菜单
Use when reviewing ava-labs/firewood code changes — pull request or local workspace. Invoke with phrases like "review PR", "review my changes", "review this branch", "firewood review". Handles PR checkout, staged/unstaged changes, Rust toolchain, Go FFI, CI workflow, and multi-agent parallel review.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ava-labs/firewood --skill firewood-review命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | firewood-review |
| description | Use when reviewing ava-labs/firewood code changes — pull request or local workspace. Invoke with phrases like "review PR", "review my changes", "review this branch", "firewood review". Handles PR checkout, staged/unstaged changes, Rust toolchain, Go FFI, CI workflow, and multi-agent parallel review. |
| compatibility | Requires Claude Code with git, cargo/rustup, go, and the gh CLI. Designed for use in the ava-labs/firewood workspace. |
| allowed-tools | Read Write Bash Agent AskUserQuestion EnterWorktree ExitWorktree mcp__github__pull_request_read mcp__github__list_pull_requests mcp__github__get_commit mcp__github__list_commits |
Firewood-specific local code review. Dispatches parallel agents covering Rust toolchain validation, Go FFI checks, CI workflow analysis, and deep code quality. Never publishes review output to GitHub unless explicitly instructed.
Code quality review rules are defined in AGENTS.md — Code Review Guidelines. Agents should read that section in full before beginning scrutiny.
Create a session-scoped scratch directory at the start. All fetched API data is written here so subagents can read it without re-fetching or needing extra permissions:
SCRATCH=/tmp/firewood-review/$$
mkdir -p "$SCRATCH"
Pass $SCRATCH to every subagent as part of its prompt.
PR review — prompt names a PR number, branch, or URL:
If worktrees are available and no worktree is already checked out for this PR, create a dedicated worktree via EnterWorktree for the PR branch. Otherwise fall back to gh pr co <PR_NUMBER>.
Skip checkout if the prompt says the workspace is already on the PR branch.
Fetch PR metadata — prefer the github MCP if available, otherwise fall back to gh CLI:
MCP (preferred):
mcp__github__pull_request_read → write JSON to $SCRATCH/pr.json
CLI fallback:
gh pr view <PR_NUMBER> --json title,body,comments,reviews \
> "$SCRATCH/pr.json"
gh api repos/ava-labs/firewood/pulls/<PR_NUMBER>/comments \
> "$SCRATCH/pr-comments.json"
Record the base and head SHAs from the fetched data for use in diffs.
Local changes — prompt says "local changes", "unpublished", or similar:
git status --porcelain to detect staged vs unstaged files.AskUserQuestion: "Review workspace as-is, or stash unstaged changes first?"git stash push --include-untracked -m "firewood-review-stash". After synthesis completes, git stash pop unconditionally.From the diff (git diff --name-only <base>..<head> for PR, or git diff --name-only HEAD for local), determine which agents to activate:
*.rs or Cargo.* changes → activate rust-toolchainffi/** Go files → activate go-toolchain.github/** changes → activate ci-reviewWrite the changed file list to $SCRATCH/changed-files.txt.
Dispatch all applicable agents in a single message (parallel execution). Each agent receives: $SCRATCH path, diff output, list of changed files, and the instructions from its section below. PR metadata is read from files in $SCRATCH.
Precondition: Any *.rs or Cargo.* files changed.
uname -s
cargo fmt --all --check
cargo fetch --locked
# All subsequent cargo commands: add --frozen --target-dir target/review
Every feature set below builds on both platforms. Note that io-uring is inert off Linux, so a green macOS --all-features run does not exercise storage/src/linear/io_uring.rs; only Linux compiles it.
Use --frozen after the initial fetch to ensure we're not accidentally pulling new dependencies that aren't covered by the lockfile.
Use --target-dir target/review for ALL cargo commands to avoid conflicting with the workspace target dir.
Run cargo clippy --workspace --all-targets and cargo nextest run --workspace --all-targets for each feature set:
| Feature flags | macOS | Linux |
|---|---|---|
| (none) | ✓ | ✓ |
--no-default-features | ✓ | ✓ |
--features ethhash,logger | ✓ | ✓ |
--all-features | ✓ | ✓ |
--all-features requires rustc 1.94.1+, above the 1.94.0 workspace MSRV, because it enables fwdctl's launch feature and its AWS SDK dependencies declare that floor. If cargo reports rustc 1.94.0 is not supported by the following packages, that is this constraint and not a defect in the PR under review.
If the PR adds any new Cargo feature, test with and without it across the applicable matrix rows.
Header drift check:
cargo build -p firewood-ffi --frozen --target-dir target/review
git diff --exit-code ffi/firewood.h
Non-zero exit = stale header → blocking finding. Write results to $SCRATCH/rust-toolchain.txt.
Precondition: Any Go files in ffi/ changed.
Go bindings require the FFI build to use the workspace target/ dir (the LDFLAGS in ffi/firewood.go reference it directly — do not use --target-dir here):
cargo build -p firewood-ffi -F ethhash,logger --frozen
cd ffi
./scripts/lint.sh
go test ./... -race
Write results to $SCRATCH/go-toolchain.txt. Report any lint violations or test failures as blocking.
Always activated. Read all changed files in full. Do not skim diffs — read the actual source for context. Read PR metadata from $SCRATCH/pr.json and $SCRATCH/pr-comments.json if present.
Read the Code Review Guidelines section of AGENTS.md in full and apply every rule to the changed code.
Additionally:
CLAUDE.md and CONTRIBUTING.md in full and check whether this PR's changes create any gaps or inconsistencies in those documents. Suggest specific wording updates if helpful.AGENTS.md or CONTRIBUTING.md.Write findings to $SCRATCH/code-quality.txt.
Precondition: Any .github/** files changed.
Read all changed workflow YAML files in full.
Efficiency: redundant steps, matrix entries that duplicate coverage, jobs that could parallelize, missing cache steps.
Correctness: wrong trigger events, missing permissions, broken if: conditionals, misuse of continue-on-error, pinned action SHA vs tag inconsistencies.
act validation (if available):
act --list
act -n -j <modified-job>
Report output. If act is unavailable, note it.
Realized CI results — prefer the github MCP if available, otherwise fall back to gh:
MCP (preferred):
mcp__github__pull_request_read → extract check_runs/status fields → write to $SCRATCH/ci-checks.json
CLI fallback:
gh pr checks <PR_NUMBER> --json name,state,conclusion \
> "$SCRATCH/ci-checks.json"
Compare expected jobs vs actual. Flag failures, unexpected skips, or jobs that didn't run.
Write findings to $SCRATCH/ci-review.txt.
After all agents complete, read $SCRATCH/*.txt and aggregate into a single report:
Three severity tiers:
unwrap in prod, unsafe soundness issues)Present the report in the current conversation. Do not publish to GitHub unless the user explicitly asks.
If a worktree was created: offer to exit and clean it up.
If changes were stashed: git stash pop unconditionally before finishing.
Clean up: rm -rf "$SCRATCH" after presenting the report.