fix: restore Claude adaptive effort and OAuth refresh handling

[i1hwan]

Summary

• promote Claude effort-only or thinkingLevel-only requests to adaptive thinking with provider effort passthrough instead of converting them into fixed budget tokens
• fix Claude OAuth health checks to pass the resolved proxy payload correctly
• align Claude OAuth refresh requests with the repository's working auth-code token exchange contract before trying compatibility fallbacks, removing Claude API-specific headers from the token endpoint path
• add focused coverage for the Claude thinking-budget, translator, token refresh, and health-check paths

Why this surfaced now

• the refresh breakage is on the proactive Claude OAuth health-check path, which only runs when a stored Claude token is close to expiry and the sweep attempts a refresh
• PR7 changed Claude request handling, but the failing refresh path is src/lib/tokenHealthCheck.ts -> getAccessToken() -> open-sse/services/tokenRefresh.ts
• the remaining 400s pointed to the Claude refresh request itself still not matching the repo's known-good Claude auth exchange contract, so this PR now normalizes that contract too

Verification

• node --import tsx/esm --test tests/unit/thinking-budget.test.mjs
• node --import tsx/esm --test tests/unit/translator-openai-to-claude.test.mjs
• node --import tsx/esm --test tests/unit/token-health-check.test.mjs tests/unit/token-refresh-service.test.mjs
• node --import tsx/esm --test tests/unit/token-refresh-route-service.test.mjs
• npm run typecheck:core

[Copilot]

Pull request overview

This PR restores Claude “adaptive thinking” behavior when requests only provide an effort signal (e.g., reasoning_effort) or a thinkingLevel signal, and hardens Claude OAuth refresh + health-check proxy handling, with added unit coverage for the affected paths.

Changes:

• Promote Claude effort-only / thinkingLevel-only requests to thinking: { type: "adaptive" } with upstream output_config.effort passthrough.
• Fix Claude OAuth health checks to pass the resolved proxy payload correctly and retry token refresh using form-encoding when JSON payloads are rejected.
• Add unit tests covering thinking-budget promotion, translation, proxy unwrap behavior, and Claude OAuth refresh fallback.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
open-sse/services/thinkingBudget.ts	Promotes Claude effort/thinkingLevel signals into adaptive thinking and resolves effort from multiple sources.
open-sse/services/tokenRefresh.ts	Sends Claude OAuth refresh as JSON first with a form-encoded fallback on specific 400 errors.
src/lib/tokenHealthCheck.ts	Unwraps the resolved proxy config before calling token refresh, matching expected proxy payload shape.
tests/unit/thinking-budget.test.mjs	Adds coverage for Claude promotion behavior in PASSTHROUGH mode and updates existing expectations.
tests/unit/translator-openai-to-claude.test.mjs	Adds coverage ensuring translateRequest promotes Claude thinkingLevel-only requests to adaptive effort.
tests/unit/token-refresh-service.test.mjs	Adds coverage for Claude OAuth refresh JSON-first behavior and form-encoded fallback.
tests/unit/token-health-check.test.mjs	Adds coverage for unwrapping resolved proxy configs for refresh calls.

[Copilot]

hasThinkingLevelSignal() treats any non-empty thinkingLevel string as a promotion signal. Because applyThinkingBudget() uses originalHasThinkingLevelSignal to force Claude requests into adaptive mode, a client value like thinkingLevel: "none" / "disabled" will now be promoted to thinking: {type:"adaptive"} instead of preserving the existing behavior where normalizeThinkingLevel() converts it to disabled thinking. Consider excluding disable values (and possibly unknown values) from this signal, or explicitly short-circuiting to a disabled thinking config when the level maps to none/disabled.

Suggested change

	return typeof level === "string" && level.trim().length > 0;
	if (typeof level !== "string") return false;
	const normalized = level.trim().toLowerCase();
	return (
	normalized === "low" ||
	normalized === "medium" ||
	normalized === "high" ||
	normalized === "max" ||
	normalized === "xhigh"
	);