[kylehgc]

Follow-up to the 2026-10-09 Conflicted Sync (fork PR #168). That PR left one item unaudited: fork code that auto-merged outside the conflict hunks. Auditing it found two kinds of difference from upstream. Gaps that upstream has not addressed stay. Behaviour upstream decided on purpose now follows upstream.

Dropped: upstream decided this on purpose

Fork behaviour	Upstream's decision	Result
Claude hook sets permissionDecision: "ask" on Default/Ask rewrites (e93cde8)	KuSh on rtk-ai#3018: upstream omits it on purpose so Claude Code asks and can remember the answer. rtk-ai#3031 was closed as no longer applicable.	hook_cmd.rs payload handling is upstream's again
Fallback for hook payloads keyed by input (c8d6e28)	rtk-ai#2535: no Claude Code hook event sends that payload, and the fallback widens what can trigger the hook	removed; tests/hook_proof_test.rs and both of its claims withdrawn
rtk init also registers a PowerShell matcher (cc5b6d3)	rtk-ai#2075: passing PowerShell through the bash rewriter mangles \ paths, takes over aliases like ls/ps/curl, and bypasses PowerShell deny rules	init/claude.rs, init/mod.rs byte-identical to e4f0509
-- before "$CMD" in the legacy hooks/{claude,cursor}/rtk-rewrite.sh (7db14a3)	rtk-ai#2475 was closed as legacy-only, with no more contributions wanted there. Upstream never had -- there. #168's claim that c6f484f dropped it was wrong and is corrected in the sync doc.	byte-identical to e4f0509
Prettier filter reads the stderr report and summarises it on stdout (fork PR #117)	rtk-ai#3772: upstream forwards that report on its own stream. KuSh chose rtk-ai#3571 for the remaining rtk format case.	upstream's filter, plus an adoption of rtk-ai#3571

Kept or fixed: gaps upstream has not decided

• git machine-output passthrough (46d16ea by 李冠辰, from fix(git): keep machine output raw rtk-ai/rtk#2573, and da407dc). This branch first dropped it (ab5d0f5) on a misreading of KuSh's close of fix(git): keep machine output raw rtk-ai/rtk#2573, then reverted that (afd4595): he lists the porcelain and --pretty=format: newlines as open bugs, and his "log --format=%H already prints exactly what git prints" holds only for short logs. On e4f0509, git log --format=%H silently stops at 50 commits, and git log -z --format=%H -5 comes back as 121 bytes of 205. Both cases are now in the proof test. KuSh's open fix(git): announce the cap on the filtered git log path rtk-ai/rtk#4152 would keep user formats capped but announce the cap; once it merges, the log half becomes a design difference and goes. That trigger is recorded in the sync doc. Related: without the passthrough, status -z goes through with_trailing_newline, and upstream's open fix(git): keep the trailing newline on filtered 'git status' output rtk-ai/rtk#3682 then adds a stray newline after the last NUL. I reproduced that on a build of its head and commented there.
• Tests:
  • detect_linter's tests no longer create directories in the checkout.
  • The hermes Windows fake now expects the -- the plugin sends.
  • The adopted prettier test now launches rtk through common::rtk_command().
• CI: stale.yml runs only on rtk-ai/rtk. On the fork the job's token step failed every week, and falling back to GITHUB_TOKEN would start closing the fork's own issues and PRs.

Commits

Commit	Author	Kind
c14c608 revert(hooks): drop the Claude ask decision and the input-key fallback	kylehgc	fork
42edde8 revert(init): drop the PowerShell matcher from rtk init	kylehgc	fork
88566c4 revert(hooks): restore upstream's legacy shell-script hooks	kylehgc	fork
ab5d0f5 fix(git): follow upstream on log formats, keep -z NUL-terminated	kylehgc	fork, reverted by afd4595
2abf372 test(lint): stop the detect_linter tests touching the checkout	kylehgc	fork
439a9dc test(hermes): expect -- in the Windows fake rtk invocation	kylehgc	fork
fb91e9c ci(stale): run the stale job only on rtk-ai/rtk	kylehgc	fork
3f2275e revert(prettier): restore upstream's prettier filter	kylehgc	fork
3d27402 fix(prettier): never report a failed run as formatted	guyoron1	cherry-pick of rtk-ai#3571 (aa5f31e9)
89df848 test(prettier): spawn rtk through common::rtk_command()	kylehgc	amendment
9cebd10 docs(fork): record the design alignment in claims, delta and sync notes	kylehgc	fork
afd4595 Revert "fix(git): follow upstream on log formats, keep -z NUL-terminated"	kylehgc	fork
09016a5 test(git): pin the stash help passthrough on the path run() takes	kylehgc	fork
67106bd test(git): prove log -z and uncapped --format reach the caller whole	kylehgc	fork
57e09a9, 51bf9ae docs(fork): keep the passthrough, with its trigger; heal the reverted commit	kylehgc	fork

The cherry-pick is byte-identical to aa5f31e9 except for one conflict line. The tracked label uses &display_args(args) (upstream's current form, which the sync brought in) instead of &args.join(" "), plus that function's import. The amendment is needed because the repo's own isolation scan, tests_that_spawn_rtk_isolate_their_data, fails on CARGO_BIN_EXE_rtk in tests/prettier_failed_run_test.rs. I watched it fail before the change.

Repro: release builds, upstream e4f0509 vs this branch

The prettier --check here is a fake: it prints Checking formatting... on stdout and [warn] src/bad.js on stderr, then exits 1.

upstream:  rtk format prettier --check src/bad.js   (exit 1)
           Prettier: All files formatted correctly
branch:    rtk format prettier --check src/bad.js   (exit 1)
           Prettier: 1 files need formatting
           1. src/bad.js
both:      rtk prettier --check src/bad.js -> stdout "Checking formatting...",
           report on stderr, exit 1

Machine output compared byte for byte with native git. The first block is a scratch repo; the last two rows are this checkout's full history:

                                      upstream  branch
git status --porcelain                DIFFERS   same
git status --porcelain=v2             DIFFERS   same
git status -z                         same      same
git log --format=%H -3                same      same
git log --pretty=format:%H -3         DIFFERS   same
git log -z --name-only -3             same      same
git diff --name-only/--name-status/--numstat/--raw   same  same

                              native   upstream   branch
git log -z --format=%H -5        205        121      205
git log --format=%H           107789       2050   107789   (upstream: 50 commits, no notice)

Hook and init output are identical to upstream on both binaries:

rtk hook claude  (git status, default mode)
  {"hookSpecificOutput":{"hookEventName":"PreToolUse","permissionDecisionReason":"RTK auto-rewrite","updatedInput":{"command":"rtk git status"}}}
rtk hook claude  ("input"-keyed payload)  -> no output, exit 0
rtk init -g --auto-patch  -> "matcher": "Bash" only
rtk rewrite "git log | wc -l"  -> exit 1 (pipeline producer stays raw)

Quality gate

Windows x64 host toolchain, on 51bf9ae (the head; the origin/develop merge touches only the README delta block).

• cargo fmt --all -- --check: clean
• cargo clippy --all-targets: 0 warnings
• cargo test --all: 0 failures (4049 unit tests plus every integration suite)
• cargo build --release: ok

tests/prettier_failed_run_test.rs and tests/stderr_only_failure_test.rs are #![cfg(unix)], so Linux CI is their first run.

Not done here

• Hermes/pi --: kept in the fork's plugins and not filed upstream. It only matters when a command is exactly -h or --help. Upstream's hermes rewrite doesn't apply inside Hermes yet (Hermes pre_tool_call hook: in-place args mutation is ignored, RTK rewrites never apply rtk-ai/rtk#3797, fix in fix(hermes): return modify directives for terminal rewrites rtk-ai/rtk#3833).
• Remaining fork gaps each already have an open upstream PR: fix(rewrite): never rewrite yadm to rtk git; keep bare git add a no-op rtk-ai/rtk#3414, fix(hook): delegate lint scripts to package managers rtk-ai/rtk#3289, fix(jest): consume the values a stripped --reporters flag would bind rtk-ai/rtk#3777, fix(cli): stop aborting on non-UTF-8 argv bytes rtk-ai/rtk#3224, fix(cargo): preserve compiler warnings in cargo test output on passing runs rtk-ai/rtk#2877, fix(hooks): honor permission rules for already-rtk-prefixed commands rtk-ai/rtk#3195, fix(cli): hand -h, --help and help to the tool each subcommand wraps rtk-ai/rtk#4348. See claudedocs/fork-improvements-candidates-2026-10-09.md.
• Proof table in .github/README.md: generated by fork-proof.yml, which I'll dispatch after merge.

🤖 Generated with Claude Code

[kylehgc]

Review round 1. An independent review agent read the branch with upstream's code-reviewer.md standards, and each finding was checked on real binaries.

Finding	Disposition
Dropping the git machine-output passthrough rested on a misreading of KuSh's rtk-ai#2573 close. On e4f0509, log --format=%H silently caps at 50 commits and log -z --format=%H -5 comes back as 121 of 205 bytes.	Fixed: reverted (afd4595) and both cases added to the proof test (67106bd). The trigger for dropping it later is rtk-ai#4152 merging, recorded in the sync doc.
Two sync-doc rows (prettier, PowerShell) were stale, and the candidates doc gave the follow-up credit for the is_rtk_prefixed fix, which landed in the sync.	Fixed in 57e09a9.
The README proof block still lists the withdrawn hook claims.	That block is generated, so I'll dispatch fork-proof.yml after merge.
git_machine_output_test runs in the checkout, so on a clean checkout the porcelain cases are vacuous.	This predates the branch. The unit tests cover it, and fork-proof's upstream worktree is dirty, so the claim still discriminates there. Not changed here.
hooks/README.md OpenCode section shows rtk rewrite --, but upstream's rtk.ts doesn't call it.	Already on develop, out of scope.

The reviewer also confirmed the following:

• Restored files are byte-identical to e4f0509.
• hook_cmd.rs keeps only the deny path and the stdin fail-open.
• The cherry-pick differs from aa5f31e9 only by display_args.
• no_invented_stdout_diagnostic_when_the_tool_used_stderr still passes under fix(prettier): never report a failed run as formatted rtk-ai/rtk#3571 (traced through guard_stdout).
• Every claim test and HEALED_SHAS entry resolves.

Re-gated on 51bf9ae: fmt, clippy and tests all green.