resolutionflow/docs/superpowers/specs/2026-03-28-tasklane-minimize-and-resolve-docs-design.md

Task Lane Minimize/Reopen + Resolve Documentation Streaming

Date: 2026-03-28 Status: Approved

---

Problem

Two UX issues in the Assistant Chat page:

1. Task Lane close destroys state. The X button calls clearTaskState() and hides the panel. There's no way to bring it back without getting a new AI response with markers. Engineers lose in-progress task responses.

2. Conclude → Resolved hangs. The resolve flow blocks on an LLM call to generate documentation. The modal shows "Generating..." for 2-5s+ with no progressive feedback. The generated output needs to be structured ticket notes engineers can copy into their PSA.

---

Feature 1: Task Lane Minimize/Reopen

Close Button Change

• Replace the X icon (X from Lucide) with PanelRightClose to signal "collapse" not "destroy."
• On click: set showTaskLane = false. Do NOT call clearTaskState(sessionId).
• Task state (questions, actions, user responses) remains in sessionStorage and backend pending_task_lane.

Reopen Pill on Input Toolbar

• New pill button in the chat input toolbar row, alongside Attach / Paste Logs / Conclude.
• Visibility condition: (activeQuestions.length > 0 || activeActions.length > 0) && !showTaskLane
• Label: Tasks (N) where N = activeQuestions.length + activeActions.length
• Icon: ListChecks from Lucide
• Action: setShowTaskLane(true)
• Style: Same ghost button style as Attach/Paste Logs — muted text, hover highlight.

Files Changed

File	Change
frontend/src/components/assistant/TaskLane.tsx	Replace X icon with PanelRightClose in header
frontend/src/pages/AssistantChatPage.tsx	Remove clearTaskState() from onClose handler; add Tasks pill to input toolbar

---

Feature 2: Resolve with Streaming Documentation

Two-Phase Resolve Flow

Phase 1 — Instant resolve:

1. User clicks Resolve in ConcludeSessionModal, enters optional notes, confirms.
2. Frontend calls POST /ai-sessions/{id}/resolve with { resolution_summary }.
3. Backend sets session status to resolved, saves summary, returns immediately. No LLM call on this path.
4. Modal transitions to the summary step instantly, showing a skeleton loading state for "Ticket Notes."

Phase 2 — Streamed doc generation:

1. Immediately after phase 1, frontend opens an SSE connection to GET /ai-sessions/{id}/documentation/stream.
2. Backend streams the structured ticket notes as they generate, token by token.
3. Frontend renders chunks progressively into the summary step using MarkdownContent.
4. When stream completes, show a "Copy to Clipboard" button.

Ticket Notes Format

The LLM generates four structured sections:

## Problem Summary
[What the engineer reported / intake context]

## Steps Taken
[Key diagnostic steps and findings from conversation]

## Resolution
[What fixed it / final action taken]

## Next Steps
[Follow-up items, if any — or "None"]

Fallback Behavior

• If SSE stream fails or times out (30s): fall back to non-streaming GET /ai-sessions/{id}/documentation.
• If that also fails: show "Documentation unavailable" with a "Copy Conversation" button that formats the raw conversation_messages into a basic structured summary (no LLM, pure template).

Copy to Clipboard

• Prominent button below the rendered ticket notes.
• Copies the full markdown text.
• Toast confirmation: "Ticket notes copied."

AI Optimizations

1. Streaming (SSE endpoint):

• New endpoint: GET /ai-sessions/{id}/documentation/stream
• Returns text/event-stream via FastAPI StreamingResponse.
• Uses Anthropic's client.messages.stream() for token-level streaming.
• Time-to-first-byte drops from ~2-5s to ~200ms.

2. Prompt caching:

• Apply cache_control: {"type": "ephemeral"} to the system prompt and conversation context prefix in the documentation generation call.
• Pattern already exists in assistant_chat_service.py (_call_anthropic_cached).
• Repeat calls for the same session (regenerate, different doc types) hit cache — up to 85% faster, 90% cheaper on input tokens.

3. Client reuse:

• Singleton AsyncAnthropic client instead of creating a new one per call.
• Eliminates connection setup overhead.

Model tier: Haiku (already configured as quick_action tier) — fastest model, appropriate for summarization.

Backend Changes

File	Change
backend/app/api/endpoints/ai_sessions.py	Modify resolve_session to only set status + save summary (remove the blocking get_session_documentation call); add new SSE streaming endpoint for doc generation
backend/app/services/flowpilot_engine.py	Add stream_session_documentation() generator function using client.messages.stream()
backend/app/core/ai_provider.py	Add generate_text_stream() method returning async iterator; singleton client
backend/app/services/flowpilot_engine.py	Add prompt caching to _build_status_update_prompt call path

Frontend Changes

File	Change
frontend/src/components/assistant/ConcludeSessionModal.tsx	Two-phase flow: instant resolve → streaming doc render with skeleton → copy button
frontend/src/api/aiSessions.ts	Add streamDocumentation(sessionId) using fetch + ReadableStream
frontend/src/pages/AssistantChatPage.tsx	Update handleConclude to not await documentation

ConcludeSessionModal Summary Step Layout

┌─────────────────────────────────────┐
│  Session Resolved                   │
│                                     │
│  ┌─ Ticket Notes ─────────────────┐ │
│  │ ## Problem Summary             │ │
│  │ [streaming text...]            │ │
│  │                                │ │
│  │ ## Steps Taken                 │ │
│  │ [streaming text...]            │ │
│  │                                │ │
│  │ ## Resolution                  │ │
│  │ [streaming text...]            │ │
│  │                                │ │
│  │ ## Next Steps                  │ │
│  │ [streaming text...]            │ │
│  └────────────────────────────────┘ │
│                                     │
│  [ Copy to Clipboard ]    [ Done ]  │
└─────────────────────────────────────┘

---

Out of Scope

• Client-facing update generation (future feature, different audience prompt)
• Direct PSA posting (requires ConnectWise integration to be wired to sessions)
• Task lane drag-to-reorder
• Task lane persistence across browser tabs (sessionStorage is per-tab by design)