Everything below runs on a development machine or CI host and touches nothing outside the repository and the target checkout you point it at. Use a supported Elixir/OTP pair: 1.17/27, 1.19/28, or 1.20/29. The maintainer-only analysis tools need Elixir 1.18 or newer; mix docs is a dev-only dependency like them.

Set up and run the quality gate

mix setup      # deps.get, then git config core.hooksPath .githooks
mix quality    # format, warnings-as-errors, full unit suite, prod compile,
               # production-artifact gate, Credo + ex_slop

mix quality compiles the root application in test and prod, runs the unit suite, runs scripts/verify-production-artifact against the fresh prod build, and finishes with strict Credo. mix quality.deep adds ex_dna and Dialyzer. mix test --cover enforces the 90% coverage floor and writes per-module reports to cover/.

Lab and Load are separate Mix projects that share the root lockfile:

(cd tools/insider_lab && mix deps.get --check-locked && mix test)
(cd tools/insider_load && mix deps.get --check-locked)

Generate this documentation

mix docs           # writes doc/index.html for the root application and guides
(cd tools/insider_lab && mix docs)   # Lab-only API reference in tools/insider_lab/doc/

Look for test-isolation candidates in another project

Compile Lab once, then point the inventory at a test tree. Compile first so compilation output cannot land in the JSON you parse:

(cd tools/insider_lab && mix compile)
(cd tools/insider_lab && mix insider.candidates --root /path/to/project --path test)

The command prints one JSON object and exits 0 (complete scan, no candidates), 2 (complete scan, candidates present), 3 (no _test.exs file in the selection), 4 (a file could not be analyzed), 5 (a scan bound was hit), 64 (invalid arguments), or 70 (inventory could not complete). Only 0 means clean. A candidate is a static hint ranked by weighted signals such as async: false, application-environment mutation without visible cleanup, VM-global Telemetry handlers, OpenTelemetry global capture, and timer-based synchronization. It is not a confirmed defect. The full contract is in mix insider.candidates.

Confirm a candidate with an oracle

Each candidate target may carry an oracle hint naming one of three fixed artifacts in tools/insider_lab/priv/oracles/. To use one, add a single Code.require_file/1 line to the target project's test/test_helper.exs after ExUnit.start/1, run exactly one selected test command with the named environment variables, read the one INSIDER_*_ORACLE= line it prints, then restore the helper to its exact prior contents. For example, for an application_env target on logger / default_handler:

INSIDER_ORACLE_APPLICATION=logger \
INSIDER_ORACLE_KEY=default_handler \
mix test test/path/to/selected_test.exs

The report says whether the key was changed, unchanged, or inconclusive between before-suite and after-suite captures, using per-run HMAC fingerprints and never the configured value. A state change does not by itself establish causality; the agent or engineer still owns that inference. Details, bounds, and failure modes for all three artifacts are in Oracles.

Inspect a process in the same BEAM

(cd tools/insider_lab && mix insider capabilities)
(cd tools/insider_lab && mix insider inspect --name Elixir.MyApp.Supervisor)

inspect starts the current Mix project, resolves the registered name in the same VM, and returns a bounded process.describe result: status, reductions, memory, heap sizes, queue length, registered name, and current function. It refuses textual PIDs because a PID copied from another BEAM can alias an unrelated local process. In IEx the same operation is Insider.Runtime.describe(pid, [environment: :development], host_policy).

Run the load harness

cd tools/insider_load
mix insider.load --profile config/load.json \
  --output ../../_build/insider_load/evidence/load
cd ../..
bash scripts/verify-load-evidence _build/insider_load/evidence/load

The harness compares a fixed-seed synthetic workload without and with a collector session and writes a canonical report.json, profile.json, and manifest.json to a new output directory. The verify script recomputes the digests and identities offline and prints one JSON verdict line per directory. Every report records production_capacity=false and budget_evaluation=not_approved; these are development and CI measurements, not budgets.

Run the privacy corpus example

(cd tools/insider_lab && mix insider.privacy_corpus \
  --corpus priv/privacy_corpus/example.json --environment production)

The example corpus has fake values and exists to show the schema. The report names which fields the production ingress profile would retain and where any declared canary token surfaced; it never prints retained values. Exit 0 means no declared canary appeared in projected output; exit 2 means one did. A clean exit is bounded evidence about the supplied items, not approval.