mirror of
https://github.com/farion1231/cc-switch.git
synced 2026-10-26 01:34:51 +08:00
fix(proxy): normalize DeepSeek tool replay in Codex Chat requests (#8351)
When Codex replays tool-call history, empty assistant messages, progress text or reasoning can sit between a function_call and its results. Converted to Chat, these become standalone assistant messages and DeepSeek rejects the request with HTTP 400. For explicit DeepSeek model ids only, fold such messages into the open tool-call message after conversion (running after the malformed-call recovery from #8241); well-formed histories stay byte-identical. Refs #8331.
This commit is contained in:
funkpopo authored and 
GitHub committed
1 parent
f9bc9a4f2b
commit
2a7080f14e
1 file changed
+394
-2
@@ -277,6 +277,7 @@ pub fn responses_to_chat_completions_with_reasoning(
|
||||
) -> Result<Value, ProxyError> {
|
||||
let mut result = json!({});
|
||||
let tool_context = build_codex_tool_context_from_request(&body);
|
||||
let model = body.get("model").and_then(Value::as_str).unwrap_or("");
|
||||
|
||||
if let Some(model) = body.get("model") {
|
||||
result["model"] = model.clone();
|
||||
@@ -302,10 +303,16 @@ pub fn responses_to_chat_completions_with_reasoning(
|
||||
// Counts only: historical arguments and tool results may contain secrets.
|
||||
log::warn!("[Codex Chat] Preserved {recovered} malformed historical tool call(s) as text");
|
||||
}
|
||||
// DeepSeek strictly validates tool-result adjacency. Limit this replay
|
||||
// repair to explicit DeepSeek model IDs: other providers may interpret
|
||||
// assistant turns differently, and rewriting their history changes caches.
|
||||
if is_deepseek_chat_model(model) {
|
||||
coalesce_replayed_tool_turn_messages(&mut messages);
|
||||
}
|
||||
backfill_tool_call_reasoning_placeholders(&mut messages);
|
||||
let messages = collapse_system_messages_to_head(messages);
|
||||
result["messages"] = json!(messages);
|
||||
|
||||
let model = body.get("model").and_then(|v| v.as_str()).unwrap_or("");
|
||||
if let Some(max_tokens) = body.get("max_output_tokens") {
|
||||
if super::transform::is_openai_o_series(model) {
|
||||
result["max_completion_tokens"] = max_tokens.clone();
|
||||
@@ -703,10 +710,109 @@ fn append_responses_input_as_chat_messages(
|
||||
last_assistant_index,
|
||||
&mut pending_reasoning,
|
||||
);
|
||||
backfill_tool_call_reasoning_placeholders(messages);
|
||||
Ok(())
|
||||
}
|
||||
|
||||
fn is_deepseek_chat_model(model: &str) -> bool {
|
||||
let name = model
|
||||
.rsplit('/')
|
||||
.next()
|
||||
.unwrap_or_default()
|
||||
.to_ascii_lowercase();
|
||||
name == "deepseek" || name.starts_with("deepseek-")
|
||||
}
|
||||
|
||||
/// Responses can replay assistant text (including an empty tool-only turn's
|
||||
/// trailing message) before all tool results have arrived. Chat requires those
|
||||
/// results to follow one assistant tool-call message without another assistant
|
||||
/// in between. Keep the replayed content on that message until its calls are
|
||||
/// answered, without crossing user/system boundaries or fabricating results.
|
||||
fn coalesce_replayed_tool_turn_messages(messages: &mut Vec<Value>) {
|
||||
let mut normalized: Vec<Value> = Vec::with_capacity(messages.len());
|
||||
let mut open_tool_turn: Option<usize> = None;
|
||||
let mut pending_call_ids = HashSet::new();
|
||||
|
||||
for mut message in std::mem::take(messages) {
|
||||
match message.get("role").and_then(Value::as_str) {
|
||||
Some("assistant") => {
|
||||
if let Some(calls) = message.get("tool_calls").and_then(Value::as_array) {
|
||||
for call in calls {
|
||||
register_known_tool_call_id(&mut pending_call_ids, call);
|
||||
}
|
||||
}
|
||||
if let Some(index) = open_tool_turn {
|
||||
let target = &mut normalized[index];
|
||||
merge_chat_assistant_content(target, message["content"].take());
|
||||
let mut reasoning = message
|
||||
.get("reasoning_content")
|
||||
.and_then(Value::as_str)
|
||||
.map(ToString::to_string);
|
||||
attach_pending_reasoning_to_assistant_unique(target, &mut reasoning);
|
||||
if let Some(calls) = message.get_mut("tool_calls").and_then(Value::as_array_mut)
|
||||
{
|
||||
target["tool_calls"].as_array_mut().unwrap().append(calls);
|
||||
}
|
||||
continue;
|
||||
}
|
||||
if !pending_call_ids.is_empty() {
|
||||
open_tool_turn = Some(normalized.len());
|
||||
}
|
||||
}
|
||||
Some("tool") => {
|
||||
if let Some(call_id) = message.get("tool_call_id").and_then(Value::as_str) {
|
||||
pending_call_ids.remove(call_id);
|
||||
}
|
||||
if pending_call_ids.is_empty() {
|
||||
open_tool_turn = None;
|
||||
}
|
||||
}
|
||||
_ => {
|
||||
open_tool_turn = None;
|
||||
pending_call_ids.clear();
|
||||
}
|
||||
}
|
||||
normalized.push(message);
|
||||
}
|
||||
*messages = normalized;
|
||||
}
|
||||
|
||||
fn merge_chat_assistant_content(target: &mut Value, content: Value) {
|
||||
fn is_empty(content: &Value) -> bool {
|
||||
content.is_null()
|
||||
|| content.as_str().is_some_and(str::is_empty)
|
||||
|| content.as_array().is_some_and(Vec::is_empty)
|
||||
}
|
||||
|
||||
fn into_parts(content: Value) -> Vec<Value> {
|
||||
match content {
|
||||
Value::Array(parts) => parts,
|
||||
Value::String(text) => vec![json!({"type": "text", "text": text})],
|
||||
other => vec![other],
|
||||
}
|
||||
}
|
||||
|
||||
if is_empty(&content) {
|
||||
return;
|
||||
}
|
||||
if is_empty(&target["content"]) {
|
||||
target["content"] = content;
|
||||
return;
|
||||
}
|
||||
let previous = target["content"].take();
|
||||
target["content"] = match (previous, content) {
|
||||
(Value::String(mut previous), Value::String(text)) => {
|
||||
previous.push('\n');
|
||||
previous.push_str(&text);
|
||||
Value::String(previous)
|
||||
}
|
||||
(previous, content) => {
|
||||
let mut parts = into_parts(previous);
|
||||
parts.extend(into_parts(content));
|
||||
Value::Array(parts)
|
||||
}
|
||||
};
|
||||
}
|
||||
|
||||
#[allow(clippy::too_many_arguments)]
|
||||
fn append_responses_item_as_chat_message(
|
||||
item: &Value,
|
||||
@@ -2376,6 +2482,292 @@ mod tests {
|
||||
result["messages"].as_array().unwrap()
|
||||
}
|
||||
|
||||
fn convert_deepseek_test_input(items: Vec<Value>) -> Value {
|
||||
responses_to_chat_completions(json!({
|
||||
"model": "deepseek/deepseek-v4-flash",
|
||||
"input": items
|
||||
}))
|
||||
.unwrap()
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_limits_tool_replay_repair_to_deepseek() {
|
||||
for (model, repair) in [
|
||||
("deepseek-v4-flash", true),
|
||||
("deepseek/deepseek-v4-flash", true),
|
||||
("deepseek-ai/DeepSeek-V3.2", true),
|
||||
("Pro/deepseek-ai/DeepSeek-R1", true),
|
||||
("DEEPSEEK-CHAT", true),
|
||||
("deepseek", true),
|
||||
("gpt-5.4", false),
|
||||
("kimi-k3", false),
|
||||
("glm-5.2", false),
|
||||
("MiniMax-M2.7", false),
|
||||
("qwen3.6-plus", false),
|
||||
("claude-sonnet-4-6", false),
|
||||
("custom-alias", false),
|
||||
("deepseek/gpt-5.4", false),
|
||||
("not-deepseek-v4-flash", false),
|
||||
] {
|
||||
let result = responses_to_chat_completions(json!({
|
||||
"model": model,
|
||||
"input": [
|
||||
test_function_call("call_1"),
|
||||
{"type": "message", "role": "assistant", "content": "Replay text.",
|
||||
"reasoning_content": "replay thought"},
|
||||
test_function_output("call_1", json!("ok"))
|
||||
]
|
||||
}))
|
||||
.unwrap();
|
||||
|
||||
if repair {
|
||||
assert_complete_tool_turns(&result);
|
||||
assert_eq!(message_roles(&result), vec!["assistant", "tool"], "{model}");
|
||||
assert_eq!(result["messages"][0]["content"], "Replay text.");
|
||||
assert_eq!(result["messages"][0]["reasoning_content"], "replay thought");
|
||||
} else {
|
||||
assert_eq!(
|
||||
result["messages"],
|
||||
json!([
|
||||
{"role": "assistant", "content": null, "tool_calls": [
|
||||
{"id": "call_1", "type": "function", "function": {
|
||||
"name": "view_image", "arguments": "{}"
|
||||
}}
|
||||
], "reasoning_content": "tool call"},
|
||||
{"role": "assistant", "content": "Replay text.", "reasoning_content": "replay thought"},
|
||||
{"role": "tool", "tool_call_id": "call_1", "content": "ok"}
|
||||
]),
|
||||
"{model}"
|
||||
);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
fn assert_complete_tool_turns(result: &Value) {
|
||||
let mut pending = HashSet::new();
|
||||
for message in result_messages(result) {
|
||||
let role = message["role"].as_str().unwrap();
|
||||
if !pending.is_empty() {
|
||||
assert_eq!(role, "tool", "interrupted tool turn: {result}");
|
||||
}
|
||||
if role == "tool" {
|
||||
let call_id = message["tool_call_id"].as_str().unwrap();
|
||||
assert!(pending.remove(call_id), "unexpected tool result: {result}");
|
||||
}
|
||||
if let Some(calls) = message["tool_calls"].as_array() {
|
||||
for call in calls {
|
||||
assert!(pending.insert(call["id"].as_str().unwrap()));
|
||||
}
|
||||
}
|
||||
}
|
||||
assert!(pending.is_empty(), "missing tool results: {result}");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_keeps_valid_deepseek_tool_history_bytes_stable() {
|
||||
let items = vec![
|
||||
json!({"type": "message", "role": "user", "content": "Inspect the image."}),
|
||||
json!({"type": "message", "role": "assistant", "content": "Checking."}),
|
||||
test_function_call("call_1"),
|
||||
test_function_call("call_2"),
|
||||
test_function_output("call_2", json!("")),
|
||||
test_function_output(
|
||||
"call_1",
|
||||
json!([
|
||||
{"type": "text", "text": "image result"},
|
||||
{"type": "input_image", "image_url": "https://example.com/image.png"}
|
||||
]),
|
||||
),
|
||||
json!({"type": "message", "role": "assistant", "content": "Image inspected.",
|
||||
"reasoning_content": "final answer thought"}),
|
||||
json!({"type": "message", "role": "user", "content": "Continue."}),
|
||||
test_function_call("call_next"),
|
||||
test_function_output("call_next", json!("ok")),
|
||||
];
|
||||
let original = convert_test_input(items.clone());
|
||||
let deepseek = convert_deepseek_test_input(items);
|
||||
|
||||
assert_complete_tool_turns(&deepseek);
|
||||
assert_eq!(
|
||||
serde_json::to_vec(&deepseek["messages"]).unwrap(),
|
||||
serde_json::to_vec(&original["messages"]).unwrap(),
|
||||
"well-formed history must keep its cache-sensitive serialization"
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_keeps_empty_assistant_replay_inside_tool_turn() {
|
||||
let empty_messages = [
|
||||
json!({"type": "message", "role": "assistant", "content": null}),
|
||||
json!({"type": "message", "role": "assistant", "content": null,
|
||||
"reasoning_content": "tool-only thought"}),
|
||||
json!({"type": "message", "role": "assistant", "content": ""}),
|
||||
json!({"type": "message", "role": "assistant", "content": []}),
|
||||
json!({"type": "message", "role": "assistant", "content": [
|
||||
{"type": "output_text", "text": ""}
|
||||
]}),
|
||||
json!({"type": "message", "role": "assistant"}),
|
||||
json!({"role": "assistant", "content": ""}),
|
||||
json!({"type": "future_message", "role": "assistant", "content": ""}),
|
||||
json!({"type": "input_text", "role": "assistant", "text": ""}),
|
||||
];
|
||||
|
||||
for empty_message in empty_messages {
|
||||
let reasoning = empty_message.get("reasoning_content").cloned();
|
||||
let result = convert_deepseek_test_input(vec![
|
||||
test_function_call("call_1"),
|
||||
empty_message,
|
||||
test_function_output("call_1", json!("")),
|
||||
]);
|
||||
|
||||
assert_complete_tool_turns(&result);
|
||||
assert_eq!(message_roles(&result), vec!["assistant", "tool"]);
|
||||
assert_eq!(result["messages"][0]["tool_calls"][0]["id"], "call_1");
|
||||
assert_eq!(result["messages"][1]["content"], "");
|
||||
if let Some(reasoning) = reasoning {
|
||||
assert_eq!(result["messages"][0]["reasoning_content"], reasoning);
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_coalesces_custom_and_search_tool_replay() {
|
||||
for (call, output) in [
|
||||
(
|
||||
json!({"type": "custom_tool_call", "call_id": "call_1",
|
||||
"name": "apply_patch", "input": "patch"}),
|
||||
json!({"type": "custom_tool_call_output", "call_id": "call_1", "output": "ok"}),
|
||||
),
|
||||
(
|
||||
json!({"type": "tool_search_call", "call_id": "call_1", "arguments": {"query": "shell"}}),
|
||||
json!({"type": "tool_search_output", "call_id": "call_1", "output": []}),
|
||||
),
|
||||
] {
|
||||
let expected_output = canonical_json_string(&output);
|
||||
let result = convert_deepseek_test_input(vec![
|
||||
call,
|
||||
json!({"type": "message", "role": "assistant", "content": ""}),
|
||||
output,
|
||||
]);
|
||||
assert_complete_tool_turns(&result);
|
||||
assert_eq!(message_roles(&result), vec!["assistant", "tool"]);
|
||||
assert_eq!(result["messages"][1]["content"], expected_output);
|
||||
}
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_preserves_replayed_tool_turn_text_and_reasoning() {
|
||||
let mut first_call = test_function_call("call_1");
|
||||
first_call["reasoning_content"] = json!("first thought");
|
||||
let mut second_call = test_function_call("call_2");
|
||||
second_call["reasoning_content"] = json!("third thought");
|
||||
let result = convert_deepseek_test_input(vec![
|
||||
first_call,
|
||||
json!({
|
||||
"type": "message", "role": "assistant", "content": "First update.",
|
||||
"reasoning_content": "second thought"
|
||||
}),
|
||||
second_call,
|
||||
json!({"type": "message", "role": "assistant", "content": "Second update."}),
|
||||
test_function_output("call_1", json!("first result")),
|
||||
test_function_output("call_2", json!("second result")),
|
||||
]);
|
||||
|
||||
assert_complete_tool_turns(&result);
|
||||
assert_eq!(message_roles(&result), vec!["assistant", "tool", "tool"]);
|
||||
let assistant = &result["messages"][0];
|
||||
assert_eq!(assistant["content"], "First update.\nSecond update.");
|
||||
assert_eq!(
|
||||
assistant["reasoning_content"],
|
||||
"first thought\n\nsecond thought\n\nthird thought"
|
||||
);
|
||||
assert_eq!(assistant["tool_calls"][0]["id"], "call_1");
|
||||
assert_eq!(assistant["tool_calls"][1]["id"], "call_2");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_coalesces_assistant_replay_between_parallel_results() {
|
||||
let result = convert_deepseek_test_input(vec![
|
||||
test_function_call("call_1"),
|
||||
test_function_call("call_2"),
|
||||
test_function_output("call_2", json!("second result")),
|
||||
json!({
|
||||
"type": "message", "role": "assistant", "content": "Still working.",
|
||||
"reasoning_content": "waiting for the first result"
|
||||
}),
|
||||
test_function_output("call_1", json!("first result")),
|
||||
json!({"type": "message", "role": "assistant", "content": "Done."}),
|
||||
]);
|
||||
|
||||
assert_complete_tool_turns(&result);
|
||||
assert_eq!(
|
||||
message_roles(&result),
|
||||
vec!["assistant", "tool", "tool", "assistant"]
|
||||
);
|
||||
assert_eq!(result["messages"][0]["content"], "Still working.");
|
||||
assert_eq!(
|
||||
result["messages"][0]["reasoning_content"],
|
||||
"waiting for the first result"
|
||||
);
|
||||
assert_eq!(result["messages"][1]["tool_call_id"], "call_2");
|
||||
assert_eq!(result["messages"][2]["tool_call_id"], "call_1");
|
||||
assert_eq!(result["messages"][3]["content"], "Done.");
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_preserves_multimodal_assistant_replay() {
|
||||
let result = convert_deepseek_test_input(vec![
|
||||
json!({"type": "message", "role": "assistant", "content": "Opening the image."}),
|
||||
test_function_call("call_1"),
|
||||
json!({"type": "message", "role": "assistant", "content": [
|
||||
{"type": "output_text", "text": "Image context."},
|
||||
{"type": "refusal", "refusal": "Cannot identify a person."},
|
||||
{"type": "input_image", "image_url": "https://example.com/image.png"}
|
||||
]}),
|
||||
test_function_output("call_1", json!("ok")),
|
||||
]);
|
||||
|
||||
assert_complete_tool_turns(&result);
|
||||
assert_eq!(message_roles(&result), vec!["assistant", "tool"]);
|
||||
assert_eq!(
|
||||
result["messages"][0]["content"],
|
||||
json!([
|
||||
{"type": "text", "text": "Opening the image."},
|
||||
{"type": "text", "text": "Image context."},
|
||||
{"type": "text", "text": "Cannot identify a person."},
|
||||
{"type": "image_url", "image_url": {"url": "https://example.com/image.png"}}
|
||||
])
|
||||
);
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn responses_request_to_chat_keeps_user_boundaries_and_missing_results() {
|
||||
for role in ["user", "system"] {
|
||||
let result = convert_deepseek_test_input(vec![
|
||||
test_function_call("call_missing"),
|
||||
json!({"type": "message", "role": role, "content": "New context."}),
|
||||
json!({"type": "message", "role": "assistant", "content": "Next turn."}),
|
||||
test_function_call("call_next"),
|
||||
test_function_output("call_next", json!("ok")),
|
||||
]);
|
||||
let messages = result_messages(&result);
|
||||
let assistants: Vec<_> = messages
|
||||
.iter()
|
||||
.filter(|message| message["role"] == "assistant")
|
||||
.collect();
|
||||
assert_eq!(assistants.len(), 2);
|
||||
assert_eq!(assistants[0]["tool_calls"][0]["id"], "call_missing");
|
||||
assert_eq!(assistants[1]["tool_calls"][0]["id"], "call_next");
|
||||
assert_eq!(assistants[1]["content"], "Next turn.");
|
||||
let outputs: Vec<_> = messages
|
||||
.iter()
|
||||
.filter(|message| message["role"] == "tool")
|
||||
.collect();
|
||||
assert_eq!(outputs.len(), 1);
|
||||
assert_eq!(outputs[0]["tool_call_id"], "call_next");
|
||||
}
|
||||
}
|
||||
|
||||
/// 回归：没有配对 function_call、call_id 为空的工具结果（Codex 桌面端
|
||||
/// 新建会话时注入的 create_thread 回执就是这个形态）不能序列化成 role=tool，
|
||||
/// 否则严格上游会以「Messages with role 'tool' must be a response to a
|
||||
|
||||
Reference in new issue
Block a user