Lock Claude→Cursor→free model cascade in ai-review and ai-qa

[eduly-haldarwish]

Summary

Lock model IDs at the action level (remove model inputs) for ai-review and ai-qa, with cascade:

1. Claude primary
2. Cursor (claude/cursor/...) if Claude is blocked
3. Pinned OpenCode free models, then auto/best-free as last hedge

Also stamp the model that actually ran + a re-run hint on every review/QA comment.

Plan

Canonical implementation plan: https://github.com/EdulyCom/github-actions/blob/main/docs/superpowers/plans/2026-08-17-claude-cursor-free-model-cascade.md

Locked free tail

oc/nemotron-3.5-lightning-free,oc/nemotron-3-ultra-free,oc/deepseek-v4-flash-free,oc/mimo-v2.5-free,oc/laguna-s-2.1-free,auto/best-free

• Do not use auto/* as Claude/Cursor primaries (auto/claude-* resolves to Cursor/Grok on this gateway).
• Do not pin oc/big-pickle until billing is confirmed free.

Tasks

• Task 1: ai-review publish banner — model + re-run hint (publish.js + tests)
• Task 2: Resolve actual model from execution logs; wire Publish
• Task 3: Remove sonnet-model / opus-model / haiku-model; hardcode Claude + Cursor + FREE in ai-review
• Task 4: Remove qa-model; same cascade + comment footer in ai-qa
• Task 5: Scrub stale model-input docs

Notes

• --fallback-model helps overload/block responses; it does not fix silent ~27m gateway stalls (ADR 0005).
• Prefer free models with structured_output: true earlier in the free list (review uses --json-schema).

[eduly-haldarwish]

Implementation landed on branch feat/ai-review-model-cascade-51 (not yet pushed as a PR).

• Task 1 — publish banner model + re-run hint
• Task 2 — resolve actual model from execution logs
• Task 3 — lock Claude→Cursor→free in ai-review
• Task 4 — lock cascade + footer in ai-qa
• Task 5 — docs scrub

Unit tests: 285 passing (ai-review/lib + ai-qa/lib).

[eduly-haldarwish]

Superseded/completed by merged PR #54 (a65ff19), which landed the Claude→Cursor→free model cascade for ai-review and ai-qa (originally tracked via #52 as well).