[codex] Stabilize Windows Bazel test flakes

[jgershen-oai]

Summary

This PR fixes and stabilizes several test areas that were making the Bazel workflow noisy. The first three commits address Windows test flakes; the later commits address the cross-platform app-server/TUI failures that surfaced while CI was running.

1. Marketplace local source parsing

   • Treat platform-absolute marketplace sources as local paths so Windows drive-letter paths are not parsed as invalid git sources.
   • Add Windows-only coverage for backslash-relative and rooted local path forms.
   • Keep marketplace metadata/config tests on the production config-writing path so Windows paths are TOML-escaped the same way as real config updates.

2. PowerShell-dependent timing tests

   • Add a test helper for building shell command responses from raw command strings so tests exercise the same command parsing and quoting path used on Windows.
   • Use that helper for the thread unsubscribe sleep test, where command parsing/quoting could otherwise affect Start-Sleep timing.
   • Add -NoProfile to PowerShell exec-test command vectors so user or runner profile startup cannot perturb short sleep/cancellation timing checks.
   • Run the MCP shell-approval test command as a non-login shell command and give the Windows PowerShell command more execution budget.

3. Multi-agent and agent-control interrupt/history timing

   • Add a start handshake for the synthetic never-ending task before sending followup_task interrupt=true.
   • Avoid waiting for the child's CleanBackgroundTerminals bootstrap op to emit TurnStarted, since that op does not create a regular turn.
   • Give redirected-envelope and subagent-notification history assertions more time under loaded Bazel executors, and improve abort-wait diagnostics so future failures report unexpected aborted turns.

4. App-server Bazel initialize/startup races

   • Cap the app-server integration test binary at --test-threads=2 under Bazel to reduce simultaneous codex-app-server child-process startup pressure.
   • Raise the app-server integration test default initialize/read budget from 10s to 30s, matching the macOS/fs/app-list/command-exec failures that timed out before initialize returned under load.
   • Make the plugin-list fail-open test use a local mock ChatGPT base URL for featured-plugin warming instead of reaching real chatgpt.com.
   • Make proactive auth-refresh tests wait until startup refresh has actually hit the mock server before sending the request that expects the second refresh.

5. TUI memory-mode test isolation

   • Keep the embedded app-server test's sqlite_home aligned with its temporary codex_home, so Bazel and Cargo both assert against the same isolated state DB.
   • Mark state DB backfill complete before starting the embedded app-server, matching the app-server memory-mode test setup and preventing async rollout backfill from racing the memory-mode write.
   • Poll for the state DB memory-mode update to land instead of racing the asynchronous persistence path.

Root Cause

The original deterministic Windows marketplace failure came from the marketplace source parser only recognizing POSIX-style local path prefixes such as ./, ../, /, and ~/. A source like C:\\Users\\... fell through to the git source parser and failed with invalid marketplace source format. A related config test manually interpolated Windows paths into TOML, which let backslashes be interpreted as TOML escapes.

Separately, the slow/flaky Windows tests were timing-sensitive. Some tests built shell responses without going through the command parser used by Windows command execution, and short PowerShell commands can be distorted by quoting, login-shell profile startup, or too-small command execution budgets. The multi-agent/agent-control failures were test ordering races around asynchronous history writes.

The latest macOS/Windows app-server failures all timed out waiting for JSON-RPC initialize responses across unrelated tests. That points at test-process startup contention under Bazel rather than fs/config/plugin behavior, so the targeted fixes reduce app-server integration-test concurrency and give startup/read handshakes a realistic remote-executor budget. One plugin-list test also leaked a real chatgpt.com request through featured-plugin cache warming, which made its fail-open behavior depend on external network/auth latency.

The TUI memory-mode failure was test isolation plus state backfill/persistence timing: the test moved codex_home to a temp directory but left sqlite_home on the helper's original configuration, started an embedded app-server without marking state backfill complete, and then immediately read state that is written asynchronously.

Validation

• just fmt
• cargo test -p codex-core marketplace_add
• cargo test -p codex-cli marketplace_add
• cargo test -p codex-app-server suite::v2::thread_unsubscribe::thread_unsubscribe_during_turn_keeps_turn_running -- --exact
• cargo test -p codex-app-server suite::v2::plugin_list::plugin_list_force_remote_sync_returns_remote_sync_error_on_fail_open -- --exact
• cargo test -p codex-app-server proactive_refresh -- --test-threads=2
• cargo test -p codex-app-server -- --test-threads=2
• cargo clippy -p codex-app-server --tests -- -D warnings
• cargo test -p codex-core tools::handlers::multi_agents::tests::multi_agent_v2_followup_task_interrupts_busy_child_without_losing_message -- --exact
• cargo test -p codex-core agent::control::tests::spawn_child_completion_notifies_parent_history -- --exact
• cargo test -p codex-core exec_full_buffer_capture_ignores_expiration
• cargo test -p codex-core process_exec_tool_call_preserves_full_buffer_capture_policy
• cargo test -p codex-core process_exec_tool_call_respects_cancellation_token
• cargo test -p codex-mcp-server suite::codex_tool::test_shell_command_approval_triggers_elicitation -- --exact
• cargo test -p codex-mcp-server
• cargo test -p codex-tui app::tests::update_memory_settings_updates_current_thread_memory_mode -- --exact
• git diff --check
• bazel query //codex-rs/app-server:app-server-all-test
• bazel build --config=argument-comment-lint -- //codex-rs/app-server/tests/common:common //codex-rs/app-server/tests/common:common-unit-tests-bin

Notes:

• just fix -p codex-app-server was attempted, but this local environment denied Cargo's TCP lock listener before Clippy started. The non-fixing scoped Clippy check above passed.
• The branch has been rebased onto current main, which includes the cargo-deny dependency update.
• Local Bazel execution on this machine is currently blocked by the local Bazel/Rust toolchain setup, so the next meaningful full Bazel signal should come from remote CI.

[jgershen-oai]

Closing this broad CI-stabilization PR in favor of a narrower extraction stacked on #18000: Windows filepath handling plus shell command string construction only.