Phase 15 — GeneratorWindow

Four lightweight tool windows (web_search, web_fetch, get_datetime, get_location) become half-height, stacked two-per-slot in a dedicated utilities column (col 4). This frees a full slot at col 2 row 0 for GeneratorWindow.

_tile_windows() changes

Add a _place_half(win, col, row, half) helper alongside the existing _place(win, col, row):

half_panel_h = H // 4 - tb_h

def _place_half(win, col: int, row: int, half: int) -> None:
    # half=0: top half of the row slot; half=1: bottom half
    x = x0 + main_w + col * panel_w
    y = y0 + row * (H // 2) + half * (H // 4) + tb_h
    win.setGeometry(x, y, panel_w, half_panel_h)

Updated _TOOL_PANEL_SLOTS

Slots now carry an optional half value (0 or 1). Full-height tools keep the existing 2-tuple; half-height tools use a 3-tuple:

_TOOL_PANEL_SLOTS: dict[str, tuple] = {
    "search_memory":  (0, 0),          # full height
    "search_library": (1, 0),          # full height
    "search_papers":  (3, 0),          # full height
    "web_search":     (4, 0, 0),       # half height — top of row 0
    "web_fetch":      (4, 0, 1),       # half height — bottom of row 0
    "get_datetime":   (4, 1, 0),       # half height — top of row 1
    "get_location":   (4, 1, 1),       # half height — bottom of row 1
}

The tiling loop checks tuple length to dispatch to the right helper:

for name, win in self._tool_windows.items():
    slot = _TOOL_PANEL_SLOTS.get(name)
    if slot:
        if len(slot) == 3:
            _place_half(win, col=slot[0], row=slot[1], half=slot[2])
        else:
            _place(win, col=slot[0], row=slot[1])

New file: src/local/ui/generator_window.py

Identity bar

Static — set once at window construction from generator config: instance_id · model · temperature · num_ctx. Uses the instance_id from config/system.yaml (new in this phase; defaults to hostname if absent).

State + context row

State badge colour-coded identically to CriticWindow's state transitions: IDLE=grey, RECEIVING=blue, GENERATING=amber, DISPATCHING_TOOL=purple, ERROR=red. Token count and fill bar mirror the sidebar ContextGauge but with numeric detail. Updates on every generator.status event.

Tool registry

Chip-style list of currently registered tool names. Updates live as tools announce on tool.schema. Count badge in section header. Clicking a tool name is a no-op for now (future: jump to that tool's window).

System prompt

Collapsible section. Read-only QTextEdit showing the active system prompt. Collapsed by default (saves vertical space). Content set once at startup from generator.status initial payload.

Transitions log

Scrolling strip of state transitions with timestamps — same pattern as CriticWindow.append_transition(). Subscribes to existing AGENT_TRANSITION subject (already published by GeneratorAgent state machine). Shows last 50 transitions.

Peers panel

Table stub: Instance ID | State | Load | Last seen. Shows "(none detected)" until agent.heartbeat events arrive. Hidden entirely if no peers ever appear. Forward-compatible with Phase 16 distributed work — no logic needed now beyond displaying what arrives on the heartbeat subject.

A new subject generator.status carries full generator state snapshots. Published on: startup, every state transition, every tool registration, and after each generation turn.

// subject: generator.status
{
  "instance_id":    "local2-macbook",
  "respondent_id":  "A",
  "model":          "gemma4:e4b",
  "temperature":    0.1,
  "num_ctx":        128000,
  "state":          "idle",           // matches GeneratorState enum lowercase
  "token_count":    47200,            // last prompt_eval_count, 0 before first turn
  "tool_names":     ["search_memory", "web_search", ...],
  "system_prompt":  "You are a helpful assistant..."
}

Publish points in GeneratorAgent

• Startup — after _request_schemas(), publish initial status with token_count=0
• State transitions — after each self._sm.transition() call
• Tool registration — after _register_tool_schema() updates the registry
• Post-generation — after set_token_count() is called (token_count now known)

A helper _publish_status() builds the payload from instance fields and publishes. Called from each of the above points.

instance_id source

Read from config/system.yaml:

# config/system.yaml  (new file)
instance_id: "local2-macbook"

If the key is absent or the file doesn't exist, fall back to socket.gethostname(). This config file is also where peer/distributed config will live in Phase 16.

Add generator_status signal to BusLogger and handle GENERATOR_STATUS subject in log_envelope():

generator_status = Signal(dict)   # full status snapshot

# in log_envelope():
if subject == GENERATOR_STATUS:
    self.generator_status.emit(raw)
    return

Subscribe to GENERATOR_STATUS in _start_bus_monitor(). Connect signal to self._generator_window.update_status(data).

GeneratorWindow is constructed at startup alongside CriticWindow and MemoryWindow, and placed at col 2, row 0 in _tile_windows().