docs: Add Session Scratchpad design and update project references

- Add Session Scratchpad design doc (Idea 6 from brainstorm)
- Update CLAUDE.md production URLs to resolutionflow.com
- Update CORS lesson domain refs for rebrand
- Add .claude/settings.local.json to gitignore docs
- Add docs/plans/ to project structure

Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>

CLAUDE.md

@@ -24,7 +24,7 @@ The project was rebranded from "Patherly" to "ResolutionFlow" in the frontend (P
 | Repository / directory name | `patherly` |
 | Backend (FastAPI, env vars, APP_NAME) | ResolutionFlow |
 | Database / Docker container | `patherly` / `patherly_postgres` |
-| Production URLs | `app.patherly.com` / `api.patherly.com` |
+| Production URLs | `resolutionflow.com` / `api.resolutionflow.com` |
 | **Frontend UI (header, login, register)** | **ResolutionFlow** |
 | **Browser tab title** | **ResolutionFlow - Decision Tree Platform** |
 
@@ -216,6 +216,7 @@ patherly/
 ├── LESSONS-LEARNED.md              # Bugs and fixes (READ THIS!)
 ├── PROGRESS.md                     # Detailed progress log
 ├── REBRAND-IMPLEMENTATION-GUIDE.md # Patherly → ResolutionFlow rebrand guide
+├── docs/plans/                     # Implementation plans
 ├── 01-PROJECT-OVERVIEW.md          # Vision and goals
 ├── 02-TECHNICAL-ARCHITECTURE.md    # System design, API specs
 ├── 03-DEVELOPMENT-ROADMAP.md       # Phases and timeline
@@ -468,12 +469,12 @@ if settings.ALLOW_RAILWAY_ORIGINS:
 if settings.ALLOW_RAILWAY_ORIGINS:
     app.add_middleware(
         CORSMiddleware,
-        allow_origins=settings.allowed_origins,  # Custom domains like app.patherly.com
+        allow_origins=settings.allowed_origins,  # Custom domains like resolutionflow.com
         allow_origin_regex=r"https://.*\.up\.railway\.app",  # Railway PR domains
         # ...
     )
 ```
-When using `allow_origin_regex` for wildcard patterns, also include `allow_origins` for explicit custom domains. The regex alone won't match custom domains like `app.patherly.com`.
+When using `allow_origin_regex` for wildcard patterns, also include `allow_origins` for explicit custom domains. The regex alone won't match custom domains like `resolutionflow.com`.
 
 ### findNode Requires Tree Structure Parameter
 
@@ -738,6 +739,7 @@ const response = await api.get('/api/v1/trees')
 # Secrets and local config
 backend/.env
 .claude.local.md
+.claude/settings.local.json
 
 # Dependencies
 backend/venv/
@@ -778,8 +780,8 @@ backend/**/__pycache__/
 
 ### Production
 
-- **Frontend:** https://app.patherly.com
+- **Frontend:** <https://resolutionflow.com>
-- **Backend:** https://api.patherly.com
+- **Backend:** <https://api.resolutionflow.com>
 - **Database:** Railway-managed PostgreSQL
 - Deploys automatically on push to `main`
 

docs/plans/2026-02-04-feature-ideas-brainstorm.md

@@ -12,6 +12,7 @@ Every feature should follow the core principle that makes ResolutionFlow work:
 **Engineers don't document — they troubleshoot, and the tool captures documentation as a byproduct.**
 
 New features should:
+
 - Reduce engineer cognitive load
 - Automate capture of what they're already doing
 - Make reusing knowledge frictionless
@@ -42,23 +43,28 @@ New features should:
 **Effort:** Small (backend export change, no new UI)
 
 ### What
+
 Every session automatically tracks duration (start → end) and includes it in the export.
 
 ### Export Output
+
 ```
 Session Duration: 23 minutes
 Started: 2:30 PM | Completed: 2:53 PM
 ```
 
 ### Why
+
 MSP engineers bill by the hour. They troubleshoot in ResolutionFlow, export notes, then separately log time in their PSA. This eliminates the second step.
 
 ### Implementation Notes
+
 - Timestamps already exist in session decisions — just compute elapsed time
 - Add duration to export templates (markdown, text, HTML)
 - No new UI required — purely a backend export enhancement
 
 ### Future Enhancement
+
 - When PSA integration exists: if no PSA connected, ResolutionFlow tracks time natively. If PSA connected, still track time but also push it directly to the ticket in the PSA.
 - Analytics: average resolution time per tree, per client, per engineer
 
@@ -70,6 +76,7 @@ MSP engineers bill by the hour. They troubleshoot in ResolutionFlow, export note
 **Effort:** Medium
 
 ### What
+
 Two mechanisms for sharing in-progress troubleshooting context:
 
 1. **"Share Progress" button** (available mid-session) — generates a formatted summary of steps completed so far. Copy to clipboard, paste into Teams/Slack. One click instead of typing "here's what I've tried."
@@ -77,6 +84,7 @@ Two mechanisms for sharing in-progress troubleshooting context:
 2. **Read-only session link** — shareable URL where anyone with the link can see the session state. If they sign in and get assigned, they can resume from where the previous engineer left off.
 
 ### Workflow
+
 1. Junior engineer gets stuck → clicks "Share Progress"
 2. Pastes formatted summary into Teams chat with senior
 3. Senior reads structured summary (not a wall of chat text)
@@ -84,11 +92,13 @@ Two mechanisms for sharing in-progress troubleshooting context:
 5. If escalating: ticket reassigned in PSA, senior resumes the session in ResolutionFlow
 
 ### Implementation Notes
+
 - Copy/paste version is nearly free — existing export logic on incomplete sessions + "Steps remaining" section
 - Read-only link: generate share token, create public read-only session view (no auth)
 - Resume capability: allow session reassignment to another user
 
 ### Why This Matters
+
 Eliminates the "what have you tried so far?" back-and-forth that happens on every escalation. The structured format means the senior gets context in 30 seconds instead of 10 minutes of chat.
 
 ---
@@ -99,9 +109,11 @@ Eliminates the "what have you tried so far?" back-and-forth that happens on ever
 **Effort:** Small
 
 ### What
+
 Action nodes (which show commands to run) get an optional "Paste Output" text area. Engineer runs the command, copies output, pastes it in.
 
 ### Export Output
+
 ```
 > Ran: Get-Service -Name Spooler
 > Output:
@@ -111,15 +123,18 @@ Action nodes (which show commands to run) get an optional "Paste Output" text ar
 ```
 
 ### Why
+
 Engineers already run commands and read output. Today the output is lost — the export says "ran this command" but not what it returned. This captures the evidence.
 
 ### Implementation Notes
+
 - Add optional `command_output` field to session decision JSONB
 - Add collapsible text area below commands on action nodes in TreeNavigationPage
 - Include in export with code formatting
 - Pairs well with Scratchpad (Idea 6) and Share Progress (Idea 2)
 
 ### Future Enhancement
+
 - Syntax highlighting for common output formats (PowerShell, JSON)
 - Image paste for screenshots of GUI-based evidence
 
@@ -131,9 +146,11 @@ Engineers already run commands and read output. Today the output is lost — the
 **Effort:** Medium-Large
 
 ### What
+
 A senior engineer (or anyone) can send a troubleshooting step directly to someone's active session. Flips the step library from a pull model (browse and find) to a push model (someone sends it to you).
 
 ### Workflow
+
 1. Junior shares progress link (Idea 2)
 2. Senior sees they're stuck at "VDA not registering"
 3. Senior picks a step from their personal library (or types one quickly)
@@ -143,15 +160,18 @@ A senior engineer (or anyone) can send a troubleshooting step directly to someon
 7. Step documented in export, optionally saved to junior's library
 
 ### Why
+
 Replaces unstructured Teams/Slack troubleshooting advice with structured, documented, reusable steps. The knowledge stays in the system.
 
 ### Implementation Notes
+
 - Requires lightweight notification/inbox system (polling or WebSocket)
 - Builds on: Share Progress (Idea 2) + Step Library (existing)
 - New API: `POST /api/v1/sessions/{id}/send-step`
 - Frontend: notification badge + step insertion flow
 
 ### Analytics Potential
+
 - Which seniors send the most steps (mentorship tracking)
 - Which pushed steps get reused (knowledge value)
 - Which juniors receive fewer pushes over time (skill growth)
@@ -164,6 +184,7 @@ Replaces unstructured Teams/Slack troubleshooting advice with structured, docume
 **Effort:** Medium
 
 ### What
+
 Aggregate completed session data to show statistical hints on decision nodes:
 
 - Badge: "78% of engineers chose Option B here"
@@ -171,9 +192,11 @@ Aggregate completed session data to show statistical hints on decision nodes:
 - At common stuck points: "Engineers often add a custom step here"
 
 ### Why
+
 Delivers on the tagline. For a junior engineer staring at three options, seeing "most engineers went this way" is a confidence boost. For tree authors, analytics reveal dead-end branches.
 
 ### Implementation Notes
+
 - Aggregate query on session `path_taken` and `decisions` JSONB
 - Compute per-node: choice distribution, resolve rate, custom step frequency
 - Cache aggregates (recompute daily or on-demand)
@@ -181,6 +204,7 @@ Delivers on the tagline. For a junior engineer staring at three options, seeing
 - Resolve rate: track whether session completed at a solution node + optional "did this fix it?" prompt
 
 ### Data Requirements
+
 - Needs sufficient session volume per tree to be statistically meaningful (suggest: show after 10+ sessions)
 - Weight recent sessions higher than old ones
 
@@ -193,12 +217,15 @@ Delivers on the tagline. For a junior engineer staring at three options, seeing
 **Priority:** Must-have (per Michael)
 
 ### What
+
 A persistent sidebar during active sessions for capturing ambient data: IP addresses, error codes, server names, usernames — anything that doesn't fit a specific decision node's notes field.
 
 ### Why
+
 During troubleshooting, engineers accumulate bits of data (from `ipconfig`, Event Viewer, phone conversations) that live on sticky notes or in their head. This gives it a home and includes it in the export.
 
 ### Export Output
+
 ```
 ## Evidence / Reference
 - Server IP: 192.168.1.50
@@ -208,6 +235,7 @@ During troubleshooting, engineers accumulate bits of data (from `ipconfig`, Even
 ```
 
 ### Implementation Notes
+
 - Persistent sidebar (collapsible) in TreeNavigationPage
 - Store in session JSONB as `scratchpad` array of entries
 - Each entry: text + optional label + timestamp
@@ -216,6 +244,7 @@ During troubleshooting, engineers accumulate bits of data (from `ipconfig`, Even
 - Future: structured key-value pairs, tags, image paste
 
 ### Pairs With
+
 - Command Output Capture (Idea 3): structured output at nodes + freeform notes in scratchpad = complete evidence
 - Share Progress (Idea 2): scratchpad content included in shared summary
 
@@ -227,9 +256,11 @@ During troubleshooting, engineers accumulate bits of data (from `ipconfig`, Even
 **Effort:** Large
 
 ### What
+
 When troubleshooting reveals the problem is in a different domain, branch into another tree mid-session without losing context. The export captures the entire journey.
 
 ### Workflow
+
 1. Engineer is in "VPN Issues" tree, 5 steps deep
 2. Discovers the actual problem is DNS, not VPN
 3. Clicks "Open Related Tree" → selects "DNS Resolution Issues"
@@ -238,6 +269,7 @@ When troubleshooting reveals the problem is in a different domain, branch into a
 6. Export shows unified narrative with both trees
 
 ### Export Output
+
 ```
 ## VPN Connection Issues
 1. Verified VPN client version: OK
@@ -255,6 +287,7 @@ When troubleshooting reveals the problem is in a different domain, branch into a
 ```
 
 ### Implementation Notes
+
 - Session model needs: `parent_session_id`, `branched_at_node_id`
 - "Open Related Tree" action on any node (tree selector modal)
 - Breadcrumb shows tree chain: VPN > DNS
@@ -262,6 +295,7 @@ When troubleshooting reveals the problem is in a different domain, branch into a
 - Back button returns to parent session at bookmark point
 
 ### Why
+
 Real troubleshooting rarely stays in one domain. This captures the full diagnostic story.
 
 ---
@@ -272,22 +306,27 @@ Real troubleshooting rarely stays in one domain. This captures the full diagnost
 **Effort:** Small-Medium
 
 ### What
+
 When an engineer starts a session and enters a client name, show previous sessions for that client in that tree. If the same resolution keeps being reached, prompt for root cause action.
 
 ### UI
+
 - At session start: "3 previous sessions for Warner Robins in this tree (last: Jan 28)" + link to view
 - At resolution (if recurring): "This is the 3rd time this issue was resolved the same way for this client. Consider documenting a permanent fix or escalating to address root cause."
 
 ### Why
+
 Turns ResolutionFlow from reactive (fix the ticket) to proactive (fix the root cause). For MSP managers, recurring issues per client = business intelligence for infrastructure upgrade proposals.
 
 ### Implementation Notes
+
 - Query: sessions grouped by client_name + tree_id, count + last date
 - Display at session start (inline, not blocking)
 - Recurrence prompt: compare resolution node_id across sessions
 - Future (with PSA/RMM): correlate with alert data for richer signals
 
 ### Data Model
+
 - No schema changes needed — query existing sessions table
 - Optional: normalize client names (fuzzy match or client_id foreign key)
 
@@ -299,9 +338,11 @@ Turns ResolutionFlow from reactive (fix the ticket) to proactive (fix the root c
 **Effort:** Medium
 
 ### What
+
 Data-driven health indicators for trees, surfaced to tree authors and admins.
 
 ### Signals
+
 - **Custom step frequency**: Engineers keep adding steps at the same node → tree is missing a branch
 - **Abandonment rate**: Sessions started but not completed → tree isn't leading to resolutions
 - **Low resolve rate**: Solution nodes that don't actually fix issues
@@ -309,14 +350,17 @@ Data-driven health indicators for trees, surfaced to tree authors and admins.
 - **Escalation rate**: High percentage of sessions shared/escalated from this tree
 
 ### Display
+
 - Green/yellow/red health badge on tree library cards
 - Author notification: "Your 'VPN Issues' tree has yellow health — 4 engineers added custom steps at 'Check Split Tunnel Config' this month"
 - Admin dashboard: team-wide tree health overview
 
 ### Why
+
 Creates a self-improving ecosystem. Sessions generate data → data identifies weak trees → authors improve trees → next engineer gets a better experience.
 
 ### Implementation Notes
+
 - Scheduled aggregation job (daily)
 - Health score algorithm: weighted combination of signals
 - Store as computed field on tree (or separate analytics table)
@@ -331,21 +375,25 @@ Creates a self-improving ecosystem. Sessions generate data → data identifies w
 **Status:** Ultimate goal
 
 ### What
+
 Three layers of AI, each building on the last, leveraging ResolutionFlow's unique structured troubleshooting dataset.
 
 ### Layer 1: Smart Tree Suggestions
+
 - Engineer pastes ticket description: "User at Warner Robins reports Outlook keeps crashing after latest update"
 - AI suggests: "Recommended: Outlook/Email Issues tree → Start at 'Recent Update' branch"
 - Not just which tree — which branch to start at, skipping generic initial questions
 - **Implementation:** NLP parsing of ticket text, match against tree node content and tags
 
 ### Layer 2: Session-Driven Tree Evolution
+
 - Aggregate session data reveals patterns: "35% of engineers add 'Check MFA Token' after 'Auth Failed' node, and it resolves 80% of the time"
 - Generate suggestion to tree author: "Recommended new branch based on 18 successful sessions"
 - Author reviews and approves with one click — tree evolves from real usage
 - **Implementation:** Aggregation queries + LLM formatting suggestions + author approval UI
 
 ### Layer 3: AI Tree Generation
+
 - Senior describes: "We keep getting Azure AD Sync issues, no tree exists"
 - AI generates complete tree draft using:
   - Similar trees in the system
@@ -356,9 +404,11 @@ Three layers of AI, each building on the last, leveraging ResolutionFlow's uniqu
 - **Implementation:** RAG over tree corpus + session data + LLM generation + tree editor integration
 
 ### Why This Is the Moat
+
 ConnectWise or IT Glue could build a decision tree tool. But they don't have hundreds of structured session paths with outcomes to learn from. ResolutionFlow's data is structured by design — decision trees + session paths + outcomes — not unstructured ticket notes. That's a dataset purpose-built for learning optimal troubleshooting paths.
 
 ### Phasing
+
 - Layer 1 can ship independently with basic NLP
 - Layer 2 needs sufficient session volume (6+ months of real usage)
 - Layer 3 needs Layers 1 + 2 data + LLM integration
@@ -371,9 +421,11 @@ ConnectWise or IT Glue could build a decision tree tool. But they don't have hun
 **Effort:** Medium
 
 ### What
+
 Replace the tree library as the default landing page with a troubleshooting command center.
 
 ### Sections
+
 - **Resume sessions** — "VPN Issues - Acme Corp (started 20 min ago)" for incomplete sessions
 - **Quick starts** — Frequent tree+client combos: "File Share Access for Warner Robins — Quick start?" One tap.
 - **Team activity** — "Sarah completed 'AD Replication' for Client X (12 min)" — visibility without a standup
@@ -381,6 +433,7 @@ Replace the tree library as the default landing page with a troubleshooting comm
 - **Recurring alerts** — Clients with repeat issues that need attention
 
 ### Why
+
 Turns ResolutionFlow from a tool you visit per-ticket into something you keep open all day. Reduces friction from "open app → find tree → start session" to "open app → click the obvious next action."
 
 ---
@@ -388,22 +441,26 @@ Turns ResolutionFlow from a tool you visit per-ticket into something you keep op
 ## Suggested Build Order
 
 ### Near-term (build now, small effort, immediate value)
+
 1. **Session Scratchpad** (Idea 6) — must-have per Michael
 2. **Session Time Tracking** (Idea 1) — quick win, enhances every export
 3. **Command Output Capture** (Idea 3) — quick win, pairs with scratchpad
 
 ### Mid-term (build next, medium effort, high value)
+
 4. **Share Progress / Escalation** (Idea 2) — daily use for team collaboration
 5. **Recurring Issue Detection** (Idea 8) — leverages existing data immediately
 6. **Quick Actions Dashboard** (Bonus) — improves daily UX
 7. **Path Analytics** (Idea 5) — needs session volume, start collecting data now
 
 ### Later (larger effort, needs foundation)
+
 8. **Push Steps to Active Sessions** (Idea 4) — needs notification system
 9. **Tree Health Scores** (Idea 9) — needs session volume + analytics infrastructure
 10. **Multi-Tree Sessions** (Idea 7) — complex UX, large refactor
 
 ### Long-term vision
+
 11. **AI Tree Intelligence** (Idea 10) — phased rollout, ultimate differentiator
 
 ---

docs/plans/2026-02-04-session-scratchpad-design.md

@@ -0,0 +1,276 @@
+# Session Scratchpad Design
+
+> **Date:** February 4, 2026
+> **Source:** Feature Ideas Brainstorm - Idea 6
+> **Priority:** Must-have (per Michael)
+> **Category:** Context capture
+
+---
+
+## Overview
+
+A collapsible right sidebar during active sessions for capturing ambient data — IP addresses, error codes, server names, usernames — anything that doesn't fit a specific decision node's notes field.
+
+**Core principle:** Engineers already accumulate scraps of data during troubleshooting (from `ipconfig`, Event Viewer, phone conversations) that live on sticky notes or in their head. This gives it a home and includes it automatically in the export.
+
+**Data model:** Single freeform `Text` column on the `sessions` table with markdown formatting support. Not JSONB — we're starting simple with a plain text field that the engineer writes in freely.
+
+**Save behavior:** Auto-save with 1000ms debounce. No manual save button.
+
+**Layout:** Collapsible right sidebar in `TreeNavigationPage`, 300px when open, 48px toggle strip when collapsed.
+
+---
+
+## Section 1: Data Model
+
+### Database
+
+New `scratchpad` column on `sessions` table:
+
+```python
+# SQLAlchemy model (session.py)
+scratchpad: Mapped[Optional[str]] = mapped_column(Text, nullable=True, server_default=sa.text("''"))
+```
+
+### Migration
+
+```python
+# alembic revision -m "add scratchpad to sessions"
+def upgrade():
+    op.add_column('sessions', sa.Column('scratchpad', sa.Text(), nullable=True, server_default=sa.text("''")))
+    # Backfill existing rows
+    op.execute("UPDATE sessions SET scratchpad = '' WHERE scratchpad IS NULL")
+
+def downgrade():
+    op.drop_column('sessions', 'scratchpad')
+```
+
+- `server_default=sa.text("''")` ensures new rows get `""` at the database level
+- Backfill normalizes existing rows so no code path sees `NULL`
+
+### Schemas
+
+**SessionUpdate** — add field:
+```python
+scratchpad: Optional[str] = None
+```
+
+**SessionResponse** — add field with validator:
+```python
+scratchpad: str = ""
+
+@validator('scratchpad', pre=True, always=True)
+def normalize_scratchpad(cls, v):
+    return v or ""
+```
+
+**New schema — ScratchpadUpdate:**
+```python
+class ScratchpadUpdate(BaseModel):
+    scratchpad: str
+```
+
+---
+
+## Section 2: Frontend Component Design
+
+### Layout Restructuring
+
+Current `TreeNavigationPage` is a single-column `container mx-auto` div. Restructure to:
+
+```
+┌─────────────────────────────────────┬──────────────┐
+│  Existing tree navigation content   │  Scratchpad  │
+│  (header, breadcrumb, node card,    │  sidebar     │
+│   notes, back button, shortcuts)    │  (300px)     │
+│                                     │              │
+└─────────────────────────────────────┴──────────────┘
+```
+
+- Outer wrapper: `flex` layout
+- Main content: `flex-1 min-w-0` (prevents overflow)
+- Sidebar: Fixed 300px open, 48px collapsed
+- Collapse state persisted in `localStorage` as `scratchpad-collapsed`
+
+### ScratchpadSidebar Component
+
+**File:** `frontend/src/components/session/ScratchpadSidebar.tsx`
+
+**Props:**
+```typescript
+interface ScratchpadSidebarProps {
+  sessionId: string
+  initialContent: string
+  onSave: (content: string) => Promise<void>
+}
+```
+
+**Internal state:**
+- `content: string` — textarea value, initialized from `initialContent`, **only reset when `sessionId` changes** (via `useEffect` with `[sessionId]` dependency)
+- `isCollapsed: boolean` — sidebar open/closed, persisted in localStorage
+- `isSaving: boolean` — network request in flight
+- `hasUnsavedChanges: boolean` — `content !== lastSaved`
+- `lastSaved: string` — last successfully saved content
+- `showPreview: boolean` — toggle between edit and markdown preview
+
+### Save Flow
+
+1. User types → `content` updates → `hasUnsavedChanges = (content !== lastSaved)`
+2. After 1000ms of no typing (debounce), if `hasUnsavedChanges`, fire `onSave(content)`
+3. Set `isSaving = true` during request
+4. On success: `lastSaved = content`, `isSaving = false`, `hasUnsavedChanges = false`
+5. On blur: trigger immediate save if `hasUnsavedChanges` (cancel pending debounce)
+6. On error: keep `hasUnsavedChanges = true`, show error indicator
+
+### Save Indicator (bottom of sidebar)
+
+| State | Display |
+|-------|---------|
+| Idle, no changes | Nothing shown |
+| Typing, unsaved | "Unsaved changes" (muted text) |
+| Saving | "Saving..." with subtle spinner |
+| Just saved | "Saved" with checkmark (green, fades after 2s) |
+
+### beforeunload Warning
+
+- Register `window.addEventListener('beforeunload', handler)` only when `hasUnsavedChanges === true`
+- Handler calls `e.preventDefault()` to trigger browser's "unsaved changes" dialog
+- Clean up listener on unmount
+
+### Markdown Preview
+
+- Small toggle link below textarea: "Preview" / "Edit"
+- Preview renders current `content` using existing `MarkdownContent` component
+- No new dependencies required
+
+### Dimensions & Animation
+
+- Open width: 300px
+- Collapsed width: 48px (notepad icon button)
+- Transition: `width 200ms ease`
+- Toggle icon: Notepad/pencil icon from Lucide
+
+---
+
+## Section 3: Backend & API
+
+### New Endpoint
+
+```
+PATCH /api/v1/sessions/{id}/scratchpad
+```
+
+**Request body:**
+```json
+{ "scratchpad": "- Server IP: 192.168.1.50\n- Error: 0x80070005" }
+```
+
+**Response:** `SessionResponse` (full session object)
+
+**Handler logic:**
+1. Load session by ID
+2. Verify `session.user_id == current_user.id` (ownership check)
+3. Set `session.scratchpad = data.scratchpad`
+4. Commit and return updated session
+
+This is intentionally separate from the existing PUT endpoint. Auto-saving the scratchpad every second should not send the full session payload (`path_taken`, `decisions`, `custom_steps`) — that risks overwriting concurrent changes from decision tracking in the same session.
+
+### Frontend API Client
+
+```typescript
+// sessions.ts
+updateScratchpad: (id: string, content: string) =>
+  api.patch<Session>(`/api/v1/sessions/${id}/scratchpad`, { scratchpad: content })
+```
+
+### Router Registration
+
+Add to `backend/app/api/router.py` — the endpoint lives in the existing `sessions.py` endpoints file.
+
+---
+
+## Section 4: Export Integration
+
+Add a "Evidence / Reference" section to all three export generators, inserted **between the header block and "Troubleshooting Steps"**. Only rendered when `scratchpad.strip()` is non-empty.
+
+### Markdown Export
+
+```markdown
+## Evidence / Reference
+
+- Server IP: 192.168.1.50
+- Error code: 0x80070005
+- Affected user: jsmith@contoso.com
+
+---
+
+## Troubleshooting Steps
+```
+
+Scratchpad content rendered as-is (user writes markdown, it appears as markdown).
+
+### Text Export
+
+```
+EVIDENCE / REFERENCE
+--------------------
+- Server IP: 192.168.1.50
+- Error code: 0x80070005
+- Affected user: jsmith@contoso.com
+
+TROUBLESHOOTING STEPS
+---------------------
+```
+
+### HTML Export
+
+```html
+<h2>Evidence / Reference</h2>
+<div class="scratchpad" style="white-space: pre-wrap; background: #f9f9f9; padding: 12px; border-radius: 4px; margin-bottom: 20px;">
+- Server IP: 192.168.1.50
+- Error code: 0x80070005
+</div>
+<h2>Troubleshooting Steps</h2>
+```
+
+Start with `pre-wrap` for HTML export. Future enhancement: render scratchpad markdown to HTML.
+
+---
+
+## Files Changed
+
+| File | Action | Description |
+|------|--------|-------------|
+| `backend/app/models/session.py` | Modify | Add `scratchpad` field |
+| `backend/app/schemas/session.py` | Modify | Add to `SessionUpdate`, `SessionResponse` + new `ScratchpadUpdate` |
+| `backend/app/api/endpoints/sessions.py` | Modify | Add PATCH endpoint + scratchpad in all 3 export generators |
+| `backend/app/api/router.py` | Modify | Register new PATCH route |
+| `backend/alembic/versions/xxx_add_scratchpad.py` | Create | Migration with backfill |
+| `frontend/src/components/session/ScratchpadSidebar.tsx` | Create | New sidebar component |
+| `frontend/src/pages/TreeNavigationPage.tsx` | Modify | Add flex layout + render sidebar |
+| `frontend/src/api/sessions.ts` | Modify | Add `updateScratchpad` method |
+| `frontend/src/types/session.ts` | Modify | Add `scratchpad` to Session type |
+
+---
+
+## Design Decisions
+
+| Decision | Choice | Rationale |
+|----------|--------|-----------|
+| Storage type | `Text` column (not JSONB) | Freeform text is simpler; structured entries are a future enhancement |
+| Save mechanism | Auto-save with 1000ms debounce | Reduces friction — engineer never thinks about saving |
+| API endpoint | Dedicated PATCH (not reuse PUT) | Prevents scratchpad auto-saves from overwriting concurrent decision updates |
+| Conflict detection | None (last-write-wins) | Sessions are single-user; multi-user requires broader solution |
+| Markdown rendering | Preview toggle (not live) | Keeps the editing experience simple; preview is optional |
+| Collapse persistence | localStorage | Survives page refreshes, per-browser preference |
+
+---
+
+## Pairs With (Future)
+
+- **Command Output Capture** (Idea 3): Structured output at nodes + freeform notes in scratchpad = complete evidence
+- **Share Progress** (Idea 2): Scratchpad content included in shared summary
+
+---
+
+*Design finalized during brainstorming session, February 4, 2026*