mindfork 0.11.0

//! Demo fixture — the showcase content behind the generated screenshots and
//! the interactive `mindfork demo` mode (docs/history/demo-screenshots.md).
//!
//! The content is honest fabrication: a plausible conversation that walks the
//! renderer through its range — a GFM table, a highlighted code block, a
//! Mermaid flowchart, delimiter-scoped LaTeX, an expanded "thoughts" block and
//! a tool call with its result. Ids and timestamps are fixed so a captured
//! frame is byte-stable across runs (the drift gate depends on it) and so
//! [`provision`] is idempotent.

use chrono::{DateTime, TimeZone, Utc};
use uuid::Uuid;

use crate::entities::chat::{Chat, FeedView};
#[cfg(test)]
use crate::entities::chat::{ChatSummary, ChildSummary};
use crate::entities::message::{Message, MessageRole, ToolCallRecord};
use crate::entities::profile::Profile;
use crate::entities::self_model::{Goal, GoalStatus, NarrativeSegment, SelfModel, UserModel};
use crate::entities::subagent::{RunKind, RunOutcome, SubagentRun};
use crate::features::tools::default_tool_ids;
use crate::shared::api::contract::{ChatChunk, FinishReason, TokenUsage};
#[cfg(test)]
use crate::shared::config::Theme;
use crate::shared::config::{AppConfig, ServerMode};
use crate::shared::i18n::Lang;
use crate::shared::storage::Storage;

/// Title of the showcase chat (shown in the feed header).
pub const CHAT_TITLE: &str = "Gemma 4 on a 12 GB GPU";

/// The unsent draft in the input box: the hero must show a *live* seat — an
/// empty prompt reads as a screensaver. The question also sets up the next
/// beat of the demo world: the settings capture shows exactly this draft
/// model configured, and the chat list holds the speculative-decoding chat.
pub const INPUT_DRAFT: &str = "And speculative decoding — is a 1B draft worth it on this card?";

/// Fixed chat id — captures must not depend on a fresh `Uuid::new_v4()`.
pub fn chat_id() -> Uuid {
    Uuid::from_u128(0x6d69_6e64_666f_726b_5f64_656d_6f5f_3031)
}

/// The feed view for captures: thoughts and tool-call details expanded —
/// collapsed pills demonstrate nothing.
pub fn feed_view() -> FeedView {
    FeedView {
        thoughts: true,
        tools: true,
    }
}

/// Fixed timestamp base; messages step forward a minute at a time.
fn at(minute: u32) -> DateTime<Utc> {
    Utc.with_ymd_and_hms(2026, 8, 1, 10, minute, 0).unwrap()
}

/// Fixed calendar date for the list/self-model fixtures.
fn date(month: u32, day: u32, hour: u32, minute: u32) -> DateTime<Utc> {
    Utc.with_ymd_and_hms(2026, month, day, hour, minute, 0)
        .unwrap()
}

fn message(role: MessageRole, minute: u32, text: &str) -> Message {
    let mut m = Message::new(role, text);
    // Fixed id derived from the slot, fixed time: determinism over realism.
    m.id = Uuid::from_u128(0x6d66_5f64_656d_6f00 + minute as u128);
    m.timestamp = at(minute);
    m
}

/// The showcase conversation. Composed so that the *tail* — what a capture at
/// the stage-1 frame size actually shows — contains the table's neighbourhood,
/// the flowchart, the LaTeX line, the expanded thoughts and the tool card,
/// while the opening exchange scrolls off the top (a conversation visibly has
/// history). `app/demo_shots.rs` asserts the load-bearing pieces are on
/// screen, so edits here cannot silently push them out of frame.
pub fn showcase_messages() -> Vec<Message> {
    let m1 = message(
        MessageRole::User,
        0,
        "Which Gemma 4 12B quant fits a 12 GB card with room for a 16k context?",
    );

    let m2 = message(
        MessageRole::Assistant,
        1,
        "Three candidates, weights only — the KV cache comes on top:\n\n\
         | Quant | Weights | Quality cost |\n\
         |---|---|---|\n\
         | Q4_K_M | ~7.3 GB | small but measurable |\n\
         | Q5_K_M | ~8.4 GB | negligible |\n\
         | Q6_K | ~9.7 GB | noise level |\n\n\
         At 16k context the KV cache adds roughly 1.5 GB, so **Q5_K_M** is the \
         sweet spot — quality headroom *and* a safety margin:\n\n\
         ```bash\n\
         llama-server -m gemma-4-12B-it-Q5_K_M.gguf -ngl 99 -c 16384 --jinja\n\
         ```",
    );

    let m3 = message(
        MessageRole::User,
        2,
        "Give me the bottom line — a small table plus a decision diagram, and \
         save a note about my setup.",
    );

    // The decision tree renders with three perfectly straight legs only for
    // these exact label widths (mermaid-text centers nodes by content) —
    // rewording a label can bend the middle edge into a `┌─┘` jog. Eyeball
    // the regenerated dump after any edit here.
    let mut m4 = message(
        MessageRole::Assistant,
        3,
        "The short version:\n\n\
         | Quant | 16k context? | Verdict |\n\
         |---|---|---|\n\
         | Q5_K_M | fits, ~1.5 GB spare | **sweet spot** |\n\
         | Q6_K | tight | quality first |\n\n\
         ```mermaid\n\
         flowchart TD\n\
             A{How much context?} -->|8k| B[Q6_K — quality first]\n\
             A -->|16k| C[Q5_K_M — sweet spot]\n\
             A -->|32k| D[Q4_K_M — KV headroom]\n\
         ```\n\n\
         Note saved. The cache rule to remember: $M_{kv} \\propto L$ — double \
         the window, double the cache.",
    );
    m4.thoughts = Some(
        "The note should record the 12 GB budget and the Q5_K_M choice, so \
         future sizing questions start from the hardware."
            .into(),
    );
    m4.tool_calls = vec![ToolCallRecord {
        id: "call_demo_1".into(),
        name: "note_save".into(),
        arguments: serde_json::json!({
            "title": "Hardware budget",
            "text": "12 GB GPU; Gemma 4 12B at Q5_K_M, 16k context."
        }),
        result: Some("Note saved: \"Hardware budget\".".into()),
        thought_signature: None,
        images: 0,
        subagent: None,
    }];

    vec![m1, m2, m3, m4]
}

/// Fixed profile id for the demo companion.
pub fn profile_id() -> Uuid {
    Uuid::from_u128(0x6d66_5f64_656d_6f5f_7072_6f66_0000_0001)
}

/// The demo companion: a named profile with the base tools enabled (so the
/// settings screen's Tools section shows a real toggle set), a fixed id (the
/// seeded chats and self-model reference it) and a greeting that says plainly
/// what the demo is — a new chat's first message must close the door on "is
/// this a real model?".
pub fn profile() -> Profile {
    let mut p = Profile::new("Gaia", "You are Gaia — a thoughtful local companion.");
    p.id = profile_id();
    p.enabled_tools = default_tool_ids();
    p.greeting = Some(
        "Hi! You're in the **mindfork demo** — I'm a scripted engine, not a \
         model. Everything else is the real app: browse the chats (`Esc`), \
         peek at my self-model (`F3`), fold my thoughts (`Ctrl+T`), open \
         settings (`Ctrl+P`). Connect a real engine there — local llama.cpp \
         or a cloud key — and this seat gets a mind."
            .into(),
    );
    p
}

/// The app config the settings captures depict: a healthy managed llama.cpp
/// setup consistent with the showcase conversation (the same model and
/// context the hero chat recommends). Capture-only — the interactive demo
/// boots [`demo_config`] instead.
#[cfg(test)]
pub fn app_config(theme: Theme) -> AppConfig {
    let mut c = AppConfig::default();
    c.interface.theme = theme;
    c.interface.language = Lang::En;
    c.engine.managed.binary = Some("C:\\llama.cpp\\llama-server.exe".into());
    c.engine.managed.model_path = Some("C:\\models\\gemma-4-12B-it-Q5_K_M.gguf".into());
    c.engine.managed.gpu_layers = 99;
    c.engine.managed.context_size = 16384;
    // Speculative decoding with a 1B draft: the demo user acted on the
    // "Speculative decoding: draft models" chat, and the hero's input draft
    // asks about exactly this. Also what fills the Model/server capture —
    // draft-* types reveal four more parameter rows (settings::helpers).
    c.engine.managed.spec_type = crate::shared::config::SpecType::DraftSimple;
    c.engine.managed.draft_model = Some("C:\\models\\gemma-4-1B-it-Q8_0.gguf".into());
    c.engine.managed.draft_gpu_layers = Some(99);
    c.engine.managed.draft_n_max = Some(16);
    c
}

/// The filler chats behind both the list capture and the interactive demo's
/// chat list — enough of them to fill the uniform gallery frame (PANEL_H in
/// `app/demo_shots.rs`) to the last row. Sorted by `modified` descending, the
/// list screen's default order.
///
/// A *text* table (`created | modified | msgs | title`, parsed by [`rows`])
/// rather than rows of tuples: the duplication gate compares **normalized**
/// tokens, so any repeated per-chat structure — constructor blocks, even
/// one-line tuples at this row count — reads as a sliding self-duplicate no
/// matter how different the literals are. One flat literal is a single token;
/// there is nothing to match. Same reasoning for the other tables below.
const ROWS_RAW: &str = r#"
    07-31 17:05 | 07-31 18:22 | 14 | Why is prompt eval slow on long chats?
    07-30 20:11 | 07-31 09:14 | 13 | Grammar-constrained JSON output
    07-29 09:12 | 07-30 21:40 | 18 | Sampler settings for livelier replies
    07-29 08:00 | 07-29 19:05 | 12 | Speculative decoding: draft models
    07-28 12:40 | 07-28 14:03 | 10 | Embeddings: why bge-m3 for the notes
    07-26 14:30 | 07-27 17:52 | 41 | Refactoring a god object in Rust
    07-26 09:12 | 07-26 21:35 | 33 | A tokio deadlock, found and fixed
    07-25 11:03 | 07-25 12:44 |  9 | What does the DRY penalty actually do?
    07-24 10:02 | 07-24 11:47 | 16 | MCP servers worth installing
    07-23 19:01 | 07-23 19:58 |  8 | Terminal fonts and box-drawing glyphs
    07-20 18:15 | 07-22 20:31 | 26 | Trip notes: the Dolomites in October
    07-21 14:30 | 07-21 15:26 | 11 | Shrink the system prompt, keep the voice
    07-20 21:15 | 07-20 22:44 | 19 | RAG chunk size for code search
    07-17 07:45 | 07-19 23:10 | 15 | Reading list: attention papers
    07-18 16:20 | 07-18 16:58 |  7 | Backup dry run before the update
    07-17 08:50 | 07-17 12:19 | 23 | Regex or a real parser?
    07-15 13:00 | 07-16 10:27 | 11 | Mermaid diagrams in the terminal
    07-14 13:20 | 07-14 16:31 | 27 | Naming things: a short rant
    07-13 10:05 | 07-13 11:58 |  9 | Reading GGUF metadata in Python
    07-12 20:30 | 07-12 21:12 |  5 | First week with mindfork: impressions
"#;

/// Non-empty, trimmed content lines of a raw fixture table.
fn table_lines(raw: &'static str) -> impl Iterator<Item = &'static str> {
    raw.lines().map(str::trim).filter(|l| !l.is_empty())
}

/// Parses a fixture timestamp `MM-DD HH:MM` (the fixture's fixed year 2026).
/// Panics on a malformed table row — every caller sits under the ordinary
/// test suite, so a bad edit fails loudly at `cargo test`, not in a release.
fn ts(s: &str) -> DateTime<Utc> {
    let mut parts = s.split(['-', ' ', ':']).map(|p| p.parse::<u32>().unwrap());
    let mut next = || parts.next().expect("fixture timestamp is MM-DD HH:MM");
    date(next(), next(), next(), next())
}

/// One filler-chat row out of [`ROWS_RAW`]: created, modified, the message
/// count the *list capture* shows (the interactive demo's filler chats hold a
/// two-message excerpt, so their real count is 2), title.
fn rows() -> impl Iterator<Item = (DateTime<Utc>, DateTime<Utc>, usize, &'static str)> {
    table_lines(ROWS_RAW).map(|line| {
        let mut f = line.splitn(4, '|').map(str::trim);
        let mut field = || f.next().expect("filler row is created|modified|msgs|title");
        (ts(field()), ts(field()), field().parse().unwrap(), field())
    })
}

/// A two-message excerpt per filler chat, aligned with [`ROWS_RAW`] by index —
/// the interactive demo's chats open to a real (if brief) exchange, so the
/// list, search and export all have something true to show. Two lines per
/// chat — the user's question, then the assistant's answer; blank lines are
/// for the reader and are ignored by [`bodies`].
const ROW_BODIES_RAW: &str = r#"
Generation is fine, but the pause before the first token keeps growing. What gives?
That pause is prompt eval: after an edit or a compaction the whole history is re-scored before the first new token. Keep FlashAttention on and watch `prompt eval time` in the server log — past ~10k tokens the wait is mostly that.

Can I force valid JSON out of the model for tool results?
Yes — llama.cpp compiles a `json_schema` from the request into a GBNF grammar, and the sampler simply cannot emit an invalid token. Strictness is free; the model only loses the option to ramble.

What's a good starting point to make replies less flat?
Raise the temperature a touch and add `min_p 0.05` — then one knob at a time. Dynamic temperature (`dynatemp_range`) is the fun one: it adapts to per-token entropy.

Do draft models actually help on a single GPU?
Yes, when the draft is much smaller than the target: `--spec-type draft` with a ~1B draft for a 12B model often lands 1.5-2x. The `ngram` variants need no second model — try those first.

Why bge-m3 for the knowledge base and not something newer?
Multilingual, an 8k input window, and one model covers notes and chat attachments alike. Dense retrieval does the heavy lifting; the full-text index catches the exact-match tail.

My `AppState` struct has 40 fields. Where do I start?
Group the fields by who mutates them together — each cluster is a struct candidate. Then move methods to the cluster that owns their data; the borrow checker will referee.

Two tasks each hold a lock and await the other. How do I even see this?
`tokio-console` shows both tasks parked on `Mutex::lock`. The cure is ordering: acquire the locks in one global order — or merge them; a message channel often deletes the second lock outright.

What does the DRY penalty actually do?
It penalizes verbatim repetition of recent sequences, scaled by match length — it kills loops without flattening style the way a high `repeat_penalty` does.

Which MCP servers are actually useful day to day?
Filesystem and fetch cover most of it; add a search server if you live outside the browser. Anything more exotic — write your own: the protocol fits in fifty lines.

Why do table borders fall apart in my terminal?
Your font lacks the box-drawing block, so the terminal substitutes a wider glyph and the grid drifts. JetBrains Mono or Cascadia fixes it — or flip on compat mode and the frames go ASCII.

Three days around Cortina in October — doable?
Doable, but pack for two seasons: the rifugi start closing mid-October. The Tre Cime loop, Cinque Torri and Lago di Sorapis cover the greatest hits.

My system prompt is 900 tokens. What survives a trim?
Keep the identity, the tools and the two rules you actually enforce; cut the examples first — a model infers style from the voice, not from a list of adjectives. 300 tokens reads the same.

What chunk size works for retrieving code?
Whole functions beat fixed windows: split on definitions and keep the signature with the body. For prose, 400-600 tokens with a little overlap is the boring answer that works.

Which attention papers should I read after the 2017 one?
Sparse and linear attention surveys, FlashAttention for the systems side, then RoPE and its long-context descendants — that's the spine of the modern stack.

Anything to check before I update?
Run `mindfork backup -p <password>`, then restore it into a scratch folder. A backup you never restored is a hope, not a backup.

At what point does a regex stop being enough?
The moment you reach for a lookbehind to balance brackets. Nesting means a grammar — and a forty-line recursive descent is usually clearer than the clever pattern it replaces.

Can you draw a flowchart right in the chat?
Yes — fence a ```mermaid block: flowcharts and sequence diagrams render as text graphics, and anything else falls back to the source.

Why is `Manager` always a design smell?
Because it names the absence of a decision — everything manages something. Name the responsibility (`Scheduler`, `Registry`, `Janitor`) and half the design questions answer themselves.

How do I read a GGUF header without loading the model?
The header is plain: magic, version, then key-value pairs — the `gguf` package on PyPI reads it in three lines. Quant type, context length and the tokenizer are all right there.

So, a week in — what stuck?
The notes: answers that start from what we already established feel different. And `Ctrl+T` — reading the thinking taught me how to ask better questions.
"#;

/// The filler row (0-based, in [`ROWS_RAW`] order) whose assistant delegated
/// a review to a sub-agent — the refactoring chat, where a second pair of
/// eyes on a split is the natural ask.
const REVIEWED_ROW: usize = 5;

/// The sub-agent transcript under the refactoring chat (spec §9.3.2): one
/// delegation, completed, titled the way the automatic titling would name
/// it. Fixed ids and times like everything else here. Shown nested in the
/// chat list capture, and openable read-only in the interactive demo.
pub fn reviewer_run() -> SubagentRun {
    let created = date(7, 27, 17, 40);
    let finished = date(7, 27, 17, 44);
    let mut ask = Message::user(
        "Review the proposed split of `AppState` into `Session`, `Catalog`, \
         `Scheduler` and `Janitor`: which methods still borrow across two of \
         them, and what is the smallest change that makes each borrow local?",
    );
    ask.id = Uuid::from_u128(0x6d66_5f64_656d_6f5f_7375_6261_0000_0001);
    ask.timestamp = created;
    let mut verdict = Message::assistant(
        "Two crossings survive the split. `retry_failed()` reads \
         `Catalog.entries` while it mutates `Scheduler.queue` — pass the \
         failed ids in as a `Vec` and the borrow ends at the call. \
         `sweep()` holds `Janitor.expired` across a `Session.close()` — \
         collect the ids first, then close in a second loop. Everything else \
         stays inside one struct; the borrow checker will accept the rest as is.",
    );
    verdict.id = Uuid::from_u128(0x6d66_5f64_656d_6f5f_7375_6261_0000_0002);
    verdict.timestamp = finished;
    SubagentRun {
        id: Uuid::from_u128(0x6d66_5f64_656d_6f5f_7375_6261_0000_0000),
        kind: RunKind::Subagent,
        title: "Borrow-checker review of the split".into(),
        renamed_manually: false,
        name: Some("Reviewer".into()),
        created_at: created,
        finished_at: Some(finished),
        system_message: "You are a meticulous Rust reviewer. Read the proposed \
                         module split and report only what will not compile or \
                         will fight the borrow checker, with the smallest fix."
            .into(),
        sampling_override: None,
        messages: vec![ask, verdict],
        outcome: Some(RunOutcome::Completed),
        tokens: 412,
        participants: Vec::new(),
        background: false,
    }
}

/// The `call_subagent` record that carries [`reviewer_run`] on the
/// refactoring chat's answer.
fn reviewer_record() -> ToolCallRecord {
    let run = reviewer_run();
    ToolCallRecord {
        id: "call_demo_review".into(),
        name: "call_subagent".into(),
        arguments: serde_json::json!({
            "name": "Reviewer",
            "system_message": run.system_message,
            "message": run.messages[0].text,
        }),
        result: Some(format!(
            "{}\n\nTranscript: {}",
            run.final_reply().unwrap_or_default(),
            crate::features::chat_links::uri(run.id)
        )),
        thought_signature: None,
        images: 0,
        subagent: Some(Box::new(run)),
    }
}

/// The `(question, answer)` pairs out of [`ROW_BODIES_RAW`].
fn bodies() -> impl Iterator<Item = (&'static str, &'static str)> {
    let mut lines = table_lines(ROW_BODIES_RAW);
    std::iter::from_fn(move || {
        let question = lines.next()?;
        let answer = lines
            .next()
            .expect("every excerpt is a question/answer pair");
        Some((question, answer))
    })
}

/// The chat list: the showcase chat on top (active), then a spread of
/// plausible topics with fixed dates and counts — exactly enough rows to
/// fill the uniform gallery frame (`app/demo_shots.rs::PANEL_H`) with no
/// scrolling and no trailing void. Capture-only (the interactive demo's
/// list comes from the real seeded chats).
#[cfg(test)]
pub fn chat_summaries() -> Vec<ChatSummary> {
    let mut chats = vec![ChatSummary {
        id: chat_id(),
        profile_id: profile_id(),
        title: CHAT_TITLE.into(),
        created_at: date(8, 1, 10, 0),
        modified_at: date(8, 1, 10, 3),
        message_count: 4,
        children: Vec::new(),
        children_expanded: false,
        unread: false,
    }];
    chats.extend(
        rows()
            .enumerate()
            .map(|(i, (created, modified, count, title))| {
                ChatSummary {
                    // Slots continue the stage-2 numbering (1-based after the hero).
                    id: Uuid::from_u128(0x6d66_5f64_656d_6f5f_6c69_7374_0000_0000 + i as u128 + 1),
                    profile_id: profile_id(),
                    title: title.into(),
                    created_at: created,
                    modified_at: modified,
                    message_count: count,
                    children: if i == REVIEWED_ROW {
                        vec![ChildSummary::of(&reviewer_run())]
                    } else {
                        Vec::new()
                    },
                    // Expanded, so the gallery frame keeps showing the nested
                    // transcript row (and stays byte-identical to the committed
                    // dumps); the fold itself is pinned by the widget's tests.
                    children_expanded: i == REVIEWED_ROW,
                    unread: false,
                }
            }),
    );
    chats
}

/// Goals for the self-model capture: `slot | status | created | closed | text`,
/// where `closed` is `-` for an active goal and `slot` feeds the fixed
/// per-goal id. Actives first, the way the screen groups best.
const GOALS_RAW: &str = r#"
    1 | active | 07-22 12:00 | -           | Answer sizing questions from saved notes alone — no re-asking about hardware.
    2 | active | 07-25 09:30 | -           | Keep recommendations reproducible: name the exact quant, context and flags.
    4 | active | 07-31 18:40 | -           | Measure the 1B draft's speedup and save the number, not the impression.
    3 | done   | 07-16 15:00 | 07-19 11:20 | Index the llama.cpp server docs into the knowledge base.
    5 | done   | 07-18 09:15 | 07-21 17:30 | Turn the October trip research into a packing-checklist note.
"#;

/// Narrative for the capture: `slot | created | text`, in chronological order —
/// the screen shows the narrative newest-first by *insertion*, the way a real
/// profile accretes segments. Times stay mid-day UTC: the screen renders dates
/// in local time, and a near-midnight timestamp would show a different
/// calendar day on CI (UTC) than on the machine that regenerated the dumps.
const NARRATIVE_RAW: &str = r#"
    4 | 07-21 17:35 | The packing checklist came out of three scattered chats — consolidation is where notes earn their keep.
    1 | 07-24 19:40 | Tables beat prose here: comparisons get read, paragraphs get skimmed.
    5 | 07-26 19:02 | Trip planning keeps landing between engineering questions — the terminal is a place to live, not a topic.
    2 | 07-28 10:15 | The 12 GB VRAM budget keeps coming up — saved it as a note so I stop re-asking.
    6 | 07-30 15:55 | Answers that cite the note they came from get fewer follow-ups — receipts beat repetition.
    3 | 08-01 10:03 | Showing the exact command I ran turns out to be the fastest way to build trust.
"#;

/// The self-model the `F3` capture shows: a companion that has already lived
/// a little — a voice, five goals in two states, a user model, and a run of
/// narrative observations consistent with the showcase chat.
pub fn self_model() -> SelfModel {
    let goals = table_lines(GOALS_RAW)
        .map(|line| {
            let mut f = line.splitn(5, '|').map(str::trim);
            let mut field = || {
                f.next()
                    .expect("goal row is slot|status|created|closed|text")
            };
            let slot: u128 = field().parse().unwrap();
            let done = field() == "done";
            let created_at = ts(field());
            let closed = field();
            Goal {
                id: Uuid::from_u128(0x6d66_5f64_656d_6f5f_676f_616c_0000_0000 + slot),
                description: field().into(),
                status: if done {
                    GoalStatus::Completed
                } else {
                    GoalStatus::Active
                },
                created_at,
                closed_at: (closed != "-").then(|| ts(closed)),
            }
        })
        .collect();
    let narrative = table_lines(NARRATIVE_RAW)
        .map(|line| {
            let mut f = line.splitn(3, '|').map(str::trim);
            let mut field = || f.next().expect("narrative row is slot|created|text");
            let slot: u128 = field().parse().unwrap();
            let created_at = ts(field());
            NarrativeSegment {
                id: Uuid::from_u128(0x6d66_5f64_656d_6f5f_6e61_7272_0000_0000 + slot),
                text: field().into(),
                created_at,
            }
        })
        .collect();
    SelfModel {
        profile_id: profile_id(),
        version: 7,
        summary: "I run locally and help with practical engineering — model \
                  sizing, Rust, the occasional trip plan. I prefer measured \
                  numbers to adjectives, and I write things down: advice \
                  should start from remembered facts, not fresh guesses. The \
                  current thread is fitting Gemma 4 onto a 12 GB card without \
                  giving up context; the next experiment on the list is a 1B \
                  draft for speculative decoding."
            .into(),
        goals,
        user_model: UserModel {
            perceived_traits: [
                "methodical",
                "impatient with vague answers",
                "allergic to hand-waving",
            ]
            .map(String::from)
            .into(),
            current_interests: [
                "local model tuning",
                "terminal tooling",
                "speculative decoding",
            ]
            .map(String::from)
            .into(),
            relationship_dynamic: "Collaborative and direct; jokes land better after the numbers."
                .into(),
        },
        narrative,
        updated_at: date(8, 1, 10, 3),
    }
}

/// The config the interactive demo boots with: English UI on the Auto theme
/// (a real terminal supplies the colors), the showcase chat active, and the
/// engine honestly labeled — `external` mode with the model name
/// `demo (mock engine)`, which the feed header shows as the model caption.
/// The supervisor is mocked, so no connection is ever attempted.
pub fn demo_config() -> AppConfig {
    let mut c = AppConfig::default();
    c.interface.language = Lang::En;
    c.engine.mode = ServerMode::External;
    c.engine.external.model_name = Some("demo (mock engine)".into());
    c.last_active_chat = Some(chat_id());
    c
}

/// The showcase conversation as a real, saveable chat — the one the demo
/// opens on. Fixed id/timestamps; thoughts and tool details pre-expanded,
/// like the hero screenshot.
pub fn showcase_chat() -> Chat {
    let mut chat = Chat::from_profile(&profile(), CHAT_TITLE);
    chat.id = chat_id();
    chat.created_at = at(0);
    chat.modified_at = at(3);
    chat.messages = showcase_messages();
    chat.feed_view = feed_view();
    // The same unsent question the hero screenshot shows in the input box —
    // the interactive demo opens mid-thought, ready to send.
    chat.draft = INPUT_DRAFT.into();
    chat
}

/// The filler chats as real chats: [`ROWS_RAW`] titles and dates with the
/// [`ROW_BODIES_RAW`] two-message excerpts.
pub fn filler_chats() -> Vec<Chat> {
    rows()
        .zip(bodies())
        .enumerate()
        .map(|(i, ((created, modified, _, title), (user, assistant)))| {
            let mut chat = Chat::from_profile(&profile(), title);
            chat.id = Uuid::from_u128(0x6d66_5f64_656d_6f5f_6c69_7374_0000_0000 + i as u128 + 1);
            chat.created_at = created;
            chat.modified_at = modified;
            let mut u = Message::user(user);
            u.id = Uuid::from_u128(0x6d66_5f64_656d_6f5f_6d73_6700 + (i as u128) * 2);
            u.timestamp = created;
            let mut a = Message::assistant(assistant);
            a.id = Uuid::from_u128(0x6d66_5f64_656d_6f5f_6d73_6700 + (i as u128) * 2 + 1);
            a.timestamp = modified;
            if i == REVIEWED_ROW {
                a.tool_calls = vec![reviewer_record()];
            }
            chat.messages = vec![u, a];
            chat
        })
        .collect()
}

/// Splits reply text into few-word chunks so the demo's streaming is visibly
/// a stream, the way a real server delivers tokens.
fn stream_text(text: &str, out: &mut Vec<ChatChunk>) {
    let words: Vec<&str> = text.split(' ').collect();
    for group in words.chunks(3) {
        let mut piece = group.join(" ");
        piece.push(' ');
        out.push(ChatChunk::Text(piece));
    }
    if let Some(ChatChunk::Text(last)) = out.last_mut() {
        while last.ends_with(' ') {
            last.pop();
        }
    }
}

/// One canned reply: optional thoughts, streamed text, plausible usage, stop.
fn reply(thoughts: Option<&str>, text: &str, prompt_tokens: u32) -> Vec<ChatChunk> {
    let mut chunks = Vec::new();
    let mut reasoning_tokens = 0;
    if let Some(t) = thoughts {
        chunks.push(ChatChunk::Thoughts(t.into()));
        reasoning_tokens = (t.len() / 4) as u32;
    }
    stream_text(text, &mut chunks);
    chunks.push(ChatChunk::Usage(TokenUsage {
        prompt_tokens,
        completion_tokens: (text.len() / 4) as u32 + reasoning_tokens,
        reasoning_tokens,
        prefill: None,
    }));
    chunks.push(ChatChunk::Finished(FinishReason::Stop));
    chunks
}

/// The demo engine's replies, cycled in order by `MockBackend::cycling`.
/// Self-contained on purpose: background calls (impersonation, a regenerate)
/// may consume a script out of turn, so no reply depends on which question
/// preceded it — each is honest about being scripted and shows something
/// real about the app.
pub fn demo_replies() -> Vec<Vec<ChatChunk>> {
    vec![
        reply(
            Some("Best to be upfront about what I am before pretending to be clever."),
            "Fair warning: **I'm the demo engine** — a scripted stand-in cycling \
             through a fixed set of replies, no model behind me. Everything around \
             me is real, though: this streaming, the collapsible thoughts above \
             (`Ctrl+T`), the notes, the search, the themes. Press `Ctrl+P` and \
             point mindfork at a real engine — then this seat gets a mind.",
            412,
        ),
        reply(
            Some("A quick rendering tour says more than a paragraph of claims."),
            "A few things the feed renders natively:\n\n\
             | Piece | In the terminal |\n\
             |---|---|\n\
             | GFM tables | this one |\n\
             | Code | `cargo run` with highlighting |\n\
             | Mermaid | flowcharts as text graphics |\n\
             | LaTeX | $E = mc^2$, no rasterization |\n\n\
             Ask about diagrams in the *Mermaid* chat in the list (`Esc`) to see \
             a flowchart drawn live.",
            498,
        ),
        reply(
            None,
            "Connecting a real engine takes a minute: `Ctrl+P` → Model/server. \
             Local — point the `external` mode at any OpenAI-compatible server \
             (llama.cpp, vLLM, LM Studio, Ollama), or let the `managed` mode \
             launch `llama-server` for you. Cloud — pick OpenAI, Gemini, Claude \
             or Grok and paste an API key; it is stored encrypted, bound to this \
             machine.",
            531,
        ),
        reply(
            Some(
                "The self-model is the differentiating feature — worth pointing at \
                 the living example seeded in this very demo.",
            ),
            "What makes mindfork more than a chat window is memory with a spine: \
             I keep notes, and I maintain a **self-model** — who I am, goals, a \
             model of you, observations over time. Press `F3` to read the one \
             seeded in this demo. With a real model in this seat, it grows on its \
             own as we talk.",
            577,
        ),
    ]
}

/// Provisions a demo data root: config, the companion profile, the showcase
/// chat, the filler chats and the self-model. Idempotent — every id is fixed,
/// so re-running overwrites the same records. Isolation is the caller's job
/// (`main` passes a throwaway temp root).
pub fn provision(storage: &Storage) -> anyhow::Result<()> {
    storage.json().save_config(&demo_config())?;
    storage.json().upsert_profile(&profile())?;
    storage.json().save_chat(&showcase_chat())?;
    for chat in filler_chats() {
        storage.json().save_chat(&chat)?;
    }
    storage.db().self_model_upsert(&self_model())?;
    Ok(())
}

#[cfg(test)]
mod tests {
    use super::*;

    /// The fixture is frozen: same ids, same timestamps, same text on every
    /// call — the property every capture and the drift gate stand on.
    #[test]
    fn fixture_is_deterministic() {
        let a = showcase_messages();
        let b = showcase_messages();
        assert_eq!(a, b);
        assert_eq!(a[0].timestamp, at(0));
        assert_ne!(a[0].id, a[1].id, "slot-derived ids must not collide");
    }

    /// The showcase must keep exercising the renderer's range — losing a
    /// feature from the fixture silently demotes the screenshots.
    #[test]
    fn fixture_covers_the_showcase_features() {
        let all = showcase_messages();
        let text: String = all.iter().map(|m| m.text.as_str()).collect();
        for needle in ["| Quant |", "```bash", "```mermaid", "$M_{kv}"] {
            assert!(text.contains(needle), "fixture lost {needle:?}");
        }
        let last = all.last().unwrap();
        assert!(
            last.thoughts.is_some(),
            "the thoughts block is part of the showcase"
        );
        assert_eq!(
            last.tool_calls.len(),
            1,
            "the tool card is part of the showcase"
        );
    }

    /// The list capture and the seeded chats describe one world: the row
    /// that shows a transcript is the chat whose file holds it, and the
    /// transcript is the same run in both.
    #[test]
    fn the_list_capture_and_the_seeded_chats_agree_on_the_transcript() {
        let summaries = chat_summaries();
        let with_child: Vec<&ChatSummary> = summaries
            .iter()
            .filter(|c| !c.children.is_empty())
            .collect();
        assert_eq!(with_child.len(), 1);
        assert_eq!(with_child[0].title, "Refactoring a god object in Rust");
        let run = reviewer_run();
        assert_eq!(with_child[0].children[0].id, run.id);
        assert_eq!(with_child[0].children[0].title, run.title);
        let chat = &filler_chats()[REVIEWED_ROW];
        assert_eq!(chat.id, with_child[0].id);
        assert_eq!(chat.child(run.id).unwrap().final_reply(), run.final_reply());
        assert_eq!(reviewer_run(), reviewer_run(), "frozen like the rest");
    }

    /// Provisioning a fresh root seeds everything the demo needs, and doing
    /// it twice changes nothing — every id is fixed.
    #[test]
    fn provision_seeds_a_complete_demo_root() {
        let dir = tempfile::tempdir().unwrap();
        let storage = Storage::open(crate::shared::paths::Paths::with_root(dir.path())).unwrap();
        provision(&storage).unwrap();
        provision(&storage).unwrap();

        let config = storage.json().load_config().unwrap();
        assert_eq!(config.last_active_chat, Some(chat_id()));
        assert_eq!(
            config.engine.external.model_name.as_deref(),
            Some("demo (mock engine)"),
            "the feed-header caption must say what the engine is"
        );

        let profiles = storage.json().load_profiles().unwrap();
        assert_eq!(profiles.len(), 1, "idempotent: one profile after two runs");
        assert_eq!(profiles[0].id, profile_id());
        let greeting = profiles[0].greeting.as_deref().unwrap_or_default();
        assert!(
            greeting.contains("demo"),
            "a new chat must say what this is"
        );

        let files = storage.json().chat_files().unwrap();
        assert_eq!(files.len(), 21, "the showcase chat + 20 fillers");
        let reviewed = filler_chats()[REVIEWED_ROW].id;
        let reviewed = storage.json().load_chat(reviewed).unwrap().unwrap();
        assert_eq!(
            reviewed.children().map(|r| r.id).collect::<Vec<_>>(),
            vec![reviewer_run().id],
            "the sub-agent transcript is part of the seeded world"
        );
        let showcase = storage.json().load_chat(chat_id()).unwrap().unwrap();
        assert_eq!(showcase.messages.len(), 4);
        assert!(
            showcase.feed_view.thoughts,
            "the hero look: thoughts expanded"
        );
        assert_eq!(showcase.profile_id, profile_id());

        let model = storage.db().self_model_get(profile_id()).unwrap();
        assert!(model.is_some(), "F3 must have something to show");
    }

    /// Every canned reply is a complete turn — text, usage, a terminal Stop —
    /// and the first one names what it is: a demo must close the door on
    /// "is this a real model?".
    #[test]
    fn demo_replies_are_complete_selfcontained_turns() {
        let replies = demo_replies();
        assert!(replies.len() >= 3, "cycling needs variety");
        for (i, r) in replies.iter().enumerate() {
            assert!(
                matches!(r.last(), Some(ChatChunk::Finished(FinishReason::Stop))),
                "reply {i} must end with Stop"
            );
            assert!(
                r.iter().any(|c| matches!(c, ChatChunk::Text(_))),
                "reply {i} has text"
            );
            assert!(
                r.iter().any(|c| matches!(c, ChatChunk::Usage(_))),
                "reply {i} reports usage (the token counter is part of the showcase)"
            );
        }
        let first: String = replies[0]
            .iter()
            .filter_map(|c| match c {
                ChatChunk::Text(t) => Some(t.as_str()),
                _ => None,
            })
            .collect();
        assert!(
            first.contains("demo engine"),
            "the first reply says what it is"
        );
    }
}