AI Setup Lanes

June 11, 2026 · View on GitHub

Three recommended AI setups for this repo. Setups A and B are complete triads: planner → driver → reviewer. Setup C is a lightweight driver-only lane for operational grunt work.

This is guidance, not a hard rule. Maintainer override is always allowed.

Setup A — Research Premium

Role	Model
Advisor	Fable 5 (via `advisorModel: "fable"` in project settings — auto-consults at key decisions)
Driver	Opus 4.6 max
Reviewer	Codex (GPT-5.5) xhigh
Escalation	+ Fable 5 review (security, releases, wide-blast architecture only)

Quality-first lane for research work. Fable 5 advises automatically at key decision points (source evaluation, confidence calibration, methodology design, blast-radius) via native advisorModel (v2.1.170+), Opus 4.6 max implements (stable, Max-bundled), GPT-5.5 xhigh reviews (cross-family, free on ChatGPT sub, catches blind spots Fable can't see in its own work). Escalate to Fable review for the ~5% of PRs where stakes justify it.

Effort levels: Opus driver at max (standing default). Fable advisor runs at its own effort level server-side. If switching driver to Fable temporarily (fallback), use /effort high — Fable high already exceeds prior models at max. Unset CLAUDE_CODE_EFFORT_LEVEL env var if it forces max.

Setup B — Research Saver (OpusPlan)

Role	Model
Planner	Opus 4.6 max (via Plan Mode — Shift+Tab)
Advisor	Opus 4.6 (via `advisorModel: "claude-opus-4-6"` — compensates for Sonnet driver)
Driver	Sonnet (latest, auto execute mode)
Reviewer	Codex (GPT-5.5) xhigh

Cost-efficient lane using CC's native opusplan alias. Opus reasons during Plan Mode (Shift+Tab), Sonnet executes. Both 200K context, Max-bundled — no API credit drain. Pin model: "opusplan" + advisorModel: "claude-opus-4-6" + CLAUDE_CODE_EFFORT_LEVEL=max in project settings. The Opus advisor auto-compensates for Sonnet's lighter reasoning at key decision points. GPT-5.5 xhigh is the cross-model reviewer.

When to Use Setup A

Reach for Premium when the change can damage a consumer repo or has high blast radius:

Primary research drafting (evidence dossiers, investigation reports)
Confidence-critical work (source hierarchy decisions, claim verification)
Cross-referencing multiple sources
Audience-firewall-sensitive deliverables
Methodology design or methodology changes
Hook behavior changes (affects every consumer research repo)
Template changes (propagate downstream)
Installer behavior (cli/, init, setup-wizard)
Security-sensitive behavior
Tagged release prep
Anything where source accuracy or downstream impact matters

When to Use Setup B

Setup B is sufficient for routine work where a Sonnet driver can ship with a strong reviewer:

Routine drafting of pre-researched content
Formatting and structuring existing research
Documentation
Examples
Tests
Normal CLI changes (non-installer)
Low-risk methodology edits
Mechanical refactors

Setup C — Research Lite

Role	Model	Notes
Planner	You (the user)	Task is pre-planned, no model reasoning needed
Driver	Sonnet 4.6	Same model as Setup B driver — Max-bundled, no extra model to manage
Reviewer	None	Blast radius too low for cross-model overhead

The "just do the thing" lane. No RDLC discipline needed — just fast, cheap execution. You already know what to do — you just need a fast pair of hands.

When to Use Setup C

Setup C is for work where RDLC discipline overhead exceeds the value:

Data formatting scripts
Bibliography management
File moves, renames, bulk operations
Template generation (non-structural)
Config updates, env var changes
Repo maintenance (dependency bumps, lockfile refreshes)
Anything where blast radius is low and you need speed, not depth

Not Lite — escalate to A or B: changes to hooks (enforce research quality on consumers), changes to confidence vocabulary or source hierarchy, template structural changes, anything security-sensitive. If you're unsure, it's not Lite.

What Setup C explicitly skips

No RDLC gates (no confidence calibration for running a formatting script)
No cross-model review (not worth the cost or time for grunt work)
No planning phase (you are the planner)
No effort escalation (Sonnet standard is plenty)

The discipline of knowing when NOT to use discipline. Documenting this lane tells users "here's when to switch off the heavy methodology" rather than silently tempting them to skip it. If the task turns out to be harder than expected, escalate to Setup B or A.

Final Review Policy

Setups A and B end at GPT-5.5 xhigh as the cross-model reviewer. Claude can't grade its own homework — the reviewer always belongs to a different lab with different blind spots.

Setup C has no reviewer — the blast radius doesn't justify it. If you're unsure whether a task is truly Lite, it probably isn't. Escalate.

If GPT-5.5 isn't available on your OpenAI account, Codex auto-falls back to GPT-5.4 — still keep model_reasoning_effort="xhigh". Lower reasoning misses subtle bugs that the reviewer is the last gate to catch.

Version Requirement

advisorModel in settings.json requires Claude Code v2.1.170+. Check your version with claude --version. If below v2.1.170, update from inside a CC session:

! claude update

The ! prefix runs shell commands inside your CC session — no need to exit and re-enter. After updating, restart the session for the advisor to activate.

Fable 5 as advisor also requires Fable 5 access for your organization/plan (free on Max through June 22, 2026).

When the Advisor Is Unavailable

If the advisor returns "Advisor unavailable," the server-side harness failed to initialize. No in-session action (/model, /clear) can recover it.

Step 1 — restart the session. Exit and run claude (not --resume). A fresh process re-initializes the server handshake. This resolves most advisor failures.

Step 2 — if the API incident persists:

/model fable + /effort high for the planning phase, then /model claude-opus-4-6 for implementation. Interactive — stays on your Max subscription.
Or proceed with Opus only and let the Codex xhigh PR gate catch issues.

Last resort (scripted/CI only):

claude --model fable --effort high -p "$(cat <file>)" — headless mode bills API credits, not your Max subscription.

Check status.claude.com if the advisor fails across multiple fresh sessions.

Whichever path you use, the cross-model PR review gate still applies.

Credit-Spend Warning

Setups A and B use Opus 4.6 max for at least the planner — that's the expensive half. On Max-plan subscriptions, Premium can burn the 5-hour cap faster than Saver because Opus 4.6 max drives implementation too. If you're hitting the cap mid-session:

Drop to Setup B for the remainder of the day
Or drop to Setup C for grunt work that doesn't need Opus reasoning
Or use Sonnet directly for the final mechanical edits, then run the GPT-5.5 reviewer over the whole diff at the end

Setup C uses Sonnet — same model as Setup B's driver, Max-bundled. One less model to manage.

The reviewer (GPT-5.5 xhigh) is billed against your OpenAI account, separately. Watch both bills.

Autocompact Thresholds

For recommended CLAUDE_AUTOCOMPACT_PCT_OVERRIDE values per context window and task shape, see RDLC.md → Autocompact Tuning.

How Billing Works — 1M Context, Max Plan, and the June 15 Split

A common question: "does the [1m] model alias get billed differently? Does it pull from my Max plan or from API credits?"

The short answer: Setup A uses Fable advisor + Opus driver (both Max-bundled in interactive sessions). Setup B is fully Max-bundled via opusplan + Opus advisor. Here's the detail.

1M context is free on Max — no API premium

Anthropic 2026-03-13: 1M context is GA at standard $5/$25 per million tokens for Opus 4.6 (also 4.7, 4.8). No long-context multiplier. No beta header required, requests over 200K tokens just work.

For Claude Code on Max / Team / Enterprise plans, 1M context is included automatically with no extra usage allocation. Whether you set claude-opus-4-6 or claude-opus-4-6[1m] doesn't change what you're billed — both pull from the same per-token budget on your subscription. The [1m] suffix just makes the alias explicit so it sticks across alias-resolution changes; functionally Opus 4.6 in Claude Code today is the 1M-context model on a Max plan.

(Pro plan is the exception: Pro users need "Enable usage credits" turned on in their Claude account settings to use 1M context. Max / Team / Enterprise have it on by default.)

The June 15, 2026 billing split

Anthropic moved a slice of usage off the subscription onto a separate metered credit pool that bills at full API rates:

Surface	Billing as of June 15, 2026
Interactive Claude Code in terminal (you typing into Claude Code right now)	Stays on Max subscription — unchanged
Claude.ai web / desktop / mobile chat	Stays on Max subscription
Claude Cowork	Stays on Max subscription
`claude -p` (headless / `--print`)	Moves to separate credit pool, billed at API rates
Claude Agent SDK	Moves to separate credit pool
Claude Code GitHub Actions	Moves to separate credit pool
Third-party apps via Agent SDK	Moves to separate credit pool

Credit allocations: Pro $20/mo, Max 5x $100/mo, Max 20x $200/mo. No rollover.

What this means for the lanes

Setup A — Premium (Fable advisor + Opus driver): Fable 5 advisor via advisorModel: "fable" in project settings — interactive session, Max-bundled. Opus 4.6 max driver on Max. GPT-5.5 xhigh reviewer on ChatGPT subscription.
Setup B — Saver (OpusPlan): fully Max-bundled. opusplan uses Opus (plan mode) + Sonnet (execute mode), both at 200K context — no [1m] variants, no credit drain. This is why Setup B now recommends opusplan instead of the old sonnet[1m] pin.
- Avoid sonnet[1m]: Sonnet with 1M context draws from your usage credits pool ($3/$15 per Mtok), NOT your Max subscription. The /model picker shows this explicitly. Plain sonnet (200K) or opusplan stays on Max.
Reviewer (GPT-5.5 xhigh) in both lanes: billed against your OpenAI account, completely separate from Anthropic.
CI loops that use claude -p post-June-15: these now bill against the separate Anthropic credit pool, not your Max subscription. Consumer-repo CI integrations may need to budget the new credit pool.

Caveat: Setup B's cost-saving has conditions

The savings argument for Setup B is "Sonnet driver is cheaper than Opus driver per turn." That's true at standard 200K context. If your Sonnet driver needs to load >200K tokens (large diff, multi-file refactor, monorepo audit), the bill quietly flips:

Mode	Per-token rate	Pool
Setup A — Opus 4.6 1M	$5/$25 per Mtok	Max subscription
Setup B — Sonnet 4.6 standard	$3/$15 per Mtok	Max subscription
Setup B — Sonnet 4.6 1M	$3/$15 per Mtok	Credits pool

So for context-heavy work that crosses 200K, Setup A on Max is actually cheaper than Setup B on credits — because Setup A's Opus stays on the Max pool while Setup B's Sonnet 1M draws down a separately-metered $100/mo (Max 5x) or $200/mo (Max 20x) credit budget.

Practical guidance:

Subscription-first mindset (recommended): Use Setup A unless you're confident the Sonnet driver in Setup B stays under 200K context. The Opus 4.6 max planner+driver combo lives entirely on Max — no credit-pool drawdown.
Setup B is a real cost win only when: the driver task fits in ≤ 200K (routine implementation, single-file work, small refactors, docs). Use B specifically for those, not as a blanket choice.
Watch the picker. When you swap to Sonnet 1M, Claude Code shows "Draws from usage credits" explicitly. That's your billing flip signal — choose Opus 1M instead if you want to stay on Max.

Bottom line

If you're using Claude Code interactively (you, in your terminal, doing research work), both lanes ride your existing Max subscription, and the [1m] alias is the same billable budget as plain claude-opus-4-6. No extra charges for the 1M context. The June 15 split only affects programmatic / headless / CI use of Claude Code.

Watch the headless surface if you've automated claude -p calls in your project — those now bill differently as of June 15, 2026.

Maintainer Override

Override at any time. A blanket setup choice doesn't replace judgment per change. If you're touching a hook but the change is a one-line typo, Setup B is fine. If you're touching templates but the section is the wizard's safety-critical confidence vocabulary, Setup A is the call.

The wizard does not enforce setup lane selection — it documents the recommended default per change shape. Whatever ships is your call.