Identifies natural component boundaries inside a monolith by clustering the dependency graph, finding the cuts with minimum coupling. Use when planning to modularize or extract microservices, when deciding what can be deployed independently, or when the user asks where the seams in this codebase are.

Interprets and explains counterexamples produced by model checkers or property-based testing tools to make them actionable. Use when TLC, NuSMV, CBMC, or a property-based test emits a counterexample the user doesn't understand, when a trace is too long to read, or when mapping a model-level trace back to source code.

Rewrites vague or incomplete requirements into precise, testable statements — filling in quantities, actors, conditions, and error behavior while preserving intent. Use after ambiguity-detector flags problems, when a requirement can't be turned into a test, or when engineers keep asking the same clarification questions.

Detects inconsistencies across configuration files, environments, and deployment manifests — missing keys, drifted values, type mismatches. Use when debugging why staging behaves differently from production, before a deploy to catch config drift, or when auditing multi-environment configs.

Translates pseudocode into idiomatic Python, choosing the right standard-library structures and leveraging Python idioms that pseudocode doesn't express. Use when implementing an algorithm from a paper or spec, when the user hands you pseudocode and wants Python, or when realizing a verified-pseudocode artifact.

Translates natural-language requirements into TLA+ properties — invariants for safety, temporal formulas for liveness — checkable with TLC. Use when writing the PROPERTY and INVARIANT sections of a TLA+ spec, when formalizing acceptance criteria, or when the user has a requirement and a model but no property.

Analyzes a codebase to quantify and locate technical debt — where it lives, what it costs, and what order to pay it down in. Use when planning a refactoring sprint, when justifying engineering time to stakeholders, when the user asks where the codebase hurts most, or when onboarding to a legacy system.

Shrinks a failing test input to its minimal form while preserving the failure — delta debugging and structured shrinking to find the smallest input that still triggers the bug. Use when a fuzzer or property test finds a failure with a huge input, when a bug report has an unwieldy reproduction, or when you need a minimal test case for a regression suite.

Summarizes undocumented legacy code by inferring intent from structure, naming, data flow, and calling context — explicitly flagging what's inferred vs. what's certain. Use when onboarding to inherited code, when documentation is missing or wrong, or when deciding whether legacy code is safe to change.

Finds code by meaning, structure, or text across large codebases — picks the right search strategy (grep, AST query, call graph walk, semantic search) for the question being asked. Use when the user asks where something is implemented, when navigating unfamiliar code, or when a simple grep isn't enough.

Traces regressions to the specific commit, change, or code path that introduced the behavioral breakage. Use when a previously passing test or feature now fails, when the user asks what change caused a regression, or when bisecting a regression across commits.

Reduces a TLA+ model so TLC can actually check it — shrinks constants, adds state constraints, abstracts data, or applies symmetry — when the state space is too large to enumerate. Use when TLC runs out of memory, when checking takes hours, or when a spec works at N=2 and you need confidence at larger scale.

Writes documentation for test cases — names, docstrings, and comments that explain what behavior is being tested and why, so a failing test tells you what broke without reading the assertion. Use when test names are test_1 through test_47, when tests fail and nobody knows what they mean, or when onboarding needs a readable test suite.

Generates pytest regression tests that capture current behavior as snapshot assertions, using Python's dynamism for low-friction recording. Use before refactoring untested Python, when the behavioral spec is "whatever it does now," or when migrating Python 2→3 or between framework versions.

Compares two versions of a requirements document and reports additions, removals, semantic changes, and scope drift — distinguishing clerical edits from meaning changes. Use when a spec was revised, when checking if a new version of a standard affects you, or when the user asks what changed between spec versions.

Transforms a changelog or commit range into user-friendly release notes with highlights, upgrade guidance, and narrative framing. Use when publishing a release announcement, when the changelog is too dense for users to read, or when the user needs a blog-post-shaped summary of a version.

Assists migrating a build or CI pipeline from one system to another — Jenkins to GitHub Actions, Travis to GitLab CI, Makefile to Bazel — preserving semantics and surfacing untranslatable constructs. Use when switching CI providers, when modernizing a legacy build, or when the user pastes a Jenkinsfile and asks for the GitHub Actions equivalent.

Builds a bidirectional traceability matrix linking requirements to design elements, code, and tests — so every requirement traces forward to its implementation and every test traces back to its justification. Use for compliance audits, when answering why a piece of code exists, or when checking that nothing was built without a reason.

Identifies recurring structural patterns in a codebase — idioms, copy-paste clones, homegrown abstractions — and characterizes each as a reusable template. Use when learning a codebase's conventions, when hunting for copy-paste that should be a function, or when documenting how this team does things.

Translates cryptic runtime error messages and stack traces into understandable explanations, pointing to the concrete line at fault and the most likely fix. Use when a user pastes an error they don't understand, when a stack trace is deep and the user doesn't know where to start, or when an error message misleads about the real cause.