diff --git a/packages/opencode/src/session/llm.ts b/packages/opencode/src/session/llm.ts
index 22fc9b1d5588..24af2e8ec1ad 100644
--- a/packages/opencode/src/session/llm.ts
+++ b/packages/opencode/src/session/llm.ts
@@ -333,6 +333,28 @@ export namespace LLM {
             toolName: lower,
           }
         }
+        // When the tool name is valid but args failed to parse, the output
+        // was likely truncated by the token limit.  Give the model a clear
+        // signal so it can retry with smaller input instead of looping.
+        if (tools[failed.toolCall.toolName] || tools[lower]) {
+          l.warn("truncated tool call detected", {
+            tool: failed.toolCall.toolName,
+            error: failed.error.message,
+          })
+          return {
+            ...failed.toolCall,
+            input: JSON.stringify({
+              tool: failed.toolCall.toolName,
+              error:
+                "Your output was truncated because it exceeded the token limit. " +
+                "The tool arguments were cut off and could not be parsed. " +
+                "Split your operation into smaller pieces and try again. " +
+                "For file writes, use the Edit tool for targeted changes or write smaller sections. " +
+                "For bash commands, break long heredocs into multiple shorter commands.",
+            }),
+            toolName: "invalid",
+          }
+        }
         return {
           ...failed.toolCall,
           input: JSON.stringify({
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
index 19f0850ff4c2..77de6895aa3b 100644
--- a/packages/opencode/src/session/prompt.ts
+++ b/packages/opencode/src/session/prompt.ts
@@ -1376,6 +1376,35 @@ NOTE: At any point in time through this workflow you should feel free to ask the
             const hasToolCalls =
               lastAssistantMsg?.parts.some((part) => part.type === "tool" && !part.metadata?.providerExecuted) ?? false
 
+            // When the model hit the output token limit (finishReason: "length"),
+            // its response was truncated.  Instead of silently exiting, inject a
+            // continuation message so the model can finish its work.
+            if (lastAssistant?.finish === "length" && lastUser.id < lastAssistant.id) {
+              log.warn("output truncated by token limit, auto-continuing", { sessionID })
+              const continuationMsg: MessageV2.User = {
+                id: MessageID.ascending(),
+                sessionID,
+                role: "user",
+                time: { created: Date.now() },
+                agent: lastUser.agent,
+                model: lastUser.model,
+              }
+              yield* sessions.updateMessage(continuationMsg)
+              yield* sessions.updatePart({
+                id: PartID.ascending(),
+                messageID: continuationMsg.id,
+                sessionID,
+                type: "text",
+                text:
+                  "Your previous response was truncated because it exceeded the output token limit. " +
+                  "If you were writing a file, split it into smaller pieces. " +
+                  "If you were using a tool, use smaller arguments. " +
+                  "Continue where you left off.",
+                synthetic: true,
+              })
+              continue
+            }
+
             if (
               lastAssistant?.finish &&
               !["tool-calls"].includes(lastAssistant.finish) &&