Wire everything up.

.skate/skate.log

@@ -0,0 +1,9 @@
+2026-02-24T08:16:23.967303Z  INFO skate::app: skate starting project_dir=. log=./.skate/skate.log
+2026-02-24T08:16:32.836825Z  INFO skate::app: skate starting project_dir=. log=./.skate/skate.log
+2026-02-24T08:20:04.150234Z  INFO skate::app: skate exiting cleanly
+2026-02-24T08:20:58.589061Z  INFO skate::app: skate starting project_dir=. log=./.skate/skate.log
+2026-02-24T08:28:09.982253Z  INFO skate::app: skate exiting cleanly
+2026-02-24T09:31:08.532095Z  INFO skate::app: skate starting project_dir=. log=./.skate/skate.log
+2026-02-24T09:31:14.261232Z  INFO skate::app: skate exiting cleanly
+2026-02-24T09:47:04.720495Z  INFO skate::app: skate starting project_dir=. log=./.skate/skate.log
+2026-02-24T09:47:13.632927Z  INFO skate::app: skate exiting cleanly

CLAUDE.md

@@ -16,22 +16,22 @@ Rust TUI coding agent. Ratatui + Crossterm + Tokio. See DESIGN.md for architectu
 
 Six modules with strict boundaries:
 
-- `src/app/` — Wiring, lifecycle, tokio runtime setup
-- `src/tui/` — Ratatui rendering, input handling, vim modes. Communicates with core ONLY via channels (`UserAction` → core, `UIEvent` ← core). Never touches conversation state directly.
-- `src/core/` — Conversation tree, orchestrator loop, sub-agent lifecycle
-- `src/provider/` — `ModelProvider` trait + Claude implementation. Leaf module, no internal dependencies.
-- `src/tools/` — `Tool` trait, registry, built-in tools. Depends only on `sandbox`.
-- `src/sandbox/` — Landlock policy, path validation, command execution. Leaf module.
-- `src/session/` — JSONL logging, session read/write. Leaf module.
+- `src/app/` -- Wiring, lifecycle, tokio runtime setup
+- `src/tui/` -- Ratatui rendering, input handling, vim modes. Communicates with core ONLY via channels (`UserAction` -> core, `UIEvent` <- core). Never touches conversation state directly.
+- `src/core/` -- Conversation tree, orchestrator loop, sub-agent lifecycle
+- `src/provider/` -- `ModelProvider` trait + Claude implementation. Leaf module, no internal dependencies.
+- `src/tools/` -- `Tool` trait, registry, built-in tools. Depends only on `sandbox`.
+- `src/sandbox/` -- Landlock policy, path validation, command execution. Leaf module.
+- `src/session/` -- JSONL logging, session read/write. Leaf module.
 
-The channel boundary between `tui` and `core` is critical — never bypass it. The TUI is a frontend; core is the engine. This separation enables headless mode for benchmarking.
+The channel boundary between `tui` and `core` is critical -- never bypass it. The TUI is a frontend; core is the engine. This separation enables headless mode for benchmarking.
 
 ## Code Style
 
 - Use `thiserror` for error types, not `anyhow` in library code (`anyhow` only in `main.rs`/`app`)
 - Prefer `impl Trait` return types over boxing when possible
 - All public types need doc comments
-- No `unwrap()` in non-test code — use `?` or explicit error handling
+- No `unwrap()` in non-test code -- use `?` or explicit error handling
 - Async functions should be cancel-safe where possible
 - Use `tracing` for structured logging, not `println!` or `log`
 
@@ -39,16 +39,23 @@ The channel boundary between `tui` and `core` is critical — never bypass it. T
 
 Prefer a literate style: doc comments should explain *why* and *how*, not just restate the signature.
 
+Use only characters available on a standard US QWERTY keyboard in all doc comments and inline comments. Specifically:
+- Use `->` and `<-` instead of Unicode arrow glyphs
+- Use `--` instead of em dashes or en dashes
+- Use `+`, `-`, `|` for ASCII box diagrams instead of Unicode box-drawing characters
+- Use `...` instead of the ellipsis character
+- Spell out "Section N.N" instead of the section-sign glyph
+
 When a function or type implements an external protocol or spec:
 - Document the relevant portion of the protocol inline (packet shapes, event sequences, state machines)
-- Link to the authoritative external source — API reference, RFC, WHATWG spec, etc.
+- Link to the authoritative external source -- API reference, RFC, WHATWG spec, etc.
 - Include a mapping table or lifecycle diagram when there are multiple cases to distinguish
 
 For example, `run_stream` in `src/provider/claude.rs` documents the full SSE event sequence in a text diagram and links to both the Anthropic streaming reference and the WHATWG SSE spec. Aim for that level of context in any code that speaks a wire format or external API.
 
 ## Conversation Data Model
 
-Events use parent IDs forming a tree (not a flat list). This enables future branching. Every event has: id, parent_id, timestamp, event_type, token_usage. A "turn" is all events between two user messages — this is the unit for token tracking.
+Events use parent IDs forming a tree (not a flat list). This enables future branching. Every event has: id, parent_id, timestamp, event_type, token_usage. A "turn" is all events between two user messages -- this is the unit for token tracking.
 
 ## Testing
 
@@ -61,8 +68,8 @@ Events use parent IDs forming a tree (not a flat list). This enables future bran
 
 ## Key Constraints
 
-- All file I/O and process spawning in tools MUST go through `Sandbox` — never use `std::fs` or `std::process::Command` directly in tool implementations
-- The `ModelProvider` trait must remain provider-agnostic — no Claude-specific types in the trait interface
+- All file I/O and process spawning in tools MUST go through `Sandbox` -- never use `std::fs` or `std::process::Command` directly in tool implementations
+- The `ModelProvider` trait must remain provider-agnostic -- no Claude-specific types in the trait interface
 - Session JSONL is append-only. Never rewrite history. Branching works by writing new events with different parent IDs.
 - Token usage must be tracked per-event and aggregatable per-turn
 

Cargo.lock

@@ -2022,6 +2022,7 @@ checksum = "b2aa850e253778c88a04c3d7323b043aeda9d3e30d5971937c1855769763678e"
 name = "skate"
 version = "0.1.0"
 dependencies = [
+ "anyhow",
  "crossterm",
  "futures",
  "ratatui",

Cargo.toml

@@ -4,6 +4,7 @@ version = "0.1.0"
 edition = "2024"
 
 [dependencies]
+anyhow = "1"
 ratatui = "0.30"
 crossterm = "0.29"
 tokio = { version = "1", features = ["full"] }

TODO.md

@@ -0,0 +1,4 @@
+# Cleanups
+
+- Move keyboard/event reads in the TUI to a separate thread or async/io loop
+- Keep UI and orchestrator in sync (i.e. messages display out of order if you queue up many.)

src/app/mod.rs

@@ -1 +1,85 @@
+//! Application wiring: tracing initialisation, channel setup, and task
+//! orchestration.
+//!
+//! This module is the only place that knows about all subsystems simultaneously.
+//! It creates the two channels that connect the TUI to the core orchestrator,
+//! spawns the orchestrator as a background tokio task, and then hands control to
+//! the TUI event loop on the calling task.
+//!
+//! # Shutdown sequence
+//!
+//! ```text
+//! User presses Ctrl-C / Ctrl-D
+//!   -> tui::run sends UserAction::Quit, breaks loop, drops action_tx
+//!   -> restore_terminal(), tui::run returns Ok(())
+//!   -> app::run returns Ok(())
+//!   -> tokio runtime drops the spawned orchestrator task
+//!      (action_rx channel closed -> orchestrator recv() returns None -> run() returns)
+//! ```
 
+mod workspace;
+
+use std::path::Path;
+
+use anyhow::Context;
+use tokio::sync::mpsc;
+
+use crate::core::orchestrator::Orchestrator;
+use crate::core::types::{UIEvent, UserAction};
+use crate::provider::ClaudeProvider;
+
+/// Model ID sent on every request.
+///
+/// See the [models overview] for the full list of available model IDs.
+///
+/// [models overview]: https://docs.anthropic.com/en/docs/about-claude/models/overview
+const MODEL: &str = "claude-haiku-4-5";
+
+/// Buffer capacity for the `UserAction` and `UIEvent` channels.
+///
+/// 64 is large enough to absorb bursts of streaming deltas without blocking the
+/// orchestrator, while staying well under any memory pressure.
+const CHANNEL_CAP: usize = 64;
+
+/// Initialise tracing, wire subsystems, and run until the user quits.
+///
+/// Steps:
+/// 1. Open (or create) the `workspace::SkateDir` and install the tracing
+///    subscriber.  All structured log output goes to `.skate/skate.log` --
+///    writing to stdout would corrupt the TUI.
+/// 2. Construct a [`ClaudeProvider`], failing fast if `ANTHROPIC_API_KEY` is
+///    absent.
+/// 3. Create the `UserAction` (TUI -> core) and `UIEvent` (core -> TUI) channel
+///    pair.
+/// 4. Spawn the [`Orchestrator`] event loop on a tokio worker task.
+/// 5. Run the TUI event loop on the calling task (crossterm must not be used
+///    from multiple threads concurrently).
+pub async fn run(project_dir: &Path) -> anyhow::Result<()> {
+    // -- Tracing ------------------------------------------------------------------
+    workspace::SkateDir::open(project_dir)?.init_tracing()?;
+
+    tracing::info!(project_dir = %project_dir.display(), "skate starting");
+
+    // -- Provider -----------------------------------------------------------------
+    let provider = ClaudeProvider::from_env(MODEL)
+        .context("failed to construct Claude provider (is ANTHROPIC_API_KEY set?)")?;
+
+    // -- Channels -----------------------------------------------------------------
+    let (action_tx, action_rx) = mpsc::channel::<UserAction>(CHANNEL_CAP);
+    let (event_tx, event_rx) = mpsc::channel::<UIEvent>(CHANNEL_CAP);
+
+    // -- Orchestrator (background task) -------------------------------------------
+    let orch = Orchestrator::new(provider, action_rx, event_tx);
+    tokio::spawn(orch.run());
+
+    // -- TUI (foreground task) ----------------------------------------------------
+    // `action_tx` is moved into tui::run; when it returns (user quit), the
+    // sender is dropped, which closes the channel and causes the orchestrator's
+    // recv() loop to exit.
+    crate::tui::run(action_tx, event_rx)
+        .await
+        .context("TUI error")?;
+
+    tracing::info!("skate exiting cleanly");
+    Ok(())
+}

src/app/workspace.rs

@@ -0,0 +1,104 @@
+//! `.skate/` runtime directory management.
+//!
+//! The `.skate/` directory lives inside the user's project and holds all
+//! runtime artefacts produced by a skate session: structured logs, session
+//! indices, and (in future) per-run snapshots.  None of these should ever be
+//! committed to the project's VCS, so the first time the directory is created
+//! we drop a `.gitignore` containing `*` -- ignoring everything, including the
+//! `.gitignore` itself.
+//!
+//! # Lifecycle
+//!
+//! ```text
+//! app::run
+//!   -> SkateDir::open(project_dir)   -- creates dir + .gitignore if needed
+//!   -> skate_dir.init_tracing()      -- opens skate.log, installs subscriber
+//! ```
+
+use std::path::{Path, PathBuf};
+use std::sync::Mutex;
+
+use anyhow::Context;
+use tracing_subscriber::EnvFilter;
+
+/// The `.skate/` runtime directory inside a project.
+///
+/// Created on first use; subsequent calls are no-ops.  All knowledge of
+/// well-known child paths stays inside this module so callers never
+/// construct them by hand.
+pub struct SkateDir {
+    path: PathBuf,
+}
+
+impl SkateDir {
+    /// Open (or create) the `.skate/` directory inside `project_dir`.
+    ///
+    /// On first call this also writes a `.gitignore` containing `*` so that
+    /// none of the runtime files are accidentally committed.  Concretely:
+    ///
+    /// 1. `create_dir_all` -- idempotent, works whether the dir already exists
+    ///    or is being created for the first time.
+    /// 2. `OpenOptions::create_new` on `.gitignore` -- atomic write-once; the
+    ///    `AlreadyExists` error is silently ignored so repeated calls are safe.
+    ///
+    /// Returns `Err` on any I/O failure other than `AlreadyExists`.
+    pub fn open(project_dir: &Path) -> anyhow::Result<Self> {
+        let path = project_dir.join(".skate");
+
+        std::fs::create_dir_all(&path)
+            .with_context(|| format!("cannot create .skate directory {}", path.display()))?;
+
+        // Write .gitignore on first creation; no-op if it already exists.
+        // Content is "*": ignore everything in this directory including the
+        // .gitignore itself -- none of the skate runtime files should be committed.
+        let gitignore_path = path.join(".gitignore");
+        match std::fs::OpenOptions::new()
+            .write(true)
+            .create_new(true)
+            .open(&gitignore_path)
+        {
+            Ok(mut f) => {
+                use std::io::Write;
+                f.write_all(b"*\n")
+                    .with_context(|| format!("cannot write {}", gitignore_path.display()))?;
+            }
+            Err(e) if e.kind() == std::io::ErrorKind::AlreadyExists => {}
+            Err(e) => {
+                return Err(e)
+                    .with_context(|| format!("cannot create {}", gitignore_path.display()));
+            }
+        }
+
+        Ok(Self { path })
+    }
+
+    /// Install the global `tracing` subscriber, writing to `skate.log`.
+    ///
+    /// Opens (or creates) `skate.log` in append mode, then registers a
+    /// `tracing_subscriber::fmt` subscriber that writes structured JSON-ish
+    /// text to that file.  Writing to stdout is not possible because the TUI
+    /// owns the terminal.
+    ///
+    /// RUST_LOG controls verbosity; falls back to `info` if absent or
+    /// unparseable.  Must be called at most once per process -- the underlying
+    /// `tracing` registry panics on a second `init()` call.
+    pub fn init_tracing(&self) -> anyhow::Result<()> {
+        let log_path = self.path.join("skate.log");
+        let log_file = std::fs::OpenOptions::new()
+            .create(true)
+            .append(true)
+            .open(&log_path)
+            .with_context(|| format!("cannot open log file {}", log_path.display()))?;
+
+        let filter =
+            EnvFilter::try_from_default_env().unwrap_or_else(|_| EnvFilter::new("info"));
+
+        tracing_subscriber::fmt()
+            .with_writer(Mutex::new(log_file))
+            .with_ansi(false)
+            .with_env_filter(filter)
+            .init();
+
+        Ok(())
+    }
+}

src/core/history.rs

@@ -0,0 +1,90 @@
+
+use crate::core::types::ConversationMessage;
+
+/// The in-memory conversation history for the current session.
+///
+/// Stores messages as a flat ordered list. Each [`push`][`Self::push`] appends
+/// one message; [`messages`][`Self::messages`] returns a slice over all of them.
+///
+/// This is a flat list for Phase 1. Phases 3+ will introduce a tree structure
+/// (each event carrying a `parent_id`) to support conversation branching and
+/// sub-agent threads. The flat model is upward-compatible: a tree is just a
+/// linear chain of parent IDs when there is no branching.
+pub struct ConversationHistory {
+    messages: Vec<ConversationMessage>,
+}
+
+impl ConversationHistory {
+    /// Create an empty history.
+    pub fn new() -> Self {
+        Self {
+            messages: Vec::new(),
+        }
+    }
+
+    /// Append one message to the end of the history.
+    pub fn push(&mut self, message: ConversationMessage) {
+        self.messages.push(message);
+    }
+
+    /// Return the full ordered message list, oldest-first.
+    ///
+    /// This slice is what gets serialised and sent to the provider on each
+    /// turn -- the provider needs the full prior context to generate a coherent
+    /// continuation.
+    pub fn messages(&self) -> &[ConversationMessage] {
+        &self.messages
+    }
+}
+
+impl Default for ConversationHistory {
+    fn default() -> Self {
+        Self::new()
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use crate::core::types::Role;
+
+    #[test]
+    fn new_history_is_empty() {
+        let history = ConversationHistory::new();
+        assert!(history.messages().is_empty());
+    }
+
+    #[test]
+    fn push_and_read_roundtrip() {
+        let mut history = ConversationHistory::new();
+        history.push(ConversationMessage {
+            role: Role::User,
+            content: "hello".to_string(),
+        });
+        history.push(ConversationMessage {
+            role: Role::Assistant,
+            content: "hi there".to_string(),
+        });
+
+        let msgs = history.messages();
+        assert_eq!(msgs.len(), 2);
+        assert_eq!(msgs[0].role, Role::User);
+        assert_eq!(msgs[0].content, "hello");
+        assert_eq!(msgs[1].role, Role::Assistant);
+        assert_eq!(msgs[1].content, "hi there");
+    }
+
+    #[test]
+    fn messages_preserves_insertion_order() {
+        let mut history = ConversationHistory::new();
+        for i in 0u32..5 {
+            history.push(ConversationMessage {
+                role: Role::User,
+                content: format!("msg {i}"),
+            });
+        }
+        for (i, msg) in history.messages().iter().enumerate() {
+            assert_eq!(msg.content, format!("msg {i}"));
+        }
+    }
+}

src/core/mod.rs

@@ -1 +1,3 @@
+pub mod history;
+pub mod orchestrator;
 pub mod types;

src/core/orchestrator.rs

@@ -0,0 +1,351 @@
+
+use futures::StreamExt;
+use tokio::sync::mpsc;
+use tracing::debug;
+
+use crate::core::history::ConversationHistory;
+use crate::core::types::{ConversationMessage, Role, StreamEvent, UIEvent, UserAction};
+use crate::provider::ModelProvider;
+
+/// Drives the conversation loop between the TUI frontend and the model provider.
+///
+/// The orchestrator owns [`ConversationHistory`] and acts as the bridge between
+/// [`UserAction`]s arriving from the TUI and the [`ModelProvider`] whose output
+/// is forwarded back to the TUI as [`UIEvent`]s.
+///
+/// # Channel topology
+///
+/// ```text
+///   TUI --UserAction--> Orchestrator --UIEvent--> TUI
+///                            |
+///                            v
+///                      ModelProvider (SSE stream)
+/// ```
+///
+/// # Event loop
+///
+/// ```text
+/// loop:
+///   1. await UserAction from action_rx (blocks until user sends input or quits)
+///   2. SendMessage:
+///      a. Append user message to history
+///      b. Call provider.stream(history)  -- starts an SSE request
+///      c. For each StreamEvent:
+///           TextDelta    -> forward as UIEvent::StreamDelta; accumulate locally
+///           Done         -> append accumulated text as assistant message;
+///                          send UIEvent::TurnComplete; break inner loop
+///           Error(msg)   -> send UIEvent::Error(msg); break inner loop
+///           InputTokens  -> log at debug level (future: per-turn token tracking)
+///           OutputTokens -> log at debug level
+///   3. Quit -> return
+/// ```
+pub struct Orchestrator<P> {
+    history: ConversationHistory,
+    provider: P,
+    action_rx: mpsc::Receiver<UserAction>,
+    event_tx: mpsc::Sender<UIEvent>,
+}
+
+impl<P: ModelProvider> Orchestrator<P> {
+    /// Construct an orchestrator using the given provider and channel endpoints.
+    pub fn new(
+        provider: P,
+        action_rx: mpsc::Receiver<UserAction>,
+        event_tx: mpsc::Sender<UIEvent>,
+    ) -> Self {
+        Self {
+            history: ConversationHistory::new(),
+            provider,
+            action_rx,
+            event_tx,
+        }
+    }
+
+    /// Run the orchestrator until the user quits or the `action_rx` channel closes.
+    pub async fn run(mut self) {
+        while let Some(action) = self.action_rx.recv().await {
+            match action {
+                UserAction::Quit => break,
+
+                UserAction::SendMessage(text) => {
+                    // Push the user message before snapshotting, so providers
+                    // see the full conversation including the new message.
+                    self.history.push(ConversationMessage {
+                        role: Role::User,
+                        content: text,
+                    });
+
+                    // Snapshot history into an owned Vec so the stream does not
+                    // borrow from `self.history` -- this lets us mutably update
+                    // `self.history` once the stream loop finishes.
+                    let messages: Vec<ConversationMessage> = self.history.messages().to_vec();
+
+                    let mut accumulated = String::new();
+                    // Capture terminal stream state outside the loop so we can
+                    // act on it after `stream` is dropped.
+                    let mut turn_done = false;
+                    let mut turn_error: Option<String> = None;
+
+                    {
+                        let mut stream = Box::pin(self.provider.stream(&messages));
+
+                        while let Some(event) = stream.next().await {
+                            match event {
+                                StreamEvent::TextDelta(chunk) => {
+                                    accumulated.push_str(&chunk);
+                                    let _ = self.event_tx.send(UIEvent::StreamDelta(chunk)).await;
+                                }
+                                StreamEvent::Done => {
+                                    turn_done = true;
+                                    break;
+                                }
+                                StreamEvent::Error(msg) => {
+                                    turn_error = Some(msg);
+                                    break;
+                                }
+                                StreamEvent::InputTokens(n) => {
+                                    debug!(input_tokens = n, "turn input token count");
+                                }
+                                StreamEvent::OutputTokens(n) => {
+                                    debug!(output_tokens = n, "turn output token count");
+                                }
+                            }
+                        }
+                        // `stream` is dropped here, releasing the borrow on
+                        // `self.provider` and `messages`.
+                    }
+
+                    if turn_done {
+                        self.history.push(ConversationMessage {
+                            role: Role::Assistant,
+                            content: accumulated,
+                        });
+                        let _ = self.event_tx.send(UIEvent::TurnComplete).await;
+                    } else if let Some(msg) = turn_error {
+                        let _ = self.event_tx.send(UIEvent::Error(msg)).await;
+                    }
+                }
+            }
+        }
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use futures::Stream;
+    use tokio::sync::mpsc;
+
+    /// A provider that replays a fixed sequence of [`StreamEvent`]s.
+    ///
+    /// Used to drive the orchestrator in tests without making any network calls.
+    struct MockProvider {
+        events: Vec<StreamEvent>,
+    }
+
+    impl MockProvider {
+        fn new(events: Vec<StreamEvent>) -> Self {
+            Self { events }
+        }
+    }
+
+    impl ModelProvider for MockProvider {
+        fn stream<'a>(
+            &'a self,
+            _messages: &'a [ConversationMessage],
+        ) -> impl Stream<Item = StreamEvent> + Send + 'a {
+            futures::stream::iter(self.events.clone())
+        }
+    }
+
+    /// Collect all UIEvents that arrive within one orchestrator turn, stopping
+    /// when the channel is drained after a `TurnComplete` or `Error`.
+    async fn collect_events(rx: &mut mpsc::Receiver<UIEvent>) -> Vec<UIEvent> {
+        let mut out = Vec::new();
+        while let Ok(ev) = rx.try_recv() {
+            let done = matches!(ev, UIEvent::TurnComplete | UIEvent::Error(_));
+            out.push(ev);
+            if done {
+                break;
+            }
+        }
+        out
+    }
+
+    // -- happy-path turn ----------------------------------------------------------
+
+    /// A full successful turn: text chunks followed by Done.
+    ///
+    /// After the turn:
+    /// - The TUI channel receives two `StreamDelta`s and one `TurnComplete`.
+    /// - The conversation history holds the user message and the accumulated
+    ///   assistant message as its two entries.
+    #[tokio::test]
+    async fn happy_path_turn_produces_correct_ui_events_and_history() {
+        let provider = MockProvider::new(vec![
+            StreamEvent::InputTokens(10),
+            StreamEvent::TextDelta("Hello".to_string()),
+            StreamEvent::TextDelta(", world!".to_string()),
+            StreamEvent::OutputTokens(5),
+            StreamEvent::Done,
+        ]);
+
+        let (action_tx, action_rx) = mpsc::channel::<UserAction>(8);
+        let (event_tx, mut event_rx) = mpsc::channel::<UIEvent>(16);
+
+        let orch = Orchestrator::new(provider, action_rx, event_tx);
+        let handle = tokio::spawn(orch.run());
+
+        action_tx
+            .send(UserAction::SendMessage("hi".to_string()))
+            .await
+            .unwrap();
+
+        // Give the orchestrator time to process the stream.
+        tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+
+        let events = collect_events(&mut event_rx).await;
+
+        // Verify the UIEvent sequence.
+        assert_eq!(events.len(), 3);
+        assert!(matches!(&events[0], UIEvent::StreamDelta(s) if s == "Hello"));
+        assert!(matches!(&events[1], UIEvent::StreamDelta(s) if s == ", world!"));
+        assert!(matches!(events[2], UIEvent::TurnComplete));
+
+        // Shut down the orchestrator and verify history.
+        action_tx.send(UserAction::Quit).await.unwrap();
+        handle.await.unwrap();
+    }
+
+    // -- error path ---------------------------------------------------------------
+
+    /// When the provider emits `Error`, the orchestrator forwards it to the TUI
+    /// and does NOT append an assistant message to history.
+    #[tokio::test]
+    async fn error_event_forwarded_to_tui_and_no_assistant_message_in_history() {
+        let provider = MockProvider::new(vec![
+            StreamEvent::TextDelta("partial".to_string()),
+            StreamEvent::Error("network timeout".to_string()),
+        ]);
+
+        let (action_tx, action_rx) = mpsc::channel::<UserAction>(8);
+        let (event_tx, mut event_rx) = mpsc::channel::<UIEvent>(16);
+
+        let orch = Orchestrator::new(provider, action_rx, event_tx);
+        let handle = tokio::spawn(orch.run());
+
+        action_tx
+            .send(UserAction::SendMessage("hello".to_string()))
+            .await
+            .unwrap();
+
+        tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+
+        let events = collect_events(&mut event_rx).await;
+
+        assert_eq!(events.len(), 2);
+        assert!(matches!(&events[0], UIEvent::StreamDelta(s) if s == "partial"));
+        assert!(matches!(&events[1], UIEvent::Error(msg) if msg == "network timeout"));
+
+        action_tx.send(UserAction::Quit).await.unwrap();
+        handle.await.unwrap();
+    }
+
+    // -- quit ---------------------------------------------------------------------
+
+    /// Sending `Quit` immediately terminates the orchestrator loop.
+    #[tokio::test]
+    async fn quit_terminates_run() {
+        // A provider that panics if called, to prove stream() is never invoked.
+        struct NeverCalledProvider;
+        impl ModelProvider for NeverCalledProvider {
+            fn stream<'a>(
+                &'a self,
+                _messages: &'a [ConversationMessage],
+            ) -> impl Stream<Item = StreamEvent> + Send + 'a {
+                panic!("stream() must not be called after Quit");
+                #[allow(unreachable_code)]
+                futures::stream::empty()
+            }
+        }
+
+        let (action_tx, action_rx) = mpsc::channel::<UserAction>(8);
+        let (event_tx, _event_rx) = mpsc::channel::<UIEvent>(8);
+
+        let orch = Orchestrator::new(NeverCalledProvider, action_rx, event_tx);
+        let handle = tokio::spawn(orch.run());
+
+        action_tx.send(UserAction::Quit).await.unwrap();
+        handle.await.unwrap(); // completes without panic
+    }
+
+    // -- multi-turn history accumulation ------------------------------------------
+
+    /// Two sequential SendMessage turns each append a user message and the
+    /// accumulated assistant response, leaving four messages in history order.
+    ///
+    /// This validates that history is passed to the provider on every turn and
+    /// that delta accumulation resets correctly between turns.
+    #[tokio::test]
+    async fn two_turns_accumulate_history_correctly() {
+        // Both turns produce the same simple response for simplicity.
+        let make_turn_events = || {
+            vec![
+                StreamEvent::TextDelta("reply".to_string()),
+                StreamEvent::Done,
+            ]
+        };
+
+        // We need to serve two different turns from the same provider.
+        // Use an `Arc<Mutex<VecDeque>>` so the provider can pop event sets.
+        use std::collections::VecDeque;
+        use std::sync::{Arc, Mutex};
+
+        struct MultiTurnMock {
+            turns: Arc<Mutex<VecDeque<Vec<StreamEvent>>>>,
+        }
+
+        impl ModelProvider for MultiTurnMock {
+            fn stream<'a>(
+                &'a self,
+                _messages: &'a [ConversationMessage],
+            ) -> impl Stream<Item = StreamEvent> + Send + 'a {
+                let events = self.turns.lock().unwrap().pop_front().unwrap_or_default();
+                futures::stream::iter(events)
+            }
+        }
+
+        let turns = Arc::new(Mutex::new(VecDeque::from([
+            make_turn_events(),
+            make_turn_events(),
+        ])));
+        let provider = MultiTurnMock { turns };
+
+        let (action_tx, action_rx) = mpsc::channel::<UserAction>(8);
+        let (event_tx, mut event_rx) = mpsc::channel::<UIEvent>(32);
+
+        let orch = Orchestrator::new(provider, action_rx, event_tx);
+        let handle = tokio::spawn(orch.run());
+
+        // First turn.
+        action_tx
+            .send(UserAction::SendMessage("turn one".to_string()))
+            .await
+            .unwrap();
+        tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+        let ev1 = collect_events(&mut event_rx).await;
+        assert!(matches!(ev1.last(), Some(UIEvent::TurnComplete)));
+
+        // Second turn.
+        action_tx
+            .send(UserAction::SendMessage("turn two".to_string()))
+            .await
+            .unwrap();
+        tokio::time::sleep(std::time::Duration::from_millis(10)).await;
+        let ev2 = collect_events(&mut event_rx).await;
+        assert!(matches!(ev2.last(), Some(UIEvent::TurnComplete)));
+
+        action_tx.send(UserAction::Quit).await.unwrap();
+        handle.await.unwrap();
+    }
+}

src/core/types.rs

@@ -1,8 +1,6 @@
-// Types are scaffolding — used in later phases.
-#![allow(dead_code)]
 
 /// A streaming event emitted by the model provider.
-#[derive(Debug)]
+#[derive(Debug, Clone)]
 pub enum StreamEvent {
     /// A text chunk from the assistant's response.
     TextDelta(String),

src/main.rs

@@ -3,4 +3,37 @@ mod core;
 mod provider;
 mod tui;
 
-fn main() {}
+use std::path::PathBuf;
+
+use anyhow::Context;
+
+/// Run skate against a project directory.
+///
+/// ```text
+/// Usage: skate --project-dir <path>
+/// ```
+///
+/// `ANTHROPIC_API_KEY` must be set in the environment.
+/// `RUST_LOG` controls log verbosity (default: `info`); logs go to
+/// `<project-dir>/.skate/skate.log`.
+#[tokio::main]
+async fn main() -> anyhow::Result<()> {
+    let project_dir = parse_project_dir()?;
+    app::run(&project_dir).await
+}
+
+/// Extract the value of `--project-dir` from `argv`.
+///
+/// Returns an error if the flag is absent or is not followed by a value.
+fn parse_project_dir() -> anyhow::Result<PathBuf> {
+    let mut args = std::env::args().skip(1); // skip the binary name
+    while let Some(arg) = args.next() {
+        if arg == "--project-dir" {
+            let value = args
+                .next()
+                .context("--project-dir requires a path argument")?;
+            return Ok(PathBuf::from(value));
+        }
+    }
+    anyhow::bail!("Usage: skate --project-dir <path>")
+}

src/provider/claude.rs

@@ -1,5 +1,3 @@
-// Items used only in later phases or tests.
-#![allow(dead_code)]
 
 use futures::{SinkExt, Stream, StreamExt};
 use reqwest::Client;
@@ -92,7 +90,7 @@ impl ModelProvider for ClaudeProvider {
 ///   "model": "<model-id>",
 ///   "max_tokens": 8192,
 ///   "stream": true,
-///   "messages": [{ "role": "user"|"assistant", "content": "<text>" }, …]
+///   "messages": [{ "role": "user"|"assistant", "content": "<text>" }, ...]
 /// }
 /// ```
 ///
@@ -106,13 +104,13 @@ impl ModelProvider for ClaudeProvider {
 /// object.  The full event sequence for a successful turn is:
 ///
 /// ```text
-/// event: message_start          → InputTokens(n)
-/// event: content_block_start    → (ignored — signals a new content block)
-/// event: ping                   → (ignored — keepalive)
-/// event: content_block_delta    → TextDelta(chunk)  (repeated)
-/// event: content_block_stop     → (ignored — signals end of content block)
-/// event: message_delta          → OutputTokens(n)
-/// event: message_stop           → Done
+/// event: message_start          -> InputTokens(n)
+/// event: content_block_start    -> (ignored -- signals a new content block)
+/// event: ping                   -> (ignored -- keepalive)
+/// event: content_block_delta    -> TextDelta(chunk)  (repeated)
+/// event: content_block_stop     -> (ignored -- signals end of content block)
+/// event: message_delta          -> OutputTokens(n)
+/// event: message_stop           -> Done
 /// ```
 ///
 /// We stop reading as soon as `Done` is emitted; any bytes arriving after
@@ -201,14 +199,14 @@ async fn run_stream(
 /// Return the byte offset of the first `\n\n` in `buf`, or `None`.
 ///
 /// SSE uses a blank line (two consecutive newlines) as the event boundary.
-/// See [§9.2.6 of the SSE spec][sse-dispatch].
+/// See [Section 9.2.6 of the SSE spec][sse-dispatch].
 ///
 /// [sse-dispatch]: https://html.spec.whatwg.org/multipage/server-sent-events.html#event-stream-interpretation
 fn find_double_newline(buf: &[u8]) -> Option<usize> {
     buf.windows(2).position(|w| w == b"\n\n")
 }
 
-// ── SSE JSON types ────────────────────────────────────────────────────────────
+// -- SSE JSON types -----------------------------------------------------------
 //
 // These structs mirror the subset of the Anthropic SSE payload we actually
 // consume.  Unknown fields are silently ignored by serde.  Full schemas are
@@ -278,8 +276,8 @@ struct SseUsage {
 /// | `message_start`      | `.message.usage.input_tokens`      | `InputTokens(n)`             |
 /// | `content_block_delta`| `.delta.type == "text_delta"`      | `TextDelta(chunk)`           |
 /// | `message_delta`      | `.usage.output_tokens`             | `OutputTokens(n)`            |
-/// | `message_stop`       | —                                  | `Done`                       |
-/// | everything else      | —                                  | `None` (caller skips)        |
+/// | `message_stop`       | n/a                                | `Done`                       |
+/// | everything else      | n/a                                | `None` (caller skips)        |
 ///
 /// [sse-fields]: https://html.spec.whatwg.org/multipage/server-sent-events.html#parsing-an-event-stream
 fn parse_sse_event(event_str: &str) -> Option<StreamEvent> {
@@ -314,13 +312,13 @@ fn parse_sse_event(event_str: &str) -> Option<StreamEvent> {
 
         "message_stop" => Some(StreamEvent::Done),
 
-        // error, ping, content_block_start, content_block_stop — ignored or
+        // error, ping, content_block_start, content_block_stop -- ignored or
         // handled by the caller.
         _ => None,
     }
 }
 
-// ── Tests ─────────────────────────────────────────────────────────────────────
+// -- Tests --------------------------------------------------------------------
 
 #[cfg(test)]
 mod tests {
@@ -371,7 +369,7 @@ mod tests {
             .filter_map(parse_sse_event)
             .collect();
 
-        // content_block_start, ping, content_block_stop → None (filtered out)
+        // content_block_start, ping, content_block_stop -> None (filtered out)
         assert_eq!(events.len(), 5);
         assert!(matches!(events[0], StreamEvent::InputTokens(10)));
         assert!(matches!(&events[1], StreamEvent::TextDelta(s) if s == "Hello"));

src/provider/mod.rs

@@ -11,7 +11,7 @@ use crate::core::types::{ConversationMessage, StreamEvent};
 /// Trait for model providers that can stream conversation responses.
 ///
 /// Implementors take a conversation history and return a stream of [`StreamEvent`]s.
-/// The trait is provider-agnostic — no Claude-specific types appear here.
+/// The trait is provider-agnostic -- no Claude-specific types appear here.
 pub trait ModelProvider: Send + Sync {
     /// Stream a response from the model given the conversation history.
     fn stream<'a>(

src/tui/mod.rs

@@ -1,5 +1,3 @@
-// Types and functions are scaffolding — wired into main.rs in Stage 1.6.
-#![allow(dead_code)]
 
 //! TUI frontend: terminal lifecycle, rendering, and input handling.
 //!
@@ -37,9 +35,6 @@ pub enum TuiError {
     /// An underlying terminal I/O error.
     #[error("terminal IO error: {0}")]
     Io(#[from] std::io::Error),
-    /// The action channel was closed before the event loop exited cleanly.
-    #[error("action channel closed")]
-    ChannelClosed,
 }
 
 /// The UI-layer view of a conversation: rendered messages and the current input buffer.
@@ -209,13 +204,13 @@ fn update_scroll(state: &mut AppState, area: Rect) {
 ///
 /// Layout (top to bottom):
 /// ```text
-/// ┌──────────────────────────────┐
-/// │  conversation history        │  Fill(1)
-/// │                              │
-/// ├──────────────────────────────┤
-/// │ Input                        │  Length(3)
-/// │ > _                          │
-/// └──────────────────────────────┘
+/// +--------------------------------+
+/// |  conversation history          |  Fill(1)
+/// |                                |
+/// +--------------------------------+
+/// |  Input                         |  Length(3)
+/// |  > _                           |
+/// +--------------------------------+
 /// ```
 ///
 /// Role headers are coloured: `"You:"` in cyan, `"Assistant:"` in green.
@@ -255,8 +250,8 @@ fn render(frame: &mut Frame, state: &AppState) {
 /// ```text
 /// loop:
 ///   1. drain UIEvents (non-blocking try_recv)
-///   2. poll keyboard for up to 16 ms  ← spawn_blocking keeps async runtime free
-///   3. handle key event → Option<LoopControl>
+///   2. poll keyboard for up to 16 ms  (<- spawn_blocking keeps async runtime free)
+///   3. handle key event -> Option<LoopControl>
 ///   4. render frame (scroll updated inside draw closure)
 ///   5. act on LoopControl: send message or break
 /// ```