Operations

Three endpoints, one metrics format, and one decision to make about which port they live on. eggrd exports Prometheus text on a plain HTTP path — there is no agent, no sidecar and no push.

Ops endpoints

Path	Answers	Use it for
`/__edgeguard/health`	`200` always	Liveness. Is the eggrd process itself up and serving? It does not consult the upstream, so it stays `200` during an upstream outage — restarting eggrd would not help.
`/__edgeguard/ready`	`200` / `503`	Readiness. Opens a TCP connection to the upstream with a 2-second timeout. `503` while the upstream is unreachable, so a load balancer stops sending traffic eggrd cannot serve.
`/__edgeguard/metrics`	Prometheus text	Scrape target. See below.

The split matters more than it looks. Pointing a liveness probe at readiness restarts a perfectly healthy proxy every time the app behind it hiccups, which turns one outage into two.

Keeping them off the public port

By default these paths live on the public listener. Set server.admin_port to a non-zero value and eggrd binds a second, plain-HTTP listener for them — the public port then serves only the proxy.

[server]
admin_port = 9090
admin_addr = "127.0.0.1"   # same-host only; 0.0.0.0 for a private network interface

admin_addr defaults to 127.0.0.1, so the endpoints are reachable only from the same host unless you widen it deliberately. Point your platform's health check at the admin port once you enable it, or it will be checking a port that no longer answers.

Metrics

Prometheus text format, scraped from /__edgeguard/metrics. The core series:

Series	Type	What it tells you
`edgeguard_requests_total`	counter	Requests by `outcome` label — the single most useful series here. See below.
`edgeguard_request_duration_seconds`	histogram	End-to-end latency through the proxy, with `_bucket`, `_sum` and `_count`.
`edgeguard_ratelimit_hits_total`	counter	Requests the limiter rejected.
`edgeguard_waf_hits_total`	counter	WAF matches, labelled by `rule` (`sqli`, `traversal`, and so on).
`edgeguard_csp_reports_total`	counter	CSP violation reports received at the browser-facing report sink.

Request outcomes

Every request lands in exactly one outcome bucket on edgeguard_requests_total. Alerting on the shape of this label is usually more informative than alerting on status codes, because it distinguishes which gate refused.

Outcome	Meaning
`ok`	Proxied to the upstream and returned.
`unauthorized`	Auth refused it. `401`.
`forbidden`	WAF, DLP, or a model allowlist refused it. `403`.
`ip_denied`	Client address outside the access list. `403`.
`rate_limited`	Over the configured rate. `429`.
`limiter_error`	Shared limiter store unreachable and failing closed. `503`.
`over_budget`	An LLM budget was exhausted. `429` for a token cap, `402` for a cost cap.
`unpriced_model`	Model absent from the price book while the policy requires one. `402`.
`payload_too_large`	Request body over `validation.max_body`. `413`.
`header_too_large`	Headers over the configured cap.
`method_not_allowed`	Method outside the allowlist. `405`.
`not_found`	No route matched.
`bad_gateway`	Upstream unreachable or answered unusably. `502`.
`upstream_error`	Upstream failed mid-response.
`upstream_timeout`	Upstream exceeded `validation.upstream_timeout`. `504`.
`ws_upgrade`	Connection upgraded to a WebSocket and passed through.

LLM gateway metrics

Present only when [llm] is enabled. Token counts come from the upstream's own usage object — eggrd does not tokenize anything itself, so the numbers are the provider's, not an estimate.

Series	What it tells you
`edgeguard_llm_requests_total`	Requests by `result` — including `unpriced`, which is how you find a model missing from the price book.
`edgeguard_llm_tokens_total`	Total tokens, with `_model_`, `_key_` and `_team_` variants for attribution.
`edgeguard_llm_cost_microdollars_total`	Cost in microdollars, with the same three attribution variants. Integer microdollars rather than floats, so summing does not drift.
`edgeguard_llm_cached_tokens_total`	The cached subset of prompt tokens, where the provider reports it.
`edgeguard_llm_reasoning_tokens_total`	Reasoning tokens, where the provider reports them separately.
`edgeguard_llm_budget_consumed_ratio`	Gauge, 0–1, per budget. The series to alert on before a cap starts refusing traffic.
`edgeguard_llm_budget_blocked_total`	Requests a budget refused.
`edgeguard_llm_budget_reconcile_failures_total`	Reservations that could not be reconciled to actual usage. Non-zero means recorded spend is drifting from reality — worth an alert.
`edgeguard_llm_dlp_findings_total`	DLP matches by `category`.
`edgeguard_llm_dlp_blocked_total`	Requests DLP refused outright.
`edgeguard_llm_keyvault_total`	Virtual-key lookups, including denials.
`edgeguard_llm_ttft_seconds`	Histogram: time to first token.
`edgeguard_llm_tpot_seconds`	Histogram: time per output token.

What fails closed

Two shared stores can be unreachable, and both refuse traffic rather than waving it through by default. An outage that silently disabled rate limiting, or silently uncapped spend, would be the worst possible failure for either feature.

Store	Setting	Default	On store error
Rate limiter	`ratelimit.fail_open`	`false`	`503`, and the request never reaches the upstream.
LLM budgets	`llm.fail_open`	`false`	`503`, and the provider is never called.

Setting either to true trades correctness for availability: requests are admitted unlimited while the store is down. That is a legitimate choice for a rate limiter in front of a resilient app, and a dangerous one for a spend cap.

Endpoint behaviour and metric names were read from the shipped source rather than from memory. The configuration reference is generated from it directly.