fix(provider): add Gemini 3.8, 3.7, 3.6 Flash and 3.5 Flash-Lite models

crates/forge_api/src/forge_api.rs

@@ -241,8 +241,8 @@ impl<
     }
 
     async fn update_config(&self, ops: Vec<forge_domain::ConfigOperation>) -> anyhow::Result<()> {
-        // Determine whether any op affects provider/model resolution before writing,
-        // so we can invalidate the agent cache afterwards.
+        // Determine whether any op affects provider/model resolution before
+        // writing, so we can invalidate the agent cache afterwards.
         let needs_agent_reload = ops
             .iter()
             .any(|op| matches!(op, forge_domain::ConfigOperation::SetSessionConfig(_)));

crates/forge_app/src/agent.rs

@@ -130,8 +130,9 @@ impl AgentExt for Agent {
 
         // Apply workflow compact configuration to agents
         if let Some(ref workflow_compact) = config.compact {
-            // Convert forge_config::Compact to forge_domain::Compact, then merge.
-            // Agent settings take priority over workflow settings.
+            // Convert forge_config::Compact to forge_domain::Compact, then
+            // merge. Agent settings take priority over workflow
+            // settings.
             let mut merged_compact = Compact {
                 retention_window: workflow_compact.retention_window,
                 eviction_window: workflow_compact.eviction_window.value(),
@@ -169,7 +170,8 @@ impl AgentExt for Agent {
                 exclude: config_reasoning.exclude,
                 enabled: config_reasoning.enabled,
             };
-            // Start from the agent's own settings and fill unset fields from config.
+            // Start from the agent's own settings and fill unset fields from
+            // config.
             let mut merged = agent.reasoning.clone().unwrap_or_default();
             merged.merge(config_as_domain);
             // If the config explicitly disables reasoning, honour that override
@@ -302,8 +304,8 @@ mod tests {
 
         // CURRENT BEHAVIOR: Due to merge order (workflow_compact merged with
         // agent.compact), agent's retention_window=0 overwrites workflow's 10
-        // This is the documented behavior: "Agent settings take priority over workflow
-        // settings"
+        // This is the documented behavior: "Agent settings take priority over
+        // workflow settings"
 
         // Agent default has retention_window=0, which overwrites workflow's 10
         assert_eq!(

crates/forge_app/src/agent_executor.rs

@@ -63,8 +63,9 @@ impl<S: Services + EnvironmentInfra<Config = forge_config::ForgeConfig>> AgentEx
                 .await?
                 .ok_or(Error::ConversationNotFound { id: conversation_id })?
         } else {
-            // Create context with agent initiator since it's spawned by a parent agent
-            // This is crucial for GitHub Copilot billing optimization
+            // Create context with agent initiator since it's spawned by a
+            // parent agent This is crucial for GitHub Copilot
+            // billing optimization
             let context = forge_domain::Context::default().initiator("agent".to_string());
             let conversation = Conversation::generate()
                 .title(task.clone())

crates/forge_app/src/agent_provider_resolver.rs

@@ -28,7 +28,8 @@ where
             // Load all agent definitions and find the one we need
 
             if let Some(agent) = self.0.get_agent(&agent_id).await? {
-                // If the agent definition has a provider, use it; otherwise use default
+                // If the agent definition has a provider, use it; otherwise use
+                // default
                 agent.provider
             } else {
                 // TODO: Needs review, should we throw an err here?

crates/forge_app/src/app.rs

@@ -148,8 +148,8 @@ impl<S: Services + EnvironmentInfra<Config = forge_config::ForgeConfig>> ForgeAp
         let tracing_handler = TracingHandler::new();
         let title_handler = TitleGenerationHandler::new(services.clone());
 
-        // Build the on_end hook, conditionally adding PendingTodosHandler based on
-        // config
+        // Build the on_end hook, conditionally adding PendingTodosHandler based
+        // on config
         let on_end_hook = if forge_config.verify_todos {
             tracing_handler
                 .clone()
@@ -194,7 +194,8 @@ impl<S: Services + EnvironmentInfra<Config = forge_config::ForgeConfig>> ForgeAp
                     let conversation = orch.get_conversation().clone();
                     let save_result = services.upsert_conversation(conversation).await;
 
-                    // Send any error to the stream (prioritize dispatch error over save error)
+                    // Send any error to the stream (prioritize dispatch error
+                    // over save error)
                     #[allow(clippy::collapsible_if)]
                     if let Some(err) = dispatch_result.err().or(save_result.err()) {
                         if let Err(e) = tx.send(Err(err)).await {
@@ -306,7 +307,8 @@ impl<S: Services + EnvironmentInfra<Config = forge_config::ForgeConfig>> ForgeAp
     pub async fn get_all_provider_models(&self) -> Result<Vec<ProviderModels>> {
         let all_providers = self.services.get_all_providers().await?;
 
-        // Build one future per configured provider, preserving the error on failure.
+        // Build one future per configured provider, preserving the error on
+        // failure.
         let futures: Vec<_> = all_providers
             .into_iter()
             .filter_map(|any_provider| any_provider.into_configured())

crates/forge_app/src/command_generator.rs

@@ -229,7 +229,8 @@ mod tests {
                 .map(|(path, is_dir)| File { path: path.clone(), is_dir: *is_dir })
                 .collect();
 
-            // Sort: directories first (alphabetically), then files (alphabetically)
+            // Sort: directories first (alphabetically), then files
+            // (alphabetically)
             files.sort_by(|a, b| match (a.is_dir, b.is_dir) {
                 (true, false) => std::cmp::Ordering::Less,
                 (false, true) => std::cmp::Ordering::Greater,

crates/forge_app/src/compact.rs

@@ -113,10 +113,10 @@ impl Compactor {
         // chains. After compaction, this consistency can break if the first
         // remaining assistant lacks reasoning.
         //
-        // Solution: Extract the LAST reasoning from compacted messages and inject it
-        // into the first assistant message after compaction. This preserves
-        // chain continuity while preventing exponential accumulation across
-        // multiple compactions.
+        // Solution: Extract the LAST reasoning from compacted messages and
+        // inject it into the first assistant message after compaction.
+        // This preserves chain continuity while preventing exponential
+        // accumulation across multiple compactions.
         //
         // Example: [U, A+r, U, A+r, U, A] → compact → [U-summary, A+r, U, A]
         //                                                          └─from last
@@ -133,8 +133,8 @@ impl Compactor {
                 _ => None,
             });
 
-        // Accumulate usage from all messages in the compaction range before they are
-        // destroyed
+        // Accumulate usage from all messages in the compaction range before
+        // they are destroyed
         let compacted_usage = context.messages.get(start..=end).and_then(|slice| {
             slice
                 .iter()
@@ -143,7 +143,8 @@ impl Compactor {
                 .reduce(|a, b| a.accumulate(&b))
         });
 
-        // Replace the range with the summary, transferring the accumulated usage
+        // Replace the range with the summary, transferring the accumulated
+        // usage
         let mut summary_entry = MessageEntry::from(ContextMessage::user(summary, None));
         summary_entry.usage = compacted_usage;
         context
@@ -289,7 +290,8 @@ mod tests {
 
         let context = compactor.compress_single_sequence(context, (0, 2)).unwrap();
 
-        // Verify reasoning didn't accumulate - should still be just 1 reasoning block
+        // Verify reasoning didn't accumulate - should still be just 1 reasoning
+        // block
         let first_assistant = context
             .messages
             .iter()
@@ -318,7 +320,8 @@ mod tests {
             ..Default::default()
         }];
 
-        // Most recent message in range has empty reasoning, earlier has non-empty
+        // Most recent message in range has empty reasoning, earlier has
+        // non-empty
         let context = Context::default()
             .add_message(ContextMessage::user("M1", None))
             .add_message(ContextMessage::assistant(
@@ -647,8 +650,9 @@ mod tests {
             "Summary message should carry accumulated usage from compacted messages"
         );
 
-        // accumulate_usage() must sum both the compacted range usage (on the summary
-        // message) and the surviving outside_usage — total = inside + inside2 + outside
+        // accumulate_usage() must sum both the compacted range usage (on the
+        // summary message) and the surviving outside_usage — total =
+        // inside + inside2 + outside
         let expected_total_usage = Usage {
             total_tokens: TokenCount::Actual(100000),
             prompt_tokens: TokenCount::Actual(90000),
@@ -855,17 +859,19 @@ mod tests {
         // - Safe threshold (89.6K): ~95K tokens, SHOULD compact (true)
         //
         // At turn 2:
-        // - Unsafe threshold (100K): ~110K tokens, SHOULD compact (true) - but too
-        //   late!
+        // - Unsafe threshold (100K): ~110K tokens, SHOULD compact (true) - but
+        //   too late!
         // - Safe threshold (89.6K): ~110K tokens, already compacted at turn 1
 
-        // Verify that safe threshold triggers at turn 1 (providing early warning)
+        // Verify that safe threshold triggers at turn 1 (providing early
+        // warning)
         let safe_token_count_turn1 = 95_000; // Approximate
         let safe_should_compact_turn1 =
             safe_compact.should_compact(&safe_context, safe_token_count_turn1);
 
-        // The key fix: safe threshold (89.6K) triggers at ~95K, while unsafe (100K)
-        // doesn't This provides a safety margin before we hit the 128K limit
+        // The key fix: safe threshold (89.6K) triggers at ~95K, while unsafe
+        // (100K) doesn't This provides a safety margin before we hit
+        // the 128K limit
         assert!(
             safe_should_compact_turn1 || safe_token_count_turn1 < 89_600,
             "Safe threshold (89.6K) should trigger compaction at ~95K tokens to provide safety margin"
@@ -876,8 +882,8 @@ mod tests {
         let final_unsafe = unsafe_context.token_count_approx();
         let final_safe = safe_context.token_count_approx();
 
-        // Both should be identical since we're just testing threshold logic, not actual
-        // compaction
+        // Both should be identical since we're just testing threshold logic,
+        // not actual compaction
         assert_eq!(
             final_unsafe, final_safe,
             "Both contexts should have same token count"

crates/forge_app/src/dto/anthropic/request.rs

@@ -129,8 +129,9 @@ impl TryFrom<forge_domain::Context> for Request {
         // a positive effort / `max_tokens` still emit reasoning on the wire.
         let reasoning_on = request.is_reasoning_supported();
         let (thinking, output_config) = if reasoning_on && let Some(reasoning) = request.reasoning {
-            // Adaptive thinking on 4.7 hides reasoning content by default; opting
-            // into reasoning should surface it unless the caller set `exclude`.
+            // Adaptive thinking on 4.7 hides reasoning content by default;
+            // opting into reasoning should surface it unless the
+            // caller set `exclude`.
             let adaptive_display = if reasoning.exclude == Some(true) {
                 Some(ThinkingDisplay::Omitted)
             } else {
@@ -185,7 +186,8 @@ impl TryFrom<forge_domain::Context> for Request {
             output_config,
             output_format: request.response_format.and_then(|rf| match rf {
                 forge_domain::ResponseFormat::Text => {
-                    // Anthropic doesn't have a "text" output format, so we skip it
+                    // Anthropic doesn't have a "text" output format, so we skip
+                    // it
                     None
                 }
                 forge_domain::ResponseFormat::JsonSchema(schema) => {
@@ -260,7 +262,8 @@ impl TryFrom<ContextMessage> for Message {
                     forge_domain::Role::User => Message { role: Role::User, content },
                     forge_domain::Role::Assistant => Message { role: Role::Assistant, content },
                     forge_domain::Role::System => {
-                        // note: Anthropic doesn't support system role messages and they're already
+                        // note: Anthropic doesn't support system role messages
+                        // and they're already
                         // filtered out. so this state is unreachable.
                         return Err(
                             forge_domain::Error::UnsupportedRole("System".to_string()).into()
@@ -285,7 +288,8 @@ impl Message {
             *content = std::mem::take(content).cached(false);
         }
 
-        // If enabling cache, set cache control on the last cacheable content item
+        // If enabling cache, set cache control on the last cacheable content
+        // item
         if enable_cache
             && let Some(last_cacheable_idx) =
                 self.content
@@ -592,7 +596,8 @@ mod tests {
 
     #[test]
     fn test_reasoning_max_tokens_and_effort_emit_both() {
-        // Effort and budget are independent knobs — neither should hide the other.
+        // Effort and budget are independent knobs — neither should hide the
+        // other.
         let fixture = Context::default().reasoning(ReasoningConfig {
             effort: Some(forge_domain::Effort::Low),
             enabled: Some(true),
@@ -735,9 +740,9 @@ mod tests {
 
     #[test]
     fn test_reasoning_enabled_none_with_max_tokens_still_emits_thinking() {
-        // Matches the domain's `is_reasoning_supported` rule: enabled: None with a
-        // positive budget counts as on, so inherited/merged configs don't silently
-        // disable reasoning on the wire.
+        // Matches the domain's `is_reasoning_supported` rule: enabled: None
+        // with a positive budget counts as on, so inherited/merged
+        // configs don't silently disable reasoning on the wire.
         let fixture = Context::default().reasoning(ReasoningConfig {
             enabled: None,
             max_tokens: Some(8000),

crates/forge_app/src/dto/anthropic/response.rs

@@ -164,9 +164,11 @@ impl From<Usage> for forge_domain::Usage {
     fn from(usage: Usage) -> Self {
         // Anthropic token breakdown:
         // - input_tokens: tokens NOT from cache (billed at full price)
-        // - cache_creation_input_tokens: tokens written to cache (billed at full price
+        // - cache_creation_input_tokens: tokens written to cache (billed at
+        //   full price
         //   + write cost)
-        // - cache_read_input_tokens: tokens read from cache (billed at 90% discount)
+        // - cache_read_input_tokens: tokens read from cache (billed at 90%
+        //   discount)
         // Total input = input_tokens + cache_creation_input_tokens +
         // cache_read_input_tokens
 
@@ -320,7 +322,8 @@ impl TryFrom<Event> for ChatCompletionMessage {
                 ChatCompletionMessage::try_from(content_block)?
             }
             Event::MessageStart { message } => {
-                // Extract usage from MessageStart - this contains input token counts
+                // Extract usage from MessageStart - this contains input token
+                // counts
                 ChatCompletionMessage::assistant(Content::part("")).usage(message.usage)
             }
             Event::MessageDelta { delta, usage } => {
@@ -332,7 +335,8 @@ impl TryFrom<Event> for ChatCompletionMessage {
                 return Err(error.into());
             }
             Event::Ping { cost: Some(cost) } => {
-                // OpenCode Zen sends cost in a ping event at the end of the stream
+                // OpenCode Zen sends cost in a ping event at the end of the
+                // stream
                 let cost_value = match cost {
                     StringOrF64::Number(n) => n,
                     StringOrF64::String(s) => s.parse().unwrap_or(0.0),
@@ -395,8 +399,9 @@ impl TryFrom<ContentBlock> for ChatCompletionMessage {
                 )
             }
             ContentBlock::ToolUse { id, name, input } => {
-                // note: We've to check if the input is empty or null. else we end up adding
-                // empty object `{}` as prefix to tool args.
+                // note: We've to check if the input is empty or null. else we
+                // end up adding empty object `{}` as prefix to
+                // tool args.
                 let is_empty =
                     input.is_null() || input.as_object().is_some_and(|map| map.is_empty());
                 ChatCompletionMessage::assistant(Content::part("")).add_tool_call(ToolCallPart {
@@ -565,8 +570,8 @@ mod tests {
         let expected_prompt = TokenCount::Actual(100 + 200 + 300);
         assert_eq!(actual.prompt_tokens, expected_prompt);
 
-        // cached_tokens should only include cache reads (tokens that benefited from
-        // caching)
+        // cached_tokens should only include cache reads (tokens that benefited
+        // from caching)
         let expected_cached = TokenCount::Actual(300);
         assert_eq!(actual.cached_tokens, expected_cached);
 

crates/forge_app/src/dto/anthropic/transforms/enforce_schema.rs

@@ -37,7 +37,8 @@ impl Transformer for EnforceStrictObjectSchema {
         if let Some(OutputFormat::JsonSchema { schema }) = request.output_format.take() {
             // Convert schema to JSON value for normalization
             if let Ok(mut schema_value) = serde_json::to_value(&schema) {
-                // Use non-strict mode (false) for Anthropic - only adds additionalProperties
+                // Use non-strict mode (false) for Anthropic - only adds
+                // additionalProperties
                 enforce_strict_schema(&mut schema_value, false);
 
                 // Convert back to RootSchema

crates/forge_app/src/dto/anthropic/transforms/mcp_tool_names.rs

@@ -68,15 +68,17 @@ mod tests {
     #[test]
     fn test_handles_server_name_containing_tool_substring() {
         // Server named "my_tool_server", tool named "action"
-        // rfind gives the LAST `_tool_`, so server = "my_tool_server", tool = "action"
+        // rfind gives the LAST `_tool_`, so server = "my_tool_server", tool =
+        // "action"
         let actual = to_claude_code_format("mcp_my_tool_server_tool_action");
         let expected = "mcp__my_tool_server__action";
         assert_eq!(actual, expected);
     }
 
     #[test]
     fn test_leaves_already_converted_names_unchanged() {
-        // mcp__ prefix means it was already converted; no `_tool_` in sanitized names
+        // mcp__ prefix means it was already converted; no `_tool_` in sanitized
+        // names
         let actual = to_claude_code_format("mcp__github__create_issue");
         let expected = "mcp__github__create_issue";
         assert_eq!(actual, expected);

crates/forge_app/src/dto/anthropic/transforms/set_cache.rs

@@ -70,7 +70,8 @@ mod tests {
     ) -> String {
         let mut messages = Vec::new();
 
-        // Add system messages to the regular messages array for Anthropic format
+        // Add system messages to the regular messages array for Anthropic
+        // format
         for c in system_messages.chars() {
             match c {
                 's' => messages.push(