Assistant Polish Suite

The domain and TUI state still distinguish assistant startup, active console state, skipped state, and runless presentation. Build runs can report a live assistant without inheriting the runless fixture's reduced story. Disabled assistant affordances can now explain startup or skip reasons instead of claiming that a normal build has no assistant.

The render layer still computes a composer layout from the draft and the available pane size. Long input wraps inside the assistant pane, grows until its cap, and then scrolls internally while preserving cursor visibility. The mouse hit range comes from the same layout, so focus behavior follows the rows the user can actually type into.

The current renderer still computes chat pane geometry from the body, chat state, and composer input. It returns the upper area, assistant bounds, transcript and composer rectangles, restore target, presentation state, and divider regions in one place. That keeps collapsed, restored, and maximized drawing on the same model rather than scattered height decisions.

The TUI still has divider regions for menu title space, resize controls, drag area, and status text. The pane remembers a user's restored size and treats minimize or maximize as temporary presentations rather than destructive changes. The menu-tour script and rendering code still reflect the pane controls that this sprint introduced, although later documentation reframed the assistant surface around the divider menu.

The TUI still renders a separate assistant menu system on the pane divider when that surface is present. Assistant actions dispatch through the shared action path instead of private row handling, and the normal build menu no longer needs a stale Chat title to expose pane controls. The live tree's surface inventory now documents the assistant divider separately from the main menu tour.

The assistant harness resolution still lives beside the broader harness configuration and reports its selected identity to the console. Explicit assistant choices and model pins keep precedence, while the default can follow the resolved build harness. The docs still describe that distinction and note that a failed assistant startup does not stop the deterministic build.

Assistant turns still emit identity events before reply events, and completed turns persist the assistant identity in the transcript. Pending and activated selection events remain part of the domain surface. That means a model change is visible as a next-turn state change rather than an ambiguous rewrite of the response already running.

The assistant documentation and TUI state still describe a Model menu that lists local harness choices and shows current, pending, and failed selection states. The menu queues a selection request rather than retargeting the running turn. The docs also preserve the caveat that file-based MCP config delivery can overwrite temporary harness config during an assistant turn.

The surviving TUI code still targets assistant logs from the active run instead of the runless diagnostic directory. Assistant menu actions remain tied to shared dispatch ids, and stale unavailable-chat behavior is no longer treated as the normal build-run truth. The narrated inventory gap did not fully settle here; the later surface inventory contract is the durable statement of that boundary.

Blocker investigation prompts still enter the assistant transcript rather than the ordinary modal path. The state keeps the prompt pending while showing failure context in the pane, and routing alone does not submit an assistant message. Non-blocker prompt kinds continue to use modal UI, preserving the direct operator-question path.

The renderer still draws blocked-run transcript entries with inline action buttons and a visible blocked state outside the scroll position. Those actions target the original prompt identity and stop being live once the prompt resolves. The quick stop path remains discoverable without turning blocker choices into persistent assistant menu rows.

Assistant turn failures still become transcript events instead of disappearing from the pane. The live MCP host can carry answer_prompt to the attached engine, and successful answers resolve the same pending prompt that inline choices use. Failed attempts remain visible and do not falsely clear the blocked state.

Runtime and architecture docs still say halted-run blockers are handed to the assistant pane, remain pending, and can be answered or stopped through the original prompt path. They also state that merely staging the blocker does not spend assistant tokens. Menu and screencast handover documents were later folded or superseded by the final surface inventory work, so this sprint's docs survive partly as current prose and partly as inputs to that later contract.

The handover contract still describes the normal build-run bar, the assistant divider bar, the chat-only surface, blocker affordances, and excluded stale rows. It explicitly treats the assistant divider as a separate tested surface rather than folding it into the normal menu tour. The generated menu-tour assets remain in the tree, while the contract explains how to read their scope.

The MCP tool registry still documents list_runs as full by default with an explicit active scope for non-finished runs. The server accepts the scope argument, rejects invalid values, preserves legacy wire aliases, and keeps run_bash in the advertised action surface. The handover contract records which old mirror claims were corrected and where regression coverage lives.

The assistant and MCP reference now reads as a current guide to the build-run assistant. It documents divider controls, Model selection, blocker handling, logs, transcripts, harness precedence, diagnostic assistant runs, runless bare invocation, external MCP attachment, and harness-aware MCP delivery. Adjacent CLI, config, harness, runtime, and architecture docs were adjusted to point at that same behavior.

The README now presents the assistant as available inside interactive build runs, with keyboard focus, divider resizing, Model selection, and blocker handoff. The web and project docs point readers to the assistant and MCP reference rather than repeating stale messages-bar phrasing. The menu-tour script and generated artifacts still exist, but current documentation frames them through the final surface contract.