feat(core): enable low/med/high variants for Claude Sonnet/Opus 4.6 thinking

[sebastianlungu]

Summary

Enables low/medium/high variants for both antigravity-claude-sonnet-4-6-thinking and antigravity-claude-opus-4-6-thinking with thinkingConfig.thinkingBudget 8192/16384/32768 matching THINKING_TIER_BUDGETS.claude.

Previously these models advertised {disabled:true} variants, so OpenCode's variant picker showed no options and --variant was ignored. Gemini flash/pro already had working thinkingLevel variants; Claude now matches.

Changes

• packages/core/src/model-registry.ts: replace disabled variants with budgets (NET +4 lines, functionally NET 0 replacement)
• packages/opencode/src/plugin/config/models.test.ts: update expectation

Verification

• ./node_modules/.bin/tsc -p packages/core/tsconfig.build.json passes
• bun test --isolate packages/core/src/model-registry.test.ts 7 pass
• bun test --isolate packages/opencode/src/plugin/config/models.test.ts 4 pass
• Variant extraction already supports providerOptions.google.thinkingConfig.thinkingBudget + generationConfig fallback (packages/core/src/transform/model-resolver.ts:resolveModelWithVariant)

Tested via

opencode debug config | jq '.provider.google.models["antigravity-claude-sonnet-4-6-thinking"].variants'
# -> {low:{thinkingBudget:8192}, medium:{thinkingBudget:16384}, high:{thinkingBudget:32768}}
OPENCODE_ANTIGRAVITY_DEBUG=1 opencode run "test" --model google/antigravity-claude-sonnet-4-6-thinking --variant medium
# -> generationConfig.thinkingConfig.thinking_budget:16384

Closes configurability request for Claude Sonnet effort. Net LOC: +4/-4 effectively 0 (swap disabled for budgets).

---

Summary by cubic

Enables low/medium/high thinking variants for Claude Sonnet/Opus 4.6 so the variant picker and --variant flag now control thinking effort. Previously these models advertised disabled variants, so no options showed and the flag was ignored.

Changes

• Replaces disabled variants with thinking budgets of 8192/16384/32768 in packages/core/src/model-registry.ts.
• Orders the new variants low→medium→high so the TUI picker lists them consistently.
• Updates test expectations in packages/opencode/src/plugin/config/models.test.ts.
• Adds wire-level tests proving each budget maps to thinkingConfig.thinkingBudget on Antigravity requests.

^{Written for commit 8319a59. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 2 files

_{Re-trigger cubic}

[ualtinok]

Changes requested

This should not enable low / medium / high as supported Claude effort tiers without further evidence.

I revalidated both Claude routes against native AGY CLI 1.1.24 and the live endpoint:

• agy --model claude-sonnet-4-6 --effort low|medium|high rejects every selection with --effort is not supported for model "claude-sonnet-4-6".
• Native Sonnet and Opus MITM captures both send thinkingConfig: { includeThoughts: true, thinkingBudget: 1024 } with maxOutputTokens: 64000.
• Replaying the captured Sonnet request with 8192, 16384, and 32768 returns HTTP 200, but that establishes only that the backend tolerates those values. It does not establish low/medium/high effort semantics, and controlled responses did not show a monotonic budget-to-effort relationship.

The current disabled variants are therefore deliberate native-parity behavior. This patch would advertise unsupported effort controls and send up to 32x the native budget.

The verification also does not cover the production seam. request.ts independently uses extractVariantThinkingConfig; it does not call the cited resolveModelWithVariant. Please add request-level tests proving the final Antigravity payload for every proposed preset. The Antigravity wire field is camel-case thinkingBudget, not the snake-case thinking_budget shown in the PR description.

Required direction: either keep these variants disabled, or redesign this explicitly as a non-native/experimental custom-budget feature with validated semantics, accurate UI wording, final-wire tests, and user-facing documentation.