Embedded Code Review Expert

Overview

Perform structured code review of embedded/firmware projects using dual-model cross-review: Claude Code and Codex independently review the same diff, then findings are cross-compared to catch blind spots that single-model review misses.

Target environments: bare-metal MCU, RTOS (FreeRTOS/Zephyr/ThreadX), Linux embedded, mixed C/C++ firmware.

Trigger

Activate when user asks to review embedded/firmware code changes. Examples:

• "review firmware-pro2 的改动"
• "review the NFC changes"
• /embedded-review ~/Documents/dec/firmware-pro2
• /embedded-review ~/Documents/dec/firmware-pro2 HEAD~5..HEAD
• /embedded-review \x3Cgithub-pr-url>

Severity Levels

Workflow

Mode Selection

Single-model mode (default for small diffs ≤100 lines):

• One review pass using the current session's model
• Faster, lower cost
• Suitable for trivial changes, config tweaks, documentation

Dual-model cross-review (default for diffs >100 lines, or when explicitly requested):

• Claude Code + Codex review independently via ACP
• Cross-compare findings
• Higher quality, catches heterogeneous blind spots
• Use for: new features, architecture changes, critical paths (ISR, crypto, NFC, DMA)

User can override: "用双模型 review" or "quick review 就行"

Phase 0: Preflight — Scope & Context

1. Run scripts/prepare-diff.sh \x3Crepo_path> [diff_range] to extract:

   • Repository info (branch, last commit)
   • Target identification (MCU, RTOS, compiler)
   • Diff stat and full diff content

2. Assess scope:

   • No changes: Inform user; offer to review staged changes or a commit range.
   • Small diff (≤100 lines): Default to single-model review.
   • Large diff (>500 lines): Summarize by file/module first, then review in batches by subsystem.
   • Critical path touched (ISR, DMA, crypto, NFC, boot): Always recommend dual-model.

3. Build review context package:

   REVIEW_CONTEXT = {
     repo_info: (branch, MCU, RTOS, compiler),
     diff: (full git diff text),
     references: (relevant checklist sections from references/),
     focus_areas: (user-specified or auto-detected critical paths)
   }
   

Phase 1: Single-Model Review

For small diffs or when dual-model is not requested:

1) Memory safety scan

• Load references/memory-safety.md for detailed checklist.
• Stack overflow, buffer overrun, alignment, DMA cache coherence, heap fragmentation
• Flag sprintf, strcpy, gets, strcat — suggest bounded alternatives

2) Interrupt & concurrency correctness

• Load references/interrupt-safety.md for detailed checklist.
• Shared variable access, critical sections, ISR best practices, RTOS pitfalls
• Priority inversion, reentrancy, nested interrupt handling

3) Hardware interface review

• Load references/hardware-interface.md for detailed checklist.
• Peripheral init ordering, register access, timing violations, pin conflicts
• Communication protocols: I2C/SPI/UART/NFC buffer management, timeout handling

4) C/C++ language pitfalls

• Load references/c-pitfalls.md for detailed checklist.
• Undefined behavior, integer issues, compiler assumptions, linker issues
• Preprocessor hazards, portability, type safety

5) Architecture & maintainability

• HAL/BSP layering, abstraction, coupling, testability
• Dead code, magic numbers, configuration management

6) Security scan (embedded-specific)

• Secret storage, debug interfaces, firmware update integrity
• Side channels, fault injection, input validation, stack canaries

→ Skip to Phase 3: Output for single-model results.

Phase 2: Dual-Model Cross-Review (ACP)

When dual-model review is triggered:

Step 1: Prepare review payloads

Build two independent review tasks from the same REVIEW_CONTEXT:

Claude Code task:

You are a senior embedded systems engineer reviewing firmware code changes.

[REVIEW_CONTEXT: repo info, diff, focus areas]

Review checklist (apply all that are relevant):
- Memory safety (references/memory-safety.md)
- Interrupt & concurrency (references/interrupt-safety.md)
- Hardware interfaces (references/hardware-interface.md)
- C/C++ pitfalls (references/c-pitfalls.md)
- Architecture & security

Output format: For each finding, provide:
[P0/P1/P2/P3] [file:line] Title
- Description
- Risk
- Suggested fix

Be thorough. Flag everything you find, even if uncertain — mark uncertain items with [?].

Codex task:

You are an independent code reviewer for embedded/firmware projects.
Your job is to find bugs, security issues, and correctness problems.

[REVIEW_CONTEXT: repo info, diff, focus areas]

Focus on:
1. Memory corruption risks (buffer overflow, use-after-free, stack overflow)
2. Concurrency bugs (race conditions, missing volatile, ISR safety)
3. Hardware interface errors (timing, register access, peripheral init)
4. Logic errors and edge cases
5. Security vulnerabilities

Output: List every issue found as:
[SEVERITY: critical/high/medium/low] [file:line] Issue title
- What's wrong
- What could happen
- How to fix

Do NOT skip low-severity items. Report everything.

Step 2: Spawn parallel ACP sessions

sessions_spawn(runtime="acp", agentId="claude-code", task=claude_task)
sessions_spawn(runtime="acp", agentId="codex", task=codex_task)

Both run simultaneously. Wait for both to complete.

Step 3: Cross-compare findings

After both complete, analyze results:

1. Consensus findings (both flagged same issue): HIGH CONFIDENCE — these are real bugs
2. Claude-only findings: Review for validity — may be false positive or genuine catch
3. Codex-only findings: Review for validity — heterogeneous perspective may catch Claude's blind spots
4. Contradictions: Flag for human judgment — one says it's fine, other says it's a bug

Map to unified severity levels (P0-P3).

Phase 3: Output Format

## Embedded Code Review Summary

**Target**: [MCU/Board] | [RTOS/Bare-metal] | [Compiler]
**Branch**: [branch name]
**Files reviewed**: X files, Y lines changed
**Review mode**: [Single-model / Dual-model (Claude Code + Codex)]
**Overall assessment**: [APPROVE / REQUEST_CHANGES / COMMENT]

---

## Findings

### 🔴 P0 - Critical (must block)
(none or list)

### 🟠 P1 - High (fix before merge)
1. **[file:line]** Brief title [🤝 consensus / 🔵 Claude-only / 🟢 Codex-only]
   - Description of issue
   - Risk: what can go wrong
   - Suggested fix

### 🟡 P2 - Medium (fix or follow-up)
...

### ⚪ P3 - Low (optional)
...

---

## Cross-Review Analysis (dual-model only)

| Metric | Count |
|--------|-------|
| 🤝 Consensus (both found) | X |
| 🔵 Claude-only | Y |
| 🟢 Codex-only | Z |
| ⚠️ Contradictions | W |

### Notable disagreements
(list any contradictions with both perspectives)

---

## Hardware/Timing Concerns
(register access, peripheral init, timing-sensitive code)

## Architecture Notes
(layering, testability, portability observations)

Phase 4: Next Steps

---
## Next Steps

Found X issues (P0: _, P1: _, P2: _, P3: _).

**How would you like to proceed?**
1. **Fix all** — implement all suggested fixes
2. **Fix P0/P1 only** — address critical and high priority
3. **Fix specific items** — tell me which issues to fix
4. **Re-review with dual-model** — run cross-review (if single-model was used)
5. **No changes** — review complete

Important: Do NOT implement changes until user explicitly confirms.

Resources

references/

scripts/