Standard source checks
cargo fmt --all -- --check
cargo clippy --workspace --all-targets --locked -- -D warnings
cargo test --workspace --locked./scripts/verify-windows.ps1 runs the Windows set for you: formatting, clippy with warnings denied, check, the workspace tests, the all-targets tests, the fixture build, and the pseudo terminal probe. It stops on failure and preserves logs. Run it after activating the GNU tools with . ./scripts/enter-windows-gnu.ps1.
What the tests use
Tests prefer real child processes in real pseudo terminals over mocks of the process layer. The runtime tests use small local programs that print what a provider’s screen would print, and the provider wording they react to comes from the real decoders.
Tests that talk to an installed provider only open its interface and leave. They never send a prompt, because that would spend someone’s quota. A test that needs a provider that is not installed skips and says so; it never passes silently.
Latest results on Windows
- GNU toolchain (2026-10-31): formatting and clippy with warnings denied are clean, and one run of the workspace tests passed 437 tests with none failed, including the tests that open each installed agent. Earlier runs of this version stopped in the tests that open and stop a real agent, for reasons that were fixed (see below), or were stopped by the machine running out of memory.
- MSVC toolchain (2026-10-31): 436 tests passed and 1 failed. The failing test opens every installed agent and leaves with Ctrl+C. Codex 0.166 exits with code 1 after two Ctrl+C even when it is started alone, so the test fails on the agent’s behavior. The same test passed on an earlier Codex version, and it was left as it is.
What the real-agent tests found
Running every agent installed on the machine, instead of a list of them, found three defects that were fixed: a learned program that starts a subcommand which only opens a picker of earlier sessions (it looked like an agent, and the test that only opens and leaves could not tell), an agent that needs about twenty seconds to leave after Ctrl+C and was killed after five, and a cache of learned answers that kept a stale answer after the way answers are written changed.
Against real programs
The agents installed on the test machine (four) were opened through uuldana-cli with no configuration, each was left with Ctrl+C and ended cleanly, each was stopped the way a failover stops it and shut down gracefully, and --bypass-permissions reached each real process as that agent’s own flags. In a real Windows Terminal window the header, the /agent menu, resizing, and a switch in both directions between agents were exercised earlier in development, and a real usage limit of one agent moved a session to another. On 2026-10-31 the installed build was run in a real window with a stand-in program printing limit lines in front of the real agents: the work moved from Claude Code to Codex and back, a warning of 85% used moved nothing, and the wheel, the arrow keys, and Shift+Enter worked after each move.
Not run yet
- A failover at a genuinely exhausted quota since the per-agent wording was removed. That would use up an account, so it is covered with real child processes standing in for agents, and real usage-limit messages as test inputs.
- Any agent that is not installed on the test machine.
- A real agent run through uuldana-cli on Linux and macOS, and anything on ARM64 Windows. The tests do run on hosted Ubuntu, macOS, and Windows runners and pass there; the search of the path for agents is skipped on them, because a build server has no agents.
Stress run
The three tests that stop a real agent gracefully are sensitive to load: alone they passed six times out of six, and run in parallel on a busy machine about half failed. Treat a failure there on a loaded machine as a measurement of the machine, and run the test alone before suspecting the code.
./scripts/stress-phase1.ps1 is a long repeat campaign for the Windows process layer. The full five-campaign gate has not passed: one timing-sensitive test, which collects a large burst of output from a pseudo terminal, fails under heavy load. A terminal-interrupt test against a real agent’s interface also failed once in dozens of runs, and its cause is not established.