diff --git a/.changeset/sturdy-finish-guard.md b/.changeset/sturdy-finish-guard.md
new file mode 100644
index 00000000000..c308018de97
--- /dev/null
+++ b/.changeset/sturdy-finish-guard.md
@@ -0,0 +1,5 @@
+---
+"@kilocode/cli": patch
+---
+
+Prevent an infinite agent loop when a provider ends the response stream without a terminal stop reason.
diff --git a/packages/opencode/src/session/prompt.ts b/packages/opencode/src/session/prompt.ts
index 760d9e50bec..69fa84bcac7 100644
--- a/packages/opencode/src/session/prompt.ts
+++ b/packages/opencode/src/session/prompt.ts
@@ -1677,6 +1677,21 @@ NOTE: At any point in time through this workflow you should feel free to ask the
               return "break" as const
             }
             // kilocode_change end
+            // kilocode_change start - guard against providers that end the stream
+            // without a terminal stop_reason (e.g. an Anthropic-style message_delta
+            // with stop_reason: null followed immediately by message_stop). Without
+            // a finishReason, the loop-exit check at the top of the next iteration
+            // sees a falsy `finish` (loaded from storage via filterCompactedEffect)
+            // and keeps stepping forever. Default to "unknown" and persist so the
+            // regular break condition fires when there are no tool calls. Skipped
+            // for the compact path so guardCompactionAttempt can still fill in
+            // "error" on exhaustion. Tool-call turns already get "tool-calls" from
+            // the AI SDK; even without it, !hasToolCalls keeps the break gated.
+            if (result !== "compact" && !handle.message.finish) {
+              handle.message.finish = "unknown"
+              yield* sessions.updateMessage(handle.message)
+            }
+            // kilocode_change end
             return "continue" as const
           }).pipe(Effect.ensuring(instruction.clear(handle.message.id)))
           if (outcome === "break") break