The contribution-check orchestrator carried ~70 lines of static report layout rules + a full example report in its context every run, plus ~14 lines of comment routing logic — all templated work well-suited to a cheaper model.

Changes

• Removed ### Report Layout Rules and ### Example Report subsections (~70 lines) from ## Step 2: Compile Report; replaced with a one-line report-formatter agent invocation that passes the verdict array, skipped_count, and an explicit run_url (${{ github.server_url }}/${{ github.repository }}/actions/runs/${{ github.run_id }})
• Replaced the 14-line ### Posting comments block (with examples/caveats) with a one-line comment-dispatcher agent invocation; clarified that the input is the JSON objects returned by the contribution-checker subagent in Step 1
• Added two inline ## agent: blocks at the bottom of the file (after {{#runtime-import shared/noop-reminder.md}}):

## agent: `report-formatter`
---
model: small
---
# groups verdict array → Ready/Needs-look/Off-guidelines markdown tables

## agent: `comment-dispatcher`
---
model: small
---
# filters verdict array → [{issue_number, body}] for non-lgtm, non-empty-comment entries

Both sub-agents use model: small. They are orthogonal to the existing file-based contribution-checker delegate and can run concurrently once subagent results land. Compiled clean with gh aw compile --strict (0 errors, 0 warnings).