packages/ai: adapt Anthropic thinking payloads

[masnwilliams]

Summary

• Convert legacy Anthropic thinking.type=enabled payloads to adaptive thinking for models that require or recommend the new format.
• Cover newer/adaptive Anthropic CUA families: Sonnet 5, Sonnet 4.6, Opus 4.8, Opus 4.7, Opus 4.6, Fable 5, and Mythos-style models.
• Preserve caller output_config while setting output_config.effort from the existing budget-token levels.
• Add unit coverage for adaptive conversion and older manual-thinking models that should remain unchanged.

Test plan

• npm run build --workspace @onkernel/cua-ai
• npm test --workspace @onkernel/cua-ai
• npm test --workspace @onkernel/cua-agent -- --exclude "**/*.live.test.ts"
• npm pack --workspace @onkernel/cua-ai --dry-run

---

Note

Medium Risk
Changes outbound Anthropic API payload shape for a defined set of newer models; behavior is gated by model ID and only applies when legacy thinking is enabled, but incorrect mapping could affect model reasoning/cost.

Overview
@onkernel/cua-ai 0.3.4 adds Anthropic request payload middleware so callers can keep sending legacy thinking: { type: "enabled", budget_tokens } while newer CUA models get the adaptive format.

For matching models (Sonnet 5/4.6, Opus 4.8/4.7/4.6, Fable 5, Mythos families), anthropicAdaptiveThinkingOnPayload rewrites thinking to { type: "adaptive" }, merges existing output_config, and sets output_config.effort from the old budget tiers (low through xhigh). Older models like Sonnet 4.5 are left untouched when the hook returns undefined. The hook is registered on the Anthropic providerModule, so resolveCuaRuntimeSpec now exposes onPayload for Anthropic like Yutori/Tzafon.

Unit tests cover conversion, model coverage, effort mapping, and runtime-spec expectations. The agent payload-hook test is relaxed to toMatchObject so composed payloads with extra fields still pass.

^{Reviewed by Cursor Bugbot for commit dac33b8. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes using high effort and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 5b7ea25. Configure here.}

[cursor]

Missing TSDoc on payload hook

Low Severity

The new exported anthropicAdaptiveThinkingOnPayload hook has no TSDoc describing when it rewrites thinking for Sonnet 5, how output_config.effort is derived, or that it returns undefined to leave other models and payloads unchanged.

 

^{Triggered by learned rule: Exported types, interfaces, and classes require TSDoc}

^{Reviewed by Cursor Bugbot for commit 5b7ea25. Configure here.}