GAP-L1-P1-adaptive-thinking-vs-extended-thinking

Current state

ModelVersion.supportsThinking: bool + thinkingEffortLevels + supportsThinkingBudgetTokens. Anthropic now ships TWO distinct thinking modes:

• **Extended thinking** — explicit budget/effort, visible thinking blocks, available on Sonnet 4.6 / Haiku 4.5 / older Opus.
• **Adaptive thinking** — model auto-decides depth, available on Opus 4.7 + Sonnet 4.6, NOT on Haiku 4.5. Opus 4.7 has adaptive but **not** extended thinking.

The current single boolean cannot represent "Opus 4.7 has adaptive but no extended thinking".

Desired state

Replace supportsThinking: bool with two booleans (or an enum-set):

• supportsExtendedThinking: bool
• supportsAdaptiveThinking: bool
• Update invariant: thinkingEffortLevels requires supportsExtendedThinking; supportsThinkingBudgetTokens requires supportsExtendedThinking.

Evidence

• https://platform.claude.com/docs/en/docs/about-claude/models/overview (Opus 4.7 row: Extended=No, Adaptive=Yes)
• https://platform.claude.com/docs/en/build-with-claude/adaptive-thinking
• https://platform.claude.com/docs/en/build-with-claude/extended-thinking

Propagation status

• Level 1: open
• Level 2: not-started — invariants.yaml + 02-node-kinds/compute-path.md need updates

Propagation chain

• Level 1: split attribute, update 3 example model files.
• Level 2: update validation rules referencing thinking-related invariants.

Notes

Real-world divergence the schema cannot currently encode.