lean_ctx/core/config/

sections.rs

//! Auxiliary configuration section structs.
//!
//! Nested config structs (secret-detection, setup, archive, providers,
//! autonomy, updates, cloud, gain, loop-detection, embedding, …) split out of
//! `config/mod.rs` to keep the top-level module focused on `Config` itself.
//! Re-exported via `pub use sections::*`, so external paths stay stable.

use super::serde_defaults;
#[allow(clippy::wildcard_imports)]
use super::*;
use serde::{Deserialize, Serialize};

#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct SecretDetectionConfig {
    pub enabled: bool,
    pub redact: bool,
    pub custom_patterns: Vec<String>,
    /// #718: subtractive counterpart to `custom_patterns` — a detected secret
    /// whose matched text is covered by any of these regexes is neither
    /// reported nor redacted. Lets users carve out known-safe identifiers or
    /// repo naming conventions without disabling secret detection wholesale.
    pub exclude_patterns: Vec<String>,
}

/// Controls what lean-ctx injects during `setup` and `update --rewire`.
/// Fresh installs default to non-invasive (rules/skills off, MCP on).
/// Users who ran setup interactively get explicit true/false.
/// `None` = undecided (legacy: check if rules already exist and preserve behavior).
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct SetupConfig {
    /// Inject agent rule files (CLAUDE.md, .cursor/rules/, etc.).
    /// None = undecided (legacy compat: inject if rules already present).
    /// Some(true) = always inject. Some(false) = never inject.
    pub auto_inject_rules: Option<bool>,
    /// Install SKILL.md files for supported agents.
    /// None = undecided. Some(true) = install. Some(false) = skip.
    pub auto_inject_skills: Option<bool>,
    /// Register lean-ctx as an MCP server in editor configs.
    #[serde(default = "serde_defaults::default_true")]
    pub auto_update_mcp: bool,
}

impl Default for SetupConfig {
    fn default() -> Self {
        Self {
            auto_inject_rules: None,
            auto_inject_skills: None,
            auto_update_mcp: true,
        }
    }
}

impl SetupConfig {
    /// Returns whether rules should be injected, considering legacy installs.
    /// If undecided (None), checks if lean-ctx rules markers already exist
    /// in any agent config — if so, keeps injecting for backward compat.
    pub fn should_inject_rules(&self) -> bool {
        match self.auto_inject_rules {
            Some(v) => v,
            None => Self::rules_already_present(),
        }
    }

    /// Returns whether skills should be installed.
    pub fn should_inject_skills(&self) -> bool {
        match self.auto_inject_skills {
            Some(v) => v,
            None => Self::rules_already_present(),
        }
    }

    /// Returns whether `setup`/`onboard`/`init` may (re)register the lean-ctx
    /// MCP server in editor configs. Honors `auto_update_mcp` (#281) so locked-
    /// down environments can keep MCP out of agent settings while still getting
    /// hooks, rules and skills.
    pub fn should_update_mcp(&self) -> bool {
        self.auto_update_mcp
    }

    /// Check if lean-ctx rules markers exist in any known agent config location.
    ///
    /// Delegates the per-agent path catalog to `rules_inject::any_rules_marker_present`
    /// (derived from the injector's own target list) so this never drifts behind
    /// newly supported agents again (#442). Claude Code and CodeBuddy have no
    /// rules *target* (they auto-load an inline block instead), so their legacy
    /// rule files are checked separately to keep honoring older installs.
    fn rules_already_present() -> bool {
        let Some(home) = dirs::home_dir() else {
            return false;
        };
        if crate::rules_inject::any_rules_marker_present(&home) {
            return true;
        }
        let legacy_paths = [
            crate::core::editor_registry::claude_rules_dir(&home).join("lean-ctx.md"),
            crate::core::editor_registry::codebuddy_rules_dir(&home).join("lean-ctx.md"),
        ];
        legacy_paths.iter().any(|p| {
            std::fs::read_to_string(p)
                .is_ok_and(|c| c.contains(crate::core::rules_canonical::START_MARK))
        })
    }
}

impl Default for SecretDetectionConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            redact: true,
            custom_patterns: Vec::new(),
            exclude_patterns: Vec::new(),
        }
    }
}

/// Settings for the zero-loss compression archive (large tool outputs saved to disk).
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct ArchiveConfig {
    pub enabled: bool,
    pub threshold_chars: usize,
    pub max_age_hours: u64,
    pub max_disk_mb: u64,
    pub ephemeral: bool,
    /// Minimum output tokens before the ephemeral firewall replaces an inline tool
    /// result with a summary + retrieval ref. Outputs below this stay fully inline.
    pub ephemeral_min_tokens: usize,
}

impl Default for ArchiveConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            threshold_chars: 800,
            max_age_hours: 48,
            max_disk_mb: 500,
            ephemeral: true,
            ephemeral_min_tokens: 2000,
        }
    }
}

impl ArchiveConfig {
    pub fn ephemeral_effective(&self) -> bool {
        if let Ok(v) = std::env::var("LEAN_CTX_EPHEMERAL") {
            return !matches!(v.trim(), "0" | "false" | "off");
        }
        self.ephemeral && self.enabled
    }

    pub fn ephemeral_min_tokens_effective(&self) -> usize {
        if let Ok(v) = std::env::var("LEAN_CTX_EPHEMERAL_MIN_TOKENS")
            && let Ok(n) = v.trim().parse::<usize>()
        {
            return n;
        }
        self.ephemeral_min_tokens
    }
}

/// Configuration for external context providers (GitHub, GitLab, Jira, etc.).
/// Each provider can be enabled/disabled and configured with auth tokens.
/// Override individual tokens via env vars (GITHUB_TOKEN, GITLAB_TOKEN, etc.).
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct ProvidersConfig {
    /// Master switch for the provider subsystem.
    pub enabled: bool,
    /// GitHub provider configuration.
    pub github: ProviderEntryConfig,
    /// GitLab provider configuration.
    pub gitlab: ProviderEntryConfig,
    /// Auto-ingest provider results into BM25/embedding indexes.
    pub auto_index: bool,
    /// Default cache TTL for provider results (seconds).
    pub cache_ttl_secs: u64,
    /// MCP Bridge providers: `{ "name" = { url = "...", description = "..." } }`.
    #[serde(default)]
    pub mcp_bridges: std::collections::HashMap<String, McpBridgeEntry>,
}

impl Default for ProvidersConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            github: ProviderEntryConfig::default(),
            gitlab: ProviderEntryConfig::default(),
            auto_index: true,
            cache_ttl_secs: 120,
            mcp_bridges: std::collections::HashMap::new(),
        }
    }
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct McpBridgeEntry {
    /// HTTP/SSE URL for remote MCP servers.
    #[serde(default)]
    pub url: Option<String>,
    /// Command to spawn a local MCP server (stdio transport).
    #[serde(default)]
    pub command: Option<String>,
    /// Arguments for the command.
    #[serde(default)]
    pub args: Vec<String>,
    /// Human-readable description.
    #[serde(default)]
    pub description: Option<String>,
    /// Environment variable name containing an auth token.
    #[serde(default)]
    pub auth_env: Option<String>,
}

/// Per-provider configuration entry.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct ProviderEntryConfig {
    /// Whether this specific provider is enabled.
    pub enabled: bool,
    /// Auth token (prefer env var; only use this for project-local overrides).
    pub token: Option<String>,
    /// API base URL override (for GitHub Enterprise, self-hosted GitLab, etc.).
    pub api_url: Option<String>,
    /// Default project/repo for this provider (auto-detected from git remote if empty).
    pub project: Option<String>,
}

impl Default for ProviderEntryConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            token: None,
            api_url: None,
            project: None,
        }
    }
}

/// Controls autonomous background behaviors (preload, dedup, consolidation).
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct AutonomyConfig {
    pub enabled: bool,
    pub auto_preload: bool,
    pub auto_dedup: bool,
    pub auto_related: bool,
    pub auto_consolidate: bool,
    pub silent_preload: bool,
    pub dedup_threshold: usize,
    pub consolidate_every_calls: u32,
    pub consolidate_cooldown_secs: u64,
    #[serde(default = "serde_defaults::default_true")]
    pub cognition_loop_enabled: bool,
    #[serde(default = "serde_defaults::default_cognition_loop_interval")]
    pub cognition_loop_interval_secs: u64,
    #[serde(default = "serde_defaults::default_cognition_loop_max_steps")]
    pub cognition_loop_max_steps: u8,
    /// Minimum facts an entity needs before observation synthesis (#802) writes a
    /// summary. Synthesis itself is gated by `cognition_loop_max_steps >= 9`.
    #[serde(default = "serde_defaults::default_cognition_synthesis_min_cluster")]
    pub cognition_synthesis_min_cluster: usize,
}

impl Default for AutonomyConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            auto_preload: true,
            auto_dedup: true,
            auto_related: true,
            auto_consolidate: true,
            silent_preload: true,
            dedup_threshold: 8,
            consolidate_every_calls: 25,
            consolidate_cooldown_secs: 120,
            cognition_loop_enabled: true,
            cognition_loop_interval_secs: 3600,
            cognition_loop_max_steps: 9,
            cognition_synthesis_min_cluster: 3,
        }
    }
}

/// Controls automatic update behavior. All defaults are OFF — auto-updates
/// require explicit opt-in via `lean-ctx setup` or `lean-ctx update --schedule`.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct UpdatesConfig {
    pub auto_update: bool,
    pub check_interval_hours: u64,
    pub notify_only: bool,
}

impl Default for UpdatesConfig {
    fn default() -> Self {
        Self {
            auto_update: false,
            check_interval_hours: 6,
            notify_only: false,
        }
    }
}

/// Fixed-context budget accounting (#964). The per-session footprint lean-ctx
/// adds — tool schemas + MCP instructions + auto-loaded rules files + the wakeup
/// briefing — is warned about once it crosses `budget_tokens`. The
/// `LEAN_CTX_CONTEXT_BUDGET_TOKENS` env var overrides it; `lean-ctx doctor
/// overhead --gate` turns a breach into a non-zero exit for CI.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct ContextConfig {
    pub budget_tokens: usize,
}

impl Default for ContextConfig {
    fn default() -> Self {
        Self {
            budget_tokens: 8000,
        }
    }
}

impl UpdatesConfig {
    pub fn from_env() -> Self {
        let mut cfg = Self::default();
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_UPDATE") {
            cfg.auto_update = v == "1" || v.eq_ignore_ascii_case("true");
        }
        if let Ok(v) = std::env::var("LEAN_CTX_UPDATE_INTERVAL_HOURS")
            && let Ok(h) = v.parse::<u64>()
        {
            cfg.check_interval_hours = h.clamp(1, 168);
        }
        if let Ok(v) = std::env::var("LEAN_CTX_UPDATE_NOTIFY_ONLY") {
            cfg.notify_only = v == "1" || v.eq_ignore_ascii_case("true");
        }
        cfg
    }
}

impl AutonomyConfig {
    /// Creates an autonomy config from env vars, falling back to defaults.
    pub fn from_env() -> Self {
        let mut cfg = Self::default();
        if let Ok(v) = std::env::var("LEAN_CTX_AUTONOMY")
            && (v == "false" || v == "0")
        {
            cfg.enabled = false;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_PRELOAD") {
            cfg.auto_preload = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_DEDUP") {
            cfg.auto_dedup = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_RELATED") {
            cfg.auto_related = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_CONSOLIDATE") {
            cfg.auto_consolidate = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_SILENT_PRELOAD") {
            cfg.silent_preload = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_DEDUP_THRESHOLD")
            && let Ok(n) = v.parse()
        {
            cfg.dedup_threshold = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_CONSOLIDATE_EVERY_CALLS")
            && let Ok(n) = v.parse()
        {
            cfg.consolidate_every_calls = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_CONSOLIDATE_COOLDOWN_SECS")
            && let Ok(n) = v.parse()
        {
            cfg.consolidate_cooldown_secs = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_LOOP_ENABLED") {
            cfg.cognition_loop_enabled = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_LOOP_INTERVAL_SECS")
            && let Ok(n) = v.parse()
        {
            cfg.cognition_loop_interval_secs = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_LOOP_MAX_STEPS")
            && let Ok(n) = v.parse()
        {
            cfg.cognition_loop_max_steps = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_SYNTHESIS_MIN_CLUSTER")
            && let Ok(n) = v.parse()
        {
            cfg.cognition_synthesis_min_cluster = n;
        }
        cfg
    }

    /// Loads autonomy config from disk, with env var overrides applied.
    pub fn load() -> Self {
        let file_cfg = Config::load().autonomy;
        let mut cfg = file_cfg;
        if let Ok(v) = std::env::var("LEAN_CTX_AUTONOMY")
            && (v == "false" || v == "0")
        {
            cfg.enabled = false;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_PRELOAD") {
            cfg.auto_preload = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_DEDUP") {
            cfg.auto_dedup = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_AUTO_RELATED") {
            cfg.auto_related = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_SILENT_PRELOAD") {
            cfg.silent_preload = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_DEDUP_THRESHOLD")
            && let Ok(n) = v.parse()
        {
            cfg.dedup_threshold = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_LOOP_ENABLED") {
            cfg.cognition_loop_enabled = v != "false" && v != "0";
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_LOOP_INTERVAL_SECS")
            && let Ok(n) = v.parse()
        {
            cfg.cognition_loop_interval_secs = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_LOOP_MAX_STEPS")
            && let Ok(n) = v.parse()
        {
            cfg.cognition_loop_max_steps = n;
        }
        if let Ok(v) = std::env::var("LEAN_CTX_COGNITION_SYNTHESIS_MIN_CLUSTER")
            && let Ok(n) = v.parse()
        {
            cfg.cognition_synthesis_min_cluster = n;
        }
        cfg
    }
}

/// Cloud sync and contribution settings (pattern sharing, model pulls).
#[derive(Debug, Clone, Serialize, Deserialize, Default)]
#[serde(default)]
pub struct CloudConfig {
    pub contribute_enabled: bool,
    pub last_contribute: Option<String>,
    pub last_sync: Option<String>,
    pub last_gain_sync: Option<String>,
    pub last_model_pull: Option<String>,
    /// Auto-push the Pro Personal-Cloud surfaces (knowledge, commands, CEP,
    /// gotchas, buddy, feedback) from the background task — opt-in, once per
    /// day, offline-tolerant (GL #384). Toggle: `lean-ctx cloud autosync on`.
    pub auto_sync: bool,
    pub last_auto_sync: Option<String>,
    /// Auto-push the project's encrypted retrieval-index bundle (hosted
    /// Personal Index, GL #392) alongside the daily auto-sync — separate
    /// opt-in because index bundles are orders of magnitude larger than the
    /// other surfaces. Toggle: `lean-ctx cloud autoindex on`.
    pub auto_index: bool,
    /// Per-project debounce: `project_hash → YYYY-MM-DD` of the last
    /// successful background index push.
    pub last_index_push: std::collections::HashMap<String, String>,
}

/// Settings for publishing your token-savings recap (`gain --publish` / auto-publish).
///
/// Publishing is always opt-in: it sends a small, whitelisted *aggregate* payload (tokens
/// saved, $ avoided, compression % — never code, paths or counts) to the cloud.
/// `auto_publish` simply removes the need to re-run `gain --publish` by hand; it stays off
/// until the user explicitly enables it.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct GainConfig {
    /// When true, `lean-ctx gain` automatically (re)publishes the recap, throttled to
    /// `auto_publish_interval_hours`. Off by default.
    pub auto_publish: bool,
    /// When auto-publishing, also opt into the public leaderboard.
    pub leaderboard: bool,
    /// Optional display name for the published card / leaderboard entry.
    pub display_name: Option<String>,
    /// Minimum hours between automatic publishes (throttle).
    pub auto_publish_interval_hours: u64,
    /// Runtime state — RFC3339 timestamp of the last automatic publish. Managed by the
    /// tool, not meant to be set by hand.
    pub last_auto_publish: Option<String>,
}

impl Default for GainConfig {
    fn default() -> Self {
        Self {
            auto_publish: false,
            leaderboard: true,
            display_name: None,
            auto_publish_interval_hours: 24,
            last_auto_publish: None,
        }
    }
}

/// Model declaration for **measured-vs-estimated** cost reporting.
///
/// Proxy-routed clients (Claude Code, Codex, Pi, Gemini CLI, OpenCode) report
/// their real model and billed tokens, so lean-ctx prices them *measured* with
/// no configuration. MCP-only IDEs (Cursor, Copilot, Windsurf, VS Code, Zed)
/// send their LLM traffic straight to the provider, bypassing lean-ctx — their
/// real model is invisible. Declaring it here lets those *estimated* turns be
/// priced with the correct model instead of a blended fallback.
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
#[serde(default)]
pub struct CostConfig {
    /// Per-session cost cap in USD. When accumulated cost exceeds this value,
    /// subsequent tool calls receive a `[COST CAP]` warning instead of the
    /// normal output (#794). 0 = unlimited (default).
    /// Override at runtime: `LEAN_CTX_COST_CAP_OVERRIDE=1` bypasses the cap.
    #[serde(default)]
    pub max_session_cost_usd: f64,
    /// Fallback pricing model for any client without a per-client entry.
    /// Unset/empty → lean-ctx keeps its blended heuristic.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub default_model: Option<String>,
    /// Per-client pricing model, keyed by client id (`cursor`, `copilot`,
    /// `windsurf`, `claude`, `codex`, …). Used for MCP-only IDEs whose real
    /// model lean-ctx cannot observe. Example:
    /// `[cost.models]` then `cursor = "claude-opus-4.5"`.
    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
    pub models: HashMap<String, String>,
    /// Operator price overrides (#1189), keyed by model name — for negotiated
    /// enterprise rates (committed-use discounts, Azure PTU, zero-rated
    /// internal models) that no public catalog can know. Merged into the
    /// pricing table as **exact** entries, overriding embedded and live rows;
    /// only a provider-measured bill beats them. Example:
    /// `[cost.prices."internal-llm"]` then `input_per_m = 0.10`,
    /// `output_per_m = 0.40`.
    #[serde(default, skip_serializing_if = "HashMap::is_empty")]
    pub prices: HashMap<String, PriceOverride>,
}

/// One `[cost.prices.<model>]` row: USD per million tokens. Omitted cache
/// rates default to the input rate (the same convention the catalogs use).
#[derive(Debug, Clone, Default, Serialize, Deserialize)]
#[serde(default)]
pub struct PriceOverride {
    pub input_per_m: Option<f64>,
    pub output_per_m: Option<f64>,
    pub cache_write_per_m: Option<f64>,
    pub cache_read_per_m: Option<f64>,
}

impl CostConfig {
    /// Configured pricing model for a client id: the per-client entry first, then
    /// the global default. `None` when neither is set (the caller then falls back
    /// to the env override / heuristic). Blank entries are ignored.
    pub fn model_for_client(&self, client: &str) -> Option<String> {
        self.models
            .get(client)
            .or(self.default_model.as_ref())
            .map(|s| s.trim().to_string())
            .filter(|s| !s.is_empty())
    }
}

/// Code-health engine (`[code_health]`): clean code as a token-cost lever.
///
/// Cognitive complexity, naming quality, and coupling are computed once during
/// indexing and surfaced at read- and edit-time. These switches tune the
/// thresholds and how assertively findings are surfaced.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct CodeHealthConfig {
    /// Cognitive-complexity threshold above which a function is a hotspot.
    /// Mirrors `core::code_health::DEFAULT_COGNITIVE_THRESHOLD` (15).
    pub cognitive_threshold: u32,
    /// Edit-gate behavior on complexity drift: `"warn"` (annotate, default),
    /// `"block"` (refuse clean→over-threshold edits), or `"off"`.
    pub gate: String,
    /// Annotate over-threshold functions inline in `ctx_read` output.
    pub annotate_reads: bool,
    /// Run the naming-quality heuristic.
    pub naming: bool,
    /// Compute module-coupling metrics.
    pub coupling: bool,
    /// Inject `[CODE HEALTH]` notices as `additionalContext` in PostToolUse stdout.
    /// Default: **false** — prevents prompt-cache invalidation on Anthropic models
    /// (#778: each injection causes 440-520k tokens of cache re-bills when Claude
    /// Code strips stale system-reminders retroactively).
    /// When false, notices route to `ctx_knowledge` + dashboard instead.
    #[serde(default)]
    pub inject_context: bool,
}

impl Default for CodeHealthConfig {
    fn default() -> Self {
        Self {
            cognitive_threshold: 15,
            gate: "warn".to_string(),
            annotate_reads: true,
            naming: true,
            coupling: true,
            inject_context: false,
        }
    }
}

/// Index-time file filters (#735): declare the retrieval corpus explicitly
/// instead of abusing `.gitignore` for retrieval policy.
///
/// Applies to every index builder through one shared filter layer
/// (`core::index_filter`): BM25, graph, and the watch/incremental path; the
/// semantic index chunks the BM25 corpus and inherits the same universe.
/// Excluded files never produce chunks, graph nodes, or embeddings. Globs are
/// matched against the root-relative path (forward slashes); exclude wins
/// over include. The empty default preserves today's behavior byte-for-byte.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct IndexConfig {
    /// Honor `.gitignore` / global gitignore / `.git/info/exclude` during
    /// index walks. `false` indexes ignored files too (rarely wanted; the
    /// vendor-directory guard still applies).
    pub respect_gitignore: bool,
    /// Files to drop from the index corpus, e.g. `["**/*.csv", "fixtures/**"]`.
    /// Evaluated after `include`; a file matching both is excluded.
    pub exclude: Vec<String>,
    /// When non-empty, ONLY matching files enter the index corpus, e.g.
    /// `["**/*.rs", "**/*.ts"]`. Empty = no restriction.
    pub include: Vec<String>,
}

impl Default for IndexConfig {
    fn default() -> Self {
        Self {
            respect_gitignore: true,
            exclude: Vec::new(),
            include: Vec::new(),
        }
    }
}

/// Settings for the code graph — in particular the *traversal* (co-access) edges
/// learned from real agent sessions (#289).
///
/// The static AST/import graph captures how code is wired structurally; it cannot
/// see which files an agent actually opens *together* while solving a task.
/// Traversal edges add that behavioural signal: files surfaced together are
/// associated with a decaying weight (Hebbian co-access), folded into the graph
/// as `co_access` edges and mixed into recall. The store is bounded and decays,
/// so stale associations fade.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct GraphConfig {
    /// Record co-access between files surfaced together in a session, surface them
    /// as decaying `co_access` edges in the graph, and boost recall by them.
    /// On by default; set to `false` for a purely static (AST-only) graph.
    pub traversal_edges: bool,
}

impl Default for GraphConfig {
    fn default() -> Self {
        Self {
            traversal_edges: true,
        }
    }
}

/// Skillify (#290): mine the project's session diary + knowledge facts into
/// versioned, git-committable `.cursor/rules/skillify-*.mdc` rule files.
///
/// The miner is precision-biased — it only codifies recurring or high-confidence
/// patterns and never invents content. Runs on demand (`ctx_skillify` /
/// `lean-ctx skillify`); re-running merges (bumps version) only when the distilled
/// content actually changes.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct SkillifyConfig {
    /// Master switch for the skillify miner. On by default; the miner only ever
    /// acts when explicitly invoked, so this never writes files unprompted.
    pub enabled: bool,
    /// Where generated rules are written: `project` (`<repo>/.cursor/rules`,
    /// git-committable, default) or `global` (`~/.cursor/rules`).
    pub scope: String,
    /// Minimum confidence for a single curated knowledge fact to be codified even
    /// without repetition. 0.0..=1.0.
    pub min_confidence: f32,
    /// Minimum number of reinforcements (confirmations / repeated mentions) before
    /// a pattern is codified when its confidence is below `min_confidence`.
    pub min_recurrence: u32,
}

impl Default for SkillifyConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            scope: "project".to_string(),
            min_confidence: 0.7,
            min_recurrence: 2,
        }
    }
}

/// AI session summaries (#292): periodically distil the working session into a
/// compact, *semantically recallable* summary so a future session can answer
/// "what did I do last time on X?". Deterministic and local-first — recall uses
/// embeddings when the `embeddings` feature is on, else a lexical fallback.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct SummariesConfig {
    /// Record periodic session summaries. On by default; recording is cheap and
    /// happens at most once per `every_n_turns` tool calls.
    pub enabled: bool,
    /// Tool calls between automatic summaries. The auto-checkpoint cadence still
    /// gates the check, so the effective minimum is the checkpoint interval.
    pub every_n_turns: u32,
    /// Maximum summaries kept per project (oldest pruned first).
    pub max_kept: u32,
}

impl Default for SummariesConfig {
    fn default() -> Self {
        Self {
            enabled: true,
            every_n_turns: 25,
            max_kept: 100,
        }
    }
}

/// A user-defined command alias mapping for shell compression patterns.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AliasEntry {
    pub command: String,
    pub alias: String,
}

/// Thresholds for detecting and throttling repetitive agent tool call loops.
#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(default)]
pub struct LoopDetectionConfig {
    pub normal_threshold: u32,
    pub reduced_threshold: u32,
    pub blocked_threshold: u32,
    pub window_secs: u64,
    pub search_group_limit: u32,
    pub tool_total_limits: HashMap<String, u32>,
}

impl Default for LoopDetectionConfig {
    fn default() -> Self {
        let mut tool_total_limits = HashMap::new();
        tool_total_limits.insert("ctx_read".to_string(), 100);
        tool_total_limits.insert("ctx_search".to_string(), 80);
        tool_total_limits.insert("ctx_shell".to_string(), 50);
        tool_total_limits.insert("ctx_semantic_search".to_string(), 60);
        Self {
            normal_threshold: 2,
            reduced_threshold: 4,
            blocked_threshold: 0,
            window_secs: 300,
            search_group_limit: 10,
            tool_total_limits,
        }
    }
}

/// Semantic-embedding engine settings.
///
/// `model` selects which local ONNX embedding model lean-ctx downloads and uses for
/// `ctx_semantic_search`. Accepts the same aliases as the `LEAN_CTX_EMBEDDING_MODEL` env
/// var: `minilm` (all-MiniLM-L6-v2, 384d — the default), `nomic` (768d) — or any
/// HuggingFace repo with an ONNX export via `hf:org/repo[@revision]` (GL #397), e.g.
/// `hf:jinaai/jina-embeddings-v2-base-code` for code-specialized embeddings. When the
/// env var is set it takes precedence; an
/// unset/`None` value uses the default model. Switching models triggers a one-time
/// re-index on the next semantic search (vector dimensions follow from the model).
///
/// `dimensions` is only consulted for `hf:` custom models as the declared fallback
/// width; the real width is probed from the ONNX graph at load time. Built-ins ignore it.
/// `[gateway_server]` — deployment parameters of the self-hosted org gateway
/// (enterprise#20). Distinct from `[gateway]` (the MCP tool-catalog gateway):
/// this section describes the LLM-proxy *server* deployment and its cockpit.
///
/// All fields optional; an empty section keeps every local behavior unchanged.
#[derive(Debug, Clone, Default, Serialize, Deserialize, PartialEq)]
#[serde(default)]
pub struct GatewayServerConfig {
    /// Seats the org-wide projection extrapolates to (e.g. `800`). `None`
    /// disables the projection — the cockpit never invents a seat count.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub seats: Option<u32>,
    /// Display label for the cockpit header (e.g. `"Zühlke AI Gateway"`).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub org_label: Option<String>,
    /// Central admin API base URL (e.g. `https://ai-gateway.example.com`).
    /// When set, the local cockpit's usage breakdown reads the org-wide
    /// `GET /api/admin/usage` instead of the machine-local snapshot. The
    /// bearer token comes from `LEAN_CTX_GATEWAY_ADMIN_TOKEN` (never config).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub admin_url: Option<String>,
    /// Bind address of the admin listener (dashboard + `/api/admin/*` +
    /// `/metrics`). Defaults to loopback — **secure by default** (#54/#56):
    /// exposing the console is an explicit decision. Container deployments set
    /// `"0.0.0.0"` here (the pod/compose port mapping stays the outer guard).
    /// `LEAN_CTX_GATEWAY_ADMIN_BIND_HOST` overrides. Invalid values fall back
    /// to loopback: a typo can only ever narrow exposure, never open it.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub admin_bind_host: Option<String>,
    /// Days to keep `usage_events` rows (enterprise#36). `None`/`0` = keep
    /// forever (the local-free default — retention is a deployment decision).
    /// A running gateway purges older rows periodically; typical compliance
    /// values are `365` or `3650` (EU AI Act evidence horizon).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub usage_retention_days: Option<u32>,
    /// Replace `person` with a stable keyed pseudonym (`p:<hash>`) before it
    /// reaches metering, budgets, dashboards and logs (enterprise#39, GDPR).
    /// The salt lives in `<data_dir>/gateway_pii_salt`; `gateway gdpr`
    /// re-derives pseudonyms from e-mail input, so DSGVO delete/export keep
    /// working. Default `false` (cleartext person tags).
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub pseudonymize_persons: Option<bool>,
    /// MCP upstream registry (GL#91/#99, Doc 15 §7 — the observe stage of MCP
    /// context governance). Each entry publishes a governed reverse-proxy
    /// route `/mcp/{id}` on the proxy port: same per-person key auth as the
    /// LLM channel, tool calls metered into `mcp_events`, tool definitions
    /// inventoried + hash-tracked (rug-pull detection). Observe-only: the
    /// gateway never blocks or rewrites MCP traffic in this stage.
    #[serde(default, skip_serializing_if = "Vec::is_empty")]
    pub mcp_servers: Vec<McpServerEntry>,
}

/// One `[[gateway_server.mcp_servers]]` registry entry — an MCP server the org
/// gateway fronts. Distinct from `[[gateway.servers]]` (the *local* tool-
/// catalog aggregator, #210): this registry is the org-facing reverse proxy.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq)]
pub struct McpServerEntry {
    /// Registry id, used in the `/mcp/{id}` route. Lowercase alphanumeric
    /// plus `-`/`_` (it becomes a URL path segment).
    pub id: String,
    /// Upstream Streamable-HTTP endpoint (the server's single MCP endpoint,
    /// e.g. `https://mcp.example.com/mcp`). HTTPS for any non-loopback host;
    /// plaintext HTTP needs the same explicit opt-in as LLM upstreams
    /// (`[proxy] allow_insecure_http_upstream`).
    pub url: String,
    /// Name of the environment variable holding the upstream credential. When
    /// set, the gateway sends `Authorization: Bearer <value>` upstream — the
    /// credential lives in the gateway's environment, never on laptops. The
    /// caller's own `Authorization` header (their gateway key) is **always**
    /// stripped before forwarding, with or without this field.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub auth_env: Option<String>,
    /// Set `false` to keep the entry in config but take it out of service.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub enabled: Option<bool>,
}

/// A validated, ready-to-serve MCP registry entry (runtime view of
/// [`McpServerEntry`]).
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct ResolvedMcpServer {
    pub id: String,
    pub url: String,
    pub auth_env: Option<String>,
}

impl GatewayServerConfig {
    /// Validate + resolve the `[[gateway_server.mcp_servers]]` registry.
    /// Same resilience contract as `[[proxy.providers]]`: invalid entries are
    /// logged and skipped (one typo never takes the gateway down), duplicates
    /// keep the first occurrence. `allow_insecure_http` mirrors the proxy's
    /// plaintext-HTTP opt-in so the two registries share one security posture.
    #[must_use]
    pub fn resolve_mcp_servers(&self, allow_insecure_http: bool) -> Vec<ResolvedMcpServer> {
        let mut seen: std::collections::BTreeSet<&str> = std::collections::BTreeSet::new();
        let mut out = Vec::new();
        for entry in &self.mcp_servers {
            if !entry.enabled.unwrap_or(true) {
                continue;
            }
            let id = entry.id.trim();
            if !is_valid_mcp_server_id(id) {
                tracing::warn!(
                    "[gateway_server.mcp_servers] invalid id '{id}' \
                     (lowercase alnum/-/_ only) — entry skipped"
                );
                continue;
            }
            if !seen.insert(id) {
                tracing::warn!(
                    "[gateway_server.mcp_servers] duplicate id '{id}' — keeping first entry"
                );
                continue;
            }
            match validate_mcp_upstream_url(&entry.url, allow_insecure_http) {
                Ok(url) => out.push(ResolvedMcpServer {
                    id: id.to_string(),
                    url,
                    auth_env: entry
                        .auth_env
                        .as_deref()
                        .map(str::trim)
                        .filter(|v| !v.is_empty())
                        .map(str::to_string),
                }),
                Err(e) => {
                    tracing::warn!(
                        "[gateway_server.mcp_servers] '{id}' has invalid url — skipped: {e}"
                    );
                }
            }
        }
        out
    }

    /// Effective admin bind address (see `admin_bind_host`). Precedence:
    /// `LEAN_CTX_GATEWAY_ADMIN_BIND_HOST` env > config > `127.0.0.1`.
    #[must_use]
    pub fn resolved_admin_bind_host(&self) -> std::net::IpAddr {
        let raw = std::env::var("LEAN_CTX_GATEWAY_ADMIN_BIND_HOST")
            .ok()
            .filter(|v| !v.trim().is_empty())
            .or_else(|| self.admin_bind_host.clone());
        match raw.as_deref().map(str::trim) {
            Some(v) if !v.is_empty() => v.parse().unwrap_or_else(|_| {
                tracing::warn!(
                    "gateway_server.admin_bind_host '{v}' is not a valid IP address — binding 127.0.0.1"
                );
                std::net::IpAddr::V4(std::net::Ipv4Addr::LOCALHOST)
            }),
            _ => std::net::IpAddr::V4(std::net::Ipv4Addr::LOCALHOST),
        }
    }
}

/// True when `id` is usable as an MCP registry id: non-empty, lowercase alnum
/// plus `-`/`_` (it becomes a URL path segment). Same shape rule as
/// `[[proxy.providers]]` ids; no built-in namespace exists to shadow here.
fn is_valid_mcp_server_id(id: &str) -> bool {
    !id.is_empty()
        && id
            .chars()
            .all(|c| c.is_ascii_lowercase() || c.is_ascii_digit() || c == '-' || c == '_')
}

/// Validates an MCP upstream URL. A declared registry entry is itself the
/// deliberate custom-host opt-in (same rationale as `[[proxy.providers]]`):
/// any HTTPS host is accepted; loopback HTTP is always fine; non-loopback
/// plaintext HTTP requires the explicit insecure-HTTP opt-in. This is the
/// SSRF boundary — the proxy only ever connects to URLs that passed here.
fn validate_mcp_upstream_url(url: &str, allow_insecure_http: bool) -> Result<String, String> {
    let trimmed = url.trim().trim_end_matches('/');
    if trimmed.is_empty() {
        return Err("empty url".into());
    }
    if crate::core::config::is_local_proxy_url(trimmed) {
        return Ok(trimmed.to_string());
    }
    if trimmed.starts_with("http://") {
        if allow_insecure_http {
            return Ok(trimmed.to_string());
        }
        return Err(format!(
            "MCP upstream must use HTTPS: {trimmed} (for a trusted local-network HTTP \
             upstream opt in with `[proxy] allow_insecure_http_upstream = true`)"
        ));
    }
    if trimmed.starts_with("https://") {
        return Ok(trimmed.to_string());
    }
    Err(format!(
        "MCP upstream must start with http:// or https://: {trimmed}"
    ))
}

#[derive(Debug, Clone, Default, Serialize, Deserialize)]
#[serde(default)]
pub struct EmbeddingConfig {
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub model: Option<String>,
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub dimensions: Option<usize>,
    /// Allow downloading the embedding model on first semantic need (#551).
    /// `None` (unset) means **allowed** — the soft default that activates the
    /// semantic features without manual setup. Set `false` for air-gapped
    /// machines. The `LEAN_CTX_EMBEDDINGS_AUTO_DOWNLOAD` env var, when set,
    /// overrides this in either direction.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub auto_download: Option<bool>,
    /// Pin embedding inference to a single CPU thread (no GPU EP) so vectors are
    /// bit-identical across machines, not just run-to-run on one host (#895).
    /// `None`/`false` keeps the multi-threaded GPU-capable path. Extractive prose
    /// ranking is already deterministic via score quantization + stable tiebreak;
    /// this flag is the extra hardening for cross-machine reproducibility. The
    /// `LEAN_CTX_EMBEDDING_DETERMINISTIC` env var overrides this either way.
    #[serde(default, skip_serializing_if = "Option::is_none")]
    pub deterministic: Option<bool>,
}

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

    #[test]
    fn admin_bind_defaults_to_loopback_and_rejects_garbage() {
        // Secure by default (#54/#56): unset and invalid both land on loopback.
        let cfg = GatewayServerConfig::default();
        assert!(cfg.resolved_admin_bind_host().is_loopback());

        let cfg = GatewayServerConfig {
            admin_bind_host: Some("not-an-ip".into()),
            ..Default::default()
        };
        assert!(
            cfg.resolved_admin_bind_host().is_loopback(),
            "a typo must narrow exposure, never widen it"
        );

        let cfg = GatewayServerConfig {
            admin_bind_host: Some("0.0.0.0".into()),
            ..Default::default()
        };
        assert!(
            !cfg.resolved_admin_bind_host().is_loopback(),
            "explicit opt-in widens the bind"
        );
    }

    fn mcp_entry(id: &str, url: &str) -> McpServerEntry {
        McpServerEntry {
            id: id.into(),
            url: url.into(),
            auth_env: None,
            enabled: None,
        }
    }

    #[test]
    fn mcp_registry_validates_ids_urls_and_duplicates() {
        let cfg = GatewayServerConfig {
            mcp_servers: vec![
                mcp_entry("github", "https://mcp.example.com/mcp/"),
                // invalid id (uppercase) — skipped, never panics
                mcp_entry("GitHub", "https://mcp.example.com/mcp"),
                // duplicate — first occurrence wins
                mcp_entry("github", "https://other.example.com/mcp"),
                // plaintext HTTP on a non-loopback host without the opt-in — skipped
                mcp_entry("plain", "http://mcp.example.com/mcp"),
                // loopback HTTP is always fine (local/dev)
                mcp_entry("local", "http://127.0.0.1:9200/mcp"),
                McpServerEntry {
                    enabled: Some(false),
                    ..mcp_entry("disabled", "https://mcp.example.com/mcp")
                },
                McpServerEntry {
                    auth_env: Some("  GITHUB_MCP_PAT  ".into()),
                    ..mcp_entry("authed", "https://api.githubcopilot.com/mcp")
                },
            ],
            ..Default::default()
        };
        let resolved = cfg.resolve_mcp_servers(false);
        let ids: Vec<&str> = resolved.iter().map(|s| s.id.as_str()).collect();
        assert_eq!(ids, ["github", "local", "authed"]);
        // Trailing slash normalized; the duplicate kept the first URL.
        assert_eq!(resolved[0].url, "https://mcp.example.com/mcp");
        assert_eq!(resolved[2].auth_env.as_deref(), Some("GITHUB_MCP_PAT"));

        // The insecure-HTTP opt-in admits the plaintext entry (trusted LAN).
        let with_optin = cfg.resolve_mcp_servers(true);
        assert!(with_optin.iter().any(|s| s.id == "plain"));
    }

    #[test]
    fn mcp_upstream_url_rules_match_the_proxy_posture() {
        assert!(validate_mcp_upstream_url("https://mcp.example.com/mcp", false).is_ok());
        assert!(validate_mcp_upstream_url("http://localhost:9200/mcp", false).is_ok());
        assert!(validate_mcp_upstream_url("http://mcp.example.com/mcp", false).is_err());
        assert!(validate_mcp_upstream_url("http://mcp.example.com/mcp", true).is_ok());
        assert!(validate_mcp_upstream_url("ftp://mcp.example.com", false).is_err());
        assert!(validate_mcp_upstream_url("   ", false).is_err());
    }
}