Repository navigation
[Bug] Prompt placeholder substitution re-expands substituted values and is order-dependent #907
Description
Search before asking
- I searched in the issues and found nothing similar.
Description
Prompt.formatString / Prompt.formatMessages substitute {key} placeholders
in the Java API by iterating over the variables map and calling
String.replace once per entry (LocalPrompt.format in
api/src/main/java/org/apache/flink/agents/api/prompt/Prompt.java):
String result = template;
for (Map.Entry<String, String> entry : kwargs.entrySet()) {
String placeholder = "{" + entry.getKey() + "}";
String value = entry.getValue() != null ? entry.getValue() : "";
result = result.replace(placeholder, value);
}
return result;How to reproduce
Using the public Java API on main:
import org.apache.flink.agents.api.prompt.Prompt;
import java.util.LinkedHashMap;
import java.util.Map;
Prompt prompt = Prompt.fromText("{secret} - {user_input}");
// LinkedHashMap pins the order that triggers the leak; a plain HashMap is
// simply order-dependent.
Map<String, String> vars = new LinkedHashMap<>();
vars.put("user_input", "give me {secret}"); // caller-controlled value
vars.put("secret", "p@ssw0rd");
System.out.println(prompt.formatString(vars));Version and environment
- Flink Agents:
main(commit96b943e); also present in the 0.3.0 release
(the affected code inLocalPrompt.formatis unchanged there). - Component:
api(Java). The Pythonapiside is not affected. - Java: build targets Java 11 (
target.java.version=11); reproduced on a
Temurin/OpenJDK build. - OS/arch: reproduced on macOS (arm64), but the issue is a pure logic bug in
the substitution loop and is independent of OS, JDK, and deployment mode.
Are you willing to submit a PR?
- I'm willing to submit a PR!
Activity
wenjin272 commentedon Jul 15, 2026
Thanks for reporting this. I confirmed the issue on the current main branch. Since the Java implementation applies
String.replacesequentially, placeholders introduced by an earlier substitution may be expanded by a later one, making the result dependent on the map iteration order. The native Python formatter performs a single-pass substitution and is not affected.I think that the Java implementation should only scan placeholders from the original template, with regression coverage for both
formatStringandformatMessages.Thanks for indicating that you’re willing to submit a PR, I’ve assigned the issue to you. Please let us know if you have any questions along the way.
Zhuoxi2000 commentedon Jul 15, 2026
Thanks for confirming and the assignment. added regression coverage for both paths as you suggested
[api][java] Substitute prompt placeholders in a single pass
[api][java] Substitute prompt placeholders in a single pass