test: provider error classification — overflow detection and message extraction

packages/opencode/test/provider/error.test.ts

@@ -0,0 +1,262 @@
+import { describe, test, expect } from "bun:test"
+import { ProviderError } from "../../src/provider/error"
+import { APICallError } from "ai"
+
+// Helper to construct APICallError instances for testing.
+// APICallError is from the Vercel AI SDK and wraps HTTP errors from LLM providers.
+function makeAPICallError(opts: {
+  message?: string
+  statusCode?: number
+  responseBody?: string
+  isRetryable?: boolean
+  url?: string
+  responseHeaders?: Record<string, string>
+}): APICallError {
+  return new APICallError({
+    message: opts.message ?? "",
+    statusCode: opts.statusCode,
+    responseBody: opts.responseBody,
+    isRetryable: opts.isRetryable ?? false,
+    url: opts.url,
+    responseHeaders: opts.responseHeaders,
+    requestBodyValues: {},
+  })
+}
+
+// ---------------------------------------------------------------------------
+// parseStreamError — classifies SSE streaming errors from providers
+// ---------------------------------------------------------------------------
+describe("ProviderError.parseStreamError: SSE error classification", () => {
+  test("classifies context_length_exceeded as context_overflow", () => {
+    const result = ProviderError.parseStreamError({
+      type: "error",
+      error: { code: "context_length_exceeded", message: "too long" },
+    })
+    expect(result).toBeDefined()
+    expect(result!.type).toBe("context_overflow")
+  })
+
+  test("classifies usage_not_included with upgrade URL", () => {
+    const result = ProviderError.parseStreamError({
+      type: "error",
+      error: { code: "usage_not_included", message: "not available" },
+    })
+    expect(result).toBeDefined()
+    expect(result!.type).toBe("api_error")
+    if (result!.type === "api_error") {
+      expect(result!.message).toContain("chatgpt.com/explore/plus")
+    }
+  })
+
+  test("classifies invalid_prompt with passthrough message", () => {
+    const result = ProviderError.parseStreamError({
+      type: "error",
+      error: { code: "invalid_prompt", message: "Your prompt contains disallowed content" },
+    })
+    expect(result).toBeDefined()
+    expect(result!.type).toBe("api_error")
+    if (result!.type === "api_error") {
+      expect(result!.message).toBe("Your prompt contains disallowed content")
+    }
+  })
+
+  test("invalid_prompt falls back to default when message is not a string", () => {
+    const result = ProviderError.parseStreamError({
+      type: "error",
+      error: { code: "invalid_prompt", message: 42 },
+    })
+    expect(result).toBeDefined()
+    expect(result!.type).toBe("api_error")
+    if (result!.type === "api_error") {
+      expect(result!.message).toBe("Invalid prompt.")
+    }
+  })
+
+  test("returns undefined for non-error events", () => {
+    expect(ProviderError.parseStreamError({ type: "content", text: "hello" })).toBeUndefined()
+  })
+
+  test("returns undefined for unknown error codes", () => {
+    expect(
+      ProviderError.parseStreamError({
+        type: "error",
+        error: { code: "unknown_code", message: "weird" },
+      }),
+    ).toBeUndefined()
+  })
+
+  test("returns undefined for null/undefined input", () => {
+    expect(ProviderError.parseStreamError(null)).toBeUndefined()
+    expect(ProviderError.parseStreamError(undefined)).toBeUndefined()
+  })
+
+  test("parses JSON string input (AI SDK sometimes passes SSE chunks as strings)", () => {
+    const jsonStr = JSON.stringify({
+      type: "error",
+      error: { code: "context_length_exceeded" },
+    })
+    const result = ProviderError.parseStreamError(jsonStr)
+    expect(result).toBeDefined()
+    expect(result!.type).toBe("context_overflow")
+
+    // Non-JSON strings return undefined
+    expect(ProviderError.parseStreamError("not valid json")).toBeUndefined()
+  })
+})
+
+// ---------------------------------------------------------------------------
+// parseAPICallError — classifies HTTP errors from LLM provider APIs.
+// Overflow detection does NOT depend on providerID; it uses regex matching
+// on the error message. providerID only affects retry logic.
+// ---------------------------------------------------------------------------
+describe("ProviderError.parseAPICallError: overflow detection", () => {
+  test("detects 'prompt is too long' pattern (Anthropic)", () => {
+    const result = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "prompt is too long: 200000 tokens > 100000 maximum",
+        statusCode: 400,
+      }),
+    })
+    expect(result.type).toBe("context_overflow")
+  })
+
+  test("detects 'exceeds the context window' pattern (OpenAI)", () => {
+    const result = ProviderError.parseAPICallError({
+      providerID: "openai" as any,
+      error: makeAPICallError({
+        message: "This request exceeds the context window for gpt-4o",
+        statusCode: 400,
+      }),
+    })
+    expect(result.type).toBe("context_overflow")
+  })
+
+  test("detects HTTP 413 as overflow regardless of message text", () => {
+    const result = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "something completely unrelated",
+        statusCode: 413,
+      }),
+    })
+    expect(result.type).toBe("context_overflow")
+  })
+
+  test("detects '400 (no body)' and '413 (no body)' patterns (Cerebras/Mistral)", () => {
+    for (const msg of ["400 (no body)", "413 (no body)", "400 status code (no body)"]) {
+      const result = ProviderError.parseAPICallError({
+        providerID: "cerebras" as any,
+        error: makeAPICallError({ message: msg, statusCode: 400 }),
+      })
+      expect(result.type).toBe("context_overflow")
+    }
+  })
+})
+
+describe("ProviderError.parseAPICallError: error message extraction", () => {
+  test("OpenAI 404 is treated as retryable (model may be temporarily unavailable)", () => {
+    const result = ProviderError.parseAPICallError({
+      providerID: "openai" as any,
+      error: makeAPICallError({
+        message: "Model not found",
+        statusCode: 404,
+        isRetryable: false, // SDK says not retryable, but our code overrides
+      }),
+    })
+    expect(result.type).toBe("api_error")
+    if (result.type === "api_error") {
+      expect(result.isRetryable).toBe(true)
+    }
+  })
+
+  test("non-OpenAI providers pass through isRetryable from SDK", () => {
+    const retriable = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "Internal server error",
+        statusCode: 500,
+        isRetryable: true,
+      }),
+    })
+    expect(retriable.type).toBe("api_error")
+    if (retriable.type === "api_error") expect(retriable.isRetryable).toBe(true)
+
+    const nonRetriable = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "Bad request",
+        statusCode: 400,
+        isRetryable: false,
+      }),
+    })
+    expect(nonRetriable.type).toBe("api_error")
+    if (nonRetriable.type === "api_error") expect(nonRetriable.isRetryable).toBe(false)
+  })
+
+  test("HTML 403 response yields human-readable gateway message", () => {
+    // When the SDK provides a message AND the response body is HTML,
+    // the code detects the HTML and returns a friendly message instead of raw markup.
+    const result = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "Forbidden",
+        statusCode: 403,
+        responseBody: "<html><body>Forbidden</body></html>",
+      }),
+    })
+    expect(result.type).toBe("api_error")
+    if (result.type === "api_error") {
+      expect(result.message).toContain("Forbidden")
+      expect(result.message).toContain("gateway or proxy")
+    }
+  })
+
+  test("preserves URL in metadata when present", () => {
+    const result = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "Bad request",
+        statusCode: 400,
+        url: "https://api.anthropic.com/v1/messages",
+      }),
+    })
+    expect(result.type).toBe("api_error")
+    if (result.type === "api_error") {
+      expect(result.metadata?.url).toBe("https://api.anthropic.com/v1/messages")
+    }
+  })
+
+  test("falls back to HTTP status text when message is empty and no body", () => {
+    // Many providers send empty messages on rate-limiting (429);
+    // the code falls back to Node's STATUS_CODES lookup.
+    const result = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "",
+        statusCode: 429,
+      }),
+    })
+    expect(result.type).toBe("api_error")
+    if (result.type === "api_error") {
+      expect(result.message).toBe("Too Many Requests")
+    }
+  })
+
+  test("appends plain-text responseBody to message", () => {
+    // When responseBody is not JSON and not HTML, it's appended to the status message
+    const result = ProviderError.parseAPICallError({
+      providerID: "anthropic" as any,
+      error: makeAPICallError({
+        message: "Bad Request",
+        statusCode: 400,
+        responseBody: "invalid JSON in request body at position 42",
+      }),
+    })
+    expect(result.type).toBe("api_error")
+    if (result.type === "api_error") {
+      expect(result.message).toContain("Bad Request")
+      expect(result.message).toContain("invalid JSON in request body")
+    }
+  })
+})