Flaky full-suite test: non_2xx_install_fetch races on ALLOW_LOCAL_HTTP_ENV under parallel execution

[senamakel]

Summary

workflows::ops_install::install_fetch_tests::non_2xx_install_fetch_returns_err_for_4xx_and_5xx fails intermittently in the full-suite Rust coverage run with url scheme http not allowed; https only, even though it sets OPENHUMAN_SKILL_INSTALL_ALLOW_LOCAL_HTTP=1. It passed on a plain re-run — a process-global env-var race under parallel test execution.

Problem

Observed while a TinyAgents port PR (#4539) temporarily ran the full coverage suite (a Cargo.toml/Cargo.lock change triggers the config-level full fallback; normal PRs run domain-scoped and don't hit it).

• What happens: the test locks TEST_ENV_LOCK, std::env::set_var(ALLOW_LOCAL_HTTP_ENV, "1"), then calls install_workflow_from_url_with_home on a loopback http:// mock. Intermittently allow_local_http_install_url reads the env as unset and validate_install_url (src/openhuman/workflows/ops_install.rs:717) rejects the scheme.
• Expected: with the escape-hatch env set, the loopback URL passes validation and the test asserts the non-2xx status is surfaced.
• Impact: flaky red on any full-suite run; does not reproduce in isolation.
• Cause: std::env::set_var is process-global and racy across threads; TEST_ENV_LOCK only serializes tests that take the lock, so a concurrently-running env-reading/writing test can observe a stale value.

Repro (flaky): run the full lib test suite in parallel (cargo test -p openhuman / cargo llvm-cov -p openhuman), not the single test.

Solution (optional)

Remove the process-global env dependency from the hot path — e.g. thread an explicit "allow local http" flag through InstallWorkflowFromUrlParams / a config value for tests instead of reading std::env at validation time, or ensure every env-touching test in the crate shares TEST_ENV_LOCK.

Acceptance criteria

• Repro gone — the test is deterministic under parallel full-suite execution
• Regression safety — no reliance on process-global set_var for the local-http escape hatch on the tested path
• Diff coverage ≥ 80% — fix PR meets the changed-lines gate

Related

• Surfaced in feat(tinyagents): Phase 0/1 port cutover — pin v1.7.1, crate-back SchemaCleanr / model-context / observed stream #4539 full-suite coverage run (src/openhuman/workflows/ops_install.rs)