fix(warp): Windows Git Bash/MSYS2 — probe tty, preserve paths, degrade without jq

[jacek4yang]

Problem

On Windows 11 + Git Bash/MSYS2 the Warp hooks break in three ways, all reproduced locally:

1. Every hook writes an error to stderr, which Claude Code surfaces to the user:

   .../emit-terminal-sequence.sh: line 82: /dev/tty: No such device or address
   

2. cwd and transcript_path are silently rewritten in the payload. Git Bash converts path-like arguments before jq sees them, so /Users/alice/my-project became D:/.../Users/alice/my-project and /tmp/transcript.jsonl became C:/Users/.../Temp/transcript.jsonl.
3. A missing jq surfaces as jq: command not found on stderr.

Failure 2 is also why tests/test-hooks.sh reports 55 passed / 2 failed on Windows today (noted as "unrelated" in #57).

Fix

Three small changes, 44 lines of script across 3 files:

• _tty_write() (emit-terminal-sequence.sh) — writes to /dev/tty with stderr redirected before the open: printf '%s' "$1" 2>/dev/null > /dev/tty. Redirections apply left to right, so the old > /dev/tty 2>/dev/null opened the device while stderr was still inherited and leaked the error. Probing and writing in one call also means the write stays safe if the tty goes away in between. A MSYS2 /dev/tty node exists but returns ENXIO on a detached process, so node existence alone is not a usable check.
• MSYS2_ARG_CONV_EXCL='*' (build-payload.sh) — stops Git Bash rewriting path-like arguments. jq treats them as opaque strings. No-op on Linux/macOS.
• jq guards — build_payload returns an empty body (callers read that as "no notification") and emit_terminal_sequence returns early. build-payload.sh is sourced, so it returns rather than exits.

Linux/macOS behavior is unchanged.

Verification

Windows 11, Git Bash/MSYS2 (MINGW64_NT-10.0-26200, bash 3.6.10):

Check	Result
bash tests/test-hooks.sh	74 passed / 0 failed (was 55/2)
stderr from the whole suite	0 bytes
7 hooks detached, no controlling tty	0 bytes stderr each
7 hooks with jq removed from PATH	exit 0, 0 bytes stderr each
version branches unset / 2.1.100 / 2.1.141 / 2.1.999	all 0 bytes stderr
POSIX + Windows-style paths through --arg/--argjson	preserved exactly
bash -n on all scripts	passes

Tests use the existing assert_eq/assert_json_field helpers and assert externally visible behavior: exact payload field values, exit status, and zero stderr.

Mutation testing (each retained test detects its regression)

• Remove MSYS2_ARG_CONV_EXCL → 4 failures (incl. the 2 pre-existing ones)
• Revert _tty_write to > /dev/tty 2>/dev/null → 10 failures
• Remove the jq guard → 1 failure

Running the new tests against unmodified main fails 16 of 17, so none are vacuous. One assertion is a deliberate control: it checks that the old unguarded pattern does leak on a host with no usable tty, so the zero-stderr assertions cannot pass for the wrong reason.

Existing PRs

• fix: Windows compatibility — skip /dev/tty, guard jq dependency #68 – also targets Windows and missing jq, but gates on _is_windows() (a platform guess rather than probing whether the terminal is usable), and does not address the MSYS2 path rewriting that breaks cwd/transcript_path.
• fix: silence /dev/tty redirect errors when no tty is attached #83 – correctly identifies the redirect-order problem and reorders to 2>/dev/null > /dev/tty. This PR makes the same ordering fix plus capability probing and path preservation.
• fix(warp): stop /dev/tty ENXIO error leaking to the user #75 – same redirect-order diagnosis, applied to the two emit_terminal_sequence sites; this PR routes them through a shared helper and also covers legacy/warp-notify.sh.
• fix: deliver OSC notifications when hooks have no controlling tty #44 / fix: TTY detection for hook subprocesses #19 – take a different approach (walking the process tree for an ancestor tty), which is complementary; both are larger changes.
• fix(warp): run hook scripts through bash on Windows #57 – hooks.json bash invocation, complementary and non-overlapping.

Happy to drop this in favor of any of the above if a maintainer prefers to land one first; the MSYS2 path preservation (MSYS2_ARG_CONV_EXCL) is the piece not covered elsewhere.

[jacek4yang]

Squashed to a single commit and trimmed to a 44-line script diff (3 files) plus tests.

Verified on Windows 11 + Git Bash/MSYS2: 74 passed / 0 failed (was 55 passed / 2 failed on main), detached hooks emit zero stderr, and the two pre-existing Windows path-preservation failures are fixed. No Linux/macOS behavior change.