codecraft/server/src/cli_api.rs

//! API for the local command line client (`codecraft`): the CLI runs the
//! agent loop and the file/shell tools on the user's machine; the server
//! provides the configuration (modes, models, rules, skills, theme), acts as
//! the model gateway (keys, budgets, statistics) and runs its MCP tools.

use crate::api::{ApiError, ApiResult, AppState};
use crate::provider::{ChatRequest, StreamEvent, ToolDef};
use crate::store::{UsageRecord, User};
use crate::tools::ToolCtx;
use crate::types::{Message, now_ms};
use axum::extract::{Path, State};
use axum::http::StatusCode;
use axum::response::sse::{Event as SseEvent, KeepAlive, Sse};
use axum::{Extension, Json};
use futures::stream::Stream;
use serde::Deserialize;
use serde_json::{Value, json};
use std::convert::Infallible;
use std::path::PathBuf;
use tokio_util::sync::CancellationToken;

/// Everything the CLI needs to run like the web UI.
pub async fn config(Extension(user): Extension<User>, State(app): State<AppState>) -> ApiResult {
    let cfg = app.config();
    let root = app.user_roots(Some(&user.id)).into_iter().next().unwrap_or_else(|| app.workspace());
    let agents: Vec<Value> = crate::prompts::agents_for(&app, &root, &root, Some(&user.id))
        .into_iter()
        .map(|a| {
            let mut v = json!(a);
            v["prompt"] = json!(a.prompt);
            v
        })
        .collect();
    let commands: Vec<Value> = crate::prompts::commands(&app, &root, &root)
        .into_iter()
        .filter(|c| c.description.len() < 400)
        .map(|c| json!({"name": c.name, "description": c.description, "body": c.body, "agent": c.agent}))
        .collect();
    let skills: Vec<Value> = crate::prompts::skills(&app, &root, &root).into_iter().map(|s| json!({"name": s.name, "description": s.description})).collect();
    let (direct, deferred) = app.mcp.tools(&cfg, Some(&user));
    let mut mcp_tools: Vec<ToolDef> = direct.iter().map(|t| t.def()).collect();
    if !deferred.is_empty() {
        mcp_tools.push(crate::tools::Tool::def(&crate::mcp::McpSearch));
        mcp_tools.push(crate::tools::Tool::def(&crate::mcp::McpCall));
    }
    let prefs: Value = app.store.user_setting(&user.id, "prefs")?;
    let memory: String = app.store.user_setting(&user.id, "memory")?;
    let default_model = if cfg.server.default_model.is_empty() { app.registry().first_model() } else { Some(cfg.server.default_model.clone()) };
    Ok(Json(json!({
        "version": codecraft_core::version::VERSION,
        "user": user,
        "models": app.registry().models_json(),
        "default_model": default_model,
        "small_model": cfg.server.small_model,
        "currency": cfg.server.currency,
        "base_prompt": crate::prompts::BASE_PROMPT,
        "auto_mode_prompt": crate::prompts::AUTO_MODE_PROMPT,
        "agents": agents,
        "commands": commands,
        "skills": skills,
        "permissions": cfg.permissions,
        "mcp_tools": mcp_tools,
        "deferred_mcp": app.mcp.deferred_summary(&cfg, Some(&user)).iter().map(|(s, n, names)| json!({"server": s, "tools": n, "examples": names})).collect::<Vec<_>>(),
        "prefs": prefs,
        "memory": memory,
    })))
}

/// A server skill with its supporting text files (templates).
pub async fn skill(Extension(user): Extension<User>, State(app): State<AppState>, Path(name): Path<String>) -> ApiResult {
    let root = app.user_roots(Some(&user.id)).into_iter().next().unwrap_or_else(|| app.workspace());
    let sk = crate::prompts::skills(&app, &root, &root).into_iter().find(|s| s.name == name).ok_or_else(|| ApiError(StatusCode::NOT_FOUND, "skill not found".into()))?;
    let dir = sk.path.parent().map(PathBuf::from).unwrap_or_default();
    let content = std::fs::read_to_string(&sk.path)?;
    let mut files = Vec::new();
    for e in ignore::WalkBuilder::new(&dir).hidden(false).build().flatten() {
        let p = e.path();
        if !p.is_file() || p == sk.path || files.len() >= 50 {
            continue;
        }
        if let Ok(text) = std::fs::read_to_string(p) {
            if text.len() <= 200_000 {
                files.push(json!({"path": p.strip_prefix(&dir).unwrap_or(p), "content": text}));
            }
        }
    }
    Ok(Json(json!({"name": sk.name, "description": sk.description, "content": content, "files": files})))
}

#[derive(Deserialize)]
pub struct LlmBody {
    /// "provider/model"
    model: String,
    system: String,
    messages: Vec<Message>,
    #[serde(default)]
    tools: Vec<ToolDef>,
    max_tokens: Option<u64>,
    #[serde(default)]
    reasoning: bool,
    effort: Option<String>,
    /// Local session id (statistics group the requests by it).
    session: Option<String>,
    /// Mode / agent name (statistics).
    agent: Option<String>,
}

/// Model gateway: one model request, streamed as SSE (`data: <json>`).
/// Events: text, reasoning, signature, tool_call, usage, done, error.
pub async fn llm_stream(Extension(user): Extension<User>, State(app): State<AppState>, Json(b): Json<LlmBody>) -> Result<Sse<impl Stream<Item = Result<SseEvent, Infallible>>>, ApiError> {
    let (provider, mut model) = app.registry().resolve(&b.model).map_err(|e| ApiError(StatusCode::BAD_REQUEST, format!("{e:#}")))?;
    if let Some(e) = b.effort.filter(|e| !e.is_empty()) {
        model.effort = Some(e);
    }
    let provider_name = b.model.split_once('/').map(|(p, _)| p.to_string()).unwrap_or_default();
    crate::budget::check(&app, &provider_name, Some(&user.id)).map_err(|e| ApiError(StatusCode::PAYMENT_REQUIRED, format!("{e:#}")))?;
    let max_tokens = b.max_tokens.or(model.max_output).unwrap_or(16_000);
    let session = format!("cli:{}", b.session.unwrap_or_else(|| "default".into()));
    let req = ChatRequest { model: model.clone(), system: b.system, messages: b.messages, tools: b.tools, max_tokens, reasoning: b.reasoning, session_id: Some(session.clone()) };
    let agent = format!("cli:{}", b.agent.unwrap_or_else(|| "build".into()));
    let spec = b.model.clone();
    let (out_tx, out_rx) = tokio::sync::mpsc::channel::<Result<SseEvent, Infallible>>(256);
    tokio::spawn(async move {
        let slot = app.registry().acquire(&spec).await;
        let started = std::time::Instant::now();
        let (tx, mut rx) = tokio::sync::mpsc::channel(256);
        let cancel = CancellationToken::new();
        let c2 = cancel.clone();
        let handle = tokio::spawn(async move {
            let _slot = slot;
            provider.stream(req, tx, c2).await
        });
        let mut usage = crate::types::Usage::default();
        let mut ttft = None;
        while let Some(ev) = rx.recv().await {
            let v = match ev {
                StreamEvent::TextDelta(t) => json!({"type": "text", "text": t}),
                StreamEvent::ReasoningDelta(t) => json!({"type": "reasoning", "text": t}),
                StreamEvent::ReasoningSignature(s) => json!({"type": "signature", "signature": s}),
                StreamEvent::ToolCall { id, name, input, extra } => json!({"type": "tool_call", "id": id, "name": name, "input": input, "extra": extra}),
                StreamEvent::Usage(u) => {
                    usage = u.clone();
                    json!({"type": "usage", "usage": u})
                }
                StreamEvent::Done { stop_reason } => json!({"type": "done", "stop_reason": stop_reason}),
            };
            if ttft.is_none() && v["type"] != "usage" {
                ttft = Some(started.elapsed().as_millis() as u64);
            }
            if out_tx.send(Ok(SseEvent::default().data(v.to_string()))).await.is_err() {
                // Client went away: stop the model request.
                cancel.cancel();
                break;
            }
        }
        let result = handle.await;
        let error = !matches!(result, Ok(Ok(())));
        if let Ok(Err(e)) | Err(e) = result.map_err(anyhow::Error::from) {
            let _ = out_tx.send(Ok(SseEvent::default().data(json!({"type": "error", "message": format!("{e:#}")}).to_string()))).await;
        }
        let (p, m) = spec.split_once('/').unwrap_or(("", spec.as_str()));
        let _ = app.store.add_usage(&UsageRecord {
            ts: now_ms(),
            session_id: session.clone(),
            root_session_id: session,
            provider: p.into(),
            model: m.into(),
            agent,
            input_tokens: usage.input_tokens,
            output_tokens: usage.output_tokens,
            cache_read_tokens: usage.cache_read_tokens,
            cache_write_tokens: usage.cache_write_tokens,
            ttft_ms: ttft.unwrap_or(0),
            duration_ms: started.elapsed().as_millis() as u64,
            cost: model.cost(&usage),
            error,
            user_id: Some(user.id.clone()),
            allowance: usage.allowance,
        });
    });
    let stream = futures::stream::unfold(out_rx, |mut rx| async move { rx.recv().await.map(|ev| (ev, rx)) });
    Ok(Sse::new(stream).keep_alive(KeepAlive::default()))
}

#[derive(Deserialize)]
pub struct ToolBody {
    name: String,
    input: Value,
}

/// Run one of the server's MCP tools for the CLI (permission was asked locally).
pub async fn tool(Extension(user): Extension<User>, State(app): State<AppState>, Json(b): Json<ToolBody>) -> ApiResult {
    let cfg = app.config();
    let (direct, deferred) = app.mcp.tools(&cfg, Some(&user));
    let tool: std::sync::Arc<dyn crate::tools::Tool> = match b.name.as_str() {
        "mcp_search" if !deferred.is_empty() => std::sync::Arc::new(crate::mcp::McpSearch),
        "mcp_call" if !deferred.is_empty() => std::sync::Arc::new(crate::mcp::McpCall),
        n => direct.into_iter().find(|t| t.name() == n).ok_or_else(|| ApiError(StatusCode::NOT_FOUND, format!("no MCP tool '{n}' for you")))?,
    };
    let root = app.user_roots(Some(&user.id)).into_iter().next().unwrap_or_else(|| app.workspace());
    let ctx = ToolCtx {
        app: app.clone(),
        session_id: format!("cli-{}", user.id),
        call_id: crate::types::new_id(),
        cwd: root.clone(),
        cancel: CancellationToken::new(),
        agent: "build".into(),
        roots: app.user_roots(Some(&user.id)),
        sandbox: crate::sandbox::spec_for(&app, Some(&user.id)),
        clean_room: false,
        owner_id: Some(user.id.clone()),
    };
    app.audit(Some(&user), "cli.mcp", &b.name, "");
    match tool.run(b.input, &ctx).await {
        Ok(out) => Ok(Json(json!({"content": out.content, "is_error": out.is_error, "images": out.images}))),
        Err(e) => Ok(Json(json!({"content": format!("Error: {e:#}"), "is_error": true}))),
    }
}

/// Interface preferences (theme) shared between web UI and CLI.
pub async fn get_prefs(Extension(user): Extension<User>, State(app): State<AppState>) -> ApiResult {
    let v: Value = app.store.user_setting(&user.id, "prefs")?;
    Ok(Json(if v.is_null() { json!({}) } else { v }))
}

pub async fn put_prefs(Extension(user): Extension<User>, State(app): State<AppState>, Json(b): Json<Value>) -> ApiResult {
    if !b.is_object() || b.to_string().len() > 4000 {
        return Err(ApiError(StatusCode::BAD_REQUEST, "invalid preferences".into()));
    }
    app.store.set_user_setting(&user.id, "prefs", &b)?;
    Ok(Json(b))
}

/// Personal memory (applies to all projects, web and CLI).
pub async fn get_memory(Extension(user): Extension<User>, State(app): State<AppState>) -> ApiResult {
    let m: String = app.store.user_setting(&user.id, "memory")?;
    Ok(Json(json!({"memory": m})))
}

#[derive(Deserialize)]
pub struct MemoryBody {
    memory: String,
}

pub async fn put_memory(Extension(user): Extension<User>, State(app): State<AppState>, Json(b): Json<MemoryBody>) -> ApiResult {
    if b.memory.len() > 20_000 {
        return Err(ApiError(StatusCode::BAD_REQUEST, "the memory is limited to 20,000 characters".into()));
    }
    app.store.set_user_setting(&user.id, "memory", &b.memory)?;
    Ok(Json(json!({"memory": b.memory})))
}

// --- prompt templates ---------------------------------------------------------

/// Owner id of the templates every user sees (edited by admins).
const GLOBAL_TEMPLATES: &str = "*";

#[derive(Debug, Clone, serde::Serialize, Deserialize, Default)]
pub struct Template {
    pub id: String,
    pub name: String,
    pub text: String,
}

pub async fn get_templates(Extension(user): Extension<User>, State(app): State<AppState>) -> ApiResult {
    let personal: Vec<Template> = app.store.user_setting(&user.id, "templates")?;
    let global: Vec<Template> = app.store.user_setting(GLOBAL_TEMPLATES, "templates")?;
    Ok(Json(json!({"personal": personal, "global": global, "can_edit_global": user.is_admin()})))
}

pub async fn put_templates(Extension(user): Extension<User>, State(app): State<AppState>, Path(scope): Path<String>, Json(list): Json<Vec<Template>>) -> ApiResult {
    let owner = match scope.as_str() {
        "personal" => user.id.clone(),
        "global" => {
            crate::auth::require_admin(&user)?;
            GLOBAL_TEMPLATES.to_string()
        }
        _ => return Err(ApiError(StatusCode::NOT_FOUND, "unknown scope".into())),
    };
    if list.len() > 200 || list.iter().any(|t| t.name.trim().is_empty() || t.name.len() > 100 || t.text.len() > 20_000) {
        return Err(ApiError(StatusCode::BAD_REQUEST, "templates need a name (max. 100 characters) and at most 20,000 characters of text".into()));
    }
    let list: Vec<Template> = list
        .into_iter()
        .map(|mut t| {
            if t.id.is_empty() {
                t.id = uuid::Uuid::new_v4().simple().to_string();
            }
            t.name = t.name.trim().to_string();
            t
        })
        .collect();
    app.store.set_user_setting(&owner, "templates", &list)?;
    Ok(Json(json!(list)))
}