AI Setup Lanes

July 20, 2026 · View on GitHub

Adaptive GPT-5.6 guidance for repositories installed by this wizard. Sol high is the normal standing root driver for meaningful SDLC work. Terra and Luna remain available for bounded support work and measured experiments, but the wizard does not route ordinary coding away from Sol merely because a task looks routine.

This is guidance, not a hard lock. Maintainers may choose another profile explicitly, and update preserves that choice.

Current Model Baseline

Tier	Model	Standard API token price	Best fit
Sol	`gpt-5.6-sol`	$5 / 1M input, $30 / 1M output	Ambiguous, difficult, high-value coding and research
Terra	`gpt-5.6-terra`	$2.50 / 1M input, $15 / 1M output	Bounded implementation and supporting work with explicit verification
Luna	`gpt-5.6-luna`	$1 / 1M input, $6 / 1M output	Clear, repeatable, high-volume support tasks

OpenAI's model guidance starts demanding agentic work on Sol and positions Terra and Luna for lighter or clearer work. Price alone does not determine the driver: independent coding results and developer reports remain harness-sensitive and mixed, and this repo has no completed representative local slices proving that its Terra-led profile preserves quality. The quality-first default therefore stays on Sol high.

Adaptive Decision Stack

Make the decisions in this order:

Start meaningful SDLC work on Sol high. The root agent owns planning, implementation, verification, and synthesis.
Keep clear low-risk support work bounded. Terra or Luna may handle extraction, scans, summaries, or mechanical transforms when a Sol-owned task has an explicit verification boundary.
Escalate only the difficult slice to xhigh. Good triggers include security review, migrations, destructive operations, long-running research, and challenging coding where high leaves unresolved risk.
Use Max for one exceptionally hard task. Max gives one selected model more reasoning time.
Use Ultra when parallelism is the point. Ultra uses subagents and fits work that divides cleanly into independent workstreams.

Most tasks do not need xhigh, Max, or Ultra. None is a standing consumer default.

Roles Are Optional

Codex does not require a permanent advisor/driver pair. The Sol high root normally owns the full task. Add an explorer, reviewer, or planner only when specialization removes real uncertainty or creates useful parallelism. The workflow must remain correct when heterogeneous subagent routing is unavailable or inherited from the root session.

Setup A: Sol Quality-First

Recommended default for meaningful SDLC work.

Work	Model	Effort
Root agent / driver	`gpt-5.6-sol`	`high`
Review	`gpt-5.6-sol`	`high`
Difficult or high-risk slice	`gpt-5.6-sol`	`xhigh`
Exceptional single task	`gpt-5.6-sol`	Max, only when justified
Parallel independent work	Sol root plus task-appropriate subagents	Ultra, only when the task divides cleanly

Use this lane for normal agentic coding as well as architecture, complex features, releases, security-sensitive behavior, installer changes, migrations, and work that can damage a consumer repo. Repo scanning adapts escalation triggers and verification, not the standing driver.

Setup B: Experimental Mixed

Explicit opt-in profile for measured efficiency trials, not a routine-work recommendation.

Work	Model	Effort
Main pass	`gpt-5.6-terra`	`medium`
Review	`gpt-5.6-sol`	`high` via explicit command override
Difficult review or unresolved risk	`gpt-5.6-sol`	`xhigh`

Use mixed only when the task is bounded, verification is strong, and the maintainer accepts the quality/latency tradeoff. review_model selects Sol but does not override the profile's global reasoning effort, so run codex -c 'model_reasoning_effort="high"' review ... for the required review gate. Record representative results before promoting it. Existing repos that explicitly selected mixed keep that choice during update; the wizard does not select it automatically.

Setup C: Lightweight Support

Manual support lane for clear, low-risk, repeatable work. It is not a normal SDLC driver profile.

Work	Model	Effort
Bounded support task	`gpt-5.6-terra` or `gpt-5.6-luna`	lowest reliable effort
Integration and acceptance	Sol root	`high`

Good candidates include extraction, classification, structured summaries, tagging, and mechanical transforms. Secrets, security advisories, migrations, destructive bulk operations, and production-impacting changes are not lightweight support work.

Repo-Aware Escalation

Setup scans the repository and records risk surfaces such as deployment tooling, databases, CI, and firmware. Generated AGENTS.md keeps Sol high as the normal baseline and names the detected surfaces where an agent should consider xhigh for the affected task.

This is task-scoped. A repository containing a database does not need every documentation edit to run at xhigh; a schema migration in that same repository often does.

Wizard Model Profiles

The shipped profile names remain stable:

Wizard profile	Main model	Review model	Use
`maximum`	`gpt-5.6-sol` high	Sol high in the same session	Default and normal quality-first driver
`mixed`	`gpt-5.6-terra` medium	`gpt-5.6-sol` high via explicit override	Experimental explicit opt-in efficiency trial

maximum means the maximum model tier supplied by the wizard. It does not select Max reasoning. Luna remains manual because the lightweight lane is support work, not a standing root profile.

Fresh installs and profile-less updates select maximum. Updates preserve an explicitly selected mixed profile instead of silently overriding a maintainer decision.

Evidence And Promotion Gate

This decision is quality-based, not pricing-only:

official guidance supports Sol for demanding agentic work and lighter tiers for clearer supporting tasks
public benchmark and developer reports vary by harness, token use, latency, and repository shape, so no single leaderboard or Reddit post is treated as universal proof
the local benchmarks/model-experiment.csv ledger must contain representative completed slices before mixed can become a recommendation
until that gate passes, Sol high remains the default and mixed remains experimental

Maintainer Exception

The codex-sdlc-wizard repository itself remains on gpt-5.6-sol at xhigh while its Sol high benchmark is unfinished. This is a narrow repo-maintainer exception for unusually meta, downstream-enforcement work; it is not the consumer default.