8000 fix(mfile): box the network refusal lists to shrink Error · gominimal/minimal@7e88067 · GitHub
[go: up one dir, main page]

Skip to content

fix(mfile): box the network refusal lists to shrink Error

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

FAIL

Inputs. Re-review of #3619 at 7e880675, on arch 8b824707. This is the same arch as the two FAIL rulings at 31c6b775. since-conditions.json is empty. No ruling on this PR set conditions, because both earlier rulings are FAIL. The why of those rulings is cut short in prior-rulings.json, so I read pr.diff whole and ruled on the full change again. since.diff holds two merges of main and four commits of the PR (f31d1f5e, 6bd48c84, 0458a540, 7e880675). Those four commits remove the sub-table merge of the earlier head. In its place they refuse the file. No findings.json or arch-map.md is staged, so I found the sections by search. The task marker (D3597, origin bug, fixes #3600, epic 1501) has no covers: line. I ruled on its three layers (expand, cli, docs), as the earlier ruling did.

Governs: architecture.md, minimal.toml layout. The Sections list (lines 1223-1224) and the Expansion paragraph (lines 1247-1253) govern the change. They say that each layer overrides the tables of the earlier layers. They say that network is replaced wholesale, and they name this exact case: a later network table "drops any mode an earlier layer supplied and must restate it". The only exit-3 failure that the paragraph gives in the expansion is a loadout variable conflict. The repo spec agrees with the design: BOX-048 (spec 29, line 285) replaces tables with the more specific layer's table, and only vars merges by name.

Ruling. refuse_dropped_network_keys (crates/mfile/src/expand.rs:569, called at line 415) makes a design-valid file fail with exit 3 (Error::NetworkTableDropsKeys, crates/mfile/src/error.rs:127). The design says that file expands, with the earlier keys dropped, and that min box spec renders the result for review. The PR's own test at expand.rs:1718 shows the effect. A task with network = { mode = "none" } is refused because [defaults.network.egress] exists. This adds a validation rule that the design and BOX-048 do not have, and the PR does not amend the spec. The earlier head's sub-table merge contradicted line 1253 too. Bug #3600 asks for an outcome that the design rules out: the earlier layer's egress must stay. No condition can make this task conform. If the refusal is removed, the task does nothing. If either fix stays, it contradicts line 1253. So the verdict is FAIL, and the escalation asks the owner to decide the merge rule for network. When the design changes, the task can resume against the new line.

Requirements. expand: deviates. cli: not addressed. create.rs and box_spec.rs do not change, and tests.rs pins the refusal, not the per-sub-table origins that the marker asks for. docs: not addressed. Neither spec 29 nor spec 18 changes.

PR text. The title ("keeps the network") and the Proof section say that min box spec dev shows both bep and egress. At this head the command exits 3, and box_spec_keeps_earlier_network_egress asserts that refusal. The test name says the opposite of what the test asserts.

Precedent. This follows the FAIL at 31c6b775 on this PR. The question is the same (the network merge rule of line 1253), with a different fix. I do not depart from it.

Not checked. I did not compile. agent-verify fails on otel_cli exec_run_attach_and_task_run_share_the_commands_trace, which this diff does not touch. I did not check whether other project files on main now fail to load because of the refusal.

Needs a design decision

  • Bug #3600: [defaults.session.network.bep] makes a session lose [defaults.network.egress], because architecture.md line 1253 replaces network wholesale across layers. Two fixes contradicted that line: a merge by sub-table (head 31c6b77), and an exit-3 refusal when a later table drops a key (head 7e88067). Which rule does the design want for network across [defaults], [defaults.] and the entry? Options: (a) keep the wholesale replace and close #3600, with min box spec provenance as the remedy. (b) merge network by sub-table, as [env_vars] merges by name. (c) refuse a later table that drops an earlier key, as invalid configuration (exit 3). BOX-048 must then follow the decision.

Ruled at 7e88067 against gominimal/arch@8b824707cde6 by claude-opus-5-5; arch_rulings 1455, job c77b5aad-2c87-40f9-88ce-c459dd010529.

0