copilot-api-plus 1.4.0 → 1.4.2

package/dist/main.js

@@ -3206,6 +3206,10 @@ function sanitizeForCopilotBackend(payload) {
 		delete extended.context_management;
 	}
 	sanitizeOutputConfigFormat(extended.output_config?.format);
+	if (payload.effort !== void 0) {
+		consola.debug("Stripping effort field (unsupported by Copilot backend)");
+		delete payload.effort;
+	}
 }
 function sanitizeOutputConfigFormat(format) {
 	if (!isRecord(format) || format.type !== "json_schema") return;
@@ -3240,19 +3244,24 @@ function normalizeAdaptiveThinkingForCopilot(payload) {
 }
 /**
 * If the client did not specify a `thinking` field, inject the maximum
-* thinking budget the model supports — pulled from Copilot's `/models`
+* thinking depth the model supports — pulled from Copilot's `/models`
 * capabilities. Mutates in place.
 *
-*   - Models with `adaptive_thinking: true` (claude-opus-4.7,
-*     claude-sonnet-4.6) get `{ type: "adaptive" }` so the model
-*     decides depth dynamically — recommended by Anthropic for
-*     these models.
+*   - Models with `adaptive_thinking: true` (Claude Opus 4.7,
+*     Sonnet 4.6, etc.) get `{ type: "adaptive", budget_tokens: max }`.
+*     Per Anthropic 2026 docs, `effort` is the recommended control
+*     for thinking depth on adaptive models — but Copilot's `/v1/messages`
+*     mirror rejects the top-level `effort` field with a 400. The
+*     `budget_tokens` hint on adaptive thinking is documented as still
+*     accepted, and serves as a softer nudge that keeps the model from
+*     skipping thinking entirely on simple prompts.
 *   - Other thinking-capable models get
 *     `{ type: "enabled", budget_tokens: max_thinking_budget }`.
 *   - Models without thinking capability are left untouched.
 *
 * Skipped if the client already specified `thinking` (any value) — we
-* always defer to explicit client intent.
+* always defer to explicit client intent. Also skipped when the runtime
+* `state.maxThinking` kill switch is off.
 */
 function injectMaxThinkingBudget(payload) {
 	if (!state.maxThinking) return;
@@ -3262,8 +3271,11 @@ function injectMaxThinkingBudget(payload) {
 	const maxBudget = supports.max_thinking_budget;
 	if (!maxBudget || maxBudget <= 0) return;
 	if (supports.adaptive_thinking === true) {
-		payload.thinking = { type: "adaptive" };
-		consola.debug(`Injected adaptive thinking for ${payload.model} (no client preference)`);
+		payload.thinking = {
+			type: "adaptive",
+			budget_tokens: maxBudget
+		};
+		consola.debug(`Injected adaptive thinking budget=${maxBudget} for ${payload.model} (no client preference)`);
 		return;
 	}
 	payload.thinking = {