8000 fix(minimal): --network none drops the entry's egress and ingress · gominimal/minimal@6cea830 · GitHub
[go: up one dir, main page]

Skip to content

fix(minimal): --network none drops the entry's egress and ingress

Minimal Arch / arch-review failed Nov 21, 2026 in 0s

FAIL

Inputs. Re-review of #3619 at 6cea830f, on arch 4f2eb87d, which is the same arch commit as the last ruling. All four earlier rulings on this PR are FAIL, so since-conditions.json is empty and no condition is open. Their why text stops at 400 characters, so I read pr.diff whole and ruled on the full change again. since.diff has fix edits in box_spec.rs, create.rs and tests.rs, plus base-merge content (crates/minimald/src/box_manager.rs) that is not in pr.diff. No findings.json or arch-map.md is staged, so I found the section by search.

Governs: architecture.md, minimal.toml layout. Line 1264 says that [defaults] unions lists, merges [env_vars] by name, and overrides other tables with the entry's. Line 1294 says that a table is replaced whole, not merged key by key, and uses network as its example: a later network.egress drops an earlier mode. The only exceptions it names are the create-time network flags and [env_vars].

Ruling. The change makes network merge by key (merge_network, crates/mfile/src/expand.rs line 1320 onward). It adds a new rule that a later mode = "none" clears the inherited egress and ingress. It also rewrites BOX-048 and the spec 18 comment to say so, and cites gominimal/arch#188. The design of record does not contain that decision, so the spec edits bend the spec away from the design (rubric rule 9). No condition can make this conform: reverting the merge removes the fix for #3600. This is the same outcome as the earlier FAIL rulings on this PR, and the escalation routes the design question to a person. When the design adopts a per-key network merge, this change can be ruled on again against that line.

Not checked. Whether gominimal/arch#188 exists or is decided: it is not in the checkout, and the checkout is the record. I did not build or run the tests. The ruling does not depend on test results. The min task run agent-verify failure (otel_cli trace test) is out of this ruling's scope.

Needs a design decision

  • Issue #3600 reports that [defaults.session.network.bep] drops [defaults.network.egress]. architecture.md line 1294 says this is the designed behaviour: a later layer replaces the network table whole and must restate the keys it keeps. Must network merge by key between layers, as [env_vars] merges by name, and must a later mode = "none" clear the inherited egress and ingress? PR #3619 and its spec edits cite gominimal/arch#188 for this, but the design checkout at 4f2eb87d does not state it. Until the design changes, the bug fix cannot conform.

Ruled at 6cea830 against gominimal/arch@4f2eb87d5d12 by claude-opus-5-5; arch_rulings 1591, job 2529216c-b14c-47ad-aeee-5ff5dab2007f.

0