Swe:test Gap Hunt

SWE Test Gap Hunt

What This Skill Does

Use this skill to run a bounded, evidence-led pass over a repository's tests.

The job is to:

• Detect the existing test and coverage workflow
• Rank the best test-improvement opportunities
• Add or strengthen a small number of high-value tests
• Improve weak test structure when it materially improves confidence or speed
• Leave a ranked backlog for the next incremental pass

This skill is language agnostic. Reuse the repository's own tooling, file layout, and test conventions instead of forcing a new framework.

When To Use

Use this skill when the user wants to:

• Improve coverage incrementally without broad rewrites
• Find the most important uncovered or weakly tested code paths
• Strengthen fragile, assertion-light, or over-mocked tests
• Add missing edge, error, branch, or boundary cases
• Run recurring or scheduled test-improvement passes over time

Do Not Use

Do not use this skill for:

• First-time test framework selection or migration
• Large production refactors done in the name of testability
• Coverage-number campaigns with no quality bar
• Pure performance profiling that is unrelated to test speed or test structure
• Broad QA or bug hunting without a test-improvement goal

Inputs To Confirm

Confirm or infer:

• Repository or package scope
• Whether the user wants execution or report-only output
• Runtime budget for validation and coverage commands
• Any no-touch areas or flaky suites to avoid
• Whether recent CI, coverage, or bug history is available

If the user does not specify a budget, default to a small incremental pass:

• rank the workspace
• improve the top 1-3 opportunities
• stop after targeted validation

Parallelization Rule

Create one session per cleanly separated package, module, or service when the environment supports parallel agent work.

• Run only on disjoint write surfaces
• Cap concurrency to a level the repo can validate safely
• Give each session a bounded target and explicit test files or directories
• Have each session return raw evidence, changed files, and validation output
• Keep prioritization, deduplication, and final validation in the parent session

If parallel sessions are unavailable, process the same opportunities in local batches and keep the same output shape.

Instructions

Step 1: Detect The Existing Test Ecosystem

Inspect the repository before proposing changes.

Identify:

• Test runners and entry commands
• Coverage tools and report formats
• Existing test file layout and naming patterns
• Mock, fixture, factory, and helper conventions
• CI commands or package scripts already used for tests

Do not introduce a new framework unless the user explicitly asks for it.

Tooling Stance

This skill is tool agnostic.

Use the strongest available local conventions for the detected ecosystem, such as:

• vitest, jest, mocha, or ava
• pytest, unittest, or nose
• go test
• cargo test
• rspec, minitest, or bundle exec rspec
• coverage outputs such as lcov, coverage.py, pytest-cov, nyc, gcov, jacoco, or CI-native summaries

Prefer the repository's own scripts and CI commands over ad hoc commands.

Step 2: Establish A Baseline

Run the narrowest useful existing commands to confirm the current state.

Capture:

• Whether the relevant baseline tests already pass
• Which coverage signals are available and trustworthy
• Which suites are flaky, missing, or too slow to rely on

If the baseline is already red for unrelated reasons, say so directly and avoid claiming the skill completed a safe improvement pass.

Step 3: Identify And Score Opportunities

Rank opportunities using concrete evidence, not raw coverage percentage alone.

Consider:

• Low or missing coverage on exposed, central, or high-churn code
• Recently changed code with little nearby test support
• Existing bug fixes with missing regression tests
• Weak tests that create false confidence
• Validation cost and expected runtime impact

Step 4: Inspect Existing Tests For Weakness

Treat weak tests as real gaps even when line coverage is technically present.

Look for:

• Assertion-light tests
• Duplicated happy-path tests with no edge or error coverage
• Snapshot-only coverage of meaningful behavior
• Brittle global fixtures or hidden shared state
• Over-mocked tests that verify calls instead of outcomes
• Slow tests that could be replaced by cheaper coverage at a lower seam

If a test is weak but still the best local starting point, augment it instead of creating duplicate coverage elsewhere.

Step 5: Choose A Small Incremental Batch

Pick the smallest high-value set of opportunities that can be improved safely in one pass.

Prefer:

• missing branch, error, boundary, and regression coverage
• augmenting existing test modules before creating many new ones
• improvements that reduce brittleness and runtime cost together
• helper, fixture, or factory cleanup only when it directly supports stronger tests

Avoid:

• padding coverage with trivial assertions
• huge fixture abstractions up front
• production changes unless a tiny testability seam is required

Step 6: Add Or Strengthen Tests

When implementing tests:

• Prefer behavior-focused assertions over incidental implementation checks
• Use real collaborators when they are cheap and deterministic
• Use fakes, stubs, or mocks at external boundaries only where they reduce brittleness
• Keep fixtures explicit, local, and easy to read
• Optimize for fast repeatable validation, not maximum abstraction

If a tiny production seam is necessary, keep it minimal and explain why the test could not be written cleanly without it.

Step 7: Validate Narrow To Broad

Validate the touched surface with the narrowest useful commands first, then expand only if the result strengthens confidence materially.

Prefer:

• targeted test files or directories
• touched-package coverage commands
• broader suite checks only when the local pass is green and the repo pattern makes it practical

Do not rerun the entire workspace unnecessarily.

Step 8: Leave A Residual Backlog

At the end of the pass, leave the next ranked opportunities for the following loop or scheduled run.

If the user asked for report-only output, stop after ranking and validation planning.

Output Requirements

Produce a report with these sections:

1. Scope reviewed
2. Detected test tooling and baseline
3. Ranked opportunities
4. Execution plan
5. Changes applied
6. Validation run
7. Remaining backlog

For each ranked opportunity, include:

• Area or surface
• Why it matters now
• Whether the gap is missing coverage, a weak test, or both
• Expected files or test modules to touch
• Validation strategy
• Priority: high, medium, or low

For each applied change, include:

• Files touched
• What test gap or weak-test issue was addressed
• Whether mocks, fixtures, or helpers were introduced or tightened
• Validation commands and outcomes

If no safe improvement should be made, say so directly and return only the ranked backlog.

Examples

Example 1

User says: Find the top test coverage gaps in this workspace and fix the best two without making a huge mess.

Actions:

1. Detect the repo's existing test and coverage commands
2. Rank the best opportunities using coverage and test-quality evidence
3. Improve the top two disjoint surfaces
4. Run targeted validation and leave the next backlog

Result: The user gets a small, validated batch of test improvements instead of an unbounded coverage project.

Example 2

User says: Run a weekly pass that strengthens weak tests and adds missing edge cases, but keep it fast.

Actions:

1. Inspect the current test layout and identify weak tests
2. Pick a small batch that improves confidence without broad rewrites
3. Prefer local fixtures, narrow validation, and fast-running tests
4. Return the remaining backlog for the next scheduled pass

Result: The workflow behaves like an incremental maintenance loop rather than a one-shot campaign.

Troubleshooting

Problem: Coverage Artifacts Are Missing Or Stale

Use the repository's existing test layout, recent changes, and CI signals to rank opportunities, but say clearly that the ranking had weaker coverage evidence.

Problem: The Baseline Suite Is Already Failing

Call that out and avoid claiming the pass is fully validated. Limit work to report-only mode or a tightly scoped area if the failures are clearly unrelated.

Problem: The Fastest Coverage Win Requires Heavy Mocking

Do not take the easy percentage win. Prefer a smaller, more behavior-focused test or leave the item in the backlog.

Quality Bar

• Be evidence-led and language agnostic
• Reuse the repository's local test conventions before inventing new ones
• Prioritize risk reduction and weak-test repair over raw coverage gain
• Keep each pass small enough to run repeatedly or on a schedule
• Use parallel sessions only on cleanly separated write surfaces
• Optimize for readable, maintainable, and fast-running tests