Skip to content

Claude review stack conversion (roadmap 2026-07-26) #27

Description

@SSFSKIM

Problem & intent

Claude models have overtaken the GPT models the review stack was built around. This epic converts the stack's defaults to Claude end to end without losing review quality: methodology (codex exec review → argus-review, improved first), effort auto-routing, engine swap in the reviewing-prs loop + execplan exit-gate, and worker default routes (clodex gateway → plain Claude).

Roadmap spec (authority for children, contracts, ordering): docs/doperpowers/specs/2026-07-26-claude-review-stack-roadmap.md (v1.1, commit 5ecdd5b).

Constraints

  • Two repos: SSFSKIM/argus-review (methodology children C2, C3) and doperpowers (C4, C5, C6). The board lives here.
  • Non-regression is gated on the X1 benchmark (owned by C1); contracts X1–X4 in the roadmap bind the children — re-litigate none of them.

Success criteria

Parent-Level Acceptance in the roadmap spec, verified as a recomposition check when all children land (label-less review dispatch = Claude worker + argus engine + auto effort + bench pass; execplan/docs name argus first; label-less implement dispatch = Claude worker).

Roadmap

  • C1 spike: methodology comparison + X1 benchmark + headless probe (child ticket)
  • C2: argus methodology improvements (child ticket, blocked-by C1)
  • C3: effort auto-routing (child ticket, blocked-by C2)
  • C4: review engine swap in doperpowers (child ticket, blocked-by C2, C3)
  • C5: review-loop worker default → Claude (child ticket, blocked-by C4)
  • C6: implement worker default → Claude (child ticket, independent)

Decision log

  • 2026-07-26: 6-child cut approved by the human after grill + external review (roadmap Decision Log has the full record incl. rejected alternatives).

Metadata

Metadata

Assignees

No one assigned

    Labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions