Coverage Analysis

Coverage Analysis

Purpose

Explain what .NET coverage evidence proves, reconcile target arithmetic, and identify the code blocking progress. Add complexity/CRAP ranking only when the user explicitly asks for risk hotspots, CRAP, priorities by risk, or refactoring safety.

When to Use

Use this skill for interpreting supplied .NET line/branch/condition evidence, coverage gaps and plateaus, target arithmetic, or explicit project-wide coverage-backed risk analysis.

When Not to Use

• Named method, class, or file CRAP/refactoring-safety analysis — use the crap-score skill instead
• Static source-to-test pairing or listing files with no tests — use find-untested-sources
• Behavioral or pseudo-mutation gaps in existing tests — use test-gap-analysis
• Test trait/category distributions or coverage shape by test type — use test-tagging
• Writing or generating tests — this skill identifies where tests are needed, not write them
• General test execution unrelated to coverage or CRAP analysis
• Only collecting .NET coverage or printing a raw percentage with no diagnosis — use run-tests; use native tooling for non-.NET coverage collection or analysis. Interpreting .NET line/branch gaps remains in scope here

Inputs

Discover optional inputs from the workspace. Do not ask for a project path when the current directory or a supplied report is sufficient.

Choose the smallest matching path

Words such as analyze coverage, what is blocking coverage, or why is coverage stuck do not by themselves request CRAP. Explicit signals include risk hotspot, CRAP, complexity-weighted priority, safe to refactor, or an equivalent request to combine complexity with coverage.

Existing-data fast path

When the user supplies a coverage excerpt, summary, or valid Cobertura path:

• Treat it as authoritative input and start there.
• Do not discover the solution or test projects unless source mapping is necessary.
• Do not run dotnet test, install ReportGenerator, add a coverage package, or read references/setup-discovery.md or references/report-generation.md.
• Do not write coverage-analysis.md or create a report directory unless the user requested a saved/full report.
• For interpretation and plateau questions, parse only the evidence needed to answer. For explicit project-wide risk requests, use the bundled scripts as described in references/guidelines.md.

A failed read/view operation is not proof that a named path does not exist. Classify the failure, then make one targeted existence probe and use a normalized path or alternate reader only for confirmed tool availability, transport, or path-normalization failures and only after verifying the canonical path remains inside the workspace. Stop on content-exclusion, permission/policy, workspace-boundary, or unknown failures. Report a missing path only when the independent probe also fails; do not broaden the search or substitute another artifact.

Collection path

Use this path only when no usable coverage evidence exists and the user asked for analysis that requires it.

1. Read references/setup-discovery.md.
2. Prefer existing Cobertura discovered under the requested root.
3. If none exists, invoke run-tests for repository overlay, platform, runner, and command selection, and have it collect Cobertura once per entry point.
4. Analyze the resulting Cobertura. Compute CRAP only if risk analysis was explicitly requested.

When the workspace contains the project/test manifests needed by this path, perform discovery directly and delegate collection to run-tests. Do not ask the user to attach files or pre-run coverage that the current workspace lets you inspect and generate.

Do not modify production code. Any incidental test-project change needed for collection is owned by run-tests; include that change and its revert command in the final analysis.

The automatic collection path is for SDK-style projects. For classic non-SDK or packages.config projects, use only a repository-owned coverage command. If none exists, stop and request Cobertura XML. Never migrate the project, inject an SDK-style provider, create a wrapper project, or report substitute coverage from another assembly.

Interpretation and arithmetic invariants

• A line hit proves execution, not both decision outcomes.
• condition-coverage="50% (1/2)" proves one reported outcome ran, but not which one. Recommend forcing the opposite outcome. Without source, never invent likely predicates or claim whether true or false is missing. State that compound predicates need independently exercised operands and short-circuit combinations when applicable; do not infer the exact combinations without source or fuller XML.
• Derive overall totals from Cobertura's covered/valid line counts. For target T, required covered lines are ceiling(valid lines × T).
• Projected coverage is (current covered lines + newly covered distinct lines) / valid lines. State assumptions such as fully covering a method.
• When asked whether one member can reach a target, show its maximum projected total and at least one concrete sufficient combination of supplied members or line gains. If no supplied combination is sufficient, say so.
• For a multi-member target, name the exact combination and resulting covered / valid total; saying only "combine it with another member" is incomplete.
• Reconcile member gaps against project totals. Method line ranges can overlap or omit class-level lines, so do not sum method counts as project truth.
• Never call one member the sole, entire, or all remaining gap unless its distinct uncovered lines exactly reconcile with the project total and no other below-threshold member remains.
• Name every supplied or extracted below-threshold member, but keep detail proportional: lead with the blockers, summarize the remainder in one sentence or a compact table.

Response contract

Answer the user's question in the first 2–4 sentences.

• Excerpt or arithmetic question: one explanation plus the next test or member priority. No dashboard.
• Existing-report interpretation or plateau: overall line/branch coverage, blocking members, reconciled target impact, and 1–3 recommendations. Use at most one compact table.
• Explicit risk/CRAP request: top 3 actual hotspots by default, supporting complexity/coverage/CRAP values, remaining flagged count, and 1–3 priorities. Exclude fully covered low-risk methods from the hotspot table. Never exceed 10 rows unless the user requests a larger count. For refactoring safety, rank risky methods by CRAP rather than raw coverage alone, then name comparatively safe well-covered methods separately.
• Explicit full report request: read references/output-format.md. Save the report only then.

Report only artifacts that exist. Do not announce inaccessible output paths or failed optional file writes when no file was requested.

Optional reports

HTML/CSV/markdown files are not part of normal analysis. Generate them only when the user explicitly requests report files or a CI artifact.

1. Deliver the direct coverage/risk answer first.
2. For a full markdown report, read references/output-format.md and save it.
3. For HTML/CSV, then read references/report-generation.md. Do not install ReportGenerator before the direct answer, and do not retry a failed install.

Validation

• Confirm a supplied report was used without test execution or tool installation.
• Reconcile covered, valid, and uncovered line totals before projecting impact.
• Confirm every stated blocker comes from supplied or extracted evidence.
• If CRAP was requested, spot-check one score using references/guidelines.md.
• If files were requested, verify they exist before reporting their paths.

Common Pitfalls

• Existing report triggers setup work — stop. Analyze the named artifact first.
• Generic analysis triggers CRAP — stop. CRAP requires explicit risk intent.
• One method is called the entire gap — reconcile distinct uncovered lines and the full below-threshold set first.
• A small request gets a dashboard — scale down to a direct answer or one table.
• No Cobertura from collection — report the collection failure; do not invent substitute coverage.
• Compiler-generated method names — verify suspicious async, lambda, or local function names against source before presenting them as user-authored members.
• Test exit code 1 — coverage may still exist; proceed with a warning. Other nonzero build failures stop the collection path.