[Yathiskumar]

Summary

• Commands::Ruff handed its args to ruff_cmd::run without restore_double_dash, so clap's leading -- was gone before routing. rtk ruff -- check became ["check"], got routed as the check subcommand with no path, and linted the whole directory.
• ruff_cmd::run now restores it first, the same way golangci_cmd and dotnet_cmd do, so the line reaches ruff as typed.

Fixes #4702

command	ruff 0.16.6	rtk before	rtk after
ruff -- check	usage error, exit 2	lints all 60 modules, 120 issues, exit 1	usage error, exit 2
ruff check -- check	All checks passed!, exit 0	Ruff: No issues found, exit 0	Ruff: No issues found, exit 0

Test plan

• cargo fmt --all && cargo clippy --all-targets && cargo test
• Manual testing: rtk <command> output inspected (table above, real ruff 0.16.6 on macOS)

New tests/ruff_double_dash_test.rs stubs ruff on PATH and checks the exact argv it receives for -- check, check -- check and --isolated -- check. The first one fails on develop (["check", "--output-format=json", "."]) and passes with the fix. The other two pin that a -- already kept by clap isn't duplicated by the restore.

One unrelated lib test failed once in five local runs of cargo test --lib and passed on rerun. It doesn't touch ruff_cmd::run.

[CLAassistant]

All committers have signed the CLA.

[Protocol-zero-0]

Independently checked b1701fb5 against base 5ab62913 with real Ruff 0.17.0 on Linux. This fixes the leading-separator routing issue from #4702 without duplicating a separator already retained by clap.

• 21 scenarios / 63 native, base and head processes. ruff -- check, -- format, -- bad.py, -- --help and the other leading-separator usage/error cases now match native exit 2 and both output streams byte-for-byte. In particular, base invokes formatting for -- format and handles -- --help as help; head correctly preserves their native errors.
• Explicit check -- check, an ordinary failing file, a file named format, a path containing a space, format --check -- <file>, normal check and implicit-check controls retain the expected exits and diagnostics. The PR's three argv-capture tests also pass.
• 18 Ruff unit tests, 3 CLI integration tests, fmt and warnings-denied all-target clippy pass. No full-suite, cross-platform or pending-PR integration run is claimed.

Separate existing issue: check --isolated -- --statistics and check --isolated -- -- still append . and scan the directory (native exit 0, both base/head exit 1 in the fixture). This PR does not change that heuristic; I am keeping it separate from this leading--- fix and the routing work in #3968/#4654.

Approved for the scoped fix.

[Yathiskumar]

Thanks for the thorough check, especially the -- format case, which I hadn't caught. Agreed on keeping the --isolated -- --statistics path heuristic out of this one.

[KuSh]

Round 1 — APPROVED, follow-ups filed

Claim: rtk ruff restores a leading -- that clap strips, so rtk ruff -- <arg> reaches ruff as typed instead of being routed as a subcommand.
Scope: accept. #3933 and #3968 edit the same function and conflict with this by one line; this lands first and they rebase over it.
Ran: real ruff 0.16.6 (LC_ALL=C) × 15 argument forms × native / develop 5ab6291 / this head, on a 61-module project with files named check, format, -- and sub/; plus rtk -v/-vv before the subcommand, rtk rewrite on 4 forms, v1.0.1 on 4 rows. Invariants: under the hook, split streams (stderr byte-identical to ruff's on the 6 error rows), every decision. Savings unchanged (filter untouched): 86% on the 122-issue run, 28% on the 2-issue run, vs the 20% floor (CONTRIBUTING.md).
Pins: 1 decision (restore before routing), 2 mutated (restore removed; routing on the stripped args with the restored ones forwarded), 0 survived — both caught by ruff_leading_double_dash_reaches_ruff.
Matrix: executed this round; pin check: executed this round.
CI @b1701fb5: all green (15 pass, check-target skipped; run approved this round)

Blocks merge (0)

Filed as follow-ups

• #4752 rtk ruff check appends . when every path is dash-prefixed or contains = (rtk ruff check a=b.py lints the whole tree) — outside this PR's hunks, same on v1.0.1 (also noted in the review above); filed.
• #4753 A leading -- is dropped for 19 other passthrough commands (rtk pytest -- -k.py runs the wrong tests) — the same bug class, outside this PR; filed.
• #3025 — with this change, rtk ruff also reaches the std::env::args() panic when rtk is invoked by an absolute path from a non-UTF-8 directory (same as rtk git on develop; invoked by name it does not happen). Comment added; #3224 covers it.

Checked and correct — no need to re-verify

• rtk ruff -- check, -- format, -- sub, -- -- check, -- --version, -- alone: exit code and stderr identical to ruff. On develop -- check linted 61 files and -- format reformatted 61 files (also on v1.0.1), where ruff refuses both.
• check -- check, check -- format, check -- sub, format -- check, format --check -- check, --isolated -- check: same output and exit code as develop; clap already kept a -- that follows an argument.
• rtk -v ruff -- check and rtk -vv ruff -- check: the restore counts only the user region, Running: ruff -- check.
• ruff_cmd::run has a single caller (Commands::Ruff), so the read of the process argv cannot see another command's arguments.
• cargo fmt, clippy --all-targets (0 warnings), cargo test --all (5033 passed) on this head, which already contains develop 5ab6291.

Next

Nothing — approving.
Rounds: 1. Threads: 0 open.