Add Windows support

[staminaoo]

Summary

• Add Windows PowerShell installer with .cmd wrappers.
• Make sub2cli and sub2cli-inject work on Windows by using msvcrt instead of termios/tty.
• Add Windows Codex CLI provider switching by copying slot auth files into ~/.codex/auth.json.
• Keep macOS symlink/App profile behavior unchanged.
• Fix sub2cli-redeem --help to exit successfully.

Windows Usage Docs Included

• PowerShell install flow: ./install.ps1 plus python -m pip install --user requests websocket-client.
• ExecutionPolicy fallback: powershell -ExecutionPolicy Bypass -File ./install.ps1.
• Custom Python path: ./install.ps1 -Python "C:\Path\To\python.exe".
• PATH note for %USERPROFILE%\.local\bin and generated .cmd wrappers.
• Direct Codex CLI injection example: sub2cli-inject add-api https://relay.example.com/v1 sk-....
• Windows behavior note: writes %USERPROFILE%\.codex\auth.<slot>.json, copies active slot to %USERPROFILE%\.codex\auth.json, updates %USERPROFILE%\.codex\config.toml, and requires manually reopening Codex App if it is running.

Validation

• python -m py_compile sub2cli sub2cli-redeem sub2cli-inject
• install.ps1 generated runnable .cmd wrappers
• sub2cli --help
• sub2cli-redeem --help
• Tested Windows sub2cli-inject workflow with temporary CODEX_HOME: init -> add-api -> use local -> use relay -> current -> remove

[r266-tech]

Thanks for the Windows port — the structure (gating macOS-specific paths behind IS_WINDOWS) is clean and doesn't disturb existing macOS behavior. Found a few correctness issues that block merge.

Must fix (correctness / credential safety)

1. Slot switch is not atomic on Windows — sub2cli-inject cmd_switch

The Windows path copies auth.<slot>.json → auth.json before save_slots() persists data["current"]. If the process is interrupted between these two steps, active auth.json reflects the new slot but slots metadata still points to the old slot. The next switch will then run save_current_windows_auth() and write the new slot's auth back to the old slot's auth_file, corrupting the old slot's credentials.

Fix: write auth.json to a temp file first, persist slots.json, then rename — or save_slots(data) before the copy and recover from inconsistency on next launch.

2. install.ps1 writes .cmd wrappers in ASCII

Set-Content -LiteralPath ... -Encoding ASCII -Value $cmd

The wrapper embeds $Python (the Python interpreter path). If the path contains any non-ASCII character (Chinese username, non-Latin install dir), it's silently mangled and the .cmd points at a nonexistent executable. Chinese-locale Windows installs hit this routinely.

Fix: -Encoding UTF8.

3. _run_inject invokes python <bin> but _resolve_inject_bin now accepts .cmd — sub2cli

_resolve_inject_bin adds sub2cli-inject.cmd to candidates, but _run_inject always prepends the Python interpreter:

subprocess.call([sys.executable, inject_bin, "add-api", url, key])

If resolution picks the .cmd, this runs python sub2cli-inject.cmd which fails.

Fix: detect .cmd suffix and invoke directly, or skip .cmd from candidates and rely on the raw script on Windows too.

4. save_current_windows_auth() runs before quit_codex() — cmd_switch

Codex.exe may still be writing auth.json while we snapshot it; the read can be partial. Swap the order: quit first, then snapshot.

Should fix

5. Windows cmd_init accepts dirty state — early-returns "already initialized" as soon as data.get("current") and data.get("slots") is truthy, without verifying that auth.json actually matches the current slot's auth_file. The macOS path verifies via symlink target; on Windows a divergent state never self-repairs.

6. PR body references sub2cli-redeem — that script was removed from main before this PR rebased. The --help fix mention and py_compile validation line referencing sub2cli-redeem should be dropped to avoid confusion.

Minor (non-blocking)

• quit_codex Windows branch has unreachable return True after fail() (which sys.exits).
• read_input Windows path skips the 200ms multi-line drain — pasted multi-line input is truncated to the first line. Acceptable for URL/apikey entry but worth a comment.
• read_key Windows: unmapped special-key scan codes after \x00/\xe0 fall through to "back" — pressing F-keys unexpectedly exits a menu. Map unknown codes to "" (ignore) instead.
• .cmd wrapper embeds the resolved $Python path — moving Python's install location breaks the wrapper until install.ps1 is re-run. Document or default to python from PATH.
• APP_SUPPORT on Windows becomes %LOCALAPPDATA%\Codex, APP_PROFILE nests to Codex\Codex. Path is never accessed on Windows (gated) but is stored in slots.json's app_profile_dir — dead data. Consider blanking on Windows.

---

Happy to merge once #1–#4 are addressed. Thanks again for the contribution.

[staminaoo]

Thanks for the careful review. Addressed the blocking items and the relevant should/minor items in aa65b2d:

• Windows slot switching now prepares the target auth in a temp file, saves provider-slots.json first, then atomically installs/removes auth.json. This avoids auth.json pointing at the new slot while metadata still points at the old slot.
• cmd_switch now quits Codex before snapshotting the current Windows auth, so it does not read while Codex may still be writing.
• Windows startup/switch paths now recover a dirty auth.json by restoring it from the persisted current slot before any snapshot can overwrite the old slot.
• install.ps1 writes .cmd wrappers as UTF-8 instead of ASCII.
• _run_inject detects .cmd injectors and invokes them directly instead of running python sub2cli-inject.cmd.
• Removed the stale sub2cli-redeem references from the PR body earlier; validation now only covers the scripts currently present on main.
• Unknown Windows special keys now return "" instead of acting like Back.
• Windows slot metadata now stores a blank app_profile_dir, since that path is not used on Windows.
• README notes that the generated wrapper embeds the Python path and should be regenerated if Python moves.

Validation run locally on Windows:

• python -m py_compile sub2cli sub2cli-inject
• git diff --check
• install.ps1 generated UTF-8 .cmd wrappers with a non-ASCII Python path preserved.
• .cmd wrapper invocation path: sub2cli-inject.cmd add-api ... and current.
• Windows switch workflow with temporary CODEX_HOME: init -> add-api -> use local -> use relay.
• Dirty-state recovery: simulated metadata current=local while auth.json contained relay auth; cmd_init restored auth.json from auth.local.json without corrupting the local slot.

[r266-tech]

Thanks for the quick turnaround — slot atomicity, .cmd detection, quit-before-snapshot, and the install.ps1 UTF-8 are all correctly addressed in aa65b2d. A re-read surfaced one class of issue both of us missed on the first round, plus a small leftover.

Must fix (new — credential/data integrity on non-UTF-8 Windows)

1. Text I/O in sub2cli-inject runs in locale encoding, not UTF-8

Every read_text() / write_text() / atomic_write_text call in sub2cli-inject omits encoding=, so Python uses locale.getpreferredencoding(). On a Chinese Windows install that's CP936; Western European is CP1252. Codex CLI writes config.toml / auth.json as UTF-8 (Rust default). Cross-tool round-trip on a non-UTF-8 Windows machine corrupts any non-ASCII byte the moment we patch config.toml (custom base_url, user-added comments, a Chinese display_name that ends up in slots metadata, etc).

Sites (line numbers on aa65b2d):

• atomic_write_json — tmp.write_text(...) at L91
• atomic_write_text — tmp.write_text(text) at L107
• load_slots — SLOTS_FILE.read_text() at L255
• decode_auth_email — auth_file.read_text() at L318
• write_apikey_auth json compare — path.read_text() at L454
• patch_config reader — CONFIG_TOML.read_text() at L488
• config_clean_for_slot — CONFIG_TOML.read_text() at L537
• relay api_key compare — auth_target.read_text() at L809

Pass encoding="utf-8" consistently. prepare_windows_auth_temp uses read_bytes/write_bytes which is fine.

Should fix

2. cmd_remove --hard leaves auth.json.*.tmp behind

prepare_windows_auth_temp writes plaintext auth to auth.json.<pid>.<ns>.tmp. Normal exit paths clean up, but a crashed or kill-9'd switch leaves the tmp on disk with credentials. Suggest globbing CODEX_HOME.glob("auth.json.*.tmp") in cmd_remove --hard, or unconditionally on cmd_init.

Still outstanding from last round

3. PR body still references sub2cli-redeem (Summary, Validation × 2). That script was removed from main pre-rebase. Earlier you said this was cleaned up but the body wasn't actually edited.

Deferred (pre-existing, not introduced by this PR — happy to handle in a follow-up)

• No cross-process lock on slots.json + config.toml + auth.json. Two sub2cli-inject invocations racing can leave the three files describing different slots; Windows recover_windows_current_slot only realigns auth.json against current slot. macOS has the same hole. Not a blocker for this PR.
• patch_config → auth install window: if the process dies between the two, config.toml reflects the new slot but auth.json is still the old one. recover fixes auth, not config. Same shape on macOS pre-existing.

Once #1–#3 are in, happy to merge. Thanks again.