Use Codex CLI request shape for all ChatGPT subscription models

The ChatGPT Codex backend load-sheds requests from unrecognized
client identities with "Our servers are currently overloaded"
stream errors even when capacity is fine. Extend the existing
Codex CLI routing metadata (originator, user-agent, session-id,
prompt_cache_key) from preview models to every model on the
subscription backend. Public Responses API paths are unchanged.

commit: b58450d9a0577d71baa0007af889d6d6a5f6a8a8
author: patriceckhart · 2mo ago · 2026-07-30 10:05 UTC
parent: ad8d624
signature: unsigned
refs: tag: v0.3.28

2 files · +29 -11

M packages/provider/openai_codex.go +14 -5


@@ -45,6 +45,7 @@ type codexClient struct {
45	45		providerName string
46	46		modelName func(string) string
47	47		disableCLIRouting bool
	48	+	cliRoutingAll bool
48	49		http *http.Client
49	50		}
50	51

@@ -61,7 +62,14 @@ func NewOpenAICodex(token, accountID, baseURL string) Client {
61	62		baseURL: strings.TrimRight(baseURL, "/"),
62	63		errorLabel: "codex",
63	64		providerName: "openai-codex",
64		-	http: &http.Client{Timeout: 0},
	65	+	// The ChatGPT Codex backend load-sheds requests from client
	66	+	// identities it does not recognize: unknown originator/user-agent
	67	+	// pairs receive "Our servers are currently overloaded" stream
	68	+	// errors near-deterministically while Codex CLI requests succeed.
	69	+	// Send the Codex CLI request shape for every model on this
	70	+	// backend.
	71	+	cliRoutingAll: true,
	72	+	http: &http.Client{Timeout: 0},
65	73		}
66	74		}
67	75

@@ -363,7 +371,7 @@ func (c *codexClient) Stream(ctx context.Context, req Request) (<-chan Event, er
363	371		wire.Model = c.modelName(wire.Model)
364	372		}
365	373		var codexCLISessionID string
366		-	if !c.disableCLIRouting && usesCodexCLIRouting(wire.Model) {
	374	+	if !c.disableCLIRouting && (c.cliRoutingAll \|\| usesCodexCLIRouting(wire.Model)) {
367	375		codexCLISessionID = newCodexSessionID()
368	376		wire.PromptCacheKey = codexCLISessionID
369	377		}

@@ -383,9 +391,10 @@ func (c *codexClient) Stream(ctx context.Context, req Request) (<-chan Event, er
383	391		httpReq.Header.Set("chatgpt-account-id", c.accountID)
384	392		httpReq.Header.Set("openai-beta", "responses=experimental")
385	393		if codexCLISessionID != "" {
386		-	// Some preview models are only admitted or reliably served by the
387		-	// ChatGPT Codex backend when the request follows Codex CLI routing
388		-	// metadata. Keep this narrow so Sol retains zot's proven shape.
	394	+	// The ChatGPT Codex backend only admits (and reliably serves)
	395	+	// requests that follow Codex CLI routing metadata; other client
	396	+	// identities are load-shed with "servers are currently
	397	+	// overloaded" errors even when capacity is fine.
389	398		httpReq.Header.Set("originator", "codex_cli_rs")
390	399		httpReq.Header.Set("session-id", codexCLISessionID)
391	400		httpReq.Header.Set("user-agent", "codex_cli_rs/0.0.0")

 	providerName      string
 	modelName         func(string) string
 	disableCLIRouting bool
+	cliRoutingAll     bool
 	http              *http.Client
 }
 		baseURL:      strings.TrimRight(baseURL, "/"),
 		errorLabel:   "codex",
 		providerName: "openai-codex",
-		http:         &http.Client{Timeout: 0},
+		// The ChatGPT Codex backend load-sheds requests from client
+		// identities it does not recognize: unknown originator/user-agent
+		// pairs receive "Our servers are currently overloaded" stream
+		// errors near-deterministically while Codex CLI requests succeed.
+		// Send the Codex CLI request shape for every model on this
+		// backend.
+		cliRoutingAll: true,
+		http:          &http.Client{Timeout: 0},
 	}
 }
 		wire.Model = c.modelName(wire.Model)
 	}
 	var codexCLISessionID string
-	if !c.disableCLIRouting && usesCodexCLIRouting(wire.Model) {
+	if !c.disableCLIRouting && (c.cliRoutingAll || usesCodexCLIRouting(wire.Model)) {
 		codexCLISessionID = newCodexSessionID()
 		wire.PromptCacheKey = codexCLISessionID
 	}
 		httpReq.Header.Set("chatgpt-account-id", c.accountID)
 		httpReq.Header.Set("openai-beta", "responses=experimental")
 		if codexCLISessionID != "" {
-			// Some preview models are only admitted or reliably served by the
+			// The ChatGPT Codex backend only admits (and reliably serves)
-			// ChatGPT Codex backend when the request follows Codex CLI routing
+			// requests that follow Codex CLI routing metadata; other client
-			// metadata. Keep this narrow so Sol retains zot's proven shape.
+			// identities are load-shed with "servers are currently
+			// overloaded" errors even when capacity is fine.
 			httpReq.Header.Set("originator", "codex_cli_rs")
 			httpReq.Header.Set("session-id", codexCLISessionID)
 			httpReq.Header.Set("user-agent", "codex_cli_rs/0.0.0")

M packages/provider/openai_codex_test.go +15 -6


@@ -196,7 +196,13 @@ func TestCodexNestedStreamError(t *testing.T) {
196	196		}
197	197		}
198	198
199		-	func TestCodexSolKeepsZotShape(t *testing.T) {
	199	+	// TestCodexSubscriptionAlwaysUsesCodexCLIShape pins the ChatGPT
	200	+	// subscription backend to the Codex CLI request identity for every
	201	+	// model. The backend load-sheds unrecognized originator/user-agent
	202	+	// pairs with "Our servers are currently overloaded" stream errors
	203	+	// even when capacity is fine, so the CLI shape is required for
	204	+	// reliable service, not just for preview-model admission.
	205	+	func TestCodexSubscriptionAlwaysUsesCodexCLIShape(t *testing.T) {
200	206		c := NewOpenAICodex("token", "acct", "https://example.test/backend-api/codex/responses").(*codexClient)
201	207		var gotReq *http.Request
202	208		var body bytes.Buffer

@@ -223,13 +229,16 @@ func TestCodexSolKeepsZotShape(t *testing.T) {
223	229		if gotReq == nil {
224	230		t.Fatal("request was not sent")
225	231		}
226		-	if gotReq.Header.Get("originator") != "zot" {
	232	+	if gotReq.Header.Get("originator") != "codex_cli_rs" {
227	233		t.Fatalf("originator = %q", gotReq.Header.Get("originator"))
228	234		}
229		-	if gotReq.Header.Get("session-id") != "" {
230		-	t.Fatalf("session-id = %q", gotReq.Header.Get("session-id"))
	235	+	if gotReq.Header.Get("user-agent") != "codex_cli_rs/0.0.0" {
	236	+	t.Fatalf("user-agent = %q", gotReq.Header.Get("user-agent"))
	237	+	}
	238	+	if gotReq.Header.Get("session-id") == "" {
	239	+	t.Fatal("session-id header missing")
231	240		}
232		-	if strings.Contains(body.String(), "prompt_cache_key") {
233		-	t.Fatalf("Sol request unexpectedly included prompt_cache_key: %s", body.String())
	241	+	if !strings.Contains(body.String(), "prompt_cache_key") {
	242	+	t.Fatalf("request missing prompt_cache_key: %s", body.String())
234	243		}
235	244		}

 	}
 }
-func TestCodexSolKeepsZotShape(t *testing.T) {
+// TestCodexSubscriptionAlwaysUsesCodexCLIShape pins the ChatGPT
+// subscription backend to the Codex CLI request identity for every
+// model. The backend load-sheds unrecognized originator/user-agent
+// pairs with "Our servers are currently overloaded" stream errors
+// even when capacity is fine, so the CLI shape is required for
+// reliable service, not just for preview-model admission.
+func TestCodexSubscriptionAlwaysUsesCodexCLIShape(t *testing.T) {
 	c := NewOpenAICodex("token", "acct", "https://example.test/backend-api/codex/responses").(*codexClient)
 	var gotReq *http.Request
 	var body bytes.Buffer
 	if gotReq == nil {
 		t.Fatal("request was not sent")
 	}
-	if gotReq.Header.Get("originator") != "zot" {
+	if gotReq.Header.Get("originator") != "codex_cli_rs" {
 		t.Fatalf("originator = %q", gotReq.Header.Get("originator"))
 	}
-	if gotReq.Header.Get("session-id") != "" {
+	if gotReq.Header.Get("user-agent") != "codex_cli_rs/0.0.0" {
-		t.Fatalf("session-id = %q", gotReq.Header.Get("session-id"))
+		t.Fatalf("user-agent = %q", gotReq.Header.Get("user-agent"))
+	}
+	if gotReq.Header.Get("session-id") == "" {
+		t.Fatal("session-id header missing")
 	}
-	if strings.Contains(body.String(), "prompt_cache_key") {
+	if !strings.Contains(body.String(), "prompt_cache_key") {
-		t.Fatalf("Sol request unexpectedly included prompt_cache_key: %s", body.String())
+		t.Fatalf("request missing prompt_cache_key: %s", body.String())
 	}
 }