fix(ai): remove <thinking> tag wrapping, convert to plain text on cross-model handoff

- Remove <thinking> tag generation from google-shared.ts, transorm-messages.ts, openai-completions.ts
- Thinking blocks now convert to plain text when switching models (prevents models mimicking tags)
- Skip empty thinking blocks to avoid API errors
- Keep thinking blocks only when same provider AND same model

fixes #561

packages/ai/CHANGELOG.md

@@ -6,6 +6,7 @@
 
 - Fixed Gemini CLI abort handling: detect native `AbortError` in retry catch block, cancel SSE reader when abort signal fires ([#568](https://github.com/badlogic/pi-mono/pull/568) by [@tmustier](https://github.com/tmustier))
 - Fixed Antigravity provider 429 errors by aligning request payload with CLIProxyAPI v6.6.89: inject Antigravity system instruction with `role: "user"`, set `requestType: "agent"`, and use `antigravity` userAgent. Added bridge prompt to override Antigravity behavior (identity, paths, web dev guidelines) with Pi defaults. ([#571](https://github.com/badlogic/pi-mono/pull/571) by [@ben-vargas](https://github.com/ben-vargas))
+- Fixed thinking block handling for cross-model conversations: thinking blocks are now converted to plain text (no `<thinking>` tags) when switching models. Previously, `<thinking>` tags caused models to mimic the pattern and output literal tags. Also fixed empty thinking blocks causing API errors. ([#561](https://github.com/badlogic/pi-mono/issues/561))
 
 ## [0.38.0] - 2026-01-08
 

packages/ai/src/providers/google-shared.ts

@@ -77,6 +77,8 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 			}
 		} else if (msg.role === "assistant") {
 			const parts: Part[] = [];
+			// Check if message is from same provider and model - only then keep thinking blocks
+			const isSameProviderAndModel = msg.provider === model.provider && msg.model === model.id;
 
 			for (const block of msg.content) {
 				if (block.type === "text") {
@@ -84,17 +86,19 @@ export function convertMessages<T extends GoogleApiType>(model: Model<T>, contex
 					if (!block.text || block.text.trim() === "") continue;
 					parts.push({ text: sanitizeSurrogates(block.text) });
 				} else if (block.type === "thinking") {
-					// Thinking blocks require signatures for Claude via Antigravity.
-					// If signature is missing (e.g. from GPT-OSS), convert to regular text with delimiters.
-					if (block.thinkingSignature) {
+					// Skip empty thinking blocks
+					if (!block.thinking || block.thinking.trim() === "") continue;
+					// Only keep as thinking block if same provider AND same model
+					// Otherwise convert to plain text (no tags to avoid model mimicking them)
+					if (isSameProviderAndModel) {
 						parts.push({
 							thought: true,
 							text: sanitizeSurrogates(block.thinking),
-							thoughtSignature: block.thinkingSignature,
+							...(block.thinkingSignature && { thoughtSignature: block.thinkingSignature }),
 						});
 					} else {
 						parts.push({
-							text: `<thinking>\n${sanitizeSurrogates(block.thinking)}\n</thinking>`,
+							text: sanitizeSurrogates(block.thinking),
 						});
 					}
 				} else if (block.type === "toolCall") {

packages/ai/src/providers/openai-completions.ts

@@ -490,10 +490,8 @@ function convertMessages(
 			const nonEmptyThinkingBlocks = thinkingBlocks.filter((b) => b.thinking && b.thinking.trim().length > 0);
 			if (nonEmptyThinkingBlocks.length > 0) {
 				if (compat.requiresThinkingAsText) {
-					// Convert thinking blocks to text with <thinking> delimiters
-					const thinkingText = nonEmptyThinkingBlocks
-						.map((b) => `<thinking>\n${b.thinking}\n</thinking>`)
-						.join("\n");
+					// Convert thinking blocks to plain text (no tags to avoid model mimicking them)
+					const thinkingText = nonEmptyThinkingBlocks.map((b) => b.thinking).join("\n\n");
 					const textContent = assistantMsg.content as Array<{ type: "text"; text: string }> | null;
 					if (textContent) {
 						textContent.unshift({ type: "text", text: thinkingText });

packages/ai/src/providers/transorm-messages.ts

@@ -45,12 +45,13 @@ export function transformMessages<TApi extends Api>(messages: Message[], model:
 				assistantMsg.api !== model.api;
 
 			// Transform message from different provider/model
-			const transformedContent = assistantMsg.content.map((block) => {
+			const transformedContent = assistantMsg.content.flatMap((block) => {
 				if (block.type === "thinking") {
-					// Convert thinking block to text block with <thinking> tags
+					// Skip empty thinking blocks, convert others to plain text
+					if (!block.thinking || block.thinking.trim() === "") return [];
 					return {
 						type: "text" as const,
-						text: `<thinking>\n${block.thinking}\n</thinking>`,
+						text: block.thinking,
 					};
 				}
 				// Normalize tool call IDs for github-copilot cross-API switches

packages/coding-agent/CHANGELOG.md

@@ -22,6 +22,7 @@
 - ESC key now works during "Working..." state after auto-retry ([#568](https://github.com/badlogic/pi-mono/pull/568) by [@tmustier](https://github.com/tmustier))
 - Abort messages now show correct retry attempt count (e.g., "Aborted after 2 retry attempts") ([#568](https://github.com/badlogic/pi-mono/pull/568) by [@tmustier](https://github.com/tmustier))
 - Fixed Antigravity provider returning 429 errors despite available quota ([#571](https://github.com/badlogic/pi-mono/pull/571) by [@ben-vargas](https://github.com/ben-vargas))
+- Fixed malformed thinking text in Gemini/Antigravity responses where thinking content appeared as regular text or vice versa. Cross-model conversations now properly convert thinking blocks to plain text. ([#561](https://github.com/badlogic/pi-mono/issues/561))
 
 ## [0.38.0] - 2026-01-08