logicprobe

Design documents say "thread-safe", "max 3 retries", "no deadlock possible" — but is the code actually doing that? Logic Probe bridges the trust gap between documentation and implementation. It enumerates every verifiable claim in design docs, architecture specs, and refactoring plans (API names, file paths, enum values, mechanism feasibility), cross-checks each against the actual codebase, and produces file:line-precise evidence with severity grading and remediation direction — never modifying code during verification.

When a claim involves behavioral logic, the plugin extracts a state-transition model, presents it to the user for confirmation, then runs 22 automated checks: 8 structural (reachability, deadlock, liveness, determinism, event and guard completeness, invariant validity, monotonic variables) plus 14 adversarial probes (race interleavings, order reordering, boundary bombardment, pairing symmetry, idempotent replay, probabilistic reachability lower bounds, worst-path cost with positive-cost cycle detection, deadline bounds, and more). Refactoring workflows add before/after model diffing to flag undeclared behavioral changes; data-migration reviews verify copy consistency, migration coverage, and breaking-change regressions. The engine can also export LogicModelV1 as native input for UPPAAL, TLA+, PRISM, and SPIN, mapping one-to-one with the external-tool routing dimensions in coverageNotes.

Logic Probe is built for engineers and tech leads who need evidence over intuition during code review, architecture design, or refactor planning. It follows the Agent Skills open standard and is validated across Claude Code, Codex CLI, Cursor, Kimi CLI, OpenCode, ZCode, and DeepSeek Harness. The Python engine is optional and provides an offline fallback with zero hard dependencies.