mindfork 0.18.3

//! Orchestrator tests — generation, the agentic loop, cancellation, readiness gates. Part of the [`super`]
//! module (fixtures in mod.rs). See docs/history/refactoring-god-objects.md, stage 3.

use super::*;

#[test]
fn effective_sampling_resolves_three_tiers() {
    let (_d, mut orch) = bare_orch();
    let mut profile = Profile::new("P", "sys");
    profile.default_sampling = Some(SamplingConfig {
        temperature: Some(0.5),
        ..Default::default()
    });
    let chat = Chat::from_profile(&profile, "c");
    let chat_id = chat.id;
    orch.profiles.push(profile);
    orch.chats.push(chat);

    // override = None → the profile's default is used.
    assert_eq!(orch.effective_sampling(chat_id).temperature, Some(0.5));

    // override = Some → it's used (the chat takes priority).
    orch.chats[0].sampling_override = Some(SamplingConfig {
        temperature: Some(0.9),
        ..Default::default()
    });
    assert_eq!(orch.effective_sampling(chat_id).temperature, Some(0.9));

    // Neither an override nor a profile default → the global one.
    orch.chats[0].sampling_override = None;
    orch.profiles[0].default_sampling = None;
    assert_eq!(orch.effective_sampling(chat_id).temperature, Some(0.1));
}

#[tokio::test]
async fn send_streams_and_persists_assistant_message() {
    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::Thoughts("думаю".into()),
        ChatChunk::Text("Привет".into()),
        ChatChunk::Finished(FinishReason::Stop),
    ])) as Arc<dyn EngineBackend>;
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = _d.path().to_path_buf();

    // Wait for activation and learn the active chat's id.
    let active = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    let chat_id = match active {
        AppEvent::ChatActivated { id, .. } => id,
        _ => unreachable!(),
    };

    cmd_tx
        .send(AppCommand::SendMessage("привет".into()))
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();

    // Finish and check that the chat was saved with two messages.
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened.json().load_chat(chat_id).unwrap().unwrap();
    assert_eq!(chat.messages.len(), 2);
    assert_eq!(chat.messages[0].role, MessageRole::User);
    assert_eq!(chat.messages[1].role, MessageRole::Assistant);
    assert_eq!(chat.messages[1].text, "Привет");
    assert_eq!(chat.messages[1].thoughts.as_deref(), Some("думаю"));
}

#[tokio::test]
async fn assistant_metadata_records_mode_model_and_filtered_sampling() {
    use crate::entities::sampling::SamplingConfig;
    use crate::shared::config::{CloudSettings, EngineSettings, ServerMode};

    // Cloud mode (OpenAI) + a model name; global sampling with top_k, which
    // the strict cloud dialect doesn't accept → it must not land in the metadata snapshot.
    let config = AppConfig {
        engine: EngineSettings {
            mode: ServerMode::OpenAi,
            openai: CloudSettings {
                model_name: Some("gpt-test".into()),
                ..Default::default()
            },
            ..Default::default()
        },
        default_sampling: SamplingConfig {
            temperature: Some(0.7),
            top_k: Some(40),
            max_tokens: Some(128),
            ..Default::default()
        },
        ..Default::default()
    };

    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::Text("Привет".into()),
        ChatChunk::Finished(FinishReason::Stop),
    ])) as Arc<dyn EngineBackend>;
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), config);
    let root = _d.path().to_path_buf();

    let active = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    let chat_id = match active {
        AppEvent::ChatActivated { id, .. } => id,
        _ => unreachable!(),
    };
    cmd_tx
        .send(AppCommand::SendMessage("привет".into()))
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened.json().load_chat(chat_id).unwrap().unwrap();
    let meta = chat.messages[1]
        .metadata
        .as_ref()
        .expect("expected a metadata snapshot");
    assert_eq!(meta.mode, ServerMode::OpenAi);
    assert_eq!(meta.model.as_deref(), Some("gpt-test"));
    // Fields available in the mode are kept; unavailable top_k and temperature (rejected
    // by GPT 5.5/5.6) are cleared.
    assert_eq!(meta.sampling.max_tokens, Some(128));
    assert_eq!(meta.sampling.top_k, None);
    assert_eq!(meta.sampling.temperature, None);
}

#[tokio::test]
async fn emits_token_counter_during_generation() {
    use crate::shared::api::contract::TokenUsage;
    // Two text deltas (the live count = 2), then exact usage from the server (= 5).
    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::Text("При".into()),
        ChatChunk::Text("вет".into()),
        ChatChunk::Usage(TokenUsage {
            prompt_tokens: 12,
            completion_tokens: 5,
            reasoning_tokens: 0,
            prefill: None,
        }),
        ChatChunk::Finished(FinishReason::Stop),
    ])) as Arc<dyn EngineBackend>;
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));

    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("привет".into()))
        .unwrap();

    // Right after the start — an estimate of the conversation (prompt): context=Some, inexact,
    // no reply yet (completion=0).
    let est = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::TokenUsage { .. }))
        .await
        .unwrap();
    assert!(
        matches!(
            est,
            AppEvent::TokenUsage {
                completion: 0,
                context: Some(c),
                context_exact: false,
                ..
            } if c > 0
        ),
        "conversation estimate: {est:?}"
    );

    // Reply deltas → the reply counter grows, the conversation estimate is untouched (None).
    let first = wait_for(&mut evt_rx, |e| {
        matches!(e, AppEvent::TokenUsage { context: None, .. })
    })
    .await
    .unwrap();
    assert!(
        matches!(
            first,
            AppEvent::TokenUsage {
                completion: 1,
                context: None,
                ..
            }
        ),
        "reply counter: {first:?}"
    );

    // The exact count from the server's usage arrives before completion: completion=5, context=12.
    let exact = wait_for(&mut evt_rx, |e| {
        matches!(
            e,
            AppEvent::TokenUsage {
                context_exact: true,
                ..
            }
        )
    })
    .await
    .unwrap();
    assert!(
        matches!(
            exact,
            AppEvent::TokenUsage {
                completion: 5,
                context: Some(12),
                context_exact: true,
                ..
            }
        ),
        "exact count from usage: {exact:?}"
    );

    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();
}

#[tokio::test]
async fn regenerate_replaces_last_assistant_message() {
    // Two different replies in turn: the original turn gives the first, regeneration gives the second.
    let backend = Arc::new(MockBackend::sequence(vec![
        vec![
            ChatChunk::Text("первый".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
        vec![
            ChatChunk::Text("второй".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])) as Arc<dyn EngineBackend>;
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = _d.path().to_path_buf();
    let active = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    let chat_id = match active {
        AppEvent::ChatActivated { id, .. } => id,
        _ => unreachable!(),
    };

    cmd_tx
        .send(AppCommand::SendMessage("вопрос".into()))
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    // ChatList after Finished — a sign that handle_done applied the reply (state Idle).
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatList(_)))
        .await
        .unwrap();

    cmd_tx.send(AppCommand::RegenerateLast).unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatList(_)))
        .await
        .unwrap();

    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // The old reply is replaced by the new one; the user's message isn't duplicated.
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened.json().load_chat(chat_id).unwrap().unwrap();
    assert_eq!(chat.messages.len(), 2, "{:?}", chat.messages);
    assert_eq!(chat.messages[0].role, MessageRole::User);
    assert_eq!(chat.messages[0].text, "вопрос");
    assert_eq!(chat.messages[1].role, MessageRole::Assistant);
    assert_eq!(chat.messages[1].text, "второй");
}

#[tokio::test]
async fn delete_last_exchange_restores_user_text() {
    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::Text("ответ".into()),
        ChatChunk::Finished(FinishReason::Stop),
    ])) as Arc<dyn EngineBackend>;
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = _d.path().to_path_buf();
    let active = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    let chat_id = match active {
        AppEvent::ChatActivated { id, .. } => id,
        _ => unreachable!(),
    };

    cmd_tx
        .send(AppCommand::SendMessage("забудь это".into()))
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatList(_)))
        .await
        .unwrap();

    cmd_tx.send(AppCommand::DeleteLastExchange).unwrap();
    let restore = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::RestoreInput(_)))
        .await
        .unwrap();
    assert!(matches!(restore, AppEvent::RestoreInput(t) if t == "забудь это"));

    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // The exchange is deleted entirely (a default chat with no greeting → empty).
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened.json().load_chat(chat_id).unwrap().unwrap();
    assert!(chat.messages.is_empty(), "{:?}", chat.messages);
    // The deleted exchange is saved for manual recovery (the user's message +
    // the assistant's reply); the input draft was empty (spec §11.7).
    assert_eq!(chat.deleted.len(), 1);
    let removed = &chat.deleted[0];
    assert_eq!(removed.messages.len(), 2);
    assert_eq!(removed.messages[0].role, MessageRole::User);
    assert_eq!(removed.messages[0].text, "забудь это");
    assert_eq!(removed.messages[1].role, MessageRole::Assistant);
    assert_eq!(removed.draft, "");
}

#[tokio::test]
async fn regenerate_without_user_message_is_noop() {
    // A chat with an assistant greeting but no user message — there's nothing to
    // regenerate; the command must not start generation.
    let backend = Arc::new(MockBackend::scripted(vec![ChatChunk::Finished(
        FinishReason::Stop,
    )])) as Arc<dyn EngineBackend>;
    let (_d, mut orch) = bare_orch();
    orch.engines.backend = Some(backend);
    let mut profile = Profile::new("P", "sys");
    profile.greeting = Some("Привет!".into());
    let mut chat = Chat::from_profile(&profile, "c");
    chat.push_message(Message::assistant("Привет!"));
    let chat_id = chat.id;
    orch.profiles.push(profile);
    orch.chats.push(chat);
    orch.active_id = Some(chat_id);

    orch.handle_regenerate();
    // State stayed Idle (generation wasn't started), history untouched.
    assert!(orch.gen_state.is_idle());
    assert_eq!(orch.chats[0].messages.len(), 1);
}

#[test]
fn regenerate_on_not_ready_server_keeps_reply() {
    // The server is still connecting (managed is loading the model) — regeneration must
    // neither wipe the previous reply nor go out as a request to a not-ready server (otherwise
    // a 503 "engine returned an error status" and the reply is lost). Regression test.
    let backend = Arc::new(MockBackend::scripted(vec![ChatChunk::Finished(
        FinishReason::Stop,
    )])) as Arc<dyn EngineBackend>;
    let (_d, mut orch) = bare_orch();
    orch.engines.backend = Some(backend);
    orch.engines.server_status = ServerStatus::Connecting; // not ready yet
    let profile = Profile::new("P", "sys");
    let mut chat = Chat::from_profile(&profile, "c");
    chat.push_message(Message::user("вопрос"));
    chat.push_message(Message::assistant("старый ответ"));
    let chat_id = chat.id;
    orch.profiles.push(profile);
    orch.chats.push(chat);
    orch.active_id = Some(chat_id);

    orch.handle_regenerate();

    // The reply is preserved, generation didn't start (history wasn't truncated).
    assert!(orch.gen_state.is_idle());
    assert_eq!(
        orch.chats[0].messages.len(),
        2,
        "the previous reply must not be wiped on a not-ready server"
    );
    assert_eq!(orch.chats[0].messages[1].text, "старый ответ");
}

#[test]
fn send_on_not_ready_server_restores_input() {
    // The server is still connecting — the send is rejected, but the text is returned to the
    // input box (RestoreInput), and no message is added to the chat.
    let backend = Arc::new(MockBackend::scripted(vec![])) as Arc<dyn EngineBackend>;
    let (_d, mut orch, mut rx) = bare_orch_rx();
    orch.engines.backend = Some(backend);
    orch.engines.server_status = ServerStatus::Connecting;
    let profile = Profile::new("P", "sys");
    let chat = Chat::from_profile(&profile, "c");
    let chat_id = chat.id;
    orch.profiles.push(profile);
    orch.chats.push(chat);
    orch.active_id = Some(chat_id);

    orch.handle_send("привет".into());

    assert!(orch.gen_state.is_idle());
    assert!(
        orch.chats[0].messages.is_empty(),
        "no message should be added"
    );
    // Among the emitted events — an error and the text returning to the input box.
    let mut got_error = false;
    let mut restored = None;
    while let Ok(ev) = rx.try_recv() {
        match ev {
            AppEvent::Error(_) => got_error = true,
            AppEvent::RestoreInput(t) => restored = Some(t),
            _ => {}
        }
    }
    assert!(got_error, "an error about not-readiness should be emitted");
    assert_eq!(restored.as_deref(), Some("привет"));
}

#[test]
fn ready_backend_gates_by_status() {
    let (_d, mut orch) = bare_orch();
    orch.engines.backend = Some(Arc::new(MockBackend::scripted(vec![])) as Arc<dyn EngineBackend>);

    orch.engines.server_status = ServerStatus::Ready;
    assert!(orch.ready_backend().is_some());

    for status in [
        ServerStatus::Connecting,
        ServerStatus::NotConfigured,
        ServerStatus::Disconnected("боль".into()),
    ] {
        orch.engines.server_status = status;
        assert!(orch.ready_backend().is_none());
    }
}

#[tokio::test]
async fn cancel_stops_generation_and_saves_partial() {
    let backend = Arc::new(MockBackend::cancellable(vec![ChatChunk::Text(
        "часть".into(),
    )])) as Arc<dyn EngineBackend>;
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();

    cmd_tx.send(AppCommand::SendMessage("hi".into())).unwrap();

    let mut sent_cancel = false;
    let mut cancelled = false;
    while let Some(ev) = evt_rx.recv().await {
        match ev {
            AppEvent::Chunk { .. } if !sent_cancel => {
                cmd_tx.send(AppCommand::Cancel).unwrap();
                sent_cancel = true;
            }
            AppEvent::Finished { reason, .. } => {
                assert_eq!(reason, FinishReason::Cancelled);
                cancelled = true;
                break;
            }
            _ => {}
        }
    }
    assert!(cancelled);

    drop(cmd_tx);
    handle.await.unwrap();
}

#[tokio::test]
async fn agentic_loop_executes_tool_then_finalizes() {
    use crate::shared::api::contract::ToolCallDelta;
    // Round 1: a call to note_save → round 2: the final text.
    let backend = Arc::new(MockBackend::sequence(vec![
        vec![
            ChatChunk::ToolCall(ToolCallDelta {
                thought_signature: None,
                index: 0,
                id: Some("c1".into()),
                name: Some("note_save".into()),
                arguments: "{\"content\":\"любит чай\"}".into(),
            }),
            ChatChunk::Finished(FinishReason::ToolCalls),
        ],
        vec![
            ChatChunk::Text("Запомнил.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])) as Arc<dyn EngineBackend>;

    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = _d.path().to_path_buf();
    let active = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    let chat_id = match active {
        AppEvent::ChatActivated { id, .. } => id,
        _ => unreachable!(),
    };

    cmd_tx
        .send(AppCommand::SendMessage("запомни про чай".into()))
        .unwrap();

    // The tool-execution event reaches the UI.
    let tool_ev = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ToolCall { .. }))
        .await
        .unwrap();
    assert!(matches!(tool_ev, AppEvent::ToolCall { name, .. } if name == "note_save"));
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();

    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // History: user → assistant(tool_calls) → tool → assistant(final).
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened.json().load_chat(chat_id).unwrap().unwrap();
    assert_eq!(chat.messages.len(), 4, "{:?}", chat.messages);
    assert_eq!(chat.messages[0].role, MessageRole::User);
    assert_eq!(chat.messages[1].role, MessageRole::Assistant);
    assert_eq!(chat.messages[1].tool_calls.len(), 1);
    assert_eq!(chat.messages[1].tool_calls[0].name, "note_save");
    assert_eq!(chat.messages[2].role, MessageRole::Tool);
    assert_eq!(chat.messages[3].text, "Запомнил.");

    // The note is actually saved by the tool (isolation by profile).
    let notes = reopened
        .db()
        .note_list(chat.profile_id, None, &[], None)
        .unwrap();
    assert_eq!(notes.len(), 1);
    assert!(notes[0].content.contains("любит чай"));
}

#[tokio::test]
async fn disabled_tool_is_refused_by_loop() {
    use crate::shared::api::contract::ToolCallDelta;
    // python_exec is disabled globally (spawn_orch: python_enabled = false) —
    // even if the model calls it, the loop will refuse without executing it.
    let backend = Arc::new(MockBackend::sequence(vec![
        vec![
            ChatChunk::ToolCall(ToolCallDelta {
                thought_signature: None,
                index: 0,
                id: Some("c1".into()),
                name: Some("python_exec".into()),
                arguments: "{\"code\":\"print(1)\"}".into(),
            }),
            ChatChunk::Finished(FinishReason::ToolCalls),
        ],
        vec![
            ChatChunk::Text("ок".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])) as Arc<dyn EngineBackend>;

    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("посчитай".into()))
        .unwrap();

    let tool_ev = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ToolCall { .. }))
        .await
        .unwrap();
    match tool_ev {
        AppEvent::ToolCall { name, result, .. } => {
            assert_eq!(name, "python_exec");
            assert!(result.contains("недоступен"), "got: {result}");
        }
        _ => unreachable!(),
    }
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    drop(cmd_tx);
    handle.await.unwrap();
}

#[tokio::test]
async fn tool_round_limit_is_respected() {
    use crate::shared::api::contract::ToolCallDelta;
    // The engine always requests a tool — the round limit should kick in.
    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::ToolCall(ToolCallDelta {
            thought_signature: None,
            index: 0,
            id: Some("c1".into()),
            name: Some("get_sampling".into()),
            arguments: "{}".into(),
        }),
        ChatChunk::Finished(FinishReason::ToolCalls),
    ])) as Arc<dyn EngineBackend>;

    let config = AppConfig {
        max_tool_rounds: 2,
        ..Default::default()
    };
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), config);
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("зациклись".into()))
        .unwrap();

    // Reach Finished; the limit produces an error marker, but generation completes.
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    drop(cmd_tx);
    handle.await.unwrap();
}

#[tokio::test]
async fn round_limit_forces_final_synthesis_without_tools() {
    use crate::shared::api::contract::ToolCallDelta;
    // With the round limit, the model must not leave the user with no reply: after
    // exhausting the rounds a final round runs WITHOUT tools, where the model
    // sums up. Scripts: rounds 1 and 2 — tool calls, round 3 (forced synthesis) — text.
    let toolcall = || {
        vec![
            ChatChunk::ToolCall(ToolCallDelta {
                thought_signature: None,
                index: 0,
                id: Some("c1".into()),
                name: Some("get_sampling".into()),
                arguments: "{}".into(),
            }),
            ChatChunk::Finished(FinishReason::ToolCalls),
        ]
    };
    let backend = Arc::new(MockBackend::sequence(vec![
        toolcall(),
        toolcall(),
        vec![
            ChatChunk::Text("Итог по собранному материалу.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])) as Arc<dyn EngineBackend>;

    let config = AppConfig {
        max_tool_rounds: 1,
        ..Default::default()
    };
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), config);
    let root = _d.path().to_path_buf();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("собери и проанализируй".into()))
        .unwrap();

    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // The last message — the assistant's final summary text, not emptiness.
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap();
    let last = chat.messages.last().unwrap();
    assert_eq!(last.role, MessageRole::Assistant);
    assert_eq!(last.text, "Итог по собранному материалу.");
}

/// The final round says what it took away. Its request is the one that just
/// produced a call, less the tool list, and Gemma 4 answered that with the
/// same call written out as text — a raw `<|tool_call>` as the user's reply
/// (docs/research/e2e-gate-budget.md §7). The note rides on the last tool
/// result, on the wire only: the stored chat never carries it.
#[tokio::test]
async fn the_final_round_tells_the_model_the_tools_are_gone() {
    use super::subagent::{ScriptRecorder, call, text};
    let backend = ScriptRecorder::new(vec![
        call("c1", "get_sampling", "{}"),
        call("c2", "get_sampling", "{}"),
        text("Summary."),
    ]);
    let config = AppConfig {
        max_tool_rounds: 1,
        ..no_auto_cfg()
    };
    let (dir, cmd_tx, mut evt_rx, handle) =
        spawn_orch_cfg(Some(backend.clone() as Arc<dyn EngineBackend>), config);
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("gather and sum up".into()))
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    let loc = crate::shared::i18n::locale(crate::shared::i18n::Lang::default());
    let note = loc.tf("prompt.round_limit_final", &[("max_rounds", "1")]);
    let reqs = backend.requests();
    assert_eq!(
        reqs.len(),
        3,
        "round 1, the refused round 2, the final round"
    );
    // The refused round's request carried the tools and no note…
    assert!(!reqs[1].tools.is_empty());
    assert!(!reqs[1].messages.last().unwrap().content.contains(&note));
    // …the final one has neither tools nor silence about it.
    let fin = &reqs[2];
    assert!(fin.tools.is_empty());
    let last = fin.messages.last().unwrap();
    assert_eq!(last.role, crate::shared::api::contract::ApiRole::Tool);
    assert!(last.content.ends_with(&note), "{}", last.content);
    assert_eq!(
        fin.messages.len(),
        reqs[1].messages.len(),
        "no turn of its own"
    );

    let chat = Storage::open(Paths::with_root(dir.path()))
        .unwrap()
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap();
    assert!(
        chat.messages.iter().all(|m| !m.text.contains(&note)),
        "the note is the wire's, not the chat's"
    );
    assert_eq!(chat.messages.last().unwrap().text, "Summary.");
}

/// Sends one message on a [`ScriptRecorder`](super::subagent::ScriptRecorder)
/// engine and returns the requests it saw and the chat's last message text.
async fn turn_on_recorder(
    scripts: Vec<super::subagent::Script>,
) -> (Vec<crate::shared::api::ChatRequest>, String) {
    let backend = super::subagent::ScriptRecorder::new(scripts);
    let (dir, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(
        Some(backend.clone() as Arc<dyn EngineBackend>),
        no_auto_cfg(),
    );
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("look it up".into()))
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();
    let chat = Storage::open(Paths::with_root(dir.path()))
        .unwrap()
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap();
    (
        backend.requests(),
        chat.messages.last().unwrap().text.clone(),
    )
}

/// A clean stop with nothing in it — no text, no call.
fn empty_stop() -> super::subagent::Script {
    super::subagent::Script {
        chunks: vec![ChatChunk::Finished(FinishReason::Stop)],
        hang: false,
    }
}

/// An answer that comes back empty right after a tool round is asked once more,
/// muted, and the second answer is what the chat keeps
/// (docs/research/empty-round-after-tools.md). Without it the turn ended on the
/// round's narration — Qwen 3.6, `code_workspace_navigate_e2e_live` on the gate.
#[tokio::test]
async fn an_empty_answer_after_a_tool_round_is_asked_once_more_muted() {
    use super::subagent::{call, text};
    let (reqs, last) = turn_on_recorder(vec![
        call("c1", "get_sampling", "{}"),
        empty_stop(),
        text("The temperature is 0.7."),
    ])
    .await;
    assert_eq!(
        reqs.len(),
        3,
        "the tool round, the empty answer, the re-ask"
    );
    assert_eq!(
        reqs[2].messages.len(),
        reqs[1].messages.len(),
        "the re-ask adds nothing to the conversation"
    );
    assert_ne!(reqs[1].sampling.reasoning_budget, Some(0));
    assert_eq!(
        reqs[2].sampling.reasoning_budget,
        Some(0),
        "the re-ask is muted"
    );
    assert_eq!(last, "The temperature is 0.7.");
}

/// The mute is the re-ask's alone: when it answers with another call, the
/// rounds after it think again.
#[tokio::test]
async fn the_muted_re_ask_does_not_mute_the_rounds_after_it() {
    use super::subagent::{call, text};
    let (reqs, last) = turn_on_recorder(vec![
        call("c1", "get_sampling", "{}"),
        empty_stop(),
        call("c2", "get_sampling", "{}"),
        text("Done."),
    ])
    .await;
    assert_eq!(reqs.len(), 4);
    assert_eq!(reqs[2].sampling.reasoning_budget, Some(0));
    assert_eq!(
        reqs[3].sampling.reasoning_budget, reqs[1].sampling.reasoning_budget,
        "the turn's own budget is back"
    );
    assert_eq!(last, "Done.");
}

/// Once per turn: a second empty answer stands, and no third request is made.
#[tokio::test]
async fn a_second_empty_answer_after_tools_is_not_asked_again() {
    use super::subagent::{call, text};
    let (reqs, last) = turn_on_recorder(vec![
        call("c1", "get_sampling", "{}"),
        empty_stop(),
        empty_stop(),
        text("never asked for"),
    ])
    .await;
    assert_eq!(reqs.len(), 3);
    assert_ne!(last, "never asked for");

    // …and once per turn, not once per empty answer: when the re-ask goes on
    // with another call and the round after it is empty too, that one stands.
    let (reqs, last) = turn_on_recorder(vec![
        call("c1", "get_sampling", "{}"),
        empty_stop(),
        call("c2", "get_sampling", "{}"),
        empty_stop(),
        text("never asked for"),
    ])
    .await;
    assert_eq!(reqs.len(), 4);
    assert_ne!(last, "never asked for");
}

/// Only after tools: an empty first answer, with no tool round before it, is the
/// turn's answer as before — and so is one cut by the cap (`Length`).
#[tokio::test]
async fn an_empty_answer_with_no_tool_round_or_cut_by_the_cap_is_not_asked_again() {
    use super::subagent::{Script, call, text};
    let (reqs, _) = turn_on_recorder(vec![empty_stop(), text("never asked for")]).await;
    assert_eq!(reqs.len(), 1, "no tool round, no re-ask");

    let cut = Script {
        chunks: vec![ChatChunk::Finished(FinishReason::Length)],
        hang: false,
    };
    let (reqs, _) = turn_on_recorder(vec![
        call("c1", "get_sampling", "{}"),
        cut,
        text("never asked for"),
    ])
    .await;
    assert_eq!(reqs.len(), 2, "a cut answer is not an empty one");
}

/// Where the note goes: onto a trailing `tool` or `user` turn, never as a
/// `user` turn after one — a template that enforces alternation refuses that —
/// and as a turn of its own only when the history ends otherwise.
#[test]
fn the_final_note_joins_the_last_turn_it_can() {
    use crate::app::orchestrator::generation::note_on_last;
    use crate::shared::api::contract::{ApiMessage, ApiRole};
    let mut tool = vec![ApiMessage::user("q"), ApiMessage::tool("c1", "result")];
    note_on_last(&mut tool, "NOTE");
    assert_eq!(tool.len(), 2);
    assert_eq!(tool[1].content, "result\n\nNOTE");

    let mut user = vec![ApiMessage::user("q")];
    note_on_last(&mut user, "NOTE");
    assert_eq!(user.len(), 1);
    assert_eq!(user[0].content, "q\n\nNOTE");

    let mut empty_tool = vec![ApiMessage::tool("c1", "")];
    note_on_last(&mut empty_tool, "NOTE");
    assert_eq!(empty_tool[0].content, "NOTE");

    let mut assistant = vec![ApiMessage::user("q"), ApiMessage::assistant("a")];
    note_on_last(&mut assistant, "NOTE");
    assert_eq!(assistant.len(), 3);
    assert_eq!(assistant[2].role, ApiRole::User);
    assert_eq!(assistant[2].content, "NOTE");

    let mut none = Vec::new();
    note_on_last(&mut none, "NOTE");
    assert_eq!(none.len(), 1);
}

#[tokio::test]
async fn followup_tool_makes_two_assistant_messages() {
    use crate::shared::api::contract::ToolCallDelta;
    // Round 1: text + a call to send_followup_message → round 2: the second message.
    let backend = Arc::new(MockBackend::sequence(vec![
        vec![
            ChatChunk::Text("Первое сообщение.".into()),
            ChatChunk::ToolCall(ToolCallDelta {
                thought_signature: None,
                index: 0,
                id: Some("c1".into()),
                name: Some("send_followup_message".into()),
                arguments: "{}".into(),
            }),
            ChatChunk::Finished(FinishReason::ToolCalls),
        ],
        vec![
            ChatChunk::Text("Второе сообщение.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])) as Arc<dyn EngineBackend>;

    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = _d.path().to_path_buf();
    enable_all_tools(&cmd_tx, &mut evt_rx).await;
    cmd_tx
        .send(AppCommand::SendMessage("давай".into()))
        .unwrap();

    // The UI receives the signal to start a new bubble.
    wait_for(&mut evt_rx, |e| {
        matches!(e, AppEvent::AssistantContinue { .. })
    })
    .await
    .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // History: user → assistant(followup tool_call) → tool → assistant(2nd, new_bubble).
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap();
    assert_eq!(chat.messages.len(), 4, "{:?}", chat.messages);
    assert_eq!(chat.messages[1].role, MessageRole::Assistant);
    assert_eq!(chat.messages[1].text, "Первое сообщение.");
    assert_eq!(chat.messages[1].tool_calls[0].name, "send_followup_message");
    assert!(!chat.messages[1].new_bubble);
    assert_eq!(chat.messages[2].role, MessageRole::Tool);
    assert_eq!(chat.messages[3].text, "Второе сообщение.");
    assert!(
        chat.messages[3].new_bubble,
        "the second message should be a separate bubble"
    );
    // The control tool discards nothing.
    assert!(chat.deleted.is_empty());
}

#[tokio::test]
async fn rewrite_tool_discards_partial_and_saves_it() {
    use crate::shared::api::contract::ToolCallDelta;
    // Round 1: an incorrect text + a call to rewrite_current_message → round 2: the rewritten one.
    let backend = Arc::new(MockBackend::sequence(vec![
        vec![
            ChatChunk::Text("Неправильный ответ".into()),
            ChatChunk::ToolCall(ToolCallDelta {
                thought_signature: None,
                index: 0,
                id: Some("c1".into()),
                name: Some("rewrite_current_message".into()),
                arguments: "{}".into(),
            }),
            ChatChunk::Finished(FinishReason::ToolCalls),
        ],
        vec![
            ChatChunk::Text("Правильный ответ.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])) as Arc<dyn EngineBackend>;

    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = _d.path().to_path_buf();
    enable_all_tools(&cmd_tx, &mut evt_rx).await;
    cmd_tx
        .send(AppCommand::SendMessage("вопрос".into()))
        .unwrap();

    // The UI receives the signal to discard the current bubble.
    wait_for(&mut evt_rx, |e| {
        matches!(e, AppEvent::AssistantRewrite { .. })
    })
    .await
    .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // History: user → assistant(rewritten). The incorrect version is in the deleted archive.
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap();
    assert_eq!(chat.messages.len(), 2, "{:?}", chat.messages);
    assert_eq!(chat.messages[0].role, MessageRole::User);
    assert_eq!(chat.messages[1].role, MessageRole::Assistant);
    assert_eq!(chat.messages[1].text, "Правильный ответ.");
    // The discarded (incorrect) reply + its tool message are saved for recovery.
    assert_eq!(chat.deleted.len(), 1);
    let discarded = &chat.deleted[0].messages;
    assert_eq!(discarded[0].role, MessageRole::Assistant);
    assert_eq!(discarded[0].text, "Неправильный ответ");
    assert_eq!(discarded[0].tool_calls[0].name, "rewrite_current_message");
}

/// A bounded [`wait_for`].
///
/// The shared helper blocks until the event channel *closes*, so a regression that
/// simply stops emitting an event makes a test hang rather than fail — and in CI a
/// hang reads as broken infrastructure instead of a broken promise. Measured while
/// mutation-testing the note below: with its `send` removed the test ran past ten
/// minutes; bounded, it fails in five seconds.
async fn wait_for_bounded<F: Fn(&AppEvent) -> bool>(
    rx: &mut UnboundedReceiver<AppEvent>,
    what: &str,
    pred: F,
) -> AppEvent {
    tokio::time::timeout(std::time::Duration::from_secs(5), wait_for(rx, pred))
        .await
        .unwrap_or_else(|_| panic!("timed out waiting for {what}"))
        .unwrap_or_else(|| panic!("the event stream closed before {what}"))
}

/// The regression this whole change exists for: a provider that dies **after** the
/// stream opened used to end the turn silently — `Finished(Error)` pushes no note
/// (only `Cancelled` does), so a reply cut off mid-sentence was indistinguishable
/// from a finished one, and the only trace was a line in the file log.
#[tokio::test]
async fn a_mid_stream_error_reaches_the_feed_and_keeps_the_partial_reply() {
    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::Text("Половина отв".into()),
        ChatChunk::Error {
            message: "overloaded_error: Overloaded".into(),
            transient: true,
        },
        ChatChunk::Finished(FinishReason::Error),
    ])) as Arc<dyn EngineBackend>;
    let (dir, cmd_tx, mut evt_rx, handle) = spawn_orch(Some(backend));
    let root = dir.path().to_path_buf();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("вопрос".into()))
        .unwrap();

    // The note comes before Finished, and `wait_for` drains — so it must be
    // pulled first or the wait below eats it.
    let AppEvent::Error(note) = wait_for_bounded(&mut evt_rx, "the failure note", |e| {
        matches!(e, AppEvent::Error(_))
    })
    .await
    else {
        unreachable!("filtered on Error")
    };
    assert!(
        note.contains("overloaded_error"),
        "the note must carry what the provider said, got {note:?}"
    );
    wait_for_bounded(&mut evt_rx, "Finished", |e| {
        matches!(e, AppEvent::Finished { .. })
    })
    .await;
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    // What did arrive is still there — the note explains it, it does not replace it.
    let reopened = Storage::open(Paths::with_root(&root)).unwrap();
    let chat = reopened
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap();
    assert_eq!(chat.messages.len(), 2, "{:?}", chat.messages);
    assert_eq!(chat.messages[1].role, MessageRole::Assistant);
    assert_eq!(chat.messages[1].text, "Половина отв");
}

/// Which advice a failed turn gets. Asserted as the *decision* (the bundle key)
/// rather than the prose, so a translation edit cannot break it and a wrong branch
/// cannot pass.
#[test]
fn the_failure_note_is_chosen_by_condition() {
    use crate::app::orchestrator::generation::engine_error_key;

    let overflow =
        "engine returned status 400: {\"error\":{\"type\":\"exceed_context_size_error\"}}";
    let plain = "overloaded_error: Overloaded";

    // Overflow wins over everything: it is the one failure with a specific way
    // out, and which way out depends on whether compression is switched on.
    assert_eq!(
        engine_error_key(overflow, true, false, false),
        "ui.err.context_overflow"
    );
    assert_eq!(
        engine_error_key(overflow, false, false, false),
        "ui.err.context_overflow_off"
    );
    // ...including when text was already on screen — pointing at `/compact` beats
    // "press Ctrl+R", which would just overflow again — and regardless of
    // whether `/continue` could pick the fragment up: continuing would just
    // overflow again too, so `/compact` stays the named way out.
    assert_eq!(
        engine_error_key(overflow, true, true, true),
        "ui.err.context_overflow"
    );
    // Anything else: text already on screen means a fragment, so say so and name
    // the way to a whole reply — the resuming one when the mode can resume
    // (fork F9), the regenerating one when it cannot; nothing on screen is a
    // plain failure, whatever the mode.
    assert_eq!(
        engine_error_key(plain, true, true, true),
        "ui.err.generation_interrupted_continuable"
    );
    assert_eq!(
        engine_error_key(plain, true, true, false),
        "ui.err.generation_interrupted"
    );
    assert_eq!(
        engine_error_key(plain, true, false, true),
        "ui.err.generation_failed"
    );
    assert_eq!(
        engine_error_key(plain, true, false, false),
        "ui.err.generation_failed"
    );
}

#[tokio::test]
async fn send_without_backend_emits_error() {
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch(None);
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::SendMessage("hi".into())).unwrap();

    let err = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Error(_)))
        .await
        .unwrap();
    assert!(matches!(err, AppEvent::Error(_)));

    drop(cmd_tx);
    handle.await.unwrap();
}

/// The code workspace is exempt from `max_tool_rounds` (spec §9.12), and the
/// exemption has to be **both** halves of a promise: a turn survives far past a
/// budget that would have cut it, and it still ends.
///
/// The engine here asks for `code_list` forever, with a budget of one round and
/// a project ceiling of three — without the exemption the turn would stop at the
/// first round, and without the ceiling it would never stop at all. The ceiling
/// is read from `workspace.max_rounds` rather than from a constant, which is
/// what the count below pins: three is a number only the setting can produce.
/// The wait is bounded, so a regression fails instead of hanging
/// (docs/lessons.md §2).
#[tokio::test]
async fn workspace_rounds_do_not_spend_the_budget_but_still_end() {
    use crate::shared::api::contract::ToolCallDelta;
    let backend = Arc::new(MockBackend::scripted(vec![
        ChatChunk::ToolCall(ToolCallDelta {
            thought_signature: None,
            index: 0,
            id: Some("c1".into()),
            name: Some(crate::features::tools::code::CODE_LIST_ID.into()),
            arguments: "{}".into(),
        }),
        ChatChunk::Finished(FinishReason::ToolCalls),
    ])) as Arc<dyn EngineBackend>;

    let project = tempfile::tempdir().unwrap();
    std::fs::write(project.path().join("a.rs"), "fn main() {}\n").unwrap();

    let config = AppConfig {
        // One round: enough that an unexempt tool would be cut off immediately.
        max_tool_rounds: 1,
        workspace: crate::shared::config::WorkspaceSettings {
            // Well below the default of 500, so the turn ends quickly *and* the
            // count proves the setting is what bounds it.
            max_rounds: 3,
            ..Default::default()
        },
        ..Default::default()
    };
    let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), config);
    let profile = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ProfileList(_)))
        .await
        .and_then(|e| match e {
            AppEvent::ProfileList(ps) => ps.first().map(|p| p.id),
            _ => None,
        })
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::UpdateProfile {
            id: profile,
            edit: Box::new(ProfileEdit {
                enabled_tools: Some(vec![crate::features::tools::code::CODE_LIST_ID.into()]),
                ..Default::default()
            }),
        })
        .unwrap();
    cmd_tx
        .send(AppCommand::ProjectAttach {
            path: project.path().to_string_lossy().into_owned(),
        })
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ProjectProgress(_)))
        .await
        .unwrap();

    cmd_tx
        .send(AppCommand::SendMessage("посмотри проект".into()))
        .unwrap();
    let mut calls = 0usize;
    let mut errors: Vec<String> = Vec::new();
    let finished = tokio::time::timeout(std::time::Duration::from_secs(60), async {
        while let Some(ev) = evt_rx.recv().await {
            match ev {
                AppEvent::ToolCall { .. } => calls += 1,
                AppEvent::Error(msg) => errors.push(msg),
                AppEvent::Finished { .. } => return true,
                _ => {}
            }
        }
        false
    })
    .await
    .expect("the turn must end rather than loop forever");
    assert!(finished, "the event stream closed before the turn finished");
    assert!(
        calls > 1,
        "a budget of one round must not stop an exempt tool: {calls} calls"
    );
    assert_eq!(
        calls, 3,
        "the ceiling must come from workspace.max_rounds: {calls} calls"
    );
    // And the note that ends such a turn must name the limit that fired. It
    // quoted `max_tool_rounds` for both until stage 3 — sending the user to
    // change a setting that was not the problem (docs/lessons.md §4).
    let loc = crate::shared::i18n::locale(crate::shared::i18n::Lang::Ru);
    let expected = loc.tf("loop.workspace_round_limit_reached", &[("max_rounds", "3")]);
    assert!(
        errors.contains(&expected),
        "the project limit must name itself: {errors:?}"
    );
    drop(cmd_tx);
    handle.await.unwrap();
}

// ---------- /continue (spec §6.4, docs/research/continue-generation.md) ----------

/// The echo filter in isolation: an exact echo yields only the continuation, a
/// non-echoing stream loses nothing (including one that first coincides with
/// part of the seed), and a multi-byte character split across deltas keeps its
/// boundaries. Mutating the mismatch flush to drop the withheld bytes fails
/// the third case; mutating full-match detection fails the first.
#[test]
fn the_echo_filter_strips_an_echo_and_passes_real_content() {
    use super::super::generation::EchoFilter;

    let mut f = EchoFilter::new("Per the atlas".into());
    assert_eq!(f.push("Per the "), "");
    assert_eq!(f.push("atlas"), "");
    assert_eq!(f.push(" Paris."), " Paris.");

    let mut f = EchoFilter::new("Per the atlas".into());
    assert_eq!(f.push(" Paris."), " Paris.", "no echo: everything flows");

    let mut f = EchoFilter::new("Per the atlas".into());
    assert_eq!(f.push("Per"), "", "still ambiguous: withheld");
    assert_eq!(
        f.push(" it goes"),
        "Per it goes",
        "divergence flushes the withheld prefix"
    );

    let mut f = EchoFilter::new("По".into());
    assert_eq!(f.push("П"), "");
    assert_eq!(
        f.push("о!"),
        "!",
        "the split multi-byte echo is consumed whole"
    );
}

/// The in-place fold (fork F8) and the model-name rule (fork F7): the text
/// joins at the model's own seam, the id and timestamp stay the seed's, the
/// end state becomes the round's — and the model name stays the seed's unless
/// the continuation outgrew what it continued.
#[test]
fn merge_continuation_appends_in_place_and_keeps_the_model_name() {
    use super::super::generation::merge_continuation;
    use crate::entities::message::{Message, MessageFinish, MessageMetadata};

    let meta = |model: &str, finish: MessageFinish| MessageMetadata {
        sampling: Default::default(),
        mode: Default::default(),
        model: Some(model.into()),
        provider: None,
        cost_nanos: None,
        finish: Some(finish),
    };
    let mut seed = Message::assistant("Начало было длинным и обстоятельным");
    seed.metadata = Some(meta("gemma-4-31b", MessageFinish::Cancelled));
    let (id, stamp) = (seed.id, seed.timestamp);

    let mut round = Message::assistant(", конец.");
    round.metadata = Some(meta("qwen-3.6", MessageFinish::Stop));
    merge_continuation(&mut seed, round);

    assert_eq!(seed.text, "Начало было длинным и обстоятельным, конец.");
    assert_eq!(seed.id, id, "the same reply, finished later");
    assert_eq!(seed.timestamp, stamp);
    let md = seed.metadata.as_ref().unwrap();
    assert_eq!(
        md.finish,
        Some(MessageFinish::Stop),
        "the end state is the round's"
    );
    assert_eq!(
        md.model.as_deref(),
        Some("gemma-4-31b"),
        "a short continuation keeps the header on the model that wrote the bulk"
    );

    // The continuation outgrew the partial: the header follows the larger share.
    let mut seed = Message::assistant("Нач");
    seed.metadata = Some(meta("gemma-4-31b", MessageFinish::Error));
    let mut round = Message::assistant("ало, середина и весь длинный конец ответа.");
    round.metadata = Some(meta("qwen-3.6", MessageFinish::Stop));
    merge_continuation(&mut seed, round);
    assert_eq!(
        seed.metadata.unwrap().model.as_deref(),
        Some("qwen-3.6"),
        "an outgrown partial takes the continuing model's name"
    );
}

/// What [`interrupted_turn`] hands back: the harness, plus the first turn's
/// `Finished` event (some tests assert on what it announced).
type InterruptedOrch = (
    tempfile::TempDir,
    tokio::sync::mpsc::UnboundedSender<AppCommand>,
    tokio::sync::mpsc::UnboundedReceiver<AppEvent>,
    tokio::task::JoinHandle<()>,
    AppEvent,
);

/// The shared road into a `/continue` scenario: one scripted turn that leaves
/// its interruption on disk, waited past the list re-emit — the proof the tail
/// has been applied before a test reads it (the result lands on its own
/// channel). Each test supplies only its script and its distinctive middle:
/// when the third test starts the same way as the first two, that opening is
/// a fixture (docs/lessons.md §2).
async fn interrupted_turn(script: Vec<Vec<ChatChunk>>) -> InterruptedOrch {
    let backend = Arc::new(MockBackend::sequence(script)) as Arc<dyn EngineBackend>;
    let (dir, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), no_auto_cfg());
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
        .await
        .unwrap();
    cmd_tx
        .send(AppCommand::SendMessage("вопрос".into()))
        .unwrap();
    let finished = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatList(_)))
        .await
        .unwrap();
    (dir, cmd_tx, evt_rx, handle, finished)
}

/// The chat as the closed app left it on disk.
fn reload_chat(root: &std::path::Path) -> crate::entities::chat::Chat {
    Storage::open(Paths::with_root(root))
        .unwrap()
        .json()
        .load_chats()
        .unwrap()
        .into_iter()
        .next()
        .unwrap()
}

/// The whole route: a cancelled reply, then `/continue` — the turn announces
/// itself as a continuation (the feed re-opens the bubble on that flag), and
/// on disk there is still **one** assistant message, joined at the model's own
/// seam, with the end state re-recorded from `Cancelled` to `Stop`.
#[tokio::test]
async fn continue_appends_into_the_same_message() {
    use crate::entities::message::MessageFinish;

    let (dir, cmd_tx, mut evt_rx, handle, _) = interrupted_turn(vec![
        vec![
            ChatChunk::Text("Нача".into()),
            ChatChunk::Finished(FinishReason::Cancelled),
        ],
        vec![
            ChatChunk::Text("ло готово.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])
    .await;

    cmd_tx.send(AppCommand::ContinueLast).unwrap();
    let started = wait_for(&mut evt_rx, |e| {
        matches!(e, AppEvent::GenerationStarted { .. })
    })
    .await
    .unwrap();
    assert!(
        matches!(
            started,
            AppEvent::GenerationStarted {
                continuation: true,
                ..
            }
        ),
        "the feed must be told to re-open the bubble instead of pushing one"
    );
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    let chat = reload_chat(dir.path());
    assert_eq!(chat.messages.len(), 2, "{:?}", chat.messages);
    assert_eq!(
        chat.messages[1].text, "Начало готово.",
        "joined at the seam, with no separator"
    );
    assert_eq!(
        chat.messages[1].metadata.as_ref().and_then(|m| m.finish),
        Some(MessageFinish::Stop),
        "the end state is the continuation's"
    );
}

/// llama.cpp returns prefill + continuation (research §7.1): the echoed prefix
/// must reach neither the screen nor the stored message. The second turn's
/// script echoes the partial verbatim; the feed sees only the new text, and
/// the disk holds it exactly once.
#[tokio::test]
async fn continue_strips_the_echoed_prefill() {
    let (dir, cmd_tx, mut evt_rx, handle, _) = interrupted_turn(vec![
        vec![
            ChatChunk::Text("Нача".into()),
            ChatChunk::Finished(FinishReason::Cancelled),
        ],
        vec![
            ChatChunk::Text("Нача".into()),
            ChatChunk::Text("ло готово.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])
    .await;

    cmd_tx.send(AppCommand::ContinueLast).unwrap();
    let mut chunks: Vec<String> = Vec::new();
    while let Some(ev) = evt_rx.recv().await {
        match ev {
            AppEvent::Chunk { text, .. } => chunks.push(text),
            AppEvent::Finished { .. } => break,
            _ => {}
        }
    }
    assert_eq!(
        chunks,
        vec!["ло готово.".to_string()],
        "the echoed prefix must not re-stream into the feed"
    );
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    assert_eq!(
        reload_chat(dir.path()).messages[1].text,
        "Начало готово.",
        "the echo must not double into the stored text"
    );
}

/// A length-cut reply records `Length` as its end state, and the turn's
/// `Finished` carries `continuable` so the feed's truncation note can name
/// `/continue` (fork F9). The mock supervisor runs as managed — the mode that
/// supports continuation.
#[tokio::test]
async fn a_length_cut_reply_is_recorded_and_announced_as_continuable() {
    use crate::entities::message::MessageFinish;

    let (dir, cmd_tx, _evt_rx, handle, finished) = interrupted_turn(vec![vec![
        ChatChunk::Text("Полов".into()),
        ChatChunk::Finished(FinishReason::Length),
    ]])
    .await;
    assert!(
        matches!(
            finished,
            AppEvent::Finished {
                reason: FinishReason::Length,
                continuable: true,
                ..
            }
        ),
        "a managed-mode length cut must announce the resume route: {finished:?}"
    );
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    assert_eq!(
        reload_chat(dir.path()).messages[1]
            .metadata
            .as_ref()
            .and_then(|m| m.finish),
        Some(MessageFinish::Length)
    );
}

/// A reply the provider's filter stopped keeps its fragment, is announced as
/// filtered and never as continuable — resuming meets the same filter — and is
/// stored as a stop, so `/continue` refuses it the way it refuses any finished
/// reply (docs/research/content-filter-finish.md, fork C2). The mock supervisor runs
/// as managed, where a length cut *would* be continuable.
#[tokio::test]
async fn a_filtered_reply_is_announced_and_stored_as_a_stop() {
    use crate::entities::message::MessageFinish;

    let (dir, cmd_tx, _evt_rx, handle, finished) = interrupted_turn(vec![vec![
        ChatChunk::Text("Half of it".into()),
        ChatChunk::Finished(FinishReason::Filtered),
    ]])
    .await;
    assert!(
        matches!(
            finished,
            AppEvent::Finished {
                reason: FinishReason::Filtered,
                continuable: false,
                ..
            }
        ),
        "a filtered reply must not offer /continue: {finished:?}"
    );
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    let chat = reload_chat(dir.path());
    assert_eq!(chat.messages[1].text, "Half of it");
    assert_eq!(
        chat.messages[1].metadata.as_ref().and_then(|m| m.finish),
        Some(MessageFinish::Stop)
    );
}

/// A turn interrupted between tool rounds leaves a tool-result tail;
/// `/continue` then resumes the agentic loop as an ordinary round — a **new**
/// assistant message, no prefill, no bubble re-opening.
#[tokio::test]
async fn continue_resumes_the_loop_on_a_tool_result_tail() {
    use crate::shared::api::contract::ToolCallDelta;

    let (dir, cmd_tx, mut evt_rx, handle, _) = interrupted_turn(vec![
        vec![
            ChatChunk::ToolCall(ToolCallDelta {
                thought_signature: None,
                index: 0,
                id: Some("c1".into()),
                name: Some("note_save".into()),
                arguments: "{\"content\":\"факт\"}".into(),
            }),
            ChatChunk::Finished(FinishReason::ToolCalls),
        ],
        // The round after the tool: nothing arrives — the turn ends as
        // cancelled, leaving the tool message as the chat's tail.
        vec![ChatChunk::Finished(FinishReason::Cancelled)],
        // The resumed round.
        vec![
            ChatChunk::Text("Записал.".into()),
            ChatChunk::Finished(FinishReason::Stop),
        ],
    ])
    .await;

    cmd_tx.send(AppCommand::ContinueLast).unwrap();
    let started = wait_for(&mut evt_rx, |e| {
        matches!(e, AppEvent::GenerationStarted { .. })
    })
    .await
    .unwrap();
    assert!(
        matches!(
            started,
            AppEvent::GenerationStarted {
                continuation: false,
                ..
            }
        ),
        "a loop resume opens a bubble of its own: {started:?}"
    );
    wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
        .await
        .unwrap();
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();

    let chat = reload_chat(dir.path());
    let last = chat.messages.last().unwrap();
    assert_eq!(last.role, MessageRole::Assistant);
    assert_eq!(last.text, "Записал.", "{:?}", chat.messages);
}

/// Every refusal answers with the route that works (docs/lessons.md §4), and
/// each is chosen by the tail's actual state. Driven on a bare orchestrator:
/// the refusals need no turn, only the answer.
#[tokio::test]
async fn continue_refusals_answer_with_the_route_that_works() {
    use crate::entities::chat::Chat;
    use crate::entities::message::{Message, MessageFinish, MessageMetadata};
    use crate::entities::profile::Profile;
    use crate::shared::config::ServerMode;

    let loc = crate::shared::i18n::locale(crate::shared::i18n::Lang::default());
    let (_d, mut orch, mut rx) = bare_orch_rx();
    orch.engines.backend = Some(Arc::new(MockBackend::scripted(vec![])) as Arc<dyn EngineBackend>);
    orch.engines.server_status = ServerStatus::Ready;
    let mut chat = Chat::from_profile(&Profile::new("P", "sys"), "Чат");
    let chat_id = chat.id;
    chat.push_message(Message::user("вопрос"));
    orch.chats.push(chat);
    orch.active_id = Some(chat_id);

    let mut expect_note = |orch: &mut super::super::Orchestrator, key: &str| {
        orch.handle_continue();
        let mut note = None;
        while let Ok(ev) = rx.try_recv() {
            if let AppEvent::Error(text) = ev {
                note = Some(text);
            }
        }
        assert_eq!(note.as_deref(), Some(loc.t(key)), "expected {key}");
    };

    // A user tail: nothing to continue.
    expect_note(&mut orch, "ui.cmd.continue_nothing");

    // An assistant tail that finished on its own.
    let mut done = Message::assistant("Готовый ответ.");
    done.metadata = Some(MessageMetadata {
        sampling: Default::default(),
        mode: Default::default(),
        model: None,
        provider: None,
        cost_nanos: None,
        finish: Some(MessageFinish::Stop),
    });
    orch.chat_mut(chat_id).unwrap().push_message(done);
    expect_note(&mut orch, "ui.cmd.continue_complete");

    // Cut inside the reasoning, before any visible text (fork F4).
    {
        let chat = orch.chat_mut(chat_id).unwrap();
        let last = chat.messages.last_mut().unwrap();
        last.text = String::new();
        last.thoughts = Some("обрывок мысли".into());
    }
    expect_note(&mut orch, "ui.cmd.continue_thoughts");

    // A cloud mode: the capability gate answers before anything else.
    {
        let chat = orch.chat_mut(chat_id).unwrap();
        let last = chat.messages.last_mut().unwrap();
        last.text = "Оборванный отв".into();
        last.thoughts = None;
        if let Some(md) = last.metadata.as_mut() {
            md.finish = Some(MessageFinish::Cancelled);
        }
    }
    orch.config.engine.mode = ServerMode::OpenAi;
    expect_note(&mut orch, "ui.cmd.continue_unsupported");

    // Anthropic is per-model (stage 2): a 4.6+ id refuses — the wider matrix
    // is pinned in `shared::config`'s own capability test.
    orch.config.engine.mode = ServerMode::Claude;
    orch.config.engine.claude.model_name = Some("claude-opus-4-8".into());
    expect_note(&mut orch, "ui.cmd.continue_unsupported");

    // Back on a supporting mode with a continuable tail — and a message stored
    // before the bookkeeping existed (`finish: None`) is continuable too
    // (fork F1): the turn actually starts.
    orch.config.engine.mode = ServerMode::Managed;
    if let Some(md) = orch
        .chat_mut(chat_id)
        .unwrap()
        .messages
        .last_mut()
        .unwrap()
        .metadata
        .as_mut()
    {
        md.finish = None;
    }
    orch.handle_continue();
    let mut started = None;
    while let Ok(ev) = rx.try_recv() {
        if let AppEvent::GenerationStarted { continuation, .. } = ev {
            started = Some(continuation);
        }
    }
    assert_eq!(
        started,
        Some(true),
        "a legacy partial without the end-state field must still continue"
    );
}

/// Through a gateway `/continue` follows the route table measured in
/// docs/history/gateway-images-and-continue.md §1.3 (fork H2): a refusing slug answers with
/// the gateway's own note — not the generic one, which says external engines
/// continue — while the same slug on an endpoint that published no catalogue keeps
/// the behaviour that shipped, and a family the table allows starts the turn.
#[tokio::test]
async fn continue_through_a_gateway_follows_the_route_table() {
    use crate::app::orchestrator::compaction::EngineFacts;
    use crate::entities::chat::Chat;
    use crate::entities::message::{Message, MessageFinish, MessageMetadata};
    use crate::entities::profile::Profile;
    use crate::shared::api::contract::ModelCapabilities;
    use crate::shared::config::ServerMode;

    let loc = crate::shared::i18n::locale(crate::shared::i18n::Lang::default());
    let (_d, mut orch, mut rx) = bare_orch_rx();
    orch.engines.backend = Some(Arc::new(MockBackend::scripted(vec![])) as Arc<dyn EngineBackend>);
    orch.engines.server_status = ServerStatus::Ready;
    orch.config.engine.mode = ServerMode::External;
    orch.config.engine.external.model_name = Some("google/gemma-4-31b-it".into());
    let mut chat = Chat::from_profile(&Profile::new("P", "sys"), "Чат");
    let chat_id = chat.id;
    chat.push_message(Message::user("вопрос"));
    let mut partial = Message::assistant("Оборванный отв");
    partial.metadata = Some(MessageMetadata {
        sampling: Default::default(),
        mode: Default::default(),
        model: None,
        provider: None,
        cost_nanos: None,
        finish: Some(MessageFinish::Cancelled),
    });
    chat.push_message(partial);
    orch.chats.push(chat);
    orch.active_id = Some(chat_id);
    let land_catalogue = |orch: &mut super::super::Orchestrator| {
        let epoch = orch.context.epoch();
        orch.handle_budget_result(
            epoch,
            EngineFacts {
                budget: None,
                caps: Some(ModelCapabilities {
                    context_length: Some(262_144),
                    sampling_fields: None,
                }),
                server: Default::default(),
            },
        );
    };
    let drain = |rx: &mut UnboundedReceiver<AppEvent>| {
        let (mut note, mut started) = (None, None);
        while let Ok(ev) = rx.try_recv() {
            match ev {
                AppEvent::Error(text) => note = Some(text),
                AppEvent::GenerationStarted { continuation, .. } => started = Some(continuation),
                _ => {}
            }
        }
        (note, started)
    };

    // A gateway (the catalogue answered) and a slug that restarts on every route.
    land_catalogue(&mut orch);
    orch.handle_continue();
    assert_eq!(
        drain(&mut rx),
        (
            Some(loc.t("ui.cmd.continue_unsupported_gateway").to_string()),
            None
        ),
        "a restarting family is refused through a gateway, with the gateway's note"
    );

    // The same slug on an endpoint that published nothing: silence is not a claim.
    orch.context.invalidate();
    assert!(
        orch.continuation_supported(),
        "without a catalogue external answers as it always did"
    );

    // A family the table allows continues through the same gateway.
    land_catalogue(&mut orch);
    orch.config.engine.external.model_name = Some("anthropic/claude-haiku-4.5".into());
    orch.handle_continue();
    assert_eq!(
        drain(&mut rx),
        (None, Some(true)),
        "Claude up to 4.5 continues through a gateway"
    );
}

/// The gateway's own mode follows the same route table **without waiting for the
/// catalogue**: `external` needs the answer to know it is speaking to a gateway,
/// and this mode is one by name (docs/research/openrouter-mode.md §2.3). So the
/// first command after a start is already answered by the slug — and refused
/// with the gateway's note, not the one that says external engines continue.
#[tokio::test]
async fn continue_in_the_gateways_own_mode_needs_no_catalogue() {
    use crate::entities::chat::Chat;
    use crate::entities::message::{Message, MessageFinish, MessageMetadata};
    use crate::entities::profile::Profile;
    use crate::shared::config::ServerMode;

    let loc = crate::shared::i18n::locale(crate::shared::i18n::Lang::default());
    let (_d, mut orch, mut rx) = bare_orch_rx();
    orch.engines.backend = Some(Arc::new(MockBackend::scripted(vec![])) as Arc<dyn EngineBackend>);
    orch.engines.server_status = ServerStatus::Ready;
    orch.config.engine.mode = ServerMode::OpenRouter;
    orch.config.engine.openrouter.model_name = Some("qwen/qwen3.6-27b".into());
    let mut chat = Chat::from_profile(&Profile::new("P", "sys"), "c");
    let chat_id = chat.id;
    chat.push_message(Message::user("a question"));
    let mut partial = Message::assistant("A reply cut sh");
    partial.metadata = Some(MessageMetadata {
        sampling: Default::default(),
        mode: ServerMode::OpenRouter,
        model: None,
        provider: None,
        cost_nanos: None,
        finish: Some(MessageFinish::Cancelled),
    });
    chat.push_message(partial);
    orch.chats.push(chat);
    orch.active_id = Some(chat_id);
    let drain = |rx: &mut UnboundedReceiver<AppEvent>| {
        let (mut note, mut started) = (None, None);
        while let Ok(ev) = rx.try_recv() {
            match ev {
                AppEvent::Error(text) => note = Some(text),
                AppEvent::GenerationStarted { continuation, .. } => started = Some(continuation),
                _ => {}
            }
        }
        (note, started)
    };

    assert!(!orch.endpoint_catalogued(), "nothing has answered yet");
    orch.handle_continue();
    assert_eq!(
        drain(&mut rx),
        (
            Some(loc.t("ui.cmd.continue_unsupported_gateway").to_string()),
            None
        ),
        "an open-weight route restarts, and the mode knows before the catalogue says so"
    );

    orch.config.engine.openrouter.model_name = Some("anthropic/claude-haiku-4.5".into());
    orch.handle_continue();
    assert_eq!(
        drain(&mut rx),
        (None, Some(true)),
        "Claude up to 4.5 continues"
    );
}

/// A backend that publishes a catalogue entry for its model, the way a gateway does.
struct CataloguedBackend {
    inner: MockBackend,
}

#[async_trait::async_trait]
impl EngineBackend for CataloguedBackend {
    async fn chat_stream(
        &self,
        req: crate::shared::api::ChatRequest,
        cancel: tokio_util::sync::CancellationToken,
    ) -> anyhow::Result<crate::shared::api::contract::ChatStream> {
        self.inner.chat_stream(req, cancel).await
    }

    async fn model_capabilities(&self) -> Option<crate::shared::api::contract::ModelCapabilities> {
        Some(crate::shared::api::contract::ModelCapabilities {
            context_length: Some(262_144),
            sampling_fields: Some(vec!["temperature".to_string()].into()),
        })
    }
}

/// The whole route on a running orchestrator: the gateway's catalogue lands
/// **before any turn** — `/continue` as the first command after a restart is the
/// command's main case, and it must not meet an unanswered question — and then
/// it governs all three readers: a length-cut reply is announced as *not*
/// continuable, and the command refuses with the gateway's note
/// (docs/history/gateway-images-and-continue.md §4, H2).
#[tokio::test]
async fn a_gateways_catalogue_lands_before_the_first_turn_and_governs_continue() {
    use crate::shared::config::ServerMode;
    use std::time::Duration;

    let loc = crate::shared::i18n::locale(crate::shared::i18n::Lang::default());
    let backend = Arc::new(CataloguedBackend {
        inner: MockBackend::sequence(vec![vec![
            ChatChunk::Text("Нача".into()),
            ChatChunk::Finished(FinishReason::Length),
        ]]),
    }) as Arc<dyn EngineBackend>;
    let mut cfg = no_auto_cfg();
    cfg.engine.mode = ServerMode::External;
    cfg.engine.external.model_name = Some("google/gemma-4-31b-it".into());
    let (_dir, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), cfg);

    let landed = tokio::time::timeout(
        Duration::from_secs(5),
        wait_for(&mut evt_rx, |e| {
            matches!(e, AppEvent::EngineSamplingFields(Some(_)))
        }),
    )
    .await;
    assert!(
        matches!(landed, Ok(Some(_))),
        "the catalogue must be asked when the engine is applied, not at the first turn"
    );

    cmd_tx
        .send(AppCommand::SendMessage("вопрос".into()))
        .unwrap();
    let finished = tokio::time::timeout(
        Duration::from_secs(5),
        wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. })),
    )
    .await
    .expect("the turn finishes")
    .unwrap();
    assert!(
        matches!(
            finished,
            AppEvent::Finished {
                reason: FinishReason::Length,
                continuable: false,
                ..
            }
        ),
        "a gateway's restarting family must not be announced as continuable: {finished:?}"
    );

    cmd_tx.send(AppCommand::ContinueLast).unwrap();
    let note = tokio::time::timeout(
        Duration::from_secs(5),
        wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Error(_))),
    )
    .await
    .expect("the command answers");
    assert!(
        matches!(&note, Some(AppEvent::Error(text)) if text == loc.t("ui.cmd.continue_unsupported_gateway")),
        "{note:?}"
    );
    cmd_tx.send(AppCommand::Quit).unwrap();
    handle.await.unwrap();
}

// ---------- the slow-prefill note (docs/research/slow-prefill-detection.md) ----------

mod slow_prefill {
    use super::*;
    use crate::shared::api::contract::{Prefill, TokenUsage};
    use crate::shared::api::managed::LLAMA_DEFAULT_BATCH;
    use crate::shared::config::ServerMode;
    use crate::shared::server::ServerStatus;

    /// The CPU build at the default batch: ~90 tok/s over a real prompt.
    const SLOW: Prefill = Prefill {
        tokens: 1800,
        ms: 20_000,
    };

    /// A managed server at llama.cpp's default batch, typed rather than left
    /// to Auto: on a Mac an empty field is the knee (`managed::auto_batch`,
    /// docs/research/macos.md §14.8), and a server at the knee is told
    /// nothing — these tests would wait for a note that never comes.
    fn default_batch_cfg() -> AppConfig {
        let mut config = no_auto_cfg();
        config.engine.managed.batch_size = Some(LLAMA_DEFAULT_BATCH);
        config
    }

    fn notices(rx: &mut UnboundedReceiver<AppEvent>) -> Vec<String> {
        let mut out = Vec::new();
        while let Ok(e) = rx.try_recv() {
            if let AppEvent::Notice(text) = e {
                out.push(text);
            }
        }
        out
    }

    /// A managed server at the default batch: one note per server session,
    /// naming the field, the hold and the batch; the next turn says nothing;
    /// a server that reaches `Ready` again — a relaunch — may be told once
    /// more (§3.3).
    #[test]
    fn a_managed_server_is_told_once_per_session() {
        let (_d, mut orch, mut rx) = bare_orch_rx();
        orch.config.engine.mode = ServerMode::Managed;
        orch.config.engine.managed.gpu_layers = 99;
        // The server's default, typed: an empty field is the knee on a Mac
        // (`managed::auto_batch`), where nothing would be said.
        orch.config.engine.managed.batch_size = Some(LLAMA_DEFAULT_BATCH);

        orch.note_slow_prefill(Some(SLOW));
        let notes = notices(&mut rx);
        assert_eq!(notes.len(), 1, "{notes:?}");
        assert!(
            notes[0].contains("Batch (-b)") && notes[0].contains("23") && notes[0].contains("2048"),
            "{}",
            notes[0]
        );

        orch.note_slow_prefill(Some(SLOW));
        assert!(notices(&mut rx).is_empty(), "once per session");

        orch.engines.set_chat_status(ServerStatus::Ready);
        orch.note_slow_prefill(Some(SLOW));
        assert_eq!(notices(&mut rx).len(), 1, "a new session, told again");
    }

    /// Nothing to say: the batch already at the knee (typed, or the CPU auto),
    /// a fast prefill, a cloud, no sample.
    #[test]
    fn nothing_is_said_where_the_rule_does_not_hold() {
        let (_d, mut orch, mut rx) = bare_orch_rx();
        orch.config.engine.mode = ServerMode::Managed;
        orch.config.engine.managed.gpu_layers = 99;
        orch.config.engine.managed.batch_size = Some(256);
        orch.note_slow_prefill(Some(SLOW));
        assert!(notices(&mut rx).is_empty(), "typed 256: the change is made");

        orch.config.engine.managed.batch_size = None;
        orch.config.engine.managed.gpu_layers = 0;
        orch.note_slow_prefill(Some(SLOW));
        assert!(notices(&mut rx).is_empty(), "the CPU auto: 256 already");

        orch.config.engine.managed.gpu_layers = 99;
        orch.note_slow_prefill(Some(Prefill {
            tokens: 1800,
            ms: 900,
        }));
        assert!(notices(&mut rx).is_empty(), "a GPU's second");

        orch.config.engine.mode = ServerMode::OpenAi;
        orch.note_slow_prefill(Some(SLOW));
        assert!(notices(&mut rx).is_empty(), "a cloud has no batch");

        orch.config.engine.mode = ServerMode::Managed;
        orch.note_slow_prefill(None);
        assert!(notices(&mut rx).is_empty(), "no sample");
    }

    /// An external server: the launch line's wording, at the default batch the
    /// app assumes (fork F3).
    #[test]
    fn an_external_server_is_told_the_launch_line() {
        let (_d, mut orch, mut rx) = bare_orch_rx();
        orch.config.engine.mode = ServerMode::External;
        orch.slots.pretend_known(Some(4));
        orch.note_slow_prefill(Some(SLOW));
        let notes = notices(&mut rx);
        assert_eq!(notes.len(), 1, "{notes:?}");
        assert!(
            notes[0].contains("-b 256 -ub 256") && notes[0].contains("2048"),
            "{}",
            notes[0]
        );
    }

    /// An external server that has not answered llama.cpp's `/props` is not
    /// told llama-server's flags: Ollama sends the same `timings`, its first
    /// request's with the model's load in them, and takes no `-b`. The session's
    /// one note is not spent on it — the same server, once it answers, is told.
    #[test]
    fn an_external_server_that_is_not_a_llama_server_is_told_nothing() {
        let (_d, mut orch, mut rx) = bare_orch_rx();
        orch.config.engine.mode = ServerMode::External;
        orch.slots.pretend_known(None);
        orch.note_slow_prefill(Some(SLOW));
        assert!(notices(&mut rx).is_empty(), "no /props, no launch line");

        orch.slots.pretend_known(Some(1));
        orch.note_slow_prefill(Some(SLOW));
        assert_eq!(notices(&mut rx).len(), 1, "a llama-server, told");
    }

    /// A tool's own request is a stream of the turn (spec §9.3.1): the
    /// timing it reports on its outcome folds into the turn's largest sample
    /// and reaches the note like a round's — the page summary's path without
    /// the page (docs/research/page-summary-usage.md §3.2). The engine's own
    /// rounds carry a warm sample under the floor; the tool's is the cold one.
    #[tokio::test]
    async fn a_tools_own_request_is_the_turns_sample_too() {
        use crate::app::orchestrator::tests::subagent::{Script, ScriptRecorder};
        let warm = ChatChunk::Usage(TokenUsage {
            prompt_tokens: 40,
            completion_tokens: 1,
            reasoning_tokens: 0,
            prefill: Some(Prefill { tokens: 40, ms: 20 }),
        });
        let call = Script {
            chunks: vec![
                ChatChunk::ToolCall(crate::shared::api::contract::ToolCallDelta {
                    thought_signature: None,
                    index: 0,
                    id: Some("s1".into()),
                    name: Some("sampled".into()),
                    arguments: "{}".into(),
                }),
                warm.clone(),
                ChatChunk::Finished(FinishReason::ToolCalls),
            ],
            hang: false,
        };
        let reply = Script {
            chunks: vec![
                ChatChunk::Text("done".into()),
                warm,
                ChatChunk::Finished(FinishReason::Stop),
            ],
            hang: false,
        };
        let backend = ScriptRecorder::new(vec![call, reply]);
        // The default managed server: its note needs no answer from `/props`,
        // which a scripted backend never gives.
        let config = default_batch_cfg();
        let tool = Arc::new(crate::app::orchestrator::tests::SampledTool {
            id: "sampled",
            sample: Some(SLOW),
        });
        let (_d, cmd_tx, mut evt_rx, handle) =
            spawn_orch_tools(Some(backend as Arc<dyn EngineBackend>), config, vec![tool]);
        // The startup emits the profile list and the active chat in whichever
        // order; both are needed before the tool can be enabled and the
        // message sent.
        let (mut pid, mut chat) = (None, None);
        while pid.is_none() || chat.is_none() {
            match tokio::time::timeout(std::time::Duration::from_secs(5), evt_rx.recv())
                .await
                .expect("startup events")
            {
                Some(AppEvent::ProfileList(v)) if !v.is_empty() => pid = Some(v[0].id),
                Some(AppEvent::ChatActivated { id, .. }) => chat = Some(id),
                Some(_) => {}
                None => panic!("the orchestrator went away during startup"),
            }
        }
        cmd_tx
            .send(AppCommand::UpdateProfile {
                id: pid.unwrap(),
                edit: Box::new(ProfileEdit {
                    enabled_tools: Some(vec!["sampled".into()]),
                    ..Default::default()
                }),
            })
            .unwrap();
        cmd_tx.send(AppCommand::SendMessage("go".into())).unwrap();
        let note = wait_for(
            &mut evt_rx,
            |e| matches!(e, AppEvent::Notice(t) if t.contains("Batch (-b)")),
        )
        .await;
        assert!(note.is_some(), "the tool's cold sample reached the note");
        cmd_tx.send(AppCommand::Quit).unwrap();
        handle.await.unwrap();
    }

    /// The whole path: the engine's `timings` on the stream's usage chunk
    /// travel through the round, the turn and the landing to the note
    /// (§3.1) — and the second turn on the same server says nothing.
    #[tokio::test]
    async fn a_turn_carries_the_engines_figure_to_the_note() {
        let backend = Arc::new(MockBackend::scripted(vec![
            ChatChunk::Text("hi".into()),
            ChatChunk::Usage(TokenUsage {
                prompt_tokens: 1850,
                completion_tokens: 1,
                reasoning_tokens: 0,
                prefill: Some(SLOW),
            }),
            ChatChunk::Finished(FinishReason::Stop),
        ])) as Arc<dyn EngineBackend>;
        // The default managed server, for the reason the tool's path above gives.
        let config = default_batch_cfg();
        let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), config);
        wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
            .await
            .unwrap();

        cmd_tx.send(AppCommand::SendMessage("one".into())).unwrap();
        let note = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Notice(_)))
            .await
            .unwrap();
        let AppEvent::Notice(text) = note else {
            unreachable!()
        };
        assert!(text.contains("Batch (-b)"), "{text}");

        cmd_tx.send(AppCommand::SendMessage("two".into())).unwrap();
        wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
            .await
            .unwrap();
        // The landing's note, had there been one, follows `Finished` at once;
        // a quit flushes the channel in order.
        cmd_tx.send(AppCommand::Quit).unwrap();
        handle.await.unwrap();
        let mut second = Vec::new();
        while let Ok(e) = evt_rx.try_recv() {
            if let AppEvent::Notice(t) = e {
                second.push(t);
            }
        }
        assert!(second.is_empty(), "the same server, told once: {second:?}");
    }
}

/// The estimate counts the tool schemas as the wire sends them
/// (docs/research/roll-usage-calibration.md §3.1): one part at the text's
/// bytes-per-token, plus the per-message overhead once for the block; a
/// request without tools estimates as before.
#[test]
fn the_estimate_counts_the_tool_schemas() {
    use crate::shared::api::contract::ToolSchema;
    let bare = crate::shared::api::ChatRequest {
        continue_final: false,
        system: Some("sys".into()),
        messages: vec![crate::shared::api::ApiMessage::user("hi")],
        sampling: Default::default(),
        tools: Vec::new(),
    };
    let schema = ToolSchema {
        name: "note_save".into(),
        description: "Saves a note.".into(),
        parameters: serde_json::json!({
            "type": "object",
            "properties": { "text": { "type": "string" } }
        }),
    };
    let with = crate::shared::api::ChatRequest {
        tools: vec![schema],
        ..bare.clone()
    };
    let json = crate::shared::api::openai::tools_json(&with.tools);
    assert!(
        json.starts_with(r#"[{"type":"function","function":{"name":"note_save""#),
        "the wire's own wrapper: {json}"
    );
    let bare_estimate = super::super::generation::estimate_prompt_tokens(&bare);
    let with_estimate = super::super::generation::estimate_prompt_tokens(&with);
    assert_eq!(
        with_estimate - bare_estimate,
        (json.len() as u64).div_ceil(4) + 4,
        "the block's bytes over four, plus one message overhead"
    );
}

/// Fork F11 of docs/research/openrouter-mode.md, end to end: what the gateway
/// said about a request — who served it, and for how much — is on the message
/// that request produced, and a reply from an engine that says neither carries
/// neither (the file stays as it was before the fields existed).
#[tokio::test]
async fn who_served_a_reply_and_what_it_cost_are_recorded_on_the_message() {
    use crate::shared::api::contract::Served;
    use crate::shared::config::{CloudSettings, EngineSettings, ServerMode};

    async fn reply_through(chunks: Vec<ChatChunk>) -> (Message, String) {
        let config = AppConfig {
            engine: EngineSettings {
                mode: ServerMode::OpenRouter,
                openrouter: CloudSettings {
                    model_name: Some("anthropic/claude-haiku-4.5".into()),
                    ..Default::default()
                },
                ..Default::default()
            },
            ..Default::default()
        };
        let backend = Arc::new(MockBackend::scripted(chunks)) as Arc<dyn EngineBackend>;
        let (_d, cmd_tx, mut evt_rx, handle) = spawn_orch_cfg(Some(backend), config);
        let root = _d.path().to_path_buf();
        let active = wait_for(&mut evt_rx, |e| matches!(e, AppEvent::ChatActivated { .. }))
            .await
            .unwrap();
        let chat_id = match active {
            AppEvent::ChatActivated { id, .. } => id,
            _ => unreachable!(),
        };
        cmd_tx.send(AppCommand::SendMessage("hi".into())).unwrap();
        wait_for(&mut evt_rx, |e| matches!(e, AppEvent::Finished { .. }))
            .await
            .unwrap();
        cmd_tx.send(AppCommand::Quit).unwrap();
        handle.await.unwrap();

        let reopened = Storage::open(Paths::with_root(&root)).unwrap();
        let chat = reopened.json().load_chat(chat_id).unwrap().unwrap();
        // Read as it is on disk, and a missing file fails here: an empty string
        // would satisfy "contains neither key" without a reply ever being saved.
        let file = std::fs::read_to_string(Paths::with_root(&root).chat_file(&chat_id.to_string()))
            .expect("the chat file");
        assert!(file.contains("Paris"), "the reply is in the file: {file}");
        (chat.messages[1].clone(), file)
    }

    let (said, _) = reply_through(vec![
        ChatChunk::Text("Paris".into()),
        ChatChunk::Served(Served {
            provider: Some("Amazon Bedrock".into()),
            cost_nanos: Some(37_000),
        }),
        ChatChunk::Finished(FinishReason::Stop),
    ])
    .await;
    let meta = said.metadata.expect("a metadata snapshot");
    assert_eq!(meta.mode, ServerMode::OpenRouter);
    assert_eq!(meta.model.as_deref(), Some("anthropic/claude-haiku-4.5"));
    assert_eq!(meta.provider.as_deref(), Some("Amazon Bedrock"));
    assert_eq!(meta.cost_nanos, Some(37_000));

    let (silent, file) = reply_through(vec![
        ChatChunk::Text("Paris".into()),
        ChatChunk::Finished(FinishReason::Stop),
    ])
    .await;
    let meta = silent.metadata.expect("a metadata snapshot");
    assert_eq!((meta.provider, meta.cost_nanos), (None, None));
    assert!(
        !file.contains("cost_nanos") && !file.contains("\"provider\""),
        "absent stays absent on disk: {file}"
    );
}

/// The tool message a request carries last — what the model reads of its call.
fn last_tool_result(req: &crate::shared::api::ChatRequest) -> String {
    req.messages.last().unwrap().content.clone()
}

/// A call whose arguments are not JSON is answered as broken, with serde's reason, and
/// nothing runs: read as `{}` it reached the tool, which reported a missing field, and
/// the model sent the same call again (docs/journal/engine.md, "a broken call is
/// answered as broken"). An empty string is still a call with no arguments.
#[tokio::test]
async fn a_call_with_broken_arguments_is_answered_as_broken_and_not_run() {
    use super::subagent::{call, text};
    let broken = r#"{"field": "x""#;
    let reason = serde_json::from_str::<serde_json::Value>(broken)
        .unwrap_err()
        .to_string();
    let (reqs, _) = turn_on_recorder(vec![
        call("c1", "get_sampling", broken),
        call("c2", "get_sampling", ""),
        text("done"),
    ])
    .await;
    let refused = last_tool_result(&reqs[1]);
    assert!(
        refused.contains("get_sampling") && refused.contains(&reason),
        "{refused}"
    );
    let ran = last_tool_result(&reqs[2]);
    assert!(!ran.contains(&reason) && !ran.contains("JSON"), "{ran}");
}

/// The sub-agent's path is refused the same way, before a child is prepared: the
/// parent's two rounds are the only requests the engine sees.
#[tokio::test]
async fn a_sub_agent_call_with_broken_arguments_starts_no_child() {
    use super::subagent::{call, text};
    let (reqs, last) = turn_on_recorder(vec![
        call(
            "c1",
            "call_subagent",
            r#"{"system_message": "x", "message": "#,
        ),
        text("done"),
    ])
    .await;
    assert_eq!(reqs.len(), 2, "no child request");
    assert!(last_tool_result(&reqs[1]).contains("JSON"));
    assert_eq!(last, "done");
}