Do not retry an HTTP 408; lock retry cap and timeout values
CI / ci (push) Successful in 1m30s
Independent review Codex gpt-6-sol high: APPROVED (0 findings)
CI / ci (pull_request) Successful in 1m30s

The OpenAI and Anthropic adapters now report an HTTP 408 as
llm.TimeoutStatusError, which still matches llm.ErrTimeout but carries
its status. Retrying no longer retries any error carrying a 4xx status,
so a vendor's 408 is returned at once while a transport timeout, which
has no status, is still retried.

Tests: 408 (bare and wrapped) is not retried and a transport timeout is;
FromSettings integration cases for an OpenAI and Anthropic 408 and an
OpenAI transport timeout; the cap and default-timeout tests assert the
literal 20 s and 60 s instead of the constants they lock.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
2026-10-04 10:58:59 -04:00
co-authored by Claude Opus 5.5
parent d90886ddc2
commit f8e975e284
7 changed files with 133 additions and 15 deletions
+57 -1
View File
@@ -523,7 +523,8 @@ func TestConfigErrorsNeverLeakKey(t *testing.T) {
}
// scriptedTransport answers each request with the next scripted status and
// body, echoing the key in every failure body, and counts the requests.
// body, echoing the key in every failure body, and counts the requests. A
// scripted status of 0 is a transport timeout instead of a response.
type scriptedTransport struct {
statuses []int
key string
@@ -536,6 +537,9 @@ func (st *scriptedTransport) RoundTrip(r *http.Request) (*http.Response, error)
}
status := st.statuses[st.sent]
st.sent++
if status == 0 {
return nil, transportTimeout{}
}
header := http.Header{"Content-Type": {"application/json"}}
body := `{"error":{"message":"bad key ` + st.key + `"}}`
switch status {
@@ -552,6 +556,13 @@ func (st *scriptedTransport) RoundTrip(r *http.Request) (*http.Response, error)
}, nil
}
// transportTimeout is a net.Error that timed out.
type transportTimeout struct{}
func (transportTimeout) Error() string { return "i/o timeout" }
func (transportTimeout) Timeout() bool { return true }
func (transportTimeout) Temporary() bool { return true }
// useTransport makes the adapters send through rt for the rest of the test.
func useTransport(t *testing.T, rt http.RoundTripper) {
t.Helper()
@@ -604,3 +615,48 @@ func TestFromSettingsRetryErrorNeverLeaksKey(t *testing.T) {
}
}
}
// An HTTP 408 from the vendor is a 4xx answer and is not retried, though it
// matches llm.ErrTimeout.
func TestFromSettingsDoesNotRetryHTTP408(t *testing.T) {
for _, provider := range []string{"openai", "anthropic"} {
t.Run(provider, func(t *testing.T) {
st := &scriptedTransport{statuses: []int{http.StatusRequestTimeout}, key: testKey}
useTransport(t, st)
p, err := FromSettings(llm.Settings{Provider: provider, Model: "m"}, testKey)
if err != nil {
t.Fatalf("FromSettings: %v", err)
}
_, err = p.Generate(context.Background(), plainRequest())
var ts *llm.TimeoutStatusError
if !errors.Is(err, llm.ErrTimeout) || !errors.As(err, &ts) || ts.Status() != http.StatusRequestTimeout {
t.Fatalf("Generate = %v, want an HTTP 408 ErrTimeout", err)
}
if st.sent != 1 {
t.Errorf("sent %d requests, want 1", st.sent)
}
})
}
}
// A transport timeout carries no status and is retried.
func TestFromSettingsRetriesTransportTimeout(t *testing.T) {
st := &scriptedTransport{statuses: []int{0, http.StatusOK}, key: testKey}
useTransport(t, st)
p, err := FromSettings(llm.Settings{Provider: "openai", Model: "m"}, testKey)
if err != nil {
t.Fatalf("FromSettings: %v", err)
}
resp, err := p.Generate(context.Background(), plainRequest())
if err != nil {
t.Fatalf("Generate: %v", err)
}
if resp.Text != "tacos" {
t.Errorf("Text = %q, want %q", resp.Text, "tacos")
}
if st.sent != 2 {
t.Errorf("sent %d requests, want 2", st.sent)
}
}