Skip to content
OpenAgentSkill
Browsing is open · Catalog captured 2026-11-21 · Automatic updates and account services are paused.
Development

test-value

Improve test value and feedback cost when auditing slow, flaky, redundant or brittle tests; choosing tests during implementation; or reviewing test changes before a commit or PR. Keep ordinary implementation and pre-commit use scoped to the affected behavior, not a repository-wide audit.

tt-a1i · Static checks on file

Installation

Read the source and dependencies before running a command. Copying does not install or execute anything.

npx skills add tt-a1i/archify --skill test-value

About this skill


name: test-value description: Improve test value and feedback cost when auditing slow, flaky, redundant or brittle tests; choosing tests during implementation; or reviewing test changes before a commit or PR. Keep ordinary implementation and pre-commit use scoped to the affected behavior, not a repository-wide audit. metadata: internal: true

Test Value

Protect outcomes worth preserving at a cost proportionate to their risk. Test counts, coverage percentages and green runs are evidence, not substitutes for product judgment. When an AI writes both implementation and tests, check whether they share the same mistaken assumption.

Choose the scope

Preserve the user's existing authorization and operational boundaries.

  • Audit existing tests: investigate cost, false alarms and protection within the requested scope. An audit request is read-only; an optimization request permits scoped changes. Read audit.md.
  • Implementation or pre-commit review: examine the behavior being changed and its affected tests. Decide whether to add, adapt, reuse or remove tests. Reviewing before a commit does not itself authorize committing or pushing.
  • Diagnose a failure: identify the outcome the failing check protects, then distinguish a product regression, obsolete assertion, environment failure or unstable test. Keep a narrowly requested fix narrow.

Read decision-examples.md when dealing with automatic-layout snapshots, shared matrices, source assertions, asynchronous waits, test doubles, benchmarks or coverage gaps.

Establish an independent basis for correctness

Explain what must remain true using user requirements, public contracts, compatibility commitments or a real failure. The current implementation and its own generated output cannot, by themselves, establish the correct expectation.

For tests being added or changed, answer three questions in the reasoning or a short change explanation:

  1. Which concrete mistake would fail this test, and what loss would it cause a user or maintainer?
  2. Would an equivalent refactor or a better valid result also fail? Which restrictions are actually contractual?
  3. Which existing test, type check or other check already catches this mistake? Does the additional protection justify its execution and maintenance cost?

Prefer observable output, state, data integrity and failure behavior. Private-function tests, snapshots and static checks can also protect real boundaries; their form alone does not determine their value.

Make the smallest effective testing decision

EvidenceDecision
A worthwhile failure boundary is uncoveredAdd a regression at the cheapest layer that exposes it; retain a real integration path where needed
An existing test already fails for the mistakeReuse or extend it instead of building a parallel fixture family
A test freezes noncontractual implementation or automatic outputAssert outcome properties, justified tolerances or boundaries while preserving explicit constraints
Several tests protect the same failure through the same implementation pathConsolidate or remove duplicates and identify the surviving protection
A shared rule table is repeated across entrypointsCover the rule set through shared logic and verify each entrypoint's integration; retain combinations with interactions or independent implementations
A costly test protects a distinct riskPreserve coverage; improve fixtures or select execution by risk
A historical contract has explicitly ended or a capability was removedRetire obsolete tests after checking for remaining consumers
Ownership of protection or the correct expectation is unclearRetain the test and identify missing evidence rather than guessing to reduce failures

Adding a test for every edit is not the objective. Pure copy changes, mechanical edits and changes already covered effectively may need no new test. When the user authorizes removing a product restriction, update the tests too; an old assertion does not veto an approved product change.

Preserve effective protection against data loss or corruption, permission and trust-boundary failures, promised compatibility, essential accessibility failures, resource leaks and broken recovery. Simplifying these checks still requires evidence of surviving protection. Bulk skips, retries, wider tolerances or refreshed snapshots must not conceal regressions. Changes to product commitments or mandatory release-platform gates must be within the user's authorized scope.

Check the test's ability to discriminate

Use evidence proportionate to the risk instead of creating another test suite for every test:

  • Bug fix: where practical, reproduce the original failure with the regression and verify the fix passes. State the gap if reproduction is unavailable.
  • Relaxed assertion: try a valid alternative and the original bad result; the former should pass and the latter fail. Temporary fault injection can help for consequential errors; a mutation score is not an end in itself.
  • Consolidation or deletion: trace actual call paths and establish that surviving tests detect the same mistake. Similar names, a shared helper or overlapping line coverage do not prove duplication.
  • Shared fixtures: reset mutable state and retain isolation and cleanup assertions between failure attempts. Gains must not depend on test order or accidental cache hits.

Run destructive probes in temporary copies or controlled fixtures, preserving existing user changes. An unavailable environment is a validation gap; simulated evidence cannot stand in for a real browser, operating system or distribution package.

Control feedback cost

Read the project's actual test commands, CI selection rules and current changes before selecting the smallest set of checks that covers the risk. Affected behavior determines whether integration, browser,

Read the complete source ↗