[guijacquemet]

command.quote(windows=True) broke texts with backslashes before a quote: C:\My Data\ came out as "C:\My Data\", whose closing quote is swallowed.

Now (_quote_windows, CommandLineToArgvW/MSVCRT rules): backslashes are literal; a run of backslashes before a " is doubled and the quote escaped; a trailing run is doubled before the closing quote. Text made only of letters, digits and -_.:/\=+, stays bare; everything else, including empty text, is double-quoted. Examples: C:\My Data\ -> "C:\My Data\\", x y\"z -> "x y\\\"z". The POSIX path (shlex.quote) and the signature are unchanged; output for texts without a backslash before a quote is unchanged, and test_shell_quoting and test_command pass unmodified.

Shared vectors for every host: tests/quote_vectors.json (31 {label, text, posix, windows} vectors: backslash, quote, space, unicode, percent, caret, ampersand, pipe/redirect, UNC, emoji, CJK, tab, !) and tests/test_quote_vectors.py (the function must equal every vector for both shells; every posix value must shlex.split back to the text; every windows value must be one argument for the existing parser oracle). F10 and its winquote:*, winproc:* and property entries are removed from the known failures, so those tests run normally.

Decision recorded (cmd.exe): the Windows terminal line targets cmd.exe, not PowerShell (use the Python snippet in PowerShell). Inside the quotes ^ & | < > are literal and need nothing; %NAME% is still expanded by cmd.exe when NAME is defined, an odd number of " upsets cmd.exe's quote tracking for later operators, and a line break cannot be pasted into cmd.exe. These limits are documented in docs/HOST_FEATURES.md, not escaped: a caret-based escape cannot be unit-tested without cmd.exe, and it would break the tests that parse quote output through the C-runtime oracle. If cmd-hardened output is wanted, those tests need a cmd stage and a Windows CI run.

tests/test_sabotage.py: the POSIX break patch moved with the line; the old Windows break is replaced by two (no escape of ", no doubling of trailing backslashes, guarded by the vectors test); the sabotage copy also copies the vectors file. Docs: ledger row, HOST_FEATURES.md.

Host follow-ups (Fiji and QuPath Groovy shellQuote, same change in both): the Windows branch returns the text unchanged if non-empty and every character is a letter, digit or one of -_.:/\=+,; otherwise " + body + " where a backslash is emitted as-is unless it is followed by " or ends the text (a run before a quote is doubled, plus one extra backslash before the quote; a trailing run is doubled); keep the POSIX branch equal to shlex.quote; copy quote_vectors.json into each repo's tests and add one test that loads it as UTF-8 and compares both shells. Vectors that fail on the current Groovy copies if they only escape ": spaced-trailing-backslash, spaced-two-trailing-backslashes, quote-after-backslash, quote-after-two-backslashes, quote-after-three-backslashes, unc. Napari uses Tools' command_line: no change.

Verified by the author on Linux, Python 3.11: black, ruff, mypy; the full suite (402 tests) has the two environmental failures plus one test (test_generic_integration.test_7_cooperative_cancel_no_leaks) that compares the global /tmp/lcjob_* folders while other agents ran in parallel (passes alone). I re-ran lint and the quoting, command, round-trip command and known-failure suites (34 tests OK). Not verified: anything on real Windows. The evidence is the parser oracle and the vectors; the real-process-start Windows tests run in this PR's CI.

🤖 Generated with Claude Code

https://claude.ai/code/session_01Y2SN8SuEakTdiWy3F2WUqN

---

Generated by Claude Code

[guijacquemet]

Closed: included in #47 (merged with a merge commit, so these commits are in main).

---

Generated by Claude Code