Add authentication-review skill

[richardmhope]

Summary

• New authentication-review skill (0.1.0): a deep-slice review of authentication changes — password storage/policy and recovery, MFA bypass and OTP handling, sessions & cookies, JWT/API-token verification, OAuth 2.0, OIDC, SAML, and LDAP sign-in — classified against OWASP ASVS 5.0 (V6/V7/V9/V10, with V11/V3 for KDF and cookie-attribute findings) and grounded in RFC 9700, NIST SP 800-63B, and the OWASP Authentication, Password Storage, Session Management, MFA, Forgot Password, and SAML Security cheat sheets (all fetched during authoring).
• security-review 0.3.2: reciprocal cross-reference (it keeps breadth; the new skill takes depth on the auth slice).
• docs/finding-output.md: classifier rows for authentication-review, plus the previously missing ci-workflow-review/iac-review rows.
• Eval fixtures: authentication-review (hand-rolled OAuth code flow with no state/PKCE) and authentication-review-saml variant (ACS verifies only the response-envelope signature, then reads NameID from the re-parsed raw document — XSW-shaped).
• Plugin tree regenerated via scripts/build_plugin.py.

Test plan

• uv run --extra dev ruff check . && uv run --extra dev ruff format --check . && uv run --extra dev mypy && uv run --extra dev pytest — 132 passed (structure lint, fixture validation, plugin freshness)
• skilldeck install authentication-review --agent claude --scope project smoke-tested; rendered SKILL.md has correct frontmatter and stamp
• Manual paid eval run (python evals/run_evals.py) against the two new fixtures — deferred, per evals/README

🤖 Generated with Claude Code

[icebergai-review-bot]

IcebergAutoReview

Verdict: request_changes

The change is mechanically sound, but the new security guidance contains a blocking SAML correctness error that would generate false-positive findings for valid implementations.

Findings

• src/skilldeck/skills/authentication-review/skill.md:196: The checklist says an unsigned Assertion inside a signed Response is attacker-writable and mandates rejecting unsigned assertions. A correctly verified signature over the enclosing Response also protects its contained Assertion; the essential requirement is to consume claims only from the verified signed subtree. The fixture’s own base implementation follows that model. Revise this guidance to accept either a signed Assertion or a signed Response whose signature reference covers the trusted Assertion, while retaining the signed-subtree/XSW checks. This is consistent with the cited OWASP SAML guidance.

Validation

• git diff --check main...HEAD — passed.
• python scripts/build_plugin.py --check — passed; generated plugin tree is current.
• Targeted fixture preparation probe for authentication-review and authentication-review-saml — passed.
• Targeted pytest invocation could not start because uv attempted to create .venv on the read-only checkout; system Python lacks pytest. Current-head GitHub CI is green as stated.

Residual risks / optional notes

• The paid agent evaluation for both new fixtures remains deferred, so their real detection and false-positive behavior has not been demonstrated.

Automated review by Codex 76eaed20bdbd using IcebergAutoReview.