Shift+Tab thinking cycle hides the max effort tier on Opus 4.7/4.8 (and Sonnet 4.6, Fable 5) — single-top-tier gating can't express xhigh + max

[aheritier]

Summary

The interactive Shift+Tab thinking-level cycle (TUI model picker) hides the max effort tier for several Claude models that actually support it — most visibly Claude Opus 4.7 / 4.8. The cycle offers off / low / medium / high / xhigh and never reaches max.

This is a gating bug only: authored thinking_budget values (e.g. adaptive/max) are still sent correctly to the API. It affects only the user-selectable Shift+Tab cycle.

Reproduction

1. Configure an agent:

   models:
     anthropic-large:
       provider: anthropic
       model: claude-opus-4-8
       thinking_budget: adaptive/max

2. Run the TUI; the header correctly shows thinking auto adaptive.
3. Press Shift+Tab to cycle thinking levels.
4. Observed: cycle is off → low → medium → high → xhigh → (wrap). max is never offered.
5. Expected: max is offered, since Anthropic lists max as supported on Opus 4.8.

Root cause

anthropicTopEffort() in pkg/modelinfo/thinking_levels.go returns a single highest effort Level, implicitly assuming the selectable tiers form a linear ladder. For Opus 4.7+ it returns XHigh:

switch {
case minor >= 7:
    return effort.XHigh
case minor == 6:
    return effort.Max
default:
    return ""
}

thinkingLevelMap() then declares only that one tier as supported:

m := effort.LevelMap{effort.Minimal: false}
if top := anthropicTopEffort(modelID); top != "" {
    m[top] = true
}

Because Max is an explicit-only tier (explicitOnlyLevels = {XHigh, Max} in pkg/effort/effort.go) and the map only declares XHigh: true, SupportedLevels() drops Max. Net result for Opus 4.7/4.8: off / low / medium / high / xhigh.

The deeper problem: xhigh and max are independent capabilities, not a single ladder. A single returned "top tier" cannot represent a model that supports both xhigh and max (Opus 4.7/4.8, Fable 5), nor distinguish it from a model that supports max without xhigh (Opus 4.6, Sonnet 4.6).

Authoritative reference

Anthropic — Effort: https://platform.claude.com/docs/en/build-with-claude/effort

Per-level model availability, quoted from that page:

• max — Claude Fable 5, Claude Mythos 5, Claude Opus 4.8, Claude Mythos Preview, Claude Opus 4.7, Claude Opus 4.6, Claude Sonnet 4.6
• xhigh — Claude Fable 5, Claude Mythos 5, Claude Opus 4.8, Claude Opus 4.7

Resulting capability matrix:

Model	low	medium	high	xhigh	max
Opus 4.5	✓	✓	✓	—	—
Sonnet 4.6	✓	✓	✓	—	✓
Opus 4.6	✓	✓	✓	—	✓
Opus 4.7	✓	✓	✓	✓	✓
Opus 4.8	✓	✓	✓	✓	✓
Fable 5	✓	✓	✓	✓	✓

Affected models (cycle gating incorrect)

• Opus 4.7, Opus 4.8 — max missing (cycle caps at xhigh)
• Sonnet 4.6 — max missing (cycle caps at high)
• Fable 5 — max missing (cycle caps at xhigh)
• Mythos 5 / Mythos Preview — unrecognized (cycle caps at high)

Correct today: Opus 4.6 (max, no xhigh) and Opus 4.5 (high).

Not affected (scope clarification)

• The request path in pkg/model/provider/anthropic/thinking.go sends adaptive/<effort> verbatim (e.g. effort max), so authored configs already behave correctly.
• RejectsTokenThinking() correctly coerces Opus 4.6/4.7/4.8 onto adaptive thinking.
• agent-schema.json already accepts max / adaptive/max.

Suggested fix

• Replace the single-Level return of anthropicTopEffort() with independent capability flags (e.g. supportsXHigh, supportsMax), and have thinkingLevelMap() set both effort.XHigh and effort.Max in the LevelMap where applicable.
• Add table-driven tests covering each Claude tier (Opus 4.5 / 4.6 / 4.7 / 4.8, Sonnet 4.6, Fable 5, and the Mythos variants).

Relevant files:

• pkg/modelinfo/thinking_levels.go
• pkg/effort/effort.go
• pkg/runtime/model_switcher.go (consumer of SupportedThinkingLevels)

[aheritier]

FYI @rumpl