harn-vm 0.7.34

//! LLM provider trait and registry.
//!
//! Defines the `LlmProvider` trait that all LLM backends implement, plus a
//! thread-local registry that tracks which providers are available. Provider
//! dispatch happens in `api.rs` via `dispatch_to_registered_provider()`.

use std::cell::RefCell;
use std::collections::HashSet;

use super::api::{DeltaSender, LlmRequestPayload, LlmResult};
use crate::value::VmError;

/// Trait that all LLM providers implement.
///
/// Dispatch currently goes through concrete types in
/// `api::dispatch_to_registered_provider`. The trait exists so that
/// custom/external providers can be registered at runtime once the
/// `provider_register()` Harn builtin is exposed.
#[allow(dead_code)]
pub(crate) trait LlmProvider {
    /// Provider name (e.g. "anthropic", "openai", "ollama", "mock").
    fn name(&self) -> &str;

    /// Whether this provider uses Anthropic-style messages API (vs OpenAI-style).
    fn is_anthropic_style(&self) -> bool {
        false
    }

    /// Whether this provider supports prompt caching.
    fn supports_cache(&self) -> bool {
        false
    }

    /// Whether this provider supports extended thinking.
    fn supports_thinking(&self) -> bool {
        false
    }

    /// Whether this is the mock provider (deterministic test responses, no API).
    fn is_mock(&self) -> bool {
        false
    }

    /// Whether this is a local provider (e.g. Ollama) that uses NDJSON streaming.
    fn is_local(&self) -> bool {
        false
    }

    /// Whether the provider requires a model to be specified.
    fn requires_model(&self) -> bool {
        true
    }

    /// Apply provider-specific transformations to the request body after it has
    /// been built by `build_request_body()`. Default is a no-op.
    fn transform_request(&self, _body: &mut serde_json::Value) {}

    /// Whether this provider's native API accepts a `defer_loading: true`
    /// flag on tool definitions — keeping their schema out of the model's
    /// context until a tool-search call surfaces them. See Anthropic's tool
    /// search docs and OpenAI's Responses API `tool_search` guide.
    ///
    /// Default impl reads from `capabilities.toml` so new providers and
    /// new model generations ship as data, not code. Override only to
    /// short-circuit the lookup (e.g. the mock provider, which the
    /// capability layer already handles specially).
    fn supports_defer_loading(&self, model: &str) -> bool {
        super::capabilities::lookup(self.name(), model).defer_loading
    }

    /// Native tool-search variants this provider supports at the given
    /// model. Reads the capabilities matrix:
    ///   - `[]` — no native support; callers fall back to the client-
    ///     executed search (harn#70).
    ///   - `["bm25", "regex"]` — Anthropic's two
    ///     `tool_search_tool_*_20251119` types.
    ///   - `["hosted", "client"]` — OpenAI Responses API `tool_search`
    ///     execution modes.
    ///
    /// Ordering is the provider's recommended default first. Callers
    /// that don't care which variant they get pick element 0.
    fn native_tool_search_variants(&self, model: &str) -> Vec<String> {
        super::capabilities::lookup(self.name(), model).tool_search
    }
}

/// Async chat operation. Uses explicit lifetime parameters because providers
/// are constructed on-the-fly, to avoid RefCell-across-await issues.
#[allow(dead_code)]
pub(crate) trait LlmProviderChat: LlmProvider {
    /// Execute an LLM chat call, optionally streaming text deltas.
    fn chat<'a>(
        &'a self,
        request: &'a LlmRequestPayload,
        delta_tx: Option<DeltaSender>,
    ) -> std::pin::Pin<Box<dyn std::future::Future<Output = Result<LlmResult, VmError>> + 'a>>;
}

thread_local! {
    /// Thread-local for !Send VM compatibility. Provider objects are
    /// constructed on-the-fly to avoid RefCell-across-await issues.
    static PROVIDER_NAMES: RefCell<HashSet<String>> = RefCell::new(HashSet::new());
}

/// Register all built-in providers. Called once per thread at VM startup.
pub(crate) fn register_default_providers() {
    PROVIDER_NAMES.with(|names| {
        let mut names = names.borrow_mut();
        if !names.is_empty() {
            return;
        }
        names.insert("mock".to_string());
        names.insert("anthropic".to_string());
        names.insert("ollama".to_string());
        for name in [
            "openai",
            "openrouter",
            "together",
            "groq",
            "deepseek",
            "fireworks",
            "huggingface",
            "local",
            "vllm",
            "tgi",
            "dashscope",
        ] {
            names.insert(name.to_string());
        }
    });
}

/// Register a custom provider name at runtime.
#[allow(dead_code)]
pub(crate) fn register_provider_name(name: &str) {
    PROVIDER_NAMES.with(|names| {
        names.borrow_mut().insert(name.to_string());
    });
}

/// Check whether a named provider is registered.
pub(crate) fn is_provider_registered(name: &str) -> bool {
    PROVIDER_NAMES.with(|names| names.borrow().contains(name))
}

/// Return all registered provider names (used by diagnostics and tests).
#[allow(dead_code)]
pub(crate) fn registered_provider_names() -> Vec<String> {
    PROVIDER_NAMES.with(|names| names.borrow().iter().cloned().collect())
}

/// Module-level dispatch for `LlmProvider::supports_defer_loading`.
///
/// Thin wrapper over `capabilities::lookup`. Kept as a named export
/// because `options.rs` reads better with the predicate form than with
/// an inline `.defer_loading` field access, and because custom
/// providers registered at runtime (via `provider_register`) still
/// flow through this function rather than carrying a trait object.
pub(crate) fn provider_supports_defer_loading(provider: &str, model: &str) -> bool {
    super::capabilities::lookup(provider, model).defer_loading
}

/// Module-level dispatch for `LlmProvider::native_tool_search_variants`.
/// Reads the capability matrix — the single source of truth since
/// `capabilities.toml` replaced the per-provider hard-coded gates.
pub(crate) fn provider_tool_search_variants(provider: &str, model: &str) -> Vec<String> {
    super::capabilities::lookup(provider, model).tool_search
}

/// Which wire shape to emit for the native tool-search meta-tool. Kept
/// in one place so the options layer, the tools builder, and the
/// response parser all agree on who emits what. Anthropic emits
/// `tool_search_tool_*_20251119` meta-tools; OpenAI-shape providers
/// emit `{"type": "tool_search"}`.
#[derive(Clone, Copy, Debug, PartialEq, Eq)]
pub(crate) enum NativeToolSearchShape {
    /// Anthropic's `{"type": "tool_search_tool_{bm25,regex}_20251119"}`.
    Anthropic,
    /// OpenAI Responses API's `{"type": "tool_search"}`.
    OpenAi,
}