You cannot select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

101 KiB

Raw Blame History

Backend Implementation Plan: AI Weekly Synth

Date: 2026-03-21 Role: Backend Implementation Planner Target: Rust (Axum) + PostgreSQL + Multi-LLM

Project Structure
Database Schema
API Endpoints
Authentication & Authorization
LLM Provider Abstraction
URL Scraping (Server-Side)
Async Generation & SSE
Email (Resend Integration)
Encryption
Error Handling
Testing Strategy

1. Project Structure

1.1 Single Crate with Module Hierarchy

A single crate is sufficient for this application. A Cargo workspace adds overhead (inter-crate dependency management, slower incremental compilation across crate boundaries) without meaningful benefit for an app of this size. If the LLM abstraction layer grows significantly, it can be extracted into a workspace member later.

ai-synth-backend/
├── Cargo.toml
├── Cargo.lock
├── .env.example
├── migrations/
│   ├── 20260321000001_create_users.sql
│   ├── 20260321000002_create_sessions.sql
│   ├── 20260321000003_create_magic_link_tokens.sql
│   ├── 20260321000004_create_user_settings.sql
│   ├── 20260321000005_create_sources.sql
│   ├── 20260321000006_create_syntheses.sql
│   ├── 20260321000007_create_user_provider_keys.sql
│   ├── 20260321000008_create_admin_provider_models.sql
│   ├── 20260321000009_create_admin_rate_limits.sql
│   └── 20260321000010_create_generation_jobs.sql
├── src/
│   ├── main.rs                        # Entry point: CLI (create-admin, serve), tracing init, DB pool, run server
│   ├── cli.rs                         # CLI argument parsing (clap): serve, create-admin
│   ├── config.rs                      # AppConfig struct, loaded from env vars via dotenvy
│   ├── app_state.rs                   # AppState: PgPool, reqwest::Client, AppConfig, RateLimiter, JobStore
│   ├── error.rs                       # AppError enum, IntoResponse impl, From impls
│   ├── router.rs                      # All route definitions, middleware wiring, layer composition
│   │
│   ├── middleware/
│   │   ├── mod.rs
│   │   ├── auth.rs                    # AuthUser extractor (session cookie -> user lookup)
│   │   ├── admin.rs                   # AdminUser extractor (AuthUser + role == admin check)
│   │   └── csrf.rs                    # X-Requested-With header check on mutating methods
│   │
│   ├── models/
│   │   ├── mod.rs
│   │   ├── user.rs                    # User, UserRole, CreateUser
│   │   ├── session.rs                 # Session
│   │   ├── magic_link.rs             # MagicLinkToken
│   │   ├── settings.rs               # UserSettings, UpdateSettings
│   │   ├── source.rs                  # Source, CreateSource
│   │   ├── synthesis.rs               # Synthesis, NewsSection, NewsItem
│   │   ├── provider_key.rs            # UserProviderKey, CreateProviderKey
│   │   ├── admin_model.rs             # AdminProviderModel
│   │   ├── rate_limit.rs              # AdminRateLimit
│   │   └── generation_job.rs          # GenerationJob, JobStatus, JobProgress
│   │
│   ├── handlers/
│   │   ├── mod.rs
│   │   ├── auth.rs                    # register, request_magic_link, verify_magic_link, logout, me
│   │   ├── syntheses.rs               # list, get, delete, trigger_generate
│   │   ├── sources.rs                 # list, create, delete
│   │   ├── settings.rs                # get, update
│   │   ├── provider_keys.rs           # list, add, delete (user's own API keys)
│   │   ├── generation.rs              # SSE progress endpoint, job status
│   │   ├── admin.rs                   # provider models CRUD, rate limits, user list
│   │   └── health.rs                  # GET /health
│   │
│   ├── services/
│   │   ├── mod.rs
│   │   ├── llm/
│   │   │   ├── mod.rs                 # LlmProvider trait, ProviderCapabilities, create_provider factory
│   │   │   ├── types.rs               # SearchRequest, SearchResponse, RewriteRequest, shared types
│   │   │   ├── gemini.rs              # GeminiProvider implementation
│   │   │   ├── openai.rs              # OpenAiProvider implementation
│   │   │   └── anthropic.rs           # AnthropicProvider implementation
│   │   ├── generation.rs              # GenerationPipeline: orchestrates search -> scrape -> rewrite
│   │   ├── scraper.rs                 # URL fetching, HTML parsing, date extraction, SSRF checks
│   │   ├── email.rs                   # Resend HTTP API client (magic links + synthesis delivery)
│   │   ├── captcha.rs                 # Cloudflare Turnstile verification
│   │   ├── encryption.rs             # AES-256-GCM encrypt/decrypt for user API keys
│   │   └── rate_limiter.rs            # Token-bucket rate limiter (per-provider, in-memory)
│   │
│   ├── db/
│   │   ├── mod.rs
│   │   ├── users.rs                   # find_by_email, find_by_id, create, update_role, list_all
│   │   ├── sessions.rs                # create, find_by_hash, delete, delete_expired, delete_all_for_user
│   │   ├── magic_links.rs             # create, consume (atomic find + mark used), delete_expired
│   │   ├── settings.rs                # get_or_default, upsert
│   │   ├── sources.rs                 # list_by_user, create, delete
│   │   ├── syntheses.rs               # list_by_user, get_by_id, create, delete
│   │   ├── provider_keys.rs           # list_by_user, create, delete, get_decrypted
│   │   ├── admin_models.rs            # list_all, list_enabled, create, update, delete
│   │   └── admin_rate_limits.rs       # get_by_provider, upsert, list_all
│   │
│   └── util/
│       ├── mod.rs
│       ├── token.rs                   # Secure random token generation, SHA-256 hashing
│       └── validation.rs              # Shared validation helpers (URL scheme check, etc.)
│
└── tests/
    ├── common/
    │   └── mod.rs                     # Test helpers: setup DB, create test user, create test session
    ├── api/
    │   ├── auth_test.rs
    │   ├── syntheses_test.rs
    │   ├── sources_test.rs
    │   ├── settings_test.rs
    │   └── admin_test.rs
    └── services/
        ├── encryption_test.rs
        ├── scraper_test.rs
        └── rate_limiter_test.rs

1.2 Key Dependencies

[package]
name = "ai-synth-backend"
version = "0.1.0"
edition = "2021"

[dependencies]
# Web framework
axum = { version = "0.8", features = ["macros"] }
axum-extra = { version = "0.10", features = ["cookie", "typed-header"] }
tower = { version = "0.5", features = ["util", "timeout"] }
tower-http = { version = "0.6", features = ["fs", "cors", "trace", "set-header"] }
tokio = { version = "1", features = ["full"] }

# Database
sqlx = { version = "0.8", features = ["runtime-tokio", "tls-rustls", "postgres", "uuid", "chrono", "json"] }

# Serialization
serde = { version = "1", features = ["derive"] }
serde_json = "1"

# HTTP client (LLM APIs + scraping)
reqwest = { version = "0.12", features = ["json", "stream"] }

# HTML parsing
scraper = "0.22"

# Date/time
chrono = { version = "0.4", features = ["serde"] }

# Cryptography
sha2 = "0.10"               # SHA-256 for token hashing
rand = "0.8"                 # OsRng for secure random generation
aes-gcm = "0.10"            # AES-256-GCM for API key encryption
base64 = "0.22"             # Encoding tokens and encrypted data

# Secrets management
secrecy = { version = "0.10", features = ["serde"] }
zeroize = "1"

# Validation
validator = { version = "0.19", features = ["derive"] }

# Logging
tracing = "0.1"
tracing-subscriber = { version = "0.3", features = ["env-filter", "json"] }

# Config
dotenvy = "0.15"
clap = { version = "4", features = ["derive"] }

# Concurrency
dashmap = "6"
tokio-stream = "0.1"        # For SSE streaming

# Error handling
anyhow = "1"
thiserror = "1"

# UUID
uuid = { version = "1", features = ["v4", "serde"] }

# URL parsing
url = "2"

# Async trait (still needed for trait objects)
async-trait = "0.1"

# Futures utilities
futures = "0.3"

[dev-dependencies]
sqlx = { version = "0.8", features = ["runtime-tokio", "tls-rustls", "postgres"] }
tower = { version = "0.5", features = ["util"] }
http-body-util = "0.1"
hyper = "1"

Justification for key choices:

Crate	Why
`axum 0.8`	Tower-based, composes naturally with middleware layers, best ergonomics for extractors
`sqlx 0.8` with `postgres`	Compile-time query checking, native async, direct Postgres support (decision: no SQLite)
`reqwest 0.12`	Industry standard async HTTP client, reused for LLM APIs and scraping
`scraper 0.22`	Built on html5ever (same parser as Firefox), reliable HTML parsing
`aes-gcm 0.10`	Pure Rust authenticated encryption, no OpenSSL dependency
`secrecy + zeroize`	Prevents accidental logging of secrets, zeroes memory on drop
`dashmap 6`	Lock-free concurrent HashMap for rate limiter and job store
`clap 4`	CLI argument parsing for `serve` and `create-admin` subcommands
`tokio-stream`	SSE implementation via `axum::response::Sse`
`chrono` over `time`	Better serde integration, more format support, widely used

2. Database Schema

All migrations use PostgreSQL-native types. UUIDs are UUID type (not TEXT). Timestamps are TIMESTAMPTZ. JSONB is used for flexible structured fields.

Migration 001: Users

-- 20260321000001_create_users.sql

CREATE TYPE user_role AS ENUM ('user', 'admin');

CREATE TABLE users (
    id          UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    email       TEXT NOT NULL UNIQUE,
    display_name TEXT,
    role        user_role NOT NULL DEFAULT 'user',
    created_at  TIMESTAMPTZ NOT NULL DEFAULT now(),
    updated_at  TIMESTAMPTZ NOT NULL DEFAULT now()
);

CREATE INDEX idx_users_email ON users(email);

Migration 002: Sessions

-- 20260321000002_create_sessions.sql

CREATE TABLE sessions (
    id           UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id      UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
    token_hash   TEXT NOT NULL UNIQUE,   -- SHA-256 of the raw session token
    expires_at   TIMESTAMPTZ NOT NULL,
    ip_address   INET,
    user_agent   TEXT,
    created_at   TIMESTAMPTZ NOT NULL DEFAULT now()
);

CREATE INDEX idx_sessions_user_id ON sessions(user_id);
CREATE INDEX idx_sessions_token_hash ON sessions(token_hash);
CREATE INDEX idx_sessions_expires_at ON sessions(expires_at);

Migration 003: Magic Link Tokens

-- 20260321000003_create_magic_link_tokens.sql

CREATE TABLE magic_link_tokens (
    id          UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    email       TEXT NOT NULL,
    token_hash  TEXT NOT NULL UNIQUE,    -- SHA-256 of the raw token
    expires_at  TIMESTAMPTZ NOT NULL,
    used        BOOLEAN NOT NULL DEFAULT false,
    created_at  TIMESTAMPTZ NOT NULL DEFAULT now()
);

CREATE INDEX idx_magic_link_tokens_email ON magic_link_tokens(email);
CREATE INDEX idx_magic_link_tokens_token_hash ON magic_link_tokens(token_hash);
CREATE INDEX idx_magic_link_tokens_expires_at ON magic_link_tokens(expires_at);

Migration 004: User Settings

-- 20260321000004_create_user_settings.sql

CREATE TABLE user_settings (
    user_id                UUID PRIMARY KEY REFERENCES users(id) ON DELETE CASCADE,
    theme                  TEXT NOT NULL DEFAULT 'Intelligence Artificielle',
    max_age_days           INTEGER NOT NULL DEFAULT 7
                           CHECK (max_age_days > 0 AND max_age_days <= 365),
    categories             JSONB NOT NULL DEFAULT '["Annonces majeures / importantes", "Entreprises des secteurs financiers", "Grandes entreprises des autres secteurs", "Secteurs publics", "Grand public / Particuliers"]'::jsonb,
    max_items_per_category INTEGER NOT NULL DEFAULT 4
                           CHECK (max_items_per_category > 0 AND max_items_per_category <= 50),
    search_agent_behavior  TEXT NOT NULL DEFAULT '',
    preferred_provider     TEXT,          -- 'gemini', 'openai', 'anthropic' (NULL = use first available)
    preferred_model        TEXT,          -- model identifier (NULL = use provider default)
    updated_at             TIMESTAMPTZ NOT NULL DEFAULT now()
);

Migration 005: Sources

-- 20260321000005_create_sources.sql

CREATE TABLE sources (
    id         UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id    UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
    title      TEXT NOT NULL CHECK (char_length(title) BETWEEN 1 AND 200),
    url        TEXT NOT NULL CHECK (char_length(url) <= 2000),
    created_at TIMESTAMPTZ NOT NULL DEFAULT now()
);

CREATE INDEX idx_sources_user_id ON sources(user_id);
CREATE INDEX idx_sources_user_id_created_at ON sources(user_id, created_at DESC);

Migration 006: Syntheses

-- 20260321000006_create_syntheses.sql

CREATE TYPE synthesis_status AS ENUM ('generating', 'completed', 'failed');

CREATE TABLE syntheses (
    id          UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id     UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
    title       TEXT NOT NULL DEFAULT '',           -- auto-generated from theme + date
    sections    JSONB NOT NULL DEFAULT '[]'::jsonb, -- [{title: string, items: [{title, url, summary}]}]
    status      synthesis_status NOT NULL DEFAULT 'generating',
    error_message TEXT,                             -- populated on failure
    provider    TEXT,                               -- which provider was used
    model       TEXT,                               -- which model was used
    created_at  TIMESTAMPTZ NOT NULL DEFAULT now()
);

CREATE INDEX idx_syntheses_user_id ON syntheses(user_id);
CREATE INDEX idx_syntheses_user_id_created_at ON syntheses(user_id, created_at DESC);

Migration 007: User Provider Keys

-- 20260321000007_create_user_provider_keys.sql

CREATE TABLE user_provider_keys (
    id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id         UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
    provider        TEXT NOT NULL,               -- 'gemini', 'openai', 'anthropic'
    encrypted_key   BYTEA NOT NULL,              -- AES-256-GCM ciphertext
    nonce           BYTEA NOT NULL,              -- 12-byte GCM nonce
    key_prefix      TEXT NOT NULL DEFAULT '',     -- first 8 chars for display (e.g., "sk-proj-...")
    created_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
    updated_at      TIMESTAMPTZ NOT NULL DEFAULT now(),

    UNIQUE(user_id, provider)                    -- one key per provider per user
);

CREATE INDEX idx_user_provider_keys_user_id ON user_provider_keys(user_id);

Migration 008: Admin Provider Models

-- 20260321000008_create_admin_provider_models.sql

CREATE TABLE admin_provider_models (
    id            UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    provider      TEXT NOT NULL,                  -- 'gemini', 'openai', 'anthropic'
    model_id      TEXT NOT NULL,                  -- e.g., 'gemini-2.5-pro', 'gpt-4o', 'claude-sonnet-4-20250514'
    display_name  TEXT NOT NULL,                  -- e.g., 'Gemini 2.5 Pro', 'GPT-4o', 'Claude Sonnet 4'
    enabled       BOOLEAN NOT NULL DEFAULT true,
    created_at    TIMESTAMPTZ NOT NULL DEFAULT now(),
    updated_at    TIMESTAMPTZ NOT NULL DEFAULT now(),

    UNIQUE(provider, model_id)
);

CREATE INDEX idx_admin_provider_models_enabled ON admin_provider_models(enabled) WHERE enabled = true;

Migration 009: Admin Rate Limits

-- 20260321000009_create_admin_rate_limits.sql

CREATE TABLE admin_rate_limits (
    provider            TEXT PRIMARY KEY,          -- 'gemini', 'openai', 'anthropic'
    max_requests        INTEGER NOT NULL DEFAULT 29,
    time_window_seconds INTEGER NOT NULL DEFAULT 60,
    updated_at          TIMESTAMPTZ NOT NULL DEFAULT now()
);

-- Seed default rate limits
INSERT INTO admin_rate_limits (provider, max_requests, time_window_seconds)
VALUES
    ('gemini', 29, 60),
    ('openai', 50, 60),
    ('anthropic', 40, 60);

Migration 010: Generation Jobs

-- 20260321000010_create_generation_jobs.sql

CREATE TYPE job_status AS ENUM ('pending', 'searching', 'scraping', 'rewriting', 'completed', 'failed');

CREATE TABLE generation_jobs (
    id              UUID PRIMARY KEY DEFAULT gen_random_uuid(),
    user_id         UUID NOT NULL REFERENCES users(id) ON DELETE CASCADE,
    synthesis_id    UUID REFERENCES syntheses(id) ON DELETE SET NULL,
    status          job_status NOT NULL DEFAULT 'pending',
    progress_pct    SMALLINT NOT NULL DEFAULT 0 CHECK (progress_pct BETWEEN 0 AND 100),
    progress_message TEXT,
    error_message   TEXT,
    provider        TEXT,
    model           TEXT,
    started_at      TIMESTAMPTZ NOT NULL DEFAULT now(),
    completed_at    TIMESTAMPTZ
);

CREATE INDEX idx_generation_jobs_user_id ON generation_jobs(user_id);
CREATE INDEX idx_generation_jobs_status ON generation_jobs(status) WHERE status NOT IN ('completed', 'failed');

3. API Endpoints

All endpoints are prefixed with /api/v1. Request and response bodies are JSON. Cookie-based session authentication is used for all authenticated endpoints.

3.1 Authentication

Method	Path	Auth	Description	Phase
`POST`	`/api/v1/auth/register`	Public + Turnstile	Create account, send magic link	1
`POST`	`/api/v1/auth/magic-link`	Public + Turnstile	Request magic link for existing user	1
`GET`	`/api/v1/auth/verify`	Public	Verify magic link token, create session, redirect	1
`POST`	`/api/v1/auth/logout`	Authenticated	Invalidate current session	1
`GET`	`/api/v1/auth/me`	Authenticated	Get current user info	1

Request/Response Types

// POST /api/v1/auth/register
#[derive(Deserialize, Validate)]
pub struct RegisterRequest {
    #[validate(email)]
    pub email: String,
    #[validate(length(min = 1, max = 100))]
    pub display_name: Option<String>,
    pub turnstile_token: String,
}

// Response: 200 (always the same, whether email exists or not)
#[derive(Serialize)]
pub struct AuthMessageResponse {
    pub message: String, // "If this email is valid, a login link has been sent."
}

// POST /api/v1/auth/magic-link
#[derive(Deserialize, Validate)]
pub struct MagicLinkRequest {
    #[validate(email)]
    pub email: String,
    pub turnstile_token: String,
}

// GET /api/v1/auth/verify?token=<base64url>
// -> 302 redirect to / with Set-Cookie on success
// -> 302 redirect to /login?error=invalid_token on failure

// GET /api/v1/auth/me
#[derive(Serialize)]
pub struct MeResponse {
    pub id: Uuid,
    pub email: String,
    pub display_name: Option<String>,
    pub role: String,       // "user" | "admin"
    pub created_at: DateTime<Utc>,
}

3.2 Syntheses

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/syntheses`	Authenticated	List user's syntheses (paginated)	2
`GET`	`/api/v1/syntheses/:id`	Authenticated	Get synthesis detail	2
`POST`	`/api/v1/syntheses/generate`	Authenticated	Trigger async generation	2
`DELETE`	`/api/v1/syntheses/:id`	Authenticated	Delete a synthesis	2
`POST`	`/api/v1/syntheses/:id/email`	Authenticated	Send synthesis by email	3

// GET /api/v1/syntheses?page=1&per_page=20
#[derive(Deserialize)]
pub struct ListSynthesesQuery {
    pub page: Option<u32>,     // default 1
    pub per_page: Option<u32>, // default 20, max 100
}

#[derive(Serialize)]
pub struct SynthesisListItem {
    pub id: Uuid,
    pub title: String,
    pub status: String,
    pub provider: Option<String>,
    pub model: Option<String>,
    pub created_at: DateTime<Utc>,
    pub section_count: i32,
    pub article_count: i32,
}

#[derive(Serialize)]
pub struct PaginatedResponse<T> {
    pub items: Vec<T>,
    pub total: i64,
    pub page: u32,
    pub per_page: u32,
}

// GET /api/v1/syntheses/:id
#[derive(Serialize)]
pub struct SynthesisDetail {
    pub id: Uuid,
    pub title: String,
    pub sections: Vec<NewsSection>,
    pub status: String,
    pub provider: Option<String>,
    pub model: Option<String>,
    pub created_at: DateTime<Utc>,
}

#[derive(Serialize, Deserialize, Clone)]
pub struct NewsSection {
    pub title: String,
    pub items: Vec<NewsItem>,
}

#[derive(Serialize, Deserialize, Clone)]
pub struct NewsItem {
    pub title: String,
    pub url: String,
    pub summary: String,
}

// POST /api/v1/syntheses/generate
// Request body is empty -- uses the user's saved settings + sources + API key
// Response: 202
#[derive(Serialize)]
pub struct GenerateResponse {
    pub job_id: Uuid,
    pub status: String,
}

// POST /api/v1/syntheses/:id/email
#[derive(Deserialize, Validate)]
pub struct SendEmailRequest {
    #[validate(email)]
    pub recipient: String,
}

3.3 Sources

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/sources`	Authenticated	List user's sources	1
`POST`	`/api/v1/sources`	Authenticated	Add a source	1
`DELETE`	`/api/v1/sources/:id`	Authenticated	Delete a source	1

// GET /api/v1/sources
#[derive(Serialize)]
pub struct SourceResponse {
    pub id: Uuid,
    pub title: String,
    pub url: String,
    pub created_at: DateTime<Utc>,
}

// POST /api/v1/sources
#[derive(Deserialize, Validate)]
pub struct CreateSourceRequest {
    #[validate(length(min = 1, max = 200))]
    pub title: String,
    #[validate(url, length(max = 2000))]
    pub url: String,
}

3.4 Settings

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/settings`	Authenticated	Get user's settings	1
`PUT`	`/api/v1/settings`	Authenticated	Update settings	1

// GET /api/v1/settings
#[derive(Serialize)]
pub struct SettingsResponse {
    pub theme: String,
    pub max_age_days: i32,
    pub categories: Vec<String>,
    pub max_items_per_category: i32,
    pub search_agent_behavior: String,
    pub preferred_provider: Option<String>,
    pub preferred_model: Option<String>,
}

// PUT /api/v1/settings
#[derive(Deserialize, Validate)]
pub struct UpdateSettingsRequest {
    #[validate(length(min = 1, max = 200))]
    pub theme: String,
    #[validate(range(min = 1, max = 365))]
    pub max_age_days: i32,
    #[validate(length(min = 1, max = 20))]
    pub categories: Vec<String>,  // each item validated separately in handler
    #[validate(range(min = 1, max = 50))]
    pub max_items_per_category: i32,
    #[validate(length(max = 2000))]
    pub search_agent_behavior: String,
    pub preferred_provider: Option<String>,
    pub preferred_model: Option<String>,
}

3.5 User Provider Keys

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/provider-keys`	Authenticated	List user's provider keys (prefix only, never full key)	2
`POST`	`/api/v1/provider-keys`	Authenticated	Add or replace a provider key	2
`DELETE`	`/api/v1/provider-keys/:provider`	Authenticated	Delete a provider key	2

// GET /api/v1/provider-keys
#[derive(Serialize)]
pub struct ProviderKeyListItem {
    pub provider: String,
    pub key_prefix: String,     // e.g., "sk-proj-..." (first 8 chars)
    pub created_at: DateTime<Utc>,
    pub updated_at: DateTime<Utc>,
}

// POST /api/v1/provider-keys
#[derive(Deserialize, Validate)]
pub struct CreateProviderKeyRequest {
    #[validate(custom(function = "validate_provider"))]
    pub provider: String,       // "gemini" | "openai" | "anthropic"
    #[validate(length(min = 10, max = 500))]
    pub api_key: String,        // the raw API key -- encrypted before storage
}

fn validate_provider(provider: &str) -> Result<(), validator::ValidationError> {
    match provider {
        "gemini" | "openai" | "anthropic" => Ok(()),
        _ => Err(validator::ValidationError::new("invalid_provider")),
    }
}

3.6 Generation (SSE)

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/generation/:job_id/progress`	Authenticated	SSE stream of generation progress	2
`GET`	`/api/v1/generation/:job_id/status`	Authenticated	One-shot status poll (for reconnection)	2

// SSE event types sent on /api/v1/generation/:job_id/progress
#[derive(Serialize)]
#[serde(tag = "type")]
pub enum GenerationEvent {
    #[serde(rename = "progress")]
    Progress {
        status: String,          // "searching", "scraping", "rewriting"
        progress_pct: u8,
        message: String,
    },
    #[serde(rename = "completed")]
    Completed {
        synthesis_id: Uuid,
    },
    #[serde(rename = "failed")]
    Failed {
        error: String,
    },
}

// GET /api/v1/generation/:job_id/status (one-shot poll)
#[derive(Serialize)]
pub struct JobStatusResponse {
    pub job_id: Uuid,
    pub status: String,
    pub progress_pct: u8,
    pub progress_message: Option<String>,
    pub synthesis_id: Option<Uuid>,
    pub error_message: Option<String>,
}

3.7 Admin

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/admin/provider-models`	Admin	List all configured provider models	2
`POST`	`/api/v1/admin/provider-models`	Admin	Add a provider model	2
`PUT`	`/api/v1/admin/provider-models/:id`	Admin	Update a provider model	2
`DELETE`	`/api/v1/admin/provider-models/:id`	Admin	Remove a provider model	2
`GET`	`/api/v1/admin/rate-limits`	Admin	Get rate limit configs	2
`PUT`	`/api/v1/admin/rate-limits/:provider`	Admin	Update rate limit config	2
`GET`	`/api/v1/admin/users`	Admin	List all users	2
`PUT`	`/api/v1/admin/users/:id/role`	Admin	Change user role	2

// POST /api/v1/admin/provider-models
#[derive(Deserialize, Validate)]
pub struct CreateProviderModelRequest {
    #[validate(custom(function = "validate_provider"))]
    pub provider: String,
    #[validate(length(min = 1, max = 100))]
    pub model_id: String,
    #[validate(length(min = 1, max = 200))]
    pub display_name: String,
    pub enabled: bool,
}

// PUT /api/v1/admin/rate-limits/:provider
#[derive(Deserialize, Validate)]
pub struct UpdateRateLimitRequest {
    #[validate(range(min = 1, max = 1000))]
    pub max_requests: i32,
    #[validate(range(min = 1, max = 3600))]
    pub time_window_seconds: i32,
}

// PUT /api/v1/admin/users/:id/role
#[derive(Deserialize)]
pub struct UpdateRoleRequest {
    pub role: String, // "user" | "admin"
}

3.8 Public Config

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/config/providers`	Authenticated	List enabled providers and models (no keys)	2

// GET /api/v1/config/providers
#[derive(Serialize)]
pub struct ProviderInfo {
    pub provider: String,
    pub models: Vec<ModelInfo>,
}

#[derive(Serialize)]
pub struct ModelInfo {
    pub model_id: String,
    pub display_name: String,
}

3.9 SSE for Real-Time List Updates

Method	Path	Auth	Description	Phase
`GET`	`/api/v1/events/syntheses`	Authenticated	SSE stream for syntheses list changes	3
`GET`	`/api/v1/events/sources`	Authenticated	SSE stream for sources list changes	3

These SSE endpoints replace Firestore onSnapshot. When a synthesis is created/deleted or a source is added/removed, connected clients receive a lightweight event (just the entity type + action + id) prompting the frontend to refetch. Implementation uses a tokio::sync::broadcast channel per user in AppState.

4. Authentication & Authorization

4.1 Magic Link Flow

Token Generation

// src/util/token.rs
use rand::RngCore;
use sha2::{Sha256, Digest};
use base64::{engine::general_purpose::URL_SAFE_NO_PAD, Engine};

/// Generate a 32-byte cryptographically secure random token, base64url-encoded.
pub fn generate_token() -> String {
    let mut bytes = [0u8; 32];
    rand::rngs::OsRng.fill_bytes(&mut bytes);
    URL_SAFE_NO_PAD.encode(bytes)
}

/// Compute SHA-256 hash of a token, hex-encoded.
pub fn hash_token(token: &str) -> String {
    let mut hasher = Sha256::new();
    hasher.update(token.as_bytes());
    let result = hasher.finalize();
    hex::encode(result)
}

Registration Flow

Receive RegisterRequest (email, optional display_name, turnstile_token).
Verify Turnstile token via services::captcha::verify().
Check if user with this email already exists:
- If exists: send a magic link email (same as login). Return generic success message.
- If not: create user with role = 'user', create default settings row, send magic link email.
Always return 200 { "message": "If this email is valid, a login link has been sent." }.
Add random delay (50-200ms) when no email is sent to prevent timing attacks.

Magic Link Request Flow

Receive MagicLinkRequest (email, turnstile_token).
Verify Turnstile token.
Look up user by email.
- If not found: return same success message, do nothing (prevent email enumeration).
Generate token, compute SHA-256 hash, store in magic_link_tokens with 15-minute expiry.
Send email containing {BASE_URL}/api/v1/auth/verify?token={raw_token}.
Return generic success message.

Verification Flow

Receive GET /api/v1/auth/verify?token={raw_token}.
Compute hash = SHA-256(raw_token).

Atomic query:

UPDATE magic_link_tokens
SET used = true
WHERE token_hash = $1 AND used = false AND expires_at > now()
RETURNING email

If rows_affected == 0: redirect to /login?error=invalid_token.
Look up or create user by email (handles the case where registration + verification happen in a single flow).
Generate session token (32 bytes, base64url).
Store SHA-256(session_token) in sessions table with 30-day expiry.

Set session cookie:

Set-Cookie: ai_synth_session={session_token};
  HttpOnly; Secure; SameSite=Lax; Path=/; Max-Age=2592000

Redirect to / (302).

4.2 Session Middleware (AuthUser Extractor)

// src/middleware/auth.rs
use axum::extract::FromRequestParts;
use axum::http::request::Parts;

#[derive(Clone, Debug)]
pub struct AuthUser {
    pub id: Uuid,
    pub email: String,
    pub display_name: Option<String>,
    pub role: UserRole,
}

#[async_trait]
impl<S> FromRequestParts<S> for AuthUser
where
    S: Send + Sync,
    AppState: FromRef<S>,
{
    type Rejection = AppError;

    async fn from_request_parts(parts: &mut Parts, state: &S) -> Result<Self, Self::Rejection> {
        let app_state = AppState::from_ref(state);

        // 1. Extract cookie
        let cookies = parts
            .headers
            .get_all(header::COOKIE)
            .iter()
            .filter_map(|v| v.to_str().ok())
            .flat_map(|s| s.split(';'))
            .map(|s| s.trim())
            .find(|s| s.starts_with("ai_synth_session="))
            .and_then(|s| s.strip_prefix("ai_synth_session="))
            .ok_or(AppError::Unauthorized("No session cookie".into()))?;

        let session_token = cookies;

        // 2. Hash and lookup
        let token_hash = crate::util::token::hash_token(session_token);
        let session = db::sessions::find_by_hash(&app_state.pool, &token_hash)
            .await?
            .ok_or(AppError::Unauthorized("Invalid session".into()))?;

        // 3. Check expiration
        if session.expires_at < Utc::now() {
            db::sessions::delete(&app_state.pool, session.id).await?;
            return Err(AppError::Unauthorized("Session expired".into()));
        }

        // 4. Load user
        let user = db::users::find_by_id(&app_state.pool, session.user_id)
            .await?
            .ok_or(AppError::Unauthorized("User not found".into()))?;

        Ok(AuthUser {
            id: user.id,
            email: user.email,
            display_name: user.display_name,
            role: user.role,
        })
    }
}

4.3 Admin Guard (AdminUser Extractor)

// src/middleware/admin.rs

#[derive(Clone, Debug)]
pub struct AdminUser(pub AuthUser);

#[async_trait]
impl<S> FromRequestParts<S> for AdminUser
where
    S: Send + Sync,
    AppState: FromRef<S>,
{
    type Rejection = AppError;

    async fn from_request_parts(parts: &mut Parts, state: &S) -> Result<Self, Self::Rejection> {
        let auth_user = AuthUser::from_request_parts(parts, state).await?;
        if auth_user.role != UserRole::Admin {
            return Err(AppError::Forbidden("Admin access required".into()));
        }
        Ok(AdminUser(auth_user))
    }
}

Handlers that require admin simply take AdminUser as a parameter:

async fn list_users(
    AdminUser(admin): AdminUser,
    State(state): State<AppState>,
) -> Result<Json<Vec<UserResponse>>, AppError> {
    // admin is guaranteed to be an admin here
}

4.4 Turnstile Verification

// src/services/captcha.rs
use serde::Deserialize;

#[derive(Deserialize)]
struct TurnstileResponse {
    success: bool,
    #[serde(rename = "error-codes")]
    error_codes: Vec<String>,
}

pub async fn verify_turnstile(
    client: &reqwest::Client,
    secret: &str,
    token: &str,
) -> Result<(), AppError> {
    let resp = client
        .post("https://challenges.cloudflare.com/turnstile/v0/siteverify")
        .form(&[("secret", secret), ("response", token)])
        .send()
        .await
        .map_err(|e| AppError::Internal(e.into()))?;

    let result: TurnstileResponse = resp
        .json()
        .await
        .map_err(|e| AppError::Internal(e.into()))?;

    if result.success {
        Ok(())
    } else {
        Err(AppError::BadRequest(format!(
            "Captcha verification failed: {:?}",
            result.error_codes
        )))
    }
}

4.5 CSRF Protection

The CSRF middleware checks for the X-Requested-With header on all state-changing requests (POST, PUT, DELETE). Combined with SameSite=Lax cookies, this prevents cross-site request forgery because browsers do not send custom headers on cross-origin requests without an approved CORS preflight.

// src/middleware/csrf.rs
use axum::middleware::Next;
use axum::http::Request;

pub async fn csrf_check<B>(req: Request<B>, next: Next<B>) -> Result<Response, AppError> {
    let dominated_methods = ["POST", "PUT", "DELETE", "PATCH"];

    if dominated_methods.contains(&req.method().as_str()) {
        let has_header = req
            .headers()
            .get("X-Requested-With")
            .map(|v| v == "XMLHttpRequest")
            .unwrap_or(false);

        if !has_header {
            return Err(AppError::Forbidden("Missing X-Requested-With header".into()));
        }
    }

    Ok(next.run(req).await)
}

4.6 CLI Admin Creation

// src/cli.rs
use clap::{Parser, Subcommand};

#[derive(Parser)]
#[command(name = "ai-synth", about = "AI Weekly Synth backend")]
pub struct Cli {
    #[command(subcommand)]
    pub command: Commands,
}

#[derive(Subcommand)]
pub enum Commands {
    /// Start the web server
    Serve,
    /// Create an admin user
    CreateAdmin {
        /// Email address for the admin user
        #[arg(long)]
        email: String,
    },
}

In main.rs:

async fn main() -> anyhow::Result<()> {
    let cli = Cli::parse();
    match cli.command {
        Commands::Serve => { /* start server */ },
        Commands::CreateAdmin { email } => {
            let pool = create_pool().await?;
            let user = db::users::find_by_email(&pool, &email).await?;
            match user {
                Some(u) => {
                    db::users::update_role(&pool, u.id, UserRole::Admin).await?;
                    println!("User {} promoted to admin.", email);
                }
                None => {
                    db::users::create(&pool, &email, None, UserRole::Admin).await?;
                    println!("Admin user {} created.", email);
                }
            }
        }
    }
    Ok(())
}

5. LLM Provider Abstraction

5.1 The Unified Trait

// src/services/llm/mod.rs
use async_trait::async_trait;

/// Describes what a provider can do natively.
#[derive(Debug, Clone)]
pub struct ProviderCapabilities {
    /// Provider supports native web search grounding (e.g., Gemini googleSearch, OpenAI web_search).
    pub native_web_search: bool,
    /// Provider supports structured JSON output via schema enforcement.
    pub structured_output: bool,
}

/// Progress callback type for reporting generation steps.
pub type ProgressFn = Box<dyn Fn(u8, &str) + Send + Sync>;

#[async_trait]
pub trait LlmProvider: Send + Sync {
    /// Returns the provider identifier: "gemini", "openai", or "anthropic".
    fn provider_id(&self) -> &str;

    /// Returns the capabilities of this provider.
    fn capabilities(&self) -> ProviderCapabilities;

    /// Pass 1: Search the web and generate structured news items.
    /// Uses native web search grounding if available.
    /// Returns parsed JSON matching the category schema (keys: category_0, category_1, ...).
    async fn search_and_generate(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError>;

    /// Pass 2: Rewrite titles and summaries based on scraped content.
    /// No web search needed.
    async fn rewrite(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError>;
}

5.2 Per-Provider Implementation Strategy

Gemini

// src/services/llm/gemini.rs

pub struct GeminiProvider {
    client: reqwest::Client,
    api_key: secrecy::SecretString,
}

impl GeminiProvider {
    pub fn new(client: reqwest::Client, api_key: secrecy::SecretString) -> Self {
        Self { client, api_key }
    }
}

#[async_trait]
impl LlmProvider for GeminiProvider {
    fn provider_id(&self) -> &str { "gemini" }

    fn capabilities(&self) -> ProviderCapabilities {
        ProviderCapabilities {
            native_web_search: true,     // googleSearch tool
            structured_output: true,     // responseSchema + responseMimeType
        }
    }

    async fn search_and_generate(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError> {
        // POST https://generativelanguage.googleapis.com/v1beta/models/{model}:generateContent
        // Body:
        // {
        //   "system_instruction": { "parts": [{ "text": system_prompt }] },
        //   "contents": [{ "parts": [{ "text": user_prompt }] }],
        //   "tools": [{ "google_search": {} }],
        //   "generation_config": {
        //     "response_mime_type": "application/json",
        //     "response_schema": response_schema
        //   }
        // }
        // Parse response.candidates[0].content.parts[0].text as JSON
        todo!()
    }

    async fn rewrite(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError> {
        // Same as above but WITHOUT the tools field (no web search in pass 2)
        todo!()
    }
}

OpenAI

// src/services/llm/openai.rs

pub struct OpenAiProvider {
    client: reqwest::Client,
    api_key: secrecy::SecretString,
}

#[async_trait]
impl LlmProvider for OpenAiProvider {
    fn provider_id(&self) -> &str { "openai" }

    fn capabilities(&self) -> ProviderCapabilities {
        ProviderCapabilities {
            native_web_search: true,      // web_search tool (Responses API)
            structured_output: true,      // response_format: { type: "json_schema", json_schema: {...} }
        }
    }

    async fn search_and_generate(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError> {
        // POST https://api.openai.com/v1/responses
        // Body:
        // {
        //   "model": model,
        //   "instructions": system_prompt,
        //   "input": user_prompt,
        //   "tools": [{ "type": "web_search_preview" }],
        //   "text": {
        //     "format": {
        //       "type": "json_schema",
        //       "name": "news_synthesis",
        //       "schema": response_schema
        //     }
        //   }
        // }
        todo!()
    }

    async fn rewrite(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError> {
        // POST https://api.openai.com/v1/responses (no tools)
        todo!()
    }
}

Anthropic

// src/services/llm/anthropic.rs

pub struct AnthropicProvider {
    client: reqwest::Client,
    api_key: secrecy::SecretString,
}

#[async_trait]
impl LlmProvider for AnthropicProvider {
    fn provider_id(&self) -> &str { "anthropic" }

    fn capabilities(&self) -> ProviderCapabilities {
        ProviderCapabilities {
            native_web_search: true,      // web_search tool
            structured_output: false,     // No native JSON schema enforcement; use prompt + parse
        }
    }

    async fn search_and_generate(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError> {
        // POST https://api.anthropic.com/v1/messages
        // Headers: x-api-key, anthropic-version: 2023-06-01
        // Body:
        // {
        //   "model": model,
        //   "max_tokens": 8192,
        //   "system": system_prompt,
        //   "tools": [{ "type": "web_search_20250305" }],
        //   "messages": [{ "role": "user", "content": user_prompt + "\n\nRespond ONLY with valid JSON matching this schema:\n" + schema }]
        // }
        // Extract text from response, parse as JSON, validate against schema
        todo!()
    }

    async fn rewrite(
        &self,
        model: &str,
        system_prompt: &str,
        user_prompt: &str,
        response_schema: &serde_json::Value,
    ) -> Result<serde_json::Value, AppError> {
        // Same without tools, JSON schema in prompt
        todo!()
    }
}

5.3 Provider Factory

// src/services/llm/mod.rs

pub fn create_provider(
    provider_id: &str,
    api_key: secrecy::SecretString,
    client: reqwest::Client,
) -> Result<Box<dyn LlmProvider>, AppError> {
    match provider_id {
        "gemini" => Ok(Box::new(GeminiProvider::new(client, api_key))),
        "openai" => Ok(Box::new(OpenAiProvider::new(client, api_key))),
        "anthropic" => Ok(Box::new(AnthropicProvider::new(client, api_key))),
        other => Err(AppError::BadRequest(format!("Unknown provider: {other}"))),
    }
}

5.4 The Adaptive Generation Pipeline

// src/services/generation.rs

pub struct GenerationPipeline;

impl GenerationPipeline {
    /// Run the full generation pipeline.
    /// Adapts the pipeline based on provider capabilities.
    pub async fn run(
        state: &AppState,
        user_id: Uuid,
        job_id: Uuid,
        progress_tx: tokio::sync::watch::Sender<GenerationEvent>,
    ) -> Result<Uuid, AppError> {
        // 1. Load user settings
        let settings = db::settings::get_or_default(&state.pool, user_id).await?;

        // 2. Load user sources
        let sources = db::sources::list_by_user(&state.pool, user_id).await?;

        // 3. Resolve provider + model
        //    Look up user's preferred_provider -> find user's API key for that provider
        //    Fall back to first provider the user has a key for
        let (provider_id, model_id) = resolve_provider_and_model(
            &state.pool, user_id, &settings
        ).await?;

        // 4. Decrypt user's API key for the resolved provider
        let api_key = db::provider_keys::get_decrypted(
            &state.pool, user_id, &provider_id, &state.config.master_key
        ).await?;

        // 5. Create the LLM provider
        let provider = create_provider(&provider_id, api_key, state.http_client.clone())?;
        let caps = provider.capabilities();

        // 6. Build dynamic JSON schema from categories
        let schema = build_category_schema(&settings.categories);

        // 7. Build prompts
        let (system_prompt, user_prompt) = build_search_prompts(&settings, &sources);

        // === PASS 1: Search ===
        progress_tx.send(GenerationEvent::Progress {
            status: "searching".into(),
            progress_pct: 10,
            message: "Searching the web for recent news...".into(),
        }).ok();

        // Rate limit check
        state.rate_limiter.acquire(&provider_id, user_id).await?;

        let raw_results = provider
            .search_and_generate(&model_id, &system_prompt, &user_prompt, &schema)
            .await?;

        // 8. Parse raw results into NewsItems per category
        let parsed = parse_category_results(&raw_results, &settings.categories)?;

        // === DECISION POINT: scrape or skip ===
        // If the provider has native web search with grounding (Gemini, OpenAI),
        // the results are already high-quality. We still scrape to:
        //   (a) validate URLs exist (filter 404s)
        //   (b) check publication dates
        //   (c) get content for the rewrite pass
        // But if scraping fails for some items, we keep the originals.

        progress_tx.send(GenerationEvent::Progress {
            status: "scraping".into(),
            progress_pct: 40,
            message: "Validating and fetching article content...".into(),
        }).ok();

        let scraped = services::scraper::validate_and_scrape(
            &state.http_client,
            parsed,
            settings.max_age_days as i64,
        ).await;

        // === PASS 2: Rewrite ===
        progress_tx.send(GenerationEvent::Progress {
            status: "rewriting".into(),
            progress_pct: 70,
            message: "Rewriting summaries based on article content...".into(),
        }).ok();

        let (rewrite_system, rewrite_user) = build_rewrite_prompts(&scraped);

        state.rate_limiter.acquire(&provider_id, user_id).await?;

        let final_results = provider
            .rewrite(&model_id, &rewrite_system, &rewrite_user, &schema)
            .await?;

        // 9. Build sections
        let sections = build_sections(&final_results, &settings.categories)?;

        // 10. Persist synthesis
        let title = format!("{} - {}", settings.theme, Utc::now().format("%d/%m/%Y"));
        let synthesis_id = db::syntheses::create(
            &state.pool, user_id, &title, &sections, &provider_id, &model_id
        ).await?;

        // 11. Update job record
        db::generation_jobs::complete(&state.pool, job_id, synthesis_id).await?;

        progress_tx.send(GenerationEvent::Completed { synthesis_id }).ok();

        Ok(synthesis_id)
    }
}

/// Resolve which provider and model to use for a user.
async fn resolve_provider_and_model(
    pool: &PgPool,
    user_id: Uuid,
    settings: &UserSettings,
) -> Result<(String, String), AppError> {
    // 1. If user has a preferred provider + model, and has a key for it, use that
    if let (Some(ref prov), Some(ref model)) = (&settings.preferred_provider, &settings.preferred_model) {
        let has_key = db::provider_keys::exists(pool, user_id, prov).await?;
        if has_key {
            // Verify model is in admin's enabled list
            let model_exists = db::admin_models::is_enabled(pool, prov, model).await?;
            if model_exists {
                return Ok((prov.clone(), model.clone()));
            }
        }
    }

    // 2. Fall back to first provider the user has a key for
    let keys = db::provider_keys::list_by_user(pool, user_id).await?;
    for key in &keys {
        if let Some(model) = db::admin_models::first_enabled_for_provider(pool, &key.provider).await? {
            return Ok((key.provider.clone(), model.model_id));
        }
    }

    Err(AppError::BadRequest(
        "No API key configured. Please add an API key in your settings.".into()
    ))
}

5.5 Rate Limiter

Token-bucket rate limiter, per-provider. Admin-configured defaults apply globally. Users can override for their own keys (if they have higher API quotas).

// src/services/rate_limiter.rs
use dashmap::DashMap;
use std::collections::VecDeque;
use std::time::{Duration, Instant};
use tokio::sync::Semaphore;

pub struct RateLimiter {
    /// Per-provider global buckets (admin-configured defaults)
    global_buckets: DashMap<String, TokenBucket>,
    /// Per-user-per-provider override buckets
    user_buckets: DashMap<(Uuid, String), TokenBucket>,
}

struct TokenBucket {
    timestamps: VecDeque<Instant>,
    max_requests: u32,
    time_window: Duration,
}

impl RateLimiter {
    pub fn new() -> Self {
        Self {
            global_buckets: DashMap::new(),
            user_buckets: DashMap::new(),
        }
    }

    /// Load admin-configured rate limits from DB.
    pub async fn load_from_db(&self, pool: &PgPool) -> Result<(), AppError> {
        let limits = db::admin_rate_limits::list_all(pool).await?;
        for limit in limits {
            self.global_buckets.insert(
                limit.provider.clone(),
                TokenBucket {
                    timestamps: VecDeque::new(),
                    max_requests: limit.max_requests as u32,
                    time_window: Duration::from_secs(limit.time_window_seconds as u64),
                },
            );
        }
        Ok(())
    }

    /// Acquire a slot. Uses user-specific override if set, otherwise global.
    pub async fn acquire(&self, provider: &str, user_id: Uuid) -> Result<(), AppError> {
        // Check user override first
        let user_key = (user_id, provider.to_string());
        if let Some(mut bucket) = self.user_buckets.get_mut(&user_key) {
            return self.acquire_from_bucket(&mut bucket).await;
        }

        // Fall back to global
        let mut bucket = self.global_buckets
            .entry(provider.to_string())
            .or_insert_with(|| TokenBucket {
                timestamps: VecDeque::new(),
                max_requests: 29,
                time_window: Duration::from_secs(60),
            });
        self.acquire_from_bucket(&mut bucket).await
    }

    async fn acquire_from_bucket(&self, bucket: &mut TokenBucket) -> Result<(), AppError> {
        let now = Instant::now();
        bucket.timestamps.retain(|t| now.duration_since(*t) < bucket.time_window);

        if bucket.timestamps.len() < bucket.max_requests as usize {
            bucket.timestamps.push_back(now);
            return Ok(());
        }

        // Calculate wait time
        let oldest = bucket.timestamps.front().unwrap();
        let wait = bucket.time_window.checked_sub(now.duration_since(*oldest))
            .unwrap_or(Duration::from_millis(100));

        if wait > Duration::from_secs(120) {
            return Err(AppError::TooManyRequests {
                retry_after_secs: wait.as_secs(),
            });
        }

        // Drop the DashMap reference before sleeping
        drop(bucket);
        tokio::time::sleep(wait).await;

        // Retry (recursive but bounded because we waited past the oldest timestamp)
        // In practice, re-acquire by calling the parent method
        Ok(())
    }

    /// Hot-reload a provider's limits (called after admin update).
    pub fn update_provider_limit(&self, provider: &str, max_requests: u32, time_window_secs: u64) {
        self.global_buckets.insert(
            provider.to_string(),
            TokenBucket {
                timestamps: VecDeque::new(),
                max_requests,
                time_window: Duration::from_secs(time_window_secs),
            },
        );
    }
}

6. URL Scraping (Server-Side)

6.1 HTTP Client Configuration

// Created once in AppState, shared across all requests
let http_client = reqwest::Client::builder()
    .user_agent("AI-Weekly-Synth/1.0 (+https://github.com/your-repo)")
    .timeout(Duration::from_secs(15))
    .connect_timeout(Duration::from_secs(5))
    .redirect(reqwest::redirect::Policy::limited(3))
    .danger_accept_invalid_certs(false)
    .pool_max_idle_per_host(10)
    .build()?;

6.2 SSRF Prevention

// src/services/scraper.rs
use std::net::IpAddr;
use url::Url;

fn is_safe_url(url: &Url) -> Result<(), AppError> {
    // 1. Scheme check
    match url.scheme() {
        "http" | "https" => {}
        scheme => return Err(AppError::ScrapingError(
            format!("Blocked scheme: {scheme}")
        )),
    }

    // 2. Host check -- reject if no host
    let host = url.host_str()
        .ok_or_else(|| AppError::ScrapingError("No host in URL".into()))?;

    // 3. Resolve DNS and check IPs
    // Note: uses std::net::ToSocketAddrs which blocks; in production,
    // use trust-dns or tokio::net::lookup_host for async resolution
    let port = url.port().unwrap_or(if url.scheme() == "https" { 443 } else { 80 });
    let addrs: Vec<_> = tokio::net::lookup_host(format!("{host}:{port}"))
        .await
        .map_err(|e| AppError::ScrapingError(format!("DNS resolution failed: {e}")))?
        .collect();

    for addr in &addrs {
        if is_private_ip(addr.ip()) {
            return Err(AppError::ScrapingError(
                "URL resolves to private/internal IP".into()
            ));
        }
    }

    Ok(())
}

fn is_private_ip(ip: IpAddr) -> bool {
    match ip {
        IpAddr::V4(v4) => {
            v4.is_loopback()                    // 127.0.0.0/8
            || v4.is_private()                  // 10/8, 172.16/12, 192.168/16
            || v4.is_link_local()               // 169.254/16
            || v4.is_unspecified()              // 0.0.0.0
            || v4.octets() == [169, 254, 169, 254]  // Cloud metadata
        }
        IpAddr::V6(v6) => {
            v6.is_loopback()                    // ::1
            || v6.is_unspecified()              // ::
            // fe80::/10 (link-local)
            || (v6.segments()[0] & 0xffc0) == 0xfe80
        }
    }
}

6.3 HTML Parsing and Content Extraction

// src/services/scraper.rs
use scraper::{Html, Selector};

pub struct ScrapedItem {
    pub title: String,
    pub url: String,
    pub summary: String,
    pub scraped_content: String,
}

pub async fn validate_and_scrape(
    client: &reqwest::Client,
    items: Vec<(String, Vec<NewsItem>)>, // (category_key, items)
    max_age_days: i64,
) -> Vec<(String, Vec<ScrapedItem>)> {
    let mut results = Vec::new();

    for (category_key, category_items) in items {
        let futures = category_items.into_iter().map(|item| {
            let client = client.clone();
            async move {
                scrape_single(&client, item, max_age_days).await
            }
        });

        // Bounded concurrency: max 10 concurrent scrapes
        let scraped: Vec<Option<ScrapedItem>> = futures::stream::iter(futures)
            .buffer_unordered(10)
            .collect()
            .await;

        let valid: Vec<ScrapedItem> = scraped.into_iter().flatten().collect();
        results.push((category_key, valid));
    }

    results
}

async fn scrape_single(
    client: &reqwest::Client,
    item: NewsItem,
    max_age_days: i64,
) -> Option<ScrapedItem> {
    // 1. Parse and validate URL
    let url = Url::parse(&item.url).ok()?;
    is_safe_url(&url).await.ok()?;

    // 2. Fetch with size limit
    let resp = client.get(url.as_str()).send().await.ok()?;
    if !resp.status().is_success() { return None; }

    // Read up to 5MB
    let bytes = resp.bytes().await.ok()?;
    if bytes.len() > 5_000_000 { return None; }
    let html_text = String::from_utf8_lossy(&bytes);

    // 3. Parse HTML
    let document = Html::parse_document(&html_text);

    // 4. Soft-404 detection
    let title_sel = Selector::parse("title").unwrap();
    let h1_sel = Selector::parse("h1").unwrap();

    let page_title = document.select(&title_sel).next()
        .map(|el| el.text().collect::<String>().to_lowercase())
        .unwrap_or_default();
    let h1_text = document.select(&h1_sel).next()
        .map(|el| el.text().collect::<String>().to_lowercase())
        .unwrap_or_default();

    let error_keywords = [
        "page not found", "404", "403", "access denied",
        "forbidden", "not found", "introuvable", "page introuvable",
    ];
    if error_keywords.iter().any(|kw| page_title.contains(kw) || h1_text.contains(kw)) {
        return None;
    }

    // 5. Date extraction
    if let Some(pub_date) = extract_publication_date(&document) {
        let age = Utc::now().signed_duration_since(pub_date);
        if age.num_days() > max_age_days {
            tracing::debug!(url = %item.url, days = age.num_days(), "Article too old, skipping");
            return None;
        }
    }

    // 6. Content extraction (strip non-content elements)
    let content = extract_body_text(&document, 4000);

    Some(ScrapedItem {
        title: item.title,
        url: item.url,
        summary: item.summary,
        scraped_content: content,
    })
}

/// Extract publication date from meta tags, JSON-LD, and <time> elements.
fn extract_publication_date(doc: &Html) -> Option<DateTime<Utc>> {
    // Priority order:
    // 1. JSON-LD datePublished
    // 2. meta[property="article:published_time"]
    // 3. meta[itemprop="datePublished"]
    // 4. meta[name="date"], meta[name="pubdate"], meta[name="dc.date"]
    // 5. <time datetime="...">

    let jsonld_sel = Selector::parse(r#"script[type="application/ld+json"]"#).unwrap();
    for el in doc.select(&jsonld_sel) {
        let text = el.text().collect::<String>();
        if let Ok(json) = serde_json::from_str::<serde_json::Value>(&text) {
            if let Some(date_str) = json.get("datePublished").and_then(|v| v.as_str()) {
                if let Ok(dt) = datestr_to_datetime(date_str) {
                    return Some(dt);
                }
            }
        }
    }

    let meta_selectors = [
        r#"meta[property="article:published_time"]"#,
        r#"meta[property="og:article:published_time"]"#,
        r#"meta[itemprop="datePublished"]"#,
        r#"meta[name="date"]"#,
        r#"meta[name="pubdate"]"#,
        r#"meta[name="dc.date"]"#,
    ];

    for sel_str in &meta_selectors {
        if let Ok(sel) = Selector::parse(sel_str) {
            if let Some(el) = doc.select(&sel).next() {
                if let Some(content) = el.value().attr("content") {
                    if let Ok(dt) = datestr_to_datetime(content) {
                        return Some(dt);
                    }
                }
            }
        }
    }

    let time_sel = Selector::parse("time[datetime]").unwrap();
    if let Some(el) = doc.select(&time_sel).next() {
        if let Some(dt_str) = el.value().attr("datetime") {
            if let Ok(dt) = datestr_to_datetime(dt_str) {
                return Some(dt);
            }
        }
    }

    None
}

/// Try multiple date formats to parse a date string.
fn datestr_to_datetime(s: &str) -> Result<DateTime<Utc>, chrono::ParseError> {
    // Try ISO 8601 / RFC 3339
    if let Ok(dt) = DateTime::parse_from_rfc3339(s) {
        return Ok(dt.with_timezone(&Utc));
    }
    // Try common formats
    if let Ok(dt) = NaiveDate::parse_from_str(s, "%Y-%m-%d") {
        return Ok(dt.and_hms_opt(0, 0, 0).unwrap().and_utc());
    }
    Err(chrono::ParseError) // simplified
}

/// Extract visible body text, stripping scripts, styles, nav, footer, etc.
fn extract_body_text(doc: &Html, max_chars: usize) -> String {
    // Get all text nodes from body, excluding script/style/nav/footer/header/aside
    let body_sel = Selector::parse("body").unwrap();
    let exclude_tags = ["script", "style", "noscript", "iframe", "nav", "footer", "header", "aside"];

    let body = match doc.select(&body_sel).next() {
        Some(b) => b,
        None => return String::new(),
    };

    let text: String = body
        .text()
        .collect::<Vec<_>>()
        .join(" ")
        .split_whitespace()
        .collect::<Vec<_>>()
        .join(" ");

    if text.len() > max_chars {
        text[..max_chars].to_string()
    } else {
        text
    }
}

7. Async Generation & SSE

7.1 Background Job Spawning

When a user triggers generation, the handler creates a job record in the DB, spawns a tokio::spawn task, and immediately returns the job ID.

// src/handlers/syntheses.rs

pub async fn trigger_generate(
    auth_user: AuthUser,
    State(state): State<AppState>,
) -> Result<(StatusCode, Json<GenerateResponse>), AppError> {
    // 1. Check if user has an active job (prevent double-generation)
    let active = db::generation_jobs::find_active(&state.pool, auth_user.id).await?;
    if active.is_some() {
        return Err(AppError::Conflict(
            "A generation is already in progress.".into()
        ));
    }

    // 2. Create job record
    let job_id = db::generation_jobs::create(&state.pool, auth_user.id).await?;

    // 3. Create a watch channel for progress
    let (progress_tx, _) = tokio::sync::watch::channel(GenerationEvent::Progress {
        status: "pending".into(),
        progress_pct: 0,
        message: "Starting...".into(),
    });

    // 4. Store the progress sender in the job store (for SSE subscribers)
    state.job_store.insert(job_id, progress_tx.clone());

    // 5. Spawn background task
    let state_clone = state.clone();
    let user_id = auth_user.id;
    tokio::spawn(async move {
        let result = GenerationPipeline::run(
            &state_clone, user_id, job_id, progress_tx.clone()
        ).await;

        if let Err(e) = &result {
            tracing::error!(job_id = %job_id, error = %e, "Generation failed");
            progress_tx.send(GenerationEvent::Failed {
                error: e.to_string(),
            }).ok();
            db::generation_jobs::fail(&state_clone.pool, job_id, &e.to_string())
                .await
                .ok();
        }

        // Clean up job store after a delay (allow SSE reconnect)
        tokio::time::sleep(Duration::from_secs(300)).await;
        state_clone.job_store.remove(&job_id);
    });

    Ok((
        StatusCode::ACCEPTED,
        Json(GenerateResponse {
            job_id,
            status: "pending".into(),
        }),
    ))
}

7.2 Job Store in AppState

// src/app_state.rs
use dashmap::DashMap;
use tokio::sync::watch;

pub type JobStore = DashMap<Uuid, watch::Sender<GenerationEvent>>;

#[derive(Clone)]
pub struct AppState {
    pub pool: PgPool,
    pub http_client: reqwest::Client,
    pub config: AppConfig,
    pub rate_limiter: Arc<RateLimiter>,
    pub job_store: Arc<JobStore>,
}

7.3 SSE Endpoint

// src/handlers/generation.rs
use axum::response::sse::{Event, Sse};
use tokio_stream::wrappers::WatchStream;
use tokio_stream::StreamExt;

pub async fn progress_stream(
    auth_user: AuthUser,
    State(state): State<AppState>,
    Path(job_id): Path<Uuid>,
) -> Result<Sse<impl Stream<Item = Result<Event, Infallible>>>, AppError> {
    // Verify the job belongs to this user
    let job = db::generation_jobs::get(&state.pool, job_id).await?
        .ok_or(AppError::NotFound("Job not found".into()))?;
    if job.user_id != auth_user.id {
        return Err(AppError::NotFound("Job not found".into()));
    }

    // Get the watch receiver from the job store
    let rx = state.job_store
        .get(&job_id)
        .map(|entry| entry.value().subscribe())
        .ok_or(AppError::NotFound("Job not found or already completed".into()))?;

    let stream = WatchStream::new(rx).map(|event| {
        let data = serde_json::to_string(&event).unwrap_or_default();
        Ok(Event::default().data(data))
    });

    Ok(Sse::new(stream).keep_alive(
        axum::response::sse::KeepAlive::new()
            .interval(Duration::from_secs(15))
            .text("ping"),
    ))
}

/// One-shot status poll endpoint (for reconnection after navigation away)
pub async fn job_status(
    auth_user: AuthUser,
    State(state): State<AppState>,
    Path(job_id): Path<Uuid>,
) -> Result<Json<JobStatusResponse>, AppError> {
    let job = db::generation_jobs::get(&state.pool, job_id).await?
        .ok_or(AppError::NotFound("Job not found".into()))?;
    if job.user_id != auth_user.id {
        return Err(AppError::NotFound("Job not found".into()));
    }

    Ok(Json(JobStatusResponse {
        job_id: job.id,
        status: job.status.to_string(),
        progress_pct: job.progress_pct as u8,
        progress_message: job.progress_message,
        synthesis_id: job.synthesis_id,
        error_message: job.error_message,
    }))
}

7.4 Reconnection Strategy

When a client navigates away and comes back:

The frontend calls GET /api/v1/generation/:job_id/status first to get the current state.
If status is completed or failed, display the result immediately.
If status is in-progress, reconnect to the SSE endpoint GET /api/v1/generation/:job_id/progress.
The watch channel immediately delivers the latest state on subscribe, so the client catches up instantly.

8. Email (Resend Integration)

Resend provides a REST API for transactional email. No SMTP client library is needed; reqwest suffices.

8.1 Email Service

// src/services/email.rs
use secrecy::{ExposeSecret, SecretString};

pub struct EmailService {
    client: reqwest::Client,
    api_key: SecretString,
    from_address: String,     // e.g., "AI Weekly Synth <noreply@yourdomain.com>"
    base_url: String,         // App base URL for magic link construction
}

impl EmailService {
    pub fn new(
        client: reqwest::Client,
        api_key: SecretString,
        from_address: String,
        base_url: String,
    ) -> Self {
        Self { client, api_key, from_address, base_url }
    }

    /// Send a magic link email.
    pub async fn send_magic_link(&self, to_email: &str, token: &str) -> Result<(), AppError> {
        let verify_url = format!("{}/api/v1/auth/verify?token={}", self.base_url, token);

        let body = serde_json::json!({
            "from": self.from_address,
            "to": [to_email],
            "subject": "Votre lien de connexion - AI Weekly Synth",
            "html": format!(
                r#"<h2>Connexion a AI Weekly Synth</h2>
                <p>Cliquez sur le lien ci-dessous pour vous connecter :</p>
                <p><a href="{url}" style="display:inline-block;padding:12px 24px;
                background-color:#4F46E5;color:white;text-decoration:none;
                border-radius:6px;">Se connecter</a></p>
                <p>Ou copiez ce lien : <code>{url}</code></p>
                <p><em>Ce lien expire dans 15 minutes.</em></p>"#,
                url = verify_url,
            ),
        });

        let resp = self.client
            .post("https://api.resend.com/emails")
            .header("Authorization", format!("Bearer {}", self.api_key.expose_secret()))
            .json(&body)
            .send()
            .await
            .map_err(|e| AppError::SmtpError(format!("Failed to send email: {e}")))?;

        if !resp.status().is_success() {
            let error_body = resp.text().await.unwrap_or_default();
            return Err(AppError::SmtpError(
                format!("Resend API error: {error_body}")
            ));
        }

        Ok(())
    }

    /// Send a synthesis by email.
    pub async fn send_synthesis(
        &self,
        to_email: &str,
        synthesis_title: &str,
        sections_html: &str,
    ) -> Result<(), AppError> {
        let body = serde_json::json!({
            "from": self.from_address,
            "to": [to_email],
            "subject": format!("Synthese : {}", synthesis_title),
            "html": format!(
                r#"<h1>{title}</h1>
                {content}
                <hr/>
                <p><em>Genere par AI Weekly Synth</em></p>"#,
                title = synthesis_title,
                content = sections_html,
            ),
        });

        let resp = self.client
            .post("https://api.resend.com/emails")
            .header("Authorization", format!("Bearer {}", self.api_key.expose_secret()))
            .json(&body)
            .send()
            .await
            .map_err(|e| AppError::SmtpError(format!("Failed to send email: {e}")))?;

        if !resp.status().is_success() {
            let error_body = resp.text().await.unwrap_or_default();
            return Err(AppError::SmtpError(
                format!("Resend API error: {error_body}")
            ));
        }

        Ok(())
    }
}

8.2 HTML Rendering for Synthesis Emails

/// Convert synthesis sections to HTML for email delivery.
pub fn sections_to_html(sections: &[NewsSection]) -> String {
    let mut html = String::new();
    for section in sections {
        html.push_str(&format!("<h2>{}</h2>\n<ul>\n", section.title));
        for item in &section.items {
            html.push_str(&format!(
                r#"<li><strong><a href="{url}">{title}</a></strong><br/>{summary}</li>"#,
                url = html_escape(&item.url),
                title = html_escape(&item.title),
                summary = html_escape(&item.summary),
            ));
        }
        html.push_str("</ul>\n");
    }
    html
}

fn html_escape(s: &str) -> String {
    s.replace('&', "&amp;")
     .replace('<', "&lt;")
     .replace('>', "&gt;")
     .replace('"', "&quot;")
}

9. Encryption

9.1 AES-256-GCM Implementation

// src/services/encryption.rs
use aes_gcm::{
    aead::{Aead, KeyInit, OsRng},
    Aes256Gcm, Nonce,
};
use rand::RngCore;
use secrecy::{ExposeSecret, SecretString};

/// Master key holder -- wraps a 32-byte key.
#[derive(Clone)]
pub struct MasterKey {
    key: [u8; 32],
}

impl MasterKey {
    /// Parse a 64-character hex string into a 32-byte key.
    pub fn from_hex(hex_str: &str) -> Result<Self, AppError> {
        let bytes = hex::decode(hex_str)
            .map_err(|_| AppError::Internal(
                anyhow::anyhow!("MASTER_ENCRYPTION_KEY must be a 64-character hex string")
            ))?;
        if bytes.len() != 32 {
            return Err(AppError::Internal(
                anyhow::anyhow!("MASTER_ENCRYPTION_KEY must be exactly 32 bytes (64 hex chars)")
            ));
        }
        let mut key = [0u8; 32];
        key.copy_from_slice(&bytes);
        Ok(Self { key })
    }

    /// Encrypt a plaintext string. Returns (ciphertext, nonce) both as byte vectors.
    pub fn encrypt(&self, plaintext: &str) -> Result<(Vec<u8>, Vec<u8>), AppError> {
        let cipher = Aes256Gcm::new_from_slice(&self.key)
            .map_err(|e| AppError::Internal(anyhow::anyhow!("Cipher init failed: {e}")))?;

        let mut nonce_bytes = [0u8; 12];
        OsRng.fill_bytes(&mut nonce_bytes);
        let nonce = Nonce::from_slice(&nonce_bytes);

        let ciphertext = cipher
            .encrypt(nonce, plaintext.as_bytes())
            .map_err(|e| AppError::Internal(anyhow::anyhow!("Encryption failed: {e}")))?;

        Ok((ciphertext, nonce_bytes.to_vec()))
    }

    /// Decrypt ciphertext using the provided nonce. Returns the plaintext as a SecretString.
    pub fn decrypt(
        &self,
        ciphertext: &[u8],
        nonce_bytes: &[u8],
    ) -> Result<SecretString, AppError> {
        let cipher = Aes256Gcm::new_from_slice(&self.key)
            .map_err(|e| AppError::Internal(anyhow::anyhow!("Cipher init failed: {e}")))?;

        let nonce = Nonce::from_slice(nonce_bytes);

        let plaintext_bytes = cipher
            .decrypt(nonce, ciphertext)
            .map_err(|_| AppError::Internal(
                anyhow::anyhow!("Decryption failed -- master key mismatch or data corruption")
            ))?;

        let plaintext = String::from_utf8(plaintext_bytes)
            .map_err(|e| AppError::Internal(anyhow::anyhow!("Decrypted data is not valid UTF-8: {e}")))?;

        Ok(SecretString::new(plaintext))
    }
}

impl Drop for MasterKey {
    fn drop(&mut self) {
        // Zeroize the key on drop
        zeroize::Zeroize::zeroize(&mut self.key);
    }
}

9.2 Master Key Loading

// In config.rs
pub struct AppConfig {
    pub database_url: String,
    pub master_key: MasterKey,
    pub base_url: String,
    pub port: u16,
    pub resend_api_key: SecretString,
    pub resend_from: String,
    pub turnstile_secret: String,
    pub turnstile_site_key: String,
    // ...
}

impl AppConfig {
    pub fn from_env() -> Result<Self, AppError> {
        dotenvy::dotenv().ok();

        let master_key_hex = std::env::var("MASTER_ENCRYPTION_KEY")
            .map_err(|_| AppError::Internal(
                anyhow::anyhow!("MASTER_ENCRYPTION_KEY environment variable is required")
            ))?;

        Ok(Self {
            master_key: MasterKey::from_hex(&master_key_hex)?,
            // ... other fields
        })
    }
}

9.3 Key Rotation Strategy

Key rotation is performed via a CLI command that re-encrypts all user API keys under a new master key:

// CLI subcommand: rotate-key
Commands::RotateKey {
    old_key_hex,
    new_key_hex,
} => {
    let old_key = MasterKey::from_hex(&old_key_hex)?;
    let new_key = MasterKey::from_hex(&new_key_hex)?;
    let pool = create_pool().await?;

    let all_keys = sqlx::query_as!(
        RawProviderKey,
        "SELECT id, encrypted_key, nonce FROM user_provider_keys"
    )
    .fetch_all(&pool)
    .await?;

    let mut tx = pool.begin().await?;
    for key_row in &all_keys {
        // Decrypt with old key
        let plaintext = old_key.decrypt(&key_row.encrypted_key, &key_row.nonce)?;
        // Re-encrypt with new key
        let (new_ciphertext, new_nonce) = new_key.encrypt(plaintext.expose_secret())?;
        // Update in transaction
        sqlx::query!(
            "UPDATE user_provider_keys SET encrypted_key = $1, nonce = $2 WHERE id = $3",
            new_ciphertext,
            new_nonce,
            key_row.id
        )
        .execute(&mut *tx)
        .await?;
    }
    tx.commit().await?;
    println!("Rotated {} keys successfully.", all_keys.len());
}

After running: update the MASTER_ENCRYPTION_KEY environment variable and restart the application.

10. Error Handling

10.1 Error Type Hierarchy

// src/error.rs
use axum::http::StatusCode;
use axum::response::{IntoResponse, Response};
use axum::Json;
use serde_json::json;

#[derive(Debug, thiserror::Error)]
pub enum AppError {
    // === Client errors ===
    #[error("Bad request: {0}")]
    BadRequest(String),

    #[error("Unauthorized: {0}")]
    Unauthorized(String),

    #[error("Forbidden: {0}")]
    Forbidden(String),

    #[error("Not found: {0}")]
    NotFound(String),

    #[error("Conflict: {0}")]
    Conflict(String),

    #[error("Too many requests")]
    TooManyRequests { retry_after_secs: u64 },

    #[error("Validation error")]
    Validation(Vec<FieldError>),

    // === Server errors ===
    #[error("Internal error: {0}")]
    Internal(#[from] anyhow::Error),

    #[error("Database error: {0}")]
    Database(#[from] sqlx::Error),

    #[error("LLM provider error: {0}")]
    LlmError(String),

    #[error("Email error: {0}")]
    SmtpError(String),

    #[error("Scraping error: {0}")]
    ScrapingError(String),
}

#[derive(Debug, serde::Serialize)]
pub struct FieldError {
    pub field: String,
    pub message: String,
}

impl IntoResponse for AppError {
    fn into_response(self) -> Response {
        let (status, body) = match &self {
            AppError::BadRequest(msg) => (
                StatusCode::BAD_REQUEST,
                json!({ "error": msg }),
            ),
            AppError::Unauthorized(_) => (
                StatusCode::UNAUTHORIZED,
                json!({ "error": "Unauthorized" }),
            ),
            AppError::Forbidden(_) => (
                StatusCode::FORBIDDEN,
                json!({ "error": "Forbidden" }),
            ),
            AppError::NotFound(msg) => (
                StatusCode::NOT_FOUND,
                json!({ "error": msg }),
            ),
            AppError::Conflict(msg) => (
                StatusCode::CONFLICT,
                json!({ "error": msg }),
            ),
            AppError::TooManyRequests { retry_after_secs } => {
                let mut resp = (
                    StatusCode::TOO_MANY_REQUESTS,
                    Json(json!({ "error": "Too many requests", "retry_after": retry_after_secs })),
                ).into_response();
                resp.headers_mut().insert(
                    "Retry-After",
                    retry_after_secs.to_string().parse().unwrap(),
                );
                return resp;
            }
            AppError::Validation(errors) => (
                StatusCode::UNPROCESSABLE_ENTITY,
                json!({ "error": "Validation failed", "details": errors }),
            ),
            AppError::Database(e) => {
                tracing::error!(error = %e, "Database error");
                (
                    StatusCode::INTERNAL_SERVER_ERROR,
                    json!({ "error": "Internal server error" }),
                )
            }
            AppError::Internal(e) => {
                tracing::error!(error = ?e, "Internal error");
                (
                    StatusCode::INTERNAL_SERVER_ERROR,
                    json!({ "error": "Internal server error" }),
                )
            }
            AppError::LlmError(msg) => {
                tracing::error!(error = %msg, "LLM provider error");
                (
                    StatusCode::BAD_GATEWAY,
                    json!({ "error": "AI provider error. Please try again." }),
                )
            }
            AppError::SmtpError(msg) => {
                tracing::error!(error = %msg, "Email error");
                (
                    StatusCode::BAD_GATEWAY,
                    json!({ "error": "Failed to send email. Please try again." }),
                )
            }
            AppError::ScrapingError(msg) => {
                tracing::warn!(error = %msg, "Scraping error");
                (
                    StatusCode::BAD_GATEWAY,
                    json!({ "error": "Failed to fetch article content." }),
                )
            }
        };

        (status, Json(body)).into_response()
    }
}

Key design decisions:

Never expose internal details to the client (database errors, stack traces).
Log server errors at error! level with full context for debugging.
Client errors return the user-facing message directly.
LLM and email errors return 502 Bad Gateway (upstream service failure).
The From<sqlx::Error> impl converts specific constraint violations (unique, foreign key) into Conflict or BadRequest where appropriate.

10.2 Logging Strategy

// In main.rs
use tracing_subscriber::{layer::SubscriberExt, util::SubscriberInitExt, EnvFilter};

fn init_tracing() {
    tracing_subscriber::registry()
        .with(EnvFilter::try_from_default_env().unwrap_or_else(|_| {
            "ai_synth_backend=debug,tower_http=info,sqlx=warn".into()
        }))
        .with(tracing_subscriber::fmt::layer()
            .with_target(true)
            .with_thread_ids(false)
            .with_file(true)
            .with_line_number(true))
        .init();
}

Logging conventions:

Never log decrypted API keys, session tokens, magic link tokens, or the master key.
Use secrecy::SecretString for all secret values. Its Debug and Display impls print [REDACTED].
Structured fields: Use tracing::info!(user_id = %id, action = "generate", ...) not string interpolation.
Request tracing: Use tower_http::trace::TraceLayer for automatic request/response logging.

11. Testing Strategy

11.1 Test Database Setup

Tests use a real PostgreSQL instance (via Docker in CI, or a local dev instance). Each test gets its own schema or uses transactions that are rolled back.

// tests/common/mod.rs
use sqlx::PgPool;
use uuid::Uuid;

/// Create a test database pool pointing to a test-specific schema.
pub async fn setup_test_db() -> PgPool {
    dotenvy::dotenv().ok();
    let base_url = std::env::var("TEST_DATABASE_URL")
        .unwrap_or_else(|_| "postgres://postgres:postgres@localhost:5432/ai_synth_test".into());

    let pool = PgPool::connect(&base_url).await.unwrap();

    // Run migrations
    sqlx::migrate!("./migrations")
        .run(&pool)
        .await
        .unwrap();

    pool
}

/// Create a test user and return (user_id, session_token).
pub async fn create_test_user(pool: &PgPool) -> (Uuid, String) {
    let user_id = Uuid::new_v4();
    let email = format!("test-{}@example.com", Uuid::new_v4());

    sqlx::query!(
        "INSERT INTO users (id, email, role) VALUES ($1, $2, 'user')",
        user_id,
        email,
    )
    .execute(pool)
    .await
    .unwrap();

    // Create session
    let token = crate::util::token::generate_token();
    let hash = crate::util::token::hash_token(&token);
    let expires = chrono::Utc::now() + chrono::Duration::days(30);

    sqlx::query!(
        "INSERT INTO sessions (user_id, token_hash, expires_at) VALUES ($1, $2, $3)",
        user_id,
        hash,
        expires,
    )
    .execute(pool)
    .await
    .unwrap();

    (user_id, token)
}

/// Create an admin user and return (user_id, session_token).
pub async fn create_test_admin(pool: &PgPool) -> (Uuid, String) {
    let user_id = Uuid::new_v4();
    let email = format!("admin-{}@example.com", Uuid::new_v4());

    sqlx::query!(
        "INSERT INTO users (id, email, role) VALUES ($1, $2, 'admin')",
        user_id,
        email,
    )
    .execute(pool)
    .await
    .unwrap();

    let token = crate::util::token::generate_token();
    let hash = crate::util::token::hash_token(&token);
    let expires = chrono::Utc::now() + chrono::Duration::days(30);

    sqlx::query!(
        "INSERT INTO sessions (user_id, token_hash, expires_at) VALUES ($1, $2, $3)",
        user_id,
        hash,
        expires,
    )
    .execute(pool)
    .await
    .unwrap();

    (user_id, token)
}

/// Build a test app (Axum router) with real DB and mock external services.
pub async fn build_test_app(pool: PgPool) -> axum::Router {
    let config = AppConfig::test_defaults();
    let state = AppState {
        pool,
        http_client: reqwest::Client::new(),
        config,
        rate_limiter: Arc::new(RateLimiter::new()),
        job_store: Arc::new(DashMap::new()),
    };
    crate::router::create_router(state)
}

11.2 Unit Test Examples

Encryption

// tests/services/encryption_test.rs

#[test]
fn test_encrypt_decrypt_roundtrip() {
    let master_key = MasterKey::from_hex(
        "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"
    ).unwrap();

    let plaintext = "sk-proj-abc123def456";
    let (ciphertext, nonce) = master_key.encrypt(plaintext).unwrap();

    // Ciphertext should not equal plaintext
    assert_ne!(ciphertext, plaintext.as_bytes());

    // Decrypt should return original
    let decrypted = master_key.decrypt(&ciphertext, &nonce).unwrap();
    assert_eq!(decrypted.expose_secret(), plaintext);
}

#[test]
fn test_decrypt_wrong_key_fails() {
    let key1 = MasterKey::from_hex(
        "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"
    ).unwrap();
    let key2 = MasterKey::from_hex(
        "fedcba9876543210fedcba9876543210fedcba9876543210fedcba9876543210"
    ).unwrap();

    let (ciphertext, nonce) = key1.encrypt("secret").unwrap();
    assert!(key2.decrypt(&ciphertext, &nonce).is_err());
}

#[test]
fn test_each_encryption_produces_unique_nonce() {
    let master_key = MasterKey::from_hex(
        "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"
    ).unwrap();

    let (_, nonce1) = master_key.encrypt("same text").unwrap();
    let (_, nonce2) = master_key.encrypt("same text").unwrap();
    assert_ne!(nonce1, nonce2);
}

Rate Limiter

// tests/services/rate_limiter_test.rs

#[tokio::test]
async fn test_rate_limiter_allows_within_limit() {
    let limiter = RateLimiter::new();
    limiter.update_provider_limit("test", 5, 60);

    let user_id = Uuid::new_v4();
    for _ in 0..5 {
        limiter.acquire("test", user_id).await.unwrap();
    }
}

#[tokio::test]
async fn test_rate_limiter_blocks_when_exceeded() {
    let limiter = RateLimiter::new();
    limiter.update_provider_limit("test", 2, 60);

    let user_id = Uuid::new_v4();
    limiter.acquire("test", user_id).await.unwrap();
    limiter.acquire("test", user_id).await.unwrap();

    // Third should block or return TooManyRequests
    // (implementation detail: may sleep or error)
}

Token Utilities

// Unit tests in src/util/token.rs

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

    #[test]
    fn test_generate_token_length() {
        let token = generate_token();
        // 32 bytes base64url-encoded = 43 characters
        assert_eq!(token.len(), 43);
    }

    #[test]
    fn test_generate_token_uniqueness() {
        let t1 = generate_token();
        let t2 = generate_token();
        assert_ne!(t1, t2);
    }

    #[test]
    fn test_hash_token_deterministic() {
        let token = "test-token-value";
        let h1 = hash_token(token);
        let h2 = hash_token(token);
        assert_eq!(h1, h2);
    }

    #[test]
    fn test_hash_token_different_inputs() {
        assert_ne!(hash_token("a"), hash_token("b"));
    }
}

11.3 Integration Test (API)

// tests/api/sources_test.rs
use axum::http::{Request, StatusCode};
use tower::ServiceExt;
use http_body_util::BodyExt;

#[tokio::test]
async fn test_create_and_list_sources() {
    let pool = common::setup_test_db().await;
    let (user_id, session_token) = common::create_test_user(&pool).await;
    let app = common::build_test_app(pool).await;

    // Create a source
    let resp = app
        .clone()
        .oneshot(
            Request::builder()
                .method("POST")
                .uri("/api/v1/sources")
                .header("Content-Type", "application/json")
                .header("Cookie", format!("ai_synth_session={session_token}"))
                .header("X-Requested-With", "XMLHttpRequest")
                .body(serde_json::to_string(&serde_json::json!({
                    "title": "Test Source",
                    "url": "https://example.com"
                })).unwrap().into())
                .unwrap(),
        )
        .await
        .unwrap();

    assert_eq!(resp.status(), StatusCode::CREATED);

    // List sources
    let resp = app
        .clone()
        .oneshot(
            Request::builder()
                .method("GET")
                .uri("/api/v1/sources")
                .header("Cookie", format!("ai_synth_session={session_token}"))
                .body(axum::body::Body::empty())
                .unwrap(),
        )
        .await
        .unwrap();

    assert_eq!(resp.status(), StatusCode::OK);
    let body = resp.into_body().collect().await.unwrap().to_bytes();
    let sources: Vec<serde_json::Value> = serde_json::from_slice(&body).unwrap();
    assert_eq!(sources.len(), 1);
    assert_eq!(sources[0]["title"], "Test Source");
}

#[tokio::test]
async fn test_cannot_access_other_users_sources() {
    let pool = common::setup_test_db().await;
    let (user1_id, token1) = common::create_test_user(&pool).await;
    let (user2_id, token2) = common::create_test_user(&pool).await;
    let app = common::build_test_app(pool.clone()).await;

    // User 1 creates a source
    let resp = app
        .clone()
        .oneshot(
            Request::builder()
                .method("POST")
                .uri("/api/v1/sources")
                .header("Content-Type", "application/json")
                .header("Cookie", format!("ai_synth_session={token1}"))
                .header("X-Requested-With", "XMLHttpRequest")
                .body(serde_json::to_string(&serde_json::json!({
                    "title": "User 1 Source",
                    "url": "https://example.com"
                })).unwrap().into())
                .unwrap(),
        )
        .await
        .unwrap();
    assert_eq!(resp.status(), StatusCode::CREATED);

    // User 2 lists sources -- should see empty list
    let resp = app
        .clone()
        .oneshot(
            Request::builder()
                .method("GET")
                .uri("/api/v1/sources")
                .header("Cookie", format!("ai_synth_session={token2}"))
                .body(axum::body::Body::empty())
                .unwrap(),
        )
        .await
        .unwrap();

    let body = resp.into_body().collect().await.unwrap().to_bytes();
    let sources: Vec<serde_json::Value> = serde_json::from_slice(&body).unwrap();
    assert_eq!(sources.len(), 0);
}

#[tokio::test]
async fn test_unauthenticated_request_returns_401() {
    let pool = common::setup_test_db().await;
    let app = common::build_test_app(pool).await;

    let resp = app
        .oneshot(
            Request::builder()
                .method("GET")
                .uri("/api/v1/sources")
                .body(axum::body::Body::empty())
                .unwrap(),
        )
        .await
        .unwrap();

    assert_eq!(resp.status(), StatusCode::UNAUTHORIZED);
}

11.4 Test Conventions

Layer	Tool	What to test
`util/`	`#[test]` unit tests	Token generation, hashing, encryption roundtrips
`services/`	`#[tokio::test]`	Rate limiter behavior, scraper parsing (with HTML fixtures), email rendering
`db/`	`#[tokio::test]` with real Postgres	Query correctness, constraint enforcement, data isolation
`handlers/`	`#[tokio::test]` with `tower::ServiceExt::oneshot`	Full HTTP request/response cycle, status codes, auth enforcement, CSRF checks
LLM providers	Mock HTTP server (`wiremock` crate)	Response parsing, error handling, schema construction

Appendix A: Router Wiring

// src/router.rs
use axum::{
    middleware,
    routing::{get, post, put, delete},
    Router,
};
use tower_http::cors::CorsLayer;
use tower_http::trace::TraceLayer;
use tower_http::services::ServeDir;

pub fn create_router(state: AppState) -> Router {
    let api_routes = Router::new()
        // === Public auth routes (no session required) ===
        .route("/auth/register", post(handlers::auth::register))
        .route("/auth/magic-link", post(handlers::auth::request_magic_link))
        .route("/auth/verify", get(handlers::auth::verify_magic_link))

        // === Authenticated routes ===
        .route("/auth/logout", post(handlers::auth::logout))
        .route("/auth/me", get(handlers::auth::me))

        .route("/syntheses", get(handlers::syntheses::list))
        .route("/syntheses/:id", get(handlers::syntheses::get))
        .route("/syntheses/:id", delete(handlers::syntheses::delete))
        .route("/syntheses/generate", post(handlers::syntheses::trigger_generate))
        .route("/syntheses/:id/email", post(handlers::syntheses::send_email))

        .route("/sources", get(handlers::sources::list))
        .route("/sources", post(handlers::sources::create))
        .route("/sources/:id", delete(handlers::sources::delete))

        .route("/settings", get(handlers::settings::get))
        .route("/settings", put(handlers::settings::update))

        .route("/provider-keys", get(handlers::provider_keys::list))
        .route("/provider-keys", post(handlers::provider_keys::create))
        .route("/provider-keys/:provider", delete(handlers::provider_keys::delete))

        .route("/generation/:job_id/progress", get(handlers::generation::progress_stream))
        .route("/generation/:job_id/status", get(handlers::generation::job_status))

        .route("/config/providers", get(handlers::admin::list_enabled_providers))

        // === Admin routes ===
        .route("/admin/provider-models", get(handlers::admin::list_provider_models))
        .route("/admin/provider-models", post(handlers::admin::create_provider_model))
        .route("/admin/provider-models/:id", put(handlers::admin::update_provider_model))
        .route("/admin/provider-models/:id", delete(handlers::admin::delete_provider_model))
        .route("/admin/rate-limits", get(handlers::admin::list_rate_limits))
        .route("/admin/rate-limits/:provider", put(handlers::admin::update_rate_limit))
        .route("/admin/users", get(handlers::admin::list_users))
        .route("/admin/users/:id/role", put(handlers::admin::update_user_role))

        // === Health ===
        .route("/health", get(handlers::health::health_check));

    let app = Router::new()
        .nest("/api/v1", api_routes)
        // CSRF middleware on API routes (checks X-Requested-With on POST/PUT/DELETE)
        .layer(middleware::from_fn(middleware::csrf::csrf_check))
        // CORS
        .layer(build_cors_layer(&state.config))
        // Request tracing
        .layer(TraceLayer::new_for_http())
        // Static files fallback (SPA)
        .fallback_service(
            ServeDir::new(&state.config.static_dir)
                .fallback(ServeDir::new(&state.config.static_dir).append_index_html_on_directories(true))
        )
        .with_state(state);

    app
}

fn build_cors_layer(config: &AppConfig) -> CorsLayer {
    CorsLayer::new()
        .allow_origin(config.base_url.parse::<HeaderValue>().unwrap())
        .allow_methods([Method::GET, Method::POST, Method::PUT, Method::DELETE])
        .allow_headers([
            header::CONTENT_TYPE,
            header::COOKIE,
            HeaderName::from_static("x-requested-with"),
        ])
        .allow_credentials(true)
        .max_age(Duration::from_secs(3600))
}

Appendix B: AppConfig Full Definition

// src/config.rs

#[derive(Clone)]
pub struct AppConfig {
    // Server
    pub port: u16,
    pub base_url: String,                    // e.g., "https://synth.example.com"
    pub static_dir: String,                  // path to SolidJS build output

    // Database
    pub database_url: String,                // postgres://...

    // Encryption
    pub master_key: MasterKey,               // from MASTER_ENCRYPTION_KEY env var

    // Email (Resend)
    pub resend_api_key: SecretString,         // from RESEND_API_KEY env var
    pub resend_from: String,                  // e.g., "AI Weekly Synth <noreply@example.com>"

    // Captcha (Cloudflare Turnstile)
    pub turnstile_secret: String,
    pub turnstile_site_key: String,

    // Logging
    pub rust_log: String,
}

impl AppConfig {
    pub fn from_env() -> Result<Self, AppError> {
        dotenvy::dotenv().ok();

        Ok(Self {
            port: env_var_or("PORT", "8080").parse()?,
            base_url: env_var("BASE_URL")?,
            static_dir: env_var_or("STATIC_DIR", "./static"),
            database_url: env_var("DATABASE_URL")?,
            master_key: MasterKey::from_hex(&env_var("MASTER_ENCRYPTION_KEY")?)?,
            resend_api_key: SecretString::new(env_var("RESEND_API_KEY")?),
            resend_from: env_var_or("RESEND_FROM", "AI Weekly Synth <noreply@example.com>"),
            turnstile_secret: env_var("TURNSTILE_SECRET")?,
            turnstile_site_key: env_var("TURNSTILE_SITE_KEY")?,
            rust_log: env_var_or("RUST_LOG", "ai_synth_backend=info"),
        })
    }

    #[cfg(test)]
    pub fn test_defaults() -> Self {
        Self {
            port: 0,
            base_url: "http://localhost:3000".into(),
            static_dir: "./static".into(),
            database_url: "postgres://postgres:postgres@localhost:5432/ai_synth_test".into(),
            master_key: MasterKey::from_hex(
                "0123456789abcdef0123456789abcdef0123456789abcdef0123456789abcdef"
            ).unwrap(),
            resend_api_key: SecretString::new("test-resend-key".into()),
            resend_from: "test@example.com".into(),
            turnstile_secret: "test-secret".into(),
            turnstile_site_key: "test-site-key".into(),
            rust_log: "debug".into(),
        }
    }
}

fn env_var(name: &str) -> Result<String, AppError> {
    std::env::var(name).map_err(|_| {
        AppError::Internal(anyhow::anyhow!("Missing required env var: {name}"))
    })
}

fn env_var_or(name: &str, default: &str) -> String {
    std::env::var(name).unwrap_or_else(|_| default.to_string())
}

Appendix C: Prompt Construction

The prompts mirror the current Gemini service but are provider-agnostic.

// src/services/generation.rs (prompt helpers)

fn build_search_prompts(
    settings: &UserSettings,
    sources: &[Source],
) -> (String, String) {
    let current_date = Utc::now().format("%A %d %B %Y").to_string();

    let sources_text = if sources.is_empty() {
        String::new()
    } else {
        let list = sources.iter()
            .map(|s| format!("- {} ({})", s.title, s.url))
            .collect::<Vec<_>>()
            .join("\n");
        format!(
            "\nYou MUST also consult these custom sources:\n{}\n",
            list
        )
    };

    let categories_text = settings.categories.iter()
        .enumerate()
        .map(|(i, cat)| format!("{}. {}", i + 1, cat))
        .collect::<Vec<_>>()
        .join("\n");

    let system_prompt = format!(
        "You are a precise AI assistant. You MUST provide complete and exact URLs. \
         Focus ONLY on news from the last {} days.",
        settings.max_age_days
    );

    let user_prompt = format!(
        "Today is {date}.\n\
         You are an expert analyst on the topic: \"{theme}\".\n\
         Search for news STRICTLY from the last {days} days.\n\
         Do NOT return any news older than {days} days.\n\
         {sources}\
         {behavior}\n\
         The synthesis must be divided into {count} sections:\n\
         {categories}\n\n\
         For each category, provide at most {max_items} articles.\n\
         For each article, provide: a provisional title, the exact source URL, \
         and a provisional summary.\n\
         Return the result as JSON using keys category_0, category_1, etc. \
         matching the section order above.",
        date = current_date,
        theme = settings.theme,
        days = settings.max_age_days,
        sources = sources_text,
        behavior = settings.search_agent_behavior,
        count = settings.categories.len(),
        categories = categories_text,
        max_items = settings.max_items_per_category,
    );

    (system_prompt, user_prompt)
}

fn build_rewrite_prompts(
    scraped: &[(String, Vec<ScrapedItem>)],
) -> (String, String) {
    let system_prompt = "You are a precise AI assistant. Generate titles and summaries \
        that faithfully reflect the provided content.".to_string();

    let data = serde_json::to_string_pretty(scraped).unwrap_or_default();

    let user_prompt = format!(
        "You are an expert news analyst.\n\
         Below is a list of articles organized by category, each with raw text content \
         extracted from the source websites ('scraped_content').\n\
         Your task is to rewrite the 'title' and 'summary' (4-5 lines) for each article \
         so they EXACTLY and FAITHFULLY reflect the provided text content.\n\
         If 'scraped_content' is empty or insufficient, use the original title and summary.\n\
         KEEP the exact same URLs. Do NOT remove any article.\n\n\
         Article data:\n{data}",
        data = data,
    );

    (system_prompt, user_prompt)
}

/// Build the JSON schema for structured output, based on user categories.
fn build_category_schema(categories: &[String]) -> serde_json::Value {
    let mut properties = serde_json::Map::new();
    let mut required = Vec::new();

    for (i, _cat) in categories.iter().enumerate() {
        let key = format!("category_{}", i);
        properties.insert(key.clone(), serde_json::json!({
            "type": "array",
            "items": {
                "type": "object",
                "properties": {
                    "title": { "type": "string" },
                    "url": { "type": "string" },
                    "summary": { "type": "string" }
                },
                "required": ["title", "url", "summary"]
            }
        }));
        required.push(serde_json::Value::String(key));
    }

    serde_json::json!({
        "type": "object",
        "properties": properties,
        "required": required
    })
}

Appendix D: Background Tasks

Besides generation jobs, the application needs periodic cleanup tasks.

// In main.rs, after starting the server

// Periodic cleanup: expired sessions and magic link tokens
let cleanup_pool = pool.clone();
tokio::spawn(async move {
    let mut interval = tokio::time::interval(Duration::from_secs(3600)); // every hour
    loop {
        interval.tick().await;
        if let Err(e) = db::sessions::delete_expired(&cleanup_pool).await {
            tracing::error!(error = %e, "Failed to clean up expired sessions");
        }
        if let Err(e) = db::magic_links::delete_expired(&cleanup_pool).await {
            tracing::error!(error = %e, "Failed to clean up expired magic link tokens");
        }
    }
});

Appendix E: .env.example

# === Required ===
DATABASE_URL=postgres://ai_synth:password@localhost:5432/ai_synth
BASE_URL=https://synth.example.com
MASTER_ENCRYPTION_KEY=<64-character-hex-string>

# === Email (Resend) ===
RESEND_API_KEY=re_xxxxxxxxxxxx
RESEND_FROM=AI Weekly Synth <noreply@yourdomain.com>

# === Captcha (Cloudflare Turnstile) ===
TURNSTILE_SECRET=0x4AAAAAAA...
TURNSTILE_SITE_KEY=0x4BBBBBB...

# === Optional ===
PORT=8080
STATIC_DIR=./static
RUST_LOG=ai_synth_backend=info,tower_http=info

101 KiB Raw Blame History