Runtime evidence for the BEAM, structured for coding agents.

Insider is an Elixir library and a set of development tools that turn the signals a running BEAM system already produces (OTP crash and supervisor reports, Logger events, Telemetry events, process and VM counters) into bounded, redacted, content-addressed evidence. That evidence can be detected against, captured into a portable capsule, reproduced on a disposable target, minimized, bisected, and used to verify a candidate repair, with every step producing a record that says exactly what was observed and what was not.

Insider is a pre-release proof of concept. Its protocol, APIs, and package layout are not stable. Production use is explicitly outside the current phase, and the project's own readiness artifacts conclude that staging and production observation remain denied.

The problem

A failing test, a restart storm, or a growing mailbox is easy to notice and hard to prove. The usual artifacts are an exit status, a stack trace, or a log line: enough to suspect a cause, not enough to confirm one, and not enough for a coding agent to act on without guessing. Agents given raw access to a live node can observe too much, change too much, and mistake a plausible hypothesis for a confirmed one.

Insider's position is that the loop around the signals is the valuable part:

  • correlate evidence from several sources into one normalized stream with explicit identity and ordering;
  • keep the causal window around a trigger in a fixed-size recorder;
  • export a sealed capsule that can be validated offline, without the original node;
  • reproduce the failure on a clean, disposable target, and record it as nonreproduced when it does not reproduce;
  • turn a confirmed reproduction into a regression, shrink it, and locate the boundary where it starts reproducing;
  • verify a candidate repair before and after, and produce an inert draft proposal for a human to review.

Every one of those steps is bounded by time, events, bytes, and targets, reports truncation and uncertainty, and refuses to grant itself authority it was not configured with.

Who it is for

Engineers who want a precise, reviewable record of a runtime failure rather than a hunch, and who want a coding agent to work from that record instead of from the live system.

Coding agents that need machine-readable candidates, oracles, and verification artifacts with closed vocabularies, explicit limits, and no ambiguity about what has and has not been proven. The agent-facing pages in this documentation are written for them.

What it is not

Insider does not replace Sentry, Datadog, or an observability platform; its provider mappings are deterministic mock data marked mock: true and delivery_status: not_attempted. It does not evaluate arbitrary code in production, inject faults, trace every function or message, or replay the VM. It does not merge or deploy anything. It does not find business-logic bugs that have no observable failure or declared invariant, and it does not claim to find bugs in the BEAM itself. A full list is in Safety and non-goals.

Where to go next