Invariant Tools

Why this matters

AI agents write code confidently but frequently drift from the stated goal — adding unrelated changes, deleting important functions, or introducing subtle dependency shifts. Static analysis catches syntax errors but can't evaluate intent. Invariant bridges the gap: it uses LLM inference to understand what the code does semantically, then uses deterministic symbolic reasoning to verify whether that matches what was intended. Add a few lines to your agent's system prompt and every code change gets verified before it's presented as complete.

Capabilities

• ✓invariant.code_lens — Extract structural and semantic facts from source code: functions, calls, dependencies, intent, side effects, and patterns
• ✓invariant.code_query — Run predefined or custom queries over lensed facts: orphans, test gaps, dependency cycles, intent mismatches, security concerns, hotspots
• ✓invariant.diff_analyzer — Compare code before and after changes against a stated goal. Returns alignment score, unexpected changes, concerns, and actionable suggestions
• ✓invariant.review — Comprehensive code review combining diff analysis, Prolog-based fact queries, and criteria/constraint evaluation into a single structured verdict with pass/fail determination

Agent feedback loop

The most powerful use of Invariant is as an automatic verification step in an agent's coding loop. After making code changes, the agent calls invariant.diff_analyzer with its stated goal. If the alignment score is low or unexpected changes are flagged, the agent revises before presenting its work. This creates a neuro-symbolic feedback loop: fuzzy reasoning writes the code, deterministic reasoning verifies it, and fuzzy reasoning interprets the feedback to self-correct.