embedded-review_skill

Overview

This skill performs expert code review for embedded and firmware projects using a dual-model cross-review approach (Claude + Codex via ACP) to surface blind spots single-model reviews miss. It targets bare-metal MCUs, RTOS-based systems (FreeRTOS/Zephyr/ThreadX), embedded Linux, and mixed C/C++ firmware. Reviews prioritize memory safety, interrupt and concurrency hazards, hardware interface correctness, and common C/C++ anti-patterns.

How this skill works

The skill prepares a REVIEW_CONTEXT from the repository and diff, detects scope and critical paths, then runs either a single-model pass for small diffs or a dual-model cross-review for larger or sensitive changes. In dual-model mode it spawns two independent reviewers (Claude Code and Codex), waits for both results, and cross-compares findings into unified severity levels (P0–P3). Output includes actionable findings, risk assessment, suggested fixes, and a cross-review analysis showing consensus and disagreements.

When to use it

• New features touching ISRs, DMA, boot, crypto, or NFC
• Large diffs (>100 lines) or architecture changes
• Pull requests modifying HAL/BSP, peripheral init, or concurrency primitives
• Security-sensitive updates: firmware updates, secrets, debug interfaces
• Quick checks for tiny patches or documentation changes (single-model)

Best practices

• Provide target context: MCU, RTOS, compiler, and hardware dependencies
• Run prepare-diff to build full REVIEW_CONTEXT before requesting review
• Request dual-model cross-review for critical paths or >100-line changes
• Prioritize fixing P0/P1 findings before merging; treat P2 as follow-ups
• When models disagree, inspect the flagged lines and request human judgment

Example use cases

• Review a PR that adds an ISR handler and DMA-based transfer
• Audit firmware update path for signature/rollback issues
• Assess RTOS API usage and potential priority-inversion or race conditions
• Scan drivers for buffer overruns, alignment or DMA cache-coherency bugs
• Cross-compare two independent reviewers on a major refactor of HAL layers