GeezerVison

LLM Description written by Claude Code:claude-opus-5-5

Refs #823 (slice 16: scan, socket.yml policy and apply-network suites). Uses Refs, so #823 stays open for the remaining PENDING_RAW_SPAWNS files.

Summary

e2e_scan.rs, e2e_socket_yml_policy.rs and apply/apply_network.rs now spawn socket-patch through tests/common/hermetic.rs and leave PENDING_RAW_SPAWNS (12 → 9 on main; 6 once #1510 also lands, and the two edits don't touch adjacent lines). The ratchet in spawn_env_hygiene.rs keeps them hermetic. Test-only; no product code and no workflow changes.

Root cause

Each file started the CLI with a bare Command::new and its own SOCKET_* handling, which had drifted from the shared builder:

• apply_network removed only SOCKET_API_TOKEN in 8 of its 9 spawns. On main, a shell that exports SOCKET_DRY_RUN=true SOCKET_ECOSYSTEMS=cargo fails 9 of its tests.
• All three sweeps let an ambient SOCKET_NO_CONFIG=0 through, so a developer's real socket login could reach the child, and none of them forced the update check off.
• The policy suite's fixture includes a gem but never scrubbed BUNDLE_*. It now uses Extra::Venv + Extra::Bundle.
• e2e_scan's npm children now drop ambient SOCKET_* too.

Two coverage fixes so the move doesn't weaken guards (from /code-review):

• offline_apply_with_token_makes_zero_network_requests: the builder forces telemetry off, which would hide a telemetry POST under --offline. The test sets SOCKET_TELEMETRY_DISABLED=0. Checked by removing is_offline_env() from is_telemetry_disabled: the test fails with mock server saw: ["POST /patch/telemetry"].
• apply_uses_locally_cached_blob_without_fetching pinned only SOCKET_API_URL to a dead port, but with no token the child uses the public proxy. It now pins SOCKET_PROXY_URL as well.

Test evidence

• Red, without the conversions (only the PENDING_RAW_SPAWNS removal) and with a hostile shell (SOCKET_DRY_RUN=true SOCKET_ECOSYSTEMS=cargo SOCKET_ORG_SLUG=hostile SOCKET_NO_CONFIG=0): spawn_env_hygiene fails no_new_bare_binary_spawns for all 3 files, and apply fails 9 apply_network tests.
• Green under the same hostile shell (plus VIRTUAL_ENV=/nonexistent): apply 135/135, e2e_socket_yml_policy 26/26, spawn_env_hygiene 16/16, e2e_scan non-ignored 20/20.
• e2e_scan's 11 #[ignore]d live-API tests need patches-api.socket.dev, which the agent sandbox can't reach, so they weren't run. CI doesn't run them on PRs either.
• cargo fmt --all -- --check and cargo clippy --locked --workspace --all-features -- -D warnings are clean. The test-target clippy notes are all pre-existing in prebuilt_common.

/code-review high

Fixed: the telemetry guard, the proxy pin, Extra::Venv in the pypi test, and the e2e_scan comment. Not fixed:

• Moving e2e_scan's private run / npm_run / assert_run_ok onto the common copies, and scrubbing SOCKET_* inside common::run_toolchain for every package-manager child: both are refactors of common beyond this slice. They're worth a follow-up slice under Spawn every CLI test child through one hermetic Command builder; 10 test files inherit ambient SOCKET_* today #823.
• Empty start commit: the claim protocol requires it, and squash-merge drops it.

CI cost

No workflow changes. A lean PR run is the same before and after (~28 Linux jobs, ~8 min).

🤖 Generated with Claude Code

https://claude.ai/code/session_01NSWn8iEQppkpg6sqyjhEx6

Generated by Claude Code