Skip to content

Every agent's model. One place. Codex on DeepSeek, Claude Code on Kimi, from the menu bar.

gateway: Factory's usage policy refusal reaches agents as a refused request, not a failed sign-in

yetone750bd1bb695d
diff --git a/internal/gateway/fallback.go b/internal/gateway/fallback.go
index 8148e9bff..80538c230 100644
--- a/internal/gateway/fallback.go
+++ b/internal/gateway/fallback.go
@@ -2907,17 +2907,23 @@ const refusedStatus = http.StatusBadRequest
// filterReasons are how a vendor says its safety filter stopped the reply:
// Anthropic's stop_reason, OpenAI's finish_reason and incomplete reason,
-// an OpenAI or Azure error's code, and Gemini's finishReason and
-// blockReason.
+// an OpenAI or Azure error's code, Factory's error code, and Gemini's
+// finishReason and blockReason.
var filterReasons = map[string]string{
"refusal": "refusal",
"content_filter": "content_filter",
"content_policy_violation": "content_filter",
- "SAFETY": "safety",
- "PROHIBITED_CONTENT": "safety",
- "BLOCKLIST": "safety",
- "SPII": "safety",
- "IMAGE_SAFETY": "safety",
+ // Factory's usage policy check, a 403 whose code droid 0.245.0 tells
+ // apart from an account's 403 the same way (the Factory plugin keeps
+ // the code on the error it hands magpie): handed on as its 403, the
+ // agent took it for a failed sign-in (sweanng on Discord, ZCode:
+ // "Provider authentication failed", reason=auth_failed status=403)
+ "content_policy_rejected": "content_policy_rejected",
+ "SAFETY": "safety",
+ "PROHIBITED_CONTENT": "safety",
+ "BLOCKLIST": "safety",
+ "SPII": "safety",
+ "IMAGE_SAFETY": "safety",
}
// refusedNote is what a refusal is noted as: the vendor's own reason.
@@ -3345,8 +3351,11 @@ func policyRefusal(raw []byte) (string, bool) {
}
note := func(reason string) (string, bool) {
if msg = strings.TrimSpace(msg); msg != "" {
- if len(msg) > 300 {
- msg = msg[:300] + "…"
+ // long enough for the vendor's own advice after its words:
+ // Factory's "start a new session…" and OpenAI's Daybreak
+ // link were cut off at 300
+ if len(msg) > 1000 {
+ msg = msg[:1000] + "…"
}
return refusedNote(reason) + " — " + msg, true
}
diff --git a/internal/gateway/refusal_forbidden_test.go b/internal/gateway/refusal_forbidden_test.go
new file mode 100644
index 000000000..7faa2285a
--- /dev/null
+++ b/internal/gateway/refusal_forbidden_test.go
@@ -0,0 +1,111 @@
+package gateway
+
+import (
+ "encoding/json"
+ "net/http"
+ "strings"
+ "testing"
+
+ "github.com/yetone/magpie/internal/provider"
+)
+
+// factoryPolicyMsg is what the Factory plugin says of Factory's usage
+// policy check refusing a request, as sweanng's ZCode showed it from magpie
+// (Discord #general, 10-29), magpie's "Factory: " left off.
+const factoryPolicyMsg = `This request was blocked by Factory's usage policy check. (upstream request id 8d1f0c2a3b4e5f60::LAX, code content_policy_rejected) — Factory's usage policy check reads the whole conversation, so this session's requests are refused again until the text that tripped it is gone: start a new session, or remove or reword what the conversation quotes (another agent's system prompt or instructions pasted in a message or a tool's output are the usual cause)`
+
+// factoryPolicy403 is that refusal as the Factory plugin answers it, in
+// the shape of the API the request was for: Anthropic's on /llm/a, OpenAI's
+// on the others, the code kept on the error.
+func factoryPolicy403(anthropic bool) string {
+ e := map[string]any{"type": "permission_error", "message": factoryPolicyMsg, "code": "content_policy_rejected"}
+ var v any = map[string]any{"error": e}
+ if anthropic {
+ v = map[string]any{"type": "error", "error": e, "request_id": "8d1f0c2a3b4e5f60::LAX"}
+ }
+ b, _ := json.Marshal(v)
+ return string(b)
+}
+
+// Factory's 403 content_policy_rejected is its usage policy check, not the
+// account: in Factory's own two shapes and in the plugin's. Its other 403s
+// (an org the user can't reach, a model the org's policy leaves out) are
+// the account's still.
+func TestFactoryPolicyRejectedIsARefusal(t *testing.T) {
+ for body, want := range map[string]bool{
+ `{"code":"content_policy_rejected","detail":"This request was blocked by Factory's usage policy check."}`: true,
+ `{"error":{"code":"content_policy_rejected","message":"This request was blocked by Factory's usage policy check."}}`: true,
+ factoryPolicy403(true): true,
+ factoryPolicy403(false): true,
+ `{"detail":"Requested active organization is not accessible by this user"}`: false,
+ `{"error":{"type":"permission_error","message":"Model not allowed by your organization's model policy"}}`: false,
+ `{"error":{"type":"permission_error","code":"org_policy_denied","message":"Model not allowed by your org"}}`: false,
+ } {
+ note, got := policyRefusal([]byte(body))
+ if got != want {
+ t.Errorf("%s: refusal %v, want %v", body, got, want)
+ }
+ if got && strings.Contains(body, "start a new session") && !strings.Contains(note, "are the usual cause") {
+ t.Errorf("the vendor's advice was cut from the note: %q", note)
+ }
+ }
+}
+
+// sweanng's ZCode, asking Factory through magpie, was handed Factory's
+// usage policy refusal as the plugin's 403 permission_error, and showed
+// "Provider authentication failed" (reason=auth_failed status=403). Each
+// agent protocol is now told it as a refusal of the request: 400, the
+// protocol's invalid-request error, Factory's words and advice kept, not
+// asked again of the same account, and the account not rested.
+func TestFactoryPolicyRefusalReachesAgentsAsARefusal(t *testing.T) {
+ asks := []struct{ name, path, body, inShape string }{
+ {"messages", "/v1/messages", `{"model":"f/m","max_tokens":64,"messages":[{"role":"user","content":"hi"}]}`, `"type":"invalid_request_error"`},
+ {"messages stream", "/v1/messages", `{"model":"f/m","max_tokens":64,"stream":true,"messages":[{"role":"user","content":"hi"}]}`, `"type":"invalid_request_error"`},
+ {"chat", "/v1/chat/completions", `{"model":"f/m","messages":[{"role":"user","content":"hi"}]}`, `"type":"invalid_request_error"`},
+ {"responses", "/v1/responses", `{"model":"f/m","input":"hi"}`, `"type":"invalid_request_error"`},
+ {"responses stream", "/v1/responses", `{"model":"f/m","stream":true,"input":"hi"}`, `"type":"invalid_request_error"`},
+ {"gemini", "/v1beta/models/f/m:generateContent", `{"contents":[{"role":"user","parts":[{"text":"hi"}]}]}`, `"status":"INVALID_ARGUMENT"`},
+ }
+ for _, up := range []provider.Protocol{provider.Anthropic, provider.Chat} {
+ for _, ask := range asks {
+ t.Run(string(up)+"/"+ask.name, func(t *testing.T) {
+ fresh(t)
+ f := &scripted{replies: []reply{{403, "application/json", factoryPolicy403(up == provider.Anthropic)}}}
+ scriptedOn(t, "f", up, f)
+ s := New()
+ code, out := sendTo(s, ask.path, ask.body)
+ if code != http.StatusBadRequest {
+ t.Fatalf("status %d, want 400: %s", code, out)
+ }
+ if !strings.Contains(out, ask.inShape) || strings.Contains(out, "permission_error") || strings.Contains(out, "PERMISSION_DENIED") {
+ t.Errorf("not the protocol's request error: %s", out)
+ }
+ if !strings.Contains(out, "content_policy_rejected") || !strings.Contains(out, "are the usual cause") {
+ t.Errorf("Factory's words or advice lost: %s", out)
+ }
+ if f.n != 1 {
+ t.Errorf("Factory asked %d times, want once", f.n)
+ }
+ r := lastRoute(s)
+ if n := len(r.Tries); n == 0 || r.Tries[n-1].Fail != failRefused || r.Tries[n-1].Rest != nil {
+ t.Errorf("tries: %+v", r.Tries)
+ }
+ if rests := liveRests(); len(rests) != 0 {
+ t.Errorf("rested for a refusal of the request: %v", rests)
+ }
+ })
+ }
+ }
+}
+
+// Any route that rests a candidate after a failure leaves it be for
+// Factory's 403 refusal, as for a 400 one.
+func TestFactoryPolicyRefusalRestsNobody(t *testing.T) {
+ fresh(t)
+ s := &Server{}
+ p := provider.Provider{ID: "f", Name: "Factory", Chat: "https://api.factory.ai/api/llm/o/v1", Key: "k"}
+ c := candidate{p: p, model: "glm-5.3", rest: p.ID}
+ if r := s.restAfter(c, 403, http.Header{}, []byte(factoryPolicy403(false))); r.Why != failRefused || r.By != "off" {
+ t.Errorf("rests as %s by %s, want %s by off", r.Why, r.By, failRefused)
+ }
+}
diff --git a/internal/gateway/routing.go b/internal/gateway/routing.go
index dd37d7c29..172048546 100644
--- a/internal/gateway/routing.go
+++ b/internal/gateway/routing.go
@@ -1042,7 +1042,10 @@ func (s *Server) restAfterMarked(c candidate, status int, header http.Header, bo
// as soon as the switch is off
return Rest{Why: why, Status: status, By: "off", Until: now}
}
- if status == http.StatusBadRequest {
+ if status == http.StatusBadRequest || status == http.StatusForbidden {
+ // a 403 too: Factory's usage policy check refuses at 403
+ // (content_policy_rejected), which failureOf alone reads as the
+ // account's
if _, ok := policyRefusal(body); ok {
// the safety filter refused the request's content (#3274:
// Codex's 400 cyber_policy), read from the vendor's error
diff --git a/internal/provider/migrate_factory.go b/internal/provider/migrate_factory.go
index d9176e82c..23e8e7214 100644
--- a/internal/provider/migrate_factory.go
+++ b/internal/provider/migrate_factory.go
@@ -14,7 +14,7 @@ import (
func init() {
movers["factory"] = &mover{
pkg: "@magpie-community/opencode-factory-auth",
- min: "0.1.39", // DeepSeek V4.1 Flash 1017, from droid 0.251.0, on its registry default upstream Fireworks (sweanng on Discord; plugins #480); 0.1.38: an account stops once its Standard window is used up when the Standard usage only option is on (plugins #349); 0.1.37: a billing header joined with instructions loses only its own line, so Claude Code auto mode keeps its policy (plugins #302); Claude Code 2.1.292's instructions re-read after compaction (changed: true), and a 403 that keeps Factory's request id (plugins #243); the update-config entry in a system prompt's configuration section, as claude-code-guide renders it with a slash, adapted the way Factory takes it (plugins #198); a model's token multiplier reaching magpie as its rate, with a running promotion's discount (plugins #158); an OpenAI model's request without temperature and top_p, which OpenAI's reasoning models turn away with 400 (sweanng424 on Discord, Nowledge Mem on gpt-6.1-sol); requests saying droid 0.242.0, the current release (sky杨 on Discord: every model 403 "Factory: Forbidden" from Claude Code); Claude Code on GLM and other /api/llm/o models without the phrases Factory refuses (Echoo on Discord); the Fast Mode models listed last, so GPT-5.3-Codex, Grok and GLM are in the first 24 again (plugins #119); droid's Fast Mode models (Opus 5.5/5/4.8, GPT-5.5/5.4/5.3-Codex), and fast asked of a base model sent to its Fast Mode model (plugins #119); Gemini 4 Argon and Gemini 3.x on Factory's Gemini route (#2080); a 413 asked once more with older images left out, then said as Factory's byte limit (#2082); Haiku 5.5 and Mistral Large 4, in droid 0.240.0's picker (JL); a session continued after Claude Code compaction, whose restored context opens with the invoked-skills reminder (plugins #66); Claude Code's streamed context_management left out, which Factory's streaming route refuses (plugins #71); Claude Code's skill list sent alone as a system message, as text blocks or a string, gets update-config's line Factory takes (plugins #72); Claude Code 2.1.280's restored-context turn after /compact (plugins #68); pi's compaction and branch summary openings reworded (#1316); hook notifications kept in Claude Code compaction context only stop changed-file quoting, summaries quoted like tool output, a cut after a hunk separator (plugins #42); Claude Code compaction/runtime metadata and quoted session summaries (#1132); pi's opening sentence cut to its first half (#952); DeepSeek V4.1 Flash, Kimi K3 and GLM-5.2 offering none, so reasoning turned off stops their thinking (#899); requests saying droid 0.233.0 (JL: every model 403); Claude Code 2.1.288's runtime context as a system message of text blocks (#658); a tool whose input_schema has anyOf/oneOf/allOf at its root (Codex's automation_update) wrapped for Claude models (#646); a failure's status and its sign-in mark as the built-in's; Claude 5's system-role metadata as Factory takes it; Claude Code's standalone model-switch system updates; tool results quoting Claude Code's fixed phrases, and Claude Code's directory updates carrying a model; the sign-in renewed ahead of time through auth.refresh; Claude Code with a global CLAUDE.md or a SessionStart hook, and Claude Desktop's model line (#634)
+ min: "0.1.40", // Factory's usage policy refusal (content_policy_rejected) is a 400 invalid_request_error with its code, not a 403 permission_error an agent reads as a failed sign-in (sweanng on Discord; plugins #494); 0.1.39: DeepSeek V4.1 Flash 1017, from droid 0.251.0, on its registry default upstream Fireworks (sweanng on Discord; plugins #480); 0.1.38: an account stops once its Standard window is used up when the Standard usage only option is on (plugins #349); 0.1.37: a billing header joined with instructions loses only its own line, so Claude Code auto mode keeps its policy (plugins #302); Claude Code 2.1.292's instructions re-read after compaction (changed: true), and a 403 that keeps Factory's request id (plugins #243); the update-config entry in a system prompt's configuration section, as claude-code-guide renders it with a slash, adapted the way Factory takes it (plugins #198); a model's token multiplier reaching magpie as its rate, with a running promotion's discount (plugins #158); an OpenAI model's request without temperature and top_p, which OpenAI's reasoning models turn away with 400 (sweanng424 on Discord, Nowledge Mem on gpt-6.1-sol); requests saying droid 0.242.0, the current release (sky杨 on Discord: every model 403 "Factory: Forbidden" from Claude Code); Claude Code on GLM and other /api/llm/o models without the phrases Factory refuses (Echoo on Discord); the Fast Mode models listed last, so GPT-5.3-Codex, Grok and GLM are in the first 24 again (plugins #119); droid's Fast Mode models (Opus 5.5/5/4.8, GPT-5.5/5.4/5.3-Codex), and fast asked of a base model sent to its Fast Mode model (plugins #119); Gemini 4 Argon and Gemini 3.x on Factory's Gemini route (#2080); a 413 asked once more with older images left out, then said as Factory's byte limit (#2082); Haiku 5.5 and Mistral Large 4, in droid 0.240.0's picker (JL); a session continued after Claude Code compaction, whose restored context opens with the invoked-skills reminder (plugins #66); Claude Code's streamed context_management left out, which Factory's streaming route refuses (plugins #71); Claude Code's skill list sent alone as a system message, as text blocks or a string, gets update-config's line Factory takes (plugins #72); Claude Code 2.1.280's restored-context turn after /compact (plugins #68); pi's compaction and branch summary openings reworded (#1316); hook notifications kept in Claude Code compaction context only stop changed-file quoting, summaries quoted like tool output, a cut after a hunk separator (plugins #42); Claude Code compaction/runtime metadata and quoted session summaries (#1132); pi's opening sentence cut to its first half (#952); DeepSeek V4.1 Flash, Kimi K3 and GLM-5.2 offering none, so reasoning turned off stops their thinking (#899); requests saying droid 0.233.0 (JL: every model 403); Claude Code 2.1.288's runtime context as a system message of text blocks (#658); a tool whose input_schema has anyOf/oneOf/allOf at its root (Codex's automation_update) wrapped for Claude models (#646); a failure's status and its sign-in mark as the built-in's; Claude 5's system-role metadata as Factory takes it; Claude Code's standalone model-switch system updates; tool results quoting Claude Code's fixed phrases, and Claude Code's directory updates carrying a model; the sign-in renewed ahead of time through auth.refresh; Claude Code with a global CLAUDE.md or a SessionStart hook, and Claude Desktop's model line (#634)
agents: []string{"factory"},
out: func() ([]Moving, error) {
ls := factoryLogins()