Stream reasoning tokens into interactive CLI

AGENTS.md

@@ -40,6 +40,41 @@ npm test            # Run tests
 ./cli/bin/mew.js space up
 ```
 
+## Quick Testing Guide
+```bash
+# Create test space (from repo root)
+cd tests && mkdir test-feature && cd test-feature
+../../cli/bin/mew.js space init --template coder-agent .
+cd ../.. && npm install  # Link local packages
+
+# Start space
+cd tests/test-feature
+../../cli/bin/mew.js space up  # Default port 8080
+
+# Send messages via HTTP API (replace test-feature with your folder name)
+curl -X POST 'http://localhost:8080/participants/human/messages?space=test-feature' \
+  -H 'Content-Type: application/json' \
+  -H 'Authorization: Bearer human-token' \
+  -d '{"type": "chat", "content": "Hello"}'
+
+# View logs (non-blocking for coding agents)
+pm2 logs --nostream --lines 50     # Last 50 lines, exits immediately
+pm2 logs gateway --nostream         # Gateway logs only, no streaming
+pm2 logs coder --nostream --err     # Agent errors only
+tail -n 100 ~/.pm2/logs/gateway-out.log  # Direct file access
+
+# Check responses in curl (add -v for verbose)
+curl -v -X POST ...  # Shows request/response headers
+
+# Interactive mode (better for development)
+../../cli/bin/mew.js space connect
+# Paste JSON directly to send protocol messages
+# Type normally for chat messages
+
+# Clean up
+../../cli/bin/mew.js space down
+```
+
 ## Development Workflow
 1. Read existing patterns before implementing
 2. Update types first, then implementations
@@ -58,6 +93,7 @@ npm test            # Run tests
 - DO follow TypeScript strict mode
 - DO run relevant tests after changes
 - DO update specs if changing protocol
+- DO send stream/close when done with a stream (agents must clean up)
 
 ## Making Changes
 1. **Protocol changes**: Update spec/draft → types → implementations → tests

CLAUDE.md

@@ -40,6 +40,41 @@ npm test            # Run tests
 ./cli/bin/mew.js space up
 ```
 
+## Quick Testing Guide
+```bash
+# Create test space (from repo root)
+cd tests && mkdir test-feature && cd test-feature
+../../cli/bin/mew.js space init --template coder-agent .
+cd ../.. && npm install  # Link local packages
+
+# Start space
+cd tests/test-feature
+../../cli/bin/mew.js space up  # Default port 8080
+
+# Send messages via HTTP API (replace test-feature with your folder name)
+curl -X POST 'http://localhost:8080/participants/human/messages?space=test-feature' \
+  -H 'Content-Type: application/json' \
+  -H 'Authorization: Bearer human-token' \
+  -d '{"type": "chat", "content": "Hello"}'
+
+# View logs (non-blocking for coding agents)
+pm2 logs --nostream --lines 50     # Last 50 lines, exits immediately
+pm2 logs gateway --nostream         # Gateway logs only, no streaming
+pm2 logs coder --nostream --err     # Agent errors only
+tail -n 100 ~/.pm2/logs/gateway-out.log  # Direct file access
+
+# Check responses in curl (add -v for verbose)
+curl -v -X POST ...  # Shows request/response headers
+
+# Interactive mode (better for development)
+../../cli/bin/mew.js space connect
+# Paste JSON directly to send protocol messages
+# Type normally for chat messages
+
+# Clean up
+../../cli/bin/mew.js space down
+```
+
 ## Development Workflow
 1. Read existing patterns before implementing
 2. Update types first, then implementations
@@ -58,6 +93,7 @@ npm test            # Run tests
 - DO follow TypeScript strict mode
 - DO run relevant tests after changes
 - DO update specs if changing protocol
+- DO send stream/close when done with a stream (agents must clean up)
 
 ## Making Changes
 1. **Protocol changes**: Update spec/draft → types → implementations → tests

cli/spec/draft/SPEC.md

@@ -1316,7 +1316,8 @@ The default interactive mode uses Ink (React for CLI) to provide a modern termin
 - Chat acknowledgement and cancellation shortcuts (`/ack`, `/cancel`) to manage UI queues
 - Participant control shortcuts (`/status`, `/pause`, `/resume`, `/forget`, `/clear`, `/restart`, `/shutdown`)
 - Stream negotiation helpers (`/stream request`, `/stream close`, `/streams`) with live frame previews
-- Reasoning status card with `/reason-cancel` hint and automatic stream mirroring for thoughts
+- Reasoning bar with token streaming, action display, and `/reason-cancel` control
+- Fixed-height UI components to prevent jitter during real-time updates
 
 **Architecture:**
 - Message history uses Ink's `Static` component for native scrolling
@@ -1330,6 +1331,90 @@ The default interactive mode uses Ink (React for CLI) to provide a modern termin
 - **Participant Status**: Displays most recent `participant/status` payloads (tokens, context occupancy, latency) per sender
 - **Pause State**: Highlights active pauses applied to the CLI participant with time remaining indicators
 - **Stream Monitor**: Lists current `stream/open` sessions and the latest raw frames decoded from `#streamId#` WebSocket traffic
+- **Collapsed Summary**: When the board is hidden, the status bar surfaces a compact `acks`, `status`, and `paused` summary so operators can monitor high-level state without losing transcript space. The board remains collapsed unless toggled with `/ui board open`; `/ui board close` re-collapses it and `/ui board auto` re-enables auto-collapse when no activity is present.
+
+#### Reasoning Bar Display
+
+The advanced interactive mode includes a dedicated reasoning bar that appears when agents engage in reasoning sessions. This component provides real-time feedback about the agent's thought process, even when the underlying model doesn't stream the actual reasoning content.
+
+**Visual Design:**
+- Cyan rounded border distinguishes it from regular messages
+- Fixed height (8 lines) to prevent UI jitter during streaming
+- Animated spinner indicates active processing
+- Horizontal layout spreads information across the width
+
+**Information Display:**
+```
+╭─────────────────────────────────────────────────────────────╮
+│ ⠋ mew is thinking     90s   32 tokens                      │
+│ Using tool mcp-fs-bridge /read_text_file                    │
+│ ...preview of reasoning text (max 3 lines)...               │
+│                                                              │
+│ Use /reason-cancel to interrupt reasoning.                  │
+╰─────────────────────────────────────────────────────────────╯
+```
+
+**Content Elements:**
+1. **Status Line**:
+   - Animated spinner (⠋⠙⠹⠸⠼⠴⠦⠧⠇⠏) for visual activity
+   - Participant name ("mew is thinking")
+   - Elapsed time in seconds
+   - Token count or thought count depending on available data
+
+2. **Action Line** (when available):
+   - Shows current tool or action being executed
+   - Extracted from `reasoning/thought` payload's `action` field
+   - Example: "Using tool mcp-fs-bridge /read_text_file"
+
+3. **Text Preview**:
+   - Limited to 3 lines and 400 characters to prevent height changes
+   - Shows latest reasoning text or message
+   - Italicized gray text for visual distinction
+   - Ellipsis prefix when truncated
+
+4. **Control Hint**:
+   - Always displayed at bottom: "Use /reason-cancel to interrupt reasoning."
+   - Provides user with clear action to stop long-running reasoning
+
+**Streaming and Token Counting:**
+
+The reasoning bar handles two types of progress indication:
+
+1. **Token Streaming** (OpenAI-style models):
+   - Receives token count updates via `stream/data` frames
+   - Displays running total as "X tokens"
+   - Updates in real-time as tokens are generated
+   - No actual reasoning content streamed (model limitation)
+
+2. **Thought Counting** (legacy/fallback):
+   - Counts number of `reasoning/thought` messages
+   - Displays as "X thoughts"
+   - Used when token information not available
+
+**Implementation Details:**
+
+- **Fixed Layout**: All elements use fixed heights to prevent layout shifts:
+  - Main container: `height: 8`
+  - Action line: `height: 1`
+  - Text preview: `height: 3` (matches `maxLines`)
+  - Always renders elements with space character fallback to avoid Ink rendering errors
+
+- **State Management**:
+  - Tracks `activeReasoning` in UI state
+  - Updates on `reasoning/start` messages
+  - Clears on `reasoning/conclusion` or `/reason-cancel`
+  - Merges token metrics from multiple sources (thoughts and streams)
+
+- **Stream Integration**:
+  - Automatically subscribes to reasoning streams via `stream/request`
+  - Processes `stream/open` responses to track stream IDs
+  - Parses `#streamId#data` frames for token updates
+  - Handles `stream/close` for cleanup
+
+**User Interactions:**
+- `/reason-cancel [reason]` - Sends cancellation request to stop reasoning
+- Escape key can be used in some contexts to cancel
+- Visual feedback shows cancellation was sent
 
 ### Debug Mode (--debug flag)
 
@@ -1519,6 +1604,7 @@ Verbose mode (`/verbose`) shows full JSON messages.
 /help                       Show available commands and shortcuts
 /verbose                    Toggle verbose output (show full JSON)
 /ui-clear                   Clear local UI buffers (history, status board)
+/ui board [open|close|auto] Control Signal Board docking behaviour
 /exit                       Disconnect and exit
 /ack [selector] [status]    Acknowledge pending chat message(s)
 /cancel [selector] [reason] Cancel pending chat message(s)
@@ -1628,4 +1714,4 @@ This CLI successfully implements the test plan when:
 2. Gateway starts and accepts connections
 3. Agents respond appropriately
 4. FIFO mode enables test automation
-5. Messages flow correctly between participants
+5. Messages flow correctly between participants

cli/src/commands/gateway.js

@@ -204,7 +204,7 @@ gateway
     const wss = new WebSocket.Server({ server });
 
     // Track spaces and participants
-    const spaces = new Map(); // spaceId -> { participants: Map(participantId -> ws) }
+    const spaces = new Map(); // spaceId -> { participants: Map(participantId -> ws), streamCounter: number, activeStreams: Map }
 
     // Track participant info
     const participantTokens = new Map(); // participantId -> token
@@ -563,7 +563,34 @@ gateway
 
       ws.on('message', async (data) => {
         try {
-          const message = JSON.parse(data.toString());
+          const dataStr = data.toString();
+
+          // Check if this is a stream data frame (format: #streamID#data)
+          if (dataStr.startsWith('#') && dataStr.indexOf('#', 1) > 0) {
+            const secondHash = dataStr.indexOf('#', 1);
+            const streamId = dataStr.substring(1, secondHash);
+
+            // Forward stream data to all participants in the space
+            if (spaceId && spaces.has(spaceId)) {
+              const space = spaces.get(spaceId);
+
+              // Verify stream exists and belongs to this participant
+              const streamInfo = space.activeStreams.get(streamId);
+              if (streamInfo && streamInfo.participantId === participantId) {
+                // Forward to all participants
+                for (const [pid, pws] of space.participants.entries()) {
+                  if (pws.readyState === WebSocket.OPEN) {
+                    pws.send(data); // Send raw data frame
+                  }
+                }
+              } else {
+                console.log(`[GATEWAY WARNING] Invalid stream ID ${streamId} from ${participantId}`);
+              }
+            }
+            return; // Don't process as JSON message
+          }
+
+          const message = JSON.parse(dataStr);
 
           // Validate message
           const validationError = validateMessage(message);
@@ -593,7 +620,11 @@ gateway
 
             // Create space if it doesn't exist
             if (!spaces.has(spaceId)) {
-              spaces.set(spaceId, { participants: new Map() });
+              spaces.set(spaceId, {
+                participants: new Map(),
+                streamCounter: 0,
+                activeStreams: new Map()
+              });
             }
 
             // Add participant to space
@@ -942,6 +973,65 @@ gateway
             envelope.correlation_id = [envelope.correlation_id];
           }
 
+          // Handle stream/request - gateway must respond with stream/open
+          if (envelope.kind === 'stream/request' && spaceId && spaces.has(spaceId)) {
+            const space = spaces.get(spaceId);
+
+            // Generate unique stream ID
+            space.streamCounter++;
+            const streamId = `stream-${space.streamCounter}`;
+
+            // Track the stream
+            space.activeStreams.set(streamId, {
+              requestId: envelope.id,
+              participantId: participantId,
+              direction: envelope.payload?.direction || 'unknown',
+              created: new Date().toISOString()
+            });
+
+            // Send stream/open response
+            const streamOpenResponse = {
+              protocol: 'mew/v0.3',
+              id: `stream-open-${Date.now()}-${Math.random().toString(36).substr(2, 9)}`,
+              ts: new Date().toISOString(),
+              from: 'gateway',
+              to: [participantId],
+              kind: 'stream/open',
+              correlation_id: [envelope.id],
+              payload: {
+                stream_id: streamId,
+                encoding: 'text'
+              }
+            };
+
+            // Send stream/open to ALL participants (MEW Protocol visibility)
+            for (const [pid, pws] of space.participants.entries()) {
+              if (pws.readyState === WebSocket.OPEN) {
+                pws.send(JSON.stringify(streamOpenResponse));
+              }
+            }
+
+            if (options.logLevel === 'debug') {
+              console.log(`[GATEWAY DEBUG] Assigned stream ID ${streamId} for request from ${participantId}`);
+            }
+          }
+
+          // Handle stream/close - clean up active streams
+          if (envelope.kind === 'stream/close' && spaceId && spaces.has(spaceId)) {
+            const space = spaces.get(spaceId);
+
+            // Find and remove the stream (may be referenced by correlation_id)
+            if (envelope.payload?.stream_id) {
+              const streamId = envelope.payload.stream_id;
+              if (space.activeStreams.has(streamId)) {
+                space.activeStreams.delete(streamId);
+                if (options.logLevel === 'debug') {
+                  console.log(`[GATEWAY DEBUG] Closed stream ${streamId}`);
+                }
+              }
+            }
+          }
+
           // ALWAYS broadcast to ALL participants - MEW Protocol requires all messages visible to all
           if (spaceId && spaces.has(spaceId)) {
             const space = spaces.get(spaceId);