Repository navigation
fork-ci: build workspace bins before cargo test - #17
Conversation
Tests find helper binaries in other packages through cargo_bin(), and cargo test -p never builds another package's bins. CARGO_BIN_EXE_* is unset and the fallback lookup in target/debug misses, so every PR touching codex-app-server fails thread_start_unloads_after_requesting_connection with "could not locate binary test_stdio_server". There are about twenty such bins, so the test step builds every workspace bin once first. codex-voice-host is excluded: it needs GStreamer >= 1.28, which Ubuntu 24.04 does not ship. Upstream's Bazel build declares the same bins as data dependencies instead.
thedancingdeveloper
left a comment
thedancingdeveloper
left a comment
There was a problem hiding this comment.
Review: no blockers (head 18a8048)
One commit on fork/main feaf20e, +7 lines in fork-ci's test step. It adds a single cargo build --workspace --bins --exclude codex-voice-host before the per-crate test loop, so every helper bin the tests resolve through codex_utils_cargo_bin::cargo_bin() (I counted 20 distinct names, many from other packages, e.g. codex-rmcp-client's test_stdio_server) is in target/debug for the assert_cmd fallback. That covers the whole class, not just the one test. The codex-voice-host exclusion matches the existing workspace check. Tests that need it would fail only for a PR touching those crates, which we don't plan.
Merge condition: this PR's own fork-ci run is green. The real proof needs a PR that touches codex-app-server, which this one doesn't, so right after merge re-run #6 or #12 and confirm thread_start_unloads_after_requesting_connection_closes_during_mcp_startup passes. If it doesn't, this gets reverted or fixed before anything else merges.
Summary
Test plan