fix(core): forward stderr from stdout-only filters on every exit (adopt upstream #3772)

[kylehgc]

Closes #135. Adoption of upstream rtk-ai#3772 (Nicolas Le Cam / KuSh): every RunOptions::stdout_only() filter dropped the tool's stderr, and with stdout empty the never-worse guard collapsed the run to silence on both streams. Around 30 call sites (ruff, pytest, rspec, rubocop, gh, glab, go test, golangci-lint, tree, wc, psql, …).

Commits

Commit	Author	Kind
12db983 fix(core): forward stderr from stdout-only filters and count it	Nicolas Le Cam	Cherry-pick of 2eee910. One deviation, folded into the pick: the import hunk conflicted with the fork's use regex::Regex; (from upstream's bun/deno runner work); both imports kept. Body otherwise byte-identical.
4d3b417 test(core): pin stdout-only stderr handling	Nicolas Le Cam	Cherry-pick of 7b89028, byte-identical (tests/stderr_only_failure_test.rs, unix-only).
9ed2e49 fix(runner): drop the failure-only stderr forward superseded by rtk-ai#3772	kylehgc	Amendment. The fork carries an adoption of upstream rtk-ai#3029 (georgyia, c27bbd0, fork PR #17) that forwards stderr on non-zero exits only. rtk-ai#3772 subsumes it; left in place the two blocks print the diagnostics twice on failure. Its test file tests/stderr_passthrough_test.rs keeps the failure case; the success case flips from "stays suppressed" to "reaches the user" — the old assertion encoded rtk-ai#3029's scoping, which is exactly what rtk-ai#3772 changes.
90008ee test(core): pin the prettier stderr case to this fork's prettier contract	kylehgc	Amendment. Upstream's prettier is stdout-only, so rtk-ai#3772's prettier test expects the stderr report forwarded verbatim and an empty stdout. This fork's prettier (PR #117, upstream rtk-ai#2878) runs on the combined stream and summarises the [warn] report on stdout, so that test would fail on fork CI. Re-pinned to the intent: the report reaches the user (either stream) and no clean run is invented; exit code still checked.

Why the rtk-ai#3029 adoption goes

Upstream rtk-ai#3029 is still open, so this is not a heal. It is a partial supersession inside the fork: rtk-ai#3772 does everything rtk-ai#3029 did plus the success-exit case, and the two cannot coexist without a double print. git show upstream/develop:src/core/runner.rs | grep 3029 is blank — the block was fork-only.

Repro, real binaries (before = installed develop build, after = this branch), fake ruff.cmd / prettier.cmd first on PATH

ruff: stderr "ruff diagnostics on stderr", exit 0
before: stdout=[] stderr=[]                                  ← the bug: silence on both streams
after:  stdout=[] stderr=[ruff diagnostics on stderr]

ruff: stderr "ruff failed hard", exit 2
before: stderr=[ruff failed hard]           (#3029 path)
after:  stderr=[ruff failed hard]  1 line   (once — the amendment matters here)

ruff: stdout "stdout line" + stderr "stderr diagnostic", exit 1
after:  stdout=[stdout line] stderr=[stderr diagnostic]      (no replay onto stdout)

ruff: nothing, exit 0
after:  stdout=[] stderr=[]                                  (silent stays silent)

prettier: stderr-only "[warn] src/a.js" report, exit 1
before/after: stdout=[Prettier: 1 files need formatting / 1. src/a.js] stderr=[]   (fork #117 contract, unchanged)

Quality gate

x64 host toolchain: cargo fmt --all --check clean · cargo clippy --all-targets 0 warnings · cargo test --all 3233 unit tests passed, 0 failed, 8 ignored; all integration suites green. The two stderr integration test files are #![cfg(unix)] and compile away on Windows; fork CI (ubuntu, macos) is their verification. git diff --check clean.

Not done

• Upstream fix(core): stdout-only filters silently swallow a tool's stderr rtk-ai/rtk#3772 is KuSh's own and still REVIEW_REQUIRED; Adoption, no upstream obligation. When it merges, the next sync's conflict is confined to the import line and the removed fix(runner): surface a failing tool's stderr under stdout-only filtering rtk-ai/rtk#3029 block.
• Fork ticket for fix(runner): surface a failing tool's stderr under stdout-only filtering rtk-ai/rtk#3029's adoption is already closed (PR Adopt rtk-ai/rtk#3029: runner - surface failing tool's stderr under stdout-only filters #17); nothing to reopen — its capability survives inside fix(core): stdout-only filters silently swallow a tool's stderr rtk-ai/rtk#3772.
• Tracking: fix(core): stdout-only filters silently swallow a tool's stderr rtk-ai/rtk#3772 measures savings against the full raw output and counts forwarded stderr as emitted, so rtk gain for stdout-only filters will report slightly lower savings than before. Intended by the PR; not adjusted here.

🤖 Generated with Claude Code

[kylehgc]

Review round

Independent read-only reviewer pass over the branch, applying .claude/agents/code-reviewer.md. Verdict: approve, 0 blocking, 2 nits.

Verified

• Cherry-pick fidelity: 12db983 differs from upstream 2eee910 only in the import hunk's context lines; 4d3b417 is byte-identical to 7b89028. Authorship preserved.
• Every path through run_captured_filter emits stderr exactly once and never onto stdout: skip_filter_on_failure (early return), stdout-only with and without tee, no_trailing_newline, and the default combined-stream mode (stderr travels inside the filter input, not forwarded separately). Guard baseline stays stdout; tracking baseline is the full raw output plus forwarded stderr, as the PR states.
• All 30 RunOptions::stdout_only() call sites hand off to the runner and print nothing themselves; the eprint! hits elsewhere are exec_capture paths that never enter the runner. No double print anywhere. The removed fix(runner): surface a failing tool's stderr under stdout-only filtering rtk-ai/rtk#3029 block (9ed2e49) was the only other eprint! of raw_stderr on the filtered path, so its removal was required, not optional.
• tests/stderr_passthrough_test.rs::successful_tool_stderr_reaches_the_user fails on the pre-branch binary and passes after: a real regression test for the success-exit case, which fix(core): stdout-only filters silently swallow a tool's stderr rtk-ai/rtk#3772's own tests do not cover (all its fake tools exit non-zero except the silent one).
• No unix-gated unit test in src/** asserts stderr suppression or tracking numbers through the runner, so nothing the Windows gate skipped would flip.

Said plainly: the re-pinned prettier test (90008ee) passes both before and after the runner change, because this fork's prettier runs on the combined stream. It now pins the fork's prettier contract (PR #117), not the runner. The runner's success-exit case is covered by the passthrough test above.

Nits taken (e98a14b): upstream's six adopted tests spawned rtk without RTK_DB_PATH and booked fake-tool savings into the developer's real history database; they now use a per-test file in the tempdir. The prettier comment names rtk-ai#2878 as an upstream issue.

Nits declined: no Windows-runnable success-exit stderr test (both files are #![cfg(unix)]; fork CI runs ubuntu and macos). Behavioural note for the record: go test / golangci-lint stderr chatter on a success exit is now forwarded verbatim, as intended by rtk-ai#3772.

Gate re-run after the amendment: fmt clean, clippy 0 warnings, all tests green on the x64 host toolchain.