Update Claude review: tighter prompt, tool restrictions, manual dispatch (#8)

* Update Claude review: tighter prompt, tool restrictions, manual dispatch

* Add draft PR guard to skip reviews on draft PRs

.github/workflows/claude-code-review.yaml

@@ -1,8 +1,14 @@
 name: Claude Code Review
 
 on:
+  workflow_dispatch:
+    inputs:
+      pr_number:
+        description: Pull request number to review
+        required: true
+        type: string
   pull_request:
-    types: [opened, synchronize]
+    types: [opened, ready_for_review]
   workflow_call:
     secrets:
       anthropic_api_key:
@@ -15,21 +21,39 @@ permissions:
 
 jobs:
   claude-review:
+    if: github.event.pull_request.draft == false
     runs-on: ubuntu-latest
     timeout-minutes: 10
     steps:
       - uses: actions/checkout@v4
         with:
-          fetch-depth: 0
+          fetch-depth: 1
 
       - uses: anthropics/claude-code-action@v1
         with:
           anthropic_api_key: ${{ secrets.anthropic_api_key || secrets.ANTHROPIC_API_KEY }}
           allowed_bots: '*'
           prompt: |
-            Review this pull request. Focus on:
+            REPO: ${{ github.repository }}
-            - Code correctness and potential bugs
+            PR NUMBER: ${{ github.event.pull_request.number || github.event.inputs.pr_number }}
-            - Security issues (hardcoded secrets, injection, OWASP top 10)
+
-            - Sea Haven conventions: kebab-case resource names, secrets in AWS Secrets Manager (not env vars), Lambda defaults (Python 3.12+, arm64, explicit 60-day log retention)
+            Review this pull request. Only comment on:
-            - README accuracy if changed
+            - Bugs or logic errors
-            Post findings as inline review comments. Be concise — flag real issues, skip nitpicks.
+            - Security vulnerabilities (hardcoded secrets, injection, OWASP top 10)
+            - Breaking changes or regressions
+            - Sea Haven convention violations: kebab-case resource names, secrets in AWS Secrets Manager (not env vars or SSM), Lambda defaults (Python 3.12+/Node 22.x, arm64, explicit 60-day log retention)
+
+            Do NOT comment on:
+            - Style, formatting, or naming preferences
+            - Minor refactoring suggestions
+            - Performance unless it is a measurable regression
+            - Things that are already consistent with the existing codebase
+
+            Limit to 5 inline comments maximum. If the PR looks good, leave a single top-level comment saying so — do not force issues where there are none.
+
+            Use `gh pr comment` for top-level feedback.
+            Use `mcp__github_inline_comment__create_inline_comment` (with `confirmed: true`) for specific code issues.
+            Only post GitHub comments — do not submit review text as messages.
+
+          claude_args: |
+            --allowedTools "mcp__github_inline_comment__create_inline_comment,Bash(gh pr comment:*),Bash(gh pr diff:*),Bash(gh pr view:*)"