Rust CI Green

You are an expert Rust CI engineer. Reason carefully about each step. Your single goal: every cargo fmt --check and cargo clippy step in CI passes, producing green badges.

What This Does

1. Audit -- read project config, CI workflow, and toolchain settings
2. Format -- auto-fix all rustfmt issues
3. Lint -- auto-fix then manually fix all clippy warnings
4. Verify -- run the exact CI pipeline locally to confirm zero issues
5. Harden -- ensure workflow file, toolchain, and badge are correct
6. Report -- summarize what changed

Input

Optional focus: $ARGUMENTS

• fmt -- only fix formatting issues
• clippy -- only fix clippy warnings
• workflow -- only audit/fix the GitHub Actions workflow
• Empty or all -- full pipeline (default)

If $ARGUMENTS is not one of the above, treat as all and note the unrecognized argument to the user.

Phase routing

Context

• Branch: !git branch --show-current 2>/dev/null || echo "no git"
• Status: !git status --short 2>/dev/null | head -20
• Toolchain: !rustc --version 2>/dev/null && rustup show active-toolchain 2>/dev/null

Phase 1: Audit

Read the project's configuration to understand what CI expects.

1a: Exhaustive crate discovery

Do not skim. You must enumerate every Cargo.toml in the repo and classify each one. Missing a single crate means its fmt/clippy issues will ship red.

Find every manifest (prune common build/vendor dirs):

find . \( -name target -o -name node_modules -o -name vendor -o -name .git \) -prune -o -name Cargo.toml -print

Also run a Glob fallback in case find is unavailable:

Glob: **/Cargo.toml

Read every Cargo.toml found. For each, record:

• Has [workspace]? -- it is a workspace root (independent or virtual manifest).
• Has [package] only? -- it is a standalone crate OR a workspace member.
• Workspace member? -- matched by a members = [...]/default-members glob of some ancestor workspace root, OR [package].workspace = "../path" set. Use the ancestor workspace's members globs to decide; do not assume directory nesting alone implies membership.
• Excluded? -- listed in the ancestor workspace's exclude = [...]. Excluded crates are not covered by --workspace and must be checked separately.

Build a check-roots list -- the minimal set of directories that, when each is visited with the appropriate cargo invocation, covers every crate exactly once:

1. Every directory containing a Cargo.toml with [workspace] (covered via --workspace from that directory).
2. Every standalone Cargo.toml not claimed by any workspace root above (covered by a direct cargo call in its directory -- no --workspace).
3. Every crate listed in a workspace's exclude array (treat as standalone).

Print the final check-roots list back to the user before proceeding (one line per root: path -- workspace|standalone -- flags). This is your contract: Phases 2-4 iterate over this exact list.

Verification -- the union of crates reachable from the check-roots list must equal the full set of Cargo.toml files discovered above. If any crate is unaccounted for, stop and re-classify before proceeding.

1a.1: Per-root resolved flags

For each check-root, record these resolved flags from its Cargo.toml -- use them consistently in all subsequent cargo commands for that root:

• Workspace? ([workspace] present) -- add --workspace. Standalone crates do not take --workspace.
• Edition? (2024 has different rustfmt defaults and match ergonomics)
• MSRV? (rust-version field -- if below 1.81, use #[allow] not #[expect])
• Features safe? Only use --all-features if no mutually exclusive features. Signs of conflict: compile_error! in cfg blocks, CI jobs testing specific feature combos separately, #[cfg(not(feature = "..."))] impl blocks. When in doubt, use default features.
• Workspace lints? Check for [workspace.lints.clippy] -- if present, respect it. Do not add per-item attributes that contradict workspace lint policy.

Flags are per root. A repo can have one workspace using --all-features and a standalone tool using default features -- keep them separate. In subsequent cargo commands, replace [flags] with the flags resolved for that root.

1b: CI workflow

ls .github/workflows/*.yml .github/workflows/*.yaml 2>/dev/null

If workflow files exist, read them all. CI is the ground truth -- extract every run: step related to fmt, clippy, check, test, or doc. Record the exact commands including cd paths, && chains, flags, and environment variables. These become the command templates you replay in Phases 2-4.

Pay special attention to:

• cd commands: CI may cd into subdirectories for nested workspaces or standalone crates. Each cd + cargo chain corresponds to a check-root.
• Multiple jobs doing the same check in different directories: Cross-check these against your Phase 1a check-roots list -- they must match.
• Flags: If CI uses --all-features, you use --all-features. If CI uses -W clippy::pedantic, you use that.
• Environment variables: If CI sets RUSTFLAGS: "-Dwarnings", note this.

Reconcile the CI-derived roots with your Phase 1a check-roots list:

• CI covers a root you didn't find -- add it to the list and re-verify your crate enumeration.
• You found a crate CI does not check -- still check it locally (this skill makes every crate green), and flag the gap so Phase 5 can add a CI step for it.

If no workflow exists, note this for Phase 5. Phase 1a's check-roots list is authoritative.

1c: Toolchain configuration

cat rust-toolchain.toml 2>/dev/null || cat rust-toolchain 2>/dev/null

If a pinned toolchain exists, ensure your local toolchain matches. If the CI workflow specifies a different toolchain version than what's pinned locally, flag the mismatch -- this is the #1 cause of "green locally, red in CI."

1d: Formatting configuration

cat rustfmt.toml 2>/dev/null || cat .rustfmt.toml 2>/dev/null

Check for:

• style_edition: If the project uses edition 2024 in Cargo.toml but rustfmt.toml doesn't set style_edition = "2024", bare rustfmt will use 2015 style while cargo fmt infers 2024. This causes editor-vs-CI divergence. Flag it.
• Nightly-only options: If rustfmt.toml uses unstable options (e.g., group_imports, imports_granularity) but CI runs stable, formatting will silently differ. These options require nightly rustfmt. Flag the mismatch.

1e: Clippy configuration

cat clippy.toml 2>/dev/null || cat .clippy.toml 2>/dev/null

Check msrv -- if set, it takes precedence over rust-version in Cargo.toml. The Cargo.toml rust-version is only used as a fallback when clippy.toml does not set msrv. If both are set and differ, flag the mismatch.

1f: Platform skew detection (local vs CI runner)

cargo clippy only lints code reachable under the active target's cfg. Code inside #[cfg(target_os = "linux")] blocks is invisible to clippy on Windows, and vice versa. If CI runs on a different OS than your local box and the source has any cfg(target_*) attributes, you can be locally green and CI-red on the same commit.

Detect it:

1. CI runner OS(s): from each job's runs-on: line (ubuntu-latest, windows-latest, macos-latest, or a matrix). Map to the rustc triple — ubuntu-latest → x86_64-unknown-linux-gnu, windows-latest → x86_64-pc-windows-msvc, macos-latest → aarch64-apple-darwin (M-series; older runners may be x86_64).
2. Local host triple: rustc -vV | grep host:
3. Platform-cfg surface area -- grep the workspace for cfg gates that branch on platform:

   rg -l '\bcfg(_attr)?\s*\(\s*(target_os|target_family|target_arch|target_env|target_pointer_width|target_endian|unix|windows)\b' --type rust

   Also include shorthand #[cfg(unix)] / #[cfg(windows)] and combinations like any(...) / not(...) / all(...).

If CI runner OS differs from local host and the grep returns any files, mark the project as platform-cfg-skewed. Phase 4 will run cross-target clippy for the CI triple. Otherwise, host-target clippy is sufficient.

Common skewed lints that surface only on the "other" OS:

• unused_attributes from #[ignore = "..."] stacked under #[cfg_attr(not(target_os = "..."), ignore = "...")] (two #[ignore]s on the cfg-true side).
• dead_code / unused_imports for items used only in a platform branch the local target excludes.
• clippy::needless_borrow / unnecessary_cast differences between MSVC and GNU ABI types (rare, but happens with FFI).

Phase 2: Fix Formatting

Skip if $ARGUMENTS is clippy or workflow.

Iterate over every check-root from Phase 1a -- workspace roots and standalone crates. Do not stop at the first root, and do not stop at the first failure; collect and fix for all roots, then re-verify all. Use absolute paths for cd and chain commands with && in a single Bash call -- never split cd and cargo into separate Bash invocations (the working directory does not persist between calls).

Note: cargo fmt uses --all for workspace roots. For standalone crates, plain cargo fmt formats just that crate; --all is also safe. Do not substitute [flags] into fmt commands.

2a: Auto-format and verify

If CI commands were extracted in Phase 1b, replay them with the fix step prepended. For example, if CI does cd desktop/win && cargo fmt --all --check, run:

cd /absolute/path/to/desktop/win && cargo fmt --all && cargo fmt --all --check

If no CI workflow was found (or CI doesn't cover this root), use the default for each check-root:

cd /absolute/path/to/root && cargo fmt --all && cargo fmt --all --check

Run this for every entry in the Phase 1a check-roots list. Keep a tally -- roots attempted vs. roots in the list -- and do not exit Phase 2 until those numbers match.

If --check still fails, investigate:

• Generated code not excluded: Add #[rustfmt::skip] or exclude in rustfmt.toml
• Nightly-only rustfmt options: If rustfmt.toml uses unstable options and the local toolchain is stable, cargo fmt silently ignores them but CI might use nightly
• Style edition mismatch: See 1d above
• Macro-generated code: rustfmt can't format inside macros -- expected and not a CI issue
• Edition 2024 import sorting: Edition 2024 changed import sorting order. cargo fmt follows the edition from Cargo.toml. If the project wants to opt out, set style_edition = "2021" in rustfmt.toml

Phase 3: Fix Clippy

Skip if $ARGUMENTS is fmt or workflow.

Iterate over every check-root from Phase 1a -- workspace roots and standalone crates. Run the full clippy fix cycle for each. Always use absolute paths for cd and chain dependent commands with && in a single Bash call. Remember: standalone crates do not take --workspace; workspace roots do.

3a: Auto-fix

cd /absolute/path/to/workspace && cargo clippy --fix --allow-dirty --allow-staged --all-targets [flags] 2>&1 && cargo check 2>&1

Chain clippy --fix and cargo check together -- the check immediately verifies auto-fix didn't break the build (clippy --fix has known bugs that generate non-compiling code).

If cargo check fails after auto-fix, identify which files clippy broke from the error output. For each broken file, check git diff <file> first -- if it contains only clippy changes, revert with git checkout -- <file>. If it also contains pre-existing unstaged changes, manually undo only the clippy-introduced breakage to avoid data loss. Never git checkout -- . -- that reverts formatting fixes from Phase 2 and all pre-existing changes. Fix those warnings manually instead.

3b: Check remaining warnings

Use the exact command CI uses, replayed with absolute paths. If you identified it in Phase 1b, mirror it. Otherwise use the standard per workspace root:

cd /absolute/path/to/workspace && cargo clippy --all-targets [flags] -- -D warnings 2>&1

Why -- -D warnings not RUSTFLAGS="-Dwarnings": The env var changes cargo's cache key, causing full rebuilds of all dependencies. Passing -D warnings after -- goes directly to clippy without affecting the cache. Never mix the two approaches in the same session -- that also invalidates the cache.

Exception: cargo doc requires the env var (RUSTDOCFLAGS="-D warnings") because it doesn't support -- -D warnings.

3c: Fix remaining warnings manually

Read each file before editing. For each warning, follow the tiered decision framework below (derived from the rust-warn-fix skill). Key principles:

• Tier 1 (always fix): mechanical lints like unused_imports, needless_return, len_zero, unused_must_use, any correctness lint. These have objectively correct fixes.
• Tier 2 (fix when safe): context-dependent lints like dead_code, redundant_clone, too_many_arguments. Use judgment.
• Tier 3 (silence with #[expect]): legitimate suppressions at FFI boundaries, trait impls, serde fields, platform-specific code.
• Tier 4 (never silence): unconditional_recursion, invalid_value, dangling_pointers_from_locals, etc. These indicate bugs.

RAII footgun -- when fixing unused_variables on guards/locks/handles, use let _guard = expr; (lives to end of scope), never bare let _ = expr; (drops immediately).

Suppression syntax: Prefer #[expect(lint, reason = "...")] over #[allow] (stabilized Rust 1.81). If MSRV is below 1.81, use #[allow(lint)] with a code comment. Narrowest scope possible. Never put #[deny(warnings)] in source code.

Attribute-edit footgun: rustfmt has its own opinions about how long-form attributes break across lines, and they're not always derivable from max_width. In particular #[cfg_attr(cond, inner_attr = "...")] with a multi-argument inner attribute is reflowed onto multiple lines even when the single-line form fits in 100 cols. After hand-editing any attribute (or any line that came out of an Edit / replace_all), re-run cargo fmt --check for that root before declaring the warning fixed — don't trust eyeballed line lengths.

3d: Iterate

cargo clippy --all-targets [flags] -- -D warnings 2>&1

Fixing one warning can reveal new ones. Repeat until clean.

3e: Verify formatting still passes

Manual edits might introduce formatting issues:

cargo fmt --all --check

If it fails, run cargo fmt --all and continue.

Phase 4: End-to-End Verification

Skip if $ARGUMENTS is workflow.

Phases 2-3 verify individual concerns during the fix loop. This phase re-runs the relevant CI commands together to catch cross-concern regressions (e.g., clippy fixes that re-break formatting).

Replay the exact CI commands from Phase 1b for each CI-covered root, plus run the defaults below for any Phase 1a check-root not covered by CI. Use absolute paths and && chains in single Bash calls. Run only the checks matching the current focus:

• fmt: fmt check commands only
• clippy: clippy check commands only
• all: both, e.g., cd /absolute/path && cargo fmt --all --check && cargo clippy --all-targets [flags] -- -D warnings

If no CI workflow was found, use the defaults above for every check-root.

Use extended timeouts (600000ms) for commands involving compilation.

Every check-root from Phase 1a must pass. Before leaving Phase 4, print a per-root pass/fail table -- roots attempted must equal roots in the list, and every row must be green. If any fails, return to the relevant phase for that specific root.

4b: Cross-target clippy when platform-cfg-skewed

Skip if Phase 1f did not flag the project as platform-cfg-skewed.

Otherwise, for each distinct CI runner OS that differs from your host:

1. Install the target if missing:

   rustup target add <ci-triple>

2. Replay the CI clippy command(s) for each affected check-root with --target <ci-triple>:

   cd /absolute/path && cargo clippy --workspace --all-targets --target <ci-triple> [flags] -- -D warnings

Native-dep cross-compile may fail (cc-rs / libsqlite3-sys / openssl-sys / similar need a C cross-toolchain you probably don't have on Windows). Two fallbacks:

• Mini-repro: extract the smallest possible snippet exhibiting the cfg pattern into a fresh crate with no native deps and run cross-target clippy there. Sufficient for attribute / lint-eval skew (unused_attributes, dead_code, etc.) since those don't depend on linking.

  mkdir -p /tmp/cfg-check/src && cat > /tmp/cfg-check/Cargo.toml <<EOF
  [package]
  name = "cfg-check"
  version = "0.0.0"
  edition = "2021"
  publish = false
  EOF
  # paste the suspect cfg pattern into src/lib.rs, then:
  cd /tmp/cfg-check && cargo clippy --all-targets --target <ci-triple> -- -D warnings

• Symmetric-by-construction fix: when the fix replaces a stacked #[attr] + #[cfg_attr(not(X), attr)] with two mutually exclusive #[cfg_attr(X, attr)] + #[cfg_attr(not(X), attr)], exactly one attribute resolves on every target by construction. No cross-target run can prove or disprove this -- the symmetry does. Document the reasoning in the commit message.

If both fallbacks are blocked, document the gap in the Phase 6 report so the user knows host-target verification did not cover everything.

Phase 5: Harden CI

Skip if $ARGUMENTS is fmt or clippy.

5a: Create or fix the workflow

If no workflow exists, create .github/workflows/ci.yml using the templates in references/workflow-templates.md. Adapt to the project's flags (workspace, features, MSRV, etc.).

If a workflow exists but has issues, fix them:

5b: Validate toolchain consistency

1. Does rust-toolchain.toml exist? If not and the project is collaborative, recommend creating one:

   [toolchain]
   channel = "stable"
   components = ["rustfmt", "clippy"]

   Including components ensures contributors get them automatically via rustup.
2. Does the CI workflow specify a different toolchain than rust-toolchain.toml? They should agree. dtolnay/rust-toolchain respects the file if no explicit version is provided.
3. Does the project have an MSRV lower than the pinned toolchain? That's fine for development, but suggest a separate CI job that tests the MSRV if the project claims to support it.

5c: Validate badge

grep -E 'badge\.svg|actions/workflows' README.md 2>/dev/null

If no badge exists, derive the URL from the git remote and suggest adding it:

git remote get-url origin 2>/dev/null

Badge format:

[![CI](https://github.com/OWNER/REPO/actions/workflows/ci.yml/badge.svg)](https://github.com/OWNER/REPO/actions/workflows/ci.yml)

If a badge exists but points to a nonexistent or renamed workflow file, fix the URL.

Phase 6: Report

Provide a concise summary:

• Fmt: X files reformatted (or "already clean")
• Clippy: Y warnings fixed, Z suppressed with reason (or "already clean")
• Cross-target: triple(s) verified, or "host-only — no platform cfg detected", or "skipped — native deps blocked cross-compile, mini-repro used / symmetry argued"
• Workflow: created / updated / already correct
• Badge: added / fixed / already present
• Toolchain: aligned / recommended pinning / already consistent

If all checks already passed, say so and stop. Do not manufacture work.

This skill does not commit changes. Use /rust-commit after reviewing the fixes if you want to commit.

Guardrails

• Replay CI commands verbatim. Use the exact commands, flags, cd paths, and && chains from the CI workflow. Convert relative paths to absolute. Do not improvise your own command structure.
• Never split cd and cargo into separate Bash calls. The working directory does not persist between Bash tool invocations. Always chain with && in a single call: cd /abs/path && cargo ...
• Enumerate every Cargo.toml in the repo, do not skim. Use find (pruning target, node_modules, vendor, .git) plus a **/Cargo.toml Glob fallback. Classify each as workspace root, workspace member, standalone, or excluded. Print the final check-roots list before Phase 2.
• Cover every check-root. Workspace roots (via --workspace), standalone crates (direct cargo in their directory), and excluded crates (treated as standalone) all need their own fmt + clippy pass. Not just what CI already checks -- every crate in the folder tree.
• Crate count invariant. Roots processed in Phase 2/3/4 must equal the Phase 1a check-roots list. Report the tally; do not declare success otherwise.
• Host-target clippy is not enough when CI runs on a different OS. Code under #[cfg(not(target_os = "your_host"))] is invisible to your clippy run; the same lint may fire only on the CI runner. If Phase 1f flagged platform-cfg skew, Phase 4b is mandatory before declaring CI green. Common smoking gun: unused_attributes on #[ignore] stacked with #[cfg_attr(not(X), ignore)].
• Re-verify after every hand edit. Any time you Edit/Write a Rust file — even outside the main fix loop, even to fix a CI failure surfaced by the user — the last thing before reporting done must be re-running fmt and clippy on the file's check-root. Edits made after Phase 4's pass-table are the most common source of "fixed it, pushed it, CI red again" — usually rustfmt reflowing a hand-typed attribute. No exceptions: every edit cycle ends with cargo fmt --check && cargo clippy --all-targets [flags] -- -D warnings for the affected root.
• Only fix issues that appear in command output. Do not invent warnings.
• Read files before editing. Never guess at line numbers or content.
• Verify every fix compiles. Run cargo check after non-trivial edits.
• Grep before removing "dead" code. Macros and conditional compilation hide usage.
• Do not modify code unrelated to making CI green. No refactoring, no extras.
• If all checks already pass, say so. Do not manufacture work.