Redesign the Netclaw chat TUI

Directory.Packages.props

@@ -76,7 +76,7 @@
     <PackageVersion Include="Cronos" Version="0.13.0" />
     <PackageVersion Include="Netclaw.SkillClient" Version="0.4.1" />
     <PackageVersion Include="ShellSyntaxTree" Version="0.3.3" />
-    <PackageVersion Include="Termina" Version="0.16.1" />
+    <PackageVersion Include="Termina" Version="0.17.0-beta.6" />
   </ItemGroup>
   <!-- Serialization -->
   <ItemGroup>
@@ -142,4 +142,4 @@
   <ItemGroup>
     <NuGetAuditSuppress Include="https://github.com/advisories/GHSA-2m69-gcr7-jv3q" />
   </ItemGroup>
-</Project>
+</Project>

IMPLEMENTATION_PLAN.md

@@ -110,6 +110,33 @@ the smallest repeatable manual script plus expected output.
 
 ## NOW
 
+### Priority: Redesign The Developer Chat TUI
+
+**PRDs:** `docs/prd/PRD-004-cli-onboarding-and-config.md`, `docs/prd/PRD-009-input-adapters-and-unified-input.md`
+**Specs:** `docs/spec/SPEC-002-session-lifecycle-and-protocol.md`, `docs/spec/SPEC-004-cli-contract.md`, `docs/spec/SPEC-010-testing-and-smoke-strategy.md`, `docs/spec/SPEC-011-daemon-architecture.md`
+**OpenSpec:** `openspec/changes/redesign-netclaw-chat-tui/`
+**Surface area:** session output, SignalR, Termina, chat command, resume, input, approvals, copy
+**Verification:** L3 plus the Termina cross-platform matrix
+
+The user promoted this work into `NOW`. The chat command must expose the
+daemon's structured event model through a clear developer interface.
+
+The design is extend-only. Existing protocol fields, enum values, presentation
+modes, and non-chat applications keep their current behavior and defaults.
+
+Done when:
+
+- [ ] Termina adds an opt-in inline mode while full-screen mode stays the
+  default.
+- [ ] Netclaw shows thoughts, parallel tools, sub-agents, approvals, file
+  changes, errors, usage, and compaction as structured events.
+- [ ] Structured resume data extends `RecentMessages` without removing it.
+- [ ] The composer supports `Shift+Enter`, draft restoration, and double
+  Escape.
+- [ ] Native terminal scrollback, mouse-wheel input, semantic copy, and the
+  inspector pass their defined proof matrix.
+- [ ] Headless tests and native smoke tests prove the critical chat flows.
+
 ### Priority: Keep MCP HTTP Protocol Fallback Deterministic
 
 **PRD:** `docs/prd/PRD-006-mcp-tool-integration.md`

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

@@ -11,6 +11,7 @@
   offline vs daemon-required command categorization)
 - Revised: 2026-05-24 (bootstrap-only `init`, domain-oriented `config`,
   init-owned identity re-entry, explicit reset flow)
+- Revised: 2026-08-11 (inline chat, structured output, input, approval, and copy contracts)
 - Depends on: `PRD-001`, `PRD-002`
 
 ## Goal
@@ -38,7 +39,7 @@ Netclaw ships as two binaries (see PRD-001 for full architecture):
 
 - **Simple arg routing** in `Program.cs` for command 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`)
+- **Termina** for interactive TUI commands, with an explicit mode for each application
 - All other commands use plain console output
 - Commands that need the daemon connect via `Microsoft.AspNetCore.SignalR.Client`
 - If the daemon isn't running and a command requires it, print an error with
@@ -133,7 +134,8 @@ Command ownership stays explicit:
   daemon over SignalR. Renders `SessionOutput` stream, sends `ChannelInput`.
   Session entity key: `tui/{uuid}`. If `netclaw.json` is absent, the command
   SHALL fail before contacting the daemon with
-  `daemon not configured - please run netclaw init`. See TUI-001 wireframes.
+  `daemon not configured - please run netclaw init`. Chat SHALL use the primary
+  terminal buffer after its full-screen session picker exits. See TUI-001.
 
 ### TUI-Interactive Commands (Termina, offline)
 
@@ -297,11 +299,13 @@ Results are persisted to the environment inventory file.
 
 ### CLI-010 TUI Commands
 
-`netclaw init`, `netclaw config`, and `netclaw chat` SHALL use Termina 0.5.1
-for interactive TUI rendering. Bare `netclaw provider` and `netclaw model`
-SHALL also use Termina. All other commands SHALL use plain console output. TUI
-commands SHALL launch Termina as a hosted service within the mode-selected host
-builder.
+`netclaw init`, `netclaw config`, and `netclaw chat` SHALL use the pinned Termina
+package. Bare `netclaw provider` and `netclaw model` SHALL also use Termina. All
+other commands SHALL use plain console output.
+
+Each Termina application SHALL select one presentation mode for its lifetime.
+Chat SHALL select `Inline` and `NativeTerminal`. Setup, config, provider, model,
+and session picker applications SHALL retain `FullScreen`.
 
 ### CLI-011 Chat Thin Client
 
@@ -312,12 +316,26 @@ interactive TUI for agent conversations. The TUI SHALL:
 - Create a session via the hub and receive a session ID
 - Send `ChannelInput` messages via SignalR
 - Subscribe to `SessionOutput` stream for rendering
-- Render session output as streaming text via StreamingTextNode
-- Display tool invocation status inline (completed with duration, in-progress
-  with spinner)
-- Show model name, token usage, and context percentage in status bar
+- Reduce each typed output into immutable settled content and a bounded live deck
+- Show thought, parallel tool, sub-agent, approval, file, error, usage,
+  compaction, and turn outcome forms with stable identities
+- Keep the settled transcript borderless in native terminal scrollback
+- Show model name, token usage, and context percentage in the Session Header
+- Use bare `Enter` to submit and `Shift+Enter` to add a newline
+- Keep the Composer active while the agent works
+- Show every later prompt in the ordered Queue Shelf
+- Send all later prompts through the session queue for one FIFO follow-up model call
+- Show assistant text as each streaming delta arrives
+- Restore a saved draft after prompt history reaches its newest entry
+- Clear prompt text only after two Escape keys inside a `TimeProvider` window
+- Give a pending approval input priority and preserve `Ctrl+O` detail expansion
+- Provide an Inspector and semantic copy for complete safe event detail
+- Report copy, terminal, and unsupported-event failures visibly
 - Print a clear error if the daemon is not running
 
+The client SHALL preserve `RecentMessages` compatibility. It SHALL prefer the
+additive structured transcript when the daemon supplies it.
+
 ### CLI-012 Daemon Management
 
 The CLI SHALL provide commands to manage the daemon lifecycle:

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

@@ -6,6 +6,7 @@
 - Owner: Netclaw engineering
 - Date: 2026-02-21
 - Revised: 2026-02-23 (daemon + thin client split, TUI as SignalR client)
+- Revised: 2026-08-11 (correlated activity and structured resume output)
 - Depends on: `PRD-001`, `PRD-002`, `PRD-008`
 
 ## Goal
@@ -53,9 +54,11 @@ child session actor.
 - Receives keyboard input via Termina TextInputNode
 - Sends `ChannelInput` to daemon over SignalR
 - Subscribes to `SessionOutput` stream over SignalR for rendering
-- Renders responses as streaming text via StreamingTextNode
-- Displays tool invocation status inline (name, duration, spinner)
-- Shows model name, token usage, and context percentage in status bar
+- Reduces typed output into settled transcript blocks and one bounded live deck
+- Correlates parallel tools by `CallId` and sub-agents by `RunId`
+- Shows thought, approval, file, error, usage, compaction, and turn outcome data
+- Restores settled event chronology from the structured resume contract
+- Keeps transient tool progress outside model context and persisted history
 
 **Slack Socket Mode Adapter** (Phase 1):
 - Runs in-process within the daemon
@@ -167,6 +170,14 @@ The TUI adapter is a pure thin client running in the `Netclaw.Cli` binary —
 all agent logic lives in the daemon. The TUI adapter SHALL display tool
 invocation status inline between user message and response.
 
+The output contract SHALL preserve all security-safe typed fields across
+SignalR. Tool activity SHALL retain `CallId` and turn identity. Sub-agent
+activity SHALL retain `RunId` and parent `CallId`.
+
+The daemon SHALL emit both the current `RecentMessages` field and an additive
+structured settled transcript during the compatibility period. The TUI SHALL
+not restore an old settled record as active work.
+
 ## Acceptance Criteria (MVP)
 
 1. TUI adapter receives input, routes through session actor, renders streaming

docs/spec/SPEC-002-session-lifecycle-and-protocol.md

@@ -1,6 +1,6 @@
 # SPEC-002: Session Lifecycle and Protocol
 
-Source PRDs: `PRD-001`
+Source PRDs: `PRD-001`, `PRD-009`
 Research: `docs/research/context-management-patterns.md`
 
 ## Purpose
@@ -106,6 +106,19 @@ always delivered regardless of filter.
 `UsageOutput` includes `ContextWindowTokens` and `UsagePercent` so subscribers
 can display context consumption without duplicating session config.
 
+`ToolActivityOutput` carries a stable `CallId`, turn identity, safe phase, and
+safe summary. A terminal tool result uses the same `CallId`.
+
+`SubAgentOutput` carries an additive `RunId` and parent `CallId`. Start,
+activity, and completion output for one run uses the same identities.
+
+Tool and sub-agent activity uses `OutputFilter.ToolCalls`. Thought activity
+uses `OutputFilter.Thinking`. Transient activity does not enter model context
+or the actor journal.
+
+Every supported output field has an explicit SignalR mapper disposition. A
+mapper test fails when a new field lacks mapped or security-omitted handling.
+
 ## Behavior States
 
 ```
@@ -184,7 +197,31 @@ are summarized as "Used {tool} for {purpose} → {outcome}".
 | `SessionTitleSet` | Title generated or updated |
 | `SessionCompacted` | History compacted with summary + retained messages |
 
+## Structured Resume Timeline
+
+The session keeps a bounded settled timeline for the recent turn window. The
+timeline uses framework-owned records with stable discriminators.
+
+The timeline can contain these settled entries:
+
+- user and assistant text
+- disclosed thought summary
+- tool call and result
+- sub-agent result
+- file metadata
+- error and usage detail
+- compaction detail
+- approval and turn outcome
+
+`SessionJoined` keeps `RecentMessages`. It adds a nullable `RecentTranscript`.
+The daemon emits both fields during the compatibility period.
+
+New readers prefer `RecentTranscript` when present. An absent timeline selects
+an explicit legacy conversion path. Unsupported legacy detail produces a
+diagnostic entry and never creates a false active state.
+
 ## Snapshot
 
-`SessionSnapshot` captures `History`, `TurnCount`, `Title` for fast recovery.
-Taken periodically per `SessionConfig.SnapshotInterval` and after compaction.
+`SessionSnapshot` captures `History`, `TurnCount`, `Title`, and the additive
+settled timeline for fast recovery. New Protobuf tags preserve current tags.
+Snapshots occur per `SessionConfig.SnapshotInterval` and after compaction.

docs/spec/SPEC-004-cli-contract.md

@@ -94,6 +94,31 @@ Behavior:
 - smoke test command runs optional live integration checks outside CI-required
   test suite
 
+### 7) Interactive Chat
+
+- `netclaw chat [--session <id>]`
+
+Behavior:
+
+- starts a Termina application with `Inline` presentation
+- selects `NativeTerminal` scroll input
+- leaves settled output in the primary terminal buffer
+- exits any full-screen session picker before chat starts
+- fails visibly when inline mode cannot start
+- never selects full-screen chat as a silent fallback
+
+Setup, config, provider, model, and session picker applications retain
+`FullScreen` presentation.
+
+The chat composer uses bare `Enter` for submit and `Shift+Enter` for a newline.
+A model call does not disable or hide the composer. Later prompts enter the
+session actor queue while the current turn runs. The actor retains all accepted
+prompts in FIFO order and includes them in one follow-up model call. The client
+does not start one turn for each queued prompt. The live region shows assistant
+text as each stream delta arrives.
+A pending approval owns input before the composer. `Ctrl+O` changes approval
+detail without a decision.
+
 ## Output and Exit Codes
 
 - default output: human readable text
@@ -110,6 +135,7 @@ Behavior:
 - read-only default for all inspection commands
 - mutating commands require explicit confirmation or `--yes`
 - no command may silently broaden exposure policy
+- no TUI command may silently change its terminal presentation mode
 
 ## Onboarding State Persistence
 

docs/spec/SPEC-010-testing-and-smoke-strategy.md

@@ -29,6 +29,36 @@ tests can validate real provider integrations.
 - explicit opt-in tests using real endpoints (for example, local Ollama)
 - intended for developer or pre-release validation
 
+## Chat TUI Proof Matrix
+
+The chat redesign requires deterministic headless proof and visual review.
+Unit tests alone cannot validate the visual grammar.
+
+Headless tests SHALL cover:
+
+- every supported `SessionOutput` disposition
+- parallel tools that finish out of order
+- same-name sub-agents with distinct `RunId` values
+- structured resume and old payload conversion
+- layouts at 40, 60, 80, and 120 columns
+- `Shift+Enter`, history draft restore, and double Escape with virtual time
+- approval priority, `Ctrl+O`, detail scroll, paste, and semantic copy
+
+Development review SHALL use three disposable video checkpoints outside the
+repository. The checkpoints SHALL cover the core chat, rich activity with
+approval, and the Inspector with responsive layout.
+
+Each checkpoint SHALL produce a temporary video and selected lossless frame
+images. A developer SHALL review these files and record material visual defects.
+The tapes SHALL not enter CI or the permanent smoke suite.
+
+The Termina package SHALL provide separate primary-buffer proof for Linux,
+macOS, Windows Terminal, and tmux. The proof SHALL cover resize, paste,
+selection, scrollback, and exit recovery.
+
+The full-screen smoke suite SHALL prove that init, config, provider, model, and
+picker applications retain their current terminal lifecycle.
+
 ## Critical Producer/Consumer Contract Inventory
 
 The contracts below are the minimum cross-boundary producer/consumer pairs that
@@ -45,6 +75,7 @@ proof is not complete yet, the gap is assigned to an explicit `NOW` task in
 | Scheduler -> delivery gateway | `SetReminderTool` and reminder persistence write `ReminderDefinition.Delivery` and later emit trusted delivery messages | Reminder execution actor and provider session binding actors that deliver without re-running inbound ACL | `Delivery.Kind` is `Channel` for channel delivery, `Delivery.Transport` is the lowercase provider key such as `slack`, and `Delivery.Address` is a canonical provider channel/user ID resolved before persistence. Runtime trusted delivery uses the stored target rather than a display name. | `src/Netclaw.Daemon.Tests/Reminder/ReminderTargetResolutionPathTests.cs` proves display target resolution to canonical channel/user IDs and unresolved target rejection. `src/Netclaw.Actors.Tests/Reminders/ReminderExecutionActorTests.cs` proves delivery success/failure reporting. Full gateway-chain and no-inbound-ACL re-entry coverage remains an explicit gap in Task 5.3. |
 | Tool schemas -> model/tool dispatcher | Built-in tool registrations and MCP tool adapters expose tool declarations and schemas | Provider serializers, `SessionToolExecutionPipeline`, `McpToolAdapter`, and MCP client manager | Model-facing tools serialize as OpenAI-compatible function tools with stable names, descriptions, JSON Schema parameters, and required fields. MCP tool names use `server/tool`. Dispatcher arguments preserve schema-declared string values and reconstruct structured JSON values only when the schema requires them. | `src/Netclaw.Daemon.Tests/Configuration/OpenAiCompatibleChatClientTests.cs` proves OpenAI function-tool serialization and tool-call history shape. `src/Netclaw.Daemon.Tests/Mcp/SmokeMcpServerArgumentCoercionTests.cs` proves schema-driven MCP argument reconstruction over the real stdio JSON-RPC path. Approval allow/deny/prompt and malformed metadata coverage remains an explicit gap in Task 4.2. |
 | Memory persistence -> prompt assembly | Memory curation, SQLite memory store, session events, and compaction events persist memory and conversation state | `SQLiteMemoryRecallCoordinator`, `SessionMessageAssembler`, and system prompt/session state assembly | Persisted memory uses framework-owned SQLite records and wire enum strings such as trust audience wire values. Session history uses `SerializableChatMessage` records, not provider SDK chat types. Recall appears as volatile context/nudges and does not mutate the stable system prompt prefix. | `src/Netclaw.Actors.Tests/Memory/SQLiteMemoryStoreTests.cs` proves memory persistence/search filtering and audience boundaries. `src/Netclaw.Actors.Tests/Memory/MemoryRedesignedEvalSuiteTests.cs` proves formation -> persistence -> recall. `src/Netclaw.Actors.Tests/Sessions/SessionMessageAssemblerTests.cs`, `SessionStateTests.cs`, and `src/Netclaw.Actors.Tests/Protocol/SerializationRoundTripTests.cs` prove prompt assembly placement and serialization-safe session records. Restart/recovery and corrupt/missing state coverage remains an explicit gap in Task 5.2. |
+| Session output -> SignalR -> chat reducer | Session actor output relay and `SessionOutputMapper` | `DaemonClient`, chat presentation reducer, and inline output owner | Typed output keeps all security-safe fields. Tools use `CallId`. Sub-agents use `RunId` and parent `CallId`. Resume uses settled framework-owned transcript entries. | The `redesign-netclaw-chat-tui` change requires output parity, legacy payload, reducer, and native chat tests before completion. |
 
 ## CI Rules
 

docs/spec/SPEC-011-daemon-architecture.md

@@ -1,6 +1,6 @@
 # SPEC-011: Daemon Architecture and Process Model
 
-Source PRDs: `PRD-001`, `PRD-002`, `PRD-004`
+Source PRDs: `PRD-001`, `PRD-002`, `PRD-004`, `PRD-009`
 
 ## Purpose
 
@@ -135,6 +135,10 @@ ReceiveOutput(output: SessionOutputDto) → void
 
 `SessionOutputDto` is a wire-safe mapping of the `SessionOutput` discriminated
 union. The mapper handles union → flat DTO conversion for SignalR serialization.
+It preserves every security-safe output field and all correlation identities.
+
+The DTO keeps `RecentMessages`. It adds nullable structured transcript and
+activity fields. Old clients ignore the additive properties.
 
 ### Connection Lifecycle
 
@@ -300,13 +304,22 @@ SignalR Client Adapter
 Daemon SignalR Hub (/hub/session)
 ```
 
-The `ChatViewModel` interface remains the same as the current in-process
-implementation — it exposes `IObservable<SessionOutput>` and accepts
-`SubmitAsync(text)`. The only change is the backend: SignalR client instead of
-direct `SessionPipeline`.
+`ChatViewModel` exposes the SignalR output stream and accepts prompt submission.
+It also owns immutable presentation state for stable event identities.
+
+`ChatPage` uses these semantic regions: Session Header, Transcript, Activity
+Rail, Decision Gate, Composer, and Status Line. It does not flatten typed output
+into one mutable text value.
+
+The chat process uses Termina inline presentation and native terminal scroll.
+The session picker remains a separate full-screen Termina application. The
+picker exits before the inline application starts.
+
+All chat output uses the inline output owner. Diagnostic logs use the configured
+file or structured sink while the live region is active.
 
-`ChatPage` does not change at all. Same rendering, same paste debounce, same
-status bar, same tool call spinners.
+The daemon keeps transient tool and sub-agent activity outside persistence. It
+persists only framework-owned settled transcript records with additive tags.
 
 ## Tool Execution Model
 

docs/ui/README.md

@@ -8,6 +8,8 @@ This directory contains management UI planning artifacts for Netclaw.
   component behavior
 - `TUI-001-command-wireframes.md` - Termina TUI wireframes for `netclaw init`,
   `netclaw chat`, and plain CLI commands
+- `chat-reply-stack-v2/index.html` - interactive Netclaw chat hierarchy,
+  state, motion, and responsive mock-ups
 - `ops-console-v1.html` - static high-fidelity mockup for visual direction
 
 ## Design Intent