Pre-merge performance gate

Merged in minutes.
Reviewed by no one.

ShadowTel replays real production traffic against every code change and blocks the merge if it makes your system slower, less reliable, or more expensive to run.

ShadowTel / traffic-replay  · PR #482 Refactor payment processing loop Required
Running traffic replay…
  • p95 latency18ms412msFAIL
  • error rate0.02%1.80%FAIL
  • cost / 1k req$0.11$0.34FAIL
  • memory (rss)512MB754MBWARN
Replayed 12,450 requests over 180s against production baseline. Blocks only on regressions outside normal variance.

Nobody is reading the code

AI writes code much faster than people can read it. So a lot of it now ships without anyone looking at it. The senior engineer who used to catch the expensive mistake in review can't keep up with the volume.

+31%

Code merging with no review at all. Not human, not agentic.

3.4×

Production incidents per code change.

+441%

Median time senior engineers spend in code review.

Faros AI Engineering Report 2026. Two years of telemetry from 22,000 developers across 4,000 teams.

How it works

No new agent in your request path. No staging environment to maintain.

SHADOWTEL live traffic your proxy envoy · nginx · mesh your users served normally, never held mirrored copy PR sandbox writes + egress held compare vs your baseline PASS BLOCK
01 / MIRROR

Point a slice of live traffic at us

Request mirroring you already have in Envoy, nginx, or your service mesh. No code changes.

02 / REPLAY

Every PR gets a disposable sandbox

We replay the mirrored traffic against it, holding writes and outbound calls inside the sandbox.

03 / GATE

Pass or fail on the pull request

Compared against a rolling baseline from your OpenTelemetry data, with the dollar delta derived from measured CPU, memory, and egress at your own cloud rates. You set the budgets, and it runs advisory until you turn enforcement on.

Safe to put in the merge path

This thing can stop your deploys. So it has to be safe, and it has to be right.

Writes stay in the sandbox

Mutating requests are held or served by sandbox-local stand-ins. They never reach production data stores.

Outbound calls are intercepted

Payment providers, email, and third-party APIs are stubbed at the sandbox boundary. Nothing bills, nothing sends.

Production is untouched

Mirroring is fire-and-forget at your proxy. Your users' requests are never held waiting on us.

Customer data is redacted before it leaves

Headers, bodies, and query strings pass through a redaction step at capture time. Tokens, credentials, and configurable PII fields are replaced before anything is stored or replayed.

Want to be one of the first five
teams to run it?

I'm looking for teams already shipping AI-written code who have been bitten by it. Half an hour, no deck.

Book a call
Deterministic Sandboxed Cost-aware Blocking