Make the comment pass part of every PR's read

allod-agent commented

2026-09-27 04:57:07 +01:00

Member

Cross-family reviewers are now asked for comment findings as a named dimension beside correctness: added comments that explain what the code does, narrate history or an issue, or belong in a README or the PR body. The overseer folds such a finding into the one fix round it already runs. No length rule and no new read for the overseer. Nothing on a machine changes. If it is wrong, revert the two commits.

Why: in one session an R2 PR shipped multi-line comment blocks past a codex reviewer whose prompt asked for correctness only and said not to raise style, so the one reader with fresh context was told to stay quiet about exactly this. The owner caught it.

Second commit, at the owner's request on this PR: an earlier version also gave the integrator a comment pass on every PR, including R0 and R1. That put the judgment on the most expensive tier for every change and is dropped; those PRs get the integrator's ordinary read and the owner's, as before.

Not added to memory.md: the index takes a line only on two ledger sightings or a safety reason, and this is one session's lesson filed in the skill that owns it.

Risk

R0. Process text only.

Validation

.hooks/memory-budget --worktree: passed.

Cross-family reviewers are now asked for comment findings as a named dimension beside correctness: added comments that explain what the code does, narrate history or an issue, or belong in a README or the PR body. The overseer folds such a finding into the one fix round it already runs. No length rule and no new read for the overseer. Nothing on a machine changes. If it is wrong, revert the two commits. Why: in one session an R2 PR shipped multi-line comment blocks past a codex reviewer whose prompt asked for correctness only and said not to raise style, so the one reader with fresh context was told to stay quiet about exactly this. The owner caught it. Second commit, at the owner's request on this PR: an earlier version also gave the integrator a comment pass on every PR, including R0 and R1. That put the judgment on the most expensive tier for every change and is dropped; those PRs get the integrator's ordinary read and the owner's, as before. Not added to `memory.md`: the index takes a line only on two ledger sightings or a safety reason, and this is one session's lesson filed in the skill that owns it. ## Risk R0. Process text only. ## Validation - `.hooks/memory-budget --worktree`: passed.

allod-agent added 1 commit

2026-09-27 04:57:07 +01:00

Make the comment pass part of every PR's read 93afc7c3de

The integrator reads the comments a branch adds against the code-comment
rule before writing the PR body, since R0 and R1 PRs have no other reader
and an author does not see their own comment bloat. Cross-family reviewers
are asked for the same as a named dimension; a prompt that asks for
correctness only tells them not to raise it. Two PRs in one session shipped
paragraphs past both reads before the owner caught them.

allod-agent commented

2026-09-27 05:13:58 +01:00

Author

Member

Owner request before merge: the integrator's comment pass on every PR is too much; drop it from git-workflow.md and from the overseer section of the skill. Keep the cross-family reviewer's comment dimension.

allod-agent added 1 commit

2026-09-27 05:14:23 +01:00

Drop the integrator's per-PR comment pass; keep the reviewer's 9dc07d66ff

The owner's read: a comment pass by the overseer on every PR puts the
judgment on the most expensive tier for every change. The cross-family
reviewer keeps comments as a named dimension; the overseer folds a comment
finding into the one fix round.

allod-agent added 1 commit

2026-09-27 05:14:38 +01:00

git-workflow.md: restore the line break the removal swallowed 5d5af331e9

vnprc merged commit 5d5af331e9 into master

2026-09-27 05:21:30 +01:00

vnprc deleted branch agent/review-comment-pass

2026-09-27 05:21:30 +01:00

Rows
Columns

Make the comment pass part of every PR's read #69

Risk

Validation