feat(llm, google): Add support for Gemini models (#133)

This commit introduces support for Google's Gemini models as a new LLM
provider. This allows users to leverage the Gemini family of models,
including `gemini-1.5-pro` and `gemini-1.5-flash`, for chat completions.

This was possible already, using the `openrouter` provider, but this
adds first-party support for Google's API (reducing cost, not having to
agree to Openrouter's terms of service, improving latency, leveraging
Google-specific APIs).

The integration is built on top of the `gemini_client_rs` crate (which
was fairly heavily modified for this purpose in
Adriftdev/gemini-client#8). A new `Google`
provider struct implements the `Provider` trait, handling the
translation between the internal conversation format and the Gemini
API's request and response structure.

This includes full support for streaming responses, tool use, and model
parameters like temperature and token limits. The provider also
correctly handles system prompts and attachments by formatting them into
the multi-turn conversation format expected by Gemini.

Configuration for the Google provider, including the API key environment
variable (`GEMINI_API_KEY`) and a customizable `base_url`, has been
added to `jp_config`.

As a proof-of-concept, this commit was generated using the following
command:

```sh
jp query \
    --model=google/gemini-2.5-pro-preview-06-05 \
    --no-persist \
    --new \
    --context=commit \
    --hide-reasoning \
    --no-tool "Give me a commit message"
```

Signed-off-by: Jean Mertz <git@jeanmertz.com>

Author: JeanMertz

Cargo.lock

@@ -39,6 +39,8 @@ duct = { version = "1", default-features = false }
 dyn-clone = { version = "1", default-features = false }
 dyn-hash = { version = "0.2", default-features = false }
 futures = { version = "0.3", default-features = false }
+# See: <https://github.com/Adriftdev/gemini-client/pull/8>
+gemini_client_rs = { git = "https://github.com/JeanMertz/gemini-client", default-features = false }
 glob = { version = "0.3", default-features = false }
 httpmock = { git = "https://github.com/alexliesenfeld/httpmock", default-features = false }
 ignore = { version = "0.4", default-features = false }
@@ -114,6 +116,7 @@ format_push_string = "allow"
 missing_errors_doc = "allow" # Temporary
 option_option = "allow"
 pedantic = { level = "warn", priority = -1 }
+result_large_err = "allow"
 similar_names = "allow"
 struct_excessive_bools = "allow"
 struct_field_names = "allow"

Cargo.toml

@@ -399,6 +399,11 @@ impl From<jp_llm::Error> for Error {
                 ("error", error.to_string()),
             ]
             .into(),
+            Gemini(error) => [
+                ("message", "Gemini error".into()),
+                ("error", error.to_string()),
+            ]
+            .into(),
         };
 
         Self::from(metadata)

crates/jp_cli/src/cmd.rs

@@ -8,12 +8,20 @@ pub struct Config {
     /// Environment variable that contains the API key.
     #[config(default = "GEMINI_API_KEY", env = "JP_LLM_PROVIDER_GOOGLE_API_KEY_ENV")]
     pub api_key_env: String,
+
+    /// The base URL to use for API requests.
+    #[config(
+        default = "https://generativelanguage.googleapis.com/v1beta",
+        env = "JP_LLM_PROVIDER_GOOGLE_BASE_URL"
+    )]
+    pub base_url: String,
 }
 
 impl Default for Config {
     fn default() -> Self {
         Self {
             api_key_env: "GEMINI_API_KEY".to_owned(),
+            base_url: "https://generativelanguage.googleapis.com/v1beta".to_owned(),
         }
     }
 }
@@ -23,6 +31,7 @@ impl Config {
     pub fn set(&mut self, path: &str, key: &str, value: impl Into<String>) -> Result<()> {
         match key {
             "api_key_env" => self.api_key_env = value.into(),
+            "base_url" => self.base_url = value.into(),
             _ => return crate::set_error(path, key),
         }
 

crates/jp_config/src/llm/provider/google.rs

@@ -23,6 +23,7 @@ async-anthropic = { workspace = true }
 async-stream = { workspace = true }
 async-trait = { workspace = true }
 futures = { workspace = true }
+gemini_client_rs = { workspace = true }
 ollama-rs = { workspace = true, features = ["rustls", "stream"] }
 openai_responses = { workspace = true, features = ["stream"] }
 reqwest = { workspace = true }

crates/jp_llm/Cargo.toml

@@ -35,6 +35,9 @@ pub enum Error {
         response: String,
     },
 
+    #[error("Gemini error: {0}")]
+    Gemini(#[from] gemini_client_rs::GeminiError),
+
     #[error("Ollama error: {0}")]
     Ollama(#[from] ollama_rs::error::OllamaError),
 

crates/jp_llm/src/error.rs

@@ -1,5 +1,5 @@
 // pub mod deepseek;
-// pub mod google;
+pub mod google;
 // pub mod xai;
 pub mod anthropic;
 pub mod ollama;
@@ -11,6 +11,7 @@ use std::{mem, pin::Pin};
 use anthropic::Anthropic;
 use async_trait::async_trait;
 use futures::{Stream, StreamExt as _};
+use google::Google;
 use jp_config::llm::provider;
 use jp_conversation::{message::ToolCallRequest, model::ProviderId, Model};
 use jp_query::query::{ChatQuery, StructuredQuery};
@@ -40,13 +41,21 @@ pub struct ModelDetails {
     /// The maximum output tokens, if known.
     pub max_output_tokens: Option<u32>,
 
-    /// Whether the model supports reasoning, if known.
-    pub reasoning: Option<bool>,
+    /// Whether the model supports reasoning, if unknown, it is assumed to not
+    /// be supported.
+    pub reasoning: Option<ReasoningDetails>,
 
     /// The knowledge cutoff date, if known.
     pub knowledge_cutoff: Option<Date>,
 }
 
+/// Details about the reasoning capabilities of a model.
+#[derive(Debug, Clone, Copy, Default, PartialEq)]
+pub struct ReasoningDetails {
+    pub min_tokens: u32,
+    pub max_tokens: Option<u32>,
+}
+
 /// Represents an event yielded by the chat completion stream.
 #[derive(Debug, Clone)]
 pub enum StreamEvent {
@@ -98,6 +107,19 @@ impl Event {
     }
 }
 
+impl From<Event> for StreamEvent {
+    fn from(event: Event) -> Self {
+        match event {
+            Event::Content(content) => StreamEvent::ChatChunk(CompletionChunk::Content(content)),
+            Event::Reasoning(reasoning) => {
+                StreamEvent::ChatChunk(CompletionChunk::Reasoning(reasoning))
+            }
+            Event::ToolCall(call) => StreamEvent::ToolCall(call),
+            Event::Metadata(key, metadata) => StreamEvent::Metadata(key, metadata),
+        }
+    }
+}
+
 impl From<Delta> for Option<Result<Event>> {
     fn from(delta: Delta) -> Self {
         if let Some(content) = delta.content {
@@ -248,14 +270,13 @@ pub trait Provider: std::fmt::Debug + Send + Sync {
 
 pub fn get_provider(id: ProviderId, config: &provider::Config) -> Result<Box<dyn Provider>> {
     let provider: Box<dyn Provider> = match id {
-        // ProviderId::Deepseek => Box::new(Deepseek::try_from(&config.deepseek)?),
-        // ProviderId::Google => Box::new(Google::try_from(&config.google)?),
-        // ProviderId::Xai => Box::new(Xai::try_from(&config.xai)?),
-        ProviderId::Ollama => Box::new(Ollama::try_from(&config.ollama)?),
         ProviderId::Anthropic => Box::new(Anthropic::try_from(&config.anthropic)?),
+        ProviderId::Deepseek => todo!(),
+        ProviderId::Google => Box::new(Google::try_from(&config.google)?),
+        ProviderId::Ollama => Box::new(Ollama::try_from(&config.ollama)?),
         ProviderId::Openai => Box::new(Openai::try_from(&config.openai)?),
         ProviderId::Openrouter => Box::new(Openrouter::try_from(&config.openrouter)?),
-        _ => todo!(),
+        ProviderId::Xai => todo!(),
     };
 
     Ok(provider)

crates/jp_llm/src/provider.rs

@@ -19,7 +19,7 @@ use serde_json::Value;
 use time::macros::date;
 use tracing::{trace, warn};
 
-use super::{Event, EventStream, ModelDetails, Provider, StreamEvent};
+use super::{Event, EventStream, ModelDetails, Provider, ReasoningDetails, StreamEvent};
 use crate::{
     error::{Error, Result},
     provider::{handle_delta, AccumulationState, Delta},
@@ -163,31 +163,31 @@ fn map_model(model: types::Model) -> ModelDetails {
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(32_000),
-            reasoning: Some(true),
+            reasoning: Some(ReasoningDetails::default()),
             knowledge_cutoff: Some(date!(2025 - 3 - 1)),
         },
         "claude-sonnet-4-0" | "claude-sonnet-4-20250514" => ModelDetails {
             provider: ProviderId::Anthropic,
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(64_000),
-            reasoning: Some(true),
+            reasoning: Some(ReasoningDetails::default()),
             knowledge_cutoff: Some(date!(2025 - 3 - 1)),
         },
         "claude-3-7-sonnet-latest" | "claude-3-7-sonnet-20250219" => ModelDetails {
             provider: ProviderId::Anthropic,
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(64_000),
-            reasoning: Some(true),
+            reasoning: Some(ReasoningDetails::default()),
             knowledge_cutoff: Some(date!(2024 - 11 - 1)),
         },
         "claude-3-5-haiku-latest" | "claude-3-5-haiku-20241022" => ModelDetails {
             provider: ProviderId::Anthropic,
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(8_192),
-            reasoning: Some(false),
+            reasoning: None,
             knowledge_cutoff: Some(date!(2024 - 7 - 1)),
         },
         "claude-3-5-sonnet-latest"
@@ -197,23 +197,23 @@ fn map_model(model: types::Model) -> ModelDetails {
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(8_192),
-            reasoning: Some(false),
+            reasoning: None,
             knowledge_cutoff: Some(date!(2024 - 4 - 1)),
         },
         "claude-3-opus-latest" | "claude-3-opus-20240229" => ModelDetails {
             provider: ProviderId::Anthropic,
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(4_096),
-            reasoning: Some(false),
+            reasoning: None,
             knowledge_cutoff: Some(date!(2023 - 8 - 1)),
         },
         "claude-3-haiku-20240307" => ModelDetails {
             provider: ProviderId::Anthropic,
             slug: model.id,
             context_window: Some(200_000),
             max_output_tokens: Some(4_096),
-            reasoning: Some(false),
+            reasoning: None,
             knowledge_cutoff: Some(date!(2024 - 8 - 1)),
         },
         id => {