Content Review Gate
Review Edits Get Proven Again
The sprint pipeline stopped trusting path dirtiness when review changes content.
0 milestones 1 tasks 4 agents
2 plan 1 execute 1 review
113k peak context
16m 39s total 7m 53s per task
1m 59s plan 6m 36s execute 54s review
15 edits 75 commands
codex harness
This run fixed a narrow but important invariant in Gantry's sprint flow: if review changes project content, the gate has to bless that changed tree before the sprint can commit. The failure mode was subtle because execution had already dirtied the path, so a reviewer appending to the same file could leave the path-status view unchanged.
The work was small enough to collapse into one sprint, but it crossed production code and the fake worlds used by the tests. The resulting change made review edits observable through the same content-diff signal that production already had available, then taught the test doubles to distinguish a clean review from one that actually changes project files.
How this walkthrough is structured
Feature
What did the run build and ship?
The original driver files are gone, but the current pattern pipeline still verifies reviewer edits before commit and documents the repair path for a red post-review gate.
Build
How did Gantry structure the work?
The run collapsed to one sprint and stayed green, with the useful stress located in the extra after-review gate the brief required rather than in a repair loop.
Content Review Gate
2 agents 7m 53s wall time 113k peak context
The commit changed the old sprint driver to decide the re-gate from the review diff captured against a pre-review snapshot, and it added fake-world cases for clean reviews, editing reviews, red post-review gates, and the interpreted gated-task pattern. Those exact source files have since been removed, so their introduced lines no longer survive as isolated code in today's tree. The behavior is still visible in the active pattern stack: `config/patterns/sprint.toml` marks the tree before review, routes complete reviews through `verify_review_edits`, repairs or reverts a red review gate, and commits only after the post-review path is green. Current interpreter tests also exercise the pre-review mark and revert path, so the invariant was absorbed into the newer pattern engine rather than left as an obsolete driver-only patch.
This sprint carried the whole fix because the production decision and the fake-world contract were the same problem: review edits had to be represented as content, not as a dirty-path set. The brief kept the unit bounded by asking for one invariant, one test-double signal, and focused regressions around editing and non-editing reviews. The run facts show execution, review, and the additional gate all clearing without a repair pass, so the cut held even though it touched both runtime behavior and test scaffolding. The ledger later records a single-milestone collapse, which matches the digest's shape: there was no useful smaller handoff after this contract was proven.