Index · Glossary · Cheat sheet

What works at this revision, and the phases of the roadmap.

Contents

Current status

Each table lists one capability per row. "Core" is the root :insider application. "Lab" is the separate tools/insider_lab project, published as insider_lab. Notes under each table hold the limits that a row does not show. The status continues in Live observation status and Lab investigation status.

Core evidence pipeline

AreaWhat works todayWhere it livesSection
Build and gatesThe Mix application and the quality gateCoreChecks and gates
ProtocolThe bounded v1 protocolCoreModule boundaries
Runtime policyDeny-by-default authorization from trusted runtime context, with bounded audit evidenceCoreModule boundaries
Fault fixturesPaired faulty and healthy fixtures that emit raw runtime measurementsLabRepository layout
Event schemaThe normalized event v1 schemaCoreModule boundaries
IdentityExplicit same-node identityCoreModule boundaries
RecorderThe bounded priority recorderCoreModule boundaries
Logger ingressStructured OTP Logger ingressCoreSupervised collector sessions
Restart recurrenceThe first detectorCoreRestart-recurrence detector
RedactionThe fail-closed redaction primitive and its Event-compatible manifestCoreRedaction boundary
ProvenanceBounded portable provenanceCorePortable provenance
CapsulesOffline-validating, content-addressed incident capsulesCoreIncident capsules
Runtime pollingBounded scalar polling of explicit same-node processesCoreBounded runtime polling
Mailbox growthA conservative bounded detectorCoreMailbox-growth detection
System eventsAn opt-in system-event collectorCoreSystem-event collector
  • The capsule boundary accepts restart-recurrence, mailbox-growth, and declared semantic-invariant findings. They arrive through one detector-neutral wire envelope. Legacy restart capsules still verify.
  • The fixtures emit raw runtime measurements without detector verdicts.
  • The policy evaluates requests against trusted runtime context. It ignores environment or capability fields that a caller claims.
  • The mailbox-growth detector separates sustained growth without reductions from bursty recovery and from declared intentional buffering.
  • The system-event collector uses isolated OTP 28-29 trace sessions. On OTP 27 it has an explicit legacy mode that refuses conflicts. It projects hostile messages into a closed shape, disables itself on overload, and accounts for loss within bounds.
  • Generic adopter adapters and additional detectors are future work.

Live observation status

AreaWhat works todayWhere it livesSection
Live sessionOne leased collector session inside a development, test, or CI nodeCoreLive observation
ReadsRedacted events after a recorder cursor, with eviction and loss gapsCoreCursors and gaps
SubscriptionsMonitored subscriber batchesCoreSubscriptions
TidewaveLive through Tidewave project_evalCoreTidewave project_eval
Observer nodemix insider.live and insider-lab live over Erlang distributionLabObserver node
WelcomeA welcome document generated from the codeCoreWelcome document
Agent reportsBounded, deduplicated reports about Insider itselfCoreAgent reports
DisclosureHost-chosen rich or strict profileCoreDisclosure profiles
Process and ETSPer-interval rankings of processes, ETS tables, and portsCoreProcess and ETS evidence
One-shot inspectionsprocess, bin_leak, ports, allocCoreOne-shot inspections
Call tracesCalls and failures of named functions (OTP 28 and later)CoreCall traces
Telemetry discoverydiscover lists the Telemetry event names a node can emitCoreTelemetry discovery
Exemplars and trendsSummary exemplars and trendsCoreSummary trends
Logger burstsFloods counted per message template and call siteCoreFilters, summaries, and views
Digestdigest over a recent window, anomalies firstCoreDigest
BaselineAnomalies in a known-normal baseline are hiddenCoreKnown-normal baseline
Pairing marksmark, marks, and --since-markCorePairing marks
Pairing chartchart writes a page of the session for the programmer; chart --serve serves it live on 127.0.0.1LabPairing chart
OpenTelemetrySpans of the node's SDK, joined into traces (OTP 28 and later)CoreOpenTelemetry spans and traces
Frontend tracesA frontend's spans added through the OTLP receiverLabFrontend traces
GraphQL outcomeThe outcome on spans and Telemetry eventsCoreGraphQL outcome
Query shapesA trace groups its repeated query shapesCoreRepeated query shapes
Request waitsEach request's waitCoreRequest waits and bursts
Mutation recordsA kept record of each GraphQL mutationCoreKept mutation records
Value changeschanges names the writer of an in-memory valueCoreChanges of in-memory values
  • A supervised owner holds the Live session. Insider.Live has a canonical-JSON facade.
  • The observer attaches a hidden, short-lived node to a local development, test, or CI target. It loads Insider into the target only with an explicit --load. --load replaces another Insider build only when Insider is not started. --force-load replaces a started Insider and ends its session.
  • The welcome document carries a content hash, and a drift-guard test keeps it in step with the code.
  • Agent reports reach GitHub only when an operator or the host opts in.
  • The rich profile is the default in development and test. It shows Logger text and metadata, Telemetry metadata, and all measurements. The strict profile applies in ci and wherever live observation is denied. Callers may only narrow the profile. Every profile masks credentials.
  • The poller component ranks processes by mailbox length, memory, and reductions since the previous scan. It ranks ETS tables by memory and size, in the style of recon but without depending on it. It counts ports by type and owner, with each type's growth (port_top). It samples named processes. It reports Insider's own processes, tables, and ports apart from the application's.
  • One-shot inspections answer one bounded question about the node now. Their results join the session's history.
  • The calls component traces in an isolated trace session it owns. It pairs each failure with the arguments that failed, rendered and capped in the rich profile. It removes the trace at stop, lease expiry, or amend. The receipt shows that the functions are no longer traced.
  • discover reads attached handlers and the literal :telemetry.execute and span names in the debug info of loaded modules. It runs no application code. A session can subscribe to a discovered prefix such as shop.*. status reports each subscribed event's seen count, so an accepted but silent name reads as not emitted.
  • A Telemetry summary names the receipts of its interval's slowest events and first error as exemplars. It reports the exemplar candidates it lost.
  • trends lists each summary series across its windows in fixed units. It shows a duration's handler or caller scope only when the session declared it.
  • Logger bursts keep at most 64 bursts an interval. The rest fold into a counted overflow row.
  • digest reads what the Recorder admitted in the last minute by default. latest-DURATION selects another span. A cursor or latest-N selects receipts. The digest says which window it read, when that window starts, and how old its newest event is. Each hint names the window it was worked out over. The digest gives next-command hints.
  • mark records the moment before a programmer acts in the app. digest, trends, events, and timeline then read only what came after it (--since-mark). marks lists the marks.
  • chart reads with its own reader. chart --serve renews no lease, reads the window again when the host narrows the profile, and says on the page what it could not read.
  • Mutation records need the telemetry component. They are kept apart from the Recorder.
  • changes needs watch_writes.

Lab investigation status

AreaWhat works todayWhere it livesSection
Sentinel manifestA content-addressed manifest that binds exact capsule bytesLabCollector-to-Sentinel handoff
ScenariosPortable bounded scenarios with deterministic semantic recipesLabSentinel and proof workflow
Sandbox lifecycleA provider-driven lifecycle with owner-scoped leasesLabSentinel and proof workflow
ExUnit candidatesA bounded, deterministic inventory of investigation candidatesLabAgent-requested ExUnit candidates
Fixed oraclesValue-free oracles for application env, Telemetry, and OpenTelemetry globalsLabLab candidates and oracles
Oracle runsmix insider.oracle runs one oracle against a test commandLabLab candidates and oracles
Actor differentialOne operation run as several actors, with a value-free access verdictLabLab candidates and oracles
Fault and rerunA worker run with a declared provider fault and its retryLabLab candidates and oracles
Privacy corpusSynthetic Logger and Telemetry items through the real ingress codeLabPrivacy corpus runner
Draft repairPhase 6 offline draft repair orchestration as inert artifactsLabProof workflow
  • The Sentinel manifest reports external confirmation prerequisites. It conveys no execution authority.
  • Scenarios define cleanup-reserved disposable execution. They do not claim deterministic BEAM replay.
  • The sandbox lifecycle binds exact observed provenance to its leases. It reports cleanup complete only after an empty independent inventory. It quarantines any reclamation it cannot prove.
  • Candidates have static application-env, Telemetry handler, and Telemetry event targets. Each carries oracle input hints.
  • The fixed oracles cover up to sixteen application-env keys, dead :telemetry_test destinations on one exact event, and identity-only change detection of the OpenTelemetry default tracer and text-map injector/extractor globals.
  • mix insider.oracle does not edit the target. It saves a versioned run record.
  • An agent's probe test calls the actor differential. It runs the operation behind healthy controls.
  • The fault-and-rerun oracle runs behind a healthy control. It reports orphaned or repeated effects.
  • The privacy corpus runner takes caller-supplied items. It reports retained field names and canary coordinates without values. The application corpus and the field review stay human work.
  • Broader target-local operations stay separate slices, each gated on evidence.
  • Draft repair artifacts: a bounded coding-agent handoff built from a confirmed regression plus two fresh failed confirmations, a candidate submission bound to that exact handoff, a recomputed repair report, and an offline draft proposal that carries no disclosure approval.
  • More draft repair artifacts: a source-ownership routing companion, a bounded in-memory dedup ledger, a whole-step scenario minimizer, and an offline bisection. The bisection runs over a host-declared ordered target list and isolates one adjacent nonreproduced/reproduced boundary.
  • The candidate submission and the ownership routing are declared but not verified: Insider records what the submitter declares and does not check it.
  • The draft-creation gate is disabled by default. Its only providers are test fixtures.
  • One Lab integration test runs this chain end to end on a live sandbox executor. It checks that the same candidate identity threads through submission, verification, proposal, ownership, and the executor request.
  • Before any provider authority, three items remain: a candidate-disclosure attestation, a real draft provider, and dedup persistence. Human review stays required.

Roadmap

  1. Protocol and fixtures (done): establish typed operations, environment policies, schemas, and deliberately faulty applications.
  2. Evidence pipeline (done): normalize signals and produce a valid bounded capsule while a healthy control stays quiet.
  3. Reproduction loop (done): reproduce on a clean target, persist a failing regression, and prove before/after repair evidence.
  4. Semantic and ecosystem expansion (in progress): deepen contracts and add optional observer_cli, recon, Tidewave/MCP, property-testing, and Mob adapters. Mob is a mobile app framework on the BEAM.
  5. Production observation readiness (in progress): threat-model and load-test a read-only collector, audit redaction, and design Sentry/Datadog integration. Production stays policy-denied. A bounded staging canary and a later production release each need a separate, revision-bound human decision.
  6. Offline draft repair orchestration (in progress): add declared but unverified ownership routing, deduplication, bounded coding-agent handoffs, bisection, and evidence-gated draft proposals. Do this before granting any Git or PR provider authority. Human review stays required.
  7. Agent-directed investigation (in progress): expose bounded candidate inventories and target-local evidence operations one proven slice at a time. Agents choose candidates and investigation steps. Insider supplies structured signals. It never treats a source heuristic as a confirmed defect.