lean_ctx/proxy/
openai.rs

1use axum::{
2    body::Body,
3    extract::State,
4    http::{Request, StatusCode},
5    response::Response,
6};
7use serde_json::Value;
8
9use super::compress::compress_tool_result;
10use super::forward;
11use super::tool_kind::{self, should_protect, ToolResultKind};
12use super::ProxyState;
13
14pub async fn handler(
15    State(state): State<ProxyState>,
16    req: Request<Body>,
17) -> Result<Response, StatusCode> {
18    let upstream = state.openai_upstream.clone();
19    forward::forward_request(
20        State(state),
21        req,
22        &upstream,
23        "/v1/chat/completions",
24        compress_request_body,
25        "OpenAI",
26        &[],
27    )
28    .await
29}
30
31fn compress_request_body(parsed: Value, original_size: usize) -> (Vec<u8>, usize, usize) {
32    let mut doc = parsed;
33    let mut modified = false;
34
35    if let Some(messages) = doc.get_mut("messages").and_then(|m| m.as_array_mut()) {
36        let tool_names = tool_kind::openai_tool_names(messages);
37
38        // OpenAI's automatic prompt caching is prefix-based like Anthropic's,
39        // so history is pruned at the same frozen, cache-aware boundary.
40        let mode = crate::core::config::Config::load()
41            .proxy
42            .resolved_history_mode();
43        let boundary = super::history_prune::prune_boundary(mode, messages.len());
44        modified |= super::history_prune::prune_history(messages, boundary, &tool_names);
45
46        for msg in messages.iter_mut() {
47            let role = msg.get("role").and_then(|r| r.as_str()).unwrap_or("");
48            if role != "tool" {
49                continue;
50            }
51
52            let name = msg
53                .get("tool_call_id")
54                .and_then(|v| v.as_str())
55                .and_then(|id| tool_names.get(id))
56                .map(String::as_str);
57            let kind = name.map_or(ToolResultKind::Other, tool_kind::classify_tool_name);
58
59            if let Some(content) = msg
60                .get_mut("content")
61                .and_then(|c| c.as_str().map(String::from))
62            {
63                if should_protect(kind, &content) {
64                    continue;
65                }
66                let compressed = compress_tool_result(&content, name);
67                if compressed.len() < content.len() {
68                    msg["content"] = Value::String(compressed);
69                    modified = true;
70                }
71            }
72        }
73    }
74
75    let out = serde_json::to_vec(&doc).unwrap_or_default();
76    let compressed_size = if modified { out.len() } else { original_size };
77    (out, original_size, compressed_size)
78}
79
80#[cfg(test)]
81mod tests {
82    use super::*;
83
84    #[test]
85    fn read_file_tool_result_protected() {
86        let code = (0..60)
87            .map(|i| format!("    const value{i} = computeValue{i}(ctx, opts);"))
88            .collect::<Vec<_>>()
89            .join("\n");
90        let body = serde_json::json!({
91            "model": "gpt-5",
92            "messages": [
93                {"role": "assistant", "tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "read_file"}}]},
94                {"role": "tool", "tool_call_id": "call_1", "content": code}
95            ]
96        });
97        let bytes = serde_json::to_vec(&body).unwrap();
98        let (out, _orig, _comp) = compress_request_body(body, bytes.len());
99        let parsed: Value = serde_json::from_slice(&out).unwrap();
100        assert!(parsed["messages"][1]["content"]
101            .as_str()
102            .unwrap()
103            .contains("value59"));
104    }
105}
lean_ctx/proxy/openai.rs

lean_ctx/proxy/
openai.rs
