feat(codex): support native Anthropic Messages protocol as upstream (#5071)

* feat(codex): support native Anthropic Messages protocol as upstream

Allow gateways that only expose the native Anthropic Messages protocol
(/v1/messages) to be used by Codex: the local proxy performs bidirectional
request/response/streaming conversion between Responses and Anthropic.

Backend:
- Add two conversion modules: transform_codex_anthropic / streaming_codex_anthropic
- codex.rs: add routing detection and auth: ANTHROPIC_AUTH_TOKEN→Bearer (default),
  ANTHROPIC_API_KEY→x-api-key, mutually exclusive
- handlers.rs: add handle_codex_anthropic_to_responses_transform
- forwarder.rs: support optional Claude Code client fingerprint impersonation
  (User-Agent / anthropic-beta / x-app / system prompt first-line injection)
  and /responses→/v1/messages rewriting
- codex_config: the anthropic format reuses the NativeResponses profile to strip
  custom tools
- ProviderMeta: add impersonateClaudeCode

Frontend:
- CodexApiFormat: add "anthropic"; the form adds auth field selection and an
  impersonation toggle
- Add en/ja/zh/zh-TW copy

Robustness:
- Downgrade when tool history / forced tool_choice conflicts with extended
  thinking, avoiding upstream 400s
- Emit cache_creation_input_tokens in usage and use saturating_add to guard
  against overflow
- Append a unique suffix to non-streaming/streaming output-item ids to avoid
  multi-segment text/thinking overwriting each other

* fix(codex): harden Anthropic bridge against empty text blocks and truncated streams

- Drop empty/whitespace-only text content blocks when rebuilding Anthropic
  messages from Responses history; Anthropic 400s on empty text blocks (e.g. an
  empty assistant text emitted alongside a tool_use), which broke follow-up and
  tool-result requests. Also drop messages left without content.
- Do not report a truncated Anthropic stream as completed: when the SSE
  connection ends before message_stop with no stop_reason, emit an incomplete
  response if partial output exists, or a failed (stream_truncated) response
  otherwise, mirroring the chat converter's EOF handling.

* fix(codex): resolve Anthropic-bridge review findings

Blocking:
1. Defer stripping the [1m] long-context marker until after catalog matching and model write-back, re-stripping it on the final Codex→Anthropic body and setting a flag to emit the context-1m-2025-08-07 beta header, so the marker is no longer lost or overridden by the default model.
2. Gate Anthropic thinking on the trailing turn only (via trailing_turn_allows_thinking) instead of scanning full history, so a Codex session that resends history each round no longer permanently loses thinking after the first tool call.
3. Inject 5m ephemeral cache_control on the Codex→Anthropic body by reusing cache_injector (handling the system string→array conversion), so system/tools/history are cached instead of re-sent at full price every round.
4. Add a shared base_url_is_full_endpoint helper (normalizing whitespace/query/fragment/trailing slash) used by both the Anthropic and Chat paths, so a base URL already ending in /v1/messages is treated as a full endpoint instead of double-appending to /v1/messages/v1/messages.
5. Align the catalog tool-profile predicate with the routing predicate so resolve_codex_catalog_tool_profile returns the Anthropic profile whenever the request converts to Anthropic, preventing freeform tools like apply_patch from being silently filtered by a ProxyChat catalog.
6. Explicitly disable native web_search for the Anthropic profile (including the no-catalog branch) via set_codex_native_web_search_field, so Codex no longer treats it as available while the transform silently drops it.
7. Only forward tool_choice when tools survive filtering, dropping it otherwise, to avoid a non-retryable 400 from upstream when tool_choice is sent with no tools.
8. Lower the fallback default to max_tokens=8192 (only when max_output_tokens is omitted) and clamp the thinking budget to max_tokens/2, disabling thinking below the 1024 floor, to avoid hard 400s on low-output-ceiling models/gateways.

Minor:
9. Centralize the Codex/OpenAI fingerprint-header denylist in is_codex_client_fingerprint_header so impersonating Claude Code uniformly drops originator/session_id/conversation_id/chatgpt-account-id/x-client-request-id/openai-* and the x-stainless-*/x-codex-* prefixes.
10. Retain content_block_start.input as start_input and fall back to it at block close when no input_json_delta arrived, so a gateway that carries the full tool input on the start event no longer yields empty tool arguments.
11. Extract a shared codex_responses_sse module as the single Responses SSE envelope builder that both the chat and anthropic streaming emitters delegate to, with byte-for-byte-unchanged wire output, so future event-format fixes touch one place instead of two.

* fix(codex): add per-provider max_output_tokens override for Codex→Anthropic path

Codex does not forward model_max_output_tokens in the request body,
causing the proxy to fall back to a conservative 8192 default. This
truncates long or thinking-heavy responses (stop_reason=max_tokens).

- Add maxOutputTokens field to ProviderMeta (Rust + TypeScript)
- Inject the value into the Anthropic request body before transform,
  taking precedence over request-supplied and default values
- Add numeric input in Codex form fields (Anthropic format only)
- Add i18n entries for label, placeholder, and hint (en/zh/zh-TW/ja)
- Include roundtrip and omission unit tests for the new field

* fix(codex): harden Anthropic protocol bridge

* fix(codex): address Anthropic bridge review

* fix(codex): preserve flattened Anthropic inputs

* fix(codex): harden Anthropic recovery paths

* fix(ci): satisfy Rust clippy

---------

Co-authored-by: Jason <farion1231@gmail.com>
This commit is contained in:
Yeeyzy
2026-07-10 23:06:30 +08:00
committed by GitHub
parent 98ccde0050
commit 99e11e0851
29 changed files with 5733 additions and 371 deletions
+300 -1
View File
@@ -84,6 +84,72 @@ pub fn should_convert_codex_responses_to_chat(provider: &Provider, endpoint: &st
) && codex_provider_uses_chat_completions(provider)
}
/// Whether this Codex provider's real upstream speaks the native Anthropic
/// Messages protocol (`/v1/messages`). The local Codex client always talks to CC
/// Switch through the Responses API, so CC Switch bridges Responses ⇄ Anthropic.
///
/// Determined solely from explicit config (apiFormat / wire_api); no base_url
/// guessing — Anthropic gateway addresses vary widely and guessing easily misfires.
pub fn codex_provider_uses_anthropic(provider: &Provider) -> bool {
if let Some(api_format) = provider
.meta
.as_ref()
.and_then(|meta| meta.api_format.as_deref())
.or_else(|| {
provider
.settings_config
.get("api_format")
.and_then(|v| v.as_str())
})
.or_else(|| {
provider
.settings_config
.get("apiFormat")
.and_then(|v| v.as_str())
})
{
return is_anthropic_wire_api(api_format);
}
provider
.settings_config
.get("config")
.and_then(|v| v.as_str())
.and_then(extract_codex_wire_api_from_toml)
.map(|wire_api| is_anthropic_wire_api(&wire_api))
.unwrap_or(false)
}
pub fn should_convert_codex_responses_to_anthropic(provider: &Provider, endpoint: &str) -> bool {
let path = endpoint
.split_once('?')
.map_or(endpoint, |(path, _query)| path);
matches!(
path,
"/responses" | "/v1/responses" | "/responses/compact" | "/v1/responses/compact"
) && codex_provider_uses_anthropic(provider)
}
/// Resolve the model-catalog tool profile for a Codex provider using the SAME
/// Anthropic detection as the proxy router ([`codex_provider_uses_anthropic`]), so the
/// generated catalog never disagrees with the routed transform. A provider whose
/// Anthropic upstream is declared only via settings `apiFormat` or TOML `wire_api`
/// (not `meta.api_format`) would otherwise get a `ProxyChat` catalog and emit the
/// freeform `apply_patch` tool that the Anthropic transform then silently drops.
/// Non-Anthropic providers keep the existing `meta.api_format` classification.
pub fn resolve_codex_catalog_tool_profile(
provider: &Provider,
) -> crate::codex_config::CodexCatalogToolProfile {
use crate::codex_config::CodexCatalogToolProfile;
if codex_provider_uses_anthropic(provider) {
return CodexCatalogToolProfile::Anthropic;
}
CodexCatalogToolProfile::from_api_format(
provider.meta.as_ref().and_then(|m| m.api_format.as_deref()),
)
}
/// Extract the real upstream model configured for a Codex provider.
pub fn codex_provider_upstream_model(provider: &Provider) -> Option<String> {
provider
@@ -129,7 +195,13 @@ pub fn apply_codex_chat_upstream_model(
if !codex_provider_uses_chat_completions(provider) {
return None;
}
apply_codex_upstream_model(provider, body)
}
/// Same model-substitution logic as `apply_codex_chat_upstream_model`, but without
/// the chat gating check. Reused by the anthropic conversion path (the forwarder has
/// already confirmed this provider uses anthropic).
pub fn apply_codex_upstream_model(provider: &Provider, body: &mut JsonValue) -> Option<String> {
let catalog_model_ids = codex_provider_catalog_model_ids(provider);
if let Some(request_model) = body
.get("model")
@@ -345,6 +417,13 @@ fn is_chat_wire_api(value: &str) -> bool {
)
}
fn is_anthropic_wire_api(value: &str) -> bool {
matches!(
value.trim().to_ascii_lowercase().as_str(),
"anthropic" | "anthropic_messages" | "anthropic-messages" | "claude" | "messages"
)
}
fn is_chat_completions_url(value: &str) -> bool {
value
.trim_end_matches('/')
@@ -527,8 +606,28 @@ impl ProviderAdapter for CodexAdapter {
}
fn extract_auth(&self, provider: &Provider) -> Option<AuthInfo> {
// Anthropic upstream: the auth field is chosen by the user in the UI (meta.apiKeyField).
// ANTHROPIC_API_KEY → x-api-key (AuthStrategy::Anthropic)
// ANTHROPIC_AUTH_TOKEN → Authorization: Bearer (default, AuthStrategy::Bearer)
// The two are mutually exclusive to avoid a 401 from the gateway receiving
// both auth headers at once. All other Codex upstreams stay pure Bearer.
let strategy = if codex_provider_uses_anthropic(provider) {
let uses_x_api_key = provider
.meta
.as_ref()
.and_then(|meta| meta.api_key_field.as_deref())
.map(|field| field.eq_ignore_ascii_case("ANTHROPIC_API_KEY"))
.unwrap_or(false);
if uses_x_api_key {
AuthStrategy::Anthropic
} else {
AuthStrategy::Bearer
}
} else {
AuthStrategy::Bearer
};
self.extract_key(provider)
.map(|key| AuthInfo::new(key, AuthStrategy::Bearer))
.map(|key| AuthInfo::new(key, strategy))
}
fn build_url(&self, base_url: &str, endpoint: &str) -> String {
@@ -569,6 +668,15 @@ impl ProviderAdapter for CodexAdapter {
) -> Result<Vec<(http::HeaderName, http::HeaderValue)>, ProxyError> {
use super::adapter::auth_header_value;
let bearer = format!("Bearer {}", auth.api_key);
// Anthropic gateway: send only x-api-key (anthropic-version is filled in by
// the forwarder). Mutually exclusive with Bearer to avoid a 401 from the
// gateway receiving both auth headers at once.
if auth.strategy == AuthStrategy::Anthropic {
return Ok(vec![(
http::HeaderName::from_static("x-api-key"),
auth_header_value(&auth.api_key)?,
)]);
}
Ok(vec![(
http::HeaderName::from_static("authorization"),
auth_header_value(&bearer)?,
@@ -662,6 +770,197 @@ experimental_bearer_token = "sk-config-key"
assert_eq!(url, "https://api.openai.com/v1/responses");
}
// ==================== anthropic upstream detection ====================
#[test]
fn test_uses_anthropic_from_settings_api_format() {
let provider = create_provider(json!({ "apiFormat": "anthropic" }));
assert!(codex_provider_uses_anthropic(&provider));
let provider = create_provider(json!({ "api_format": "anthropic_messages" }));
assert!(codex_provider_uses_anthropic(&provider));
}
#[test]
fn test_uses_anthropic_from_meta_api_format() {
let mut provider = create_provider(json!({}));
provider.meta = Some(crate::provider::ProviderMeta {
api_format: Some("anthropic".to_string()),
..Default::default()
});
assert!(codex_provider_uses_anthropic(&provider));
}
#[test]
fn test_uses_anthropic_from_toml_wire_api() {
let provider = create_provider(json!({
"config": r#"model_provider = "custom"
[model_providers.custom]
wire_api = "anthropic"
"#
}));
assert!(codex_provider_uses_anthropic(&provider));
}
#[test]
fn test_anthropic_false_for_chat_and_responses() {
let chat = create_provider(json!({ "apiFormat": "openai_chat" }));
assert!(!codex_provider_uses_anthropic(&chat));
let responses = create_provider(json!({ "apiFormat": "openai_responses" }));
assert!(!codex_provider_uses_anthropic(&responses));
}
#[test]
fn test_anthropic_and_chat_are_mutually_exclusive() {
let anth = create_provider(json!({ "apiFormat": "anthropic" }));
assert!(codex_provider_uses_anthropic(&anth));
assert!(!codex_provider_uses_chat_completions(&anth));
let chat = create_provider(json!({ "apiFormat": "openai_chat" }));
assert!(codex_provider_uses_chat_completions(&chat));
assert!(!codex_provider_uses_anthropic(&chat));
}
#[test]
fn test_should_convert_responses_to_anthropic_path_guard() {
let provider = create_provider(json!({ "apiFormat": "anthropic" }));
assert!(should_convert_codex_responses_to_anthropic(
&provider,
"/responses"
));
assert!(should_convert_codex_responses_to_anthropic(
&provider,
"/v1/responses/compact"
));
assert!(should_convert_codex_responses_to_anthropic(
&provider,
"/responses?x=1"
));
assert!(!should_convert_codex_responses_to_anthropic(
&provider,
"/chat/completions"
));
}
#[test]
fn test_resolve_catalog_profile_matches_router() {
use crate::codex_config::CodexCatalogToolProfile;
// Anthropic declared only via TOML wire_api (no meta.api_format) must still
// resolve to the Anthropic catalog profile — this is the routing/catalog
// divergence that let apply_patch leak through.
let toml_anthropic = create_provider(json!({
"config": r#"model_provider = "custom"
[model_providers.custom]
wire_api = "anthropic"
"#
}));
assert_eq!(
resolve_codex_catalog_tool_profile(&toml_anthropic),
CodexCatalogToolProfile::Anthropic
);
// Anthropic via settings apiFormat.
let settings_anthropic = create_provider(json!({ "apiFormat": "anthropic" }));
assert_eq!(
resolve_codex_catalog_tool_profile(&settings_anthropic),
CodexCatalogToolProfile::Anthropic
);
// Native openai_responses (meta) → NativeResponses; chat → ProxyChat.
let mut native = create_provider(json!({}));
native.meta = Some(crate::provider::ProviderMeta {
api_format: Some("openai_responses".to_string()),
..Default::default()
});
assert_eq!(
resolve_codex_catalog_tool_profile(&native),
CodexCatalogToolProfile::NativeResponses
);
let chat = create_provider(json!({ "apiFormat": "openai_chat" }));
assert_eq!(
resolve_codex_catalog_tool_profile(&chat),
CodexCatalogToolProfile::ProxyChat
);
}
#[test]
fn test_apply_codex_upstream_model_preserves_one_m_catalog_model() {
// Regression for the [1m] path: a request model carrying the [1m] marker must
// match its catalog entry and be preserved (not overridden by the provider
// default) so the transform can later strip [1m] and emit the context-1m beta.
// This only works because the forwarder no longer strips [1m] before this call
// on the Anthropic path.
let provider = create_provider(json!({
"config": r#"model_provider = "custom"
model = "claude-opus-4-1"
[model_providers.custom]
wire_api = "anthropic"
"#,
"modelCatalog": {
"models": [
{ "model": "claude-opus-4-1[1m]" }
]
}
}));
let mut body = json!({ "model": "claude-opus-4-1[1m]", "input": "hi" });
let result = apply_codex_upstream_model(&provider, &mut body);
assert_eq!(result.as_deref(), Some("claude-opus-4-1[1m]"));
assert_eq!(
body.get("model").and_then(|v| v.as_str()),
Some("claude-opus-4-1[1m]")
);
}
#[test]
fn test_anthropic_auth_defaults_to_bearer() {
// No meta.apiKeyField (defaults to ANTHROPIC_AUTH_TOKEN) → Authorization: Bearer only
let adapter = CodexAdapter::new();
let provider = create_provider(json!({
"apiFormat": "anthropic",
"auth": { "OPENAI_API_KEY": "sk-anthropic-key-123" }
}));
let auth = adapter.extract_auth(&provider).unwrap();
assert_eq!(auth.strategy, AuthStrategy::Bearer);
let headers = adapter.get_auth_headers(&auth).unwrap();
let names: Vec<String> = headers
.iter()
.map(|(name, _)| name.as_str().to_string())
.collect();
assert_eq!(names, vec!["authorization".to_string()]);
}
#[test]
fn test_anthropic_auth_x_api_key_when_selected() {
// meta.apiKeyField = ANTHROPIC_API_KEY → x-api-key only
let adapter = CodexAdapter::new();
let mut provider = create_provider(json!({
"apiFormat": "anthropic",
"auth": { "OPENAI_API_KEY": "sk-anthropic-key-123" }
}));
provider.meta = Some(crate::provider::ProviderMeta {
api_format: Some("anthropic".to_string()),
api_key_field: Some("ANTHROPIC_API_KEY".to_string()),
..Default::default()
});
let auth = adapter.extract_auth(&provider).unwrap();
assert_eq!(auth.strategy, AuthStrategy::Anthropic);
let headers = adapter.get_auth_headers(&auth).unwrap();
let names: Vec<String> = headers
.iter()
.map(|(name, _)| name.as_str().to_string())
.collect();
assert_eq!(names, vec!["x-api-key".to_string()]);
}
#[test]
fn test_build_url_origin_adds_v1() {
let adapter = CodexAdapter::new();
@@ -0,0 +1,399 @@
//! Shared builders for the OpenAI Responses SSE envelope.
//!
//! The two Codex streaming converters — `streaming_codex_chat` (Chat Completions SSE →
//! Responses SSE) and `streaming_codex_anthropic` (Anthropic Messages SSE → Responses
//! SSE) — have completely different *input* state machines but must emit the identical
//! Responses event stream the Codex client understands. This module owns that output
//! envelope so the two converters cannot drift when an event's shape changes: a wire fix
//! lands here once instead of being mirrored in both files.
//!
//! Each function is pure — it takes primitives or a caller-built `item` `Value` and
//! returns the exact bytes the converters previously constructed inline. Item shapes that
//! vary per converter (including function, namespace, custom, and tool-search calls)
//! are supplied by the caller via the generic
//! `output_item_added` / `output_item_done` helpers.
use bytes::Bytes;
use serde_json::{json, Value};
/// Serialize one Responses SSE event with the standard `event:`/`data:` framing.
pub(crate) fn sse_event(event: &str, data: Value) -> Bytes {
Bytes::from(format!(
"event: {event}\ndata: {}\n\n",
serde_json::to_string(&data).unwrap_or_default()
))
}
// ---------------------------------------------------------------------------
// Response lifecycle (created / in_progress / completed / failed)
// ---------------------------------------------------------------------------
/// `response.created`, wrapping a caller-built `response` object (usage/created_at differ
/// per converter, so the caller supplies the whole object).
pub(crate) fn response_created(response: &Value) -> Bytes {
sse_event(
"response.created",
json!({ "type": "response.created", "response": response }),
)
}
/// `response.in_progress`.
pub(crate) fn response_in_progress(response: &Value) -> Bytes {
sse_event(
"response.in_progress",
json!({ "type": "response.in_progress", "response": response }),
)
}
/// `response.completed`.
pub(crate) fn response_completed(response: &Value) -> Bytes {
sse_event(
"response.completed",
json!({ "type": "response.completed", "response": response }),
)
}
/// `response.failed`.
pub(crate) fn response_failed(response: &Value) -> Bytes {
sse_event(
"response.failed",
json!({ "type": "response.failed", "response": response }),
)
}
// ---------------------------------------------------------------------------
// Generic output-item add/done (item value supplied by the caller)
// ---------------------------------------------------------------------------
/// `response.output_item.added` with a caller-built item (message / reasoning /
/// function_call / custom_tool_call).
pub(crate) fn output_item_added(output_index: u32, item: &Value) -> Bytes {
sse_event(
"response.output_item.added",
json!({
"type": "response.output_item.added",
"output_index": output_index,
"item": item
}),
)
}
/// `response.output_item.done` with a caller-built item.
pub(crate) fn output_item_done(output_index: u32, item: &Value) -> Bytes {
sse_event(
"response.output_item.done",
json!({
"type": "response.output_item.done",
"output_index": output_index,
"item": item
}),
)
}
// ---------------------------------------------------------------------------
// Assistant message (text) lifecycle
// ---------------------------------------------------------------------------
/// `response.output_item.added` for an in-progress assistant message.
pub(crate) fn message_item_added(output_index: u32, item_id: &str) -> Bytes {
output_item_added(
output_index,
&json!({
"id": item_id,
"type": "message",
"status": "in_progress",
"role": "assistant",
"content": []
}),
)
}
/// `response.content_part.added` for the (empty) output_text part of a message.
pub(crate) fn message_content_part_added(output_index: u32, item_id: &str) -> Bytes {
sse_event(
"response.content_part.added",
json!({
"type": "response.content_part.added",
"item_id": item_id,
"output_index": output_index,
"content_index": 0,
"part": { "type": "output_text", "text": "", "annotations": [] }
}),
)
}
/// `response.output_text.delta`.
pub(crate) fn output_text_delta(output_index: u32, item_id: &str, delta: &str) -> Bytes {
sse_event(
"response.output_text.delta",
json!({
"type": "response.output_text.delta",
"item_id": item_id,
"output_index": output_index,
"content_index": 0,
"delta": delta
}),
)
}
/// The completed assistant-message item value.
pub(crate) fn message_item(item_id: &str, text: &str) -> Value {
json!({
"id": item_id,
"type": "message",
"status": "completed",
"role": "assistant",
"content": [{ "type": "output_text", "text": text, "annotations": [] }]
})
}
/// Close an assistant message: emits `output_text.done` → `content_part.done` →
/// `output_item.done`, and returns the completed item so the caller can record it.
pub(crate) fn message_close(output_index: u32, item_id: &str, text: &str) -> (Vec<Bytes>, Value) {
let item = message_item(item_id, text);
let events = vec![
sse_event(
"response.output_text.done",
json!({
"type": "response.output_text.done",
"item_id": item_id,
"output_index": output_index,
"content_index": 0,
"text": text
}),
),
sse_event(
"response.content_part.done",
json!({
"type": "response.content_part.done",
"item_id": item_id,
"output_index": output_index,
"content_index": 0,
"part": { "type": "output_text", "text": text, "annotations": [] }
}),
),
output_item_done(output_index, &item),
];
(events, item)
}
// ---------------------------------------------------------------------------
// Reasoning (summary) lifecycle
// ---------------------------------------------------------------------------
/// `response.output_item.added` for an in-progress reasoning item.
pub(crate) fn reasoning_item_added(output_index: u32, item_id: &str) -> Bytes {
output_item_added(
output_index,
&json!({
"id": item_id,
"type": "reasoning",
"status": "in_progress",
"summary": []
}),
)
}
/// `response.reasoning_summary_part.added` for the (empty) summary part.
pub(crate) fn reasoning_summary_part_added(output_index: u32, item_id: &str) -> Bytes {
sse_event(
"response.reasoning_summary_part.added",
json!({
"type": "response.reasoning_summary_part.added",
"item_id": item_id,
"output_index": output_index,
"summary_index": 0,
"part": { "type": "summary_text", "text": "" }
}),
)
}
/// `response.reasoning_summary_text.delta`.
pub(crate) fn reasoning_summary_text_delta(output_index: u32, item_id: &str, delta: &str) -> Bytes {
sse_event(
"response.reasoning_summary_text.delta",
json!({
"type": "response.reasoning_summary_text.delta",
"item_id": item_id,
"output_index": output_index,
"summary_index": 0,
"delta": delta
}),
)
}
/// The completed reasoning item value (note: no `status` field, matching both converters).
pub(crate) fn reasoning_item(item_id: &str, text: &str) -> Value {
json!({
"id": item_id,
"type": "reasoning",
"summary": [{ "type": "summary_text", "text": text }]
})
}
/// Close a reasoning item: emits `reasoning_summary_text.done` →
/// `reasoning_summary_part.done` → `output_item.done`, and returns the completed item.
pub(crate) fn reasoning_close(output_index: u32, item_id: &str, text: &str) -> (Vec<Bytes>, Value) {
let item = reasoning_item(item_id, text);
let events = reasoning_close_with_item(output_index, item_id, text, &item, true);
(events, item)
}
/// Close a reasoning item whose completed shape is supplied by the converter.
/// Anthropic uses this to attach opaque signed/redacted thinking in
/// `encrypted_content` while keeping the standard Responses event lifecycle.
pub(crate) fn reasoning_close_with_item(
output_index: u32,
item_id: &str,
text: &str,
item: &Value,
has_visible_summary: bool,
) -> Vec<Bytes> {
let mut events = Vec::new();
if has_visible_summary {
events.extend([
sse_event(
"response.reasoning_summary_text.done",
json!({
"type": "response.reasoning_summary_text.done",
"item_id": item_id,
"output_index": output_index,
"summary_index": 0,
"text": text
}),
),
sse_event(
"response.reasoning_summary_part.done",
json!({
"type": "response.reasoning_summary_part.done",
"item_id": item_id,
"output_index": output_index,
"summary_index": 0,
"part": { "type": "summary_text", "text": text }
}),
),
]);
}
events.push(output_item_done(output_index, item));
events
}
// ---------------------------------------------------------------------------
// Tool-call argument streaming (item value supplied by the caller)
// ---------------------------------------------------------------------------
/// `response.function_call_arguments.delta`.
pub(crate) fn function_call_arguments_delta(
output_index: u32,
item_id: &str,
delta: &str,
) -> Bytes {
sse_event(
"response.function_call_arguments.delta",
json!({
"type": "response.function_call_arguments.delta",
"item_id": item_id,
"output_index": output_index,
"delta": delta
}),
)
}
/// `response.function_call_arguments.done`.
pub(crate) fn function_call_arguments_done(
output_index: u32,
item_id: &str,
arguments: &str,
) -> Bytes {
sse_event(
"response.function_call_arguments.done",
json!({
"type": "response.function_call_arguments.done",
"item_id": item_id,
"output_index": output_index,
"arguments": arguments
}),
)
}
/// `response.custom_tool_call_input.delta` (Chat freeform tools only).
pub(crate) fn custom_tool_call_input_delta(output_index: u32, item_id: &str, delta: &str) -> Bytes {
sse_event(
"response.custom_tool_call_input.delta",
json!({
"type": "response.custom_tool_call_input.delta",
"item_id": item_id,
"output_index": output_index,
"delta": delta
}),
)
}
/// `response.custom_tool_call_input.done` (Chat freeform tools only).
pub(crate) fn custom_tool_call_input_done(output_index: u32, item_id: &str, input: &str) -> Bytes {
sse_event(
"response.custom_tool_call_input.done",
json!({
"type": "response.custom_tool_call_input.done",
"item_id": item_id,
"output_index": output_index,
"input": input
}),
)
}
#[cfg(test)]
mod tests {
use super::*;
fn body(bytes: &Bytes) -> String {
String::from_utf8(bytes.to_vec()).unwrap()
}
#[test]
fn sse_event_framing() {
let ev = sse_event("response.created", json!({ "a": 1 }));
assert_eq!(body(&ev), "event: response.created\ndata: {\"a\":1}\n\n");
}
#[test]
fn message_close_shapes_match_legacy() {
let (events, item) = message_close(2, "resp_1_msg", "hi");
assert_eq!(events.len(), 3);
assert!(body(&events[0]).contains("\"type\":\"response.output_text.done\""));
assert!(body(&events[0]).contains("\"text\":\"hi\""));
assert!(body(&events[1]).contains("\"type\":\"response.content_part.done\""));
assert!(body(&events[2]).contains("\"type\":\"response.output_item.done\""));
assert_eq!(item["type"], "message");
assert_eq!(item["status"], "completed");
assert_eq!(item["content"][0]["text"], "hi");
}
#[test]
fn reasoning_close_item_has_no_status() {
let (events, item) = reasoning_close(0, "rs_1", "because");
assert_eq!(events.len(), 3);
assert!(body(&events[0]).contains("\"type\":\"response.reasoning_summary_text.done\""));
assert!(body(&events[1]).contains("\"type\":\"response.reasoning_summary_part.done\""));
// The completed reasoning item intentionally carries no `status` field.
assert!(item.get("status").is_none());
assert_eq!(item["summary"][0]["text"], "because");
}
#[test]
fn message_item_added_is_in_progress() {
let ev = message_item_added(0, "m1");
let s = body(&ev);
assert!(s.contains("\"type\":\"response.output_item.added\""));
assert!(s.contains("\"status\":\"in_progress\""));
assert!(s.contains("\"role\":\"assistant\""));
}
#[test]
fn function_call_argument_events() {
assert!(body(&function_call_arguments_delta(1, "fc_x", "{\"a\":"))
.contains("\"type\":\"response.function_call_arguments.delta\""));
assert!(body(&function_call_arguments_done(1, "fc_x", "{\"a\":1}"))
.contains("\"arguments\":\"{\\\"a\\\":1}\""));
}
}
+6 -2
View File
@@ -18,6 +18,7 @@ mod codex;
pub(crate) mod codex_chat_common;
pub mod codex_chat_history;
pub mod codex_oauth_auth;
pub(crate) mod codex_responses_sse;
pub mod copilot_auth;
pub mod copilot_model_map;
mod gemini;
@@ -25,10 +26,12 @@ pub(crate) mod gemini_schema;
pub mod gemini_shadow;
pub mod models;
pub mod streaming;
pub mod streaming_codex_anthropic;
pub mod streaming_codex_chat;
pub mod streaming_gemini;
pub mod streaming_responses;
pub mod transform;
pub mod transform_codex_anthropic;
pub mod transform_codex_chat;
pub mod transform_gemini;
pub mod transform_responses;
@@ -47,8 +50,9 @@ pub use claude::{
};
pub use codex::CodexAdapter;
pub use codex::{
apply_codex_chat_upstream_model, codex_provider_upstream_model,
resolve_codex_chat_reasoning_config, should_convert_codex_responses_to_chat,
apply_codex_chat_upstream_model, apply_codex_upstream_model, codex_provider_upstream_model,
resolve_codex_catalog_tool_profile, resolve_codex_chat_reasoning_config,
should_convert_codex_responses_to_anthropic, should_convert_codex_responses_to_chat,
};
pub use gemini::GeminiAdapter;
File diff suppressed because it is too large Load Diff
@@ -1,5 +1,6 @@
//! OpenAI Chat Completions SSE → OpenAI Responses SSE conversion.
use super::codex_responses_sse as sse;
use super::{
codex_chat_common::{
extract_reasoning_field_text, split_leading_think_block, strip_leading_think_open_tag,
@@ -270,20 +271,8 @@ impl ChatToResponsesState {
let response = self.base_response("in_progress", Vec::new());
vec![
sse_event(
"response.created",
json!({
"type": "response.created",
"response": response
}),
),
sse_event(
"response.in_progress",
json!({
"type": "response.in_progress",
"response": self.base_response("in_progress", Vec::new())
}),
),
sse::response_created(&response),
sse::response_in_progress(&response),
]
}
@@ -297,45 +286,16 @@ impl ChatToResponsesState {
self.reasoning.item_id = item_id.clone();
self.reasoning.added = true;
events.push(sse_event(
"response.output_item.added",
json!({
"type": "response.output_item.added",
"output_index": output_index,
"item": {
"id": item_id,
"type": "reasoning",
"status": "in_progress",
"summary": []
}
}),
));
events.push(sse_event(
"response.reasoning_summary_part.added",
json!({
"type": "response.reasoning_summary_part.added",
"item_id": self.reasoning.item_id,
"output_index": output_index,
"summary_index": 0,
"part": {
"type": "summary_text",
"text": ""
}
}),
));
events.push(sse::reasoning_item_added(output_index, &item_id));
events.push(sse::reasoning_summary_part_added(output_index, &item_id));
}
self.reasoning.text.push_str(delta);
let output_index = self.reasoning.output_index.unwrap_or(0);
events.push(sse_event(
"response.reasoning_summary_text.delta",
json!({
"type": "response.reasoning_summary_text.delta",
"item_id": self.reasoning.item_id,
"output_index": output_index,
"summary_index": 0,
"delta": delta
}),
events.push(sse::reasoning_summary_text_delta(
output_index,
&self.reasoning.item_id,
delta,
));
events
@@ -351,47 +311,16 @@ impl ChatToResponsesState {
self.text.item_id = item_id.clone();
self.text.added = true;
events.push(sse_event(
"response.output_item.added",
json!({
"type": "response.output_item.added",
"output_index": output_index,
"item": {
"id": item_id,
"type": "message",
"status": "in_progress",
"role": "assistant",
"content": []
}
}),
));
events.push(sse_event(
"response.content_part.added",
json!({
"type": "response.content_part.added",
"item_id": self.text.item_id,
"output_index": output_index,
"content_index": 0,
"part": {
"type": "output_text",
"text": "",
"annotations": []
}
}),
));
events.push(sse::message_item_added(output_index, &item_id));
events.push(sse::message_content_part_added(output_index, &item_id));
}
self.text.text.push_str(delta);
let output_index = self.text.output_index.unwrap_or(0);
events.push(sse_event(
"response.output_text.delta",
json!({
"type": "response.output_text.delta",
"item_id": self.text.item_id,
"output_index": output_index,
"content_index": 0,
"delta": delta
}),
events.push(sse::output_text_delta(
output_index,
&self.text.item_id,
delta,
));
events
@@ -485,36 +414,21 @@ impl ChatToResponsesState {
&self.tool_context,
);
events.push(sse_event(
"response.output_item.added",
json!({
"type": "response.output_item.added",
"output_index": assigned,
"item": item
}),
));
events.push(sse::output_item_added(assigned, &item));
if !pending_arguments.is_empty() && !is_custom_tool {
events.push(sse_event(
"response.function_call_arguments.delta",
json!({
"type": "response.function_call_arguments.delta",
"item_id": state.item_id,
"output_index": assigned,
"delta": pending_arguments
}),
events.push(sse::function_call_arguments_delta(
assigned,
&state.item_id,
&pending_arguments,
));
}
} else if !args_delta.is_empty() && !is_custom_tool {
if let Some(output_index) = output_index {
events.push(sse_event(
"response.function_call_arguments.delta",
json!({
"type": "response.function_call_arguments.delta",
"item_id": item_id,
"output_index": output_index,
"delta": args_delta
}),
events.push(sse::function_call_arguments_delta(
output_index,
&item_id,
&args_delta,
));
}
}
@@ -567,13 +481,7 @@ impl ChatToResponsesState {
response["incomplete_details"] = json!({ "reason": "max_output_tokens" });
}
events.push(sse_event(
"response.completed",
json!({
"type": "response.completed",
"response": response
}),
));
events.push(sse::response_completed(&response));
self.completed = true;
events
}
@@ -586,50 +494,10 @@ impl ChatToResponsesState {
let output_index = self.reasoning.output_index.unwrap_or(0);
let item_id = self.reasoning.item_id.clone();
let text = self.reasoning.text.clone();
let item = json!({
"id": item_id,
"type": "reasoning",
"summary": [{
"type": "summary_text",
"text": text
}]
});
self.output_items.push((output_index, item.clone()));
let (events, item) = sse::reasoning_close(output_index, &item_id, &text);
self.output_items.push((output_index, item));
self.reasoning.done = true;
vec![
sse_event(
"response.reasoning_summary_text.done",
json!({
"type": "response.reasoning_summary_text.done",
"item_id": self.reasoning.item_id,
"output_index": output_index,
"summary_index": 0,
"text": self.reasoning.text
}),
),
sse_event(
"response.reasoning_summary_part.done",
json!({
"type": "response.reasoning_summary_part.done",
"item_id": self.reasoning.item_id,
"output_index": output_index,
"summary_index": 0,
"part": {
"type": "summary_text",
"text": self.reasoning.text
}
}),
),
sse_event(
"response.output_item.done",
json!({
"type": "response.output_item.done",
"output_index": output_index,
"item": item
}),
),
]
events
}
fn finalize_text(&mut self) -> Vec<Bytes> {
@@ -638,54 +506,12 @@ impl ChatToResponsesState {
}
let output_index = self.text.output_index.unwrap_or(0);
let item = json!({
"id": self.text.item_id,
"type": "message",
"status": "completed",
"role": "assistant",
"content": [{
"type": "output_text",
"text": self.text.text,
"annotations": []
}]
});
self.output_items.push((output_index, item.clone()));
let item_id = self.text.item_id.clone();
let text = self.text.text.clone();
let (events, item) = sse::message_close(output_index, &item_id, &text);
self.output_items.push((output_index, item));
self.text.done = true;
vec![
sse_event(
"response.output_text.done",
json!({
"type": "response.output_text.done",
"item_id": self.text.item_id,
"output_index": output_index,
"content_index": 0,
"text": self.text.text
}),
),
sse_event(
"response.content_part.done",
json!({
"type": "response.content_part.done",
"item_id": self.text.item_id,
"output_index": output_index,
"content_index": 0,
"part": {
"type": "output_text",
"text": self.text.text,
"annotations": []
}
}),
),
sse_event(
"response.output_item.done",
json!({
"type": "response.output_item.done",
"output_index": output_index,
"item": item
}),
),
]
events
}
fn finalize_tools(&mut self) -> Vec<Bytes> {
@@ -742,14 +568,7 @@ impl ChatToResponsesState {
Some(&state.reasoning_content),
&self.tool_context,
);
add_event = Some(sse_event(
"response.output_item.added",
json!({
"type": "response.output_item.added",
"output_index": assigned,
"item": item
}),
));
add_event = Some(sse::output_item_added(assigned, &item));
}
if let Some(event) = add_event {
@@ -777,44 +596,25 @@ impl ChatToResponsesState {
if is_custom_tool {
let input = custom_tool_input_from_chat_arguments(&arguments);
if !input.is_empty() {
events.push(sse_event(
"response.custom_tool_call_input.delta",
json!({
"type": "response.custom_tool_call_input.delta",
"item_id": state.item_id,
"output_index": output_index,
"delta": input.clone()
}),
events.push(sse::custom_tool_call_input_delta(
output_index,
&state.item_id,
&input,
));
}
events.push(sse_event(
"response.custom_tool_call_input.done",
json!({
"type": "response.custom_tool_call_input.done",
"item_id": state.item_id,
"output_index": output_index,
"input": input
}),
events.push(sse::custom_tool_call_input_done(
output_index,
&state.item_id,
&input,
));
} else {
events.push(sse_event(
"response.function_call_arguments.done",
json!({
"type": "response.function_call_arguments.done",
"item_id": state.item_id,
"output_index": output_index,
"arguments": arguments
}),
events.push(sse::function_call_arguments_done(
output_index,
&state.item_id,
&arguments,
));
}
events.push(sse_event(
"response.output_item.done",
json!({
"type": "response.output_item.done",
"output_index": output_index,
"item": item
}),
));
events.push(sse::output_item_done(output_index, &item));
}
events
@@ -864,13 +664,7 @@ impl ChatToResponsesState {
let mut response = self.base_response("failed", self.completed_output_items());
response["error"] = error;
sse_event(
"response.failed",
json!({
"type": "response.failed",
"response": response
}),
)
sse::response_failed(&response)
}
}
@@ -1030,13 +824,6 @@ fn extract_chat_sse_error(value: &Value) -> (String, Option<String>) {
(message, error_type)
}
fn sse_event(event: &str, data: Value) -> Bytes {
Bytes::from(format!(
"event: {event}\ndata: {}\n\n",
serde_json::to_string(&data).unwrap_or_default()
))
}
#[cfg(test)]
mod tests {
use super::*;
File diff suppressed because it is too large Load Diff
@@ -80,7 +80,11 @@ impl CodexToolContext {
.is_some_and(|spec| matches!(&spec.kind, CodexToolKind::Custom))
}
fn chat_name_for_response_function(&self, name: &str, namespace: Option<&str>) -> String {
pub(crate) fn chat_name_for_response_function(
&self,
name: &str,
namespace: Option<&str>,
) -> String {
if let Some(namespace) = namespace.filter(|value| !value.is_empty()) {
if let Some(chat_name) = self
.namespace_name_to_chat_name