multi-agent: paired model-routing eval for retrieval, review and mechanical-edit roles #6901

Problem

No published measurement covers Claude 5.5 models on this repo's subagent roles: codebase retrieval, code review and verification, and bounded mechanical edits. Routing defaults in plugins/multi-agent/reference/defaults.yaml therefore rest on judgment. Haiku 5.5 may take no role until that role passes this eval.

Proposal

After the Haiku 5.5 documentation PRs land, build a paired eval comparing these setups on the same tasks from this repo:

• Haiku 5.5 at high and xhigh
• Haiku 5.5 with an Opus 5.5 advisor, using Anthropic's Haiku coding prompt
• Sonnet 5.5 at low, medium and high
• Opus 5.5 at low and medium

Measure:

• pass rate, or verified findings
• tokens per completed task (input, output and thinking)
• advisor call rate
• wall time

State no prices. Use a judge from outside the candidate set, or Opus with its self-preference noted.

Acceptance

• An eval suite covers the retrieval, review and mechanical-edit roles
• Results are recorded and the routing defaults are confirmed or changed
• A role moves to Haiku only when it matches the current model's accuracy at no higher token cost per completed task