TUI polish: paste debounce, status bar, crash logging, tool timers

Directory.Packages.props

@@ -16,6 +16,7 @@
     <PackageVersion Include="Microsoft.Extensions.Configuration.Json" Version="10.0.2" />
     <PackageVersion Include="Microsoft.Extensions.Hosting" Version="10.0.2" />
     <PackageVersion Include="OllamaSharp" Version="5.4.16" />
+    <PackageVersion Include="Termina" Version="0.5.1" />
   </ItemGroup>
   <!-- Serialization -->
   <ItemGroup>

IMPLEMENTATION_PLAN.md

@@ -200,9 +200,9 @@ Full task breakdowns:
 
 > **Gateway note:** Netclaw.App was temporarily changed from `Microsoft.NET.Sdk.Web`
 > to `Microsoft.NET.Sdk` (plain console host) for the proof-of-concept console
-> adapter. The web gateway (health endpoints, future API surface) must be
-> restored before Slack adapter or any multi-client scenario. Track as part of
-> Task 1.11 (CLI scaffold) or earlier if needed.
+> adapter. Task 1.11 restores `Microsoft.NET.Sdk.Web` for daemon modes per
+> SPEC-011. Single-process architecture validated by research in
+> `docs/research/agent-gateway-architecture.md`.
 
 ### Task 1.1: Framework protocol and persistence-safe message envelopes (DONE)
 
@@ -364,43 +364,52 @@ Done when:
 - [ ] CI tests pass without live provider credentials.
 - [ ] Tests for provider switching, fallback activation, tool calling round-trip.
 
-### Task 1.11: CLI scaffold with Cocona
+### Task 1.11: Daemon architecture scaffold with mode routing
 
 **PRD:** `docs/prd/PRD-004-cli-onboarding-and-config.md` (CLI-010, CLI-012)
+**Spec:** `docs/spec/SPEC-011-daemon-architecture.md`
 **OpenSpec:** `openspec/specs/netclaw-cli/spec.md`
 **OpenSpec Changes:** `openspec/changes/add-tui-adapter-and-config-hot-reload/` (Section 1)
-**Surface area:** CLI framework
+**Surface area:** CLI framework + gateway
 **Verification:** L1
 **Previously:** Task 1.13
 
-Replaces the bare console loop with a proper CLI framework. Restore web gateway
-hosting (`Microsoft.NET.Sdk.Web`) if needed for daemon mode.
+Replaces the bare console loop with mode-selected startup per SPEC-011.
+Daemon modes use `WebApplication.CreateBuilder()` (ASP.NET host for SignalR
+and health endpoints). Lightweight modes use `Host.CreateApplicationBuilder()`.
+Cocona is archived (Dec 2025) — replaced with simple arg routing.
 
 Done when:
-- [ ] Cocona and Termina package references added to `Directory.Packages.props` and `Netclaw.App.csproj`.
-- [ ] `Program.cs` rewritten as Cocona entry point with DI registration.
-- [ ] `RunCommand.cs` created for daemon mode (`netclaw run`).
-- [ ] Termina wired as hosted service for TUI commands.
+- [ ] Termina and System.Reactive package references added to `Directory.Packages.props`.
+- [ ] `Netclaw.App.csproj` SDK changed to `Microsoft.NET.Sdk.Web`, Termina added, Hosting removed.
+- [ ] `Program.cs` rewritten with mode routing: `run`/`chat`/headless use `WebApplication`, `init`/`doctor` use `Host`.
+- [ ] Shared config services extracted to `ConfigureConfigServices()` method.
+- [ ] Daemon-only services extracted to `ConfigureDaemonServices()` method.
+- [ ] SignalR hub mapped at `/hub/session` (stub — Phase 1 placeholder).
+- [ ] Health probe mapped at `/api/health/ready`.
+- [ ] `ConsoleChannel.cs` deleted (replaced by Termina TUI).
 - [ ] `dotnet build` passes with new dependencies.
 
 ### Task 1.12: TUI chat adapter (`netclaw chat`)
 
 **PRD:** `docs/prd/PRD-004-cli-onboarding-and-config.md` (CLI-011), `docs/prd/PRD-009-input-adapters-and-unified-input.md` (INPUT-005)
+**Spec:** `docs/spec/SPEC-011-daemon-architecture.md`
 **OpenSpec:** `openspec/specs/netclaw-input-adapters/spec.md`, `openspec/specs/netclaw-cli/spec.md`
 **OpenSpec Changes:** `openspec/changes/add-tui-adapter-and-config-hot-reload/` (Section 2)
 **Surface area:** TUI + adapter
 **Verification:** L3
 **Wireframe:** `docs/ui/TUI-001-command-wireframes.md` (netclaw chat)
 **Previously:** Task 1.14
 
+The TUI uses `SessionPipeline` directly — no SignalR indirection. Same
+in-process pattern as the current ConsoleChannel/HeadlessChannel.
+
 Done when:
-- [ ] `TuiInputAdapter` implementing adapter contract (`SendUserMessage` with entity key `tui/{sessionId}`).
-- [ ] `ChatCommand.cs` hosts actor system in-process and launches TUI.
 - [ ] `ChatPage.cs` with `StreamingTextNode` (scrollable history) and `TextInputNode` (multi-line input).
-- [ ] `ChatViewModel.cs` with session lifecycle and broadcast subscription.
+- [ ] `ChatViewModel.cs` uses `SessionPipeline` directly via DI — calls `CreateAsync()`, pushes `ChannelInput`, subscribes to `SessionOutput`.
 - [ ] Inline tool activity panel (completed with duration, in-progress with spinner).
 - [ ] MCP status indicator in status bar (green/yellow/red).
-- [ ] E2E: user types → `SendUserMessage` → session actor → LLM → streaming response in TUI.
+- [ ] E2E: user types → `ChannelInput` → `SessionPipeline` → session actor → LLM → streaming response in TUI.
 
 ---
 
@@ -703,3 +712,6 @@ the linked research documents.
 - `docs/research/actor-llm-optimization-patterns.md` — Prompt caching,
   safety circuit breakers, parallel execution, streaming, retry, and
   sub-agent isolation patterns for actor-based LLM systems
+- `docs/research/agent-gateway-architecture.md` — Single-process vs
+  multi-process architecture analysis (OpenClaw, IronClaw, ZeroClaw).
+  Validates single-process model for homelab/personal agent use.

docs/prd/PRD-001-netclaw-mvp.md

@@ -55,9 +55,37 @@ irrelevant — the differentiator is the instructions attached to the context.
 10. Agent can modify its own configuration through conversation.
 11. MCP integration provides external memory (Memorizer) and tool capabilities.
 
+## Daemon Architecture
+
+Netclaw runs as a single OS process. All components — Akka actor system,
+persistence, gateway endpoints, TUI, and tool execution — share one process
+boundary. This is validated by architecture research across comparable projects
+(see `docs/research/agent-gateway-architecture.md`).
+
+The executable supports multiple modes selected by command-line arguments:
+
+| Mode | Command | Description |
+|------|---------|-------------|
+| Daemon | `netclaw run` | Full service stack. Slack, schedules, SignalR hub, health endpoint. |
+| Chat | `netclaw chat` | Full service stack + Termina TUI. Interactive chat via `SessionPipeline`. |
+| Headless | `netclaw -p "..."` | Full service stack + headless client. Single turn, exits on completion. |
+| Init | `netclaw init` | Lightweight. Config services only, no Akka/persistence. Reentrant TUI wizard. |
+| Doctor | `netclaw doctor` | Lightweight. Config services only. Health checks and diagnostics. |
+
+In-process channels (TUI, headless) use `SessionPipeline` directly — no network
+hop. The SignalR hub exists for future remote clients (Blazor ops console) and
+uses the same `SessionPipeline` internally.
+
+Tools execute on the host process. This is the only model that works for Slack
+(no client to delegate to), scheduled tasks (autonomous), and Docker deployment
+(tools need access to the host's Docker socket).
+
+See `SPEC-011-daemon-architecture.md` for full specification.
+
 ## Non-Goals (MVP)
 
-- Multi-process gateway/agent split
+- Multi-process gateway/agent split (validated by architecture research —
+  single-process is the correct model for homelab/personal agent use)
 - Ambient channel monitoring with per-channel instructions
 - Webhook ingress (Tailscale Serve / Cloudflare Tunnel)
 - Sub-agent model routing (cheaper models for high-token tasks)
@@ -269,3 +297,5 @@ schedule changes → timer reconfiguration).
 - Agent Personality and Memory: PRD-007
 - Scheduling: PRD-008
 - Input Adapters: PRD-009 (post-MVP)
+- Daemon Architecture: SPEC-011
+- Architecture Research: `docs/research/agent-gateway-architecture.md`

docs/prd/PRD-004-cli-onboarding-and-config.md

@@ -23,19 +23,34 @@ diagnostics using CLI commands and guided output.
 
 ## CLI Framework
 
-- **Cocona** for command routing (lightweight, convention-based, DI integration)
+- **Simple arg routing** in `Program.cs` for mode selection (Cocona is archived
+  as of Dec 2025 — replaced with direct `args[0]` routing)
 - **Termina 0.5.1** for interactive TUI commands (`netclaw init`, `netclaw chat`)
-- All other commands use plain console output via Cocona
+- All other commands use plain console output
 - `netclaw run` is the explicit daemon entry point (Slack + timers + health
   endpoints, no TUI)
+- All CLI modes are in-process — no REST client in Phase 1
+- Configuration is privileged local file I/O, never exposed over the wire
+  (contains API keys/secrets)
 
 ## Two-Phase Onboarding
 
 ### Phase 1: CLI Wizard (`netclaw init`)
 
-Technical setup, no LLM required:
+Technical setup, no LLM required. `netclaw init` runs as a **lightweight mode**
+— no Akka actor system, no persistence, no SignalR. Only config services are
+booted. Provider testing uses direct DI service calls (`ChatClientFactory`),
+not REST endpoints.
 
-1. LLM provider configuration (OpenRouter API key, model selection)
+The wizard is **reentrant** — re-running `netclaw init` detects existing config
+and shows a section dashboard with status per section. Each section is
+independently enterable for modification. First-run guides linearly through
+all steps.
+
+Steps:
+
+1. LLM provider configuration (endpoint URL, API key, model selection,
+   connectivity test via direct HTTP to provider)
 2. Slack app setup (bot token, app token for Socket Mode)
 3. PostgreSQL connection string
 4. ACL bootstrap (owner identity, initial channel rules)
@@ -175,7 +190,7 @@ Results are persisted to the environment inventory file.
 
 `netclaw init` and `netclaw chat` SHALL use Termina 0.5.1 for interactive TUI
 rendering. All other commands SHALL use plain console output. TUI commands SHALL
-launch Termina as a hosted service within the Cocona command handler.
+launch Termina as a hosted service within the mode-selected host builder.
 
 ### CLI-011 Local Chat Adapter
 
@@ -219,3 +234,5 @@ rendering. This is the primary production entry point.
 - MCP setup: PRD-006
 - Memory and personality: PRD-007
 - Scheduling: PRD-008
+- Daemon architecture: SPEC-011
+- TUI wireframes: TUI-001

docs/prd/PRD-009-input-adapters-and-unified-input.md

@@ -37,10 +37,11 @@ child session actor.
 ### MVP Input Adapters
 
 **Local TUI Adapter** (Phase 1):
-- Hosted in-process by `netclaw chat` command
+- Hosted in-process by `netclaw chat` command (daemon mode — full service stack)
+- Uses `SessionPipeline` directly — no SignalR indirection for in-process channels
 - Receives keyboard input via Termina TextInputNode
-- Produces `SendUserMessage` commands with entity key `tui/{sessionId}`
-- Subscribes to session broadcasts for reply delivery
+- Pushes `ChannelInput` to `SessionPipeline` input sink
+- Subscribes to `SessionOutput` source for rendering
 - Renders responses as streaming text via StreamingTextNode
 - Displays tool invocation status inline (name, duration, spinner)
 - Shows MCP server connectivity in status bar
@@ -75,9 +76,12 @@ child session actor.
 - Each webhook hit creates a new session with source-specific instructions
 
 **Web UI Adapter** (Phase 5):
-- WebSocket connection from Blazor Server ops console
+- Connects via SignalR hub (`/hub/session`) — the gateway surface documented in
+  SPEC-011
 - Provides interactive session control and real-time updates
-- Same pub/sub broadcast pattern as Slack adapter
+- Same `SessionPipeline` abstraction as all other channels
+- SignalR hub is mapped from Phase 1 (for future remote clients) but not used
+  by TUI or headless modes
 
 ## Adapter Contract
 
@@ -174,3 +178,5 @@ SHALL display tool invocation status inline between user message and response.
 - Security (ACL per source): PRD-002
 - Scheduling (timer source): PRD-008
 - Ops console (web UI source): PRD-003
+- Daemon architecture and gateway surface: SPEC-011
+- SessionPipeline and channel abstraction: `src/Netclaw.Actors/Channels/ChannelPipeline.cs`