Ship Loop Watch

ship-loop-watch — Chrome MCP browser-watch child loop

Activation

• Standalone: /loop ship-loop-watch (production debugging)
• Orchestrated: invoked by ship-loop-master every 15 min (most frequent inner loop — production state can change without our action)

Pre-condition

• Chrome MCP must be connected (verified via mcp__Claude_in_Chrome__tabs_context_mcp succeeding)
• If not connected: write event {"reason":"chrome_mcp_unavailable","skipped":true} and return without changing watch state

Process

Step 1: Surface inventory

Surfaces to watch (ordered most-critical first):

Step 2: Per-surface capture

For each browser surface (not the API ones):

1. mcp__Claude_in_Chrome__navigate to URL
2. Wait 3s for hydration
3. mcp__Claude_in_Chrome__read_console_messages filtered with pattern error|warn|CSP|Content Security
4. mcp__Claude_in_Chrome__read_network_requests filtered with urlPattern matching the host; check for any 4xx/5xx
5. mcp__Claude_in_Chrome__read_page filter interactive to assert at least 1 element exists (proves hydration)
6. For pages with JSON-LD requirement: read_page and grep for application/ld+json element

For API surfaces: simple curl with --ssl-no-revoke (Windows compat) and JSON parse.

Step 3: Update state

Atomic update to layers.watch:

{
  "state": "green" | "amber" | "red",
  "last_run": "<iso>",
  "details": {
    "surfaces": [
      {
        "url": "https://synthex.social/login",
        "http_code": 200,
        "console_errors": 0,
        "csp_violations": 0,
        "network_5xx": 0,
        "rendered": true,
        "json_ld_present": null
      },
      ...
    ],
    "api_health_status": "healthy" | "degraded" | "unhealthy"
  }
}

State levels:

• green: all surfaces 200, 0 console errors, 0 CSP violations, 0 5xx
• amber: API health degraded (cold-start latency tolerated) OR ≤2 console warnings
• red: any surface 5xx OR any CSP violation OR API health unhealthy OR hydration failure

Step 4: Recovery / classify on red or amber

Amber (degraded health, cold start): retry once after 60s sleep. If still amber, log to events but do not escalate (production cold-start is normal).

Red triggers classification (not auto-fix — production code is not loop's to edit):

1. Pull the failing surface details
2. Match against browser-debug skill's failure-mode classifier (CSP / hydration / 5xx / cold-start)
3. Check Vercel deploy status via gh api repos/CleanExpo/Synthex/deployments for last deploy
4. If last Vercel deploy is FAILURE: escalate as ## P0 — Production deploy failure, master will halt
5. If deploy is READY but production red: escalate as ## P1 — Production regression on <surface> with the recipe lookup result
6. If recipe match: include the recipe's "Suggested fix" in the escalation entry

Production-fix work happens via subsequent PR cycles, not from this loop.

Recipe priorities for this loop

• #13 (cold-start latency) — auto-retry without escalating
• #14 (production 5xx) — Vercel deploy state check + escalate

Verification

• Run when production is healthy; expect state === 'green' with all surfaces clean
• Manually break production (don't actually do this!); simulate by editing the surface list to include https://synthex.social/this-route-does-not-exist; expect red state + escalation

Out of scope

• Production fixes (loop reports; subsequent PR cycles fix)
• Visual regression (screenshot diff — separate concern, requires baseline images)
• Performance benchmarking beyond cold-start latency (handled by site-smoke-test skill or Lighthouse)
• E2E user journey verification (Playwright suite, separate from this skill)