Core
Use this when the review needs the general debt model and no language-specific lens.
Read core docsDebt skills
The Vølven skills teach agents to find debt pressure: hidden contracts, brittle tests, unsafe dynamic behavior, deploy friction, upgrade friction, and trapped knowledge.
Skill map
The core skill defines what counts as debt. The language skills translate that model into the places each ecosystem tends to hide cost.
Use this when the review needs the general debt model and no language-specific lens.
Read core docsWeak arrays, unsafe mixed, magic workflow strings, misleading PHPDoc, and shallow tests.
Observers, jobs, migrations, validation drift, policy drift, factories, and lifecycle traps.
Read Laravel docsImplicit schemas, mutable globals, swallowed exceptions, async drift, and scripts that became systems.
Read Python docsUnsafe casts, trusted runtime data, duplicated state, React business logic, and config drift.
Read TypeScript docsHidden object shapes, async mixing, mutable shared state, utility drift, and framework magic.
Read JavaScript docsHow to use them
Vølven works best after normal checks have passed. Let formatters, type checkers, tests, and framework tools clear the cheap noise. Then ask what still makes the next change expensive.
Use the core skill for general debt review. Add the language skill when the risk lives in local framework or runtime behavior.
Focus on the code that moved and the contracts around it. Broad archaeology belongs in a separate task.
Every finding needs evidence, future cost, the smallest useful fix, and the case where ignoring it is acceptable.
Use the Pest plugin for deterministic PHP signals so CI blocks growth while the skill handles judgment-heavy review.
Prompt shape
Good debt review prompts are narrow. They tell the agent what changed, which lens to use, and which kinds of findings are out of scope.
Review this diff with the Vølven PHP debt skill. Focus on future maintenance cost, hidden contracts, weak runtime boundaries, misleading tests, and suppressions. Ignore style, naming taste, and static-analysis duplicates. Findings first. Max 7 bullets.