fix: improve review skill

.claude/commands/review.md

@@ -16,9 +16,11 @@ Read the following files to understand the project's standards and structure. Th
 
 Also use Glob to find any additional guideline files:
 - `.cursor/rules/*.mdc` or `.cursor/*.md` (coding standards, review checklists)
-- Any `CLAUDE.md` files in subdirectories relevant to the changed files
+- Any `CLAUDE.md` or `BUGBOT.md` files in subdirectories relevant to the changed files
 - `CONTRIBUTING.md` if it exists
 
+**Subdirectory guidelines**: If changed files are in subdirectories with their own guideline files, read those as well for module-specific review criteria.
+
 ## Step 2: Gather PR data
 
 Run these commands in parallel:
@@ -33,7 +35,7 @@ Run these commands in parallel:
 Count the number of changed files from step 2.
 
 **If fewer than 100 files changed:**
-Review all changed files directly. Read each changed file using the Read tool to understand surrounding context beyond the diff.
+Review all changed files directly. Read each changed file using the Read tool to understand surrounding context beyond the diff. Cache these file contents for reuse in later validation steps.
 
 **If 100 or more files changed:**
 This is a large PR. Deploy parallel sub-agents to gather context efficiently:
@@ -43,7 +45,7 @@ This is a large PR. Deploy parallel sub-agents to gather context efficiently:
 
 ## Step 4: Four review passes
 
-Execute four independent review passes. For PRs with fewer than 100 files, do these sequentially. For large PRs, launch agents in parallel.
+Execute four independent review passes **in parallel** using separate tool calls. Each pass operates independently on the file context gathered in Step 3.
 
 ### Pass 1 + 2: Code standards and style compliance
 
@@ -93,18 +95,17 @@ For each issue found across all passes, assign a confidence score from 0 to 100:
 - **75**: Highly confident, real and important
 - **100**: Absolutely certain, definitely real
 
-**Validation**: For each finding scored 50 or above, verify it by:
-- Re-reading the relevant code in full context (not just the diff)
+**Filter**: Discard all findings below confidence 80.
+
+**Validation**: For each finding scored **80 or above**, verify it by:
+- Re-reading the relevant code in full context (not just the diff) if not already cached from Step 3
 - Checking if the pattern is intentionally used elsewhere in the codebase
 - For style violations: confirming the project actually enforces this rule
 
-**Filter**: Discard all findings below confidence 80.
-
 **Always discard (false positives):**
 - Pre-existing issues not introduced in this PR
 - Code that appears buggy but is actually correct in context
 - Pedantic nitpicks a senior engineer would not flag
-- Issues that linters (ruff, mypy) will catch — do not run the linter to verify
 - General code quality concerns not explicitly required by project conventions
 - Issues silenced in code via lint-ignore comments
 
@@ -141,14 +142,16 @@ sequenceDiagram
     <interactions showing the primary flow affected by this PR>
 ```
 
-<Generate a Mermaid sequence diagram that shows the key interaction flow introduced or modified by this PR. Rules:>
+<Generate a Mermaid sequence diagram **if the change affects architectural flow or component interactions**. Skip for simple bug fixes (typos, off-by-one errors, single-function changes). Rules:>
 <- Maximum 8 participants>
 <- Maximum 15 interactions>
 <- For refactors: show before/after with labeled boxes>
 <- For new features: show the end-to-end flow>
-<- For bug fixes: show the incorrect flow crossed out and the corrected flow>
+<- For bug fixes affecting flow: show the corrected flow>
 <- Use descriptive labels on arrows>
 
+<If skipping the diagram, replace this section with a brief explanation of why it's not needed.>
+
 ### Findings
 
 <If findings exist above threshold:>
@@ -174,7 +177,7 @@ No issues found. Checked for bugs, security, and code standards compliance.
 For each finding in the Findings table, post an inline comment using `mcp__github_inline_comment__create_inline_comment`.
 
 Rules for inline comments:
-- Maximum 10 inline comments total (prioritize by severity)
+- Maximum 10 inline comments total (prioritize by severity), **unless there are more than 10 Critical severity findings**
 - Each comment includes: severity tag, issue description, why it matters, and the suggested fix
 - For small, self-contained fixes (< 6 lines): include a committable suggestion block
 - For larger fixes: describe the issue and suggested approach without a suggestion block

.github/workflows/claude.yml

@@ -1,6 +1,8 @@
 name: Claude Code
 
 on:
+  pull_request:
+    types: [opened, synchronize, reopened]
   issue_comment:
     types: [created]
   pull_request_review_comment:
@@ -17,6 +19,7 @@ permissions:
 jobs:
   claude:
     if: |
+      (github.event_name == 'pull_request' && github.event.pull_request.draft == false) ||
       (github.event_name == 'issue_comment' && contains(github.event.comment.body, '@claude')) ||
       (github.event_name == 'pull_request_review_comment' && contains(github.event.comment.body, '@claude')) ||
       (github.event_name == 'issues' && contains(github.event.issue.body, '@claude'))
@@ -47,7 +50,9 @@ jobs:
       - uses: anthropics/claude-code-action@v1
         with:
           anthropic_api_key: ${{ secrets.LITELLM_API_KEY }}
+          prompt: |
+            ${{ github.event_name == 'pull_request' && format('Review PR #{0} using the /review command.', github.event.pull_request.number) || '' }}
           claude_args: |
-            --model claude-opus-4-6
-            --allowedTools "mcp__github_inline_comment__create_inline_comment"
+            --model claude-opus-4.5
+            --allowedTools "Bash(gh pr diff:*),Bash(gh pr view:*),Bash(gh pr comment:*),Bash(gh api:*),Bash(git show:*),Bash(git blame:*),mcp__github_inline_comment__create_inline_comment"
             --append-system-prompt "CRITICAL INSTRUCTION: When a user asks you to review a pull request (e.g. @claude review, @claude review this PR, @claude please review), you MUST execute the /review slash command. Do NOT attempt a freeform review. Always use the /review skill — no exceptions."