8000 Merge remote-tracking branch 'origin/main' into agent/3582-a-task-on-a-vm-backed-host-ip-box-runs-in-the-cl · gominimal/minimal@8dcc569 · GitHub
[go: up one dir, main page]

Skip to content

Merge remote-tracking branch 'origin/main' into agent/3582-a-task-on-…

Minimal Arch / arch-review succeeded Nov 19, 2026 in 0s

PASS

Inputs. Re-review of #3610 at 8dcc5698, on arch 8b824707, the same arch as every earlier ruling on this PR. The baseline is the PASS_WITH_CONDITIONS at 7a65a15d. since-conditions.json carries its C1, and since.diff starts at 7a65a15d. I checked that since.diff equals git diff 7a65a15d..8dcc5698 and that pr.diff equals git diff 2af0ca89..8dcc5698. No findings.json or arch-map.md is staged, so I found the sections by search. The task marker (D3574, fixes #3582, epic 1437) has no covers: line, so I rule on its three layers, as the baseline did.

The compare. It holds the PR commit e7c2be24 and two merges of main (#3536, #3537, #3538, #3542, #3543, #3566, #3568, #3590, #3608, #3613, #3616, #3618, #3622, #3628, #3629). Among this PR's files, the compare edits only crates/minvmd/src/net/egress_gate.rs, and only in e7c2be24. #3566 edits the hook-shell path in box_actor. It does not touch task_network_in_box_leaf, the run_task leaf in crates/minimald/src/env.rs, or FirstOpeningDrops. So the baseline status stands for the placement and cache layers.

Governs: specs/networking/deployment-and-egress-gateway.md §4.1 line 54. Every host_ip box is in the boxes cgroup subtree and takes the cohort address. Per-box attribution in the cohort is impossible. The deny-all verdict is decided on the box's own cgroup, and minimald must make the box's cgroup placement inescapable.

C1: met. (1) The FirstOpeningDrops doc comment now says "logged once per rule and source address on this relay" and "the latch does not name a box". The phrase "the source address on one relay names the box the drop is for" is gone. (2) The tracing::warn! message is the exact text C1 gave, with the fields source, destination, port and rule_matched. (3) The test is now first_opening_drop_is_logged_once_per_rule_and_source. It asserts that a second HOST_IP_BOXES_RULE opening from the node address writes no line, and it expects two lines, not three. The latch key is (rule, src), so the code matches the doc. No condition is open.

Ruling. Placement: a host_ip task joins its box's classifier leaf, and its plan follows that leaf's verdict (§4.1 line 54, own cgroup and inescapable placement). Gate: the first-opening line is keyed by rule and source, and it claims no box (§4.1 line 54, no per-box attribution). Cache: the design of record is silent on cache cleaning, so that layer passes by absence. The out-of-layer test edit in crates/minimald/src/rpc/tests.rs is unchanged since the baseline.

Precedent. I follow the baseline at 7a65a15d and depart from no ruling.

Not checked, and not design questions. The task marker and the PR body's Tests table still name node_plane_drop_is_logged_once_per_box. C1 renamed that test, so that filter matches no test at this head. Update the marker's test: line to the new name when the proof is run again. The PR body says min task run agent-verify timed out on minimald env::tests::minenv_run_forwards_stdin_to_the_task. I did not build, so I cannot say if this head clears it. CI decides that, not this ruling.

Ruled at 8dcc569 against gominimal/arch@8b824707cde6 by claude-opus-5-5; arch_rulings 1501, job e4117c1b-9cb8-4cb1-8cc8-0e7a66f02a04.

0