Two latent defects surfaced during #27's post-fix suite runs, present on main as well — filed as a dedicated pass rather than a rider on the lint PR.
1. A connection is left open at httptest.Server.Close — every run, silent.
Each full-suite run shows httptest.Server blocked in Close ... in state active with a 5.0s block before the harness gives up. This is invisible while the suite is green — it only costs 5s per run — but it means some client/transport in pkg/opencode is not closed or its request not drained.
2. TestReconcile_Good_AdoptsMatchingContainer and TestReconcile_NoMatchingContainers_ZeroRecovered depend on wall-clock timing tightly enough to fail under load.
Measured on 2026-08-08 (fix/lint-109 branch, but latent on main): 2 failures / 8 full runs — both on the first two post-edit runs where cold-cache compilation ran in parallel with the tests; 6 consecutive greens once warm. main worktree: 0 failures / 5 runs. Failure signature: 5.0s timeouts + medium not mounted: default. The only production diff in the package (proxy.go Director→Rewrite) was A/B-tested in both directions and disconfirmed as the mechanism — the signature reads as CPU starvation against a 5s budget.
The fix shape is likely: find and close the leaked connection (1), and give the reconcile tests either a synchronisation point or a budget that survives a loaded machine (2) — not a bare timeout bump.
Findings and measurements by the agent-c7 session during PR #27.
Two latent defects surfaced during #27's post-fix suite runs, present on main as well — filed as a dedicated pass rather than a rider on the lint PR.
1. A connection is left open at httptest.Server.Close — every run, silent.
Each full-suite run shows
httptest.Server blocked in Close ... in state activewith a 5.0s block before the harness gives up. This is invisible while the suite is green — it only costs 5s per run — but it means some client/transport in pkg/opencode is not closed or its request not drained.2. TestReconcile_Good_AdoptsMatchingContainer and TestReconcile_NoMatchingContainers_ZeroRecovered depend on wall-clock timing tightly enough to fail under load.
Measured on 2026-08-08 (fix/lint-109 branch, but latent on main): 2 failures / 8 full runs — both on the first two post-edit runs where cold-cache compilation ran in parallel with the tests; 6 consecutive greens once warm. main worktree: 0 failures / 5 runs. Failure signature: 5.0s timeouts +
medium not mounted: default. The only production diff in the package (proxy.go Director→Rewrite) was A/B-tested in both directions and disconfirmed as the mechanism — the signature reads as CPU starvation against a 5s budget.The fix shape is likely: find and close the leaked connection (1), and give the reconcile tests either a synchronisation point or a budget that survives a loaded machine (2) — not a bare timeout bump.
Findings and measurements by the agent-c7 session during PR #27.