Fix transient SSH config scope for ProxyJump aliases

[sunxiuguo]

Summary

Fixes #1320. An unsaved SSH destination writes a temporary config whose source Include currently remains inside the destination's Host block. OpenSSH's implicit ProxyJump child reuses that -F file for the jump alias, so its Include is inactive and the child misses the jump's HostName/Port.

Insert Host * immediately before the existing source Include, keeping the typed destination block first. Typed user/port/options retain OpenSSH first-value-wins precedence; jump aliases can read the source config without inheriting destination-only options. The existing ConnectionRepository → daemon launch provider → native OpenSSH path is preserved.

Add seven integration-marked native ssh -G cases in tests/core/test_transient_connections.py. They call actual repository creation and daemon launch preparation, evaluate both the generated config and launch argv, and compare destination/jump settings with a synthetic root. Coverage includes typed user/port precedence, no destination keepalive leakage into the jump, paths with spaces, nested Includes, Match originalhost/Match all boundaries, missing roots, source immutability, mode 0600/0700 permissions and cleanup. Baseline: six failed, one passed. Patched: all seven passed.

API and architecture

• Not applicable: this change does not affect the frontend-neutral API or core/frontend ownership boundary

No public DTO, capability, authentication resolver or alternate connection path is added. No protocol-version or generated API update is needed. Independent Codex review approved the exact patch and final commit; a separate review confirmed the SFTP race attribution below.

Testing

Cloud Linux; Python 3.12.14, OpenSSH 10.0p2 Debian-7+deb13u4 and Ruff 0.15.17. New native cases use temporary configs and .invalid hosts, no real keys or exec predicates, and only ssh -G evaluation. No remote SSH sessions, ProxyCommand execution or authentication were attempted.

Passed checks:

• PATH=/workspace/sshpilot-1320/.venv/bin:/usr/bin:/bin .venv/bin/pytest -ra tests/core/test_transient_connections.py -m integration -n 0: 7 passed, 14 deselected; independently rerun with the same result. Restricted PATH prevents the generic integration hooks from finding Docker and sweeping containers.
• .venv/bin/pytest -ra tests/core/test_transient_connections.py -m 'not integration and not gui' -n 0: 14 passed, 7 deselected.
• .venv/bin/ruff check src/sshpilot/core/connections/transient.py tests/core/test_transient_connections.py: passed.
• git diff --check: passed.
• .venv/bin/python scripts/generate_api_artifacts.py --check: passed.

Broader checks were run and are not green:

• .venv/bin/pytest -ra -m 'not integration and not gui': head 8,312 passed, 13 failed, 95 skipped, 3 warnings; untouched base 8,313 passed, 12 failed, 95 skipped, 3 warnings. The expression also selects property/stress tests and opt-in e2e cases, which skip without their environment.
• .venv/bin/pytest -ra tests/core tests/architecture tests/daemon/test_connection_launch_provider.py tests/daemon/test_ssh_launch.py tests/daemon/test_ssh_launch_parity.py -m 'not integration and not gui' -n 2: 1,311 passed, 3 failed, 1 skipped. All three failures reproduce on the untouched base.
• .venv/bin/ruff check src/ tests/ scripts/generate_api_artifacts.py: three unchanged findings: RUF022 in src/sshpilot/api/models/__init__.py:373, UP012 in src/sshpilot/window_tabs.py:85, PIE807 in tests/test_connection_dialog_fixes.py:564. These files are independently verified byte-identical to base.

The twelve shared broad failures, each observed on both head and untouched base, are:

tests/test_effective_config_diff.py::test_default_config_path_excludes_system_defaults
tests/test_protocol_suggestion_messages.py::test_real_catalogue_formats_translated_reason_without_touching_diagnostic
tests/test_transfer_error_messages.py::test_real_french_catalogue_formats_capability
tests/test_macos_menubar.py::test_menubar_installed_at_startup_not_construction
tests/test_macos_menubar.py::test_menubar_never_installed_off_macos
tests/core/test_dependency_boundary.py::test_daemon_imports_only_allowed_edges
tests/daemon/test_sftp_session_death_stuck_tab.py::test_dead_connection_fails_service_and_signals_the_tab
tests/daemon/test_socket_security.py::test_insecure_directory_is_refused
tests/architecture/test_frontend_closure.py::test_active_frontend_has_no_direct_backend_process_or_secret_ownership
tests/architecture/test_frontend_closure.py::test_plugin_gap_identities_are_still_exactly_deferred
tests/sidecar/test_sidecar_stateful.py::test_sidecar_identity_reconciliation_normal
tests/sidecar/test_sidecar_stateful.py::test_sidecar_identity_reconciliation_stress

The thirteenth head failure was tests/daemon/test_transfer_batch_real_sftp.py::test_losing_the_server_mid_batch_stops_short_with_what_landed. It passed in the uninstrumented full baseline run; that result is preserved. Twenty additional ordinary isolated invocations per checkout all pass, as does the original eight-test module on each checkout.

Controlled native scheduling independently reproduces the saved failure on both base and head: when the worker resumes before the reader marks the client closed, real pipe writes produce EPIPE and the unchanged transfer runtime counts all five items handled, failing the original 5 < 5 assertion. Three controls per checkout fail identically; the opposite reader-first order passes three times per checkout and stops at one item with no per-item failures. The temporary diagnostic plugin changes only synchronization, fabricating no packets/errors/summary values, and is excluded from this PR. Relevant SFTP code and fixtures are byte-identical to base and bypass transient creation/launch. This establishes a pre-existing disconnect race; it does not turn the historical broad run green or fix that unrelated defect.

Limits: Python 3.13, real GTK/desktop, packaged deb/Flatpak/macOS, actual SSH/ProxyJump child sessions, authentication/agents/key use, Match exec/canonicalization, packaging/resource builds and remote CI are untested. Broader headless tests use their existing synthetic local files, sockets, pipes and binaries. Optional MCP/pexpect, GTK/Cairo/gettext and other environment-dependent checks skip where unavailable. No remote CI success is claimed.

AI assistance: implementation, regression-test development and separate independent reviews used OpenAI Codex. No CLA acceptance or DCO sign-off is asserted.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T01:24:07.233233Z	46dc453	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[mfat]

Many thanks for the PR, I've already fixed the issue. Closing.