interrogate, arena, reflect, figure-it-out and architect fan out across models. Upstream
reads this roster from ~/.cursor/rules/pstack-models.mdc; on Claude Code it lives here. Tiers are
the ones in CLAUDE.md and the subagent-strategy skill (Sonnet 5.5 low / Sonnet 5.5 medium / Sonnet 5.5 high, Sonnet 5.5 high the top planning tier; reviewers per skills/HOSTS.md §Reviewer models);
OpenAI- and Gemini-backed tooling maps them through the table at the top of CLAUDE.md. Set model
explicitly on every Agent call.
Every lens uses skills/HOSTS.md §Reviewer models. The reviewer role overrides general tiers.
Each lens is its own agent, so the diversity comes from independent contexts and lenses, not from a
persona label. Reviewer independence follows the subagent-strategy skill: a lens never reviews code
it wrote, and the same lens agent persists across rounds of its own loop.
| Reviewer | Tier | Lens |
|---|---|---|
| A | Host reviewer mapping | correctness and money paths |
| B | Host reviewer mapping | security and boundary/input validation |
| C | Host reviewer mapping | concurrency, error handling, resource leaks |
| D | Host reviewer mapping | reuse and over-engineering (CLAUDE.md §1a and the sixth review dimension) |
Drop D on small diffs. Add a fifth independent host-mapped reviewer pass on anything touching money or tenant isolation.
Sonnet 5.5 high for each candidate whose design is still open. Sonnet 5.5 medium once the direction is settled and the candidate is execution against a decided shape.
skills/HOSTS.md §Reviewer models. The judge must not be the agent that produced the front-runner.
Sonnet 5.5 high for all three passes (judgment, tooling, divergent): reflection is a judgement task.