diff --git a/repi.session.sql b/repi.session.sql
new file mode 100644
index 0000000..e69de29
diff --git a/repi/api/chat.py b/repi/api/chat.py
index 7c19874..2ee4381 100644
--- a/repi/api/chat.py
+++ b/repi/api/chat.py
@@ -24,12 +24,10 @@
 import logging
 from collections import Counter
 from datetime import datetime, timedelta, timezone
-from typing import List, Literal, Optional
 from uuid import UUID, uuid4
 
 from fastapi import APIRouter
 from fastapi.responses import StreamingResponse
-from pydantic import BaseModel, Field
 from sqlalchemy import select, text as sa_text
 
 from repi.core.container import get_container
@@ -44,6 +42,7 @@
 from repi.llm.provider import Message
 from repi.models.filters import RetrievalFilters
 from repi.models.schema import ChatMessage, Conversation
+from repi.api.schemas import ChatFilters, ChatRequest, ChatTurn
 from repi.retrieval.cluster_view import cluster_chunks
 from repi.retrieval.timeline_view import build_timeline
 
@@ -52,43 +51,6 @@
 router = APIRouter()
 
 
-# ── Request / response models ─────────────────────────────────────────────────
-
-
-class ChatTurn(BaseModel):
-    role: Literal["user", "assistant"]
-    content: str
-
-
-class ChatFilters(BaseModel):
-    service: Optional[str] = None
-    time_from: Optional[datetime] = None
-    time_to: Optional[datetime] = None
-    entity: Optional[str] = None
-
-
-class ChatRequest(BaseModel):
-    query: str
-    history: List[ChatTurn] = []
-    filters: Optional[ChatFilters] = None
-    conversation_id: Optional[UUID] = None
-    # Followup-bias hint: chunk_ids the previous assistant turn cited. When
-    # the current query is missing EITHER an explicit service or an explicit
-    # time window, the chat path fills in just the missing dimension from
-    # the previous turn's chunks — service via dominant-source check, time
-    # via a `Settings.FOLLOWUP_BIAS_WINDOW_MINUTES` envelope around their
-    # timestamps. Soft — never overrides an explicit filter, silently
-    # ignored if the IDs no longer resolve.
-    #
-    # Capped at 50 to bound the indexed-PK fetch and reject malformed
-    # payloads early. The legitimate caller only ever sends the last
-    # assistant turn's citations (≤10 in practice).
-    previous_chunk_ids: List[str] = Field(default_factory=list, max_length=50)
-    # UX P1: scopes retrieval + known-services resolution to one project. If
-    # omitted but the conversation has a project, that project applies.
-    project_id: Optional[UUID] = None
-
-
 # ── Module-level constants ────────────────────────────────────────────────────
 
 # Caller-visible window on cited-chunk `text` in the SSE done payload. Locked
diff --git a/repi/api/conversations.py b/repi/api/conversations.py
index 13548d5..0c78de9 100644
--- a/repi/api/conversations.py
+++ b/repi/api/conversations.py
@@ -7,51 +7,21 @@
 from __future__ import annotations
 
 import logging
-from typing import List, Literal, Optional
+from typing import List
 from uuid import UUID
 
 from fastapi import APIRouter, HTTPException
-from pydantic import BaseModel
 from sqlmodel import select
 
 from repi.core.container import get_container
 from repi.models.schema import ChatMessage, Conversation, Investigation, Project
+from repi.api.schemas import ConversationDetail, ConversationSummary, TranscriptTurn
 
 logger = logging.getLogger("repi.api.conversations")
 
 router = APIRouter()
 
 
-class ConversationSummary(BaseModel):
-    id: str
-    title: Optional[str]
-    project_id: Optional[str] = None
-    project_name: Optional[str] = None
-    created_at: str
-    updated_at: str
-
-
-class TranscriptTurn(BaseModel):
-    mode: Literal["chat", "investigate"]
-    id: str
-    role: Optional[str] = None  # "user" | "assistant" for chat turns
-    content: str
-    chunk_ids: List[str] = []
-    confidence: Optional[str] = None
-    status: Optional[str] = None  # investigation status (chat turns leave this null)
-    created_at: str
-
-
-class ConversationDetail(BaseModel):
-    id: str
-    title: Optional[str]
-    project_id: Optional[str] = None
-    project_name: Optional[str] = None
-    created_at: str
-    updated_at: str
-    turns: List[TranscriptTurn]
-
-
 @router.get("/conversations", response_model=List[ConversationSummary])
 async def list_conversations(limit: int = 50):
     container = get_container()
diff --git a/repi/api/ingest.py b/repi/api/ingest.py
index c765fd4..e68732d 100644
--- a/repi/api/ingest.py
+++ b/repi/api/ingest.py
@@ -1,21 +1,12 @@
 import logging
 from fastapi import APIRouter, UploadFile, File, Form, Depends
-from pydantic import BaseModel
 from repi.core.container import get_container
+from repi.api.schemas import IngestResponse
 
 logger = logging.getLogger("repi.api.ingest")
 
 router = APIRouter()
 
-class IngestResponse(BaseModel):
-    service: str
-    project: str
-    chunk_count: int
-    lines_total: int
-    lines_with_timestamp: int
-    level_counts: dict[str, int]
-    message: str
-
 @router.post("/ingest", response_model=IngestResponse)
 async def ingest(
     service: str = Form(...),
diff --git a/repi/api/investigate.py b/repi/api/investigate.py
index 84dd30c..989587f 100644
--- a/repi/api/investigate.py
+++ b/repi/api/investigate.py
@@ -4,59 +4,23 @@
 from typing import Optional, List
 from fastapi import APIRouter, HTTPException, Depends
 from fastapi.responses import StreamingResponse
-from pydantic import BaseModel
 from uuid import UUID
 from datetime import datetime
 
 from repi.core.container import get_container
 from repi.investigation.react_loop import InvestigationStep
+from repi.api.schemas import (
+    ClarifyRequest,
+    InvestigateRequest,
+    InvestigationResponse,
+    InvestigationStepModel,
+    SimpleInvestigationResponse,
+)
 
 logger = logging.getLogger("repi.api.investigate")
 
 router = APIRouter()
 
-class InvestigateRequest(BaseModel):
-    query: str
-    resume: bool = True
-    # Optional thread back to a chat surface (A1/A2). If omitted, a new
-    # conversation row is created and its id returned so the UI can attach
-    # subsequent /chat turns to the same thread.
-    conversation_id: Optional[UUID] = None
-    # UX P1: scopes retrieval + every ReAct tool to one project. If omitted
-    # but the conversation has a project, the conversation's project applies.
-    project_id: Optional[UUID] = None
-
-class InvestigationStepModel(BaseModel):
-    step_number: int
-    thought: str
-    # Legacy preview fields — kept for back-compat with anything that may still
-    # read them. The list endpoint returns empty steps anyway.
-    tool_name: Optional[str] = None
-    tool_args: Optional[dict] = None
-    observation_preview: Optional[str] = None
-    # Rich shape the UI uses to render a step identically to the SSE stream.
-    action: Optional[dict] = None
-    observation: Optional[dict] = None
-    kind: Optional[str] = None
-
-class InvestigationResponse(BaseModel):
-    id: str
-    query: str
-    status: str
-    answer: Optional[str] = None
-    created_at: datetime
-    steps: List[InvestigationStepModel]
-    pending_question: Optional[str] = None
-    stats: Optional[dict] = None
-
-class SimpleInvestigationResponse(BaseModel):
-    id: str
-    status: str
-    conversation_id: Optional[str] = None
-
-class ClarifyRequest(BaseModel):
-    reply: str
-
 @router.get("/investigations", response_model=List[InvestigationResponse])
 async def list_investigations(limit: int = 20):
     """List recent investigations."""
diff --git a/repi/api/projects.py b/repi/api/projects.py
index c179be8..d27cb16 100644
--- a/repi/api/projects.py
+++ b/repi/api/projects.py
@@ -13,13 +13,13 @@
 from uuid import UUID
 
 from fastapi import APIRouter, HTTPException
-from pydantic import BaseModel
 from sqlalchemy import text as sa_text
 from sqlmodel import select
 
 from repi.core.container import get_container
 from repi.core.dates import DateHandler
 from repi.models.schema import Project
+from repi.api.schemas import ProjectCreate, ProjectRead, ProjectService, ProjectUpdate
 from repi.retrieval.event_feed import derive_events, fetch_window_aggregates, parse_window
 
 logger = logging.getLogger("repi.api.projects")
@@ -76,33 +76,6 @@ async def _get_or_create_by_name(session, name: str) -> Project:
     return row
 
 
-# ── Models ───────────────────────────────────────────────────────────────────
-
-class ProjectCreate(BaseModel):
-    name: str
-    settings: dict[str, Any] = {}
-
-
-class ProjectUpdate(BaseModel):
-    name: Optional[str] = None
-    settings: Optional[dict[str, Any]] = None
-
-
-class ProjectRead(BaseModel):
-    id: str
-    name: str
-    settings: dict[str, Any]
-    service_count: int = 0
-    created_at: datetime
-    updated_at: datetime
-
-
-class ProjectService(BaseModel):
-    name: str
-    chunk_count: int
-    last_seen: Optional[datetime] = None
-
-
 # ── Endpoints ────────────────────────────────────────────────────────────────
 
 @router.get("/projects", response_model=List[ProjectRead])
@@ -275,9 +248,6 @@ async def project_overview(
         for r in svc_rows
     ]
 
-    # Suggested actions — derived, never LLM-generated. Top clusters become
-    # Deep-Research entry points with the service + time range pre-filled so
-    # the investigation starts grounded instead of from a bare phrase.
     suggested: list[dict] = []
     for c in clusters[:3]:
         svc_part = f" on {c['services'][0]}" if c["services"] else ""
@@ -289,16 +259,6 @@ async def project_overview(
                 f"between {c['first_ts']} and {c['last_ts']}"
             ),
         })
-    suggested.append({
-        "kind": "chat",
-        "label": f"Summarize the last {window_str}",
-        "query": f"summarize what happened in the last {window_str}",
-    })
-    suggested.append({
-        "kind": "chat",
-        "label": "Show affected services",
-        "query": "which services are having problems?",
-    })
 
     return {
         "project_id": str(project_id),
diff --git a/repi/api/schemas.py b/repi/api/schemas.py
new file mode 100644
index 0000000..7d81241
--- /dev/null
+++ b/repi/api/schemas.py
@@ -0,0 +1,207 @@
+"""Shared API request/response models.
+
+Pydantic models for the HTTP surface, extracted out of the individual route
+modules so they live in one tagged, importable place (mirrors the
+``repi/models/`` convention for domain/db models). The route modules import
+these back in; FastAPI behaviour, validation and the generated OpenAPI schema
+are unchanged by the move.
+"""
+from __future__ import annotations
+
+from datetime import datetime
+from typing import Any, List, Literal, Optional
+from uuid import UUID
+
+from pydantic import BaseModel, Field
+
+
+# ── Ingest ────────────────────────────────────────────────────────────────────
+
+class IngestResponse(BaseModel):
+    service: str
+    project: str
+    chunk_count: int
+    lines_total: int
+    lines_with_timestamp: int
+    level_counts: dict[str, int]
+    message: str
+
+
+# ── Watchers ──────────────────────────────────────────────────────────────────
+
+class WatcherConfigCreate(BaseModel):
+    service_name: str
+    watch_path: str
+    env: str = "production"
+    enabled: bool = True
+    project_id: UUID | None = None
+
+
+class WatcherConfigRead(BaseModel):
+    id: UUID
+    service_name: str
+    watch_path: str
+    env: str
+    enabled: bool
+    project_id: UUID | None = None
+    created_at: datetime
+    updated_at: datetime
+
+
+class WatcherConfigUpdate(BaseModel):
+    service_name: str = None
+    watch_path: str = None
+    env: str = None
+    enabled: bool = None
+    project_id: UUID = None
+
+
+class WatcherStatus(BaseModel):
+    file_path: str
+    offset: int
+    last_seen_at: datetime
+    updated_at: datetime
+
+
+# ── Projects ──────────────────────────────────────────────────────────────────
+
+class ProjectCreate(BaseModel):
+    name: str
+    settings: dict[str, Any] = {}
+
+
+class ProjectUpdate(BaseModel):
+    name: Optional[str] = None
+    settings: Optional[dict[str, Any]] = None
+
+
+class ProjectRead(BaseModel):
+    id: str
+    name: str
+    settings: dict[str, Any]
+    service_count: int = 0
+    created_at: datetime
+    updated_at: datetime
+
+
+class ProjectService(BaseModel):
+    name: str
+    chunk_count: int
+    last_seen: Optional[datetime] = None
+
+
+# ── Conversations ─────────────────────────────────────────────────────────────
+
+class ConversationSummary(BaseModel):
+    id: str
+    title: Optional[str]
+    project_id: Optional[str] = None
+    project_name: Optional[str] = None
+    created_at: str
+    updated_at: str
+
+
+class TranscriptTurn(BaseModel):
+    mode: Literal["chat", "investigate"]
+    id: str
+    role: Optional[str] = None  # "user" | "assistant" for chat turns
+    content: str
+    chunk_ids: List[str] = []
+    confidence: Optional[str] = None
+    status: Optional[str] = None  # investigation status (chat turns leave this null)
+    created_at: str
+
+
+class ConversationDetail(BaseModel):
+    id: str
+    title: Optional[str]
+    project_id: Optional[str] = None
+    project_name: Optional[str] = None
+    created_at: str
+    updated_at: str
+    turns: List[TranscriptTurn]
+
+
+# ── Investigations ────────────────────────────────────────────────────────────
+
+class InvestigateRequest(BaseModel):
+    query: str
+    resume: bool = True
+    # Optional thread back to a chat surface (A1/A2). If omitted, a new
+    # conversation row is created and its id returned so the UI can attach
+    # subsequent /chat turns to the same thread.
+    conversation_id: Optional[UUID] = None
+    # UX P1: scopes retrieval + every ReAct tool to one project. If omitted
+    # but the conversation has a project, the conversation's project applies.
+    project_id: Optional[UUID] = None
+
+
+class InvestigationStepModel(BaseModel):
+    step_number: int
+    thought: str
+    # Legacy preview fields — kept for back-compat with anything that may still
+    # read them. The list endpoint returns empty steps anyway.
+    tool_name: Optional[str] = None
+    tool_args: Optional[dict] = None
+    observation_preview: Optional[str] = None
+    # Rich shape the UI uses to render a step identically to the SSE stream.
+    action: Optional[dict] = None
+    observation: Optional[dict] = None
+    kind: Optional[str] = None
+
+
+class InvestigationResponse(BaseModel):
+    id: str
+    query: str
+    status: str
+    answer: Optional[str] = None
+    created_at: datetime
+    steps: List[InvestigationStepModel]
+    pending_question: Optional[str] = None
+    stats: Optional[dict] = None
+
+
+class SimpleInvestigationResponse(BaseModel):
+    id: str
+    status: str
+    conversation_id: Optional[str] = None
+
+
+class ClarifyRequest(BaseModel):
+    reply: str
+
+
+# ── Chat ──────────────────────────────────────────────────────────────────────
+
+class ChatTurn(BaseModel):
+    role: Literal["user", "assistant"]
+    content: str
+
+
+class ChatFilters(BaseModel):
+    service: Optional[str] = None
+    time_from: Optional[datetime] = None
+    time_to: Optional[datetime] = None
+    entity: Optional[str] = None
+
+
+class ChatRequest(BaseModel):
+    query: str
+    history: List[ChatTurn] = []
+    filters: Optional[ChatFilters] = None
+    conversation_id: Optional[UUID] = None
+    # Followup-bias hint: chunk_ids the previous assistant turn cited. When
+    # the current query is missing EITHER an explicit service or an explicit
+    # time window, the chat path fills in just the missing dimension from
+    # the previous turn's chunks — service via dominant-source check, time
+    # via a `Settings.FOLLOWUP_BIAS_WINDOW_MINUTES` envelope around their
+    # timestamps. Soft — never overrides an explicit filter, silently
+    # ignored if the IDs no longer resolve.
+    #
+    # Capped at 50 to bound the indexed-PK fetch and reject malformed
+    # payloads early. The legitimate caller only ever sends the last
+    # assistant turn's citations (≤10 in practice).
+    previous_chunk_ids: List[str] = Field(default_factory=list, max_length=50)
+    # UX P1: scopes retrieval + known-services resolution to one project. If
+    # omitted but the conversation has a project, that project applies.
+    project_id: Optional[UUID] = None
diff --git a/repi/api/watchers.py b/repi/api/watchers.py
index 76d38e8..51f0255 100644
--- a/repi/api/watchers.py
+++ b/repi/api/watchers.py
@@ -3,46 +3,21 @@
 from uuid import UUID, uuid4
 from datetime import datetime
 from fastapi import APIRouter, HTTPException, Depends
-from pydantic import BaseModel
 from sqlmodel import select
 
 from repi.core.container import get_container
 from repi.models.schema import WatcherConfig, WatcherOffset
+from repi.api.schemas import (
+    WatcherConfigCreate,
+    WatcherConfigRead,
+    WatcherConfigUpdate,
+    WatcherStatus,
+)
 
 logger = logging.getLogger("repi.api.watchers")
 
 router = APIRouter()
 
-class WatcherConfigCreate(BaseModel):
-    service_name: str
-    watch_path: str
-    env: str = "production"
-    enabled: bool = True
-    project_id: UUID | None = None
-
-class WatcherConfigRead(BaseModel):
-    id: UUID
-    service_name: str
-    watch_path: str
-    env: str
-    enabled: bool
-    project_id: UUID | None = None
-    created_at: datetime
-    updated_at: datetime
-
-class WatcherConfigUpdate(BaseModel):
-    service_name: str = None
-    watch_path: str = None
-    env: str = None
-    enabled: bool = None
-    project_id: UUID = None
-
-class WatcherStatus(BaseModel):
-    file_path: str
-    offset: int
-    last_seen_at: datetime
-    updated_at: datetime
-
 @router.post("/watchers", response_model=WatcherConfigRead)
 async def create_watcher(config: WatcherConfigCreate):
     container = get_container()
diff --git a/web/app/globals.css b/web/app/globals.css
index ee2b7a3..bccb224 100644
--- a/web/app/globals.css
+++ b/web/app/globals.css
@@ -7,9 +7,9 @@
 @theme inline {
   --color-background: var(--background);
   --color-foreground: var(--foreground);
-  --font-sans: var(--font-jakarta);
-  --font-mono: var(--font-mono);
-  --font-heading: var(--font-jakarta);
+  --font-sans: var(--font-geist);
+  --font-mono: var(--font-geist-mono);
+  --font-heading: var(--font-geist);
   --color-sidebar-ring: var(--sidebar-ring);
   --color-sidebar-border: var(--sidebar-border);
   --color-sidebar-accent-foreground: var(--sidebar-accent-foreground);
@@ -127,4 +127,104 @@
   html {
     @apply font-sans;
   }
+}
+
+/* Markdown rendering for chat prose / compiled answers / step thoughts.
+   Tailwind's preflight strips element margins, so ReactMarkdown output needs
+   its own typographic spacing. Inline `code` gets a Notion-like warm orange
+   hue; fenced code blocks keep the neutral mono treatment. */
+@layer components {
+  .md-content {
+    line-height: 1.6;
+    word-break: break-word;
+  }
+  .md-content > :first-child {
+    margin-top: 0;
+  }
+  .md-content > :last-child {
+    margin-bottom: 0;
+  }
+  .md-content p {
+    margin: 0.5em 0;
+  }
+  .md-content ul,
+  .md-content ol {
+    margin: 0.5em 0;
+    padding-left: 1.25em;
+  }
+  .md-content ul {
+    list-style: disc;
+  }
+  .md-content ol {
+    list-style: decimal;
+  }
+  .md-content li {
+    margin: 0.15em 0;
+  }
+  .md-content li > ul,
+  .md-content li > ol {
+    margin: 0.15em 0;
+  }
+  .md-content h1,
+  .md-content h2,
+  .md-content h3,
+  .md-content h4 {
+    font-weight: 600;
+    line-height: 1.3;
+    margin: 0.75em 0 0.35em;
+  }
+  .md-content h1 {
+    font-size: 1.25em;
+  }
+  .md-content h2 {
+    font-size: 1.15em;
+  }
+  .md-content h3 {
+    font-size: 1.05em;
+  }
+  .md-content a {
+    color: var(--primary);
+    text-decoration: underline;
+    text-underline-offset: 2px;
+  }
+  .md-content strong {
+    font-weight: 600;
+  }
+  .md-content em {
+    font-style: italic;
+  }
+  .md-content blockquote {
+    border-left: 2px solid var(--border);
+    padding-left: 0.75em;
+    margin: 0.5em 0;
+    color: var(--muted-foreground);
+  }
+  .md-content pre {
+    margin: 0.5em 0;
+    padding: 0.75em;
+    border-radius: 0.5rem;
+    background: var(--muted);
+    overflow-x: auto;
+    font-family: var(--font-mono);
+    font-size: 0.85em;
+  }
+  .md-content pre code {
+    background: none;
+    color: inherit;
+    padding: 0;
+  }
+  /* Notion-like warm orange inline code (not fenced blocks). */
+  .md-content :not(pre) > code {
+    font-family: var(--font-mono);
+    font-size: 0.85em;
+    color: oklch(0.62 0.17 45);
+    background: oklch(0.62 0.17 45 / 0.1);
+    padding: 0.12em 0.34em;
+    border-radius: 0.32rem;
+    white-space: break-spaces;
+  }
+  .dark .md-content :not(pre) > code {
+    color: oklch(0.78 0.15 55);
+    background: oklch(0.78 0.15 55 / 0.14);
+  }
 }
\ No newline at end of file
diff --git a/web/app/layout.tsx b/web/app/layout.tsx
index 9352fe2..c5a66fe 100644
--- a/web/app/layout.tsx
+++ b/web/app/layout.tsx
@@ -1,13 +1,21 @@
 import type { Metadata } from "next";
-import { Plus_Jakarta_Sans } from "next/font/google";
+import { Geist, Geist_Mono } from "next/font/google";
 import "./globals.css";
 import { ThemeProvider } from "@/components/theme-provider";
 import { Toaster } from "@/components/ui/sonner";
 import { Navbar } from "@/components/navbar";
 
-const jakarta = Plus_Jakarta_Sans({
+// Mellow, developer-native pairing: Geist for body/headings, Geist Mono for
+// code/logs. Geist Mono finally backs the `--font-mono` slot that globals.css
+// referenced but nothing previously loaded.
+const geist = Geist({
   subsets: ["latin"],
-  variable: "--font-jakarta",
+  variable: "--font-geist",
+});
+
+const geistMono = Geist_Mono({
+  subsets: ["latin"],
+  variable: "--font-geist-mono",
 });
 
 export const metadata: Metadata = {
@@ -23,7 +31,7 @@ export default function RootLayout({
   return (
     <html lang="en" suppressHydrationWarning>
       <body
-        className={`${jakarta.variable} font-sans antialiased min-h-screen bg-background text-foreground`}
+        className={`${geist.variable} ${geistMono.variable} font-sans antialiased min-h-screen bg-background text-foreground`}
       >
         <ThemeProvider
           attribute="class"
diff --git a/web/app/page.tsx b/web/app/page.tsx
index af3bf3f..de6f50e 100644
--- a/web/app/page.tsx
+++ b/web/app/page.tsx
@@ -8,31 +8,16 @@ import { ConversationSidebar } from "@/components/conversations/ConversationSide
 import { InvestigationStepCard } from "@/components/investigation-step"
 import { ProjectPicker } from "@/components/projects/ProjectPicker"
 import { ProjectOverview, type SuggestedAction } from "@/components/projects/ProjectOverview"
-import { Step, useSSE } from "@/lib/sse"
+import { useSSE } from "@/lib/sse"
 import { ThinkingIndicator } from "@/components/chat/ThinkingIndicator"
 import { CompiledAnswer } from "@/components/chat/CompiledAnswer"
 import { Badge } from "@/components/ui/badge"
 import { Sparkles } from "lucide-react"
 import { toast } from "sonner"
+import type { Step, Turn } from "@/lib/types"
 
 const DR_KEY = "repi.deepResearch"
 
-// Local transcript model. `mode` discriminates chat vs investigate so we can
-// render the right component inline. Investigation turns carry an
-// investigationId; the SSE stream is mounted via <InvestigateTurnView />.
-// `finalAnswer` is populated when the investigation's SSE `done` event fires —
-// kept on the parent's state so the next /chat turn can include it as history
-// context (lite contextual chat: see `buildChatHistory`).
-type Turn =
-  | ({ mode: "chat" } & ChatMessageProps)
-  | {
-      mode: "investigate"
-      id: string
-      investigationId: string
-      query: string
-      finalAnswer?: string
-    }
-
 // Token-budget proxy: number of recent turns to send to /chat as `history`.
 // 6 covers a typical Q&A → followup → followup flow without bloating the prompt.
 // No compaction, no summarisation — when it falls off it falls off. The
@@ -80,7 +65,16 @@ export default function HomePage() {
   const [turns, setTurns] = useState<Turn[]>([])
   const [busy, setBusy] = useState(false)
   const [sidebarRefresh, setSidebarRefresh] = useState(0)
+  // Conversation id whose investigation is streaming this session — drives the
+  // sidebar's per-row loading spinner. Session-only: cleared when the stream
+  // settles or the user switches conversations (not persisted across reload).
+  const [activeInvestigatingConvId, setActiveInvestigatingConvId] = useState<string | null>(null)
   const scrollRef = useRef<HTMLDivElement>(null)
+  const contentRef = useRef<HTMLDivElement>(null)
+  // Whether the view should keep itself pinned to the bottom as new content
+  // (streaming investigation steps / chat deltas) grows. Flips to false when
+  // the user scrolls up to read, so we don't yank them back down.
+  const stickToBottom = useRef(true)
 
   // Restore the sticky toggle from localStorage on mount.
   useEffect(() => {
@@ -93,16 +87,43 @@ export default function HomePage() {
     try { localStorage.setItem(DR_KEY, deepResearch ? "1" : "0") } catch {}
   }, [deepResearch])
 
-  // Auto-scroll to bottom on new content.
+  // New turn (the user just sent something, or a transcript loaded) → snap to
+  // the bottom and re-arm bottom-sticking.
   useEffect(() => {
+    stickToBottom.current = true
     if (scrollRef.current) {
       scrollRef.current.scrollTop = scrollRef.current.scrollHeight
     }
   }, [turns])
 
+  // Follow streaming content (investigation steps / chat deltas) that grows
+  // WITHOUT changing `turns`. A ResizeObserver on the content wrapper keeps the
+  // view pinned to the bottom, but only while the user is already there.
+  useEffect(() => {
+    const el = scrollRef.current
+    const content = contentRef.current
+    if (!el || !content) return
+    const ro = new ResizeObserver(() => {
+      if (stickToBottom.current) el.scrollTop = el.scrollHeight
+    })
+    ro.observe(content)
+    return () => ro.disconnect()
+  }, [])
+
+  // Track whether the user is parked near the bottom; drives `stickToBottom`.
+  const onScroll = () => {
+    const el = scrollRef.current
+    if (!el) return
+    const distanceFromBottom = el.scrollHeight - el.scrollTop - el.clientHeight
+    stickToBottom.current = distanceFromBottom < 80
+  }
+
   // Load an existing conversation's transcript from the API.
   const loadConversation = useCallback(async (id: string | null) => {
     setConversationId(id)
+    // Switching conversations drops the live stream we were following, so any
+    // session-only "investigating" spinner no longer applies.
+    setActiveInvestigatingConvId(null)
     if (!id) {
       // New conversation → back to project selection (the picker auto-skips
       // itself when exactly one project exists).
@@ -152,9 +173,13 @@ export default function HomePage() {
       // Toggle ON → kick off a real investigation, embed its SSE stream.
       try {
         const res = await api.investigations.create(query, conversationId ?? undefined, project?.id)
+        const convId = res.conversation_id ?? conversationId
         if (!conversationId && res.conversation_id) {
           setConversationId(res.conversation_id)
         }
+        // Mark this conversation as investigating so the sidebar shows a spinner
+        // on its row until the stream settles (see InvestigateTurnView.onSettled).
+        if (convId) setActiveInvestigatingConvId(convId)
         setTurns((prev) => [
           ...prev,
           {
@@ -300,13 +325,13 @@ export default function HomePage() {
   // New chat, no project yet → step 1 of the flow: pick (or create) a project.
   if (!conversationId && !project) {
     return (
-      <div className="flex h-[calc(100vh-3.5rem)]">
+      <div className="flex h-[calc(100vh-3.5rem)] overflow-hidden">
         <ConversationSidebar
           activeId={conversationId}
           onSelect={loadConversation}
           refreshKey={sidebarRefresh}
         />
-        <main className="flex-1 flex flex-col">
+        <main className="flex-1 flex flex-col min-h-0">
           <ProjectPicker onSelect={(p) => setProject({ id: p.id, name: p.name })} />
         </main>
       </div>
@@ -314,14 +339,16 @@ export default function HomePage() {
   }
 
   return (
-    <div className="flex h-[calc(100vh-3.5rem)]">
+    <div className="flex h-[calc(100vh-3.5rem)] overflow-hidden">
       <ConversationSidebar
         activeId={conversationId}
         onSelect={loadConversation}
         refreshKey={sidebarRefresh}
+        activeInvestigatingId={activeInvestigatingConvId}
       />
-      <main className="flex-1 flex flex-col">
-        <div ref={scrollRef} className="flex-1 overflow-y-auto py-6 space-y-4">
+      <main className="flex-1 flex flex-col min-h-0">
+        <div ref={scrollRef} onScroll={onScroll} className="flex-1 overflow-y-auto">
+          <div ref={contentRef} className="py-6 space-y-4">
           {empty ? (
             // Timeline-first landing: no empty chat screen. The overview
             // answers "what happened recently?" before the user types.
@@ -353,6 +380,7 @@ export default function HomePage() {
                   key={t.id}
                   investigationId={t.investigationId}
                   alreadyHoisted={!!t.finalAnswer}
+                  onSettled={() => setActiveInvestigatingConvId(null)}
                   onComplete={(finalAnswer) => {
                     setTurns((prev) => {
                       const idx = prev.findIndex(
@@ -370,6 +398,7 @@ export default function HomePage() {
               ),
             )
           )}
+          </div>
         </div>
         <ChatInput
           deepResearch={deepResearch}
@@ -390,9 +419,10 @@ interface InvestigateTurnViewProps {
   investigationId: string
   alreadyHoisted: boolean
   onComplete: (finalAnswer: string) => void
+  onSettled?: () => void
 }
 
-function InvestigateTurnView({ investigationId, alreadyHoisted, onComplete }: InvestigateTurnViewProps) {
+function InvestigateTurnView({ investigationId, alreadyHoisted, onComplete, onSettled }: InvestigateTurnViewProps) {
   const streamUrl = `${API_BASE}/investigations/${investigationId}/stream`
   const { steps, answer, error, done, clarificationQuestion, awaitingClarification, phase } = useSSE(streamUrl)
 
@@ -402,6 +432,12 @@ function InvestigateTurnView({ investigationId, alreadyHoisted, onComplete }: In
     if (done && answer && !alreadyHoisted) onComplete(answer)
   }, [done, answer, alreadyHoisted, onComplete])
 
+  // Stream reached a terminal state (answered, errored, or otherwise done) →
+  // let the parent clear the sidebar "investigating" spinner.
+  useEffect(() => {
+    if (done || error) onSettled?.()
+  }, [done, error, onSettled])
+
   // Renders as the assistant-side response of a chat turn. The user-side
   // bubble is owned by the parent (the optimistic push in handleSend) — this
   // component is purely "what the assistant did to answer."
diff --git a/web/components/chat/ChatMessage.tsx b/web/components/chat/ChatMessage.tsx
index 02937bf..0b7a677 100644
--- a/web/components/chat/ChatMessage.tsx
+++ b/web/components/chat/ChatMessage.tsx
@@ -1,31 +1,18 @@
 "use client"
 
 import { useRef, useState } from "react"
+import ReactMarkdown from "react-markdown"
 import { Badge } from "@/components/ui/badge"
 import { Button } from "@/components/ui/button"
 import { cn } from "@/lib/utils"
 import { AlertTriangle, Clock, Layers, Microscope, Sparkles, User } from "lucide-react"
-import { EventClusters, type Cluster } from "@/components/chat/EventClusters"
-import { Timeline, type TimelineEntry } from "@/components/chat/Timeline"
-import { CitedChunks, type CitedChunk } from "@/components/chat/CitedChunks"
+import { EventClusters } from "@/components/chat/EventClusters"
+import { Timeline } from "@/components/chat/Timeline"
+import { CitedChunks } from "@/components/chat/CitedChunks"
+import type { ChatMessageProps } from "@/lib/types"
 
-export type ChatMessageProps = {
-  role: "user" | "assistant"
-  content: string
-  chunkIds?: string[]
-  confidence?: "low" | "medium" | "high" | null
-  isClarification?: boolean
-  streaming?: boolean
-  clusters?: Cluster[]
-  timeline?: TimelineEntry[]
-  citedChunks?: CitedChunk[]
-  query?: string
-  onInvestigateDeeper?: (query: string) => void
-}
+export type { ChatMessageProps }
 
-// Strip raw chunk citations the LLM may still inline despite the system
-// prompt asking it not to. Catches [chunk:abc...] and bare hex/uuid runs in
-// brackets so the user sees clean prose.
 function cleanContent(raw: string): string {
   if (!raw) return raw
   return raw
@@ -88,9 +75,9 @@ export function ChatMessageView({
       <div className={cn("max-w-[85%] space-y-2", isUser && "items-end")}>
         <div
           className={cn(
-            "rounded-2xl px-4 py-2.5 text-sm whitespace-pre-wrap border",
+            "rounded-2xl px-4 py-2.5 text-sm border",
             isUser
-              ? "bg-primary text-primary-foreground border-transparent"
+              ? "bg-primary text-primary-foreground border-transparent whitespace-pre-wrap"
               : "bg-muted text-foreground border-transparent",
           )}
         >
@@ -100,7 +87,19 @@ export function ChatMessageView({
               Need a bit more info
             </div>
           )}
-          {displayed || (streaming ? <span className="opacity-60">…</span> : "")}
+          {isUser ? (
+            displayed
+          ) : displayed ? (
+            // Assistant prose is markdown — render it. The `md-content` class
+            // carries the typographic spacing + Notion-style orange backticks.
+            <div className="md-content">
+              <ReactMarkdown>{displayed}</ReactMarkdown>
+            </div>
+          ) : streaming ? (
+            <span className="opacity-60">…</span>
+          ) : (
+            ""
+          )}
         </div>
 
         {!isUser && confidence && (
diff --git a/web/components/chat/CitedChunks.tsx b/web/components/chat/CitedChunks.tsx
index 25333ad..6edde85 100644
--- a/web/components/chat/CitedChunks.tsx
+++ b/web/components/chat/CitedChunks.tsx
@@ -6,13 +6,8 @@ import { ChevronDown, ChevronRight, FileText } from "lucide-react"
 import { cn } from "@/lib/utils"
 import { levelTone } from "@/lib/log-levels"
 
-export type CitedChunk = {
-  chunk_id: string
-  service: string | null
-  level: string | null
-  timestamp: string | null
-  text: string
-}
+import type { CitedChunk } from "@/lib/types"
+export type { CitedChunk }
 
 function shortTs(iso: string | null): string {
   if (!iso) return ""
diff --git a/web/components/chat/CompiledAnswer.tsx b/web/components/chat/CompiledAnswer.tsx
index 1d2c027..d1afb81 100644
--- a/web/components/chat/CompiledAnswer.tsx
+++ b/web/components/chat/CompiledAnswer.tsx
@@ -3,37 +3,7 @@
 import ReactMarkdown from "react-markdown"
 import { Badge } from "@/components/ui/badge"
 import { AlertTriangle, ArrowDown, Target, Ban, HelpCircle } from "lucide-react"
-
-// Shape of the compiled InvestigationAnswer (see repi/investigation/schema.py).
-// Everything is optional here because the JSON comes off the wire and we render
-// defensively — a missing section is simply skipped.
-interface TriggerEvent {
-  chunk_id?: string
-  service?: string
-  timestamp?: string
-  log_line?: string
-}
-interface PropagationHop {
-  service?: string
-  chunk_id?: string
-  ts?: string
-  what?: string
-}
-interface RuledOut {
-  hypothesis?: string
-  why_ruled_out?: string
-}
-interface InvestigationAnswer {
-  incident_window?: { start?: string; end?: string }
-  affected_services?: string[]
-  trigger_event?: TriggerEvent
-  propagation_chain?: PropagationHop[]
-  root_cause?: string
-  ruled_out_hypotheses?: RuledOut[]
-  assumptions?: string[]
-  confidence?: string
-  gaps?: string[]
-}
+import type { InvestigationAnswer } from "@/lib/types"
 
 // Strip the model's inline chunk references from prose — same cleanup the old
 // plain-text card did.
@@ -89,11 +59,12 @@ function Section({
 export function CompiledAnswer({ answer }: { answer: string }) {
   const parsed = tryParse(answer)
 
-  // Fallback: plain-text answer (clarification text or legacy prose).
+  // Fallback: plain-text answer (clarification text or legacy prose). Rendered
+  // as markdown for consistency with the rest of the chat surface.
   if (!parsed) {
     return (
-      <div className="rounded-md border bg-muted/40 px-3 py-2 text-sm whitespace-pre-wrap">
-        {stripChunkRefs(answer)}
+      <div className="md-content rounded-md border bg-muted/40 px-3 py-2 text-sm">
+        <ReactMarkdown>{stripChunkRefs(answer)}</ReactMarkdown>
       </div>
     )
   }
@@ -132,7 +103,7 @@ export function CompiledAnswer({ answer }: { answer: string }) {
       {/* Root cause — the headline */}
       {root_cause && (
         <Section icon={<Target className="h-3 w-3" />} title="Root cause">
-          <div className="text-foreground/90 leading-relaxed prose-sm">
+          <div className="md-content text-foreground/90 leading-relaxed">
             <ReactMarkdown>{stripChunkRefs(root_cause)}</ReactMarkdown>
           </div>
         </Section>
diff --git a/web/components/chat/EventClusters.tsx b/web/components/chat/EventClusters.tsx
index d7cc00f..c12ea50 100644
--- a/web/components/chat/EventClusters.tsx
+++ b/web/components/chat/EventClusters.tsx
@@ -5,13 +5,8 @@ import { Badge } from "@/components/ui/badge"
 import { ChevronDown, ChevronRight, Layers } from "lucide-react"
 import { cn } from "@/lib/utils"
 
-export type Cluster = {
-  signature: string
-  count: number
-  services: string[]
-  first_ts: string | null
-  last_ts: string | null
-}
+import type { Cluster } from "@/lib/types"
+export type { Cluster }
 
 function formatRange(first: string | null, last: string | null): string {
   if (!first && !last) return ""
diff --git a/web/components/chat/ThinkingIndicator.tsx b/web/components/chat/ThinkingIndicator.tsx
index 157f3dd..1dab4c7 100644
--- a/web/components/chat/ThinkingIndicator.tsx
+++ b/web/components/chat/ThinkingIndicator.tsx
@@ -24,6 +24,10 @@ const THINKING_WORDS = [
   "Weighing evidence",
   "Reading log clusters",
   "Forming hypotheses",
+  "Drinking water",
+  "Taking a deep breath",
+  "Crashing out",
+  "Praying to the AI gods"
 ]
 
 function contextLabel(phase: InvestigationPhase | null, lastStep?: Step): string | null {
diff --git a/web/components/chat/Timeline.tsx b/web/components/chat/Timeline.tsx
index eb5505d..0738a44 100644
--- a/web/components/chat/Timeline.tsx
+++ b/web/components/chat/Timeline.tsx
@@ -6,14 +6,8 @@ import { ChevronDown, ChevronRight, Clock } from "lucide-react"
 import { cn } from "@/lib/utils"
 import { levelTone } from "@/lib/log-levels"
 
-export type TimelineEntry = {
-  service: string | null
-  level: string | null
-  signature: string
-  first_ts: string
-  last_ts: string
-  repeat_count: number
-}
+import type { TimelineEntry } from "@/lib/types"
+export type { TimelineEntry }
 
 function formatTs(iso: string): string {
   if (!iso) return ""
diff --git a/web/components/conversations/ConversationSidebar.tsx b/web/components/conversations/ConversationSidebar.tsx
index 1f6be29..692a873 100644
--- a/web/components/conversations/ConversationSidebar.tsx
+++ b/web/components/conversations/ConversationSidebar.tsx
@@ -5,22 +5,19 @@ import { Button } from "@/components/ui/button"
 import { MessageSquare, Plus } from "lucide-react"
 import { api } from "@/lib/api"
 import { cn } from "@/lib/utils"
-
-type ConversationSummary = {
-  id: string
-  title: string | null
-  project_name?: string | null
-  created_at: string
-  updated_at: string
-}
+import { Spinner } from "@/components/ui/spinner"
+import type { ConversationSummary } from "@/lib/types"
 
 interface ConversationSidebarProps {
   activeId: string | null
   onSelect: (id: string | null) => void
   refreshKey?: number  // bump to force a re-fetch (e.g. after sending a turn)
+  // Conversation id with a live investigation this session → show a spinner on
+  // its row in place of the message icon.
+  activeInvestigatingId?: string | null
 }
 
-export function ConversationSidebar({ activeId, onSelect, refreshKey }: ConversationSidebarProps) {
+export function ConversationSidebar({ activeId, onSelect, refreshKey, activeInvestigatingId }: ConversationSidebarProps) {
   const [items, setItems] = useState<ConversationSummary[]>([])
   const [loading, setLoading] = useState(true)
   const [error, setError] = useState<string | null>(null)
@@ -67,7 +64,11 @@ export function ConversationSidebar({ activeId, onSelect, refreshKey }: Conversa
               activeId === c.id && "bg-muted font-medium",
             )}
           >
-            <MessageSquare className="size-3.5 mt-0.5 flex-shrink-0 text-muted-foreground" />
+            {activeInvestigatingId === c.id ? (
+              <Spinner size="sm" className="mt-0.5 flex-shrink-0 text-primary" />
+            ) : (
+              <MessageSquare className="size-3.5 mt-0.5 flex-shrink-0 text-muted-foreground" />
+            )}
             <span className="min-w-0">
               <span className="line-clamp-2 break-words">{c.title || "Untitled"}</span>
               {c.project_name && (
diff --git a/web/components/investigation-step.tsx b/web/components/investigation-step.tsx
index 7fffae1..f614aba 100644
--- a/web/components/investigation-step.tsx
+++ b/web/components/investigation-step.tsx
@@ -82,7 +82,7 @@ export function InvestigationStepCard({ step }: { step: Step }) {
         {/* Thought */}
         <div className="flex items-start gap-3">
           <Brain className="h-5 w-5 mt-0.5 text-primary" />
-          <div className="flex-1 text-sm leading-relaxed text-foreground/90 bg-muted/30 p-3 rounded-lg border border-border/50">
+          <div className="md-content flex-1 text-sm leading-relaxed text-foreground/90 bg-muted/30 p-3 rounded-lg border border-border/50">
             <ReactMarkdown>{step.thought}</ReactMarkdown>
           </div>
         </div>
diff --git a/web/lib/sse.ts b/web/lib/sse.ts
index 09970ab..fcdbdf1 100644
--- a/web/lib/sse.ts
+++ b/web/lib/sse.ts
@@ -1,32 +1,14 @@
 "use client"
 
 import { useState, useEffect, useCallback, useRef } from "react"
+import type {
+  InvestigationPhase,
+  InvestigationStats,
+  Step,
+  StepKind,
+} from "@/lib/types"
 
-export type StepKind = null | "reflection" | "signal" | "compile"
-
-export interface Step {
-  step_number: number
-  thought: string
-  action?: {
-    tool: string
-    args: any
-  }
-  observation?: any
-  kind?: StepKind
-}
-
-export type InvestigationPhase = "gathering" | "compiling" | "done"
-
-export interface InvestigationStats {
-  iterations_used?: number
-  reflections_used?: number
-  chunks_gathered?: number
-  tools_called?: string[]
-  compile_source?: string
-  compile_attempts?: number
-  floor_adjustments?: string[]
-  gathering_exit_reason?: string
-}
+export type { InvestigationPhase, InvestigationStats, Step, StepKind }
 
 export function useSSE(url: string | null) {
   const [steps, setSteps] = useState<Step[]>([])
diff --git a/web/lib/types.ts b/web/lib/types.ts
new file mode 100644
index 0000000..07acccc
--- /dev/null
+++ b/web/lib/types.ts
@@ -0,0 +1,132 @@
+// Central type surface for the web app. These declarations were previously
+// inlined across component and lib modules; collecting them here gives a single
+// import target (`@/lib/types`) for components and tests. The original modules
+// re-export from here for backward compatibility, so existing import paths keep
+// working. Pure type declarations only — no runtime code.
+
+// ── Retrieval / evidence views ────────────────────────────────────────────────
+
+export type Cluster = {
+  signature: string
+  count: number
+  services: string[]
+  first_ts: string | null
+  last_ts: string | null
+}
+
+export type TimelineEntry = {
+  service: string | null
+  level: string | null
+  signature: string
+  first_ts: string
+  last_ts: string
+  repeat_count: number
+}
+
+export type CitedChunk = {
+  chunk_id: string
+  service: string | null
+  level: string | null
+  timestamp: string | null
+  text: string
+}
+
+// ── Investigation SSE stream ──────────────────────────────────────────────────
+
+export type StepKind = null | "reflection" | "signal" | "compile"
+
+export interface Step {
+  step_number: number
+  thought: string
+  action?: {
+    tool: string
+    args: any
+  }
+  observation?: any
+  kind?: StepKind
+}
+
+export type InvestigationPhase = "gathering" | "compiling" | "done"
+
+export interface InvestigationStats {
+  iterations_used?: number
+  reflections_used?: number
+  chunks_gathered?: number
+  tools_called?: string[]
+  compile_source?: string
+  compile_attempts?: number
+  floor_adjustments?: string[]
+  gathering_exit_reason?: string
+}
+
+// ── Compiled investigation answer (mirrors repi/investigation/schema.py) ───────
+// Everything is optional — the JSON comes off the wire and is rendered
+// defensively; a missing section is simply skipped.
+
+export interface TriggerEvent {
+  chunk_id?: string
+  service?: string
+  timestamp?: string
+  log_line?: string
+}
+
+export interface PropagationHop {
+  service?: string
+  chunk_id?: string
+  ts?: string
+  what?: string
+}
+
+export interface RuledOut {
+  hypothesis?: string
+  why_ruled_out?: string
+}
+
+export interface InvestigationAnswer {
+  incident_window?: { start?: string; end?: string }
+  affected_services?: string[]
+  trigger_event?: TriggerEvent
+  propagation_chain?: PropagationHop[]
+  root_cause?: string
+  ruled_out_hypotheses?: RuledOut[]
+  assumptions?: string[]
+  confidence?: string
+  gaps?: string[]
+}
+
+// ── Chat / conversation ───────────────────────────────────────────────────────
+
+export type ChatMessageProps = {
+  role: "user" | "assistant"
+  content: string
+  chunkIds?: string[]
+  confidence?: "low" | "medium" | "high" | null
+  isClarification?: boolean
+  streaming?: boolean
+  clusters?: Cluster[]
+  timeline?: TimelineEntry[]
+  citedChunks?: CitedChunk[]
+  query?: string
+  onInvestigateDeeper?: (query: string) => void
+}
+
+// Local transcript model. `mode` discriminates chat vs investigate so the page
+// can render the right component inline. Investigation turns carry an
+// investigationId; `finalAnswer` is populated when the SSE `done` event fires.
+export type Turn =
+  | ({ mode: "chat" } & ChatMessageProps)
+  | {
+      mode: "investigate"
+      id: string
+      investigationId: string
+      query: string
+      finalAnswer?: string
+    }
+
+export type ConversationSummary = {
+  id: string
+  title: string | null
+  project_name?: string | null
+  created_at: string
+  updated_at: string
+}