kc-pr-review

Language rule (two layers):

• Conversation-facing text (status updates, confirmation prompts, findings tables shown to the user) — follows unified language preference. See plugin CLAUDE.md for query flow.
• PR-facing artifacts (review body, inline comment bodies, anything POSTed to GitHub) — default to English regardless of the conversation language, matching the convention for PR title / commit messages / code comments. Override only if the target repo's CLAUDE.md explicitly requires otherwise.

Process Flow

digraph review_pr {
  rankdir=TB;
  node [shape=box];

  detect [label="Detect PR\n(number, URL, or current branch)"];
  fetch [label="Fetch PR metadata + diff"];
  concerns [label="Extract user concerns\n→ verification matrix"];
  ownership [label="Check repo ownership"];
  triage [label="Triage: filter noise,\nsize + security + archetype\n→ agent selection"];

  subgraph cluster_parallel {
    label="Parallel dispatch";
    style=dashed;
    prescan [label="Pre-scan (main context)\nCLAUDE.md rules\n+ stale refs\n+ dependency chains\n+ prompt consistency\n+ runtime data shape\n+ lint gate\n+ non-code scan\n+ dead export detection\n+ helper rollout completeness\n+ doc claim verification\n+ intra-doc rule-vs-example"];
    review [label="Review agents\n(code-reviewer,\ncomment-analyzer)"];
    tests [label="Test execution\n(worktree)\nunit + eval"];
    audit_prep [label="Compliance prep\n(5a: read docs,\n5b: match skills)"];
    knowledge [label="Knowledge layer\n(5a-k: episodic memory\n+ review-lessons.md)"];
    tob_security [label="ToB security reviewer\n(differential review,\nadversarial modeling)"];
    tob_supply [label="ToB supply chain\n(dep risk + insecure defaults)\n[conditional]"];
    tob_actions [label="ToB actions auditor\n(AI agent CI/CD vectors)\n[conditional]"];
    probe [label="Break-point probe\n(failure chain + A/B/C/D\n+ residual uncertainty)\n[bugfix/cross-stack]"];
    codex [label="Codex cross-model\n(second opinion)\n[--codex / bugfix-cross-stack]"];
  }

  classify [label="5c: Cross-reference\nagent + test findings\n+ root cause classify"];
  reconcile [label="5.5: Cross-model\nreconciliation\n(buckets + conflict set)\n[CODEX present]"];
  arbitrate [label="5.6: Gemini\narbitration\n(verdict → confidence)\n[conflict ∧ gemini]"];
  draft [label="Draft review:\nCODE table\n+ DOC/NEW advisory"];

  node [shape=diamond];
  confirm [label="User confirms\nreview?"];

  node [shape=box];
  diagrams [label="Generate + preview\nsequence + architecture/status\ndiagrams (optional)"];
  post [label="Post review via\ngh pr review"];
  review_url [label="Return review URL"];

  node [shape=diamond];
  has_learning [label="Learning\ninsights exist?"];

  node [shape=box];
  capture [label="8: Learning\nD1: skill patterns\nD2: project knowledge"];
  write_knowledge [label="D1: auto-append\nD2: threshold → confirm\n→ write + commit"];
  end_node [label="Done"];

  detect -> fetch -> concerns -> ownership -> triage;
  triage -> prescan;
  triage -> review;
  triage -> tests;
  triage -> audit_prep;
  triage -> knowledge;
  triage -> tob_security;
  triage -> tob_supply [label="deps changed"];
  triage -> tob_actions [label="workflows changed"];
  triage -> probe [label="bugfix/cross-stack"];
  triage -> codex [label="--codex / bugfix-cross-stack"];
  prescan -> classify;
  review -> classify;
  tests -> classify;
  audit_prep -> classify;
  knowledge -> classify;
  tob_security -> classify;
  tob_supply -> classify;
  tob_actions -> classify;
  probe -> classify;
  codex -> classify;
  classify -> reconcile;
  reconcile -> arbitrate [label="conflict ∧ gemini"];
  reconcile -> draft [label="no conflict / no gemini"];
  arbitrate -> draft;
  draft -> confirm;
  confirm -> diagrams [label="D: preview only"];
  diagrams -> confirm [label="no post"];
  confirm -> post [label="approved"];
  confirm -> draft [label="edit requested"];
  post -> review_url -> has_learning;
  has_learning -> capture [label="yes"];
  has_learning -> end_node [label="none"];
  capture -> write_knowledge [label="approved"];
  capture -> end_node [label="skip"];
  write_knowledge -> end_node;
}