fix(langfuse): use compile() instead of .format() for prompt variable injection

Langfuse uses {{variable}} syntax in its prompt management UI, while the
hardcoded fallbacks use {variable} (Python str.format). The previous code
always called .format() which silently failed/errored when a real Langfuse
prompt was fetched.

- langfuse_client.py: add compile_prompt(template, prompt_obj, **vars)
  → uses prompt_obj.compile(**vars) when Langfuse is available
  → falls back to template.format(**vars) when using the hardcoded fallback
- agent_runner.py: replace .format() with compile_prompt() for
  unified_processing (V2 local) and batch_cloud_processing (cloud agent)
- agent_setup.py: replace .format() with compile_prompt() for journey_system

deep_agent.py prompts have no variables, so no change needed there.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

app/api/routes/agent_setup.py

@@ -32,7 +32,7 @@ from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, Tool
 
 from app.agents.filesystem_agent import FILESYSTEM_TOOLS
 from app.config.settings import settings
-from app.core.langfuse_client import extract_usage, get_langfuse, get_prompt_or_fallback
+from app.core.langfuse_client import compile_prompt, extract_usage, get_langfuse, get_prompt_or_fallback
 from app.core.llm import get_llm
 
 logger = logging.getLogger(__name__)
@@ -160,7 +160,9 @@ def _build_system_prompt(
     template, prompt_obj = get_prompt_or_fallback(
         "journey_system", _JOURNEY_SYSTEM_PROMPT
     )
-    compiled = template.format(
+    compiled = compile_prompt(
+        template,
+        prompt_obj,
         directory=directory,
         data_types=", ".join(data_types),
         template_start=_TEMPLATE_START,

app/core/agent_runner.py

@@ -45,7 +45,7 @@ from app.agents.task_agent import TASK_TOOLS
 from app.agents.timeline_agent import TIMELINE_TOOLS
 from app.config.settings import settings
 from app.core.device_manager import DeviceConnectionManager
-from app.core.langfuse_client import extract_usage, get_langfuse, get_prompt_or_fallback
+from app.core.langfuse_client import compile_prompt, extract_usage, get_langfuse, get_prompt_or_fallback
 from app.core.llm import get_llm
 from app.core.preprocessors import detect_content_type, preprocess
 from app.core.ws_context import clear_client_executor, execute_on_client, set_client_executor
@@ -661,7 +661,9 @@ async def run_local_agent(
                 )
                 metadata_section = _format_metadata(preprocessed.metadata)
 
-                system_prompt = unified_template.format(
+                system_prompt = compile_prompt(
+                    unified_template,
+                    prompt_obj,
                     filename=filename,
                     metadata_section=metadata_section,
                     projects_list=projects_block,
@@ -893,7 +895,9 @@ async def run_cloud_agent(
             cloud_template, cloud_prompt_obj = get_prompt_or_fallback(
                 "batch_cloud_processing", _BATCH_CLOUD_PROCESSING_PROMPT
             )
-            processing_prompt = cloud_template.format(
+            processing_prompt = compile_prompt(
+                cloud_template,
+                cloud_prompt_obj,
                 data_types=", ".join(config.data_types),
                 project_context="Determine the appropriate project from the message context.",
                 file_list=f"Message from {config.provider} (id: {msg.id})",

app/core/langfuse_client.py

@@ -80,10 +80,11 @@ def get_langfuse() -> Any | None:
 def get_prompt_or_fallback(name: str, fallback: str) -> tuple[str, Any]:
     """Fetch a text prompt from Langfuse; fall back to ``fallback`` on any error.
 
-    Returns ``(prompt_text, prompt_obj_or_None)``.
+    Returns ``(raw_template, prompt_obj_or_None)``.
 
-    * ``prompt_text`` — the raw template string (variables not yet substituted).
+    * ``raw_template`` — the uncompiled template string.  Do NOT call ``.format()``
-      Callers perform variable substitution with Python's ``.format()``.
+      on it directly; use :func:`compile_prompt` instead so the correct variable
+      syntax is applied (``{{var}}`` for Langfuse, ``{var}`` for the fallback).
     * ``prompt_obj`` — the Langfuse prompt object, or ``None`` when Langfuse is
       unavailable / the fetch failed.  Pass this to generation observations so
       Langfuse links the generation to the exact prompt version in the UI.
@@ -102,6 +103,38 @@ def get_prompt_or_fallback(name: str, fallback: str) -> tuple[str, Any]:
         return fallback, None
 
 
+def compile_prompt(template: str, prompt_obj: Any, **variables: Any) -> str:
+    """Compile *template* with *variables*, choosing the right syntax.
+
+    * When *prompt_obj* is a real Langfuse prompt object, calls
+      ``prompt_obj.compile(**variables)`` which handles ``{{variable}}``
+      substitution as defined in the Langfuse UI.
+    * When *prompt_obj* is ``None`` (Langfuse unavailable or fetch failed),
+      falls back to ``template.format(**variables)`` which handles the
+      ``{variable}`` syntax used in the hardcoded fallback strings.
+
+    This keeps callers oblivious to which syntax is in use.
+    """
+    if prompt_obj is not None:
+        try:
+            compiled = prompt_obj.compile(**variables)
+            # compile() returns a string for text prompts.
+            if isinstance(compiled, str):
+                return compiled
+            # Chat prompts return a list of dicts — join text parts.
+            if isinstance(compiled, list):
+                return "\n".join(
+                    m.get("content", "") for m in compiled if isinstance(m, dict)
+                )
+        except Exception as exc:
+            logger.warning(
+                "langfuse: compile failed for prompt %r: %s — falling back to .format()",
+                getattr(prompt_obj, "name", "?"),
+                exc,
+            )
+    return template.format(**variables)
+
+
 def extract_usage(response: Any) -> dict[str, int]:
     """Extract token usage from a LangChain AI message into Langfuse format."""
     meta = getattr(response, "usage_metadata", None)