vibeguard: Command for Claude Code

Core Concept

• There are blind spots in single model review, and independent verification of the second model can detect omissions
• Claude is responsible for generating structured reviews, and Codex does adversarial questioning through a structured lens
• Review depth adapts to the size of the change: small changes focus on correctness, while large changes cover architecture and over-design
• Iterate to convergence for at most two total review rounds
• Automatically downgrade to single model review when Codex is unavailable (/vibeguard:review)

Reviewer Lenses

Three complementary perspectives, each focusing on a different failure mode:

Lens Allocation Rules (automatically selected based on change size):

Steps

1. Check Codex Availability

   • Run which codex to check if Codex CLI is installed
   • If it is not available, output the prompt and downgrade the complete process of executing /vibeguard:review, and end
   • If available, continue the dual model process

2. Get guard baseline

   • Run bash guards/<language>/check_*.sh <target_dir> to get the current guard status
   • Record existing issues as a baseline (no duplication of reports)

3. Determine the scope and lens of the review

   • If a file path is specified: review the file
   • If a directory is specified: review recently modified files (git diff --name-only)
   • If no parameters: review the files in the current git staging area
   • Remember the target directory as $TARGET_DIR
   • Count the number of changed lines: git diff --stat or wc -l
   • Determine the active lens based on row number (see table above)

4. Claude generates review report (Round 1)

   • Review by priority P0-P3 (consistent with /vibeguard:review):
     • P0 security: refer to rules/claude-rules/common/security.md, OWASP Top 10, key leakage, input verification
     • P1 logic: boundary conditions, error handling, concurrency safety, data consistency (U-11~U-14)
     • P2 quality: duplicate code, naming conventions, exception handling, file size
     • P3 performance: hot paths, N+1 queries, memory allocation
   • Write review report to /tmp/vibeguard-cross-review-<timestamp>.md
   • Report format:

     ## Claude Review Report (Round N)
     
     ### Guard the baseline
     <Summary of guard script results>
     
     ### Scope of review
     - Change number of rows: <N>
     - Activate lens: <Skeptic / Skeptic+Architect / Skeptic+Architect+Minimalist>
     
     ### Discover
     | Priority | File:line number | Question | Suggestion |
     |--------|-----------|------|------|
     | P0     | ...       | ...  | ...  |
     
     ### Passed items
     - <Confirm that there are no problems>
     

5. Codex Lens Review

   • Construct a dedicated prompt for each activated lens and run the Codex (read-only sandbox):

     codex exec --skip-git-repo-check -a read-only -C "$TARGET_DIR" \
       "You are a code reviewer conducting adversarial reviews using the [LENS_NAME] lens.
     
        ## Your Lens: [LENS_NAME]
        [LENS_DESCRIPTION — see lens definition table]
     
        ## Review report (generated by another model)
        $(cat /tmp/vibeguard-cross-review-<timestamp>.md)
     
        ## Your task
        1. Validate each finding in the report through your lens (marked CONFIRMED / FALSE-POSITIVE)
        2. Check for missing issues from your camera’s perspective (list MISSED items, note camera source)
        3. Evaluate whether the repair proposal is reasonable
     
        ## Output format
        ### [LENS_NAME] review results
        #### verify
        - [Discovery 1]: CONFIRMED / FALSE-POSITIVE — Reason
        #### Omission
        - MISSED [LENS_NAME]: Description
        #### Judgment
        VERDICT: APPROVED / REVISE — reason" 2>/dev/null
     

   • Codex calls for multiple lenses merged into one execution (avoiding multiple startup overhead), listing all active lenses in prompt
   • Capture Codex output