[codex] route reviewer findings to source tasks

.claude-plugin/marketplace.json

@@ -71,7 +71,7 @@
     {
       "name": "nihil",
       "description": "Evidence-gated maintainability discipline in two layers. (1) A native plugin: four modes \u2014 /nihil:raze (root-authority, write-capable transformation of a repo you own; only secret-leak and catastrophic, unrecoverable commands are blocked) plus /nihil:review, /nihil:implement, /nihil:release with confidence scoring, scope control, and release gating \u2014 backed by five read-only review agents and PreToolUse/Stop hooks (a discipline aid, not a security boundary: heuristic command matching, fail-open; a secret / API-key brake is active in every mode). (2) A summonable pantheon of five first-principles dynamic workflows (/nihil orchestrator + /nihil-maat review, /nihil-odin research, /nihil-shiva deletion, /nihil-athena restructure), installed into .claude/workflows/ via /nihil:summon.",
-      "version": "0.3.0",
+      "version": "0.3.1",
       "source": "./plugins/nihil"
     },
     {
@@ -89,7 +89,7 @@
     {
       "name": "safety-nets",
       "description": "Two deterministic Stop-event safety nets (no LLM) that keep a session honest: overclaim-net blocks a 'done/fixed/verified' claim made without a verifying command this turn; slnx-sync blocks unregistered .csproj in a .slnx, auto-skipping VMR-synced upstream forks (dotnet/dotnet codeflow / dotnet upstream remote) and honoring a .slnx-sync-ignore opt-out.",
-      "version": "0.1.0",
+      "version": "0.1.1",
       "source": "./plugins/safety-nets"
     },
     {

.coderabbit.yaml

@@ -24,6 +24,7 @@ reviews:
   changed_files_summary: true
   sequence_diagrams: false
   poem: false
+  enable_prompt_for_ai_agents: true
   auto_review:
     enabled: true
     drafts: false
@@ -41,6 +42,15 @@ reviews:
     - "!**/*.min.css"
     - "!plugins/cc-plugin-eval/fixtures/**"   # test scaffolding, not shipped
   path_instructions:
+    - path: "**"
+      instructions: |
+        AI reviewer coordination. Comments from Codex, Claude, Copilot, or CodeRabbit are all valid
+        implementation leads. If a review point is outside the GitHub diff range, do not discard it:
+        translate it into a concrete task against the current source tree by naming the best target
+        path plus line, symbol, or search phrase. If no exact anchor exists, name the files that must
+        be inspected and the behavior to change. Multiple agents may work the same reviewer point in
+        parallel; overlapping findings are signal, not duplication. Do not transcribe reviewer text
+        into repo files as the fix: forward the finding to the source and change the root cause.
     - path: "plugins/*/skills/**/SKILL.md"
       instructions: |
         Skill definitions. The `description` frontmatter is the ONLY trigger mechanism — check it

.github/copilot-instructions.md

@@ -352,6 +352,13 @@ AIs coordinate through **shared files**, not real-time communication:
 
 Each AI does its own complete review. Overlapping findings indicate high confidence issues.
 
+Outside-diff review comments are still valid work. When GitHub cannot anchor a
+comment inline, convert it into a concrete source task: target file, line or
+symbol when possible, plus the behavior to change. Do not paste reviewer prose
+into repository files as the fix; forward the finding to the source and change
+the root cause. Multiple agents may implement the same reviewer point in
+parallel; the common denominator between their fixes is the signal to preserve.
+
 ### 9.6 Plugin Review Scope
 
 All AIs review the same things in this repo: