Affected Verification

Original AI-assisted instructions for Jordan’s collection. Companion helpers have automated fixture tests; agent-host effectiveness remains experimental.

When to use

Use after a scoped change or before handing work to an integrator. Prefer the smallest check set that proves the changed behavior while preserving mandatory repository and release checks.

When not to use

Do not replace required integration checks with a convenient unit test. Do not add new frameworks or run production mutations to verify an unrelated edit. A command runner inherits the caller’s authority; it is not a sandbox.

Procedure

Choose checks from the actual change

Read the baseline, diff, project manifests, and acceptance criteria. Use existing dependency or affected-target tooling when available. Connect each criterion to a real check; explain why shared contracts, migrations, lockfiles, or build wiring require wider coverage. Do not guess scripts or run every package by default.

Execute without losing the exit result

After inspecting the command and its side effects, run it directly as an argument array through the helper. Store output outside the target repository:

node tools/skillcheck.mjs run --repo ../target-repo --id search-tests --timeout-ms 60000 -- node --test tests/search.test.mjs > ../evidence/search-check.json

Create the evidence directory first. The runner does not invoke a shell, does not approve the command, and does not hide nonzero exits behind a display pipeline. It captures exit code, signal, elapsed time, bounded stdout/stderr, full-stream digests, runtime, and before/after Git-visible fingerprints.

Receipt outcomes are passed, failed, blocked, timed-out, or invalidated. A missing executable is blocked. A killed command does not pass. Exit zero with changed source is invalidated. Human-report categories may also include skipped or not-run, but those never become passing receipts. A printed success message cannot override the real process result.

Reconcile what was actually proved

The helper deliberately leaves cache status unknown: a successful command can have used its own cache. Check the tool’s output before calling a run fresh. Record skipped or unavailable checks with reasons. Preserve failure evidence before repairing a defect.

Before handoff, verify each passing receipt against the current state. If ignored configuration, external services, environment variables, or cache state changed, reevaluate even when the Git fingerprint matches. For long-running or writing tests, choose an isolated fixture and explicitly describe outputs.

Output

Return the impact analysis and selection rationale, baseline and final revision/state, one row per check, receipt locations, cache information actually observed, excluded checks, and a scoped conclusion. Distinguish a passing focused check from release readiness.

Failure handling

Stop claiming completion at the failed boundary, not necessarily all useful work. Reproduce a failing check before changing its assertion. Do not automatically retry until a pass hides flakiness. Run formatters before final verification because source changes invalidate earlier receipts. Review artifacts for secrets before sharing them.

Example

The command prints PASS and exits 7. The runner emits status failed and exits nonzero. A different command exits 0 but edits a tracked source file; its receipt is invalidated. Both cases are exercised automatically rather than left as prompt instructions.

Companion tools

Runnable helpers and input formats · Evaluation method and limitations. Commands above run from a full toolkit checkout; they are not standalone host-installation instructions.

Check the gate itself

Include one controlled defect and confirm the intended assertion or compatibility rule fails, then confirm the corrected case passes. Tool failure, missing dependencies and invalid input are not the desired regression signal.

Bind receipts to the accepted source and configuration. Do not drop failed/unfinished attempts from canary reports or equate independent green branches with a validated integrated result.