What Insider will refuse to do

Copy Markdown View Source

Insider fails closed. When an input, environment, or state is outside what a component can prove safe, the component returns a typed refusal and does not fall back to a weaker behavior. This page lists those refusals so an agent can recognize them, understand why they exist, and stop instead of working around them. Each entry names where the refusal lives.

Environments and authority

  • Observe in staging or production. Insider.Collector.Session.start/2, Insider.Runtime.describe/3, Insider.Collector.SystemMonitor, and the Lab runtime adapters accept only development, test, and ci. The refusal happens before any process, filter, handler, timer, or monitor exists. (The privacy corpus runner is the deliberate exception: it takes staging or production precisely to project items through the denied profiles' allowlists without observing anything.) Why: enabling either environment is a separate, revision-bound human decision recorded outside the code; the readiness dossier grants none.
  • Trust an environment or capability a request claims. Requests are evaluated against trusted runtime context supplied at host startup (Insider.Policy.Context); the environment and capability fields in a request are inputs to a policy decision, not the decision. Why: a transport cannot authenticate its caller.
  • Remove a protected target. Insider.Runtime.HostPolicy and Insider.Policy.ProtectedTargets let a host add protected PIDs and names but never remove the versioned OTP/Elixir inventory or Insider's own session processes. Why: observing a supervisor's own infrastructure is an observer effect on the thing being measured.
  • Accept textual PIDs. mix insider inspect takes --name only. Why: a PID string copied from another BEAM can alias an unrelated local process.
  • Infer a free session slot. A session start that times out while its DynamicSupervisor.start_child/2 call is queued stays quarantined until the call reports; a new session is never permitted on inferred absence. Why: the collector may still be starting.
  • Remove a Logger filter it did not install. The session uses the fixed :insider_logger primary-filter id and permits one active session globally; an unrelated holder of that id is never removed.

Evidence and redaction

  • Emit partial redaction output. Insider.Redaction.redact/3 returns :redaction_failed with nothing else on invalid terms, exceeded limits, deadlines, or a malformed or forged policy. Undeclared secret data drops by default. Why: a partially redacted artifact is indistinguishable from a fully redacted one to its consumer.
  • Let an event choose its own classification. Recorder classification is a trusted binding supplied at construction; Event fields cannot select or downgrade it.
  • Truncate a capsule silently. Insider.Incident.Capsule rejects snapshots over 64 events, 32 artifact references, or 1 MiB. Why: a truncated capsule would verify as complete.
  • Wait on a stuck recorder, or let an application process wait at all. Insider.Recorder.record/2 never sends to the recorder process and is called only from collector processes. Writers that meet at its admission lock wait for their turn in arrival order and give up with {:dropped, :contended} when the lock has not changed hands for admission_stall_ms (default 1 s); it also returns {:dropped, reason} under pressure, rate limits, invalid input, closure, or recorder loss, and the snapshot keeps the gap. Why: the host application must never block on observation; a held-up collector backs up into its own bounded inbox.
  • Report an unavailable metric as zero. The load harness and poller use typed availability; a required metric that is unavailable fails the gate.

Static analysis and oracles

  • Treat a candidate as a defect. The inventory labels every result candidate and lists static_signals_do_not_confirm_runtime_causality. It never equates a source heuristic with runtime behavior.
  • Guess a dynamic identity. Rebound aliases, multi-alias forms, multiple-module files, dynamic event lists, and non-literal arguments become unresolved targets with oracle: null. Why: a wrong identity would send an oracle at the wrong state.
  • Follow symlinks, expand globs, or leave the root. Selectors are exact paths validated component by component with lstat; a symlink anywhere in a selector is selector_symlink; a path outside the root is selector_outside_root.
  • Load or execute target code. The scanner parses with Code.string_to_quoted/2 and never compiles, evaluates, or runs tests.
  • Intern caller input. All three oracle artifacts, the privacy corpus runner, and mix insider inspect resolve names only against atoms that already exist. Why: atom creation from untrusted input is unbounded.
  • Export the HMAC key or a configured value. Oracle reports carry presence and per-run fingerprints only; a value over 1,048,576 external bytes makes the key inconclusive rather than being serialized anyway.
  • Count silence as unchanged. If no INSIDER_*_ORACLE= line appears, the after-suite callback never ran. That is an infrastructure failure.
  • Accept a canary token that occurs nowhere. The privacy corpus runner fails closed on tokens absent from every item. Why: an unused token proves nothing and would inflate a clean result.

Sentinel

  • Mutate source, read a repository, or use the network. Every portable artifact records it: manifest and regression source_mutation=prohibited; bisection repository_read=none, checkout=none, git_authority=none, network_effect=none; handoff and proposal repository_mutation=none, network_effect=none. The core itself never invokes a shell, filesystem, network, node, container, or cloud API.
  • Let portable data select code. Scenario strings resolve only through a trusted contract-module allowlist; body modules, providers, and disclosure authorities come only from host configuration; candidate bytes are data and only a host-selected body interprets them. An external-process command's arguments must exactly equal the profile's declared public fingerprint projection.
  • Report cleanup it did not observe. A lease is complete only after independent owner inventory is empty; otherwise it is partial or uncertain and quarantined. A timeout cannot retract an inspection already delivered, so timeout results carry cleanup.status uncertain and say so.
  • Classify a regression from reused runs. Regression.classify/2 is inconclusive if either record is the embedded source run or shares its observation-run identity. Verification never reuses the handoff's failure runs.
  • Dispatch a candidate before the baseline fails. Verification.run/7 will not run post-repair attempts unless classify/2 reports failed.
  • Round up an inconclusive verification. A record with any reason is invalid, and the repair report keeps it visibly so.
  • Claim an earliest culprit or a cause. Bisection reports one adjacent boundary pair under the host's monotonicity assumption and records target_causality_is_not_claimed.
  • Create a draft without a configured authority. The draft-creation executor is disabled by default, has only test providers, requires a disclosure authority to approve the exact payload, and treats a provider timeout as provider_outcome_unknown. It never exposes merge or deploy.
  • Evict from the dedup ledger. A valid input that cannot be retained is capacity_refused and the ledger is unchanged. Why: silent eviction would turn a duplicate into new evidence.

Transport

  • Parse ETF or textual Erlang terms from untrusted input. The CLI accepts JSON or NDJSON only; canonical encoding is an output guarantee, not an input requirement.
  • Return an uncorrelated result. A dispatcher result whose protocol, request id, operation, or target does not match the request is an operation_failed error with cleanup.status uncertain.
  • Exceed its byte, frame, depth, or width limits. 16 MiB input and output, 1,000 frames, depth 8, width 1,000; violations are insider.cli.error/v1 with exit 64, or correlated operation_failed frames for oversized results in invoke_ndjson/2.

Packaging

  • Ship Lab in the production artifact. scripts/verify-production-artifact fails if any Lab-only source entry, module, or release application appears in the built release, and boots the release to prove the modules cannot load. test/production_artifact_boundary_test.exs also fails if mix.exs declares an escript, insider_lab, or a provider HTTP client.
  • Publish from an agent session. A maintainer publishes the first release by hand. Later releases need the release environment's reviewers and a maintainer's approval of the staged release on Hex.

If you need one of these behaviors, the answer is a new decision record and a reviewed change, not a workaround.