From 431670ee0028afbe6eb33ccb4ddbdbf97e8af913 Mon Sep 17 00:00:00 2001 From: Eric Liu Date: Sat, 14 Nov 2026 15:22:03 -0800 Subject: [PATCH] fix(auth-cursor): stop reporting the run's usage total on replies Cursor reports token usage once per run, and one run spans every tool round of a client turn. The plugin put that run total on the final reply as if it were this request's usage, so it counted the turn's whole context again for each round. Clients that size their context from usage read it as overflow: Hermes logged in=1,746,716 for a ~58K request on a 500K-window model and compressed its context at the start of every turn. Replies now carry no usage, the same as tool-call replies already did. LiteLLM and other OpenAI-compatible callers estimate it per request. The run total stays on the `cursor request completed` log line. Co-Authored-By: Claude Opus 5.5 --- auth-cursor/README.md | 1 + auth-cursor/go/chatcompletions.go | 11 ++++++----- auth-cursor/go/executor.go | 10 +++++----- auth-cursor/go/plugin_test.go | 8 ++++---- auth-cursor/go/tools_test.go | 12 +++++------- 5 files changed, 21 insertions(+), 21 deletions(-) diff --git a/auth-cursor/README.md b/auth-cursor/README.md index 3b8d823..dfc0cc1 100644 --- a/auth-cursor/README.md +++ b/auth-cursor/README.md @@ -185,6 +185,7 @@ Tools stay off unless the client sends them. The working directory is always emp Other limits: - No token-counting endpoint. `count_tokens` is a character estimate; billing uses counts Cursor reports after a run. +- Responses carry no `usage`. Cursor counts usage per run, and one run covers every tool round of a client turn, so the total re-counts the turn's context once per round. The run total goes only to the `cursor request completed` log line. - `executor.http_request` returns 501. - Remote images: only `http`/`https` to a non-loopback, non-private, non-metadata host. Credentials in the URL are refused. The error text never repeats the path or query. The host HTTP client still follows redirects, so a public URL that 302s onto loopback is not blocked here. - Region/plan refusals fail that one request without parking the credential. A 401 or 429 that arrives as an `SdkErrorCode` still cools or retires the key. diff --git a/auth-cursor/go/chatcompletions.go b/auth-cursor/go/chatcompletions.go index a763f0c..0d763bf 100644 --- a/auth-cursor/go/chatcompletions.go +++ b/auth-cursor/go/chatcompletions.go @@ -83,7 +83,6 @@ type chatCompletion struct { Created int64 `json:"created"` Model string `json:"model"` Choices []chatCompletionChoice `json:"choices"` - Usage *chatCompletionUsage `json:"usage,omitempty"` } func parseChatRequest(payload []byte) (chatRequest, error) { @@ -381,7 +380,11 @@ func convertUsage(usage *sdkv1.TokenUsage) *chatCompletionUsage { } } -func buildCompletion(id, model, text string, usage *sdkv1.TokenUsage) ([]byte, error) { +// Replies carry no usage. Cursor reports usage once per run, and a run spans every tool +// round of a client turn, so its total re-counts the whole turn's context on the last +// reply. Callers that size their context from usage would read that as overflow; without +// usage they estimate per request instead. +func buildCompletion(id, model, text string) ([]byte, error) { finish := "stop" return json.Marshal(chatCompletion{ ID: id, @@ -393,11 +396,10 @@ func buildCompletion(id, model, text string, usage *sdkv1.TokenUsage) ([]byte, e Message: &chatCompletionMessage{Role: "assistant", Content: text}, FinishReason: &finish, }}, - Usage: convertUsage(usage), }) } -func buildStreamChunk(id, model string, delta chatCompletionDelta, finishReason *string, usage *sdkv1.TokenUsage) []byte { +func buildStreamChunk(id, model string, delta chatCompletionDelta, finishReason *string) []byte { chunk := chatCompletion{ ID: id, Object: "chat.completion.chunk", @@ -408,7 +410,6 @@ func buildStreamChunk(id, model string, delta chatCompletionDelta, finishReason Delta: &delta, FinishReason: finishReason, }}, - Usage: convertUsage(usage), } raw, errMarshal := json.Marshal(chunk) if errMarshal != nil { diff --git a/auth-cursor/go/executor.go b/auth-cursor/go/executor.go index fa5f133..819c66b 100644 --- a/auth-cursor/go/executor.go +++ b/auth-cursor/go/executor.go @@ -52,7 +52,7 @@ func execute(raw []byte) ([]byte, error) { if len(result.toolCalls) > 0 { payload, errBody = buildToolCallCompletion(newCompletionID(), model, result.text, result.toolCalls) } else { - payload, errBody = buildCompletion(newCompletionID(), model, result.text, result.usage) + payload, errBody = buildCompletion(newCompletionID(), model, result.text) } if errBody != nil { logCtx.failed(errBody.Error()) @@ -125,7 +125,7 @@ func forwardStream(ctx context.Context, streamID, model string, framing streamFr delta.Role = "assistant" roleSent = true } - chunk := buildStreamChunk(completionID, model, delta, nil, nil) + chunk := buildStreamChunk(completionID, model, delta, nil) return emitPluginStreamChunk(streamID, framing.frame(chunk)) } @@ -145,7 +145,7 @@ func forwardStream(ctx context.Context, streamID, model string, framing streamFr return nil } finish := "stop" - final := buildStreamChunk(completionID, model, chatCompletionDelta{}, &finish, result.usage) + final := buildStreamChunk(completionID, model, chatCompletionDelta{}, &finish) if errEmit := emitPluginStreamChunk(streamID, framing.frame(final)); errEmit != nil { logCtx.failed(errEmit.Error()) return errEmit @@ -168,13 +168,13 @@ func emitToolCallStream(streamID, completionID, model string, framing streamFram delta.Role = "assistant" *roleSent = true } - chunk := buildStreamChunk(completionID, model, delta, nil, nil) + chunk := buildStreamChunk(completionID, model, delta, nil) if errEmit := emitPluginStreamChunk(streamID, framing.frame(chunk)); errEmit != nil { return errEmit } } finish := "tool_calls" - final := buildStreamChunk(completionID, model, chatCompletionDelta{}, &finish, nil) + final := buildStreamChunk(completionID, model, chatCompletionDelta{}, &finish) if errEmit := emitPluginStreamChunk(streamID, framing.frame(final)); errEmit != nil { return errEmit } diff --git a/auth-cursor/go/plugin_test.go b/auth-cursor/go/plugin_test.go index 8ba0418..09581fd 100644 --- a/auth-cursor/go/plugin_test.go +++ b/auth-cursor/go/plugin_test.go @@ -83,9 +83,9 @@ func TestExecuteAssemblesCompletion(t *testing.T) { if got := completion.Choices[0].Message.Content; got != "Hello world" { t.Errorf("content = %q, want %q", got, "Hello world") } - // Cache reads are prompt tokens served from cache, so they belong on the prompt side. - if completion.Usage == nil || completion.Usage.PromptTokens != 10 || completion.Usage.CompletionTokens != 2 { - t.Errorf("usage = %+v, want prompt=10 completion=2", completion.Usage) + // The run's usage total is not this request's usage, so the reply carries none. + if usage := gjson.GetBytes(response.Payload, "usage"); usage.Exists() { + t.Errorf("usage = %s, want none", usage.Raw) } // A non-streaming request has nothing to do with deltas, so it must not ask for them. if len(bridge.enableDeltas) != 1 || bridge.enableDeltas[0] { @@ -1047,7 +1047,7 @@ func TestExecuteReportsAnUnfetchableImageAsARequestFault(t *testing.T) { func TestBuildStreamChunkEmitsBareChunkJSON(t *testing.T) { finish := "stop" - chunk := buildStreamChunk("chatcmpl-1", "fake-model", chatCompletionDelta{Content: "Hello"}, &finish, nil) + chunk := buildStreamChunk("chatcmpl-1", "fake-model", chatCompletionDelta{Content: "Hello"}, &finish) // The chat-completions SSE writer adds the framing, so the chunk itself must stay bare. if strings.HasPrefix(string(chunk), "data:") { t.Fatalf("chunk carries SSE framing: %q", chunk) diff --git a/auth-cursor/go/tools_test.go b/auth-cursor/go/tools_test.go index 65f2867..427c3da 100644 --- a/auth-cursor/go/tools_test.go +++ b/auth-cursor/go/tools_test.go @@ -44,6 +44,10 @@ func decodeCompletion(t *testing.T, raw []byte) chatCompletion { if errUnmarshal := json.Unmarshal(env.Result, &response); errUnmarshal != nil { t.Fatalf("decode executor response: %v", errUnmarshal) } + // Cursor's usage is per run, and a run spans every tool round, so no reply carries it. + if usage := gjson.GetBytes(response.Payload, "usage"); usage.Exists() { + t.Fatalf("usage = %s, want none", usage.Raw) + } var completion chatCompletion if errUnmarshal := json.Unmarshal(response.Payload, &completion); errUnmarshal != nil { t.Fatalf("decode completion: %v", errUnmarshal) @@ -259,9 +263,6 @@ func TestExecuteToolCallRoundTrip(t *testing.T) { if finishReason(t, first) != "tool_calls" { t.Fatalf("finish_reason = %q, want tool_calls", finishReason(t, first)) } - if first.Usage != nil { - t.Fatalf("usage = %+v, want none on a tool turn", first.Usage) - } if first.Choices[0].Message == nil || len(first.Choices[0].Message.ToolCalls) != 1 { t.Fatalf("tool_calls = %+v", first.Choices[0].Message) } @@ -298,9 +299,6 @@ func TestExecuteToolCallRoundTrip(t *testing.T) { if second.Choices[0].Message.Content != "Hello world" { t.Errorf("content = %q, want Hello world", second.Choices[0].Message.Content) } - if second.Usage == nil || second.Usage.PromptTokens != 10 || second.Usage.CompletionTokens != 2 { - t.Errorf("usage = %+v, want prompt=10 completion=2", second.Usage) - } if got := bridge.closedAgents(); len(got) != 1 || got[0] != "agent-1" { t.Errorf("closed agents = %v, want [agent-1]", got) } @@ -632,7 +630,7 @@ func TestBuildStreamChunkEmitsIndexedToolCalls(t *testing.T) { Arguments: `{"city":"Paris"}`, }, }}, - }, &finish, nil) + }, &finish) if gjson.GetBytes(chunk, "choices.0.delta.tool_calls.0.index").Int() != 0 { t.Errorf("index omitted: %s", chunk) }