EPS
← All tasks·#2074Completed

workflow-fix: split code-review into per-commit sub-verdicts for large rounds

kind: infra#wf-fix#wf-fix-fp:520e1490b1ce
track:

Highlight text on any card (body, plan, or an event) to anchor a comment, or leave a whole-task comment here. Mention @claude to summon a reply.

No comments yet.

Overview / Motivation

Auto-filed by the workflow-fix-on-bug protocol from a workflow-fix candidate raised on task #2054 (emitting agent: orchestrator observation after code-reviewer-lean thrash at Step 5 round 1).

Goal

Wire per-Unit-commit split-review shape into /issue Step 5 code-review dispatch for rounds with >4 commits (one code-reviewer-lean per commit against that commit's own diff).

Workflow gap

  • Bug observed: code-reviewer + code-reviewer-lean both autocompact-thrash on 200-400 KB round-scoped diffs from legitimately-sized (5+ unit commits) experiment rounds; deaths at ~4-10 tool uses. Task #2054 round 1: code-reviewer-lean thrashed at 9 tool uses / 385s / 3 autocompacts on a 244 KB diff (6,911 lines, 5 unit commits).
  • Why it is a workflow gap: CLAUDE.md § "Autocompact-thrash subagent deaths" cites this shape as the #2054 fail-loud trigger. The rule bars unbounded lean-twin retry AND reserves inline-adopt only for workflow-fix tasks fixing this failure mode. So for a legitimately-sized kind: experiment round the class produces NO durable code-review verdict — Thomas must recover manually. The Step 4b pre_split_multi_deliverable rule ALREADY splits IMPLEMENTATION over >4 deliverables into sequential units to avoid implementer thrash; the SYMMETRIC gap on the review side is unwired.
  • Confidence (emitter): medium — the shape is well-documented in CLAUDE.md; the exact split-review shape (per-commit vs per-file vs per-lens) is a design call the spawned session's planner should make.
  • verified-at-filing: grep -l "code-reviewer" .claude/agents/code-reviewer.md .claude/agents/code-reviewer-lean.md; grep -c "Step 5" .claude/skills/issue/SKILL.md → 6 hits per agent file (code-reviewer.md and code-reviewer-lean.md); 81 hits of "Step 5" in SKILL.md (2026-08-04)

Proposed change (candidate diff sketch — refine in planning)

# In .claude/skills/issue/SKILL.md Step 5, add a pre-split gate mirroring Step 4b's:
+ If the round-scoped diff exceeds N bytes (e.g. 100KB) OR the round has >4 unit
+ commits, split the review into per-commit sub-verdicts:
+   1. For each unit commit `<sha>` in `<parent>..HEAD`, spawn code-reviewer-lean
+      with a brief pointing at `git diff <sha>~..sha` (single-commit diff).
+   2. Collect per-commit verdicts (each PASS/REVISE).
+   3. Compose round-level verdict: ALL PASS ⇒ round PASS; ANY REVISE ⇒ round
+      REVISE (union blockers across commits).

Scope / surfaces

  • Primary targets: .claude/agents/code-reviewer.md (Step 0.9 subclasses; adds a per-commit-split subclass), .claude/skills/issue/SKILL.md (Step 5 pre-split gate), .claude/agents/code-reviewer-lean.md (brief-format prescription).
  • Grep the workflow surface for the pattern before editing.

Constraints / invariants

  • Workflow-surface only — never experiment code, configs/, or tasks/.
  • scripts/workflow_lint.py --check-asks passes; ruff on touched files passes; if workflow.yaml or CLAUDE.md change, they stay consistent with the rule file.
  • This session runs under EPM_WORKFLOW_FIX_SESSION=1 and carries a workflow_fix_target: Provenance line — it MUST NOT auto-route any of its own subagents' workflow-fix candidates (recursion guard).
  • The per-commit split must NOT re-introduce whole-branch reads that already fail (Step 0 / .claude/rules/diff-size-budget.md bind).

Provenance

  • workflow_fix_target: .claude/agents/code-reviewer.md,.claude/agents/code-reviewer-lean.md,.claude/skills/issue/SKILL.md
  • fingerprint: 520e1490b1ce

target_file: .claude/agents/code-reviewer.md,.claude/agents/code-reviewer-lean.md,.claude/skills/issue/SKILL.md bug_observed: code-reviewer[-lean] autocompact-thrashes on 200-400KB round-scoped diffs from legitimately-sized (5+ unit commits) experiment rounds; the fail-loud escalation produces no durable verdict for the class. why_workflow_gap: CLAUDE.md § "Autocompact-thrash subagent deaths" cites this shape as fail-loud, and Step 4b already pre-splits implementer over >4 deliverables — the symmetric review-side gap is unwired. proposed_change: Wire per-Unit-commit split-review shape into /issue Step 5 code-review dispatch for rounds with >4 commits (one code-reviewer-lean per commit against that commit's own diff). diff_sketch: |

In .claude/skills/issue/SKILL.md Step 5, add a pre-split gate mirroring Step 4b's:

  • If the round-scoped diff exceeds N bytes (e.g. 100KB) OR the round has >4 unit
  • commits, split the review into per-commit sub-verdicts:
    1. For each unit commit <sha> in <parent>..HEAD, spawn code-reviewer-lean
  •  with a brief pointing at `git diff <sha>~..sha` (single-commit diff).
    1. Collect per-commit verdicts (each PASS/REVISE).
    1. Compose round-level verdict: ALL PASS ⇒ round PASS; ANY REVISE ⇒ round
  •  REVISE (union blockers across commits).

confidence: medium related_task: #2054

Activity