name	authz
description	Authorization-model review. Scans endpoints + object-access-by-id sites; the authz-reviewer agent finds missing authorization (CWE-862), IDOR / broken object-level authz (CWE-639), privilege escalation, and broken ownership, and promotes them into .kuzushi/findings.json (source "authz"). Complements /threat-hunt (which hunts named threats) with a dedicated authz pass.
context	fork
agent	authz-reviewer
user-invocable	false

Authz review

Check that every sensitive action is gated by the right authorization, not just authentication.

Run node "${CLAUDE_PLUGIN_ROOT}/scripts/cmd/authz-prepare.mjs" --target "<repo root>" (optionally --input '{"maxCandidates":30}'). If it reports no-candidates, say there are no handlers / object accesses to review and stop. Read the prep's prepPath.
For each candidate (endpoint or object-by-id), open the site and determine the protection — authz is often a decorator/middleware above the handler or a filter before the query, so widen with Read/Grep. Decide finding / candidate / rejected with an authzClass (missing-authz / idor / privilege-escalation / broken-ownership).
Write the { candidates: [...] } bundle to the prep's draftPath, then run the assembleCommand — it validates and promotes into .kuzushi/findings.json (source:"authz").
Report findings by class (file:line, attacker, the missing/broken check, the fix).

"It's behind login, so it's fine." → Authentication ≠ authorization; a logged-in user reaching another user's object is still IDOR. Check the ownership/tenant scope.
"There's probably a middleware." → Find it. "Probably protected" is candidate; rejected must name the actual check.
"The id is from the session." → Likely safe, but confirm the object is scoped to that principal.

name	authz
description	Authorization-model review. Scans endpoints + object-access-by-id sites; the authz-reviewer agent finds missing authorization (CWE-862), IDOR / broken object-level authz (CWE-639), privilege escalation, and broken ownership, and promotes them into .kuzushi/findings.json (source "authz"). Complements /threat-hunt (which hunts named threats) with a dedicated authz pass.
context	fork
agent	authz-reviewer
user-invocable	false

Check that every sensitive action is gated by the right authorization, not just authentication.

Run node "${CLAUDE_PLUGIN_ROOT}/scripts/cmd/authz-prepare.mjs" --target "<repo root>" (optionally --input '{"maxCandidates":30}'). If it reports no-candidates, say there are no handlers / object accesses to review and stop. Read the prep's prepPath.
For each candidate (endpoint or object-by-id), open the site and determine the protection — authz is often a decorator/middleware above the handler or a filter before the query, so widen with Read/Grep. Decide finding / candidate / rejected with an authzClass (missing-authz / idor / privilege-escalation / broken-ownership).
Write the { candidates: [...] } bundle to the prep's draftPath, then run the assembleCommand — it validates and promotes into .kuzushi/findings.json (source:"authz").
Report findings by class (file:line, attacker, the missing/broken check, the fix).

"It's behind login, so it's fine." → Authentication ≠ authorization; a logged-in user reaching another user's object is still IDOR. Check the ownership/tenant scope.
"There's probably a middleware." → Find it. "Probably protected" is candidate; rejected must name the actual check.
"The id is from the session." → Likely safe, but confirm the object is scoped to that principal.

authz