Upgrade Check

Test whether application upgrades preserve existing users' saved data. upgrade-check guides coding agents through real baseline-to-candidate upgrade tests, including persisted records, settings, boundary values, relationships, and restart behavior. It helps identify regressions that fresh-install testing can miss.

What your agent checks

• Real upgrade paths: create and save state with the baseline, then run the candidate on that state.
• Existing data: compare meaningful records, settings, relationships, and supported actions.
• Boundary values: exercise zero/falsy, empty, enum endpoints, and other values the baseline can produce; cover every value when the domain is small.
• Regressions: compare explicit expected and observed behavior, allowing documented transformations.
• Restart persistence: preserve edits and check at least two normal process starts after the initial upgrade.
• Evidence-based reports: record exact versions, commands, observations, and PASS, FAIL, or NOT TESTED outcomes.
• Disposable environments: use isolated paths and synthetic test data.

Requirements and limitations

This is a set of instructions for an agent with filesystem and shell access. The agent must be able to build or run your baseline and candidate applications and isolate their data and service endpoints. Missing prerequisites are reported as NOT TESTED.

It is not a standalone testing application or a guarantee that every migration defect will be detected. A PASS applies only to the versions, state, and behaviors exercised. Concurrent writers, interrupted migrations, and downgrade compatibility require separate application-specific tests. Use synthetic disposable data, not production databases or private user data.

The ZIP contains SKILL.md instructions and the unchanged MIT LICENSE, with no executable dependencies. Evaluation fixtures, the demonstration harness, and repository access are not included.

MIT licence and package contents

MIT License, Copyright (c) 2026 Figo5. The original MIT notice is included unchanged. The development repository is private; the purchase includes the packaged instructions and licence, not access to the repository.

This package remains MIT-licensed. MIT permits use, modification, redistribution, and commercial use subject to its notice requirements. Making the repository private does not revoke MIT rights already granted.

Local validation

The creator's separate deterministic notebook fixture was tested locally on Windows with Node 24.19.0 and SQLite 3.53.3. The good upgrade passed; the planted broken upgrade changed Sunday to Monday and was detected despite unchanged counts. All seven week-start values were checked for each candidate (14 cases). In the main Sunday scenario, records, relationships, and a Unicode note edit persisted through two explicit restarts per candidate.

These are synthetic fixture results. They do not establish better defect-detection accuracy than an agent without the skill or validate every application. Marketplace runtime delivery has not been tested.