Problem-solving guide

How to record evidence for AI coding work

Another person can review an AI coding result more easily when they can see what was requested, what changed, which checks ran, and where uncertainty remains.

Evidence is not a guarantee: a recorded result helps inspection and reproduction. It does not guarantee correctness, security, or production readiness.

Minimum evidence

  • Task and repository commit or baseline.
  • minitok version and configured provider.
  • Workflow stages and changed files.
  • Verification command and test, lint, or build results.
  • Retries, repair attempts, stop reasons, manual interventions, and limitations.

Inspect a run

minitok doctor
minitok status
minitok run list
minitok run show <run-id>

If a patch was not applied, preserve and inspect .minitok/last-run.patch.

Share safely

The CLI writes a redacted latest run summary to .minitok/last-run.json. Structured evidence artifacts, when enabled by the run path, are stored under .minitok/evidence/runs/. A preserved .minitok/last-run.patch is a recovery artifact for an unapplied change when that file exists.

  • Remove API keys, bearer tokens, passwords, private keys, customer data, and proprietary source.
  • Label illustrative fixtures separately from real execution evidence.
  • Do not publish benchmark claims without repeated comparable runs and raw support data.
  • State provider, version, repository, cost, and manual intervention context.

Data boundary

minitok is local-first and telemetry depends on plan and consent. Repository content is sent to the configured model provider when a model request is made. See the privacy policy for current terms.