wire llms, ui tweaks

2026-04-12 11:32:36 -03:00
parent 4de44baf98
commit 0f122fa8f7
22 changed files with 960 additions and 203 deletions
--- a/.env.example
+++ b/.env.example
@@ -1,11 +1,24 @@
-# AWS Bedrock (production pattern)
+# LLM Provider (groq | anthropic | bedrock | openai)
+LLM_PROVIDER=groq
+
+# Groq (default — fast, free tier)
+GROQ_API_KEY=gsk_...
+GROQ_MODEL=llama-3.3-70b-versatile
+
+# Anthropic (direct SDK)
+ANTHROPIC_API_KEY=sk-ant-...
+ANTHROPIC_MODEL=claude-sonnet-4-20250514
+
+# AWS Bedrock
 AWS_ACCESS_KEY_ID=AKIA...
 AWS_SECRET_ACCESS_KEY=...
 AWS_DEFAULT_REGION=us-east-1
-USE_BEDROCK=false
+BEDROCK_MODEL_ID=anthropic.claude-sonnet-4-20250514-v1:0

-# Anthropic (fallback for local dev)
-ANTHROPIC_API_KEY=sk-ant-...
+# OpenAI-compatible (any endpoint)
+OPENAI_API_KEY=sk-...
+OPENAI_BASE_URL=https://api.openai.com/v1
+OPENAI_MODEL=gpt-4o

 # Kong Konnect (optional)
 KONG_PROXY_URL=
--- a/agents/fce.py
+++ b/agents/fce.py
@@ -148,10 +148,25 @@ async def run_fce(
    }

    t0 = time.time()
-    notification_text = await mcp.call_tool("passenger", "generate_notification", {"context": context})
+    raw_result = await mcp.call_tool("passenger", "generate_notification", {"context": context})
    latency = int((time.time() - t0) * 1000)
    await emit("tool_call_end", tool="generate_notification", latency_ms=latency, is_live=False)

+    # Parse structured response (text + provider)
+    llm_provider = "template"
+    if isinstance(raw_result, dict) and "text" in raw_result:
+        notification_text = raw_result["text"]
+        llm_provider = raw_result.get("provider", "unknown")
+    elif isinstance(raw_result, str):
+        try:
+            parsed = json.loads(raw_result)
+            notification_text = parsed.get("text", raw_result)
+            llm_provider = parsed.get("provider", "unknown")
+        except (json.JSONDecodeError, TypeError):
+            notification_text = raw_result
+    else:
+        notification_text = str(raw_result)
+
    await emit("node_exit", node="synthesize")

    # ── Node 4: Format Output ──
@@ -164,14 +179,15 @@ async def run_fce(
    if airport_status:
        data_sources.append("faa_status_live")
    if crew_notes:
-        data_sources.append("get_crew_notes")
+        data_sources.append("crew_notes")

    notification = {
        "flight_id": flight_id,
        "type": "DELAY_NOTIFICATION" if status == "DELAYED" else f"{status}_NOTIFICATION",
        "status": status,
        "delay_minutes": delay_minutes,
-        "notification_text": notification_text if isinstance(notification_text, str) else str(notification_text),
+        "notification_text": notification_text,
+        "llm_provider": llm_provider,
        "generated_at": datetime.now(timezone.utc).isoformat(),
        "data_sources": data_sources,
        "human_approved": True,  # auto-approve in demo
--- a/agents/handover.py
+++ b/agents/handover.py
@@ -242,7 +242,7 @@ async def run_handover(
    hub_label = ", ".join(target_hubs) if len(target_hubs) < 5 else "ALL HUBS"

    t0 = time.time()
-    brief_text = await mcp.call_tool("ops", "generate_narrative", {
+    raw_result = await mcp.call_tool("ops", "generate_narrative", {
        "context": {
            "hub": hub_label,
            "shift_time": shift_time,
@@ -254,6 +254,22 @@ async def run_handover(
    latency = int((time.time() - t0) * 1000)
    await emit("tool_call_end", tool="generate_narrative", latency_ms=latency, is_live=False)

+    # Parse structured response
+    import json as _json
+    llm_provider = "template"
+    if isinstance(raw_result, dict) and "text" in raw_result:
+        brief_text = raw_result["text"]
+        llm_provider = raw_result.get("provider", "unknown")
+    elif isinstance(raw_result, str):
+        try:
+            parsed = _json.loads(raw_result)
+            brief_text = parsed.get("text", raw_result)
+            llm_provider = parsed.get("provider", "unknown")
+        except (_json.JSONDecodeError, TypeError):
+            brief_text = raw_result
+    else:
+        brief_text = str(raw_result)
+
    await emit("node_exit", node="synthesize")

    # ── Node 4: Format Output ──
@@ -263,7 +279,8 @@ async def run_handover(
    result = {
        "type": "HANDOVER_BRIEF",
        "hubs": target_hubs,
-        "brief_text": brief_text if isinstance(brief_text, str) else str(brief_text),
+        "brief_text": brief_text,
+        "llm_provider": llm_provider,
        "summary": {
            "immediate_count": len(immediate),
            "monitor_count": len(monitor),
--- a/agents/shared/mcp_client.py
+++ b/agents/shared/mcp_client.py
@@ -11,20 +11,38 @@ from typing import Any
 from fastmcp import Client


-# Server configurations for stdio transport
-SERVERS = {
-    "shared": {
+def _env() -> dict:
+    """Forward LLM-related env vars to MCP server subprocesses."""
+    import os
+
+    env = {}
+    for key in (
+        "LLM_PROVIDER", "GROQ_API_KEY", "GROQ_MODEL",
+        "ANTHROPIC_API_KEY", "ANTHROPIC_MODEL",
+        "OPENAI_API_KEY", "OPENAI_BASE_URL", "OPENAI_MODEL",
+        "AWS_ACCESS_KEY_ID", "AWS_SECRET_ACCESS_KEY", "AWS_DEFAULT_REGION",
+        "BEDROCK_MODEL_ID", "USE_BEDROCK",
+        "PATH",
+    ):
+        val = os.getenv(key)
+        if val:
+            env[key] = val
+    return env
+
+
+def _server_config(module: str) -> dict:
+    """Build server config with current env vars (called at connect time, not import time)."""
+    return {
        "command": "uv",
-        "args": ["run", "python", "-m", "mcp_servers.shared"],
-    },
-    "ops": {
-        "command": "uv",
-        "args": ["run", "python", "-m", "mcp_servers.ops"],
-    },
-    "passenger": {
-        "command": "uv",
-        "args": ["run", "python", "-m", "mcp_servers.passenger"],
-    },
+        "args": ["run", "python", "-m", module],
+        "env": _env(),
+    }
+
+
+SERVER_MODULES = {
+    "shared": "mcp_servers.shared",
+    "ops": "mcp_servers.ops",
+    "passenger": "mcp_servers.passenger",
 }

 # Agent profiles — which servers each agent connects to
@@ -43,9 +61,9 @@ class MCPMultiClient:
    async def connect(self, server_names: list[str]) -> None:
        """Connect to the specified MCP servers."""
        for name in server_names:
-            if name not in SERVERS:
-                raise ValueError(f"Unknown server: {name}. Available: {list(SERVERS.keys())}")
-            config = {"mcpServers": {"default": SERVERS[name]}}
+            if name not in SERVER_MODULES:
+                raise ValueError(f"Unknown server: {name}. Available: {list(SERVER_MODULES.keys())}")
+            config = {"mcpServers": {"default": _server_config(SERVER_MODULES[name])}}
            client = Client(config)
            await client.__aenter__()
            self._clients[name] = client
--- a/api/main.py
+++ b/api/main.py
@@ -165,6 +165,78 @@ async def set_active_scenario(req: ScenarioUpdate):
        return {"error": str(e)}


+# ── LLM config routes ──
+
+@app.get("/config/llm")
+async def get_llm_config():
+    """Current LLM provider configuration."""
+    import os
+    provider = os.getenv("LLM_PROVIDER", "groq")
+    return {
+        "provider": provider,
+        "providers": {
+            "groq": {
+                "configured": bool(os.getenv("GROQ_API_KEY")),
+                "model": os.getenv("GROQ_MODEL", "llama-3.3-70b-versatile"),
+            },
+            "anthropic": {
+                "configured": bool(os.getenv("ANTHROPIC_API_KEY")),
+                "model": os.getenv("ANTHROPIC_MODEL", "claude-sonnet-4-20250514"),
+            },
+            "bedrock": {
+                "configured": bool(os.getenv("AWS_ACCESS_KEY_ID")),
+                "model": os.getenv("BEDROCK_MODEL_ID", "anthropic.claude-sonnet-4-20250514-v1:0"),
+            },
+            "openai": {
+                "configured": bool(os.getenv("OPENAI_API_KEY")),
+                "model": os.getenv("OPENAI_MODEL", "gpt-4o"),
+                "base_url": os.getenv("OPENAI_BASE_URL", ""),
+            },
+            "template": {
+                "configured": True,
+                "model": "none (structured fallback)",
+            },
+        },
+    }
+
+
+class LLMConfigUpdate(BaseModel):
+    provider: str
+    api_key: str | None = None
+    model: str | None = None
+    base_url: str | None = None
+
+
+@app.put("/config/llm")
+async def set_llm_config(req: LLMConfigUpdate):
+    """Switch LLM provider at runtime. Sets env vars for MCP subprocesses."""
+    import os
+
+    os.environ["LLM_PROVIDER"] = req.provider
+
+    if req.provider == "groq" and req.api_key:
+        os.environ["GROQ_API_KEY"] = req.api_key
+        if req.model:
+            os.environ["GROQ_MODEL"] = req.model
+    elif req.provider == "anthropic" and req.api_key:
+        os.environ["ANTHROPIC_API_KEY"] = req.api_key
+        if req.model:
+            os.environ["ANTHROPIC_MODEL"] = req.model
+    elif req.provider == "openai" and req.api_key:
+        os.environ["OPENAI_API_KEY"] = req.api_key
+        if req.model:
+            os.environ["OPENAI_MODEL"] = req.model
+        if req.base_url:
+            os.environ["OPENAI_BASE_URL"] = req.base_url
+    elif req.provider == "bedrock":
+        if req.model:
+            os.environ["BEDROCK_MODEL_ID"] = req.model
+
+    # No need to update server configs — _server_config() reads env at connect time
+
+    return await get_llm_config()
+
+
 # ── Scenario data routes ──

@app.get("/scenarios/data/flights")
--- a/ctrl/k8s/base/configmap.yaml
+++ b/ctrl/k8s/base/configmap.yaml
@@ -5,5 +5,7 @@ metadata:
  namespace: unt
 data:
  DEFAULT_SCENARIO: "weather_disruption_ord"
-  USE_BEDROCK: "false"
+  LLM_PROVIDER: "groq"
+  GROQ_API_KEY: "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"
+  GROQ_MODEL: "llama-3.3-70b-versatile"
  LANGFUSE_HOST: "http://langfuse:3000"
--- a/ctrl/nginx.conf
+++ b/ctrl/nginx.conf
@@ -18,6 +18,10 @@ server {
        proxy_pass http://api:8000;
    }

+    location /config {
+        proxy_pass http://api:8000;
+    }
+
    location /ws/ {
        proxy_pass http://api:8000;
        proxy_http_version 1.1;
--- a/mcp_servers/ops/server.py
+++ b/mcp_servers/ops/server.py
@@ -100,17 +100,36 @@ def get_pending_rebookings(hub: str, limit: int = 20) -> list[dict]:


@mcp.tool()
-def generate_narrative(context: dict) -> str:
+async def generate_narrative(context: dict) -> str:
    """Synthesizes aggregated operational context into a structured
    handover brief for ops managers.

-    Uses Claude Sonnet via AWS Bedrock Converse API.
+    Uses Claude via Anthropic SDK (or Bedrock when USE_BEDROCK=true).
    Output: prioritized, concise, structured by IMMEDIATE / MONITOR / FYI.
-
-    NOTE: In v1, this returns a structured template from the context data.
-    LLM integration will be added when Bedrock is wired up.
+    Falls back to template if no API key is configured.
    """
-    # V1: structured template — will be replaced with Bedrock call
+    try:
+        from mcp_servers.shared_llm import generate, _get_provider
+
+        hub = context.get("hub", "ALL")
+        shift_time = context.get("shift_time", datetime.now(timezone.utc).strftime("%H:%M UTC"))
+
+        system_prompt = (
+            f"You are an airline operations shift handover briefing system. "
+            f"Generate a concise handover brief for {hub} at {shift_time}. "
+            f"Structure as: HEADER, then IMMEDIATE ACTION (items needing action within 2h), "
+            f"MONITOR (items that could escalate), FYI (resolved or low-risk). "
+            f"Be concise — ops managers scan, they don't read paragraphs. "
+            f"Use the data provided. Do not invent details."
+        )
+        text = await generate(system_prompt, json.dumps(context, indent=2))
+        return json.dumps({"text": text, "provider": _get_provider()})
+    except Exception:
+        return json.dumps({"text": _template_narrative(context), "provider": "template"})
+
+
+def _template_narrative(context: dict) -> str:
+    """Structured template fallback when LLM is unavailable."""
    sections = []
    immediate = context.get("immediate", [])
    monitor = context.get("monitor", [])
@@ -119,8 +138,7 @@ def generate_narrative(context: dict) -> str:
    hub = context.get("hub", "ALL")
    shift_time = context.get("shift_time", datetime.now(timezone.utc).strftime("%H:%M UTC"))

-    header = f"SHIFT HANDOVER BRIEF — {hub} / {shift_time}"
-    sections.append(header)
+    sections.append(f"SHIFT HANDOVER BRIEF — {hub} / {shift_time}")
    sections.append(f"Generated: {datetime.now(timezone.utc).strftime('%H:%M UTC')}")
    sections.append("")

--- a/mcp_servers/passenger/server.py
+++ b/mcp_servers/passenger/server.py
@@ -5,7 +5,6 @@ notification prompt template (multi-tone).
 """

 import json
-from datetime import datetime, timezone

 from fastmcp import FastMCP

@@ -25,17 +24,33 @@ mcp = FastMCP(


@mcp.tool()
-def generate_notification(context: dict) -> str:
+async def generate_notification(context: dict) -> str:
    """Synthesizes flight disruption context into an empathetic,
    actionable passenger notification.

-    Uses Claude Sonnet via AWS Bedrock Converse API.
+    Uses Claude via Anthropic SDK (or Bedrock when USE_BEDROCK=true).
    Output: clear, human, no jargon, includes gate/time/status.
-
-    NOTE: In v1, this returns a structured template from the context data.
-    LLM integration will be added when Bedrock is wired up.
+    Falls back to template if no API key is configured.
    """
-    # V1: structured template — will be replaced with Bedrock call
+    try:
+        from mcp_servers.shared_llm import generate, _get_provider
+
+        system_prompt = (
+            "You are a passenger notification system for Stellar Air. "
+            "Write a clear, empathetic notification about this flight disruption. "
+            "Explain WHY the delay or cancellation happened using the operational data provided. "
+            "Tell the passenger what's happening next: new boarding time, gate, status. "
+            "Be human and reassuring. No aviation jargon. No speculation. "
+            "If data is missing for a section, omit it — don't make things up."
+        )
+        text = await generate(system_prompt, json.dumps(context, indent=2))
+        return json.dumps({"text": text, "provider": _get_provider()})
+    except Exception:
+        return json.dumps({"text": _template_notification(context), "provider": "template"})
+
+
+def _template_notification(context: dict) -> str:
+    """Structured template fallback when LLM is unavailable."""
    flight_id = context.get("flight_id", "")
    origin = context.get("origin", "")
    destination = context.get("destination", "")
--- a/mcp_servers/shared_llm.py
+++ b/mcp_servers/shared_llm.py
@@ -0,0 +1,108 @@
+"""Shared LLM client for MCP server narrative tools.
+
+Multi-provider support — selected via LLM_PROVIDER env var:
+
+  groq        (default) — Groq API, OpenAI-compatible. Fast, free tier.
+                           Needs GROQ_API_KEY.
+
+  anthropic   — Direct Anthropic SDK. Needs ANTHROPIC_API_KEY.
+
+  bedrock     — AWS Bedrock Converse API. Needs AWS credentials.
+
+  openai      — Any OpenAI-compatible endpoint.
+                Set OPENAI_API_KEY and optionally OPENAI_BASE_URL.
+
+Usage:
+    LLM_PROVIDER=groq GROQ_API_KEY=gsk_... python -m mcp_servers.shared
+"""
+
+import os
+from typing import Literal
+
+Provider = Literal["groq", "anthropic", "bedrock", "openai"]
+
+
+def _get_provider() -> Provider:
+    p = os.getenv("LLM_PROVIDER", "groq").lower()
+    if p in ("groq", "anthropic", "bedrock", "openai"):
+        return p
+    return "groq"
+
+
+async def generate(system_prompt: str, user_content: str, max_tokens: int = 1024) -> str:
+    """Call an LLM and return the text response."""
+    provider = _get_provider()
+
+    if provider == "anthropic":
+        return await _generate_anthropic(system_prompt, user_content, max_tokens)
+    elif provider == "bedrock":
+        return await _generate_bedrock(system_prompt, user_content, max_tokens)
+    else:
+        # groq, openai, or any OpenAI-compatible provider
+        return await _generate_openai_compat(system_prompt, user_content, max_tokens)
+
+
+async def _generate_openai_compat(
+    system_prompt: str, user_content: str, max_tokens: int
+) -> str:
+    """OpenAI-compatible API (Groq, OpenAI, local, etc)."""
+    import openai
+
+    provider = _get_provider()
+
+    if provider == "groq":
+        api_key = os.getenv("GROQ_API_KEY")
+        base_url = "https://api.groq.com/openai/v1"
+        model = os.getenv("GROQ_MODEL", "llama-3.3-70b-versatile")
+    else:
+        api_key = os.getenv("OPENAI_API_KEY")
+        base_url = os.getenv("OPENAI_BASE_URL", "https://api.openai.com/v1")
+        model = os.getenv("OPENAI_MODEL", "gpt-4o")
+
+    client = openai.AsyncOpenAI(api_key=api_key, base_url=base_url)
+    response = await client.chat.completions.create(
+        model=model,
+        max_tokens=max_tokens,
+        temperature=0.7,
+        messages=[
+            {"role": "system", "content": system_prompt},
+            {"role": "user", "content": user_content},
+        ],
+    )
+    return response.choices[0].message.content
+
+
+async def _generate_anthropic(
+    system_prompt: str, user_content: str, max_tokens: int
+) -> str:
+    import anthropic
+
+    client = anthropic.AsyncAnthropic()
+    response = await client.messages.create(
+        model=os.getenv("ANTHROPIC_MODEL", "claude-sonnet-4-20250514"),
+        max_tokens=max_tokens,
+        temperature=0.7,
+        system=system_prompt,
+        messages=[{"role": "user", "content": user_content}],
+    )
+    return response.content[0].text
+
+
+async def _generate_bedrock(
+    system_prompt: str, user_content: str, max_tokens: int
+) -> str:
+    import json
+
+    import boto3
+
+    bedrock = boto3.client(
+        "bedrock-runtime",
+        region_name=os.getenv("AWS_DEFAULT_REGION", "us-east-1"),
+    )
+    response = bedrock.converse(
+        modelId=os.getenv("BEDROCK_MODEL_ID", "anthropic.claude-sonnet-4-20250514-v1:0"),
+        system=[{"text": system_prompt}],
+        messages=[{"role": "user", "content": [{"text": user_content}]}],
+        inferenceConfig={"maxTokens": max_tokens, "temperature": 0.7},
+    )
+    return response["output"]["message"]["content"][0]["text"]
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -10,6 +10,7 @@ dependencies = [
    "langchain-anthropic",
    "boto3",
    "anthropic",
+    "openai",
    "fastapi",
    "uvicorn[standard]",
    "pydantic>=2.0",
--- a/tests/endpoints.py
+++ b/tests/endpoints.py
@@ -25,6 +25,9 @@ class Endpoints:
    def crew_notes(flight_id: str) -> str:
        return f"/scenarios/data/crew-notes/{flight_id}"

+    # Config
+    CONFIG_LLM = "/config/llm"
+
    # Agents
    AGENT_FCE = "/agents/fce"
    AGENT_HANDOVER = "/agents/handover"
--- a/tests/test_mcp_servers.py
+++ b/tests/test_mcp_servers.py
@@ -194,9 +194,15 @@ class TestPassengerServer:
                    "gate": "H14",
                }
            })
-            text = _parse_result(result)
+            data = _parse_result(result)
+            # Response is JSON with text + provider
+            if isinstance(data, dict):
+                assert "text" in data
+                assert "provider" in data
+                text = data["text"]
+            else:
+                text = str(data)
            assert "UA432" in text
-            assert "DELAYED" in text or "delayed" in text

    @pytest.mark.asyncio
    async def test_list_prompts(self):
--- a/ui/app/src/App.vue
+++ b/ui/app/src/App.vue
@@ -1,5 +1,5 @@
 <script setup lang="ts">
-import { ref, provide } from 'vue'
+import { ref, provide, computed } from 'vue'
 import { useRouter, useRoute } from 'vue-router'
 import ScenarioSelector from './components/ScenarioSelector.vue'

@@ -7,11 +7,37 @@ const router = useRouter()
 const route = useRoute()
 const scenarioVersion = ref(0)

+// View mode for the main ops page
+const showOps = ref(true)
+const showInternals = ref(false)
+
+function toggleOps() {
+  if (route.path !== '/') router.push('/')
+  if (showOps.value && showInternals.value) {
+    showOps.value = false
+  } else {
+    showOps.value = true
+  }
+}
+
+function toggleInternals() {
+  if (route.path !== '/') router.push('/')
+  if (showInternals.value && showOps.value) {
+    showInternals.value = false
+  } else {
+    showInternals.value = true
+  }
+}
+
 function onScenarioChange() {
  scenarioVersion.value++
 }

+const isMainPage = computed(() => route.path === '/')
+
 provide('scenarioVersion', scenarioVersion)
+provide('showOps', showOps)
+provide('showInternals', showInternals)
 </script>

 <template>
@@ -22,9 +48,10 @@ provide('scenarioVersion', scenarioVersion)
        <span class="app-subtitle">NOVA Operations Platform</span>
      </div>
      <nav class="app-nav">
-        <router-link to="/" :class="{ active: route.path === '/' }">Operations</router-link>
-        <router-link to="/internals" :class="{ active: route.path === '/internals' }">Internals</router-link>
+        <a href="#" :class="{ active: isMainPage && showOps }" @click.prevent="toggleOps">Operations</a>
+        <a href="#" :class="{ active: isMainPage && showInternals }" @click.prevent="toggleInternals">Internals</a>
        <router-link to="/data" :class="{ active: route.path === '/data' }">Data</router-link>
+        <router-link to="/settings" :class="{ active: route.path === '/settings' }">Settings</router-link>
        <a href="/docs/" class="docs-link" target="_blank">Docs</a>
      </nav>
      <ScenarioSelector @change="onScenarioChange" />
--- a/ui/app/src/components/HandoverBrief.vue
+++ b/ui/app/src/components/HandoverBrief.vue
@@ -7,6 +7,7 @@ defineProps<{
    generated_at: string
    duration_ms: number
    hubs: string[]
+    llm_provider?: string
  }
 }>()
 </script>
@@ -53,7 +54,12 @@ defineProps<{

    <div class="brief-footer">
      <span>Hubs: {{ data.hubs.join(', ') }}</span>
-      <span>{{ data.duration_ms }}ms</span>
+      <span>
+        <span v-if="data.llm_provider" :style="{ color: data.llm_provider === 'template' ? 'var(--text-dim)' : 'var(--status-live)' }">
+          {{ data.llm_provider }}
+        </span>
+        {{ data.duration_ms }}ms
+      </span>
    </div>
  </div>
 </template>
--- a/ui/app/src/components/NotificationCard.vue
+++ b/ui/app/src/components/NotificationCard.vue
@@ -9,6 +9,7 @@ defineProps<{
    generated_at: string
    human_approved: boolean
    duration_ms: number
+    llm_provider?: string
  }
 }>()

@@ -38,6 +39,9 @@ const causeColor: Record<string, string> = {
        </template>
      </span>
      <span class="meta">
+        <span v-if="data.llm_provider" :class="['provider-tag', data.llm_provider === 'template' ? 'mock' : 'live']">
+          {{ data.llm_provider }}
+        </span>
        {{ data.duration_ms }}ms
        <span v-if="data.human_approved" class="approved">approved</span>
      </span>
@@ -117,6 +121,24 @@ const causeColor: Record<string, string> = {
  font-family: var(--font-mono);
 }

+.provider-tag {
+  padding: 1px 6px;
+  font-family: var(--font-mono);
+  font-size: 10px;
+  background: var(--surface-2);
+  border: 1px solid var(--surface-3);
+  margin-right: 8px;
+}
+
+.provider-tag.live {
+  border-color: var(--status-live);
+  color: var(--status-live);
+}
+
+.provider-tag.mock {
+  color: var(--text-dim);
+}
+
 .approved {
  color: var(--status-live);
  margin-left: 8px;
--- a/ui/app/src/composables/useAgentEvents.ts
+++ b/ui/app/src/composables/useAgentEvents.ts
@@ -0,0 +1,112 @@
+import { ref, onUnmounted } from 'vue'
+
+export interface GraphNode {
+  id: string
+  status: string
+}
+
+export interface AgentRun {
+  agent: string
+  run_id: string
+}
+
+export interface LogEntry {
+  level: string
+  stage: string
+  msg: string
+  ts: string
+}
+
+export function useAgentEvents() {
+  const agentStatus = ref<'idle' | 'live' | 'processing' | 'error'>('idle')
+  const entries = ref<LogEntry[]>([])
+  const graphNodes = ref<GraphNode[]>([])
+  const currentRun = ref<AgentRun | null>(null)
+
+  let ws: WebSocket | null = null
+
+  function connect() {
+    if (ws) return
+    const protocol = location.protocol === 'https:' ? 'wss:' : 'ws:'
+    ws = new WebSocket(`${protocol}//${location.host}/ws/agent-events`)
+
+    ws.onopen = () => { agentStatus.value = 'live' }
+
+    ws.onmessage = (e) => {
+      const event = JSON.parse(e.data)
+      handleEvent(event)
+    }
+
+    ws.onclose = () => {
+      agentStatus.value = 'idle'
+      ws = null
+      setTimeout(connect, 3000)
+    }
+
+    ws.onerror = () => { agentStatus.value = 'error' }
+  }
+
+  function disconnect() {
+    ws?.close()
+    ws = null
+  }
+
+  function handleEvent(event: any) {
+    const ts = event.timestamp || new Date().toISOString()
+    const time = ts.split('T')[1]?.split('.')[0] || ts
+
+    switch (event.type) {
+      case 'agent_start':
+        currentRun.value = { agent: event.agent, run_id: event.run_id }
+        agentStatus.value = 'processing'
+        graphNodes.value = []
+        entries.value = [{
+          level: 'info', stage: 'system',
+          msg: `Agent ${event.agent} started (${event.run_id})`, ts: time,
+        }]
+        break
+
+      case 'node_enter':
+        graphNodes.value.push({ id: event.node, status: 'processing' })
+        entries.value.push({
+          level: 'info', stage: event.node,
+          msg: `→ entering ${event.node}`, ts: time,
+        })
+        break
+
+      case 'node_exit': {
+        const node = graphNodes.value.find(n => n.id === event.node)
+        if (node) node.status = 'done'
+        break
+      }
+
+      case 'tool_call_end': {
+        const liveTag = event.is_live ? ' (live)' : ' (mock)'
+        entries.value.push({
+          level: 'info', stage: '',
+          msg: `${event.tool} — ${event.latency_ms}ms ✓${liveTag}`, ts: time,
+        })
+        break
+      }
+
+      case 'tool_call_error':
+        entries.value.push({
+          level: 'error', stage: '',
+          msg: `${event.tool} — FAILED: ${event.error}`, ts: time,
+        })
+        break
+
+      case 'agent_end':
+        agentStatus.value = 'live'
+        entries.value.push({
+          level: 'info', stage: 'system',
+          msg: `Agent complete: ${event.output_summary}`, ts: time,
+        })
+        break
+    }
+  }
+
+  onUnmounted(disconnect)
+
+  return { agentStatus, entries, graphNodes, currentRun, connect, disconnect }
+}
--- a/ui/app/src/main.ts
+++ b/ui/app/src/main.ts
@@ -4,15 +4,15 @@ import 'soleprint-ui/src/tokens.css'
 import './styles/mars-tokens.css'
 import App from './App.vue'
 import OpsNotifications from './pages/OpsNotifications.vue'
-import AgentInternals from './pages/AgentInternals.vue'
 import ScenarioData from './pages/ScenarioData.vue'
+import Settings from './pages/Settings.vue'

 const router = createRouter({
  history: createWebHistory(),
  routes: [
    { path: '/', component: OpsNotifications },
-    { path: '/internals', component: AgentInternals },
    { path: '/data', component: ScenarioData },
+    { path: '/settings', component: Settings },
  ],
 })

--- a/ui/app/src/pages/AgentInternals.vue
+++ b/ui/app/src/pages/AgentInternals.vue
@@ -1,103 +1,11 @@
 <script setup lang="ts">
-import { ref, onMounted, onUnmounted, nextTick } from 'vue'
+import { onMounted } from 'vue'
 import { Panel, SplitPane, LogRenderer } from 'soleprint-ui'
-import type { LogEntry } from 'soleprint-ui'
+import { useAgentEvents } from '../composables/useAgentEvents'

-const agentStatus = ref<'idle' | 'live' | 'processing' | 'error'>('idle')
-const entries = ref<LogEntry[]>([])
-const graphNodes = ref<{ id: string; status: string }[]>([])
-const currentRun = ref<{ agent: string; run_id: string } | null>(null)
+const { agentStatus, entries, graphNodes, currentRun, connect } = useAgentEvents()

-let ws: WebSocket | null = null
-
-function connectWs() {
-  const protocol = location.protocol === 'https:' ? 'wss:' : 'ws:'
-  ws = new WebSocket(`${protocol}//${location.host}/ws/agent-events`)
-
-  ws.onopen = () => {
-    agentStatus.value = 'live'
-  }
-
-  ws.onmessage = (e) => {
-    const event = JSON.parse(e.data)
-    handleEvent(event)
-  }
-
-  ws.onclose = () => {
-    agentStatus.value = 'idle'
-    setTimeout(connectWs, 3000)
-  }
-
-  ws.onerror = () => {
-    agentStatus.value = 'error'
-  }
-}
-
-function handleEvent(event: any) {
-  const ts = event.timestamp || new Date().toISOString()
-  const time = ts.split('T')[1]?.split('.')[0] || ts
-
-  switch (event.type) {
-    case 'agent_start':
-      currentRun.value = { agent: event.agent, run_id: event.run_id }
-      agentStatus.value = 'processing'
-      graphNodes.value = []
-      entries.value = [{
-        level: 'info',
-        stage: 'system',
-        msg: `Agent ${event.agent} started (${event.run_id})`,
-        ts: time,
-      }]
-      break
-
-    case 'node_enter':
-      graphNodes.value.push({ id: event.node, status: 'processing' })
-      entries.value.push({
-        level: 'info',
-        stage: event.node,
-        msg: `→ entering ${event.node}`,
-        ts: time,
-      })
-      break
-
-    case 'node_exit':
-      const node = graphNodes.value.find(n => n.id === event.node)
-      if (node) node.status = 'done'
-      break
-
-    case 'tool_call_end':
-      const liveTag = event.is_live ? ' (live)' : ' (mock)'
-      entries.value.push({
-        level: 'info',
-        stage: '',
-        msg: `${event.tool} — ${event.latency_ms}ms ✓${liveTag}`,
-        ts: time,
-      })
-      break
-
-    case 'tool_call_error':
-      entries.value.push({
-        level: 'error',
-        stage: '',
-        msg: `${event.tool} — FAILED: ${event.error}`,
-        ts: time,
-      })
-      break
-
-    case 'agent_end':
-      agentStatus.value = 'live'
-      entries.value.push({
-        level: 'info',
-        stage: 'system',
-        msg: `Agent complete: ${event.output_summary}`,
-        ts: time,
-      })
-      break
-  }
-}
-
-onMounted(connectWs)
-onUnmounted(() => { ws?.close() })
+onMounted(connect)
 </script>

 <template>
@@ -111,7 +19,7 @@ onUnmounted(() => { ws?.close() })
            </div>
            <div v-else class="graph-nodes">
              <div
-                v-for="(node, i) in graphNodes"
+                v-for="node in graphNodes"
                :key="node.id"
                :class="['graph-node', node.status]"
              >
@@ -157,10 +65,7 @@ onUnmounted(() => { ws?.close() })
  min-height: 0;
 }

-.graph-container {
-  padding: 16px;
-  height: 100%;
-}
+.graph-container { padding: 16px; height: 100%; }

 .graph-nodes {
  display: flex;
@@ -208,14 +113,8 @@ onUnmounted(() => { ws?.close() })
  box-shadow: 0 0 8px var(--status-live);
 }

-.node-label {
-  font-family: var(--font-mono);
-  font-size: 13px;
-}
-
-.summary-panel {
-  flex-shrink: 0;
-}
+.node-label { font-family: var(--font-mono); font-size: 13px; }
+.summary-panel { flex-shrink: 0; }

 .summary {
  display: flex;
--- a/ui/app/src/pages/OpsNotifications.vue
+++ b/ui/app/src/pages/OpsNotifications.vue
@@ -1,8 +1,9 @@
 <script setup lang="ts">
 import { ref, onMounted, watch, inject } from 'vue'
-import { Panel } from 'soleprint-ui'
+import { Panel, SplitPane, LogRenderer } from 'soleprint-ui'
 import NotificationCard from '../components/NotificationCard.vue'
 import HandoverBrief from '../components/HandoverBrief.vue'
+import { useAgentEvents } from '../composables/useAgentEvents'

 const flights = ref<any[]>([])
 const selectedFlight = ref('')
@@ -12,6 +13,13 @@ const notification = ref<any>(null)
 const handoverBrief = ref<any>(null)
 const scenarioVersion = inject<any>('scenarioVersion')

+const showOps = inject<any>('showOps')
+const showInternals = inject<any>('showInternals')
+const { agentStatus, entries, graphNodes, currentRun, connect } = useAgentEvents()
+
+// Connect WebSocket immediately so we don't miss events
+onMounted(connect)
+
 watch(scenarioVersion, () => {
  loadFlights()
  notification.value = null
@@ -34,6 +42,7 @@ async function runFce() {
  if (!selectedFlight.value) return
  fceStatus.value = 'processing'
  notification.value = null
+  if (!showInternals.value) showInternals.value = true

  const res = await fetch('/agents/fce', {
    method: 'POST',
@@ -42,7 +51,6 @@ async function runFce() {
  })
  const { run_id } = await res.json()

-  // Poll for result
  const poll = setInterval(async () => {
    const r = await fetch(`/agents/runs/${run_id}`)
    const data = await r.json()
@@ -60,6 +68,7 @@ async function runFce() {
 async function runHandover() {
  handoverStatus.value = 'processing'
  handoverBrief.value = null
+  if (!showInternals.value) showInternals.value = true

  const res = await fetch('/agents/handover', {
    method: 'POST',
@@ -86,7 +95,9 @@ onMounted(loadFlights)
 </script>

 <template>
-  <div class="ops-page">
+  <div :class="['ops-layout', { split: showOps && showInternals }]">
+    <!-- Ops pane (left) -->
+    <div v-show="showOps" class="ops-pane">
      <Panel title="FCE — Behind Every Departure" :status="fceStatus">
        <template #actions>
          <select v-model="selectedFlight" class="flight-select">
@@ -96,16 +107,9 @@ onMounted(loadFlights)
            {{ fceStatus === 'processing' ? 'Running...' : 'Run FCE' }}
          </button>
        </template>
-
-      <div v-if="notification" class="result-area">
-        <NotificationCard :data="notification" />
-      </div>
-      <div v-else-if="fceStatus === 'processing'" class="loading">
-        Running agent... gathering flight data, weather, crew notes...
-      </div>
-      <div v-else class="empty">
-        Select a flight and click Run FCE to generate a notification.
-      </div>
+        <div v-if="notification" class="result-area"><NotificationCard :data="notification" /></div>
+        <div v-else-if="fceStatus === 'processing'" class="loading">Running agent...</div>
+        <div v-else class="empty">Select a flight and click Run FCE.</div>
      </Panel>

      <Panel title="Shift Handover Brief" :status="handoverStatus">
@@ -114,27 +118,83 @@ onMounted(loadFlights)
            {{ handoverStatus === 'processing' ? 'Running...' : 'Run Handover' }}
          </button>
        </template>
+        <div v-if="handoverBrief" class="result-area"><HandoverBrief :data="handoverBrief" /></div>
+        <div v-else-if="handoverStatus === 'processing'" class="loading">Running agent...</div>
+        <div v-else class="empty">Click Run Handover.</div>
+      </Panel>
+    </div>

-      <div v-if="handoverBrief" class="result-area">
-        <HandoverBrief :data="handoverBrief" />
+    <!-- Internals pane (right) -->
+    <div v-show="showInternals" class="internals-pane">
+      <Panel title="Agent Graph" :status="agentStatus">
+        <div class="graph-container">
+          <div v-if="graphNodes.length === 0" class="empty">Waiting for agent run...</div>
+          <div v-else class="graph-nodes">
+            <div v-for="node in graphNodes" :key="node.id" :class="['graph-node', node.status]">
+              <div class="node-dot"></div>
+              <span class="node-label">{{ node.id }}</span>
            </div>
-      <div v-else-if="handoverStatus === 'processing'" class="loading">
-        Running agent... scanning all hubs for active issues...
+            <div class="graph-edge-line"></div>
          </div>
-      <div v-else class="empty">
-        Click Run Handover to generate a shift handover brief.
        </div>
      </Panel>
+      <Panel title="Tool Call Stream" :status="agentStatus" class="stream-panel">
+        <LogRenderer :entries="entries" :auto-scroll="true" />
+      </Panel>
+      <div v-if="currentRun" class="run-summary">
+        {{ currentRun.agent }} / {{ currentRun.run_id }} / {{ entries.length }} events
+      </div>
+    </div>
  </div>
 </template>

 <style scoped>
-.ops-page {
+.ops-layout {
+  display: flex;
+  gap: 16px;
+  height: calc(100vh - 80px);
+  position: relative;
+}
+
+/* Both visible: 50/50 with divider */
+.ops-layout.split > .ops-pane { flex: 1; }
+.ops-layout.split > .internals-pane { flex: 1; border-left: var(--panel-border); padding-left: 16px; }
+
+/* Single pane: full width */
+.ops-layout > .internals-pane { flex: 1; }
+.ops-layout > .ops-pane { flex: 1; }
+
+.ops-pane {
  display: flex;
  flex-direction: column;
  gap: 24px;
+  overflow: auto;
+  height: 100%;
+  min-width: 0;
 }

+.internals-pane {
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+  height: 100%;
+  overflow: auto;
+  min-width: 0;
+}
+
+.internals-pane > :first-child { flex-shrink: 0; }
+.internals-pane > .stream-panel { flex: 1; min-height: 0; }
+
+.run-summary {
+  padding: 4px 12px;
+  font-family: var(--font-mono);
+  font-size: 11px;
+  color: var(--text-dim);
+  flex-shrink: 0;
+}
+
+
+
 .flight-select {
  background: var(--surface-2);
  color: var(--text-primary);
@@ -152,15 +212,12 @@ onMounted(loadFlights)
  font-family: var(--font-mono);
  font-size: 12px;
  cursor: pointer;
-  transition: background 0.15s;
 }

 .run-btn:hover { background: var(--accent-dim); }
 .run-btn:disabled { opacity: 0.5; cursor: not-allowed; }

-.result-area {
-  padding: 16px;
-}
+.result-area { padding: 16px; }

 .loading, .empty {
  padding: 32px;
@@ -170,7 +227,58 @@ onMounted(loadFlights)
  font-size: 13px;
 }

-.loading {
-  color: var(--accent);
+.loading { color: var(--accent); }
+
+.graph-container { padding: 12px; }
+
+.graph-nodes {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+  position: relative;
+  padding-left: 12px;
+}
+
+.graph-edge-line {
+  position: absolute;
+  left: 17px;
+  top: 10px;
+  bottom: 10px;
+  width: 2px;
+  background: var(--surface-3);
+}
+
+.graph-node {
+  display: flex;
+  align-items: center;
+  gap: 10px;
+  padding: 6px 10px;
+  background: var(--surface-2);
+  border: var(--panel-border);
+  position: relative;
+  z-index: 1;
+}
+
+.node-dot {
+  width: 7px;
+  height: 7px;
+  border-radius: 50%;
+  background: var(--status-idle);
+  flex-shrink: 0;
+}
+
+.graph-node.processing .node-dot {
+  background: var(--status-processing);
+  box-shadow: 0 0 8px var(--status-processing);
+}
+
+.graph-node.done .node-dot {
+  background: var(--status-live);
+  box-shadow: 0 0 8px var(--status-live);
+}
+
+.node-label {
+  font-family: var(--font-mono);
+  font-size: 12px;
 }
 </style>
--- a/ui/app/src/pages/Settings.vue
+++ b/ui/app/src/pages/Settings.vue
@@ -0,0 +1,257 @@
+<script setup lang="ts">
+import { ref, onMounted } from 'vue'
+import { Panel } from 'soleprint-ui'
+
+const config = ref<any>(null)
+const selectedProvider = ref('')
+const apiKey = ref('')
+const model = ref('')
+const baseUrl = ref('')
+const saving = ref(false)
+
+async function loadConfig() {
+  const res = await fetch('/config/llm')
+  config.value = await res.json()
+  selectedProvider.value = config.value.provider
+  const p = config.value.providers[selectedProvider.value]
+  model.value = p?.model || ''
+  baseUrl.value = p?.base_url || ''
+  apiKey.value = ''
+}
+
+function onProviderChange() {
+  const p = config.value?.providers[selectedProvider.value]
+  model.value = p?.model || ''
+  baseUrl.value = p?.base_url || ''
+  apiKey.value = ''
+}
+
+async function save() {
+  saving.value = true
+  const body: any = { provider: selectedProvider.value }
+  if (apiKey.value) body.api_key = apiKey.value
+  if (model.value) body.model = model.value
+  if (baseUrl.value) body.base_url = baseUrl.value
+
+  const res = await fetch('/config/llm', {
+    method: 'PUT',
+    headers: { 'Content-Type': 'application/json' },
+    body: JSON.stringify(body),
+  })
+  config.value = await res.json()
+  saving.value = false
+}
+
+const providerLabels: Record<string, string> = {
+  groq: 'Groq (Llama 3.3 70B)',
+  anthropic: 'Anthropic (Claude)',
+  bedrock: 'AWS Bedrock (Claude)',
+  openai: 'OpenAI-compatible',
+  template: 'Template (no LLM)',
+}
+
+onMounted(loadConfig)
+</script>
+
+<template>
+  <div class="settings-page">
+    <Panel title="LLM Provider" status="idle">
+      <div v-if="config" class="config-form">
+        <div class="field">
+          <label>Provider</label>
+          <div class="provider-options">
+            <button
+              v-for="(label, key) in providerLabels"
+              :key="key"
+              :class="['provider-btn', { active: selectedProvider === key, configured: config.providers[key]?.configured }]"
+              @click="selectedProvider = key; onProviderChange()"
+            >
+              <span class="provider-name">{{ label }}</span>
+              <span v-if="config.providers[key]?.configured" class="configured-dot"></span>
+            </button>
+          </div>
+        </div>
+
+        <div v-if="selectedProvider !== 'template'" class="field">
+          <label>Model</label>
+          <input v-model="model" class="input" placeholder="model name" />
+        </div>
+
+        <div v-if="selectedProvider !== 'template' && selectedProvider !== 'bedrock'" class="field">
+          <label>API Key</label>
+          <input
+            v-model="apiKey"
+            type="password"
+            class="input"
+            :placeholder="config.providers[selectedProvider]?.configured ? '(configured — leave blank to keep)' : 'enter API key'"
+          />
+        </div>
+
+        <div v-if="selectedProvider === 'openai'" class="field">
+          <label>Base URL</label>
+          <input v-model="baseUrl" class="input" placeholder="https://api.openai.com/v1" />
+        </div>
+
+        <div class="actions">
+          <button class="save-btn" @click="save" :disabled="saving">
+            {{ saving ? 'Saving...' : 'Apply' }}
+          </button>
+          <span v-if="config.provider === selectedProvider" class="active-label">active</span>
+        </div>
+
+        <div class="status-table">
+          <div class="status-header">Provider Status</div>
+          <div v-for="(info, key) in config.providers" :key="key" class="status-row">
+            <span :class="['status-name', { active: config.provider === key }]">{{ key }}</span>
+            <span :class="['status-badge', info.configured ? 'ok' : 'missing']">
+              {{ info.configured ? 'configured' : 'no key' }}
+            </span>
+            <span class="status-model">{{ info.model }}</span>
+          </div>
+        </div>
+      </div>
+    </Panel>
+  </div>
+</template>
+
+<style scoped>
+.settings-page {
+  max-width: 700px;
+}
+
+.config-form {
+  padding: 16px;
+  display: flex;
+  flex-direction: column;
+  gap: 16px;
+}
+
+.field {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+}
+
+.field label {
+  font-family: var(--font-mono);
+  font-size: 11px;
+  color: var(--text-dim);
+  text-transform: uppercase;
+  letter-spacing: 1px;
+}
+
+.provider-options {
+  display: flex;
+  flex-direction: column;
+  gap: 4px;
+}
+
+.provider-btn {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  padding: 8px 12px;
+  background: var(--surface-2);
+  border: var(--panel-border);
+  color: var(--text-secondary);
+  font-family: var(--font-mono);
+  font-size: 12px;
+  cursor: pointer;
+  text-align: left;
+}
+
+.provider-btn:hover { background: var(--surface-3); }
+.provider-btn.active { border-color: var(--accent); color: var(--text-primary); background: var(--accent-dim); }
+
+.configured-dot {
+  width: 6px;
+  height: 6px;
+  border-radius: 50%;
+  background: var(--status-live);
+  margin-left: auto;
+}
+
+.input {
+  background: var(--surface-0);
+  color: var(--text-primary);
+  border: var(--panel-border);
+  padding: 6px 10px;
+  font-family: var(--font-mono);
+  font-size: 13px;
+}
+
+.input:focus { outline: 1px solid var(--accent); }
+
+.actions {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+}
+
+.save-btn {
+  background: var(--accent);
+  color: white;
+  border: none;
+  padding: 6px 24px;
+  font-family: var(--font-mono);
+  font-size: 12px;
+  cursor: pointer;
+}
+
+.save-btn:hover { background: var(--accent-dim); }
+.save-btn:disabled { opacity: 0.5; }
+
+.active-label {
+  font-family: var(--font-mono);
+  font-size: 11px;
+  color: var(--status-live);
+}
+
+.status-table {
+  margin-top: 8px;
+  border-top: var(--panel-border);
+  padding-top: 12px;
+}
+
+.status-header {
+  font-family: var(--font-mono);
+  font-size: 11px;
+  color: var(--text-dim);
+  text-transform: uppercase;
+  letter-spacing: 1px;
+  margin-bottom: 8px;
+}
+
+.status-row {
+  display: flex;
+  align-items: center;
+  gap: 12px;
+  padding: 4px 0;
+  font-size: 12px;
+}
+
+.status-name {
+  font-family: var(--font-mono);
+  width: 80px;
+  color: var(--text-dim);
+}
+
+.status-name.active { color: var(--accent); font-weight: 600; }
+
+.status-badge {
+  font-family: var(--font-mono);
+  font-size: 10px;
+  padding: 1px 6px;
+  width: 80px;
+  text-align: center;
+}
+
+.status-badge.ok { color: var(--status-live); border: 1px solid var(--status-live); }
+.status-badge.missing { color: var(--text-dim); border: 1px solid var(--surface-3); }
+
+.status-model {
+  font-family: var(--font-mono);
+  font-size: 11px;
+  color: var(--text-dim);
+}
+</style>
--- a/uv.lock
+++ b/uv.lock
@@ -1091,6 +1091,25 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/58/78/548fb8e07b1a341746bfbecb32f2c268470f45fa028aacdbd10d9bc73aab/numpy-2.4.4-cp314-cp314t-win_arm64.whl", hash = "sha256:ba203255017337d39f89bdd58417f03c4426f12beed0440cfd933cb15f8669c7", size = 10566643, upload-time = "2026-03-29T13:21:34.339Z" },
 ]

+[[package]]
+name = "openai"
+version = "2.31.0"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "anyio" },
+    { name = "distro" },
+    { name = "httpx" },
+    { name = "jiter" },
+    { name = "pydantic" },
+    { name = "sniffio" },
+    { name = "tqdm" },
+    { name = "typing-extensions" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/94/fe/64b3d035780b3188f86c4f6f1bc202e7bb74757ef028802112273b9dcacf/openai-2.31.0.tar.gz", hash = "sha256:43ca59a88fc973ad1848d86b98d7fac207e265ebbd1828b5e4bdfc85f79427a5", size = 684772, upload-time = "2026-04-08T21:01:41.797Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/66/bc/a8f7c3aa03452fedbb9af8be83e959adba96a6b4a35e416faffcc959c568/openai-2.31.0-py3-none-any.whl", hash = "sha256:44e1344d87e56a493d649b17e2fac519d1368cbb0745f59f1957c4c26de50a0a", size = 1153479, upload-time = "2026-04-08T21:01:39.217Z" },
+]
+
 [[package]]
 name = "openapi-pydantic"
 version = "0.5.1"
@@ -1895,6 +1914,7 @@ dependencies = [
    { name = "langfuse" },
    { name = "langgraph" },
    { name = "mcp", extra = ["cli"] },
+    { name = "openai" },
    { name = "pydantic" },
    { name = "uvicorn", extra = ["standard"] },
    { name = "websockets" },
@@ -1921,6 +1941,7 @@ requires-dist = [
    { name = "langfuse" },
    { name = "langgraph" },
    { name = "mcp", extras = ["cli"] },
+    { name = "openai" },
    { name = "pydantic", specifier = ">=2.0" },
    { name = "pytest", marker = "extra == 'dev'" },
    { name = "pytest-asyncio", marker = "extra == 'dev'" },
@@ -1939,6 +1960,18 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/d7/c1/eb8f9debc45d3b7918a32ab756658a0904732f75e555402972246b0b8e71/tenacity-9.1.4-py3-none-any.whl", hash = "sha256:6095a360c919085f28c6527de529e76a06ad89b23659fa881ae0649b867a9d55", size = 28926, upload-time = "2026-02-07T10:45:32.24Z" },
 ]

+[[package]]
+name = "tqdm"
+version = "4.67.3"
+source = { registry = "https://pypi.org/simple" }
+dependencies = [
+    { name = "colorama", marker = "sys_platform == 'win32'" },
+]
+sdist = { url = "https://files.pythonhosted.org/packages/09/a9/6ba95a270c6f1fbcd8dac228323f2777d886cb206987444e4bce66338dd4/tqdm-4.67.3.tar.gz", hash = "sha256:7d825f03f89244ef73f1d4ce193cb1774a8179fd96f31d7e1dcde62092b960bb", size = 169598, upload-time = "2026-02-03T17:35:53.048Z" }
+wheels = [
+    { url = "https://files.pythonhosted.org/packages/16/e1/3079a9ff9b8e11b846c6ac5c8b5bfb7ff225eee721825310c91b3b50304f/tqdm-4.67.3-py3-none-any.whl", hash = "sha256:ee1e4c0e59148062281c49d80b25b67771a127c85fc9676d3be5f243206826bf", size = 78374, upload-time = "2026-02-03T17:35:50.982Z" },
+]
+
 [[package]]
 name = "typer"
 version = "0.24.1"