diff --git a/crates/forge_repo/src/provider/anthropic.rs b/crates/forge_repo/src/provider/anthropic.rs
index 72fb10d46c..896c378d49 100644
--- a/crates/forge_repo/src/provider/anthropic.rs
+++ b/crates/forge_repo/src/provider/anthropic.rs
@@ -87,6 +87,10 @@ impl<H: HttpInfra> Anthropic<H> {
             headers.push(("anthropic-beta".to_string(), betas.join(",")));
         }
 
+        if let Some(custom_headers) = &self.provider.custom_headers {
+            headers.extend(custom_headers.iter().map(|(k, v)| (k.clone(), v.clone())));
+        }
+
         headers
     }
 }
diff --git a/crates/forge_repo/src/provider/google.rs b/crates/forge_repo/src/provider/google.rs
index 88e817116f..51aaf6275a 100644
--- a/crates/forge_repo/src/provider/google.rs
+++ b/crates/forge_repo/src/provider/google.rs
@@ -20,6 +20,7 @@ struct Google<T> {
     chat_url: Url,
     models: forge_domain::ModelSource<Url>,
     use_api_key_header: bool,
+    custom_headers: Vec<(String, String)>,
 }
 
 impl<H: HttpInfra> Google<H> {
@@ -30,7 +31,14 @@ impl<H: HttpInfra> Google<H> {
         models: forge_domain::ModelSource<Url>,
         use_api_key_header: bool,
     ) -> Self {
-        Self { http, api_key, chat_url, models, use_api_key_header }
+        Self {
+            http,
+            api_key,
+            chat_url,
+            models,
+            use_api_key_header,
+            custom_headers: Vec::new(),
+        }
     }
 
     fn get_headers(&self) -> Vec<(String, String)> {
@@ -45,6 +53,7 @@ impl<H: HttpInfra> Google<H> {
             ));
         }
 
+        headers.extend(self.custom_headers.iter().cloned());
         headers
     }
 }
@@ -179,13 +188,20 @@ impl<F: HttpInfra> GoogleResponseRepository<F> {
             }
         };
 
-        Ok(Google::new(
+        let mut client = Google::new(
             self.infra.clone(),
             token,
             chat_url,
             models,
             use_api_key_header,
-        ))
+        );
+        if let Some(custom_headers) = &provider.custom_headers {
+            client.custom_headers = custom_headers
+                .iter()
+                .map(|(k, v)| (k.clone(), v.clone()))
+                .collect();
+        }
+        Ok(client)
     }
 }
 #[async_trait::async_trait]
diff --git a/crates/forge_repo/src/provider/openai_responses/repository.rs b/crates/forge_repo/src/provider/openai_responses/repository.rs
index 2cd34cbed4..d09bab99a8 100644
--- a/crates/forge_repo/src/provider/openai_responses/repository.rs
+++ b/crates/forge_repo/src/provider/openai_responses/repository.rs
@@ -45,6 +45,7 @@ impl<H: HttpInfra> OpenAIResponsesProvider<H> {
 
         if provider.id == ProviderId::CODEX
             || provider.id == ProviderId::OPENCODE_ZEN
+            || provider.id == ProviderId::OPENCODE_GO
             || provider.id == ProviderId::OPENAI_RESPONSES_COMPATIBLE
         {
             // These providers already configure a complete Responses endpoint,
@@ -123,6 +124,10 @@ impl<H: HttpInfra> OpenAIResponsesProvider<H> {
                 forge_domain::AuthMethod::AwsProfile => {}
             });
 
+        if let Some(custom_headers) = &self.provider.custom_headers {
+            headers.extend(custom_headers.iter().map(|(k, v)| (k.clone(), v.clone())));
+        }
+
         // Codex provider requires the ChatGPT-Account-Id header extracted
         // from the JWT at login.
         //
diff --git a/crates/forge_repo/src/provider/opencode.rs b/crates/forge_repo/src/provider/opencode.rs
index dc534f4062..03445e9700 100644
--- a/crates/forge_repo/src/provider/opencode.rs
+++ b/crates/forge_repo/src/provider/opencode.rs
@@ -6,7 +6,7 @@ use forge_app::domain::{
     ResultStream,
 };
 use forge_app::{EnvironmentInfra, HttpInfra};
-use forge_domain::ChatRepository;
+use forge_domain::{ChatRepository, ConversationId};
 use url::Url;
 
 use crate::provider::anthropic::AnthropicResponseRepository;
@@ -61,9 +61,23 @@ impl<F: HttpInfra + EnvironmentInfra<Config = forge_config::ForgeConfig> + Sync>
     /// Derives the endpoint URL from the provider's configured base URL so that
     /// both OpenCode Zen and OpenCode Go (and any future variants) are routed
     /// to their correct endpoints.
-    fn build_provider(&self, provider: &Provider<Url>, model_id: &ModelId) -> Provider<Url> {
+    fn build_provider(
+        &self,
+        provider: &Provider<Url>,
+        model_id: &ModelId,
+        conversation_id: ConversationId,
+    ) -> Provider<Url> {
         let backend = self.get_backend(model_id);
         let mut new_provider = provider.clone();
+        // This clone belongs to one chat request, never to the shared provider
+        // config. Keep session metadata outside the body
+        // transformations used by each adapter.
+        let headers = new_provider.custom_headers.get_or_insert_default();
+        headers.retain(|name, _| !name.eq_ignore_ascii_case("x-opencode-session"));
+        headers.insert(
+            "x-opencode-session".to_string(),
+            conversation_id.to_string(),
+        );
         let base = provider.url.as_str().trim_end_matches('/');
 
         match backend {
@@ -99,7 +113,12 @@ impl<F: HttpInfra + EnvironmentInfra<Config = forge_config::ForgeConfig> + Sync>
         provider: Provider<Url>,
     ) -> ResultStream<ChatCompletionMessage, anyhow::Error> {
         let backend = self.get_backend(model_id);
-        let adapted_provider = self.build_provider(&provider, model_id);
+        // Standalone requests without a conversation still need a session, but
+        // must not share a provider-wide ID with unrelated requests.
+        let conversation_id = context
+            .conversation_id
+            .unwrap_or_else(ConversationId::generate);
+        let adapted_provider = self.build_provider(&provider, model_id, conversation_id);
 
         match backend {
             OpenCodeBackend::Anthropic => {
@@ -153,10 +172,242 @@ enum OpenCodeBackend {
 
 #[cfg(test)]
 mod tests {
+    use std::collections::{BTreeMap, HashMap};
+    use std::sync::Mutex;
+
+    use bytes::Bytes;
+    use forge_domain::{AuthCredential, AuthDetails, ContextMessage, ProviderId};
+    use forge_eventsource::EventSource;
     use pretty_assertions::assert_eq;
+    use reqwest::header::HeaderMap;
 
     use super::*;
 
+    #[derive(Default)]
+    struct RecordingInfra {
+        requests: Mutex<Vec<(Url, HeaderMap, Bytes)>>,
+    }
+
+    impl EnvironmentInfra for RecordingInfra {
+        type Config = forge_config::ForgeConfig;
+
+        fn get_config(&self) -> anyhow::Result<Self::Config> {
+            Ok(Self::Config::default())
+        }
+
+        fn get_environment(&self) -> forge_domain::Environment {
+            use fake::{Fake, Faker};
+            Faker.fake()
+        }
+
+        fn get_env_var(&self, _: &str) -> Option<String> {
+            None
+        }
+
+        fn get_env_vars(&self) -> BTreeMap<String, String> {
+            BTreeMap::new()
+        }
+
+        async fn update_environment(&self, _: Vec<forge_domain::ConfigOperation>) -> Result<()> {
+            Ok(())
+        }
+    }
+
+    #[async_trait::async_trait]
+    impl HttpInfra for RecordingInfra {
+        async fn http_delete(&self, _: &Url) -> Result<reqwest::Response> {
+            anyhow::bail!("Unexpected DELETE")
+        }
+
+        async fn http_get(&self, _: &Url, _: Option<HeaderMap>) -> Result<reqwest::Response> {
+            anyhow::bail!("Unexpected GET")
+        }
+
+        async fn http_post(
+            &self,
+            url: &Url,
+            headers: Option<HeaderMap>,
+            body: Bytes,
+        ) -> Result<reqwest::Response> {
+            self.requests
+                .lock()
+                .unwrap()
+                .push((url.clone(), headers.unwrap(), body));
+            anyhow::bail!("Recorded request")
+        }
+
+        async fn http_eventsource(
+            &self,
+            url: &Url,
+            headers: Option<HeaderMap>,
+            body: Bytes,
+        ) -> Result<EventSource> {
+            self.requests
+                .lock()
+                .unwrap()
+                .push((url.clone(), headers.unwrap(), body));
+            anyhow::bail!("Recorded request")
+        }
+    }
+
+    fn fixture_provider(id: ProviderId) -> anyhow::Result<Provider<Url>> {
+        Ok(Provider {
+            id: id.clone(),
+            provider_type: Default::default(),
+            response: Some(ProviderResponse::OpenCode),
+            url: Url::parse("https://opencode.ai/zen/go")?,
+            models: Some(forge_domain::ModelSource::Hardcoded(vec![])),
+            auth_methods: vec![forge_domain::AuthMethod::ApiKey],
+            url_params: vec![],
+            credential: Some(AuthCredential {
+                id,
+                auth_details: AuthDetails::ApiKey("fixture-key".to_string().into()),
+                url_params: HashMap::new(),
+            }),
+            custom_headers: Some(HashMap::from([
+                (
+                    "X-OpenCode-Session".to_string(),
+                    "stale-static-session".to_string(),
+                ),
+                ("x-custom".to_string(), "preserved".to_string()),
+            ])),
+        })
+    }
+
+    #[tokio::test]
+    async fn test_session_reaches_every_opencode_adapter() {
+        for provider_id in [ProviderId::OPENCODE_GO, ProviderId::OPENCODE_ZEN] {
+            let fixture = fixture_provider(provider_id).unwrap();
+            let original = fixture.clone();
+            let infra = Arc::new(RecordingInfra::default());
+            let repo = OpenCodeZenResponseRepository::new(infra.clone());
+            let conversation_id = ConversationId::generate();
+            let other_id = ConversationId::generate();
+            let context = ChatContext::default()
+                .conversation_id(conversation_id)
+                .add_message(ContextMessage::user("hello", None));
+            let models = [
+                ("deepseek-v4-flash", "/zen/go/v1/chat/completions"),
+                ("gpt-5", "/zen/go/v1/responses"),
+                ("claude-sonnet-4-5", "/zen/go/v1/messages"),
+                (
+                    "gemini-3-flash",
+                    "/zen/go/v1/models/gemini-3-flash:streamGenerateContent",
+                ),
+            ];
+
+            // Repeated turns/retries and resumed contexts use the same ID; an
+            // interleaved conversation must not overwrite that session.
+            for id in [conversation_id, other_id, conversation_id] {
+                for (model, _) in models {
+                    let result = repo
+                        .chat(
+                            &ModelId::new(model),
+                            context.clone().conversation_id(id),
+                            fixture.clone(),
+                        )
+                        .await;
+                    assert!(result.is_err());
+                }
+            }
+            let requests = infra.requests.lock().unwrap();
+            let actual = requests
+                .iter()
+                .map(|(url, headers, body)| {
+                    let json: serde_json::Value = serde_json::from_slice(body).unwrap();
+                    assert!(json.get("session_id").is_none());
+                    (
+                        url.path().to_string(),
+                        headers
+                            .get("x-opencode-session")
+                            .map(|value| value.to_str().unwrap().to_string()),
+                        headers
+                            .get("x-custom")
+                            .map(|value| value.to_str().unwrap().to_string()),
+                        headers.get_all("x-opencode-session").iter().count(),
+                    )
+                })
+                .collect::<Vec<_>>();
+            let expected = [conversation_id, other_id, conversation_id]
+                .into_iter()
+                .flat_map(|id| {
+                    models.map(|(_, path)| {
+                        (
+                            path.to_string(),
+                            Some(id.to_string()),
+                            Some("preserved".to_string()),
+                            1,
+                        )
+                    })
+                })
+                .collect::<Vec<_>>();
+            assert_eq!(actual, expected);
+            assert_eq!(fixture, original);
+        }
+    }
+
+    #[tokio::test]
+    async fn test_standalone_opencode_requests_have_independent_sessions() {
+        let fixture = fixture_provider(ProviderId::OPENCODE_GO).unwrap();
+        let infra = Arc::new(RecordingInfra::default());
+        let repo = OpenCodeZenResponseRepository::new(infra.clone());
+        for _ in 0..2 {
+            let result = repo
+                .chat(
+                    &ModelId::new("deepseek-v4-flash"),
+                    ChatContext::default(),
+                    fixture.clone(),
+                )
+                .await;
+            assert!(result.is_err());
+        }
+        let requests = infra.requests.lock().unwrap();
+        let actual = requests
+            .iter()
+            .map(|(_, headers, _)| {
+                ConversationId::parse(headers["x-opencode-session"].to_str().unwrap()).unwrap()
+            })
+            .collect::<std::collections::HashSet<_>>()
+            .len();
+        let expected = 2;
+        assert_eq!(actual, expected);
+    }
+
+    #[tokio::test]
+    async fn test_other_providers_do_not_receive_opencode_session() {
+        let mut fixture = fixture_provider(ProviderId::OPENAI).unwrap();
+        fixture.custom_headers = None;
+        let infra = Arc::new(RecordingInfra::default());
+        let context = ChatContext::default().conversation_id(ConversationId::generate());
+        let model = ModelId::new("fixture-model");
+
+        let result = OpenAIResponseRepository::new(infra.clone())
+            .chat(&model, context.clone(), fixture.clone())
+            .await;
+        assert!(result.is_err());
+        let result = OpenAIResponsesResponseRepository::new(infra.clone())
+            .chat(&model, context.clone(), fixture.clone())
+            .await;
+        assert!(result.is_err());
+        let result = AnthropicResponseRepository::new(infra.clone())
+            .chat(&model, context.clone(), fixture.clone())
+            .await;
+        assert!(result.is_err());
+        let result = GoogleResponseRepository::new(infra.clone())
+            .chat(&model, context, fixture)
+            .await;
+        assert!(result.is_err());
+        let actual = infra
+            .requests
+            .lock()
+            .unwrap()
+            .iter()
+            .map(|(_, headers, _)| headers.contains_key("x-opencode-session"))
+            .collect::<Vec<_>>();
+        let expected = vec![false, false, false, false];
+        assert_eq!(actual, expected);
+    }
+
     /// Helper function to determine backend routing (mirrors get_backend logic)
     fn get_backend_for_test(model_id: &str) -> OpenCodeBackend {
         if model_id.starts_with("claude-") {