[kurtseifried]

Agent: Gemini (SecID governing agent)

Summary

Implements independent review instructions for CloudSecurityAlliance/SecID-Service per CSA-Internal-Governance/SecID/AGENTS.md ("First tasks" #4) and CINO-Platform-Engineering#261:

1. Add REVIEW.md:
   • Starts with @AGENTS.md (the documented root include for Claude Code Review on a protected default branch).
   • Provides a complete, self-contained 6-part review checklist covering:
     1. Authority, provenance, and CI/deploy guardrails (registry-kv-upload.yml, wrangler.toml).
     2. Recursive MCP output sanitization (src/sanitize.ts: sanitizeDict + capValue) and submit_feedback untrusted envelope isolation (src/feedback.ts).
     3. /mcp HTTP method gate parity with CSA-MCP-Core (src/method-gate.ts, ADR-015).
     4. Abuse-resistance and operational limits (secid ≤ 1024 chars, MCP body ≤ 64 KiB, batch ≤ 10 and no submit_feedback in batches, 20 writes/min/isolate budget).
     5. KV --sync 50-orphan safety threshold, 25 MiB per-key limit, stateless KV runtime reads, and TYPE_REGISTRY lockstep.
     6. Vitest unit coverage, release-train test guards (test/resolver.test.ts), and @third-party tagging on Playwright E2E tests (e2e/).
2. Update AGENTS.md:
   • Adds a matching ## Code Review Rules section for OpenAI Codex (codex review / codex exec) and references REVIEW.md in the Repository Structure tree.

Verification

• Verified no bare @ tokens outside backticks (except line 1 @AGENTS.md in REVIEW.md).
• CI (ci.yml) runs on this PR.
• Independent cross-vendor review by OpenAI Codex (gpt-6-astra) in a fresh context will be recorded on this PR before merge.

Closes #50

[kurtseifried]

Agent: Gemini (SecID governing agent)

Two-Review Summary

1. Review 1 — Self-Review (Gemini, SecID governing agent):

   • Verified REVIEW.md starts with @AGENTS.md on line 1 followed by self-contained review rules for Claude Code Review, GitHub Copilot Code Review, and human/agent reviewers.
   • Verified ## Code Review Rules in AGENTS.md matches REVIEW.md.

2. Review 2 — Independent Fresh-Context Cross-Vendor Review (OpenAI Codex gpt-6-astra, session 01a19334-40e8-7f92-a948-0cb17f0817e8):

   • Findings raised (2 P2, 1 P3, 0 P1):
     1. [P2] Align sanitizer contract with src/sanitize.ts: Replace sanitizeResult() / sanitizeEnvelope(), _untrusted_prose, _prose_warning, and MAX_STRING_LEN with the exact symbols in src/sanitize.ts: sanitizeResponseForMcp(), sanitizeDict(), sanitizeValue(), UNTRUSTED_TEXT_KEYS (description, notes, note, official_name, common_name, auth, contacts, contact, scope, policy, disclosure_policy, version_disambiguation, unversioned_behavior, parsing_instructions), registry_text_untrusted._warning, capString(), capValue(), MAX_FIELD_CHARS = 4000, and MAX_ARRAY_ITEMS = 64.
     2. [P2] Preserve OPTIONS exception to the 405 rule (src/method-gate.ts): Clarify that OPTIONS passes through to CORS middleware (204), while registered non-POST, non-OPTIONS methods (GET, HEAD, PUT, PATCH, DELETE, QUERY, TRACE, CONNECT) from non-browsers receive 405 Method Not Allowed with Allow: POST, OPTIONS.
     3. [P3] Document exact isBrowserNavigation() predicate: State that isBrowserNavigation() rejects MCP-client signals (Accept: text/event-stream, Mcp-Protocol-Version, Mcp-Session-Id) first and then accepts Sec-Fetch-Mode: navigate or Accept: text/html. Also require unconditional cross-vendor review + decision record for all gate/reviewer-instruction changes.
   • Resolution: All 3 findings addressed in commit 0c1dbda.