[aryasaatvik]

Adds @executor-js/plugin-execution-metrics — an opt-in metrics sink built on the observer foundation.

• Base plugin drives Effect Metric counters + an execution-duration histogram from the event stream.
• ./api: an auth-gated GET /execution-metrics/metrics Prometheus scrape (executionMetricsHttpPlugin).
• ./cloudflare: a Workers Analytics Engine sink (writeDataPoint) — per-isolate Effect metrics being meaningless on a Worker fleet. Wired opt-in on host-cloudflare, gated on the ANALYTICS binding.

Base . entry free of @executor-js/api and @cloudflare/workers-types.

Stacked on the execution-observer foundation. Gates: format:check, lint, full-monorepo typecheck, and metrics-render tests green (an adversarial-verify pass caught + fixed a Prometheus histogram-bucket bug).

Stack

1. feat: add the execution-observer foundation #12
2. feat: add the execution-history plugin #13
3. feat: add the execution-metrics plugin #14 👈 current

[socket-security]

No dependency changes detected. Learn more about Socket for GitHub.

👍 No dependency changes detected in pull request

[greptile-apps]

Greptile Summary

Adds @executor-js/plugin-execution-metrics, a new opt-in plugin that fans execution lifecycle events into two distinct sinks: a global Effect Metric registry (counters + duration histogram) for long-lived self-hosted processes, and a Workers Analytics Engine sink for Cloudflare deployments where per-isolate metrics would be meaningless.

• Base SDK plugin (./) drives Effect Metric counters and a histogram via createExecutionMetricsObserver; start timestamps are kept in a 10,000-entry bounded LRU map so ExecutionFinished can compute durations without leaking memory.
• API sub-entry (./api) adds an auth-gated GET /execution-metrics/metrics Prometheus scrape via a hand-rolled text-exposition renderer that correctly handles all five Effect Metric types and emits the +Inf bucket exactly once (verified by adversarial test).
• Cloudflare sub-entry (./cloudflare) writes one writeDataPoint per finished execution/tool call; gated on the optional ANALYTICS binding so the Worker boots cleanly without it.

Confidence Score: 5/5

Safe to merge — the plugin is strictly additive and opt-in; no existing paths are altered and the Cloudflare host degrades gracefully to a no-op when the ANALYTICS binding is absent.

All three sub-entries are cleanly isolated (no cross-contamination of @executor-js/api or @cloudflare/workers-types into the base entry), the bounded-map correctly handles the previously noted memory-growth risk, the +Inf duplicate-bucket bug was caught and fixed before merging (with a regression test), and the only open item is a minor rendering consistency nit in the Prometheus serializer.

No files require special attention; prometheus.ts has one minor formatValue inconsistency for _count lines but it carries no practical impact given Effect's histogram/summary state types.

Important Files Changed

Filename	Overview
packages/plugins/execution-metrics/src/sdk/observer.ts	Core Effect Metric observer: drives counters and duration histogram from the execution event stream; correctly cleans up per-execution start state and degrades gracefully when a start entry is absent.
packages/plugins/execution-metrics/src/sdk/bounded-map.ts	LRU-style bounded-map helper that caps in-flight start-time entries at 10,000 and evicts the oldest on overflow; prevents unbounded memory growth in long-lived hosts.
packages/plugins/execution-metrics/src/api/prometheus.ts	Custom Prometheus text-exposition renderer over the global Effect Metric registry; handles Counter, Gauge, Histogram, Summary, and Frequency; histogram and summary _count lines use raw interpolation rather than formatValue unlike their _sum counterparts.
packages/plugins/execution-metrics/src/cloudflare/index.ts	WAE observer: writes one Analytics Engine data point per finished execution/tool call, tracking start times for both event types via the shared bounded-map helper; correctly fire-and-forget inside Effect.sync.
packages/plugins/execution-metrics/src/api/prometheus.test.ts	End-to-end test covering the observer→registry→renderPrometheus pipeline; verifies +Inf bucket is emitted exactly once (the adversarial fix), trigger_kind label propagation, and all required TYPE headers.
apps/host-cloudflare/src/plugins.ts	Wires executionMetricsPlugin with WAE observer when ANALYTICS binding is present, falling back to noopExecutionObserver; correctly avoids mounting the Prometheus HTTP route on Workers.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    EE[Execution Engine] -->|ExecutionEvent stream| CO[composeExecutionObservers]

    CO --> EMO{executionMetricsPlugin\nobserver option}
    EMO -->|default: long-lived host| EffectObs[createExecutionMetricsObserver\nSDK observer]
    EMO -->|Cloudflare + ANALYTICS binding| WAEObs[createWaeMetricsObserver\nWAE observer]
    EMO -->|Cloudflare, no binding| Noop[noopExecutionObserver]

    EffectObs -->|ExecutionStarted| BM1[rememberStart\nboundedMap 10k cap]
    EffectObs -->|ExecutionFinished| MR[Effect Metric Registry\nglobal counters + histogram]
    EffectObs -->|ToolCallStarted/Finished| MR

    WAEObs -->|ExecutionStarted| BM2[rememberStart\nboundedMap 10k cap]
    WAEObs -->|ExecutionFinished| WAE[(Workers\nAnalytics Engine)]
    WAEObs -->|ToolCallFinished| WAE

    MR -->|Metric.snapshot| Prom[renderPrometheus\ntext/plain]
    Prom -->|GET /execution-metrics/metrics\nauth-gated| Scraper[Prometheus Scraper]

Loading

_{Reviews (4): Last reviewed commit: "fix(plugin-execution-metrics): address G..." | Re-trigger Greptile}