fix: Langfuse SDK v4 migration, tracing improvements, and LLM config

- Langfuse SDK v4: fix prompt-to-trace linking (as_type=generation)
- tracing: compile_prompt with Langfuse managed prompt fallback
- journey: remove journey CLI subcommand (keep only interactive)
- LLM: add service-specific llm modules for batch-agent and chat
- gitignore: exclude eval private test data
- config: add LANGFUSE settings to shared config

.gitignore

@@ -35,3 +35,6 @@ Thumbs.db
 # Claude Code
 .claude/
 logs/
+
+# Eval private test data
+services/batch-agent/eval/fixtures/private_data/

app/config/settings.py

@@ -27,6 +27,7 @@ class Settings(BaseSettings):
     ANTHROPIC_API_KEY: str = ""
     GOOGLE_API_KEY: str = ""
     CEREBRAS_API_KEY: str = ""
+    GITHUB_TOKEN: str = ""
 
     LLM_MODEL: str = "gpt-4o"
     LLM_EMBED_MODEL: str = "text-embedding-3-small"

app/core/llm.py

@@ -50,6 +50,8 @@ def _api_key_for_model(model: str) -> str | None:
         return settings.GOOGLE_API_KEY or None
     if model.startswith("cerebras/"):
         return settings.CEREBRAS_API_KEY or None
+    if model.startswith("github/"):
+        return settings.GITHUB_TOKEN or None
     if model.startswith("github_copilot/"):
         # GitHub Copilot uses OAuth device-flow tokens managed by LiteLLM.
         # No API key is required; returning None lets LiteLLM handle auth.
@@ -83,6 +85,9 @@ def get_llm(
     if settings.GITHUB_COPILOT_TOKEN_DIR:
         os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
 
+    if settings.GITHUB_TOKEN:
+        os.environ.setdefault("GITHUB_TOKEN", settings.GITHUB_TOKEN)
+
     # Use ChatLiteLLM for provider-prefixed models (github_copilot/, anthropic/, etc.)
     # so LiteLLM handles routing and auth. ChatOpenAI for plain OpenAI model names.
     if "/" in model:

services/batch-agent/app/agent_runner.py

@@ -22,12 +22,13 @@ from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, Tool
 from sqlalchemy import select
 
 from app.agents.filesystem_agent import FILESYSTEM_TOOLS
-from app.agents.note_agent import NOTE_TOOLS
+from shared.agents.note_agent import NOTE_TOOLS
-from app.agents.project_agent import PROJECT_TOOLS
+from shared.agents.project_agent import PROJECT_TOOLS
-from app.agents.task_agent import TASK_TOOLS
+from shared.agents.task_agent import TASK_TOOLS
-from app.agents.timeline_agent import TIMELINE_TOOLS
+from shared.agents.timeline_agent import TIMELINE_TOOLS
-from app.llm import get_llm
+from shared.llm import get_llm
-from app.ws_context import execute_on_client, set_current_user, clear_current_user
+from shared.ws_context import execute_on_client, set_current_user, clear_current_user
+import app.tracing as tracing
 from shared.db import async_session
 from shared.models import AgentRunLog, CloudAgentConfig, LocalAgentConfig
 from shared.redis import redis_client, ws_out_channel
@@ -193,9 +194,11 @@ async def _run_agent_with_tools(
     user_message: str,
     tools: list[Any],
     max_steps: int,
+    langfuse_handler: Any | None = None,
 ) -> str:
     """Run an LLM agent with tool-calling, returning the final text response."""
-    llm = get_llm()
+    callbacks = [langfuse_handler] if langfuse_handler else None
+    llm = get_llm(callbacks=callbacks)
     llm_with_tools = llm.bind_tools(tools)
     messages: list[Any] = [
         SystemMessage(content=system_prompt),
@@ -396,6 +399,7 @@ async def _classify_file(
     file_content: str,
     projects: list[dict],
     config_data_types: list[str],
+    langfuse_handler: Any | None = None,
 ) -> tuple[str, list[str], str | None]:
     fallback: tuple[str, list[str], str | None] = ("new", list(config_data_types), None)
 
@@ -417,12 +421,16 @@ async def _classify_file(
         if d in _DOMAIN_DESCRIPTIONS
     )
 
-    system = _STEP1_SYSTEM_PROMPT.format(
+    system = tracing.compile_prompt(
-        domain_definitions=domain_definitions,
+        "batch_file_classifier",
-        projects_list=projects_list,
+        fallback=_STEP1_SYSTEM_PROMPT,
+        variables={
+            "domain_definitions": domain_definitions,
+            "projects_list": projects_list,
+        },
     )
 
-    llm = get_llm()
+    llm = get_llm(callbacks=[langfuse_handler] if langfuse_handler else None)
     try:
         response = await llm.ainvoke([
             SystemMessage(content=system),
@@ -458,7 +466,7 @@ async def _classify_file(
 # ── Local agent runner (two-step per file) ────────────────────────────────
 
 
-async def run_local_agent(user_id: str, trigger_data: dict[str, Any]) -> None:
+async def run_local_agent(user_id: str, trigger_data: dict[str, Any], *, langfuse_handler: Any | None = None) -> None:
     """Execute a local directory agent run.
 
     In the microservice world, trigger_data is a serialized dict from
@@ -552,6 +560,7 @@ async def run_local_agent(user_id: str, trigger_data: dict[str, Any]) -> None:
                     file_content=file_content,
                     projects=projects,
                     config_data_types=data_types,
+                    langfuse_handler=langfuse_handler,
                 )
 
                 # Step 2 — resolve project_id, fetch entities, process
@@ -593,11 +602,15 @@ async def run_local_agent(user_id: str, trigger_data: dict[str, Any]) -> None:
 
                 existing_context = "\n\n".join(existing_blocks)
 
-                system_prompt = _PROCESSING_SYSTEM_PROMPT.format(
+                system_prompt = tracing.compile_prompt(
-                    existing_context=existing_context,
+                    "batch_processing",
-                    project_context=project_context,
+                    fallback=_PROCESSING_SYSTEM_PROMPT,
-                    data_types=", ".join(domains),
+                    variables={
-                    custom_prompt_section=custom_section,
+                        "existing_context": existing_context,
+                        "project_context": project_context,
+                        "data_types": ", ".join(domains),
+                        "custom_prompt_section": custom_section,
+                    },
                 )
 
                 processing_tools = _build_processing_tools(domains)
@@ -610,6 +623,7 @@ async def run_local_agent(user_id: str, trigger_data: dict[str, Any]) -> None:
                     ),
                     tools=processing_tools,
                     max_steps=_MAX_PROCESSING_STEPS,
+                    langfuse_handler=langfuse_handler,
                 )
                 logger.info(
                     "agent_runner: run=%s file=%r result=%s",
@@ -660,7 +674,7 @@ async def run_local_agent(user_id: str, trigger_data: dict[str, Any]) -> None:
 _CLOUD_DEFAULT_LOOKBACK_DAYS: int = 7
 
 
-async def run_cloud_agent(user_id: str, config_id: str) -> None:
+async def run_cloud_agent(user_id: str, config_id: str, *, langfuse_handler: Any | None = None) -> None:
     """Execute a cloud connector agent run.
 
     Loads the CloudAgentConfig from DB, decrypts OAuth tokens, fetches
@@ -776,11 +790,15 @@ async def run_cloud_agent(user_id: str, config_id: str) -> None:
                 continue
             items_processed += 1
 
-            processing_prompt = _CLOUD_PROCESSING_PROMPT.format(
+            processing_prompt = tracing.compile_prompt(
-                data_types=", ".join(config.data_types),
+                "batch_cloud_processing",
-                project_context="Determine the appropriate project from the message context.",
+                fallback=_CLOUD_PROCESSING_PROMPT,
-                file_list=f"Message from {config.provider} (id: {msg.id})",
+                variables={
-                custom_prompt_section=custom_section,
+                    "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})",
+                    "custom_prompt_section": custom_section,
+                },
             )
 
             try:
@@ -789,6 +807,7 @@ async def run_cloud_agent(user_id: str, config_id: str) -> None:
                     user_message=f"Process this message content:\n\n{content_text[:8000]}",
                     tools=processing_tools,
                     max_steps=_MAX_PROCESSING_STEPS,
+                    langfuse_handler=langfuse_handler,
                 )
             except Exception as exc:
                 errors.append(f"LLM processing error for message {msg.id!r}: {exc}")

services/batch-agent/app/agents/filesystem_agent.py

@@ -9,7 +9,7 @@ from typing import Any
 
 from langchain_core.tools import tool
 
-from app.ws_context import execute_on_client
+from shared.ws_context import execute_on_client
 
 
 @tool

services/batch-agent/app/agents/note_agent.py

@@ -1,110 +0,0 @@
-"""Note agent — Markdown note management.
-
-Adapted for Batch Agent Service: import from app.ws_context and app.llm.
-"""
-
-from __future__ import annotations
-
-import re
-from typing import Any
-
-from langchain_core.tools import tool
-
-from app.llm import embed
-from app.ws_context import execute_on_client
-
-_UUID_RE = re.compile(
-    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
-)
-
-
-def _is_uuid(value: str) -> bool:
-    return bool(_UUID_RE.match(value))
-
-
-@tool
-async def list_notes(project_id: str = "") -> str:
-    """List notes, optionally scoped to a project by project_id."""
-    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
-    result = await execute_on_client(
-        action="select",
-        table="notes",
-        filters={"projectId": normalized_project_id or None},
-    )
-    rows = result.get("rows", [])
-    if not rows:
-        return "No notes found."
-    lines = [f"- {r['title']} (id: {r['id']})" for r in rows]
-    return f"Found {len(rows)} note(s):\n" + "\n".join(lines)
-
-
-@tool
-async def get_note(note_id: str) -> str:
-    """Fetch a single note by its UUID to read its full Markdown content."""
-    result = await execute_on_client(action="get", table="notes", data={"id": note_id})
-    row = result.get("row")
-    if not row:
-        return f"Note {note_id} not found."
-    return f"Note '{row['title']}' (id: {row['id']}):\n\n{row['content']}"
-
-
-@tool
-async def create_note(title: str, content: str, project_id: str = "") -> str:
-    """Create a new note."""
-    result = await execute_on_client(
-        action="insert",
-        table="notes",
-        data={
-            "title": title,
-            "content": content,
-            "projectId": project_id or None,
-        },
-    )
-    row = result["row"]
-    vector = await embed(content)
-    await execute_on_client(
-        action="vector_upsert",
-        data={"id": row["id"], "projectId": row.get("projectId"), "content": content},
-        vector=vector,
-    )
-    return f"Note created: '{row['title']}' (id: {row['id']})."
-
-
-@tool
-async def update_note(note_id: str, title: str = "", content: str = "") -> str:
-    """Update an existing note. Only pass fields that should change."""
-    updates: dict[str, Any] = {}
-    if title:
-        updates["title"] = title
-    if content:
-        updates["content"] = content
-    result = await execute_on_client(
-        action="update",
-        table="notes",
-        data={"id": note_id, "updates": updates},
-    )
-    row = result["row"]
-    if content:
-        vector = await embed(content)
-        await execute_on_client(
-            action="vector_upsert",
-            data={"id": note_id, "projectId": row.get("projectId"), "content": content},
-            vector=vector,
-        )
-    return f"Note updated: '{row['title']}' (id: {row['id']})."
-
-
-@tool
-async def delete_note(note_id: str) -> str:
-    """Delete a note permanently by its UUID."""
-    await execute_on_client(action="delete", table="notes", data={"id": note_id})
-    return f"Note {note_id} deleted."
-
-
-NOTE_TOOLS: list[Any] = [
-    list_notes,
-    get_note,
-    create_note,
-    update_note,
-    delete_note,
-]

services/batch-agent/app/agents/project_agent.py

@@ -1,110 +0,0 @@
-"""Project agent — full lifecycle management.
-
-Adapted for Batch Agent Service: import from app.ws_context.
-"""
-
-from __future__ import annotations
-
-from typing import Any
-
-from langchain_core.tools import tool
-
-from app.ws_context import execute_on_client
-
-
-@tool
-async def list_projects(client_id: str = "", include_archived: int = 0) -> str:
-    """List projects, optionally filtered by client_id."""
-    result = await execute_on_client(
-        action="select",
-        table="projects",
-        filters={
-            "clientId": client_id or None,
-            "includeArchived": bool(include_archived),
-        },
-    )
-    rows = result.get("rows", [])
-    if not rows:
-        return "No projects found."
-    lines = [f"- {r['name']} (status: {r['status']}, id: {r['id']})" for r in rows]
-    return f"Found {len(rows)} project(s):\n" + "\n".join(lines)
-
-
-@tool
-async def list_all_projects() -> str:
-    """List every project regardless of client or status."""
-    result = await execute_on_client(action="select", table="projects")
-    rows = result.get("rows", [])
-    if not rows:
-        return "No projects found."
-    lines = [f"- {r['name']} (status: {r['status']}, id: {r['id']})" for r in rows]
-    return f"All projects ({len(rows)}):\n" + "\n".join(lines)
-
-
-@tool
-async def get_project(project_id: str) -> str:
-    """Fetch a single project by its UUID."""
-    result = await execute_on_client(action="get", table="projects", data={"id": project_id})
-    row = result.get("row")
-    if not row:
-        return f"Project {project_id} not found."
-    return (
-        f"Project: '{row['name']}' (id: {row['id']}, status: {row['status']}, "
-        f"clientId: {row.get('clientId', 'none')})"
-    )
-
-
-@tool
-async def create_project(name: str, client_id: str = "") -> str:
-    """Create a new project."""
-    result = await execute_on_client(
-        action="insert",
-        table="projects",
-        data={"name": name, "clientId": client_id or None},
-    )
-    row = result["row"]
-    return f"Project created: '{row['name']}' (id: {row['id']})"
-
-
-@tool
-async def update_project(
-    project_id: str,
-    name: str = "",
-    client_id: str = "",
-    status: str = "",
-    ai_summary: str = "",
-) -> str:
-    """Update a project. Only pass fields that should change."""
-    updates: dict[str, Any] = {}
-    if name:
-        updates["name"] = name
-    if client_id:
-        updates["clientId"] = client_id
-    if status:
-        updates["status"] = status
-    if ai_summary:
-        updates["aiSummary"] = ai_summary
-    result = await execute_on_client(
-        action="update",
-        table="projects",
-        data={"id": project_id, "updates": updates},
-    )
-    row = result["row"]
-    return f"Project updated: '{row['name']}' (id: {row['id']}, status: {row['status']})"
-
-
-@tool
-async def delete_project(project_id: str) -> str:
-    """Permanently delete a project."""
-    await execute_on_client(action="delete", table="projects", data={"id": project_id})
-    return f"Project {project_id} permanently deleted."
-
-
-PROJECT_TOOLS: list[Any] = [
-    list_projects,
-    list_all_projects,
-    get_project,
-    create_project,
-    update_project,
-    delete_project,
-]

services/batch-agent/app/agents/task_agent.py

@@ -1,197 +0,0 @@
-"""Task agent — full CRUD for tasks and task comments.
-
-Adapted for Batch Agent Service: import from app.ws_context.
-"""
-
-from __future__ import annotations
-
-from datetime import datetime, timezone
-import re
-from typing import Any
-
-from langchain_core.tools import tool
-
-from app.ws_context import execute_on_client
-
-_UUID_RE = re.compile(
-    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
-)
-
-
-def _is_uuid(value: str) -> bool:
-    return bool(_UUID_RE.match(value))
-
-
-@tool
-async def list_tasks(
-    project_id: str = "",
-    status: str = "",
-    search: str = "",
-    order_by: str = "",
-) -> str:
-    """List tasks, optionally filtered by project_id, status, search, or order_by."""
-    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
-    result = await execute_on_client(
-        action="select",
-        table="tasks",
-        filters={
-            "projectId": normalized_project_id or None,
-            "status": status or None,
-            "search": search or None,
-            "orderBy": order_by or None,
-        },
-    )
-    rows = result.get("rows", [])
-    if not rows:
-        return "No tasks found matching the given filters."
-    lines = [
-        f"- {r['title']} (status: {r['status']}, priority: {r['priority']}, id: {r['id']})"
-        for r in rows
-    ]
-    return f"Found {len(rows)} task(s):\n" + "\n".join(lines)
-
-
-@tool
-async def create_task(
-    title: str,
-    description: str = "",
-    status: str = "todo",
-    priority: str = "medium",
-    assignees: str = "[]",
-    due_date: int = 0,
-    project_id: str = "",
-    is_ai_suggested: int = 0,
-) -> str:
-    """Create a new task."""
-    result = await execute_on_client(
-        action="insert",
-        table="tasks",
-        data={
-            "title": title,
-            "description": description or None,
-            "status": status,
-            "priority": priority,
-            "assignee": assignees,
-            "dueDate": due_date or None,
-            "projectId": project_id or None,
-            "isAiSuggested": is_ai_suggested,
-        },
-    )
-    row = result["row"]
-    return (
-        f"Task created: '{row['title']}' "
-        f"(id: {row['id']}, status: {row['status']}, priority: {row['priority']})"
-    )
-
-
-@tool
-async def update_task(
-    task_id: str,
-    title: str = "",
-    description: str = "",
-    status: str = "",
-    priority: str = "",
-    assignees: str = "",
-    due_date: int = -1,
-    project_id: str = "",
-) -> str:
-    """Update fields on an existing task. Only pass fields you want to change."""
-    updates: dict[str, Any] = {}
-    if title:
-        updates["title"] = title
-    if description:
-        updates["description"] = description
-    if status:
-        updates["status"] = status
-    if priority:
-        updates["priority"] = priority
-    if assignees:
-        updates["assignee"] = assignees
-    if due_date != -1:
-        updates["dueDate"] = due_date or None
-    if project_id:
-        updates["projectId"] = project_id
-    result = await execute_on_client(
-        action="update",
-        table="tasks",
-        data={"id": task_id, "updates": updates},
-    )
-    row = result["row"]
-    return f"Task updated: '{row['title']}' (id: {row['id']}, status: {row['status']})"
-
-
-@tool
-async def delete_task(task_id: str) -> str:
-    """Delete a task permanently by its UUID."""
-    await execute_on_client(action="delete", table="tasks", data={"id": task_id})
-    return f"Task {task_id} deleted."
-
-
-@tool
-async def list_tasks_due_today() -> str:
-    """List all tasks whose due date falls on today's date."""
-    now = datetime.now(tz=timezone.utc)
-    start_ms = int(datetime(now.year, now.month, now.day, tzinfo=timezone.utc).timestamp() * 1000)
-    end_ms = start_ms + 86_400_000 - 1
-    result = await execute_on_client(
-        action="select",
-        table="tasks",
-        filters={"dueDateFrom": start_ms, "dueDateTo": end_ms},
-    )
-    rows = result.get("rows", [])
-    if not rows:
-        return "No tasks are due today."
-    lines = [
-        f"- {r['title']} (priority: {r['priority']}, status: {r['status']}, id: {r['id']})"
-        for r in rows
-    ]
-    return f"Tasks due today ({len(rows)}):\n" + "\n".join(lines)
-
-
-@tool
-async def list_task_comments(task_id: str) -> str:
-    """List all comments on a task by its UUID."""
-    result = await execute_on_client(
-        action="select",
-        table="taskComments",
-        filters={"taskId": task_id},
-    )
-    rows = result.get("rows", [])
-    if not rows:
-        return f"No comments found for task {task_id}."
-    lines = [f"- [{r['author']}]: {r['content']} (id: {r['id']})" for r in rows]
-    return f"Found {len(rows)} comment(s):\n" + "\n".join(lines)
-
-
-@tool
-async def add_task_comment(task_id: str, author: str, content: str) -> str:
-    """Add a comment to a task."""
-    result = await execute_on_client(
-        action="insert",
-        table="taskComments",
-        data={"taskId": task_id, "author": author, "content": content},
-    )
-    row = result.get("row", {})
-    row_author = row.get("author", author)
-    row_task_id = row.get("taskId") or row.get("task_id") or task_id
-    row_comment_id = row.get("id", "unknown")
-    return f"Comment added by {row_author} on task {row_task_id} (comment id: {row_comment_id})."
-
-
-@tool
-async def delete_task_comment(comment_id: str) -> str:
-    """Delete a task comment by its UUID."""
-    await execute_on_client(action="delete", table="taskComments", data={"id": comment_id})
-    return f"Comment {comment_id} deleted."
-
-
-TASK_TOOLS: list[Any] = [
-    list_tasks,
-    create_task,
-    update_task,
-    delete_task,
-    list_tasks_due_today,
-    list_task_comments,
-    add_task_comment,
-    delete_task_comment,
-]

services/batch-agent/app/agents/timeline_agent.py

@@ -1,88 +0,0 @@
-"""Timeline agent — project milestone management.
-
-Adapted for Batch Agent Service: import from app.ws_context.
-"""
-
-from __future__ import annotations
-
-import re
-from typing import Any
-
-from langchain_core.tools import tool
-
-from app.ws_context import execute_on_client
-
-_UUID_RE = re.compile(
-    r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
-)
-
-
-def _is_uuid(value: str) -> bool:
-    return bool(_UUID_RE.match(value))
-
-
-@tool
-async def list_timelines(project_id: str = "") -> str:
-    """List timelines. Provide project_id to scope to a specific project."""
-    normalized_project_id = project_id if (project_id and _is_uuid(project_id)) else ""
-    result = await execute_on_client(
-        action="select",
-        table="timelines",
-        filters={"projectId": normalized_project_id or None},
-    )
-    rows = result.get("rows", [])
-    if not rows:
-        return "No timelines found."
-    lines = [f"- {r['title']} (date: {r['date']}, id: {r['id']})" for r in rows]
-    return f"Found {len(rows)} timeline(s):\n" + "\n".join(lines)
-
-
-@tool
-async def create_timeline(
-    project_id: str, title: str, date: int, is_ai_suggested: int = 0,
-) -> str:
-    """Create a project timeline (milestone)."""
-    result = await execute_on_client(
-        action="insert",
-        table="timelines",
-        data={
-            "projectId": project_id,
-            "title": title,
-            "date": date,
-            "isAiSuggested": is_ai_suggested,
-        },
-    )
-    row = result["row"]
-    return f"Timeline created: '{row['title']}' (id: {row['id']}, date: {row['date']})"
-
-
-@tool
-async def update_timeline(timeline_id: str, title: str = "", date: int = -1) -> str:
-    """Update a timeline. Only pass fields that should change."""
-    updates: dict[str, Any] = {}
-    if title:
-        updates["title"] = title
-    if date != -1:
-        updates["date"] = date
-    result = await execute_on_client(
-        action="update",
-        table="timelines",
-        data={"id": timeline_id, "updates": updates},
-    )
-    row = result["row"]
-    return f"Timeline updated: '{row['title']}' (id: {row['id']})"
-
-
-@tool
-async def delete_timeline(timeline_id: str) -> str:
-    """Delete a timeline permanently by its UUID."""
-    await execute_on_client(action="delete", table="timelines", data={"id": timeline_id})
-    return f"Timeline {timeline_id} deleted."
-
-
-TIMELINE_TOOLS: list[Any] = [
-    list_timelines,
-    create_timeline,
-    update_timeline,
-    delete_timeline,
-]

services/batch-agent/app/journey.py

@@ -25,7 +25,8 @@ from typing import Any
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 
 from app.agents.filesystem_agent import FILESYSTEM_TOOLS
-from app.llm import get_llm
+from shared.llm import get_llm
+import app.tracing as tracing
 
 logger = logging.getLogger(__name__)
 
@@ -79,17 +80,9 @@ def get_journey_session(session_id: str, user_id: str) -> JourneySession | None:
 _SYSTEM_PROMPT_TEMPLATE = """\
 You are a friendly assistant helping a freelancer configure a data-extraction agent.
 Your job is to understand exactly what data the user wants to extract from their
-local directory and produce a detailed prompt_template that a separate AI will use
+local directory and produce a concise prompt_template that a separate AI will use
 as its instruction set.
 
-The extraction agent already has this base behaviour built in:
-  - Reads each file using file-system tools.
-  - Creates records (tasks, notes, timelines, projects) via CRUD tools.
-  - Sets isAiSuggested=1 on every new record.
-  - Only extracts data explicitly present in the files — it never invents information.
-The user's custom prompt is appended AFTER this base behaviour, so focus on
-what to look for and how to map it — not on the general extraction mechanics.
-
 You have access to file-system tools to explore the user's directory:
 - list_directory: to see folder structure
 - read_file_content: to peek at file contents
@@ -98,38 +91,43 @@ You have access to file-system tools to explore the user's directory:
 The user's configured directory is: {directory}
 Target data types: {data_types}
 
-IMPORTANT — project assignment is handled automatically by the main agent runner
+IMPORTANT — project assignment is handled automatically.  You MUST NOT ask the user
-before the custom prompt is ever used.  You MUST NOT ask the user about projects,
+about projects, projectId, or how to link records to projects.  Never include
-projectId, or how to link records to projects.  Never include projectId logic or
+projectId logic or project creation instructions in the generated prompt_template.
-project creation instructions in the generated prompt_template.
 
 Start by exploring the directory to understand its structure.  Then ask concise,
-focused questions one at a time.  Cover these topics (not necessarily in this order):
+focused questions one at a time.  Cover only the topics relevant to the target
-  1. The type and format of the source content (confirmed by your exploration).
+data types listed above:
-  2. How fields should be mapped (e.g. filename → task title).
-  3. Priority or status rules (e.g. "urgent" keyword → high priority).
-  4. Any special handling, date extraction, or exclusions.
 
-Once you reach 90% confidence, output the final prompt_template between these exact
+  1. Content type and format — confirmed by your exploration.
-markers on their own lines:
+  2. For TASKS (if in scope): field mapping for title, status, priority, content,
+       dueDate (where is the date found? what's the fallback when absent?),
+       and assignee (is there a person name to assign?).
+  3. For NOTES when TASKS are also in scope: note vs task distinction —
+       what makes something a note rather than a task?
+  4. For TIMELINES (if in scope): the date source — what marks a milestone or event?
+  5. Exclusions and special handling applicable to the target data types.
+
+Keep asking focused questions until you are at least 90% confident.  Then stop and
+output the final prompt_template immediately, wrapped between these exact markers
+on their own lines:
 
 {template_start}
 <the complete extraction prompt here>
 {template_end}
 
-The prompt_template must be a self-contained instruction for an AI that reads files
+The prompt_template must be concise (bullet points, ~15–25 lines maximum).
-and must perform CRUD operations using tools to create records.  It should specify:
+Specify only:
-  - What entity types to create (tasks, notes, timelines) — never projects.
+  - Scope: what files/content qualify and what entity types to create.
-  - How to map file content to record fields (camelCase: title, status, priority,
+  - Field mapping rules per entity type (camelCase fields: title, status, priority,
-    dueDate, content, etc.) — never include projectId.
+    dueDate, content, assignee, etc.).
-  - That isAiSuggested must be set to 1 on every new record.
+  - dueDate rule (if tasks in scope): source and fallback behaviour.
-  - Concrete examples of mappings based on what you discovered in the directory.
+  - Note vs task rule (if both in scope): the criterion that separates them.
+  - Timeline date rule (if timelines in scope): what constitutes a timeline event.
+  - Exclusion/filtering rules.
+  - 2–3 concrete mapping examples based on what you discovered.
 
-{existing_section}\
+{existing_section}Begin by exploring the directory, then ask your first question.\
-Keep asking clarifying questions until you are at least 90% confident you have
-enough information to generate an accurate prompt_template.  Once you reach that
-confidence level, stop asking and produce the final template immediately.
-Begin by exploring the directory, then ask your first question.\
 """
 
 
@@ -144,12 +142,15 @@ def _build_system_prompt(
         if existing_template
         else ""
     )
-    return _SYSTEM_PROMPT_TEMPLATE.format(
+    # Use Langfuse compile_prompt ({{variable}} syntax) with Python .format() fallback
-        directory=directory,
+    return tracing.compile_prompt(
-        data_types=", ".join(data_types),
+        "journey_system",
-        template_start=_TEMPLATE_START,
+        fallback=_SYSTEM_PROMPT_TEMPLATE,
-        template_end=_TEMPLATE_END,
+        variables={
-        existing_section=existing_section,
+            "directory": directory,
+            "data_types": ", ".join(data_types),
+            "existing_section": existing_section,
+        },
     )
 
 
@@ -190,6 +191,7 @@ async def _call_llm_with_tools(
     system_prompt: str,
     history: list[dict[str, Any]],
     tools: list[Any],
+    langfuse_handler: Any | None = None,
 ) -> str:
     """Build LangChain messages from history and invoke the LLM with tools.
 
@@ -203,7 +205,8 @@ async def _call_llm_with_tools(
         else:
             messages.append(AIMessage(content=turn["content"]))
 
-    llm = get_llm(model=None, temperature=0.4)
+    callbacks = [langfuse_handler] if langfuse_handler else None
+    llm = get_llm(model=None, temperature=0.4, callbacks=callbacks)
     llm_with_tools = llm.bind_tools(tools)
     tool_map = {tool_def.name: tool_def for tool_def in tools}
 
@@ -247,6 +250,8 @@ async def _call_llm_with_tools(
 async def handle_journey_start(
     user_id: str,
     frame: dict[str, Any],
+    *,
+    langfuse_handler: Any | None = None,
 ) -> dict[str, Any]:
     """Handle a ``journey_start`` request.
 
@@ -277,6 +282,7 @@ async def handle_journey_start(
         system_prompt=system_prompt,
         history=seed_history,
         tools=list(FILESYSTEM_TOOLS),
+        langfuse_handler=langfuse_handler,
     )
 
     session.history.extend(seed_history)
@@ -313,6 +319,8 @@ async def handle_journey_start(
 async def handle_journey_message(
     user_id: str,
     frame: dict[str, Any],
+    *,
+    langfuse_handler: Any | None = None,
 ) -> dict[str, Any]:
     """Handle a ``journey_message`` request.
 
@@ -338,6 +346,7 @@ async def handle_journey_message(
         system_prompt=session.system_prompt,
         history=session.history,
         tools=list(FILESYSTEM_TOOLS),
+        langfuse_handler=langfuse_handler,
     )
 
     session.history.append({"role": "assistant", "content": ai_reply})
@@ -358,6 +367,7 @@ async def handle_journey_message(
                 system_prompt=session.system_prompt,
                 history=session.history,
                 tools=list(FILESYSTEM_TOOLS),
+                langfuse_handler=langfuse_handler,
             )
             session.history.append({"role": "assistant", "content": nudge_reply})
 

services/batch-agent/app/llm.py

@@ -32,6 +32,8 @@ def _api_key_for_model(model: str) -> str | None:
         return settings.GOOGLE_API_KEY or None
     if model.startswith("cerebras/"):
         return settings.CEREBRAS_API_KEY or None
+    if model.startswith("github/"):
+        return settings.GITHUB_TOKEN or None
     if model.startswith("github_copilot/"):
         return None
     return settings.OPENAI_API_KEY or None
@@ -41,29 +43,27 @@ def get_llm(
     *,
     model: str | None = None,
     temperature: float = 0,
+    callbacks: list | None = None,
 ) -> ChatOpenAI | ChatLiteLLM:
     model = model or settings.LLM_MODEL
 
     if settings.GITHUB_COPILOT_TOKEN_DIR:
         os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
 
+    if settings.GITHUB_TOKEN:
+        os.environ.setdefault("GITHUB_TOKEN", settings.GITHUB_TOKEN)
+
     if "/" in model:
-        return ChatLiteLLM(model=model, temperature=temperature)
+        return ChatLiteLLM(model=model, temperature=temperature, callbacks=callbacks)
 
     return ChatOpenAI(
         model=model,
         temperature=temperature,
         api_key=_api_key_for_model(model),
+        callbacks=callbacks,
     )
 
 
-def get_router_llm(
-    *,
-    temperature: float = 0,
-) -> ChatOpenAI | ChatLiteLLM:
-    return get_llm(model=settings.LLM_ROUTER_MODEL, temperature=temperature)
-
-
 async def embed(text: str) -> list[float]:
     model = settings.LLM_EMBED_MODEL
 

services/batch-agent/app/main.py

@@ -14,6 +14,14 @@ from __future__ import annotations
 
 import asyncio
 import logging
+import sys
+from pathlib import Path
+
+# Ensure the repo root is on sys.path so ``shared`` is importable when
+# running locally (in Docker the COPY already places it at /app/shared/).
+_repo_root = str(Path(__file__).resolve().parents[3])
+if _repo_root not in sys.path:
+    sys.path.insert(0, _repo_root)
 
 from contextlib import asynccontextmanager
 from typing import AsyncGenerator
@@ -29,6 +37,10 @@ logger = logging.getLogger(__name__)
 
 @asynccontextmanager
 async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
+    # Initialise Langfuse tracing (no-op if keys are missing)
+    from app.tracing import init_langfuse
+    init_langfuse()
+
     logger.info("batch-agent: starting Redis consumer")
     task = asyncio.create_task(start_consumer())
     yield
@@ -37,6 +49,16 @@ async def lifespan(app: FastAPI) -> AsyncGenerator[None, None]:
         await task
     except asyncio.CancelledError:
         pass
+
+    from app.tracing import shutdown as shutdown_langfuse
+    shutdown_langfuse()
+
+    from shared.db import engine
+    await engine.dispose()
+
+    from shared.redis import redis_client
+    await redis_client.aclose()
+
     logger.info("batch-agent: Redis consumer stopped")
 
 

services/batch-agent/app/redis_consumer.py

@@ -17,7 +17,8 @@ from typing import Any
 
 from shared.redis import redis_client, batch_request_channel, ws_out_channel
 
-from app.ws_context import set_current_user, clear_current_user
+import app.tracing as tracing
+from shared.ws_context import set_current_user, clear_current_user
 
 logger = logging.getLogger(__name__)
 
@@ -32,15 +33,28 @@ async def _handle_journey_start(user_id: str, data: dict[str, Any]) -> None:
     """Handle a journey_start request from WS Gateway."""
     from app.journey import handle_journey_start
 
+    session_id = data.get("session_id", "")
     set_current_user(user_id)
     try:
-        reply = await handle_journey_start(user_id, data)
+        with tracing.trace_span(
+            name="journey_start",
+            user_id=user_id,
+            session_id=session_id,
+            input=data.get("directory", ""),
+            metadata={"data_types": data.get("data_types", [])},
+            tags=["journey"],
+        ) as span:
+            langfuse_handler = tracing.get_langfuse_callback()
+            reply = await handle_journey_start(user_id, data, langfuse_handler=langfuse_handler)
+            tracing.link_prompt_to_trace(span, "journey_system")
+            span.update(output=reply.get("message", "")[:500])
             await _publish_to_user(user_id, reply)
+        tracing.flush()
     except Exception as exc:
         logger.error("batch-agent: journey_start failed user=%s: %s", user_id, exc)
         await _publish_to_user(user_id, {
             "type": "journey_reply",
-            "session_id": data.get("session_id", ""),
+            "session_id": session_id,
             "message": f"Journey setup failed: {exc}",
             "done": True,
             "prompt_template": None,
@@ -53,15 +67,27 @@ async def _handle_journey_message(user_id: str, data: dict[str, Any]) -> None:
     """Handle a journey_message from WS Gateway."""
     from app.journey import handle_journey_message
 
+    session_id = data.get("session_id", "")
     set_current_user(user_id)
     try:
-        reply = await handle_journey_message(user_id, data)
+        with tracing.trace_span(
+            name="journey_message",
+            user_id=user_id,
+            session_id=session_id,
+            input=data.get("message", "")[:200],
+            tags=["journey"],
+        ) as span:
+            langfuse_handler = tracing.get_langfuse_callback()
+            reply = await handle_journey_message(user_id, data, langfuse_handler=langfuse_handler)
+            tracing.link_prompt_to_trace(span, "journey_system")
+            span.update(output=reply.get("message", "")[:500])
             await _publish_to_user(user_id, reply)
+        tracing.flush()
     except Exception as exc:
         logger.error("batch-agent: journey_message failed user=%s: %s", user_id, exc)
         await _publish_to_user(user_id, {
             "type": "journey_reply",
-            "session_id": data.get("session_id", ""),
+            "session_id": session_id,
             "message": f"Journey processing failed: {exc}",
             "done": True,
             "prompt_template": None,
@@ -74,15 +100,29 @@ async def _handle_agent_trigger(user_id: str, data: dict[str, Any]) -> None:
     """Handle an agent_trigger request from the REST route (forwarded via Redis)."""
     from app.agent_runner import run_local_agent
 
+    run_context = data.get("run_context", {})
+    agent_id = run_context.get("agent_id", "")
     set_current_user(user_id)
     try:
-        await run_local_agent(user_id, data)
+        with tracing.trace_span(
+            name="agent_trigger",
+            user_id=user_id,
+            trace_id=run_context.get("run_id"),
+            input={"agent_id": agent_id, "directory": data.get("directory", "")},
+            metadata={"data_types": data.get("data_types", [])},
+            tags=["batch", "agent_run"],
+        ) as span:
+            langfuse_handler = tracing.get_langfuse_callback()
+            await run_local_agent(user_id, data, langfuse_handler=langfuse_handler)
+            tracing.link_prompt_to_trace(span, "batch_processing")
+            span.update(output={"status": "completed"})
+        tracing.flush()
     except Exception as exc:
         logger.error("batch-agent: agent_trigger failed user=%s: %s", user_id, exc)
         await _publish_to_user(user_id, {
             "type": "run_complete",
             "status": "error",
-            "run_context": data.get("run_context", {}),
+            "run_context": run_context,
         })
     finally:
         clear_current_user()
@@ -98,6 +138,8 @@ async def _dispatch(user_id: str, message_data: dict[str, Any]) -> None:
         await _handle_journey_message(user_id, message_data)
     elif msg_type == "agent_trigger":
         await _handle_agent_trigger(user_id, message_data)
+    elif msg_type == "device_online":
+        logger.info("batch-agent: device_online user=%s device=%s", user_id, message_data.get("device_id", "?"))
     else:
         logger.warning("batch-agent: unknown message type %r from user=%s", msg_type, user_id)
 

services/batch-agent/app/tracing.py

@@ -0,0 +1,336 @@
+"""Langfuse tracing & prompt management for the Batch Agent Service (v4 SDK).
+
+Provides:
+- ``init_langfuse()`` — initialise the singleton client at startup
+- ``trace_span()`` — context manager that creates a trace + span
+- ``get_langfuse_callback()`` — LangChain callback handler (auto-inherits trace)
+- ``get_prompt()`` — fetch a managed prompt from Langfuse by name
+- ``flush()`` / ``shutdown()`` — lifecycle management
+
+All functions gracefully degrade to no-ops when Langfuse is not configured,
+so the service works identically with or without observability keys.
+
+Requires ``langfuse >= 3.0.0`` (v4 / "Fast Preview" SDK).
+"""
+
+from __future__ import annotations
+
+import logging
+from contextlib import contextmanager
+from typing import Any
+
+from shared.config import settings
+
+logger = logging.getLogger(__name__)
+
+# ── State ────────────────────────────────────────────────────────────────
+
+_initialised: bool = False
+_disabled: bool = False
+
+
+def _is_configured() -> bool:
+    return bool(settings.LANGFUSE_SECRET_KEY and settings.LANGFUSE_PUBLIC_KEY)
+
+
+def init_langfuse() -> None:
+    """Initialise the Langfuse singleton. Call once at startup."""
+    global _initialised, _disabled
+
+    if _initialised or _disabled:
+        return
+
+    if not _is_configured():
+        _disabled = True
+        logger.info("tracing: Langfuse keys not set — tracing disabled")
+        return
+
+    try:
+        from langfuse import Langfuse
+
+        Langfuse(
+            secret_key=settings.LANGFUSE_SECRET_KEY,
+            public_key=settings.LANGFUSE_PUBLIC_KEY,
+            host=settings.LANGFUSE_HOST,
+        )
+        _initialised = True
+        logger.info("tracing: Langfuse client initialised (host=%s)", settings.LANGFUSE_HOST)
+    except Exception as exc:
+        _disabled = True
+        logger.warning("tracing: failed to initialise Langfuse: %s", exc)
+
+
+def _get_client() -> Any | None:
+    """Return the singleton Langfuse client, or *None* if disabled."""
+    if _disabled:
+        return None
+    if not _initialised:
+        init_langfuse()
+    if _disabled:
+        return None
+    try:
+        from langfuse import get_client
+        return get_client()
+    except Exception:
+        return None
+
+
+# ── Null span (no-op when Langfuse is disabled) ─────────────────────────
+
+
+class _NullSpan:
+    """Drop-in replacement when Langfuse is disabled."""
+
+    def update(self, **_: Any) -> None: ...
+    def set_trace_io(self, **_: Any) -> None: ...
+    def score_trace(self, **_: Any) -> None: ...
+
+
+# ── Trace context manager ───────────────────────────────────────────────
+
+
+@contextmanager
+def trace_span(
+    *,
+    name: str,
+    user_id: str,
+    session_id: str | None = None,
+    trace_id: str | None = None,
+    input: Any = None,
+    metadata: dict[str, Any] | None = None,
+    tags: list[str] | None = None,
+):
+    """Context manager that creates a Langfuse trace/span.
+
+    Yields the span object (or a ``_NullSpan`` if Langfuse is disabled).
+    A ``CallbackHandler`` created inside this block auto-inherits the trace
+    context, so there is no need to pass trace IDs manually.
+    """
+    lf = _get_client()
+    if lf is None:
+        yield _NullSpan()
+        return
+
+    try:
+        from langfuse import Langfuse, propagate_attributes
+
+        trace_ctx: dict[str, str] = {}
+        if trace_id is not None:
+            trace_ctx["trace_id"] = Langfuse.create_trace_id(seed=trace_id)
+
+        with lf.start_as_current_observation(
+            as_type="span",
+            name=name,
+            input=input,
+            metadata=metadata or {},
+            **({"trace_context": trace_ctx} if trace_ctx else {}),
+        ) as span:
+            with propagate_attributes(
+                user_id=user_id,
+                session_id=session_id,
+                tags=tags or [],
+            ):
+                yield span
+    except Exception as exc:
+        logger.warning("tracing: trace_span(%s) failed: %s", name, exc)
+        yield _NullSpan()
+
+
+# ── LangChain callback handler ──────────────────────────────────────────
+
+
+def get_langfuse_callback() -> Any | None:
+    """Return a LangChain ``CallbackHandler`` that auto-inherits the current trace.
+
+    Must be called inside a ``trace_span()`` block for proper linking.
+    Returns *None* when Langfuse is disabled.
+    """
+    if _disabled and not _initialised:
+        return None
+
+    try:
+        from langfuse.langchain import CallbackHandler
+        return CallbackHandler()
+    except Exception as exc:
+        logger.warning("tracing: get_langfuse_callback failed: %s", exc)
+        return None
+
+
+# ── Prompt management ────────────────────────────────────────────────────
+
+
+def get_prompt(
+    name: str,
+    *,
+    version: int | None = None,
+    label: str | None = None,
+    fallback: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> str | None:
+    """Fetch a managed prompt from Langfuse by name (without variable compilation).
+
+    Returns the raw prompt string, or *fallback* if the prompt is not
+    found or Langfuse is disabled.
+    """
+    lf = _get_client()
+    if lf is None:
+        return fallback
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        prompt = lf.get_prompt(**kwargs)
+        return prompt.prompt
+    except Exception as exc:
+        logger.warning("tracing: get_prompt(%s) failed: %s", name, exc)
+        return fallback
+
+
+def compile_prompt(
+    name: str,
+    *,
+    fallback: str,
+    variables: dict[str, str],
+    version: int | None = None,
+    label: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> str:
+    """Fetch a managed prompt from Langfuse and compile it with ``{{variables}}``.
+
+    If the prompt exists in Langfuse, uses the SDK's ``.compile(**variables)``
+    which replaces ``{{key}}`` placeholders.  If Langfuse is disabled or the
+    prompt is not found, falls back to ``fallback.format(**variables)`` (Python
+    ``{key}`` placeholders).
+
+    This means:
+      - Langfuse prompts use ``{{variable}}`` syntax.
+      - Hardcoded fallback strings use Python ``{variable}`` syntax.
+    """
+    lf = _get_client()
+    if lf is None:
+        return fallback.format(**variables)
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        prompt = lf.get_prompt(**kwargs)
+        return prompt.compile(**variables)
+    except Exception as exc:
+        logger.warning("tracing: compile_prompt(%s) failed, using fallback: %s", name, exc)
+        return fallback.format(**variables)
+
+
+def get_prompt_object(
+    name: str,
+    *,
+    version: int | None = None,
+    label: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> Any | None:
+    """Fetch the raw Langfuse prompt *object* (not the compiled string).
+
+    Returns ``None`` when Langfuse is disabled or the prompt is not found.
+    Use this when you need to pass the prompt to ``start_observation(prompt=...)``
+    for linking the prompt to a trace in the Langfuse UI.
+    """
+    lf = _get_client()
+    if lf is None:
+        return None
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        return lf.get_prompt(**kwargs)
+    except Exception as exc:
+        logger.warning("tracing: get_prompt_object(%s) failed: %s", name, exc)
+        return None
+
+
+def link_prompt_to_trace(
+    span: Any,
+    prompt_name: str,
+    *,
+    version: int | None = None,
+    label: str | None = None,
+) -> None:
+    """Link a Langfuse managed prompt to a span/observation.
+
+    Uses the SDK v4 ``prompt=`` parameter so that the prompt version
+    appears linked in the Langfuse UI with metrics tracking.
+    """
+    lf = _get_client()
+    if lf is None or isinstance(span, _NullSpan):
+        return
+
+    try:
+        prompt = get_prompt_object(prompt_name, version=version, label=label)
+        if prompt is not None:
+            span.update(prompt=prompt)
+    except Exception as exc:
+        logger.warning("tracing: link_prompt_to_trace(%s) failed: %s", prompt_name, exc)
+
+
+# ── Scoring helper ───────────────────────────────────────────────────────
+
+
+def score_trace(
+    trace_id: str,
+    name: str,
+    value: float,
+    *,
+    comment: str | None = None,
+) -> None:
+    """Post a score to a trace (e.g. user feedback, latency, quality)."""
+    lf = _get_client()
+    if lf is None:
+        return
+
+    try:
+        lf.create_score(trace_id=trace_id, name=name, value=value, comment=comment)
+    except Exception as exc:
+        logger.warning("tracing: score_trace failed: %s", exc)
+
+
+# ── Shutdown ─────────────────────────────────────────────────────────────
+
+
+def flush() -> None:
+    """Flush pending Langfuse events."""
+    lf = _get_client()
+    if lf is not None:
+        try:
+            lf.flush()
+        except Exception as exc:
+            logger.warning("tracing: flush failed: %s", exc)
+
+
+def shutdown() -> None:
+    """Flush and close the Langfuse client."""
+    global _initialised, _disabled
+    lf = _get_client()
+    if lf is not None:
+        try:
+            lf.flush()
+            lf.shutdown()
+        except Exception as exc:
+            logger.warning("tracing: shutdown failed: %s", exc)
+    _initialised = False
+    _disabled = False

services/batch-agent/eval/__init__.py

@@ -0,0 +1 @@
+"""Batch Agent E2E evaluation harness."""

services/batch-agent/eval/__main__.py

@@ -0,0 +1,5 @@
+"""Allow running the eval package as ``python -m eval``."""
+
+from eval.cli import main
+
+main()

services/batch-agent/eval/cli.py

@@ -0,0 +1,285 @@
+"""CLI entry point for the batch agent evaluation harness.
+
+Usage::
+
+    # From services/batch-agent/:
+    python -m eval run                                # all agent fixtures, default model
+    python -m eval run --fixture=classify-invoices    # single fixture
+    python -m eval run --models=gpt-4o,gpt-5.3-codex  # multiple models
+    python -m eval run --mode=step1                   # only step1 fixtures
+    python -m eval run --no-judge                     # skip LLM judge scoring
+
+    python -m eval interactive                        # interactive journey session
+    python -m eval interactive --fixture=journey-invoice-setup
+    python -m eval interactive --model=gpt-4o
+    python -m eval interactive --judge-model=github_copilot/gpt-4o-mini
+
+    python -m eval list                               # list all fixtures
+    python -m eval sync                               # sync fixtures to Langfuse datasets
+"""
+
+from __future__ import annotations
+
+import argparse
+import asyncio
+import logging
+import sys
+from pathlib import Path
+
+# Ensure the service root and repo root are in sys.path.
+# Service root must come BEFORE repo root so its ``app/`` package
+# shadows the monolith ``app/`` in the repo root.
+_SERVICE_ROOT = Path(__file__).resolve().parent.parent
+_REPO_ROOT = _SERVICE_ROOT.parent.parent
+_sr = str(_SERVICE_ROOT)
+_rr = str(_REPO_ROOT)
+if _rr not in sys.path:
+    sys.path.insert(0, _rr)
+# Always force service root to position 0 (python -m may have already
+# added CWD further down the list, which loses to repo root).
+if _sr in sys.path:
+    sys.path.remove(_sr)
+sys.path.insert(0, _sr)
+
+from eval.config import discover_fixtures, discover_journey_fixtures
+from eval.runner import run_fixture_eval, print_results
+from eval.interactive import run_interactive
+from eval import langfuse_eval
+
+
+def _setup_logging(verbose: bool) -> None:
+    level = logging.DEBUG if verbose else logging.INFO
+    logging.basicConfig(
+        level=level,
+        format="%(asctime)s %(name)-20s %(levelname)-5s %(message)s",
+        datefmt="%H:%M:%S",
+    )
+    # Quiet noisy libraries
+    for name in ("httpx", "httpcore", "openai", "litellm", "urllib3"):
+        logging.getLogger(name).setLevel(logging.WARNING)
+
+
+def _parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Batch Agent E2E evaluation harness",
+        prog="python -m eval",
+    )
+    sub = parser.add_subparsers(dest="command", required=True)
+
+    # ── run ───────────────────────────────────────────────────────
+    run_cmd = sub.add_parser("run", help="Run evaluations")
+    run_cmd.add_argument(
+        "--fixture", "-f",
+        help="Run only the named fixture (default: all)",
+    )
+    run_cmd.add_argument(
+        "--models", "-m",
+        default="github_copilot/gpt-5.3-codex",
+        help="Comma-separated list of models to test (default: github_copilot/gpt-5.3-codex)",
+    )
+    run_cmd.add_argument(
+        "--mode",
+        default=None,
+        choices=["step1", "step2", "full"],
+        help="Only run fixtures with this mode (default: all)",
+    )
+    run_cmd.add_argument(
+        "--no-judge",
+        action="store_true",
+        help="Skip LLM-as-judge scoring",
+    )
+    run_cmd.add_argument(
+        "--judge-model",
+        default="gpt-4o",
+        help="Model for LLM judge (default: gpt-4o)",
+    )
+    run_cmd.add_argument(
+        "--fixtures-dir",
+        default=None,
+        help="Path to fixtures directory (default: eval/fixtures/)",
+    )
+    run_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    # ── list ──────────────────────────────────────────────────────
+    list_cmd = sub.add_parser("list", help="List available fixtures")
+    list_cmd.add_argument("--fixtures-dir", default=None)
+    list_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    # ── sync ──────────────────────────────────────────────────────
+    sync_cmd = sub.add_parser("sync", help="Sync fixtures to Langfuse datasets")
+    sync_cmd.add_argument("--fixture", "-f", default=None, help="Sync only the named fixture")
+    sync_cmd.add_argument("--fixtures-dir", default=None)
+    sync_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    # ── interactive ───────────────────────────────────────────────
+    inter_cmd = sub.add_parser("interactive", help="Interactive journey session (human-in-the-loop)")
+    inter_cmd.add_argument(
+        "--fixture", "-f",
+        help="Journey fixture to use (default: pick interactively)",
+    )
+    inter_cmd.add_argument(
+        "--model", "-m",
+        default="github_copilot/gpt-5.3-codex",
+        help="Model for the journey AI (default: github_copilot/gpt-5.3-codex)",
+    )
+    inter_cmd.add_argument(
+        "--judge-model",
+        default="gpt-4o",
+        help="Model for LLM judge (default: gpt-4o)",
+    )
+    inter_cmd.add_argument(
+        "--fixtures-dir",
+        default=None,
+        help="Path to fixtures directory (default: eval/fixtures/)",
+    )
+    inter_cmd.add_argument(
+        "--data-dir",
+        default=None,
+        help="Override sample data directory (e.g. path to private test files not in git)",
+    )
+    inter_cmd.add_argument("-v", "--verbose", action="store_true")
+
+    return parser.parse_args()
+
+
+def _fixtures_dir(arg: str | None) -> Path | None:
+    if arg:
+        return Path(arg)
+    return None
+
+
+async def _cmd_run(args: argparse.Namespace) -> None:
+    fixtures = discover_fixtures(_fixtures_dir(args.fixtures_dir))
+    if not fixtures:
+        print("No fixtures found. Create YAML files in eval/fixtures/.")
+        return
+
+    if args.fixture:
+        fixtures = [f for f in fixtures if f.name == args.fixture]
+        if not fixtures:
+            print(f"Fixture '{args.fixture}' not found.")
+            return
+
+    models = [m.strip() for m in args.models.split(",")]
+
+    all_results = []
+    for fixture in fixtures:
+        if args.mode and fixture.mode != args.mode:
+            continue
+        results = await run_fixture_eval(
+            fixture,
+            models=models,
+            use_llm_judge=not args.no_judge,
+            judge_model=args.judge_model,
+        )
+        all_results.extend(results)
+
+    print_results(all_results)
+
+
+def _cmd_list(args: argparse.Namespace) -> None:
+    fixtures = discover_fixtures(_fixtures_dir(args.fixtures_dir))
+    journey_fixtures = discover_journey_fixtures(_fixtures_dir(args.fixtures_dir))
+
+    if not fixtures and not journey_fixtures:
+        print("No fixtures found.")
+        return
+
+    if fixtures:
+        print(f"\n{'[Agent Fixtures]'}")
+        print(f"{'Name':<30} {'Mode':<6} {'Types':<25} {'Expected'}")
+        print("-" * 90)
+        for f in fixtures:
+            types = ", ".join(f.data_types)
+            n_expected = len(f.expected) + len(f.expected_classification)
+            print(f"{f.name:<30} {f.mode:<6} {types:<25} {n_expected}")
+
+    if journey_fixtures:
+        print(f"\n{'[Journey Fixtures]'}")
+        print(f"{'Name':<30} {'Types':<25} {'Messages':<10} {'Criteria'}")
+        print("-" * 90)
+        for f in journey_fixtures:
+            types = ", ".join(f.data_types)
+            print(f"{f.name:<30} {types:<25} {len(f.user_messages):<10} {len(f.expected_template_criteria)}")
+
+    print()
+
+
+def _cmd_sync(args: argparse.Namespace) -> None:
+    fixtures = discover_fixtures(_fixtures_dir(args.fixtures_dir))
+    journey_fixtures = discover_journey_fixtures(_fixtures_dir(args.fixtures_dir))
+
+    if args.fixture:
+        fixtures = [f for f in fixtures if f.name == args.fixture]
+        journey_fixtures = [f for f in journey_fixtures if f.name == args.fixture]
+
+    if not fixtures and not journey_fixtures:
+        print("No fixtures to sync.")
+        return
+
+    for fixture in fixtures:
+        name = langfuse_eval.sync_fixture_to_dataset(fixture)
+        if name:
+            print(f"Synced: {fixture.name} → {name}")
+        else:
+            print(f"Skipped: {fixture.name} (Langfuse not configured)")
+
+    for fixture in journey_fixtures:
+        name = langfuse_eval.sync_journey_fixture_to_dataset(fixture)
+        if name:
+            print(f"Synced: {fixture.name} → {name}")
+        else:
+            print(f"Skipped: {fixture.name} (Langfuse not configured)")
+
+
+async def _cmd_interactive(args: argparse.Namespace) -> None:
+    journey_fixtures = discover_journey_fixtures(_fixtures_dir(args.fixtures_dir))
+    if not journey_fixtures:
+        print("No journey fixtures found. Create YAML files with type: journey in eval/fixtures/.")
+        return
+
+    if args.fixture:
+        fixtures = [f for f in journey_fixtures if f.name == args.fixture]
+        if not fixtures:
+            print(f"Journey fixture '{args.fixture}' not found.")
+            return
+        fixture = fixtures[0]
+    elif len(journey_fixtures) == 1:
+        fixture = journey_fixtures[0]
+    else:
+        # Let user pick
+        print("\nAvailable journey fixtures:")
+        for i, f in enumerate(journey_fixtures, 1):
+            print(f"  {i}. {f.name} — {f.description[:60]}")
+        print()
+        try:
+            choice = int(input("Pick a fixture number: ").strip()) - 1
+            fixture = journey_fixtures[choice]
+        except (ValueError, IndexError, EOFError, KeyboardInterrupt):
+            print("Invalid choice.")
+            return
+
+    await run_interactive(
+        fixture,
+        model=args.model,
+        judge_model=args.judge_model,
+        data_dir=Path(args.data_dir).resolve() if args.data_dir else None,
+    )
+
+
+def main() -> None:
+    args = _parse_args()
+    _setup_logging(args.verbose)
+
+    if args.command == "run":
+        asyncio.run(_cmd_run(args))
+    elif args.command == "interactive":
+        asyncio.run(_cmd_interactive(args))
+    elif args.command == "list":
+        _cmd_list(args)
+    elif args.command == "sync":
+        _cmd_sync(args)
+
+
+if __name__ == "__main__":
+    main()

services/batch-agent/eval/config.py

@@ -0,0 +1,219 @@
+"""Eval configuration — YAML fixture loader and dataclasses.
+
+Fixtures come in two families:
+
+1. **Agent fixtures** — test the batch agent pipeline.
+   Three modes controlled by ``mode``:
+
+   ``step1``  — classification prompt only.
+   ``step2``  — processing prompt only.
+   ``full``   — both steps in sequence.
+
+2. **Journey fixtures** — test the prompt-template builder conversation
+   (unchanged).
+"""
+
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any, Literal
+
+import yaml
+
+logger = logging.getLogger(__name__)
+
+EvalMode = Literal["step1", "step2", "full"]
+
+
+@dataclass
+class ExpectedRecord:
+    """A single expected extraction result.
+
+    Only the fields specified are checked — unspecified fields are ignored.
+    """
+
+    table: str  # tasks | notes | timelines | projects
+    fields: dict[str, Any]  # field_name → expected_value
+
+
+@dataclass
+class ExpectedClassification:
+    """Expected output of step-1 classification for one file."""
+
+    file: str  # relative path to the sample file
+    project_id: str  # expected matched project id, or "new"
+    domains: list[str]  # expected domain list
+    new_project_name: str | None = None
+
+
+@dataclass
+class EvalFixture:
+    """A complete test scenario loaded from YAML.
+
+    ``mode`` determines which pipeline steps are exercised:
+
+    - **step1**: only ``_classify_file``
+    - **step2**: only the processing LLM + tool loop
+    - **full**: both steps in sequence (``run_local_agent``)
+    """
+
+    name: str
+    description: str
+    mode: EvalMode
+    directory: str  # relative path to sample files
+    data_types: list[str]
+    file_extensions: list[str]
+    models: list[str]  # if empty, use CLI default
+    fixture_path: Path = field(default_factory=lambda: Path("."))
+
+    # ── Step-1 inputs (classification) ───────────────────────────
+    domain_definitions: str = ""
+    projects_list: list[dict[str, Any]] = field(default_factory=list)
+
+    # ── Step-2 inputs (processing) ───────────────────────────────
+    existing_context: str = ""
+    project_context: str = ""
+    custom_prompt_section: str = ""
+
+    # ── Seed records for mock executor ───────────────────────────
+    seed_records: dict[str, list[dict]] = field(default_factory=dict)
+
+    # ── Expected outputs ─────────────────────────────────────────
+    expected_classification: list[ExpectedClassification] = field(default_factory=list)
+    expected: list[ExpectedRecord] = field(default_factory=list)
+
+    @property
+    def fixture_dir(self) -> Path:
+        """Absolute path to the sample files directory."""
+        return self.fixture_path.parent / self.directory
+
+    @classmethod
+    def from_yaml(cls, path: Path) -> "EvalFixture":
+        """Load a fixture from a YAML file."""
+        raw = yaml.safe_load(path.read_text(encoding="utf-8"))
+
+        mode: EvalMode = raw.get("mode", "full")
+
+        # Parse expected records (step2/full)
+        expected: list[ExpectedRecord] = []
+        for table, records in (raw.get("expected") or {}).items():
+            for rec in records:
+                expected.append(ExpectedRecord(table=table, fields=rec))
+
+        # Parse expected classification (step1/full)
+        expected_classification: list[ExpectedClassification] = []
+        for item in raw.get("expected_classification") or []:
+            expected_classification.append(ExpectedClassification(
+                file=item["file"],
+                project_id=item["project_id"],
+                domains=item.get("domains", []),
+                new_project_name=item.get("new_project_name"),
+            ))
+
+        return cls(
+            name=raw["name"],
+            description=raw.get("description", ""),
+            mode=mode,
+            directory=raw.get("directory", "sample_files"),
+            data_types=raw.get("data_types", ["tasks"]),
+            file_extensions=raw.get("file_extensions", []),
+            models=raw.get("models", []),
+            fixture_path=path,
+            # Step-1 inputs
+            domain_definitions=raw.get("domain_definitions", ""),
+            projects_list=raw.get("projects_list", []),
+            # Step-2 inputs
+            existing_context=raw.get("existing_context", ""),
+            project_context=raw.get("project_context", ""),
+            custom_prompt_section=raw.get("custom_prompt_section", ""),
+            # Shared
+            seed_records=raw.get("seed_records", {}),
+            expected_classification=expected_classification,
+            expected=expected,
+        )
+
+
+def discover_fixtures(fixtures_dir: Path | None = None) -> list[EvalFixture]:
+    """Find and load all YAML fixtures in the fixtures directory."""
+    if fixtures_dir is None:
+        fixtures_dir = Path(__file__).parent / "fixtures"
+
+    fixtures: list[EvalFixture] = []
+    if not fixtures_dir.is_dir():
+        logger.warning("eval: fixtures directory not found: %s", fixtures_dir)
+        return fixtures
+
+    for yaml_path in sorted(fixtures_dir.glob("*.yaml")):
+        try:
+            raw = yaml.safe_load(yaml_path.read_text(encoding="utf-8"))
+            if raw.get("type") == "journey":
+                continue  # Skip journey fixtures
+            fixtures.append(EvalFixture.from_yaml(yaml_path))
+            logger.info("eval: loaded fixture %s from %s", fixtures[-1].name, yaml_path.name)
+        except Exception as exc:
+            logger.error("eval: failed to load fixture %s: %s", yaml_path.name, exc)
+
+    return fixtures
+
+
+# ── Journey fixtures ─────────────────────────────────────────────────────
+
+
+@dataclass
+class JourneyFixture:
+    """A journey test scenario — tests the prompt_template builder conversation."""
+
+    name: str
+    description: str
+    directory: str  # relative path to sample files
+    data_types: list[str]
+    expected_template_criteria: list[str]  # what the template should contain/satisfy
+    user_messages: list[str] = field(default_factory=list)  # for automated journey runs (unused in interactive mode)
+    models: list[str] = field(default_factory=list)
+    fixture_path: Path = field(default_factory=lambda: Path("."))
+
+    @property
+    def fixture_dir(self) -> Path:
+        """Absolute path to the sample files directory."""
+        return self.fixture_path.parent / self.directory
+
+    @classmethod
+    def from_yaml(cls, path: Path) -> "JourneyFixture":
+        """Load a journey fixture from a YAML file."""
+        raw = yaml.safe_load(path.read_text(encoding="utf-8"))
+
+        return cls(
+            name=raw["name"],
+            description=raw.get("description", ""),
+            directory=raw.get("directory", "sample_files"),
+            data_types=raw.get("data_types", ["tasks"]),
+            user_messages=raw.get("user_messages", []),
+            expected_template_criteria=raw.get("expected_template_criteria", []),
+            models=raw.get("models", []),
+            fixture_path=path,
+        )
+
+
+def discover_journey_fixtures(fixtures_dir: Path | None = None) -> list[JourneyFixture]:
+    """Find and load all journey YAML fixtures in the fixtures directory."""
+    if fixtures_dir is None:
+        fixtures_dir = Path(__file__).parent / "fixtures"
+
+    fixtures: list[JourneyFixture] = []
+    if not fixtures_dir.is_dir():
+        logger.warning("eval: fixtures directory not found: %s", fixtures_dir)
+        return fixtures
+
+    for yaml_path in sorted(fixtures_dir.glob("*.yaml")):
+        try:
+            raw = yaml.safe_load(yaml_path.read_text(encoding="utf-8"))
+            if raw.get("type") != "journey":
+                continue
+            fixtures.append(JourneyFixture.from_yaml(yaml_path))
+            logger.info("eval: loaded journey fixture %s from %s", fixtures[-1].name, yaml_path.name)
+        except Exception as exc:
+            logger.error("eval: failed to load journey fixture %s: %s", yaml_path.name, exc)
+
+    return fixtures

services/batch-agent/eval/fixtures/classify_invoices.yaml

@@ -0,0 +1,40 @@
+# Fixture: classify-invoices (step1)
+# Tests _STEP1_SYSTEM_PROMPT — file classification and project matching.
+# Verifies that the LLM correctly matches files to existing projects
+# and identifies the right data domains.
+
+name: classify-invoices
+mode: step1
+description: >
+  Test file classification on Italian freelance invoices and meeting notes.
+  Verifies project matching and domain identification.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines]
+file_extensions: [txt, md]
+
+# ── Step-1 prompt variables ──────────────────────────────────────
+domain_definitions: |
+  - tasks: Action items, deliverables, things to do — anything that someone needs to complete.
+  - notes: Meeting summaries, decisions, reference information — permanent knowledge entries.
+  - timelines: Project milestones, deadlines, scheduled events — specific dates that mark a point in the progress of a project.
+
+projects_list:
+  - id: "proj-web-redesign"
+    name: "Redesign Sito Web Corporate"
+    status: "active"
+    aiSummary: "Corporate website redesign for Studio Architettura Bianchi"
+  - id: "proj-ecommerce"
+    name: "E-Commerce FashionStore"
+    status: "active"
+    aiSummary: "Next.js e-commerce platform for FashionStore srl"
+
+# ── Expected classification results ─────────────────────────────
+expected_classification:
+  - file: "sample_files/invoices/fattura_042.txt"
+    project_id: "proj-web-redesign"
+    domains: [tasks, notes, timelines]
+
+  - file: "sample_files/invoices/meeting_ecommerce.md"
+    project_id: "proj-ecommerce"
+    domains: [tasks, notes, timelines]

services/batch-agent/eval/fixtures/full_invoices.yaml

@@ -0,0 +1,108 @@
+# Fixture: full-invoices (full)
+# Tests both _STEP1_SYSTEM_PROMPT and _PROCESSING_SYSTEM_PROMPT in sequence
+# via run_local_agent(). Verifies end-to-end classification + extraction.
+
+name: full-invoices
+mode: full
+description: >
+  End-to-end test: classify Italian invoices/meeting notes into the
+  correct project, then extract tasks, notes, and timeline events.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines]
+file_extensions: [txt, md]
+
+# ── Step-1 prompt variables ──────────────────────────────────────
+domain_definitions: |
+  - tasks: Action items, deliverables, things to do — anything that someone needs to complete.
+  - notes: Meeting summaries, decisions, reference information — permanent knowledge entries.
+  - timelines: Project milestones, deadlines, scheduled events — specific dates that mark a point in the progress of a project.
+
+projects_list:
+  - id: "proj-web-redesign"
+    name: "Redesign Sito Web Corporate"
+    status: "active"
+    aiSummary: "Corporate website redesign for Studio Architettura Bianchi"
+  - id: "proj-ecommerce"
+    name: "E-Commerce FashionStore"
+    status: "active"
+    aiSummary: "Next.js e-commerce platform for FashionStore srl"
+
+# ── Step-2 prompt variables ──────────────────────────────────────
+existing_context: |
+  Existing tasks:
+    (none)
+
+  Existing notes:
+    (none)
+
+  Existing timelines:
+    (none)
+
+project_context: ""
+
+custom_prompt_section: |
+  User instructions:
+  Estrai i dati dai file come segue:
+  - TASK: ogni azione da fare, deliverable, o item con scadenza.
+    Mappa "URGENTE" o "ALTA PRIORITÀ" → priority: high.
+    Mappa "media priorità" → priority: medium.
+    Mappa "bassa priorità" → priority: low.
+    Se un item è marcato come "completato" o [x], impostalo status: done.
+    Altrimenti status: todo.
+  - NOTE: riassunti di meeting, decisioni prese, note tecniche.
+  - TIMELINE: date di scadenza, milestone, meeting futuri.
+  Imposta sempre isAiSuggested=1.
+
+# ── Seed records (pre-existing DB state) ─────────────────────────
+seed_records:
+  projects:
+    - id: "proj-web-redesign"
+      name: "Redesign Sito Web Corporate"
+      status: "active"
+      aiSummary: "Corporate website redesign for Studio Architettura Bianchi"
+    - id: "proj-ecommerce"
+      name: "E-Commerce FashionStore"
+      status: "active"
+      aiSummary: "Next.js e-commerce platform for FashionStore srl"
+  tasks: []
+  notes: []
+  timelines: []
+
+# ── Expected classification (step 1) ─────────────────────────────
+expected_classification:
+  - file: "sample_files/invoices/fattura_042.txt"
+    project_id: "proj-web-redesign"
+    domains: [tasks, notes, timelines]
+
+  - file: "sample_files/invoices/meeting_ecommerce.md"
+    project_id: "proj-ecommerce"
+    domains: [tasks, notes, timelines]
+
+# ── Expected extractions (step 2) ────────────────────────────────
+expected:
+  tasks:
+    - title: "Sviluppo frontend React"
+      priority: "high"
+      status: "todo"
+    - title: "Integrazione API backend"
+      priority: "medium"
+      status: "todo"
+    - title: "Testing cross-browser e fix bug responsive"
+      status: "todo"
+    - title: "Preparare wireframe homepage"
+      priority: "high"
+      status: "todo"
+    - title: "Setup progetto Next.js e configurare CI/CD"
+      priority: "medium"
+      status: "todo"
+    - title: "Ricerca plugin Stripe per gestione abbonamenti"
+      priority: "low"
+      status: "todo"
+
+  notes:
+    - title: "Meeting Kickoff Progetto E-Commerce"
+
+  timelines:
+    - title: "MVP E-Commerce pronto"
+    - title: "Meeting di revisione"

services/batch-agent/eval/fixtures/journey_invoice_setup.yaml

@@ -0,0 +1,28 @@
+# Journey Fixture: journey-invoice-setup
+# Used by `python -m eval interactive` for human-in-the-loop testing
+# of the journey chatbot's prompt-building conversation.
+
+type: journey
+name: journey-invoice-setup
+description: >
+  Interactive test for the journey chatbot — explore a directory of
+  Italian invoices and meeting notes, answer the chatbot's questions,
+  and verify it produces a well-structured prompt_template for data
+  extraction.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines, projects]
+
+# Criteria the generated prompt_template must satisfy
+# Each is scored 0-1 by an LLM judge
+expected_template_criteria:
+  - "Mentions creating tasks from action items and work descriptions"
+  - "Mentions creating notes from meeting summaries"
+  - "Mentions extracting timeline events from deadlines and meeting dates"
+  - "Mentions creating projects from relevant information"
+  - "Sets isAiSuggested=1 on all created records"
+  - "Does NOT include projectId assignment logic"
+  - "Uses camelCase field names (title, status, priority, dueDate, content)"
+
+# Models to test (empty = use CLI --models default)
+models: []

services/batch-agent/eval/fixtures/process_invoices.yaml

@@ -0,0 +1,81 @@
+# Fixture: process-invoices (step2)
+# Tests _PROCESSING_SYSTEM_PROMPT — data extraction & tool calling.
+# The classification step is skipped; prompt variables are injected directly.
+
+name: process-invoices
+mode: step2
+description: >
+  Test data extraction from Italian freelance invoices.
+  Verifies correct record creation via tool calls with the right
+  fields, priorities, and status values.
+
+directory: sample_files/invoices
+data_types: [tasks, notes, timelines]
+file_extensions: [txt, md]
+
+# ── Step-2 prompt variables ──────────────────────────────────────
+existing_context: |
+  Existing tasks:
+    (none)
+
+  Existing notes:
+    (none)
+
+  Existing timelines:
+    (none)
+
+project_context: >
+  Project: Redesign Sito Web Corporate (id: proj-web-redesign).
+  Always set projectId to this id on every record you create.
+
+custom_prompt_section: |
+  User instructions:
+  Estrai i dati dai file come segue:
+  - TASK: ogni azione da fare, deliverable, o item con scadenza.
+    Mappa "URGENTE" o "ALTA PRIORITÀ" → priority: high.
+    Mappa "media priorità" → priority: medium.
+    Mappa "bassa priorità" → priority: low.
+    Se un item è marcato come "completato" o [x], impostalo status: done.
+    Altrimenti status: todo.
+  - NOTE: riassunti di meeting, decisioni prese, note tecniche.
+    Il titolo deve essere descrittivo. Il content deve includere tutti i dettagli.
+  - TIMELINE: date di scadenza, milestone, meeting futuri.
+  Imposta sempre isAiSuggested=1.
+
+# ── Seed records (pre-existing DB state) ─────────────────────────
+seed_records:
+  projects:
+    - id: "proj-web-redesign"
+      name: "Redesign Sito Web Corporate"
+      status: "active"
+  tasks: []
+  notes: []
+  timelines: []
+
+# ── Expected extractions ─────────────────────────────────────────
+expected:
+  tasks:
+    - title: "Sviluppo frontend React"
+      priority: "high"
+      status: "todo"
+    - title: "Integrazione API backend"
+      priority: "medium"
+      status: "todo"
+    - title: "Testing cross-browser e fix bug responsive"
+      status: "todo"
+    - title: "Preparare wireframe homepage"
+      priority: "high"
+      status: "todo"
+    - title: "Setup progetto Next.js e configurare CI/CD"
+      priority: "medium"
+      status: "todo"
+    - title: "Ricerca plugin Stripe per gestione abbonamenti"
+      priority: "low"
+      status: "todo"
+
+  notes:
+    - title: "Meeting Kickoff Progetto E-Commerce"
+
+  timelines:
+    - title: "MVP E-Commerce pronto"
+    - title: "Meeting di revisione"

services/batch-agent/eval/fixtures/sample_files/invoices/fattura_042.txt

@@ -0,0 +1,18 @@
+FATTURA N. 2026-0042
+Data: 15 Marzo 2026
+Cliente: Studio Architettura Bianchi
+
+Progetto: Redesign Sito Web Corporate
+
+Descrizione lavori:
+- Sviluppo frontend React (40 ore) — URGENTE, completare entro 20 marzo
+- Integrazione API backend (20 ore) — priorità media
+- Design UI/UX mockup homepage (8 ore) — completato
+- Testing cross-browser e fix bug responsive (12 ore) — da iniziare
+
+Totale: €4.800,00 + IVA
+
+Note:
+Meeting di revisione previsto per il 18 marzo alle 10:00.
+Il cliente ha richiesto modifiche al layout mobile della sezione contatti.
+Attendere conferma budget aggiuntivo per sezione blog.

services/batch-agent/eval/fixtures/sample_files/invoices/meeting_ecommerce.md

@@ -0,0 +1,25 @@
+# Meeting Notes - Kickoff Progetto E-Commerce
+
+**Data:** 10 Marzo 2026
+**Partecipanti:** Marco R., Giulia T., Cliente (FashionStore srl)
+
+## Decisioni prese
+
+1. **Piattaforma**: Next.js + Stripe per i pagamenti
+2. **Timeline**: MVP pronto entro 30 aprile 2026
+3. **Budget**: €12.000 totale, €4.000 anticipo già ricevuto
+
+## Action items
+
+- [ ] Marco: preparare wireframe homepage entro 14 marzo — ALTA PRIORITÀ
+- [ ] Giulia: setup progetto Next.js e configurare CI/CD — media priorità
+- [ ] Marco: ricerca plugin Stripe per gestione abbonamenti — bassa priorità
+- [x] Giulia: inviare contratto firmato al cliente — COMPLETATO
+
+## Note aggiuntive
+
+Il cliente vuole un design minimalista, ispirato a Zara.com.
+Colori primari: nero, bianco, oro.
+Font: Inter per body, Playfair Display per headings.
+
+Prossimo meeting: 24 marzo 2026 ore 15:00.

services/batch-agent/eval/interactive.py

@@ -0,0 +1,471 @@
+"""Interactive journey session — human-in-the-loop CLI conversation.
+
+Flow:
+1. Show the system prompt used by the journey AI.
+2. Start the journey (AI explores files, asks first question).
+3. User types responses in the terminal — AI replies.
+4. User types `/done` to end the conversation.
+5. User writes a comment about the interaction quality.
+6. LLM judge scores the conversation + generated template.
+7. Results are reported to Langfuse.
+
+Usage::
+
+    python -m eval interactive                        # pick a fixture interactively
+    python -m eval interactive --fixture=journey-invoice-setup
+    python -m eval interactive --model=gpt-4o
+    python -m eval interactive --judge-model=github_copilot/gpt-4o-mini
+"""
+
+from __future__ import annotations
+
+import asyncio
+import json
+import logging
+import sys
+import time
+import uuid
+from dataclasses import dataclass, field
+from typing import Any
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+from eval.config import JourneyFixture, discover_journey_fixtures
+from eval.mock_executor import MockExecutor
+from eval import langfuse_eval
+
+logger = logging.getLogger(__name__)
+
+# ── Special commands ─────────────────────────────────────────────────────
+
+_CMD_DONE = "/done"
+_CMD_QUIT = "/quit"
+_CMD_TEMPLATE = "/template"
+_CMD_HELP = "/help"
+
+_HELP_TEXT = f"""\
+  {_CMD_DONE}       — End the conversation and proceed to evaluation
+  {_CMD_QUIT}       — Abort without evaluation
+  {_CMD_TEMPLATE}   — Show the generated template (if any)
+  {_CMD_HELP}       — Show this help"""
+
+# ── Terminal colours (ANSI) ──────────────────────────────────────────────
+
+_C_RESET = "\033[0m"
+_C_BOLD = "\033[1m"
+_C_DIM = "\033[2m"
+_C_CYAN = "\033[36m"
+_C_GREEN = "\033[32m"
+_C_YELLOW = "\033[33m"
+_C_MAGENTA = "\033[35m"
+_C_RED = "\033[31m"
+_C_BLUE = "\033[34m"
+
+
+def _print_header(text: str) -> None:
+    print(f"\n{_C_BOLD}{_C_CYAN}{'═' * 80}")
+    print(f"  {text}")
+    print(f"{'═' * 80}{_C_RESET}\n")
+
+
+def _print_ai(text: str) -> None:
+    print(f"\n{_C_GREEN}{_C_BOLD}AI:{_C_RESET} {text}\n")
+
+
+def _print_system(text: str) -> None:
+    print(f"{_C_DIM}{text}{_C_RESET}")
+
+
+def _print_score(label: str, score: float) -> None:
+    if score >= 0.7:
+        color = _C_GREEN
+        tag = "PASS"
+    elif score >= 0.4:
+        color = _C_YELLOW
+        tag = "PARTIAL"
+    else:
+        color = _C_RED
+        tag = "FAIL"
+    print(f"  {color}{tag:>7}{_C_RESET} ({score:.1f}) {label}")
+
+
+# ── Result type ──────────────────────────────────────────────────────────
+
+
+@dataclass
+class InteractiveResult:
+    fixture_name: str
+    model: str
+    judge_model: str
+    prompt_template: str | None
+    conversation: list[dict[str, str]]
+    user_comment: str
+    done: bool
+    criteria_scores: dict[str, float]
+    overall_score: float
+    judge_reasoning: str
+    elapsed_seconds: float
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "fixture": self.fixture_name,
+            "model": self.model,
+            "judge_model": self.judge_model,
+            "done": self.done,
+            "turns": len([c for c in self.conversation if c["role"] == "user"]),
+            "overall_score": round(self.overall_score, 3),
+            "user_comment": self.user_comment,
+            "criteria_scores": {k: round(v, 3) for k, v in self.criteria_scores.items()},
+            "elapsed_s": round(self.elapsed_seconds, 1),
+        }
+
+
+# ── LLM judge ────────────────────────────────────────────────────────────
+
+_INTERACTIVE_JUDGE_SYSTEM = """\
+You are an evaluation judge for AI-generated prompt templates produced during
+an interactive conversation between a human and a journey chatbot.
+
+The chatbot explored a directory and through multi-turn conversation with the
+user produced a prompt_template — an instruction set for a data-extraction agent.
+
+You have access to:
+- The full conversation transcript
+- The generated prompt_template (if any)
+- The user's own comment about the interaction
+- A list of quality criteria
+
+Score each criterion from 0 to 1:
+  - 1.0: Fully satisfied
+  - 0.5: Partially satisfied
+  - 0.0: Not satisfied
+
+Also provide an overall_quality score (0-1) evaluating the conversation flow,
+how well the AI understood the user, and the template quality.
+
+Respond with ONLY a JSON object:
+{
+  "criteria_scores": {"criterion_1": 0.8, ...},
+  "overall_quality": 0.85,
+  "reasoning": "Brief explanation covering both conversation quality and template accuracy"
+}
+"""
+
+
+async def _judge_interactive(
+    conversation: list[dict[str, str]],
+    prompt_template: str | None,
+    user_comment: str,
+    criteria: list[str],
+    *,
+    judge_model: str = "gpt-4o-mini",
+) -> tuple[dict[str, float], float, str]:
+    """Score an interactive session. Returns (criteria_scores, overall_quality, reasoning)."""
+    from shared.llm import get_llm
+
+    llm = get_llm(model=judge_model, temperature=0)
+
+    conv_text = "\n".join(
+        f"{'USER' if t['role'] == 'user' else 'AI'}: {t['content']}"
+        for t in conversation
+    )
+    criteria_text = "\n".join(f"  {i+1}. {c}" for i, c in enumerate(criteria))
+
+    user_content = (
+        f"## Conversation transcript\n```\n{conv_text}\n```\n\n"
+        f"## Generated prompt_template\n```\n{prompt_template or '(none — conversation did not complete)'}\n```\n\n"
+        f"## User's comment\n{user_comment}\n\n"
+        f"## Criteria to evaluate\n{criteria_text}"
+    )
+
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=_INTERACTIVE_JUDGE_SYSTEM),
+            HumanMessage(content=user_content),
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+
+        scores_raw = parsed.get("criteria_scores", parsed.get("scores", {}))
+        criteria_scores: dict[str, float] = {}
+        for i, criterion in enumerate(criteria):
+            key_candidates = [f"criterion_{i+1}", criterion, criterion[:50], str(i + 1)]
+            score = 0.0
+            for key in key_candidates:
+                if key in scores_raw:
+                    score = float(scores_raw[key])
+                    break
+            if score == 0.0 and i < len(scores_raw):
+                score = float(list(scores_raw.values())[i])
+            criteria_scores[criterion] = score
+
+        overall = float(parsed.get("overall_quality", 0.0))
+        reasoning = str(parsed.get("reasoning", ""))
+        return criteria_scores, overall, reasoning
+
+    except Exception as exc:
+        logger.warning("interactive judge failed: %s", exc)
+        return {c: 0.0 for c in criteria}, 0.0, f"Judge error: {exc}"
+
+
+# ── Interactive session ──────────────────────────────────────────────────
+
+
+async def run_interactive(
+    fixture: JourneyFixture,
+    *,
+    model: str = "gpt-4o",
+    judge_model: str = "gpt-4o-mini",
+    data_dir: Path | None = None,
+) -> InteractiveResult:
+    """Run an interactive journey session in the terminal.
+
+    Parameters
+    ----------
+    data_dir :
+        If set, overrides the fixture's sample-file directory.  The LLM
+        will explore this folder instead of the default
+        ``fixtures/sample_files/…``.  Useful for private test data that
+        shouldn't be committed to git.
+    """
+    from shared.config import settings
+    from shared.ws_context import set_current_user, clear_current_user
+    from app.journey import (
+        handle_journey_start,
+        handle_journey_message,
+        _build_system_prompt,
+    )
+
+    # When --data-dir is given, the MockExecutor's root becomes
+    # data_dir's parent and the journey directory is data_dir's name.
+    # This way the LLM sees a meaningful directory name (not ".") and
+    # MockExecutor resolves paths correctly.
+    # Otherwise, use the fixture's YAML parent and its relative path.
+    if data_dir:
+        mock_root = data_dir.parent
+        journey_directory = data_dir.name
+    else:
+        mock_root = fixture.fixture_path.parent
+        journey_directory = fixture.directory
+
+    mock = MockExecutor(
+        fixture_dir=mock_root,
+        seed_records={},
+    )
+
+    original_model = settings.LLM_MODEL
+    settings.LLM_MODEL = model
+    eval_user_id = f"interactive-{uuid.uuid4().hex[:8]}"
+
+    # ── Show system prompt ───────────────────────────────────────
+    system_prompt = _build_system_prompt(journey_directory, fixture.data_types)
+
+    _print_header("SYSTEM PROMPT")
+    print(f"{_C_DIM}{system_prompt}{_C_RESET}")
+
+    _print_header(f"INTERACTIVE JOURNEY  |  fixture: {fixture.name}  |  model: {model}")
+    print(f"  Data dir: {mock_root}")
+    print(f"  Type your responses. Commands: {_CMD_DONE}, {_CMD_QUIT}, {_CMD_TEMPLATE}, {_CMD_HELP}")
+    print(f"  Judge model: {judge_model}")
+    print(f"  Criteria: {len(fixture.expected_template_criteria)}")
+    print()
+
+    conversation: list[dict[str, str]] = []
+    prompt_template: str | None = None
+    done = False
+    start_time = time.time()
+
+    try:
+        set_current_user(eval_user_id)
+
+        with mock.patch():
+            # ── Start ────────────────────────────────────────────
+            _print_system("Starting journey... (AI is exploring your files)")
+
+            start_frame: dict[str, Any] = {
+                "agent_type": "local",
+                "directory": journey_directory,
+                "data_types": fixture.data_types,
+                "session_id": f"interactive-{uuid.uuid4().hex[:8]}",
+            }
+
+            reply = await handle_journey_start(eval_user_id, start_frame)
+            session_id = reply["session_id"]
+            conversation.append({"role": "assistant", "content": reply["message"]})
+            _print_ai(reply["message"])
+
+            if reply["done"]:
+                prompt_template = reply.get("prompt_template")
+                done = True
+                _print_system("Journey completed on first reply (template generated).")
+
+            # ── Conversation loop ────────────────────────────────
+            while not done:
+                try:
+                    user_input = input(f"{_C_BOLD}{_C_BLUE}YOU:{_C_RESET} ").strip()
+                except (EOFError, KeyboardInterrupt):
+                    print()
+                    user_input = _CMD_QUIT
+
+                if not user_input:
+                    continue
+
+                # Handle commands
+                if user_input.lower() == _CMD_QUIT:
+                    _print_system("Aborted — no evaluation will be performed.")
+                    settings.LLM_MODEL = original_model
+                    clear_current_user()
+                    return InteractiveResult(
+                        fixture_name=fixture.name, model=model, judge_model=judge_model,
+                        prompt_template=None, conversation=conversation,
+                        user_comment="(aborted)", done=False,
+                        criteria_scores={}, overall_score=0.0,
+                        judge_reasoning="Session aborted by user.",
+                        elapsed_seconds=time.time() - start_time,
+                    )
+
+                if user_input.lower() == _CMD_HELP:
+                    print(_HELP_TEXT)
+                    continue
+
+                if user_input.lower() == _CMD_TEMPLATE:
+                    if prompt_template:
+                        print(f"\n{_C_MAGENTA}{prompt_template}{_C_RESET}\n")
+                    else:
+                        _print_system("No template generated yet.")
+                    continue
+
+                if user_input.lower() == _CMD_DONE:
+                    _print_system("Ending conversation...")
+                    break
+
+                # ── Send message to AI ───────────────────────────
+                conversation.append({"role": "user", "content": user_input})
+                _print_system("AI is thinking...")
+
+                msg_frame: dict[str, Any] = {
+                    "session_id": session_id,
+                    "message": user_input,
+                }
+                reply = await handle_journey_message(eval_user_id, msg_frame)
+                conversation.append({"role": "assistant", "content": reply["message"]})
+                _print_ai(reply["message"])
+
+                if reply["done"]:
+                    prompt_template = reply.get("prompt_template")
+                    done = True
+                    _print_system("Journey completed — template generated!")
+
+    except Exception as exc:
+        logger.error("interactive journey failed: %s", exc)
+        _print_system(f"Error: {exc}")
+    finally:
+        settings.LLM_MODEL = original_model
+        clear_current_user()
+
+    elapsed = time.time() - start_time
+    turns = len([c for c in conversation if c["role"] == "user"])
+
+    # ── Show template if generated ───────────────────────────────
+    if prompt_template:
+        _print_header("GENERATED TEMPLATE")
+        print(f"{_C_MAGENTA}{prompt_template}{_C_RESET}\n")
+    else:
+        _print_system("No template was generated during this session.")
+
+    # ── User comment ─────────────────────────────────────────────
+    _print_header("YOUR EVALUATION")
+    print("  Write your comment about this interaction (press Enter twice to finish):")
+    print()
+    comment_lines: list[str] = []
+    try:
+        while True:
+            line = input()
+            if line == "" and comment_lines and comment_lines[-1] == "":
+                comment_lines.pop()  # remove trailing empty
+                break
+            comment_lines.append(line)
+    except (EOFError, KeyboardInterrupt):
+        pass
+    user_comment = "\n".join(comment_lines).strip() or "(no comment)"
+
+    # ── Judge ────────────────────────────────────────────────────
+    _print_header("LLM JUDGE EVALUATION")
+    _print_system(f"Scoring with {judge_model}...")
+
+    criteria_scores, overall_quality, judge_reasoning = await _judge_interactive(
+        conversation=conversation,
+        prompt_template=prompt_template,
+        user_comment=user_comment,
+        criteria=fixture.expected_template_criteria,
+        judge_model=judge_model,
+    )
+
+    # ── Display scores ───────────────────────────────────────────
+    print()
+    for criterion, score in criteria_scores.items():
+        _print_score(criterion, score)
+
+    overall = (
+        sum(criteria_scores.values()) / len(criteria_scores)
+        if criteria_scores
+        else 0.0
+    )
+
+    print(f"\n  {_C_BOLD}Criteria avg:      {overall:.2f}{_C_RESET}")
+    print(f"  {_C_BOLD}Overall quality:   {overall_quality:.2f}{_C_RESET}")
+    print(f"  {_C_BOLD}Turns:             {turns}{_C_RESET}")
+    print(f"  {_C_BOLD}Time:              {elapsed:.1f}s{_C_RESET}")
+    print(f"\n  {_C_DIM}Judge: {judge_reasoning}{_C_RESET}")
+    print(f"  {_C_DIM}Your comment: {user_comment}{_C_RESET}\n")
+
+    result = InteractiveResult(
+        fixture_name=fixture.name,
+        model=model,
+        judge_model=judge_model,
+        prompt_template=prompt_template,
+        conversation=conversation,
+        user_comment=user_comment,
+        done=done,
+        criteria_scores=criteria_scores,
+        overall_score=overall_quality,
+        judge_reasoning=judge_reasoning,
+        elapsed_seconds=elapsed,
+    )
+
+    # ── Report to Langfuse ───────────────────────────────────────
+    trace_id = langfuse_eval.log_eval_trace(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_variant="interactive",
+        prompt_template=prompt_template or "(not generated)",
+        actual_mutations=[{
+            "conversation": conversation[:30],
+            "user_comment": user_comment,
+        }],
+        scores_summary=result.summary(),
+        langfuse_prompt_names=["journey_system"],
+    )
+
+    if trace_id:
+        from eval.scorer import EvalScores
+        scores_obj = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant="interactive",
+            precision=overall,
+            recall=float(done),
+            f1=overall,
+            llm_judge_score=overall_quality,
+            llm_judge_reasoning=judge_reasoning,
+        )
+        langfuse_eval.post_eval_scores(scores_obj, trace_id=trace_id)
+        _print_system(f"Results reported to Langfuse (trace: {trace_id})")
+    else:
+        _print_system("Langfuse not configured — results not reported.")
+
+    return result

services/batch-agent/eval/journey_runner.py

@@ -0,0 +1,385 @@
+"""Journey eval runner — tests the prompt_template builder conversation.
+
+For each (journey_fixture × model) combination:
+1. Build a MockExecutor (for filesystem tools used during journey)
+2. Patch execute_on_client
+3. Override LLM_MODEL
+4. Call handle_journey_start to kick off the conversation
+5. Feed simulated user_messages via handle_journey_message
+6. Collect the generated prompt_template
+7. Score it against expected_template_criteria (via LLM judge)
+8. Report to Langfuse
+"""
+
+from __future__ import annotations
+
+import asyncio
+import copy
+import json
+import logging
+import time
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+from eval.config import JourneyFixture
+from eval.mock_executor import MockExecutor
+from eval import langfuse_eval
+
+logger = logging.getLogger(__name__)
+
+
+# ── Result type ──────────────────────────────────────────────────────────
+
+
+@dataclass
+class JourneyEvalResult:
+    """Result of one journey eval run."""
+
+    fixture_name: str
+    model: str
+    prompt_template: str | None  # the generated template (None if journey failed)
+    conversation_turns: int
+    done: bool  # whether journey reached completion
+    criteria_scores: dict[str, float]  # criterion → 0-1 score
+    overall_score: float  # average of criteria scores
+    judge_reasoning: str
+    elapsed_seconds: float
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "fixture": self.fixture_name,
+            "model": self.model,
+            "done": self.done,
+            "turns": self.conversation_turns,
+            "overall_score": round(self.overall_score, 3),
+            "criteria_scores": {k: round(v, 3) for k, v in self.criteria_scores.items()},
+            "elapsed_s": round(self.elapsed_seconds, 1),
+        }
+
+
+# ── LLM judge for template quality ──────────────────────────────────────
+
+_JOURNEY_JUDGE_SYSTEM = """\
+You are an evaluation judge for AI-generated prompt templates.
+
+A journey chatbot explored a user's directory structure and through
+conversation produced a prompt_template — an instruction set for a
+data-extraction agent.
+
+Your task: evaluate the generated template against a list of criteria.
+Score each criterion from 0 to 1:
+  - 1.0: Fully satisfied, clearly present in the template
+  - 0.5: Partially satisfied or ambiguously addressed
+  - 0.0: Not satisfied, missing from the template
+
+Respond with ONLY a JSON object:
+{
+  "scores": {"criterion_1": 0.8, "criterion_2": 1.0, ...},
+  "reasoning": "Brief explanation"
+}
+"""
+
+
+async def _judge_template(
+    prompt_template: str,
+    criteria: list[str],
+    *,
+    judge_model: str = "gpt-4o-mini",
+) -> tuple[dict[str, float], str]:
+    """Use an LLM to evaluate a generated prompt_template against criteria.
+
+    Returns (criteria_scores, reasoning).
+    """
+    from shared.llm import get_llm
+
+    llm = get_llm(model=judge_model, temperature=0)
+
+    criteria_text = "\n".join(f"  {i+1}. {c}" for i, c in enumerate(criteria))
+    user_content = (
+        f"## Generated prompt_template\n```\n{prompt_template}\n```\n\n"
+        f"## Criteria to evaluate\n{criteria_text}"
+    )
+
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=_JOURNEY_JUDGE_SYSTEM),
+            HumanMessage(content=user_content),
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+
+        scores_raw = parsed.get("scores", {})
+        # Map criterion keys back to the original criteria text
+        criteria_scores: dict[str, float] = {}
+        for i, criterion in enumerate(criteria):
+            # Try matching by index key or exact criterion text
+            key_candidates = [
+                f"criterion_{i+1}",
+                criterion,
+                criterion[:50],
+                str(i + 1),
+            ]
+            score = 0.0
+            for key in key_candidates:
+                if key in scores_raw:
+                    score = float(scores_raw[key])
+                    break
+            # If no match found, try values in order
+            if score == 0.0 and i < len(scores_raw):
+                score = float(list(scores_raw.values())[i])
+            criteria_scores[criterion] = score
+
+        reasoning = str(parsed.get("reasoning", ""))
+        return criteria_scores, reasoning
+    except Exception as exc:
+        logger.warning("journey_eval: LLM judge failed: %s", exc)
+        return {c: 0.0 for c in criteria}, f"Judge error: {exc}"
+
+
+# ── Journey runner ───────────────────────────────────────────────────────
+
+
+async def run_single_journey_eval(
+    fixture: JourneyFixture,
+    model: str,
+    *,
+    judge_model: str = "gpt-4o-mini",
+    data_dir: Path | None = None,
+) -> JourneyEvalResult:
+    """Execute one journey eval: start \u2192 messages \u2192 score template."""
+    from shared.config import settings
+
+    # When data_dir is given, use its parent as MockExecutor root
+    # and its name as the journey directory so the LLM sees a
+    # meaningful path (not ".").
+    if data_dir:
+        mock_root = data_dir.parent
+        journey_directory = data_dir.name
+    else:
+        mock_root = fixture.fixture_path.parent
+        journey_directory = fixture.directory
+
+    mock = MockExecutor(
+        fixture_dir=mock_root,
+        seed_records={},
+    )
+
+    original_model = settings.LLM_MODEL
+    settings.LLM_MODEL = model
+
+    eval_user_id = f"eval-journey-{uuid.uuid4().hex[:8]}"
+
+    logger.info(
+        "journey_eval: starting %s | model=%s",
+        fixture.name, model,
+    )
+    start_time = time.time()
+
+    prompt_template: str | None = None
+    conversation: list[dict[str, str]] = []
+    done = False
+
+    try:
+        from shared.ws_context import set_current_user, clear_current_user
+        from app.journey import handle_journey_start, handle_journey_message, _sessions
+
+        set_current_user(eval_user_id)
+        with mock.patch():
+            # ── Start the journey ────────────────────────────────
+            start_frame: dict[str, Any] = {
+                "agent_type": "local",
+                "directory": journey_directory,
+                "data_types": fixture.data_types,
+                "session_id": f"eval-{uuid.uuid4().hex[:8]}",
+            }
+
+            reply = await handle_journey_start(eval_user_id, start_frame)
+            session_id = reply["session_id"]
+            conversation.append({"role": "assistant", "content": reply["message"]})
+
+            logger.info(
+                "journey_eval: start reply (%d chars), done=%s",
+                len(reply["message"]), reply["done"],
+            )
+
+            if reply["done"]:
+                prompt_template = reply.get("prompt_template")
+                done = True
+            else:
+                # ── Send user messages ───────────────────────────
+                for i, user_msg in enumerate(fixture.user_messages):
+                    if done:
+                        break
+
+                    conversation.append({"role": "user", "content": user_msg})
+
+                    msg_frame: dict[str, Any] = {
+                        "session_id": session_id,
+                        "message": user_msg,
+                    }
+                    reply = await handle_journey_message(eval_user_id, msg_frame)
+                    conversation.append({"role": "assistant", "content": reply["message"]})
+
+                    logger.info(
+                        "journey_eval: turn %d reply (%d chars), done=%s",
+                        i + 1, len(reply["message"]), reply["done"],
+                    )
+
+                    if reply["done"]:
+                        prompt_template = reply.get("prompt_template")
+                        done = True
+
+                # If not done after all user messages, send a final nudge
+                if not done:
+                    nudge = "Please generate the final prompt_template now. I'm satisfied with the configuration."
+                    conversation.append({"role": "user", "content": nudge})
+
+                    nudge_frame: dict[str, Any] = {
+                        "session_id": session_id,
+                        "message": nudge,
+                    }
+                    reply = await handle_journey_message(eval_user_id, nudge_frame)
+                    conversation.append({"role": "assistant", "content": reply["message"]})
+                    if reply["done"]:
+                        prompt_template = reply.get("prompt_template")
+                        done = True
+
+    except Exception as exc:
+        logger.error("journey_eval: pipeline failed for %s/%s: %s", fixture.name, model, exc)
+    finally:
+        settings.LLM_MODEL = original_model
+        from shared.ws_context import clear_current_user
+        clear_current_user()
+
+    elapsed = time.time() - start_time
+    turns = len([c for c in conversation if c["role"] == "user"])
+
+    logger.info(
+        "journey_eval: completed in %.1fs — %d turns, done=%s, template=%s",
+        elapsed, turns, done, "yes" if prompt_template else "no",
+    )
+
+    # ── Score the template ───────────────────────────────────────
+    criteria_scores: dict[str, float] = {}
+    judge_reasoning = ""
+
+    if prompt_template and fixture.expected_template_criteria:
+        criteria_scores, judge_reasoning = await _judge_template(
+            prompt_template,
+            fixture.expected_template_criteria,
+            judge_model=judge_model,
+        )
+    elif not prompt_template:
+        criteria_scores = {c: 0.0 for c in fixture.expected_template_criteria}
+        judge_reasoning = "No prompt_template was generated — journey did not complete."
+
+    overall = (
+        sum(criteria_scores.values()) / len(criteria_scores)
+        if criteria_scores
+        else 0.0
+    )
+
+    result = JourneyEvalResult(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_template=prompt_template,
+        conversation_turns=turns,
+        done=done,
+        criteria_scores=criteria_scores,
+        overall_score=overall,
+        judge_reasoning=judge_reasoning,
+        elapsed_seconds=elapsed,
+    )
+
+    # ── Report to Langfuse ───────────────────────────────────────
+    trace_id = langfuse_eval.log_eval_trace(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_variant="journey",
+        prompt_template=prompt_template or "(not generated)",
+        actual_mutations=[{"conversation": conversation[:20]}],
+        scores_summary=result.summary(),
+        langfuse_prompt_names=["journey_system"],
+    )
+
+    if trace_id:
+        from eval.scorer import EvalScores
+        scores_obj = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant="journey",
+            precision=overall,
+            recall=float(done),
+            f1=overall,
+            llm_judge_score=overall,
+            llm_judge_reasoning=judge_reasoning,
+        )
+        langfuse_eval.post_eval_scores(scores_obj, trace_id=trace_id)
+
+    return result
+
+
+async def run_journey_fixture_eval(
+    fixture: JourneyFixture,
+    models: list[str],
+    *,
+    judge_model: str = "gpt-4o-mini",
+    data_dir: Path | None = None,
+) -> list[JourneyEvalResult]:
+    """Run all models for a journey fixture."""
+    langfuse_eval.sync_journey_fixture_to_dataset(fixture)
+
+    results: list[JourneyEvalResult] = []
+    for model in models:
+        result = await run_single_journey_eval(
+            fixture, model, judge_model=judge_model,
+            data_dir=data_dir,
+        )
+        results.append(result)
+
+    return results
+
+
+def print_journey_results(results: list[JourneyEvalResult]) -> None:
+    """Print a formatted summary of journey eval results."""
+    if not results:
+        print("\nNo journey eval results.")
+        return
+
+    print("\n" + "=" * 95)
+    print(f"{'Fixture':<25} {'Model':<25} {'Done':>5} {'Turns':>6} {'Score':>7} {'Time':>7}")
+    print("-" * 95)
+
+    for r in results:
+        done_str = "yes" if r.done else "NO"
+        print(
+            f"{r.fixture_name:<25} {r.model:<25} {done_str:>5} "
+            f"{r.conversation_turns:>6} {r.overall_score:>7.2f} {r.elapsed_seconds:>6.1f}s"
+        )
+
+    print("=" * 95)
+
+    # Criteria breakdown
+    for r in results:
+        if r.criteria_scores:
+            print(f"\n[{r.model}] Criteria scores:")
+            for criterion, score in r.criteria_scores.items():
+                indicator = "PASS" if score >= 0.7 else "PARTIAL" if score >= 0.4 else "FAIL"
+                print(f"  {indicator:>7} ({score:.1f}) {criterion}")
+
+        if r.judge_reasoning:
+            print(f"  Judge: {r.judge_reasoning}")
+
+        if r.prompt_template:
+            preview = r.prompt_template[:200].replace("\n", " ")
+            print(f"  Template preview: {preview}...")
+
+    print()

services/batch-agent/eval/langfuse_eval.py

@@ -0,0 +1,327 @@
+"""Langfuse evaluation integration — datasets, runs, and scoring.
+
+Uses the Langfuse Python SDK v4 (OpenTelemetry-based) to:
+
+1. **Sync fixtures → Langfuse datasets**: Each YAML fixture becomes a dataset,
+   each prompt variant + expected pair becomes a dataset item.
+
+2. **Track eval runs**: Each (fixture × model × prompt_variant) execution
+   is recorded as a trace with linked scores.
+
+3. **Post scores**: precision, recall, F1, field_accuracy, llm_judge are
+   posted as numeric scores on the trace.
+"""
+
+from __future__ import annotations
+
+import logging
+import os
+from typing import Any
+
+from shared.config import settings
+from eval.config import EvalFixture
+from eval.scorer import EvalScores
+
+logger = logging.getLogger(__name__)
+
+
+def _get_langfuse():
+    """Get or create a Langfuse client instance (SDK v4)."""
+    if not settings.LANGFUSE_SECRET_KEY or not settings.LANGFUSE_PUBLIC_KEY:
+        return None
+    try:
+        os.environ.setdefault("LANGFUSE_SECRET_KEY", settings.LANGFUSE_SECRET_KEY)
+        os.environ.setdefault("LANGFUSE_PUBLIC_KEY", settings.LANGFUSE_PUBLIC_KEY)
+        if settings.LANGFUSE_HOST:
+            os.environ.setdefault("LANGFUSE_HOST", settings.LANGFUSE_HOST)
+        from langfuse import get_client
+        return get_client()
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to create client: %s", exc)
+        return None
+
+
+def sync_fixture_to_dataset(fixture: EvalFixture) -> str | None:
+    """Create or update a Langfuse dataset from a fixture.
+
+    Each prompt variant becomes a separate dataset item with:
+    - input: {directory, data_types, prompt_template, seed_records}
+    - expected_output: {expected records}
+
+    Returns the dataset name, or None if Langfuse is unavailable.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        logger.info("langfuse_eval: Langfuse not configured — skipping dataset sync")
+        return None
+
+    dataset_name = f"batch-eval-{fixture.name}"
+
+    try:
+        lf.create_dataset(
+            name=dataset_name,
+            description=fixture.description,
+            metadata={
+                "data_types": ",".join(fixture.data_types),
+                "file_extensions": ",".join(fixture.file_extensions) if fixture.file_extensions else "",
+            },
+        )
+    except Exception:
+        # Dataset may already exist — that's fine
+        pass
+
+    # Build expected_output appropriate to the fixture's mode
+    expected_output: dict[str, Any] = {}
+    if fixture.mode in ("step1", "full") and fixture.expected_classification:
+        expected_output["classifications"] = [
+            {"file": ec.file, "project_id": ec.project_id, "domains": ec.domains}
+            for ec in fixture.expected_classification
+        ]
+    if fixture.mode in ("step2", "full") and fixture.expected:
+        for rec in fixture.expected:
+            expected_output.setdefault(rec.table, []).append(rec.fields)
+
+    item_id = f"{fixture.name}--{fixture.mode}"
+    try:
+        lf.create_dataset_item(
+            dataset_name=dataset_name,
+            id=item_id,
+            input={
+                "directory": fixture.directory,
+                "data_types": fixture.data_types,
+                "mode": fixture.mode,
+                "seed_records": fixture.seed_records,
+            },
+            expected_output=expected_output,
+            metadata={"mode": fixture.mode},
+        )
+    except Exception as exc:
+        logger.warning(
+            "langfuse_eval: failed to upsert dataset item %s: %s", item_id, exc
+        )
+
+    lf.flush()
+    logger.info("langfuse_eval: synced fixture '%s' → dataset '%s'", fixture.name, dataset_name)
+    return dataset_name
+
+
+def sync_journey_fixture_to_dataset(fixture) -> str | None:
+    """Create or update a Langfuse dataset from a journey fixture.
+
+    Each journey fixture becomes a single dataset item with:
+    - input: {directory, data_types, user_messages}
+    - expected_output: {criteria}
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        logger.info("langfuse_eval: Langfuse not configured — skipping journey dataset sync")
+        return None
+
+    dataset_name = f"journey-eval-{fixture.name}"
+
+    try:
+        lf.create_dataset(
+            name=dataset_name,
+            description=fixture.description,
+            metadata={"type": "journey", "data_types": ",".join(fixture.data_types)},
+        )
+    except Exception:
+        pass  # Dataset may already exist
+
+    item_id = f"{fixture.name}--journey"
+    try:
+        lf.create_dataset_item(
+            dataset_name=dataset_name,
+            id=item_id,
+            input={
+                "directory": fixture.directory,
+                "data_types": fixture.data_types,
+                "user_messages": fixture.user_messages,
+            },
+            expected_output={
+                "criteria": fixture.expected_template_criteria,
+            },
+            metadata={"type": "journey"},
+        )
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to upsert journey dataset item %s: %s", item_id, exc)
+
+    lf.flush()
+    logger.info("langfuse_eval: synced journey fixture '%s' → dataset '%s'", fixture.name, dataset_name)
+    return dataset_name
+
+
+def create_eval_run(
+    dataset_name: str,
+    run_name: str,
+    *,
+    metadata: dict[str, Any] | None = None,
+) -> str:
+    """Create a dataset run in Langfuse. Returns the run name.
+
+    Note: In SDK v4, dataset runs are created implicitly via
+    dataset.run_experiment(). This function is kept for backwards
+    compatibility but may not create a run.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        return run_name
+
+    try:
+        if hasattr(lf, "create_dataset_run"):
+            lf.create_dataset_run(
+                dataset_name=dataset_name,
+                run_name=run_name,
+                metadata=metadata or {},
+            )
+            lf.flush()
+        else:
+            logger.debug("langfuse_eval: create_dataset_run not available in SDK v4")
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to create run %s: %s", run_name, exc)
+
+    return run_name
+
+
+def post_eval_scores(
+    scores: EvalScores,
+    *,
+    trace_id: str | None = None,
+    dataset_name: str | None = None,
+    run_name: str | None = None,
+) -> None:
+    """Post evaluation scores to Langfuse.
+
+    If trace_id is provided, scores are attached to that trace.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        return
+
+    score_data = [
+        ("precision", scores.precision),
+        ("recall", scores.recall),
+        ("f1", scores.f1),
+    ]
+    # Only post field_accuracy when there are field-level scores (step2/full)
+    if scores.field_scores:
+        score_data.append(("field_accuracy", scores.field_accuracy))
+    if scores.llm_judge_score is not None:
+        score_data.append(("llm_judge", scores.llm_judge_score))
+
+    for name, value in score_data:
+        try:
+            lf.create_score(
+                name=name,
+                value=value,
+                trace_id=trace_id,
+                data_type="NUMERIC",
+                comment=f"{scores.fixture_name} | {scores.model} | {scores.prompt_variant}",
+            )
+        except Exception as exc:
+            logger.warning("langfuse_eval: failed to post score %s: %s", name, exc)
+
+    lf.flush()
+    logger.info(
+        "langfuse_eval: posted %d scores for %s/%s/%s",
+        len(score_data), scores.fixture_name, scores.model, scores.prompt_variant,
+    )
+
+
+def log_eval_trace(
+    *,
+    fixture_name: str,
+    model: str,
+    prompt_variant: str,
+    prompt_template: str,
+    actual_mutations: list[dict],
+    scores_summary: dict[str, Any],
+    step1_results: list[dict] | None = None,
+    dataset_name: str | None = None,
+    run_name: str | None = None,
+    dataset_item_id: str | None = None,
+    langfuse_prompt_names: list[str] | None = None,
+) -> str | None:
+    """Create a Langfuse trace for one eval execution and link it to a dataset run.
+
+    Uses SDK v4 observation API (traces are created implicitly by root spans).
+    ``langfuse_prompt_names`` can contain one or two prompt names to link
+    (e.g. ``["batch_file_classifier", "batch_processing"]`` for full mode).
+    Each prompt gets its own generation-type observation for per-version
+    metrics tracking.
+
+    Returns the trace_id, or None if Langfuse is unavailable.
+    """
+    lf = _get_langfuse()
+    if lf is None:
+        return None
+
+    try:
+        from langfuse import propagate_attributes
+
+        # Fetch prompt objects for linking
+        prompt_objs: list[tuple[str, Any]] = []
+        for pname in (langfuse_prompt_names or []):
+            try:
+                obj = lf.get_prompt(name=pname, cache_ttl_seconds=300)
+                prompt_objs.append((pname, obj))
+                logger.info("langfuse_eval: linked prompt '%s' (type=%s)", pname, type(obj).__name__)
+            except Exception as exc:
+                logger.warning("langfuse_eval: prompt '%s' not found — %s", pname, exc)
+
+        # Build trace output dict
+        trace_output: dict[str, Any] = {"scores": scores_summary}
+        if step1_results:
+            trace_output["classifications"] = step1_results
+        if actual_mutations:
+            trace_output["mutations"] = actual_mutations[:50]
+
+        with propagate_attributes(
+            trace_name=f"eval-{fixture_name}",
+            metadata={
+                "eval": "true",
+                "fixture": fixture_name,
+                "model": model,
+                "prompt_variant": prompt_variant,
+            },
+            tags=["eval", f"model:{model}", f"variant:{prompt_variant}"],
+        ):
+            # Root span for the eval run
+            span = lf.start_observation(name=f"eval-{fixture_name}")
+            span.update(
+                input={
+                    "prompt_template": prompt_template,
+                    "model": model,
+                    "prompt_variant": prompt_variant,
+                },
+                output=trace_output,
+            )
+            trace_id = span.trace_id
+
+            # Create a generation-type observation per linked prompt
+            for pname, pobj in prompt_objs:
+                gen = lf.start_observation(
+                    name=f"prompt-{pname}",
+                    prompt=pobj,
+                    as_type="generation",
+                )
+                gen.end()
+
+            # Link to dataset run if available
+            if dataset_name and run_name and dataset_item_id:
+                try:
+                    dataset = lf.get_dataset(dataset_name)
+                    for item in dataset.items:
+                        if item.id == dataset_item_id:
+                            item.link(span, run_name)
+                            break
+                except Exception as exc:
+                    logger.warning("langfuse_eval: failed to link trace to dataset run: %s", exc)
+
+            span.end()
+
+        lf.flush()
+        return trace_id
+    except Exception as exc:
+        logger.warning("langfuse_eval: failed to create eval trace: %s", exc)
+        return None

services/batch-agent/eval/mock_executor.py

@@ -0,0 +1,258 @@
+"""Mock executor — intercepts execute_on_client for offline E2E testing.
+
+Patches ``execute_on_client`` at all usage sites so agent pipeline runs don't
+require a live Electron client or Redis.  Instead:
+
+- **Filesystem actions** (list_directory, read_file_content, get_file_metadata)
+  are served from local fixture files on disk.
+- **Read actions** (select, get) return preseeded records from an in-memory
+  store provided by the test fixture.
+- **Write actions** (insert, update, delete) are captured as *mutations* and
+  stored for later comparison against expected results.
+"""
+
+from __future__ import annotations
+
+import json
+import os
+import time
+import uuid
+from dataclasses import dataclass, field
+from pathlib import Path
+from typing import Any
+from contextlib import contextmanager, asynccontextmanager
+from unittest.mock import AsyncMock, patch
+
+
+@dataclass
+class Mutation:
+    """A single recorded write operation."""
+
+    action: str  # insert | update | delete
+    table: str
+    data: dict[str, Any]
+    timestamp: float = field(default_factory=time.time)
+
+
+# ── Fake DB helpers (used to bypass async_session in full mode) ───────
+
+class _FakeRow:
+    """Mimics an AgentRunLog row returned by SQLAlchemy."""
+    id = 0
+    status = "running"
+    items_processed = 0
+    items_created = 0
+    errors: list[str] = []
+    completed_at = None
+
+    def __setattr__(self, name: str, value: Any) -> None:
+        object.__setattr__(self, name, value)
+
+
+class _FakeResult:
+    """Mimics a SQLAlchemy ``Result`` with ``scalar_one_or_none``."""
+    def __init__(self, row: _FakeRow) -> None:
+        self._row = row
+
+    def scalar_one_or_none(self) -> _FakeRow:
+        return self._row
+
+
+@dataclass
+class MockExecutor:
+    """In-memory executor that replaces Redis-based tool round-trip.
+
+    Parameters
+    ----------
+    fixture_dir : Path
+        Directory containing sample files for filesystem tool calls.
+    seed_records : dict[str, list[dict]]
+        Pre-existing records per table, e.g. ``{"tasks": [...], "projects": [...]}``.
+        The executor returns these for ``select`` / ``get`` actions and auto-updates
+        them on ``insert`` / ``update`` / ``delete`` so subsequent selects reflect changes.
+    """
+
+    fixture_dir: Path
+    seed_records: dict[str, list[dict]] = field(default_factory=dict)
+    mutations: list[Mutation] = field(default_factory=list)
+    _id_counter: int = field(default=1000, repr=False)
+
+    # ── Public API ───────────────────────────────────────────────────
+
+    def reset(self) -> None:
+        """Clear recorded mutations (keep seed_records intact)."""
+        self.mutations.clear()
+
+    def get_mutations(self, *, table: str | None = None, action: str | None = None) -> list[Mutation]:
+        """Filter mutations by table and/or action."""
+        result = self.mutations
+        if table:
+            result = [m for m in result if m.table == table]
+        if action:
+            result = [m for m in result if m.action == action]
+        return result
+
+    def created_records(self, table: str) -> list[dict]:
+        """Return data dicts of all inserts into *table*."""
+        return [m.data for m in self.mutations if m.table == table and m.action == "insert"]
+
+    def updated_records(self, table: str) -> list[dict]:
+        """Return data dicts of all updates to *table*."""
+        return [m.data for m in self.mutations if m.table == table and m.action == "update"]
+
+    # ── Context manager for patching ──────────────────────────────
+
+    @contextmanager
+    def patch(self):
+        """Patch execute_on_client and DB session at all usage sites."""
+        mock_fn = AsyncMock(side_effect=self._handle)
+        targets = [
+            "shared.ws_context.execute_on_client",
+            "app.agent_runner.execute_on_client",
+            "app.agents.filesystem_agent.execute_on_client",
+        ]
+
+        # Mock async_session so run_local_agent / _finalize_run skip real DB
+        fake_row = _FakeRow()
+        fake_db = AsyncMock()
+        fake_db.commit = AsyncMock()
+        fake_db.refresh = AsyncMock()
+        fake_db.execute = AsyncMock(return_value=_FakeResult(fake_row))
+        fake_db.add = lambda obj: None  # noqa: ARG005
+
+        @asynccontextmanager
+        async def _fake_session():
+            yield fake_db
+
+        patches = [patch(t, new=mock_fn) for t in targets]
+        patches.append(patch("app.agent_runner.async_session", _fake_session))
+        for p in patches:
+            p.start()
+        try:
+            yield mock_fn
+        finally:
+            for p in patches:
+                p.stop()
+
+    # ── Internal dispatch ─────────────────────────────────────────
+
+    async def _handle(
+        self,
+        action: str,
+        table: str | None = None,
+        data: dict[str, Any] | None = None,
+        filters: dict[str, Any] | None = None,
+        vector: list[float] | None = None,
+        limit: int | None = None,
+    ) -> dict[str, Any]:
+        # Filesystem
+        if action == "list_directory":
+            return self._list_directory(data or {})
+        if action == "read_file_content":
+            return self._read_file(data or {})
+        if action == "get_file_metadata":
+            return self._get_file_metadata(data or {})
+
+        # CRUD
+        if action == "select":
+            return self._select(table or "", filters)
+        if action == "get":
+            return self._get(table or "", data or {})
+        if action == "insert":
+            return self._insert(table or "", data or {})
+        if action == "update":
+            return self._update(table or "", data or {})
+        if action == "delete":
+            return self._delete(table or "", data or {})
+
+        # Vector (no-op for eval)
+        if action in ("vector_upsert", "vector_search"):
+            return {"rows": []}
+
+        return {"error": f"Unknown action: {action}"}
+
+    # ── Filesystem handlers ───────────────────────────────────────
+
+    def _list_directory(self, data: dict) -> dict:
+        rel_path = data.get("path", "")
+        abs_path = self.fixture_dir / rel_path.lstrip("/\\")
+        if not abs_path.is_dir():
+            return {"entries": []}
+        entries: list[dict] = []
+        for child in sorted(abs_path.iterdir()):
+            entry_type = "directory" if child.is_dir() else "file"
+            # Return paths relative to fixture_dir but with the original prefix
+            entry_path = rel_path.rstrip("/\\") + "/" + child.name
+            entries.append({
+                "name": child.name,
+                "path": entry_path,
+                "type": entry_type,
+            })
+        return {"entries": entries}
+
+    def _read_file(self, data: dict) -> dict:
+        rel_path = data.get("path", "")
+        abs_path = self.fixture_dir / rel_path.lstrip("/\\")
+        if not abs_path.is_file():
+            return {"content": "", "error": f"File not found: {rel_path}"}
+        return {"content": abs_path.read_text(encoding="utf-8", errors="replace")}
+
+    def _get_file_metadata(self, data: dict) -> dict:
+        rel_path = data.get("path", "")
+        abs_path = self.fixture_dir / rel_path.lstrip("/\\")
+        if not abs_path.exists():
+            return {"error": f"Not found: {rel_path}"}
+        stat = abs_path.stat()
+        return {
+            "path": rel_path,
+            "size": stat.st_size,
+            "modifiedAt": int(stat.st_mtime * 1000),
+            "createdAt": int(stat.st_ctime * 1000),
+            "isDirectory": abs_path.is_dir(),
+        }
+
+    # ── CRUD handlers ─────────────────────────────────────────────
+
+    def _select(self, table: str, filters: dict | None) -> dict:
+        rows = list(self.seed_records.get(table, []))
+        if filters:
+            rows = [
+                r for r in rows
+                if all(r.get(k) == v for k, v in filters.items() if v is not None)
+            ]
+        return {"rows": rows}
+
+    def _get(self, table: str, data: dict) -> dict:
+        record_id = data.get("id", "")
+        rows = self.seed_records.get(table, [])
+        for r in rows:
+            if r.get("id") == record_id:
+                return {"row": r}
+        return {"row": None}
+
+    def _insert(self, table: str, data: dict) -> dict:
+        self._id_counter += 1
+        record = {**data, "id": str(self._id_counter)}
+        # Add to seed so subsequent selects can find it
+        self.seed_records.setdefault(table, []).append(record)
+        self.mutations.append(Mutation(action="insert", table=table, data=record))
+        return {"row": record}
+
+    def _update(self, table: str, data: dict) -> dict:
+        record_id = data.get("id", "")
+        rows = self.seed_records.get(table, [])
+        for r in rows:
+            if r.get("id") == record_id:
+                r.update({k: v for k, v in data.items() if v is not None and v != ""})
+                self.mutations.append(Mutation(action="update", table=table, data=dict(r)))
+                return {"row": r}
+        # Record not found — still log the mutation
+        self.mutations.append(Mutation(action="update", table=table, data=data))
+        return {"row": data}
+
+    def _delete(self, table: str, data: dict) -> dict:
+        record_id = data.get("id", "")
+        rows = self.seed_records.get(table, [])
+        self.seed_records[table] = [r for r in rows if r.get("id") != record_id]
+        self.mutations.append(Mutation(action="delete", table=table, data={"id": record_id}))
+        return {"deleted": True}

services/batch-agent/eval/requirements.txt

@@ -0,0 +1,2 @@
+# Extra dependencies for the eval harness (on top of the service requirements.txt)
+pyyaml>=6.0.0

services/batch-agent/eval/runner.py

@@ -0,0 +1,463 @@
+"""Eval runner — orchestrates fixture → mock → agent pipeline → scoring.
+
+Supports three eval modes:
+
+- **step1**: Test classification prompt only (``_STEP1_SYSTEM_PROMPT``).
+  Calls the LLM with fixture-provided ``domain_definitions`` and
+  ``projects_list`` and compares output against ``expected_classification``.
+
+- **step2**: Test processing prompt only (``_PROCESSING_SYSTEM_PROMPT``).
+  Compiles the prompt with fixture-provided ``existing_context``,
+  ``project_context``, ``data_types``, and ``custom_prompt_section``,
+  then runs the tool-calling loop.  Mutations are scored against
+  ``expected`` records.
+
+- **full**: Run ``run_local_agent()`` end-to-end (both steps).
+  Scored on both classification and extraction.
+"""
+
+from __future__ import annotations
+
+import copy
+import json
+import logging
+import time
+import uuid
+from typing import Any
+
+from eval.config import EvalFixture, ExpectedClassification
+from eval.mock_executor import MockExecutor
+from eval.scorer import (
+    EvalScores,
+    FieldScore,
+    compute_precision_recall,
+    llm_judge_score,
+    score_field_match,
+)
+from eval import langfuse_eval
+
+logger = logging.getLogger(__name__)
+
+
+# ── Step 1 runner ─────────────────────────────────────────────────────────
+
+
+async def _run_step1(
+    fixture: EvalFixture,
+    model: str,
+    mock: MockExecutor,
+) -> list[dict[str, Any]]:
+    """Run step-1 classification for each expected file.
+
+    Returns a list of result dicts:
+    ``[{file, project_id, domains, new_project_name}, ...]``
+    """
+    from app.agent_runner import _classify_file
+
+    results: list[dict[str, Any]] = []
+    for ec in fixture.expected_classification:
+        # Read the file content through the mock
+        file_result = await mock._handle(
+            action="read_file_content",
+            data={"path": ec.file},
+        )
+        file_content: str = file_result.get("content", "")
+
+        project_id, domains, new_name = await _classify_file(
+            file_path=ec.file,
+            file_content=file_content,
+            projects=fixture.projects_list,
+            config_data_types=fixture.data_types,
+        )
+        results.append({
+            "file": ec.file,
+            "project_id": project_id,
+            "domains": domains,
+            "new_project_name": new_name,
+        })
+    return results
+
+
+def _score_step1(
+    fixture: EvalFixture,
+    results: list[dict[str, Any]],
+) -> tuple[float, float, float, str]:
+    """Score step-1 results. Returns (precision, recall, f1, reasoning)."""
+    if not fixture.expected_classification:
+        return 0.0, 0.0, 0.0, "No expected classifications"
+
+    total = len(fixture.expected_classification)
+    matched = 0
+    details: list[str] = []
+
+    for ec in fixture.expected_classification:
+        actual = next((r for r in results if r["file"] == ec.file), None)
+        if actual is None:
+            details.append(f"  MISS {ec.file}: not processed")
+            continue
+
+        pid_ok = actual["project_id"] == ec.project_id
+        domains_ok = set(actual["domains"]) == set(ec.domains) if ec.domains else True
+
+        if pid_ok and domains_ok:
+            matched += 1
+            details.append(f"  OK   {ec.file}: project={actual['project_id']}, domains={actual['domains']}")
+        else:
+            parts: list[str] = []
+            if not pid_ok:
+                parts.append(f"project expected={ec.project_id} got={actual['project_id']}")
+            if not domains_ok:
+                parts.append(f"domains expected={ec.domains} got={actual['domains']}")
+            details.append(f"  FAIL {ec.file}: {'; '.join(parts)}")
+
+    precision = matched / total if total > 0 else 0.0
+    recall = precision  # in step1, precision == recall (same denominator)
+    f1 = precision  # same
+    reasoning = "\n".join(details)
+    return precision, recall, f1, reasoning
+
+
+# ── Step 2 runner ─────────────────────────────────────────────────────────
+
+
+async def _run_step2(
+    fixture: EvalFixture,
+    model: str,
+    mock: MockExecutor,
+) -> None:
+    """Run step-2 processing for each file in the fixture directory.
+
+    Compiles ``_PROCESSING_SYSTEM_PROMPT`` with fixture-provided variables
+    and runs the tool-calling loop.  Mutations are captured by the mock.
+    """
+    from app.agent_runner import (
+        _PROCESSING_SYSTEM_PROMPT,
+        _build_processing_tools,
+        _run_agent_with_tools,
+        _MAX_PROCESSING_STEPS,
+    )
+    from app import tracing
+
+    # Compile the processing prompt with fixture variables
+    system_prompt = tracing.compile_prompt(
+        "batch_processing",
+        fallback=_PROCESSING_SYSTEM_PROMPT,
+        variables={
+            "existing_context": fixture.existing_context,
+            "project_context": fixture.project_context,
+            "data_types": ", ".join(fixture.data_types),
+            "custom_prompt_section": fixture.custom_prompt_section,
+        },
+    )
+
+    tools = _build_processing_tools(fixture.data_types)
+
+    # Scan files in the fixture directory
+    file_entries = await mock._handle(
+        action="list_directory",
+        data={"path": fixture.directory},
+    )
+    for entry in file_entries.get("entries", []):
+        if entry.get("type") != "file":
+            continue
+        # Filter by extension if specified
+        if fixture.file_extensions:
+            ext = entry["name"].rsplit(".", 1)[-1] if "." in entry["name"] else ""
+            if ext not in fixture.file_extensions:
+                continue
+
+        file_result = await mock._handle(
+            action="read_file_content",
+            data={"path": entry["path"]},
+        )
+        file_content: str = file_result.get("content", "")
+        if not file_content.strip():
+            continue
+
+        await _run_agent_with_tools(
+            system_prompt=system_prompt,
+            user_message=(
+                f"Process this file and extract relevant information.\n\n"
+                f"File: {entry['path']}\n\nContent:\n{file_content}"
+            ),
+            tools=tools,
+            max_steps=_MAX_PROCESSING_STEPS,
+        )
+
+
+# ── Full runner ───────────────────────────────────────────────────────────
+
+
+async def _run_full(
+    fixture: EvalFixture,
+    model: str,
+    mock: MockExecutor,
+    user_id: str,
+) -> None:
+    """Run the full two-step pipeline via ``run_local_agent``."""
+    from app.agent_runner import run_local_agent
+
+    trigger_data: dict[str, Any] = {
+        "type": "agent_trigger",
+        "directory": fixture.directory,
+        "directory_paths": [fixture.directory],
+        "data_types": fixture.data_types,
+        "file_extensions": fixture.file_extensions,
+        "prompt_template": fixture.custom_prompt_section,
+        "device_id": "eval-harness",
+        "run_context": {
+            "agent_id": f"eval-{fixture.name}",
+            "run_id": None,
+        },
+    }
+
+    with mock.patch():
+        await run_local_agent(user_id, trigger_data)
+
+
+# ── Scoring helpers ───────────────────────────────────────────────────────
+
+
+def _score_mutations(
+    fixture: EvalFixture,
+    mock: MockExecutor,
+) -> tuple[list[FieldScore], float, float, float, int, int]:
+    """Score mutations against expected records.
+
+    Returns (field_scores, precision, recall, f1, extra, missing).
+    """
+    all_field_scores: list[FieldScore] = []
+    total_expected = 0
+    total_actual = 0
+    total_matched = 0
+    total_extra = 0
+    total_missing = 0
+
+    expected_by_table: dict[str, list[dict]] = {}
+    for rec in fixture.expected:
+        expected_by_table.setdefault(rec.table, []).append(rec.fields)
+
+    tables = set(expected_by_table.keys()) | {m.table for m in mock.mutations}
+    for table in tables:
+        expected_records = expected_by_table.get(table, [])
+        actual_records = mock.created_records(table) + mock.updated_records(table)
+
+        field_scores, extra, missing = score_field_match(expected_records, actual_records, table)
+        all_field_scores.extend(field_scores)
+
+        matched = sum(1 for s in field_scores if s.best_match is not None)
+        total_expected += len(expected_records)
+        total_actual += len(actual_records)
+        total_matched += matched
+        total_extra += extra
+        total_missing += missing
+
+    precision, recall, f1 = compute_precision_recall(total_expected, total_actual, total_matched)
+    return all_field_scores, precision, recall, f1, total_extra, total_missing
+
+
+# ── Main entry point ──────────────────────────────────────────────────────
+
+
+async def run_single_eval(
+    fixture: EvalFixture,
+    model: str,
+    *,
+    use_llm_judge: bool = True,
+    judge_model: str = "gpt-4o-mini",
+) -> EvalScores:
+    """Execute one eval run for a fixture + model.  Mode is read from the fixture."""
+    from shared.config import settings
+    from shared.ws_context import set_current_user, clear_current_user
+
+    seed = copy.deepcopy(fixture.seed_records)
+    mock = MockExecutor(
+        fixture_dir=fixture.fixture_path.parent,
+        seed_records=seed,
+    )
+
+    original_model = settings.LLM_MODEL
+    settings.LLM_MODEL = model
+    eval_user_id = str(uuid.uuid4())
+
+    logger.info(
+        "eval: starting %s | mode=%s | model=%s",
+        fixture.name, fixture.mode, model,
+    )
+    start_time = time.time()
+
+    step1_results: list[dict[str, Any]] = []
+    step1_reasoning = ""
+
+    try:
+        set_current_user(eval_user_id)
+
+        if fixture.mode == "step1":
+            with mock.patch():
+                step1_results = await _run_step1(fixture, model, mock)
+
+        elif fixture.mode == "step2":
+            with mock.patch():
+                await _run_step2(fixture, model, mock)
+
+        elif fixture.mode == "full":
+            with mock.patch():
+                # Step 1 — classification (independent from run_local_agent)
+                if fixture.expected_classification:
+                    step1_results = await _run_step1(fixture, model, mock)
+
+            # Step 2 — full pipeline (run_local_agent handles both steps)
+            await _run_full(fixture, model, mock, eval_user_id)
+
+    except Exception as exc:
+        logger.error("eval: pipeline failed for %s/%s: %s", fixture.name, model, exc)
+    finally:
+        settings.LLM_MODEL = original_model
+        clear_current_user()
+
+    elapsed = time.time() - start_time
+    logger.info("eval: completed in %.1fs — %d mutations", elapsed, len(mock.mutations))
+
+    # ── Score ─────────────────────────────────────────────────────
+
+    if fixture.mode == "step1":
+        s1_precision, s1_recall, s1_f1, step1_reasoning = _score_step1(fixture, step1_results)
+        scores = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant=fixture.mode,
+            precision=s1_precision,
+            recall=s1_recall,
+            f1=s1_f1,
+            llm_judge_reasoning=step1_reasoning,
+        )
+    else:
+        # step2 or full — score mutations
+        field_scores, precision, recall, f1, extra, missing = _score_mutations(fixture, mock)
+        scores = EvalScores(
+            fixture_name=fixture.name,
+            model=model,
+            prompt_variant=fixture.mode,
+            field_scores=field_scores,
+            precision=precision,
+            recall=recall,
+            f1=f1,
+            extra_records=extra,
+            missing_records=missing,
+        )
+
+        # Add step1 classification scores for full mode
+        if fixture.mode == "full" and fixture.expected_classification:
+            s1_p, s1_r, s1_f1, step1_reasoning = _score_step1(fixture, step1_results)
+            scores.llm_judge_reasoning = f"Step1 classification:\n{step1_reasoning}"
+
+        # Optional LLM judge for extraction quality
+        if use_llm_judge and fixture.expected:
+            all_expected = [r.fields for r in fixture.expected]
+            all_actual = [m.data for m in mock.mutations if m.action in ("insert", "update")]
+            judge_score, reasoning = await llm_judge_score(
+                all_expected, all_actual, judge_model=judge_model,
+            )
+            scores.llm_judge_score = judge_score
+            if step1_reasoning:
+                scores.llm_judge_reasoning += f"\n\nLLM judge:\n{reasoning}"
+            else:
+                scores.llm_judge_reasoning = reasoning
+
+    # ── Report to Langfuse ────────────────────────────────────────
+    prompt_names = {
+        "step1": ["batch_file_classifier"],
+        "step2": ["batch_processing"],
+        "full": ["batch_file_classifier", "batch_processing"],
+    }.get(fixture.mode, ["batch_processing"])
+
+    trace_id = langfuse_eval.log_eval_trace(
+        fixture_name=fixture.name,
+        model=model,
+        prompt_variant=fixture.mode,
+        prompt_template=fixture.custom_prompt_section or "(default)",
+        actual_mutations=[{"action": m.action, "table": m.table, "data": m.data} for m in mock.mutations],
+        scores_summary=scores.summary(),
+        step1_results=step1_results or None,
+        langfuse_prompt_names=prompt_names,
+    )
+
+    if trace_id:
+        langfuse_eval.post_eval_scores(scores, trace_id=trace_id)
+
+        # For full mode, post classification scores separately
+        if fixture.mode == "full" and fixture.expected_classification:
+            s1_p, s1_r, s1_f1, _ = _score_step1(fixture, step1_results)
+            for name, value in [
+                ("classification_precision", s1_p),
+                ("classification_recall", s1_r),
+                ("classification_f1", s1_f1),
+            ]:
+                try:
+                    from langfuse import get_client
+                    lf = get_client()
+                    if lf:
+                        lf.create_score(
+                            name=name,
+                            value=value,
+                            trace_id=trace_id,
+                            data_type="NUMERIC",
+                            comment=f"{fixture.name} | {model} | full",
+                        )
+                except Exception:
+                    pass
+
+    return scores
+
+
+async def run_fixture_eval(
+    fixture: EvalFixture,
+    models: list[str],
+    *,
+    use_llm_judge: bool = True,
+    judge_model: str = "gpt-4o-mini",
+) -> list[EvalScores]:
+    """Run all models for a fixture."""
+    langfuse_eval.sync_fixture_to_dataset(fixture)
+
+    results: list[EvalScores] = []
+    for model in models:
+        scores = await run_single_eval(
+            fixture, model,
+            use_llm_judge=use_llm_judge,
+            judge_model=judge_model,
+        )
+        results.append(scores)
+
+    return results
+
+
+def print_results(results: list[EvalScores]) -> None:
+    """Print a formatted summary table of eval results."""
+    if not results:
+        print("\nNo eval results.")
+        return
+
+    print("\n" + "=" * 95)
+    print(f"{'Fixture':<25} {'Mode':<6} {'Model':<25} {'P':>6} {'R':>6} {'F1':>6} {'FA':>6} {'LLM':>6}")
+    print("-" * 95)
+
+    for s in results:
+        llm_str = f"{s.llm_judge_score:.2f}" if s.llm_judge_score is not None else "  --"
+        print(
+            f"{s.fixture_name:<25} {s.prompt_variant:<6} {s.model:<25} "
+            f"{s.precision:>6.2f} {s.recall:>6.2f} {s.f1:>6.2f} "
+            f"{s.field_accuracy:>6.2f} {llm_str:>6}"
+        )
+
+    print("=" * 95)
+    print()
+
+    print("=" * 90)
+
+    # If LLM judge reasoning is available, print it
+    for s in results:
+        if s.llm_judge_reasoning:
+            print(f"\n[{s.model} / {s.prompt_variant}] LLM Judge: {s.llm_judge_reasoning}")
+
+    print()

services/batch-agent/eval/scorer.py

@@ -0,0 +1,268 @@
+"""Scoring functions for batch agent evaluation.
+
+Two scoring strategies:
+
+1. **FieldMatchScorer** — deterministic check: for each expected record,
+   find the best-matching actual record and compare specified fields.
+   Returns precision, recall, and per-field accuracy.
+
+2. **LLMJudgeScorer** — uses a secondary LLM to semantically evaluate
+   whether the actual extractions satisfy the expected intent, even if
+   wording differs.  Returns a 0-1 score + reasoning.
+"""
+
+from __future__ import annotations
+
+import json
+import logging
+from dataclasses import dataclass, field
+from difflib import SequenceMatcher
+from typing import Any
+
+from langchain_core.messages import HumanMessage, SystemMessage
+
+logger = logging.getLogger(__name__)
+
+
+# ── Result types ─────────────────────────────────────────────────────────
+
+
+@dataclass
+class FieldScore:
+    """Score for a single expected record against its best match."""
+
+    expected: dict[str, Any]
+    best_match: dict[str, Any] | None
+    matched_fields: dict[str, bool]
+    similarity: float  # 0-1 overall similarity
+
+    @property
+    def field_accuracy(self) -> float:
+        if not self.matched_fields:
+            return 0.0
+        return sum(self.matched_fields.values()) / len(self.matched_fields)
+
+
+@dataclass
+class EvalScores:
+    """Aggregated scores for one eval run."""
+
+    fixture_name: str
+    model: str
+    prompt_variant: str
+    field_scores: list[FieldScore] = field(default_factory=list)
+    precision: float = 0.0
+    recall: float = 0.0
+    f1: float = 0.0
+    llm_judge_score: float | None = None
+    llm_judge_reasoning: str = ""
+    extra_records: int = 0  # records created but not expected
+    missing_records: int = 0  # expected but not found
+
+    @property
+    def field_accuracy(self) -> float:
+        if not self.field_scores:
+            return 0.0
+        return sum(s.field_accuracy for s in self.field_scores) / len(self.field_scores)
+
+    def summary(self) -> dict[str, Any]:
+        return {
+            "fixture": self.fixture_name,
+            "model": self.model,
+            "prompt_variant": self.prompt_variant,
+            "precision": round(self.precision, 3),
+            "recall": round(self.recall, 3),
+            "f1": round(self.f1, 3),
+            "field_accuracy": round(self.field_accuracy, 3),
+            "llm_judge_score": round(self.llm_judge_score, 3) if self.llm_judge_score is not None else None,
+            "extra_records": self.extra_records,
+            "missing_records": self.missing_records,
+        }
+
+
+# ── Field Match Scorer ───────────────────────────────────────────────────
+
+
+def _normalize(value: Any) -> str:
+    """Normalize a value for comparison."""
+    if value is None:
+        return ""
+    return str(value).strip().lower()
+
+
+def _text_similarity(a: str, b: str) -> float:
+    """Fuzzy text similarity using SequenceMatcher."""
+    if not a and not b:
+        return 1.0
+    if not a or not b:
+        return 0.0
+    return SequenceMatcher(None, a.lower(), b.lower()).ratio()
+
+
+def _find_best_match(
+    expected: dict[str, Any],
+    actuals: list[dict[str, Any]],
+) -> tuple[dict[str, Any] | None, float]:
+    """Find the actual record most similar to expected, return (match, similarity)."""
+    if not actuals:
+        return None, 0.0
+
+    best_match = None
+    best_score = 0.0
+
+    # Primary matching key: title or name
+    expected_title = _normalize(expected.get("title", expected.get("name", "")))
+
+    for actual in actuals:
+        actual_title = _normalize(actual.get("title", actual.get("name", "")))
+        sim = _text_similarity(expected_title, actual_title)
+        if sim > best_score:
+            best_score = sim
+            best_match = actual
+
+    return best_match, best_score
+
+
+def _compare_fields(
+    expected: dict[str, Any],
+    actual: dict[str, Any],
+) -> dict[str, bool]:
+    """Compare each expected field against the actual record."""
+    results: dict[str, bool] = {}
+    for key, expected_val in expected.items():
+        actual_val = actual.get(key)
+        # Exact match for non-string types
+        if not isinstance(expected_val, str):
+            results[key] = actual_val == expected_val
+        else:
+            # Fuzzy match for strings (threshold: 0.7)
+            results[key] = _text_similarity(
+                _normalize(expected_val), _normalize(actual_val)
+            ) >= 0.7
+    return results
+
+
+def score_field_match(
+    expected_records: list[dict[str, Any]],
+    actual_records: list[dict[str, Any]],
+    table: str,
+) -> tuple[list[FieldScore], int, int]:
+    """Score actual extractions against expected records for one table.
+
+    Returns (field_scores, extra_count, missing_count).
+    """
+    field_scores: list[FieldScore] = []
+    matched_actuals: set[int] = set()
+
+    for exp in expected_records:
+        # Find best match among unmatched actuals
+        candidates = [
+            (i, a) for i, a in enumerate(actual_records) if i not in matched_actuals
+        ]
+        if not candidates:
+            field_scores.append(FieldScore(
+                expected=exp, best_match=None, matched_fields={}, similarity=0.0,
+            ))
+            continue
+
+        best_idx, best_match = None, None
+        best_sim = 0.0
+        for idx, actual in candidates:
+            _, sim = _find_best_match(exp, [actual])
+            if sim > best_sim:
+                best_sim = sim
+                best_idx = idx
+                best_match = actual
+
+        if best_sim >= 0.5 and best_match is not None:
+            matched_actuals.add(best_idx)
+            matched_fields = _compare_fields(exp, best_match)
+            field_scores.append(FieldScore(
+                expected=exp, best_match=best_match,
+                matched_fields=matched_fields, similarity=best_sim,
+            ))
+        else:
+            field_scores.append(FieldScore(
+                expected=exp, best_match=None, matched_fields={}, similarity=0.0,
+            ))
+
+    extra_count = len(actual_records) - len(matched_actuals)
+    missing_count = sum(1 for s in field_scores if s.best_match is None)
+
+    return field_scores, extra_count, missing_count
+
+
+def compute_precision_recall(
+    expected_count: int,
+    actual_count: int,
+    matched_count: int,
+) -> tuple[float, float, float]:
+    """Compute precision, recall, F1."""
+    precision = matched_count / actual_count if actual_count > 0 else 0.0
+    recall = matched_count / expected_count if expected_count > 0 else 0.0
+    f1 = (
+        2 * precision * recall / (precision + recall)
+        if (precision + recall) > 0
+        else 0.0
+    )
+    return precision, recall, f1
+
+
+# ── LLM Judge Scorer ─────────────────────────────────────────────────────
+
+_JUDGE_SYSTEM_PROMPT = """\
+You are an evaluation judge for a data extraction system.
+
+Your task is to compare the EXPECTED extractions against the ACTUAL extractions
+produced by an AI agent, and assess quality on a 0-1 scale.
+
+Scoring criteria:
+- 1.0: All expected records found with correct fields, no significant extras
+- 0.8: Most expected records found, minor field differences or extras
+- 0.6: Core extractions present but some missing or incorrect
+- 0.4: Partial match — several expected records missing or wrong
+- 0.2: Poor quality — most expected records missing or incorrect
+- 0.0: Complete failure — no meaningful overlap
+
+Consider semantic equivalence: "Send invoice" and "Email the invoice" are matches.
+Ignore field ordering and formatting differences.
+
+Respond with ONLY a JSON object:
+{"score": 0.85, "reasoning": "Brief explanation of the score"}
+"""
+
+
+async def llm_judge_score(
+    expected: list[dict[str, Any]],
+    actual: list[dict[str, Any]],
+    *,
+    judge_model: str = "gpt-4o-mini",
+) -> tuple[float, str]:
+    """Use an LLM to semantically evaluate extraction quality.
+
+    Returns (score, reasoning).
+    """
+    from shared.llm import get_llm
+
+    llm = get_llm(model=judge_model, temperature=0)
+
+    user_content = (
+        f"## Expected extractions\n```json\n{json.dumps(expected, indent=2, default=str)}\n```\n\n"
+        f"## Actual extractions\n```json\n{json.dumps(actual, indent=2, default=str)}\n```"
+    )
+
+    try:
+        response = await llm.ainvoke([
+            SystemMessage(content=_JUDGE_SYSTEM_PROMPT),
+            HumanMessage(content=user_content),
+        ])
+        raw = response.content.strip()
+        if raw.startswith("```"):
+            raw = raw.split("```")[1]
+            if raw.startswith("json"):
+                raw = raw[4:]
+        parsed = json.loads(raw.strip())
+        return float(parsed.get("score", 0.0)), str(parsed.get("reasoning", ""))
+    except Exception as exc:
+        logger.warning("eval: LLM judge failed: %s", exc)
+        return 0.0, f"Judge error: {exc}"

services/batch-agent/requirements.txt

@@ -14,6 +14,7 @@ langchain-litellm>=0.3.0
 litellm>=1.50.0
 openai>=1.50.0
 httpx>=0.27.0
+langfuse>=3.0.0
 croniter>=2.0.0
 google-api-python-client>=2.130.0
 google-auth>=2.30.0

services/chat/app/agents/__init__.py

@@ -1 +0,0 @@
-"""Chat Service domain agents."""

services/chat/app/deep_agent.py

@@ -16,13 +16,13 @@ from typing import Any, Literal
 from langchain_core.messages import AIMessage, HumanMessage, SystemMessage, ToolMessage
 from langchain_core.tools import tool
 
-from app.agents.note_agent import NOTE_TOOLS
+from shared.agents.note_agent import NOTE_TOOLS
-from app.agents.project_agent import PROJECT_TOOLS
+from shared.agents.project_agent import PROJECT_TOOLS
-from app.agents.task_agent import TASK_TOOLS
+from shared.agents.task_agent import TASK_TOOLS
-from app.agents.timeline_agent import TIMELINE_TOOLS
+from shared.agents.timeline_agent import TIMELINE_TOOLS
-from app.llm import get_llm
+from shared.llm import get_llm
 from app.memory_middleware import MemoryMiddleware
-from app.ws_context import clear_tool_result_collector, execute_on_client, set_tool_result_collector
+from shared.ws_context import clear_tool_result_collector, execute_on_client, set_tool_result_collector
 from app import tracing
 from shared.db import async_session
 

services/chat/app/llm.py

@@ -33,6 +33,8 @@ def _api_key_for_model(model: str) -> str | None:
         return settings.GOOGLE_API_KEY or None
     if model.startswith("cerebras/"):
         return settings.CEREBRAS_API_KEY or None
+    if model.startswith("github/"):
+        return settings.GITHUB_TOKEN or None
     if model.startswith("github_copilot/"):
         return None
     return settings.OPENAI_API_KEY or None
@@ -49,6 +51,9 @@ def get_llm(
     if settings.GITHUB_COPILOT_TOKEN_DIR:
         os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
 
+    if settings.GITHUB_TOKEN:
+        os.environ.setdefault("GITHUB_TOKEN", settings.GITHUB_TOKEN)
+
     if "/" in model:
         return ChatLiteLLM(model=model, temperature=temperature, callbacks=callbacks)
 

services/chat/app/redis_consumer.py

@@ -17,7 +17,7 @@ from shared.redis import redis_client, ws_out_channel
 from app.deep_agent import run_floating_stream, run_home_stream
 from app.memory_middleware import MemoryMiddleware
 from app.output_formatter import StreamFormatter
-from app.ws_context import clear_current_user, set_current_user
+from shared.ws_context import clear_current_user, set_current_user
 from app import tracing
 
 logger = logging.getLogger(__name__)

services/chat/app/routes.py

@@ -8,7 +8,7 @@ from fastapi.responses import JSONResponse
 from shared.schemas import ChatRequest
 
 from app.deep_agent import run_home
-from app.ws_context import clear_current_user, set_current_user
+from shared.ws_context import clear_current_user, set_current_user
 
 router = APIRouter(prefix="/chat", tags=["chat"])
 

services/chat/app/tracing.py

@@ -167,9 +167,9 @@ def get_prompt(
     fallback: str | None = None,
     cache_ttl_seconds: int = 300,
 ) -> str | None:
-    """Fetch a managed prompt from Langfuse by name.
+    """Fetch a managed prompt from Langfuse by name (without variable compilation).
 
-    Returns the compiled prompt string, or *fallback* if the prompt is not
+    Returns the raw prompt string, or *fallback* if the prompt is not
     found or Langfuse is disabled.
     """
     lf = _get_client()
@@ -192,6 +192,46 @@ def get_prompt(
         return fallback
 
 
+def compile_prompt(
+    name: str,
+    *,
+    fallback: str,
+    variables: dict[str, str],
+    version: int | None = None,
+    label: str | None = None,
+    cache_ttl_seconds: int = 300,
+) -> str:
+    """Fetch a managed prompt from Langfuse and compile it with ``{{variables}}``.
+
+    If the prompt exists in Langfuse, uses the SDK's ``.compile(**variables)``
+    which replaces ``{{key}}`` placeholders.  If Langfuse is disabled or the
+    prompt is not found, falls back to ``fallback.format(**variables)`` (Python
+    ``{key}`` placeholders).
+
+    This means:
+      - Langfuse prompts use ``{{variable}}`` syntax.
+      - Hardcoded fallback strings use Python ``{variable}`` syntax.
+    """
+    lf = _get_client()
+    if lf is None:
+        return fallback.format(**variables)
+
+    try:
+        kwargs: dict[str, Any] = {
+            "name": name,
+            "cache_ttl_seconds": cache_ttl_seconds,
+        }
+        if version is not None:
+            kwargs["version"] = version
+        if label is not None:
+            kwargs["label"] = label
+        prompt = lf.get_prompt(**kwargs)
+        return prompt.compile(**variables)
+    except Exception as exc:
+        logger.warning("tracing: compile_prompt(%s) failed, using fallback: %s", name, exc)
+        return fallback.format(**variables)
+
+
 def link_prompt_to_trace(
     span: Any,
     prompt_name: str,

services/chat/app/ws_context.py

@@ -1,115 +0,0 @@
-"""WebSocket context for Chat Service — Redis-based tool call round-trip.
-
-Replaces the monolith's ws_context.py. Instead of calling Electron directly
-via WebSocket, this publishes tool_call frames to Redis (ws:out:{user_id})
-and awaits the result via BRPOP on tool:result:{call_id}.
-"""
-
-from __future__ import annotations
-
-import json
-import logging
-from contextvars import ContextVar
-from typing import Any
-from uuid import uuid4
-
-from shared.redis import redis_client, tool_result_key, ws_out_channel
-
-logger = logging.getLogger(__name__)
-
-_TOOL_CALL_TIMEOUT = 30  # seconds — BRPOP timeout
-
-# Per-request user_id context var (set before agent runs)
-_current_user_id: ContextVar[str | None] = ContextVar("_current_user_id", default=None)
-
-# Optional collector for debug
-_tool_result_collector: ContextVar[list[dict] | None] = ContextVar(
-    "_tool_result_collector", default=None
-)
-
-
-def set_current_user(user_id: str) -> None:
-    _current_user_id.set(user_id)
-
-
-def clear_current_user() -> None:
-    _current_user_id.set(None)
-
-
-def set_tool_result_collector(lst: list[dict]) -> None:
-    _tool_result_collector.set(lst)
-
-
-def clear_tool_result_collector() -> None:
-    _tool_result_collector.set(None)
-
-
-async def execute_on_client(
-    action: str,
-    table: str | None = None,
-    data: dict[str, Any] | None = None,
-    filters: dict[str, Any] | None = None,
-    vector: list[float] | None = None,
-    limit: int | None = None,
-) -> dict[str, Any]:
-    """Send a tool_call to Electron via Redis and await the result.
-
-    1. Build tool_call payload
-    2. Publish to ws:out:{user_id} (WS Gateway forwards to Electron)
-    3. BRPOP on tool:result:{call_id} (WS Gateway pushes when Electron replies)
-    4. Return result dict
-
-    Raises RuntimeError if no user_id is set or if the call times out.
-    """
-    user_id = _current_user_id.get()
-    if not user_id:
-        raise RuntimeError(
-            "execute_on_client() called without a user_id — "
-            "set_current_user() must be called first."
-        )
-
-    call_id = str(uuid4())
-    payload: dict[str, Any] = {
-        "type": "tool_call",
-        "id": call_id,
-        "action": action,
-    }
-    if table is not None:
-        payload["table"] = table
-    if data is not None:
-        payload["data"] = data
-    if filters is not None:
-        payload["filters"] = {k: v for k, v in filters.items() if v is not None}
-    if vector is not None:
-        payload["vector"] = vector
-    if limit is not None:
-        payload["limit"] = limit
-
-    # Publish tool_call to WS Gateway → Electron
-    channel = ws_out_channel(user_id)
-    await redis_client.publish(channel, json.dumps(payload))
-
-    # Wait for Electron's tool_result
-    result_key = tool_result_key(call_id)
-    response = await redis_client.brpop(result_key, timeout=_TOOL_CALL_TIMEOUT)
-
-    if response is None:
-        raise RuntimeError(
-            f"Tool call {call_id} timed out after {_TOOL_CALL_TIMEOUT}s — "
-            f"device may be offline or unresponsive."
-        )
-
-    # response is (key, value) tuple
-    _, raw = response
-    result = json.loads(raw)
-
-    # Collect for debug if requested
-    collector = _tool_result_collector.get(None)
-    if collector is not None:
-        collector.append({
-            "action": action,
-            "table": table,
-            "data": result,
-        })
-
-    return result

shared/agents/__init__.py

@@ -0,0 +1 @@
+"""Shared domain agents — tool definitions used by both Chat and Batch Agent services."""

shared/agents/note_agent.py

@@ -1,6 +1,6 @@
 """Note agent — Markdown note management (list, get, create, update, delete).
 
-Adapted for Chat Service: import from app.ws_context and app.llm.
+Shared tool definitions used by both Chat and Batch Agent services.
 """
 
 from __future__ import annotations
@@ -10,8 +10,8 @@ from typing import Any
 
 from langchain_core.tools import tool
 
-from app.llm import embed
+from shared.llm import embed
-from app.ws_context import execute_on_client
+from shared.ws_context import execute_on_client
 
 _UUID_RE = re.compile(
     r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"

shared/agents/project_agent.py

@@ -1,6 +1,6 @@
 """Project agent — full lifecycle management (list, get, create, update, archive, delete).
 
-Adapted for Chat Service: import from app.ws_context instead of app.core.ws_context.
+Shared tool definitions used by both Chat and Batch Agent services.
 """
 
 from __future__ import annotations
@@ -9,7 +9,7 @@ from typing import Any
 
 from langchain_core.tools import tool
 
-from app.ws_context import execute_on_client
+from shared.ws_context import execute_on_client
 
 PROJECT_SYSTEM_PROMPT = (
     "You are a project management assistant. You help users create, find,\n"

shared/agents/task_agent.py

@@ -1,6 +1,6 @@
 """Task agent — full CRUD for tasks and task comments.
 
-Adapted for Chat Service: import from app.ws_context instead of app.core.ws_context.
+Shared tool definitions used by both Chat and Batch Agent services.
 """
 
 from __future__ import annotations
@@ -11,7 +11,7 @@ from typing import Any
 
 from langchain_core.tools import tool
 
-from app.ws_context import execute_on_client
+from shared.ws_context import execute_on_client
 
 _UUID_RE = re.compile(
     r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
@@ -32,7 +32,6 @@ TASK_SYSTEM_PROMPT = (
     "  - project_id is optional; link to a project when the user mentions one\n"
     "  - is_ai_suggested: 1 only when proactively proposing a task the user\n"
     "    did not explicitly request; 0 otherwise\n"
-    "  - is_ai_suggested: 1 only when proactively proposing a task the user did not explicitly request; 0 otherwise\n"
     "  - Use list_tasks_due_today for 'what's due today' queries\n"
     "  - For update_task, use -1 for integer fields you do not want to change\n"
     "  - Always confirm the action in plain, user-friendly language."
@@ -225,7 +224,7 @@ async def delete_task_comment(comment_id: str) -> str:
     return f"Comment {comment_id} deleted."
 
 
-# ── Agent ─────────────────────────────────────────────────────────────
+# ── Exports ───────────────────────────────────────────────────────────
 
 
 TASK_TOOLS: list[Any] = [

shared/agents/timeline_agent.py

@@ -1,6 +1,6 @@
 """Timeline agent — project milestone management (list, create, update, delete).
 
-Adapted for Chat Service: import from app.ws_context instead of app.core.ws_context.
+Shared tool definitions used by both Chat and Batch Agent services.
 """
 
 from __future__ import annotations
@@ -10,7 +10,7 @@ from typing import Any
 
 from langchain_core.tools import tool
 
-from app.ws_context import execute_on_client
+from shared.ws_context import execute_on_client
 
 _UUID_RE = re.compile(
     r"^[0-9a-fA-F]{8}-[0-9a-fA-F]{4}-[1-5][0-9a-fA-F]{3}-[89abAB][0-9a-fA-F]{3}-[0-9a-fA-F]{12}$"
@@ -28,7 +28,6 @@ TIMELINE_SYSTEM_PROMPT = (
     "  - For listing, project_id must be a UUID; never pass plain names as project_id\n"
     "  - date is a Unix timestamp in milliseconds; convert human-readable dates\n"
     "  - is_ai_suggested: 1 when proactively proposing a timeline, 0 otherwise\n"
-    "  - is_ai_suggested: 1 when proactively proposing a timeline, 0 otherwise\n"
     "  - For update_timeline, use -1 for integer fields you do not want to change\n"
     "  - Listing without a project_id returns all timelines across projects\n"
     "  - Always echo the title and formatted date in your confirmation."

shared/config.py

@@ -62,6 +62,7 @@ class Settings(BaseSettings):
     ANTHROPIC_API_KEY: str = ""
     GOOGLE_API_KEY: str = ""
     CEREBRAS_API_KEY: str = ""
+    GITHUB_TOKEN: str = ""
 
     LLM_MODEL: str = "gpt-4o"
     LLM_EMBED_MODEL: str = "text-embedding-3-small"

shared/llm.py

@@ -0,0 +1,77 @@
+"""LLM factory — centralised model instantiation via LiteLLM.
+
+Shared by Chat and Batch Agent services.
+Uses shared.config.settings for all configuration.
+"""
+
+from __future__ import annotations
+
+import os
+import warnings
+
+from openai import AsyncOpenAI
+import litellm
+
+from langchain_openai import ChatOpenAI
+from langchain_litellm import ChatLiteLLM
+
+from shared.config import settings
+
+litellm.drop_params = True
+
+warnings.filterwarnings(
+    "ignore",
+    message=r"PydanticSerializationUnexpectedValue\(Expected `ResponseAPIUsage`",
+    category=UserWarning,
+)
+
+
+def _api_key_for_model(model: str) -> str | None:
+    if model.startswith("anthropic/"):
+        return settings.ANTHROPIC_API_KEY or None
+    if model.startswith("gemini/") or model.startswith("google/"):
+        return settings.GOOGLE_API_KEY or None
+    if model.startswith("cerebras/"):
+        return settings.CEREBRAS_API_KEY or None
+    if model.startswith("github/"):
+        return settings.GITHUB_TOKEN or None
+    if model.startswith("github_copilot/"):
+        return None
+    return settings.OPENAI_API_KEY or None
+
+
+def get_llm(
+    *,
+    model: str | None = None,
+    temperature: float = 0,
+    callbacks: list | None = None,
+) -> ChatOpenAI | ChatLiteLLM:
+    model = model or settings.LLM_MODEL
+
+    if settings.GITHUB_COPILOT_TOKEN_DIR:
+        os.environ.setdefault("GITHUB_COPILOT_TOKEN_DIR", settings.GITHUB_COPILOT_TOKEN_DIR)
+
+    if settings.GITHUB_TOKEN:
+        os.environ.setdefault("GITHUB_TOKEN", settings.GITHUB_TOKEN)
+
+    if "/" in model:
+        return ChatLiteLLM(model=model, temperature=temperature, callbacks=callbacks)
+
+    return ChatOpenAI(
+        model=model,
+        temperature=temperature,
+        api_key=_api_key_for_model(model),
+        callbacks=callbacks,
+    )
+
+
+async def embed(text: str) -> list[float]:
+    model = settings.LLM_EMBED_MODEL
+
+    if model.startswith("github_copilot/") or "/" in model:
+        response = await litellm.aembedding(model=model, input=[text])
+        return response.data[0]["embedding"]
+
+    client = AsyncOpenAI(api_key=settings.OPENAI_API_KEY)
+    response = await client.embeddings.create(model=model, input=text)
+    return response.data[0].embedding

shared/ws_context.py

@@ -1,12 +1,12 @@
-"""WebSocket context for Batch Agent Service — Redis-based tool call round-trip.
+"""WebSocket context — Redis-based tool call round-trip.
 
-Same pattern as services/chat/app/ws_context.py: publishes tool_call frames
+Shared by Chat and Batch Agent services.  Publishes tool_call frames to
-to Redis ws:out:{user_id} and awaits BRPOP on tool:result:{call_id}.
+Redis ``ws:out:{user_id}`` and awaits the result via BRPOP on
+``tool:result:{call_id}``.
 
-Additionally provides set_client_executor / clear_client_executor stubs
+Also provides ``set_client_executor`` / ``clear_client_executor`` no-op
-for backward compatibility with the agent_runner code (which originally
+shims for backward compatibility with agent_runner code that originally
-used a DeviceConnectionManager callback).  In the microservice world these
+used a DeviceConnectionManager callback.
-are no-ops — execute_on_client() always uses the Redis path.
 """
 
 from __future__ import annotations
@@ -23,10 +23,10 @@ logger = logging.getLogger(__name__)
 
 _TOOL_CALL_TIMEOUT = 30  # seconds — BRPOP timeout
 
-# Per-request user_id context var (set before agent run)
+# Per-request user_id context var (set before agent runs)
 _current_user_id: ContextVar[str | None] = ContextVar("_current_user_id", default=None)
 
-# Optional collector for debug / logging
+# Optional collector for debug
 _tool_result_collector: ContextVar[list[dict] | None] = ContextVar(
     "_tool_result_collector", default=None
 )
@@ -51,17 +51,14 @@ def clear_tool_result_collector() -> None:
 # ── Compatibility shims ──────────────────────────────────────────────────
 # agent_runner.py originally called set_client_executor / clear_client_executor
 # with a DeviceConnectionManager callback.  In the microservice world the
-# Redis-based execute_on_client replaces this, so these are no-ops that
+# Redis-based execute_on_client replaces this, so these are no-ops.
-# keep the agent_runner code unchanged.
 
 def set_client_executor(fn: Callable[[dict], Coroutine[Any, Any, dict]] | None) -> None:
     """No-op — kept for agent_runner compatibility."""
-    pass
 
 
 def clear_client_executor() -> None:
     """No-op — kept for agent_runner compatibility."""
-    pass
 
 
 async def execute_on_client(