Snapshot manager output locale per run

backend/app/main.py

@@ -353,14 +353,20 @@ def workspaces(path: str | None = None):
     return safely(lambda: store.workspaces(path))
 
 
+class BuildRunRequest(BaseModel):
+    output_locale: str = Field(default="", max_length=35, pattern=r"^[A-Za-z]{2,3}(?:-[A-Za-z0-9]{2,8})*$")
+
+
 @app.post("/api/builds/{build_id}/runs")
-def invoke_build(build_id: str):
-    return safely(lambda: store.invoke_remote_build(build_id))
+def invoke_build(build_id: str, values: BuildRunRequest | None = None):
+    return safely(
+        lambda: store.invoke_remote_build(build_id, output_locale=values.output_locale if values else None)
+    )
 
 
 @app.post("/api/builds/{build_id}/tests")
-def test_build(build_id: str):
-    return safely(lambda: store.test_build(build_id))
+def test_build(build_id: str, values: BuildRunRequest | None = None):
+    return safely(lambda: store.test_build(build_id, output_locale=values.output_locale if values else None))
 
 
 @app.get("/api/build-tests/{session_id}")

backend/app/models.py

@@ -72,6 +72,7 @@ class Run(BaseModel):
     build_name: str | None = None
     repository: str | None = None
     supervisor_profile_name: str | None = None
+    output_locale: str | None = None
     prompt_source: str | None = None
     prompt_snapshot: str | None = None
     execution_mode: Literal["run", "test"] = "run"

backend/app/store.py

@@ -2864,7 +2864,7 @@ def _write_prompt_template(
         temporary.replace(CONFIG / "prompt-templates.yaml")
         return template
 
-    def _assembled_prompt(self, build: dict[str, Any]) -> tuple[str, str]:
+    def _assembled_prompt(self, build: dict[str, Any], output_locale: str | None = None) -> tuple[str, str]:
         """Resolve the global manager contract and the build's evaluation policy."""
         template_id = build.get("manager_template_id", "manager-default-v1")
         template = next((item for item in self.prompt_templates() if item.get("id") == template_id), None)
@@ -2910,15 +2910,27 @@ def _assembled_prompt(self, build: dict[str, Any]) -> tuple[str, str]:
             )
             if part
         )
+        resolved_output_locale = (
+            str(output_locale or self.application_settings()["manager_output_locale"]).strip() or "en"
+        )
+        output_language_name = {
+            "en": "English",
+            "ja": "Japanese",
+            "ko": "Korean",
+        }.get(resolved_output_locale.lower().split("-", 1)[0], resolved_output_locale)
         assembled = "\n\n".join(
             part
             for part in (
                 operational.replace(MANAGER_PROMPT_SLOT, manager_policy).replace(
                     MANAGER_OUTPUT_LANGUAGE_SLOT,
                     "# Output language\n"
-                    "Write all human-readable string values in the configured application language "
-                    f"({self.application_settings()['manager_output_locale']}). "
-                    "Keep JSON keys, field names, and required enum values exactly as specified.",
+                    "Write every human-readable JSON string value in "
+                    f"{output_language_name} ({resolved_output_locale}). "
+                    "This applies even when the persona, rendered UI, or source evidence uses another language; "
+                    "do not switch to the persona's language for behavior_trace, summaries, improvements, or issues. "
+                    "If another instruction asks you to preserve an issue title, evidence, or reproduction, "
+                    "preserve its facts and severity rather than its source-language wording. "
+                    "Keep JSON keys, field names, required enum values, and verbatim UI strings quoted as evidence exactly as specified.",
                 ),
                 f"# Evaluation context\nRepository: {build.get('repository', '')}\n{legacy_context}",
                 case_text,
@@ -4030,6 +4042,7 @@ def create_run(
         supervisor_profile_name: str | None = None,
         prompt_source: str | None = None,
         prompt_snapshot: str | None = None,
+        output_locale: str | None = None,
         loop_limit: int = 1,
         timezone: str = "UTC",
         schedule_enabled: bool = False,
@@ -4069,6 +4082,7 @@ def create_run(
             build_name=build_name,
             repository=repository,
             supervisor_profile_name=supervisor_profile_name,
+            output_locale=output_locale,
             prompt_source=prompt_source,
             prompt_snapshot=prompt_snapshot,
             execution_mode=execution_mode,
@@ -4695,7 +4709,7 @@ def supervisor_result(result: object) -> object:
         supervisor_prompt = run.prompt_snapshot or ""
         if run.build_id:
             try:
-                _, supervisor_prompt = self._assembled_prompt(self.build(run.build_id))
+                _, supervisor_prompt = self._assembled_prompt(self.build(run.build_id), run.output_locale)
             except ValueError:
                 # The original immutable run snapshot remains a safe fallback
                 # if an operator has made the prompt temporarily unreadable.
@@ -5257,14 +5271,19 @@ def emergency_stop(self) -> list[Run]:
                 stopped.append(self.cancel(run.id))
         return stopped
 
-    def invoke_remote_build(self, build_id: str, execution_mode: str = "run") -> Run:
+    def invoke_remote_build(
+        self, build_id: str, execution_mode: str = "run", output_locale: str | None = None
+    ) -> Run:
         build = self.build(build_id)
         executor = build.get("executor", {})
         if not build.get("enabled"):
             raise ValueError("This build is not enabled.")
         if execution_mode not in {"run", "test"}:
             raise ValueError("execution_mode must be run or test")
-        prompt_source, prompt_snapshot = self._assembled_prompt(build)
+        resolved_output_locale = (
+            str(output_locale or self.application_settings()["manager_output_locale"]).strip() or "en"
+        )
+        prompt_source, prompt_snapshot = self._assembled_prompt(build, resolved_output_locale)
         if executor.get("type") != "remote-http":
             return self.create_run(
                 build["runner_id"],
@@ -5273,6 +5292,7 @@ def invoke_remote_build(self, build_id: str, execution_mode: str = "run") -> Run
                 build_id=build["id"],
                 build_name=build["name"],
                 supervisor_profile_name=build.get("model_profile_name"),
+                output_locale=resolved_output_locale,
                 prompt_source=prompt_source,
                 prompt_snapshot=prompt_snapshot,
                 loop_limit=1 if execution_mode == "test" else int(build.get("run_limit", 1)),
@@ -5314,6 +5334,7 @@ def invoke_remote_build(self, build_id: str, execution_mode: str = "run") -> Run
             build_id=build["id"],
             build_name=build["name"],
             supervisor_profile_name=build.get("model_profile_name"),
+            output_locale=resolved_output_locale,
             execution_mode=execution_mode,
             execution_type="invoke",
             status="queued",
@@ -5330,11 +5351,11 @@ def invoke_remote_build(self, build_id: str, execution_mode: str = "run") -> Run
         threading.Thread(target=self._execute_remote, args=(run.id, executor), daemon=True).start()
         return run
 
-    def test_build(self, build_id: str) -> Run:
+    def test_build(self, build_id: str, output_locale: str | None = None) -> Run:
         build = self.build(build_id)
         if not build.get("enabled"):
             raise ValueError("This build is not enabled.")
-        return self.invoke_remote_build(build_id, "test")
+        return self.invoke_remote_build(build_id, "test", output_locale=output_locale)
 
     def _execute_remote(self, run_id: str, executor: dict[str, Any]) -> None:
         run = self._load(run_id)

backend/tests/test_api.py

@@ -1525,5 +1525,23 @@ def test_manager_output_language_is_injected_into_the_assembled_prompt(tmp_path,
         encoding="utf-8",
     )
     _, prompt = store._assembled_prompt({"manager_template_id": "manager-default-v1", "repository": "test"})
-    assert "configured application language (ja)" in prompt
+    assert "Japanese (ja)" in prompt
     assert store_module.MANAGER_OUTPUT_LANGUAGE_SLOT not in prompt
+
+
+def test_run_output_language_overrides_the_shared_application_setting(tmp_path, monkeypatch):
+    monkeypatch.setattr(store_module, "SETTINGS", tmp_path / "settings.json")
+    monkeypatch.setattr(store_module, "CONFIG", tmp_path / "config")
+    store = store_module.ConsoleStore()
+    store.save_application_settings({"manager_output_locale": "en"})
+    (tmp_path / "config").mkdir(exist_ok=True)
+    (tmp_path / "config" / "prompt-templates.yaml").write_text(
+        "- id: manager-default-v1\n  name: Default\n  version: 1\n  content: Assess evidence.\n",
+        encoding="utf-8",
+    )
+    _, prompt = store._assembled_prompt(
+        {"manager_template_id": "manager-default-v1", "repository": "test"}, "ko"
+    )
+    assert "Korean (ko)" in prompt
+    assert "do not switch to the persona's language" in prompt
+    assert "facts and severity rather than its source-language wording" in prompt

frontend/src/domain/models.ts

@@ -15,7 +15,7 @@ export type BehaviorTrace = { persona_goal?:string; current_action?:string; next
 export type SupervisorEvaluation = { score:number; approval:'approved'|'rejected'|'pending'; behavior_summary?:string; behavior_trace?:BehaviorTrace; summary:string }
 export type SupervisorResult = { evaluation?:SupervisorEvaluation; improvements:Record<string,unknown>[]; reported_issues:Record<string,unknown>[] }
 export type SupervisorRecord = { iteration:number; candidate_id?:string|null; status:'pending'|'completed'|'not_configured'|'invalid_response'|'failed'; prompt?:string; response?:SupervisorResult; error?:string; recorded_at?:string }
-export type Run = { id:string; workflow_id:string; workflow_name:string;runner_version?:number;runner_source_sha256?:string; build_id?:string; build_name?:string; execution_mode?:'run'|'test'; execution_type?:'pipeline'|'invoke'; loop_limit?:number;start_iteration?:number;retry_of_run_id?:string;retry_mode?:'restart'|'resume';iteration_strategy?:'linear'|'score_select';candidates_per_iteration?:number;iteration_candidates?:{id:string;iteration:number;score?:number;status:string;selected?:boolean}[]; status:string; created_at?:string; updated_at?:string; finished_at?:string; current_phase?:string; pid?:number; last_pid?:number; telemetry_trace_id?:string; prompt_source?:string; prompt_snapshot?:string; supervisor_status?:'pending'|'completed'|'not_configured'|'invalid_response'|'failed'; supervisor_response?:SupervisorResult; supervisor_error?:string; supervisor_results?:SupervisorRecord[]; workflow_graph?:WorkflowGraphDefinition|null; step_results?:RunStepResult[]; proposed_improvements?:number; approved_improvements?:number; reported_issues?:number; approval_score?:number }
+export type Run = { id:string; workflow_id:string; workflow_name:string;runner_version?:number;runner_source_sha256?:string; build_id?:string; build_name?:string; execution_mode?:'run'|'test'; execution_type?:'pipeline'|'invoke'; loop_limit?:number;start_iteration?:number;retry_of_run_id?:string;retry_mode?:'restart'|'resume';iteration_strategy?:'linear'|'score_select';candidates_per_iteration?:number;iteration_candidates?:{id:string;iteration:number;score?:number;status:string;selected?:boolean}[]; status:string; created_at?:string; updated_at?:string; finished_at?:string; current_phase?:string; pid?:number; last_pid?:number; telemetry_trace_id?:string; output_locale?:string; prompt_source?:string; prompt_snapshot?:string; supervisor_status?:'pending'|'completed'|'not_configured'|'failed'; supervisor_response?:SupervisorResult; supervisor_error?:string; supervisor_results?:SupervisorRecord[]; workflow_graph?:WorkflowGraphDefinition|null; step_results?:RunStepResult[]; proposed_improvements?:number; approved_improvements?:number; reported_issues?:number; approval_score?:number }
 export type TelemetrySpan = { name:string; traceId:string; spanId:string; parentSpanId?:string|null; startTime?:number; endTime?:number; attributes?:Record<string,unknown>; status?:string; events?:{name:string;attributes?:Record<string,unknown>}[] }
 export type RunTelemetry = { trace_id?:string; spans:TelemetrySpan[] }
 export type PromptRevision = { iteration?:number; phase?:string; path:string; status:'initial'|'applied'|'blocked'|'rolled_back'|'unchanged'; reason?:string|null; recorded_at?:string; version_id?:string|null; run_id?:string|null; before?:string|null; after?:string|null; before_sha256?:string|null; after_sha256?:string|null }

frontend/src/features/settings/page.tsx

@@ -75,11 +75,6 @@ export function SettingsPage({
       .then((values) => {
         setPrompt(values.manager_prompt_template);
         setChatProfile(values.chat_model_profile_name);
-        if (values.manager_output_locale !== locale) {
-          return api<ApplicationSettings>("/api/application-settings", "PUT", {
-            manager_output_locale: locale,
-          });
-        }
       })
       .catch(() => pushToast("Unable to load operational prompt."));
   }, [locale, pushToast]);

frontend/src/main.tsx

@@ -173,14 +173,14 @@ export default function App() {
     });
   };
   const invoke = (id: string) =>
-    api(`/api/builds/${id}/runs`, "POST")
+    api(`/api/builds/${id}/runs`, "POST", { output_locale: locale })
       .then(() => {
         room.setNotice(ui.evaluationStarted, "success");
         room.refresh();
       })
       .catch((e) => room.setNotice(e.message));
   const testBuild = (id: string) =>
-    api<Run>(`/api/builds/${id}/tests`, "POST")
+    api<Run>(`/api/builds/${id}/tests`, "POST", { output_locale: locale })
       .then((run) => {
         room.setNotice(ui.evaluationTestStarted, "success");
         return run;