High-performance browser automation bridge and multi-instance orchestrator with advanced stealth injection and real-time dashboard.
Metrics
Runtime counters, recent failures and crash diagnostics.
curl -H "Authorization: Bearer $TOKEN" http://localhost:9867/metrics
Two layers, never summed
Server mode runs two processes, and each keeps its own counters:
| Layer | Process | Sees |
|---|---|---|
frontDoor |
the orchestrator that owns the port | auth rejections, unrouted paths, everything before a request is forwarded |
instance |
a browser-control child (also what pinchtab bridge runs standalone) |
every request that reached a browser |
Every response says which layer its numbers describe:
{
"layer": "frontDoor",
"metrics": { "requestsTotal": 41, "requestsFailed": 7, "avgLatencyMs": 3.1, "...": "..." },
"failures": {
"layer": "frontDoor",
"requestsFailed": 7,
"recent": [
{ "time": "...", "requestId": "a1b2c3", "method": "GET", "path": "/tabs",
"status": 401, "type": "http_error", "layer": "frontDoor" }
]
},
"crashes": { "total": 0, "recent": [] }
}
metrics carries requestsTotal, requestsFailed, avgLatencyMs, rateLimited,
staleRefRetries, rateBucketHosts and the Go runtime figures goHeapAllocMB, goHeapSysMB,
goNumGoroutine and goHeapObjects. A failure event carries code and message when it has a reason.
A failure is not always a 4xx. POST /actions and POST /macro answer 200 with per-item
results by design, so a run whose steps failed publishes a failure reason instead of a
status: it counts in requestsFailed, appears in failures.recent with the run's path,
status: 200, and a message of the form N of M steps failed: <first step's error>, and
logs at WARN. Watching requestsFailed therefore sees an agent whose batches are all
failing, which is the shape of traffic these endpoints exist to serve.
This holds at both layers. The reason travels the proxy hop on its own response
headers, so a batch served by an instance and proxied by the front door moves the front
door's requestsFailed too — which is the number an operator running pinchtab server
curls, since GET /metrics there reports the front door's own counters. A proxied
response carrying no reason records nothing, whatever its status.
The two layers are never added together. A single requestsTotal covering both
would mean two things at once, and an operator cannot act on that number: a spike
would not say whether clients are being turned away at the door or the browser is
failing. Read each layer at its own endpoint instead:
| Endpoint | Answered by | Reports |
|---|---|---|
GET /metrics |
front door (server mode) / the bridge itself (bridge mode) | that process's own counters |
GET /instances/{id}/metrics |
proxied to that instance | that instance's counters |
GET /instances/metrics |
front door | per-instance browser memory, not request counters |
GET /tabs/{id}/metrics |
proxied to the owning instance | that instance's browser memory — the whole process tree, not the tab |
The tab id in the last row selects the instance to ask, never the measurement. There is no per-tab reading behind it: two tabs of one instance answer identically, and the figure is the same one that instance reports for itself. See Memory monitoring.
failures.recent repeats layer on each event as well as on the block, so an
event pasted into a bug report still says where it came from.
failures and crashes are always present, empty when there is nothing to
report. A key that appeared only after the first failure could not be relied on by
anything watching for one.
What the two modes share, and what they do not
Shared, by construction — both modes serve the same diagnostics payload:
metrics(request counters, latency, Go runtime figures)failureswithrequestsFailedandrecentcrasheswithtotalandrecentversionon/health
Deliberately different:
/metricsis answered locally in both modes;/healthis not. In server mode/healthis the orchestrator's own envelope (instances,profiles,defaultInstance,restartRequired), because those facts exist only at the front door. A bridge has no instances to report. The two/healthbodies therefore differ on purpose, and/metricsis the endpoint to compare across modes.memoryappears only on an instance's/metrics, since only a process holding a browser can measure one.- Browser crash events are recorded by whichever process owns the browser, so in
server mode a crash shows up under the instance layer, not the front door — the
front door's
crashesblock stays empty by design.