fix(anthropic): send adaptive thinking only to models that support it

[ericleepi314]

Summary

Testing the Claude subscription (OAuth) login flow end-to-end surfaced a pre-existing, general bug: every Anthropic request 400s with adaptive thinking is not supported on this model for any model except sonnet-4-6 / opus-4-6 / opus-4-7.

_call_model_sync (src/query/query.py) sent thinking={"type":"adaptive"} + output_config={"effort":...} to all Claude 4.x models, gated only by the broad _model_supports_extended_thinking regex. But adaptive thinking and effort are supported by only a narrow subset. This is not subscription-specific — the API-key path 400s identically — but the default model sonnet-4-6 supports adaptive, which masked it. For the subscription flow it was fatal: a Max user selecting haiku (very plausible, to conserve rate limit) or sonnet-4-5 got a hard 400 on every turn.

Fix

Ports the authoritative adaptive-vs-budget selection from the in-repo upstream Claude Code source (typescript/):

• claude.ts:1612-1640 — the selection logic
• utils/thinking.ts:117-169 — modelSupportsThinking (all Claude 4+ incl. Haiku on first-party) vs modelSupportsAdaptiveThinking (opus-4-6/4-7, sonnet-4-6 only)
• utils/effort.ts:32-51 — modelSupportsEffort (opus-4-6, sonnet-4-6 only)

Two new helpers (_model_supports_adaptive_thinking, _model_supports_effort) split the single injection into three independently-gated decisions:

model	thinking	adaptive	effort	sent
sonnet-4-6, opus-4-6	✓	✓	✓	{type:adaptive} + effort
opus-4-7	✓	✓	✗	{type:adaptive}
sonnet-4-5, haiku-4-5, opus-4-1/4-5	✓	✗	✗	{type:enabled, budget_tokens}
claude-3.x	✗	✗	✗	(none)

Budget models get budget_tokens = max_tokens-1 (clamped ≥1024, <max_tokens, matching claude.ts:1628-1635). Adaptive-capable models — including the default sonnet-4-6 — are unchanged, so there is no regression for the common path.

Validation

Live against api.anthropic.com with a Claude Max OAuth token (the subscription path #697 added):

• haiku headless — previously a hard 400, now passes ✓
• full Bash tool loop through the agent loop under subscription auth ✓
• sonnet-4-6 default — adaptive+effort unchanged ✓

Also independently confirms the #697 subscription request path works: OAuth bearer + beta headers accepted, mcp_ tool prefixing/stripping correct, billing_mode: subscription ($0), 2-turn tool exchange.

Unit tests (tests/test_query_extended_thinking.py): updated 2 that pinned the old buggy behavior (asserted adaptive for sonnet-4-5), added the opus-4-7 adaptive-without-effort regression pin (the fix's most fragile case) and a direct allowlist matrix. Full suite: 8482 passed, 3 skipped. Critic-reviewed: APPROVE.

🤖 Generated with Claude Code