claude-code-plugin

sparda-gate — the behavior gate for AI edits (Claude Code plugin)

Your agent just edited a route. Did it quietly replace an auth wrapper with an identity function, drop a route, or grow a write's blast radius? This plugin runs sparda gate after every Edit/Write/MultiEdit and tells the agent — in under two seconds, deterministically, offline, with no API key — the moment the regression lands, so it fixes it in the same session instead of it slipping into a commit. (For a hard pre-commit / CI block, the same engine runs as sparda apocalypse in a git hook or the GitHub Action.)

✗ SPARDA GATE — this edit changed the app's proven behavior (1.6s, deterministic):
  [critical] GUARD_REMOVED — POST /api/links was guarded in the baseline and is
             now reachable without any guard (app/api/links/route.ts:56)
  → fix the edit or, if intended, accept it with `sparda gate --arm`.

It is built to disappear when there's nothing to say:

• Silent when clean. No output, no interruption on a benign edit.
• Never fights your in-progress work. A route that vanished only because its file doesn't parse yet (the normal mid-multi-edit state) is held, never blocked — the gate abstains until the file parses again.
• Delta-only. It reports what this edit changed, never pre-existing findings.
• Zero config. The first run arms a baseline from your current tree; every edit after is proven against it. Re-arm intended changes with sparda gate --arm.

Install

/plugin marketplace add zyx77550/sparda
/plugin install sparda-gate@sparda

Requires Node ≥ 18. The hook shells out to npx sparda-mcp gate --hook; nothing is installed globally and nothing leaves your machine.

Monorepo: point the gate at the app dir by editing the hook command to npx -y sparda-mcp gate --hook --dir apps/web.

What it proves (and what it doesn't)

sparda gate compiles your backend (Express · Next.js · NestJS · Medusa · FastAPI) to a deterministic behavior graph and diffs it against the armed baseline. It proves the delta: no guard removed, no route dropped, no blast radius grown, no new unguarded mutation introduced. It is honest about coverage — a guard hidden in an unresolved imported helper is a blind spot, reported as such, never silently passed as safe. It does not replace review or tests; it catches the one class of regression an LLM cannot reliably catch in itself: did my own edit quietly remove a protection?

Full docs: https://github.com/zyx77550/sparda