Plan/Build/Review pattern with multi-model subagents + cross-model GPT-5.5 review

[BaseInfinity]

Summary

A proven SDLC pattern using Claude Code subagents and Codex CLI for multi-model plan/build/review. Tested and validated on a real harness (caught a config bug on first run).

The Pattern

1. Plan      → Opus 4.8[1m] subagent  (planner.md — scope, risks, test plan, stop conditions)
2. Review plan → GPT-5.5 via Codex    (gpt5-advise.sh plan — cross-model second opinion)
3. Build     → Opus 4.6[1m] session   (token-efficient builder, max effort)
4. Review code → Opus 4.8[1m] subagent (reviewer.md — severity-ranked findings)
               + GPT-5.5 via Codex    (gpt5-advise.sh review — cross-model, diff-scoped)

Model Assignments

Role	Model	Context	Why
Planner	Opus 4.8[1m] (subagent)	Fresh per call	Smarter than builder, structured output
Builder	Opus 4.6[1m] (session)	1M, max effort	Token-efficient, does the actual work
Reviewer	Opus 4.8[1m] (subagent)	Fresh per call	Catches what builder misses
Cross-model	GPT-5.5 via Codex CLI	Separate process	Different lab, different blind spots

Key Design Decisions

• Subagents over advisor: Subagents get fresh context (cheap, focused). Advisor forwards full conversation (expensive at 250k+, leaks cost into subagent runs). Advisor removed from default config.
• GPT-5.5 reviews BOTH plan and code: Not just code review — the plan gets cross-model validation before any code is written.
• Reviewer is the gate: Nothing ships with open must-fix findings. Findings are must-fix, should-fix, or can-ship.
• No LGTM without tests: Reviewer flags missing test coverage explicitly.
• [1m] suffix required on all agent model IDs: Without it, agents silently run at 200k context. The reviewer caught this bug in its own config on its first run.

Agent Files

.claude/agents/planner.md

---
name: planner
description: Plans implementation approach before code is written.
model: claude-opus-4-8[1m]
---
[structured output: scope, assumptions, risks, steps, test plan, stop conditions]

.claude/agents/reviewer.md

---
name: reviewer
description: Reviews a diff for correctness bugs.
model: claude-opus-4-8[1m]
---
[structured output: severity-ranked findings with file, line, test evidence]

GPT-5.5 Integration

Uses gpt5-advise.sh wrapper around codex exec that:

• Passes model + effort explicitly (config is what drifts)
• Asserts the banner reports the requested effort
• Default effort: high (fast). xhigh for final audit only.
• Reviewer is diff-scoped and never re-reviews the same code

What This Enables

Copy .claude/agents/ to any repo and the pattern works immediately — global settings (model, effort, gateway) carry over from ~/.claude/settings.json.

🤖 Generated with Claude Code