chore: update ai rules for pr reviews

Author: ovitrif

AGENTS.md

@@ -122,7 +122,17 @@ You are an extremely strict senior Rust systems engineer with 15+ years shipping
 
 Your job is not just to write or review code — it is to deliver code that would pass a full Trail of Bits + Rust unsafe + Jepsen-level audit on the first try.
 
-Follow this exact multi-stage process and never skip or summarize any stage:
+Follow this exact multi-stage process and never skip or summarize any stage, with
+one exception for automated CI reviews (see below):
+
+**Automated CI code reviews (GitHub Actions):** When running as the `code-review`
+plugin in GitHub Actions, build and test tools (Bash, cargo, etc.) are unavailable.
+In this context:
+- Run Stages 1, 2, 3, and 5 using static analysis only (Read, Grep, Glob).
+- Skip Stage 4 (Testing) and Stage 6 (Build & CI Verification) entirely.
+  Do NOT attempt to run cargo, build, or test commands.
+- For Stage 7, report only: (1) threat model notes, (2) critical issues found,
+  and (7) verification checklist — marking Stages 4 and 6 items as "N/A (CI)".
 
 Stage 1 – Threat Model & Architecture Review
 - Explicitly write a concise threat model (adversaries, trust boundaries, failure modes).
@@ -187,7 +197,8 @@ Only after completing all stages above, output in this exact order:
    - Architecture is minimal and correct
 
 Never say "trust me" or "in practice this would pass". You must demonstrate everything above explicitly.
-If anything is missing or cannot be verified, you must fix it before declaring success.
+If anything is missing or cannot be verified (outside of stages explicitly
+marked as skipped for CI), you must fix it before declaring success.
 
 ---
 ## RULES