sensez

name

sensez

description

Use the Sensez MCP server when checking a codebase for structural duplication, dead code candidates, circular imports, boundary violations, or design smells. Use it after you have completed a full edit turn to verify that newly written code is structurally correct. You can also use it on user triggers: "find duplicate code", "audit this project", "check dead code", "detect cycles", "enforce boundaries", "run Sensez", or "check this change". Not for per-file lint/type issues; use the language's linter and type-checker for those.

Sensez

Sensez finds project-level maintainability problems that linters and type-checkers usually cannot see across files. It currently has profiles for Python, JavaScript, TypeScript/TSX, and Rust:

Area	Output key	What to do with it
Duplication	duplication	Prefer reuse/extraction when the duplicated block is real code, not tests/fixtures.
Cycles	cycles	Treat as load-time architecture risk; break one import edge.
Boundaries	boundaries	Enforce the rule in sensez.toml; do not ignore unless the rule is wrong.
Dead code	dead_code	High confidence is actionable; Medium/Low need review.
Smells	smells	Advisory design pressure; lead with high-impact findings.

Use The MCP

Sensez is an agent-facing MCP integration. When this skill applies, use the available Sensez MCP tools directly. Always pass an absolute repository root as path unless the user explicitly asks for a partial scan.

MCP Tools

• noze_sniff: analyze duplication, dead code, cycles, boundary violations, and smells. Results are diff-focused by default so they show findings touched by current uncommitted changes while still analyzing the full graph. Pass diff=false for an explicit full-repository audit.
• get_configuration_summary: use only when the user asks to tune or adjust Sensez thresholds/configuration. Summarize the noisiest rules and offer concrete config changes before editing config files.
• noze_gate: end-of-turn quality gate for agent hooks. It is intended for hook integration, not ordinary direct calls during conversation.
• noze_explain: define a pillar or smell kind in plain English. Use it when you need to explain a finding category without guessing.
• brainz_report: summarize local-only Sensez usage metrics. Use when the user asks how Sensez helped, or before a commit.
• brainz_triage: record the user's explicit verdict on a finding. Never call it unless the user classifies a finding as debt, false positive, or cleared.

Running Analysis

Call noze_sniff with path set to the absolute repository root. Omit diff for the default diff-focused scan, or pass diff=false when the user asks for a full audit. Use limit to cap each pillar's returned findings when context is tight, and use threshold only when the user requests a different duplication token threshold.

After each edit turn, call noze_sniff to check that the code just written is structurally sound. Report only findings that touch the current change unless the user asked for a full audit.

Reading Results

Important meta fields:

• mode: full or diff.
• *_total: true counts before limit truncation.
• boundaries_configured: false means boundaries were not checked.
• external_edges vs internal_edges: a very high external ratio may mean roots/imports were not resolved well.

Dead-code confidence:

• High: module is reachable and the symbol is not referenced. Usually safe to remove after a quick domain check.
• Medium: module is imported plainly; attribute access may hide usage.
• Low: module itself is not imported. Treat as possible script/framework entrypoint.

Do not present Medium/Low as guaranteed deletion work. Phrase them as review candidates.

Agent Guidance

Lead with reliable signals: boundary violations, cycles, high-confidence dead code, and large duplication classes. For smells, explain why the finding matters in this codebase before recommending a refactor.

Prefer adding framework entrypoints or boundary rules to Sensez configuration over suppressing individual findings in source. Use get_configuration_summary when the user explicitly asks to tune noisy rules.

Do not use Sensez for unused local variables, per-file unused imports, formatting, or type errors. Those belong to tools like ruff, ESLint, TypeScript, ty, mypy, or the language compiler.