GameCult/Ghostlight
Watch
9
0
Fork
You've already forked Ghostlight
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Queue connector work at admitted capacity

Browse source
This commit is contained in:
Meta Vens 2026-08-26 06:27:35 +02:00
commit e5345ede4b
2 changed files with 28 additions and 1 deletions
Download patch file Download diff file Expand all files Collapse all files

6
crates/ghostlight-dungeon/src/main.rs
View file

@ -316,6 +316,11 @@ async fn main() -> anyhow::Result<()> {
std::env::var("GHOSTLIGHT_MODEL_FAST").unwrap_or_else(|_| default_fast_model.into());
let capable_model =
std::env::var("GHOSTLIGHT_MODEL_CAPABLE").unwrap_or_else(|_| default_capable_model.into());
let connector_max_concurrent_requests =
std::env::var("GHOSTLIGHT_CODEX_CONNECTOR_MAX_CONCURRENT_REQUESTS")
.unwrap_or_else(|_| "8".to_string())
.parse::<usize>()
.context("GHOSTLIGHT_CODEX_CONNECTOR_MAX_CONCURRENT_REQUESTS must be an integer")?;
let secret_path = std::env::var_os("GHOSTLIGHT_MODEL_CREDENTIAL")
.map(PathBuf::from)
.unwrap_or_else(|| runtime_root.join("secrets").join(default_secret_name));
@ -336,6 +341,7 @@ async fn main() -> anyhow::Result<()> {
&secret_path,
fast_model.clone(),
capable_model.clone(),
connector_max_concurrent_requests,
)?),
"codex-connector" => Arc::new(CodexConnectorModelPort::from_runtime_secret(
std::env::var("GHOSTLIGHT_MODEL_CONNECTOR")

23
crates/ghostlight-dungeon/src/model_connector.rs
View file

@ -11,6 +11,7 @@ use codex_connector::{
CodexTransportOutcome,
};
use sha2::{Digest, Sha256};
use tokio::sync::Semaphore;
use zeroize::Zeroizing;
use crate::model::{
@ -23,6 +24,7 @@ const REQUEST_EXPIRY: Duration = Duration::from_secs(150);
#[derive(Clone)]
pub struct CodexConnectorModelPort {
client: CodexConnectorClient,
request_gate: std::sync::Arc<Semaphore>,
caller_runtime_id: String,
fast_model: String,
capable_model: String,
@ -35,14 +37,19 @@ impl CodexConnectorModelPort {
caller_runtime_id: impl Into<String>,
fast_model: impl Into<String>,
capable_model: impl Into<String>,
max_concurrent_requests: usize,
) -> Result<Self> {
let caller_runtime_id = caller_runtime_id.into();
if caller_runtime_id.trim().is_empty() || caller_runtime_id.trim() != caller_runtime_id {
bail!("CodexConnector caller runtime must be a non-empty exact identity")
}
if !(1..=128).contains(&max_concurrent_requests) {
bail!("CodexConnector caller concurrency must be between 1 and 128")
}
let client = CodexConnectorClient::new(endpoint, connection_key, MAX_FRAME_BYTES, None)?;
Ok(Self {
client,
request_gate: std::sync::Arc::new(Semaphore::new(max_concurrent_requests)),
caller_runtime_id,
fast_model: fast_model.into(),
capable_model: capable_model.into(),
@ -55,6 +62,7 @@ impl CodexConnectorModelPort {
caller_runtime_id: impl Into<String>,
fast_model: impl Into<String>,
capable_model: impl Into<String>,
max_concurrent_requests: usize,
) -> Result<Self> {
let bytes = Zeroizing::new(std::fs::read(path.as_ref())?);
let raw = std::str::from_utf8(bytes.as_slice())?;
@ -68,6 +76,7 @@ impl CodexConnectorModelPort {
caller_runtime_id,
fast_model,
capable_model,
max_concurrent_requests,
)
}
@ -125,9 +134,19 @@ impl ModelPort for CodexConnectorModelPort {
}
async fn run_observed(&self, request: &ModelStageRequest) -> Result<ModelProviderOutput> {
let permit = self
.request_gate
.clone()
.acquire_owned()
.await
.map_err(|_| anyhow::anyhow!("CodexConnector request gate closed"))?;
let port = self.clone();
let request = request.clone();
tokio::task::spawn_blocking(move || port.invoke(&request)).await?
tokio::task::spawn_blocking(move || {
let _permit = permit;
port.invoke(&request)
})
.await?
}
fn provider(&self) -> &'static str {
@ -588,7 +607,9 @@ mod tests {
"ghostlight-dungeon-yggdrasil",
"gpt-5.6-luna",
"gpt-5.6-luna",
1,
)?;
assert_eq!(port.request_gate.available_permits(), 1);
let output = port
.run_observed(&ModelStageRequest {
stage: "test_interpreter".to_string(),