR
• Gemini 2.0 Flash as LLM

Checklist

• Fixes are isolated to the affected files
• No new dependencies introduced
• Backwards compatible — existing single-instance setups unaffected