be-review: consolidation degrades to a serial reconcile marathon on single-file PRs (overlap is the common case, not rare)

[srid]

Problem

The Consolidate phase took ~5 minutes on the #1165 dogfood. Root cause: the design assumes "the common case is no overlap (clean picks); the rare overlap surfaces as a cherry-pick conflict." For be-review's own PRs that assumption is inverted — the change is concentrated in one file (be-review.workflow.js + SKILL.md), so all three tracks edit the same lines:

codex  → be-review.workflow.js, SKILL.md
lens   → be-review.workflow.js, SKILL.md
police → be-review.workflow.js, SKILL.md

Consolidation replays each track's commits onto the branch in sequence (codex 3 + lens 5 + police 5 = 13 commits), and with everyone on the same lines nearly every commit after the first conflicts. The single consolidate:cherry-pick agent then hand-reconciles them serially. Measured on this run:

• 246 KB agent transcript, 110 tool calls
• 15 git cherry-pick invocations
• 48 conflict/cherry-pick --continue cycles

So consolidation is effectively O(overlapping commits) of serial read-both-sides → merge → continue, in one Sonnet agent. It's correct (preserves the per-commit ledger, never silently clobbers) but slow exactly when the PR is single-file — which is most of our agent-tooling work.

Why it's not just "rare overlap"

The "rare overlap → reconcile" path was designed for PRs that span many files where tracks touch different files (clean picks). A single-file PR makes overlap the default, turning the reconcile fast-path into the hot path.

Possible fixes (weigh tradeoffs)

1. Squash each track to its net diff, then 3-way merge the final per-track file states instead of replaying 13 individual commits. ~3 merges instead of ~13 cherry-picks → far fewer conflict cycles. Cost: loses the per-commit debate ledger in history (the per-track PR comment already carries the narrative, so this may be acceptable).
2. Detect high overlap up front (all live tracks touch a shared file set) and switch to a single reconcile of the combined result rather than per-commit cherry-pick.
3. Parallelize across non-overlapping file groups — cherry-pick file-disjoint commit subsets concurrently, reconcile only the shared files. Helps multi-file PRs, not single-file ones.
4. Keep as-is but set expectations — document that consolidation is slow for single-file/high-overlap PRs; it's the price of per-commit cherry-pick correctness.

#1 is the highest-leverage for the single-file case. Worth measuring reconcile-cycle count before/after on a be-review self-PR.

Acceptance

• A /be-review run on a single-file PR (all tracks touching the same file) consolidates in materially fewer reconcile cycles / less wall-clock, with the same final tree.

Filed from the #1165 dogfood, where Consolidate ran 48 reconcile cycles on a single-file diff.

🤖 Generated with Claude Code