fix(windows): restore CREATE_NO_WINDOW after d3d621f7c revert + add local-commit safety guard to hermes update

[Oc2cO]

Problem

Console-flash regression: commit d3d621f (June 2026, "revert(windows): roll back terminal-popup PRs #53791 #53810 #53829") bulk-reverted three earlier windows-hide-flag fixes across ~40 files (1284 lines), stripping creationflags=windows_hide_flags() from nearly every Windows subprocess spawn in the codebase. Only a handful of files were individually re-fixed since — everything else has been flashing a console window on Windows on every git/npm/ssh subprocess call for months, including on every single CLI/TUI launch via banner.py's startup version check.

Silent data loss on update: separately, hermes update's fast-forward-failure path does git reset --hard origin/<branch> when local and remote history have diverged, with no backup of the local HEAD first. Any committed-but-unpushed local work is silently discarded. This hit the same user twice in one week before the guard in this PR existed.

Fix

1. Restores creationflags=windows_hide_flags() (or windows_detach_popen_kwargs() for long-running detached processes) across every remaining file hit by the June revert: banner.py (startup path, highest priority — 7 sites), cli.py (24 sites), kanban_db.py, gateway/run.py, clipboard.py, claw.py, mcp_catalog.py, dep_ensure.py, doctor.py, setup.py, dump.py, gateway/platforms/webhook.py, plugins/platforms/raft/adapter.py, managed_uv.py, plugins/memory/mem0/_setup.py, plugins/memory/honcho/client.py, agent/proxy_sources/iron_proxy.py, tools/environments/docker.py, tools/environments/ssh.py, hermes_cli/uninstall.py, plugins_cmd.py, profiles.py, agent/secret_sources/{bitwarden,onepassword}.py, plugins/platforms/discord/voice_mixer.py, scripts/contributor_audit.py, profile_distribution.py. ~60 call sites total.
2. Also fixes plugins_cmd.py's _run_plugin_git helper — introduced by a more recent refactor of the plugin-update flow, so it never had the flag to begin with (found while rebasing this branch onto current main).
3. Also fixes apps/desktop/electron/ssh-connection.ts (2 unhidden ssh spawns for Remote/SSH backend profiles) via hiddenWindowsChildOptions().
4. Adds a safety guard to hermes update's reset path: before git reset --hard origin/<branch>, pins current HEAD to a hermes-local-fixes-backup branch and prints the recovery command, so committed local work is always recoverable after a forced reset.
5. Also includes a gateway RestartOnFailure fix (synchronous sh.Run + exit-code propagation) and an MCP _normalize_headers fix (headers config coerced to dict before use) that were bundled with the safety-guard commit originally.

Verification

• Every touched Python file re-parses cleanly (ast.parse + py_compile).
• Desktop app (apps/desktop) rebuilds clean with the restored fixes.
• Live-tested on Windows: flash confirmed gone on affected startup/git paths after rebuild.
• Deliberately left untouched: service_manager.py (Linux-only s6-overlay branch), agent/secret_sources/command.py (Windows early-return), agent/anthropic_adapter.py (Darwin-gated / intentional interactive OAuth), scripts/profile-tui.py (imports pty, doesn't import on Windows at all), hermes_cli/relaunch.py (intentional visible self-relaunch).

Reported and fixed by @Oc2cO.

[trevorgordon981]

Mostly-mechanical restoration of the bulk-reverted Windows fix, plus two substantive additions (the VBS bWaitOnReturn change and the update backup guard). The restorations are fine. The two additions need work.

1. The Run(..., True) wait may be a no-op for the GUI-subsystem target (core claim likely unproven)

gateway_windows.py runs the gateway via pythonw.exe (the test asserts "pythonw.exe" in content). WScript.Shell.Run with bWaitOnReturn=True only blocks for console-subsystem executables — for a GUI-subsystem app it returns 0 immediately once the shell launches it, without waiting or returning the real exit code. If that's the case, the entire premise of the fix ("RestartOnFailure can finally see the real exit code") is defeated: wscript still returns 0 within milliseconds and a crash still isn't propagated.

The existing test only checks the text of the generated script (", 0, True)", "WScript.Quit exitCode") — it validates the content, not the behavior, so it passes even if the wait silently no-ops on the GUI-subsystem target. Needs a real Windows runtime verification that Run(..., True) on a pythonw launch returns the actual exit code.

2. The update backup guard degrades to a warning and still wipes data

update_cmd.py: when git branch -f hermes-local-fixes-backup HEAD fails, it prints a warning and falls through to git reset --hard anyway. So in exactly the scenario the PR claims to prevent (silent data loss of local commits), data can still be wiped and the user is merely told "may be lost." If the goal is a hard guarantee, the reset should abort (or the guard retry) when the backup ref can't be created.

Also: it's a single hardcoded ref pointing only at the current checked-out HEAD — commits on other local branches are still orphaned by the reset. And it's never rolled back, so it accumulates and would silently -f overwrite a user's own ref of that name.

3. _normalize_headers silently swallows malformed config

tools/mcp_tool.py: strings are split on the first ": " with no .strip(), and dicts missing name/value are dropped silently. "Key:Value" (no space), " Key: Val" (leading space), or {key,val} keys all vanish rather than surfacing. Behavior change too: dict([...]) previously raised ValueError on misuse (loud failure); now it's quiet misbehavior. That's arguably better UX but turns a crash into a silent misconfig — explicit decision + a unit test warranted for new parsing logic with no coverage.

4. Scope creep in raft/adapter.py on POSIX

_spawn_bridge now uses windows_detach_popen_kwargs(), which on POSIX adds start_new_session=True (setsid) — a real behavioral change beyond "restore hide flags," and on the bridge's process group at that. The PR is framed as re-adding hidden windows; call this out explicitly or it'll surprise maintainers expecting a no-op on non-Windows.

Tests

Minimal: one test file touches the VBS text content only. Nothing covers the backup guard (most safety-critical), _normalize_headers branches, the Electron resolveGitBinary reorder, or the ~60 windows_hide_flags() additions. The two non-mechanical changes (#1 backup guard, #2 VBS wait) are exactly where coverage is missing.

[trevorgordon981]

Ping — I don't see a reply or new commits since my review. Four findings still stand, two of them (1 and 2) blocking from my side:

1. VBS Run(..., True) wait may be a no-op for the GUI-subsystem target (blocking). gateway_windows.py launches via pythonw.exe. WScript.Shell.Run with bWaitOnReturn=True only blocks for console-subsystem executables; for a GUI-subsystem app it returns immediately without the real exit code. If so, the whole "RestartOnFailure can see the real exit code" premise is defeated. The current test only checks the generated script text, not behavior. Needs a real Windows runtime verification that Run(..., True) on a pythonw launch propagates the actual exit code — or an explicit note on why GUI-subsystem still works.

2. Update backup guard degrades to a warning and still wipes data (blocking). update_cmd.py: when git branch -f hermes-local-fixes-backup HEAD fails it warns and falls through to git reset --hard anyway — the exact data-loss the PR claims to prevent. The reset should abort (or retry the guard) when the backup ref can't be created. Also: single hardcoded ref at current HEAD only (other local branches still orphaned), and it's never rolled back / would -f overwrite a user's own ref of that name.

3. _normalize_headers silently swallows malformed config. Strings split on first ": " with no .strip(); dicts missing name/value dropped silently. "Key:Value", " Key: Val", or {key,val} keys vanish. This also flips loud ValueError into silent misbehavior — explicit decision + unit tests for the new parsing (currently no coverage).

4. Scope creep in raft/adapter.py on POSIX. windows_detach_popen_kwargs() adds start_new_session=True (setsid) on POSIX — a real behavior change framed as "restore hide flags." Call it out explicitly, or keep it Windows-gated.

Tests: coverage is still the gap — nothing exercises the backup guard (most safety-critical), _normalize_headers branches, the Electron resolveGitBinary reorder, or the ~60 windows_hide_flags() additions. Findings 1 and 2 are the must-haves; 3 and 4 need at minimum an explicit decision and ideally tests.

[Enough1122]

AI code review — automated review for reference, author can ignore or act on any point.

fix(windows): restore CREATE_NO_WINDOW after d3d621f revert + add local-commit safety guard to hermes update

1. hermes_cli/update_cmd.py — the local-fixes backup uses a fixed ref name hermes-local-fixes-backup with git branch -f, so every later divergent update overwrites the previous backup. If update A pins the backup, the subsequent reset destroys the local commits, and a later update B re-pins the ref to the already-reset HEAD, the original local work becomes unrecoverable. Consider a timestamped backup ref (or only creating it when the ref does not already exist). Also, branch -f clobbers a user's real branch if they happen to have one with that exact name — a refs/hermes-backup/... namespace would avoid the collision.
2. bin/hermes-acp.exe and bin/hermes.exe — new binary files in the tree with no source provenance in the diff. For a supply-chain-conscious repo these deserve an explicit statement of how they are built and a repeatable build step; otherwise the binaries become opaque update vectors. If they are Windows-only, consider documenting the exact commit they were produced from.
3. apps/desktop/electron/main.ts git-binary resolution — preferring bin\git.exe over the cmd\git.exe shim is reasonable, and every root block now probes bin before cmd consistently. Minor: the candidate list can contain duplicates (e.g. localAppData\hermes\git\bin vs localAppData\Programs\Git\bin when hermes is a per-user Git install); deduplicating keeps the probe order predictable.
4. hermes_cli/gateway_windows.py VBS change — bWaitOnReturn=True lets RestartOnFailure react to the gateway exit code, but a hung gateway (process alive, no progress) still never exits, so the restart policy cannot fire. Complementary hang detection (heartbeat + timeout watchdog) would close that gap; out of scope here but worth a note in the issue this fixes.