@@ -168,66 +168,67 @@ describe("VsCodeLmHandler", () => {
168168 } ,
169169 )
170170
171- it . each ( [ "message" , "tool_result" ] as const ) (
172- "forwards image bytes in the %s path only when supported" ,
173- async ( path ) => {
171+ it . each (
172+ ( [ "message" , "tool_result" ] as const ) . flatMap ( ( path ) =>
173+ [ true , false , undefined ] . map ( ( supportsImages ) => ( { path, supportsImages } ) ) ,
174+ ) ,
175+ ) (
176+ "forwards image bytes in the $path path only when supported ($supportsImages)" ,
177+ async ( { path, supportsImages } ) => {
174178 await handler . initializeClient ( )
175- for ( const supportsImages of [ true , false , undefined ] ) {
176- const sendRequest = vi . fn < vscode . LanguageModelChat [ "sendRequest" ] > ( ) . mockResolvedValue ( {
177- stream : asyncStreamFrom ( [ new vscode . LanguageModelTextPart ( "done" ) ] ) ,
178- text : asyncStreamFrom ( [ "done" ] ) ,
179+ const sendRequest = vi . fn < vscode . LanguageModelChat [ "sendRequest" ] > ( ) . mockResolvedValue ( {
180+ stream : asyncStreamFrom ( [ new vscode . LanguageModelTextPart ( "done" ) ] ) ,
181+ text : asyncStreamFrom ( [ "done" ] ) ,
182+ } )
183+ const countTokens = vi . fn < vscode . LanguageModelChat [ "countTokens" ] > ( ) . mockResolvedValue ( 37 )
184+ handler [ "client" ] = {
185+ ...mockLanguageModelChat ,
186+ ...{ capabilities : { supportsImageToText : supportsImages } } ,
187+ sendRequest,
188+ countTokens,
189+ }
190+ const messages : Anthropic . Messages . MessageParam [ ] =
191+ path === "message"
192+ ? [ { role : "user" , content : [ { type : "text" , text : "look" } , image ] } ]
193+ : [
194+ {
195+ role : "assistant" ,
196+ content : [ { type : "tool_use" , id : "image-tool" , name : "read_file" , input : { } } ] ,
197+ } ,
198+ {
199+ role : "user" ,
200+ content : [
201+ {
202+ type : "tool_result" ,
203+ tool_use_id : "image-tool" ,
204+ content : [ { type : "text" , text : "look" } , image ] ,
205+ } ,
206+ ] ,
207+ } ,
208+ ]
209+ const chunks = await collectStream ( handler . createMessage ( "system" , messages ) )
210+ const sent = sendRequest . mock . calls [ 0 ] [ 0 ]
211+ expect ( chunks ) . toContainEqual ( { type : "usage" , inputTokens : sent . length * 37 , outputTokens : 37 } )
212+ const counted = countTokens . mock . calls . map ( ( [ content ] ) => content )
213+ if ( supportsImages ) {
214+ expect ( counted ) . toEqual ( expect . arrayContaining ( sent ) )
215+ } else {
216+ expect ( counted . every ( ( content ) => typeof content === "string" ) ) . toBe ( true )
217+ }
218+ const content = sent [ sent . length - 1 ] . content
219+ const parts =
220+ path === "tool_result" && content [ 0 ] instanceof vscode . LanguageModelToolResultPart
221+ ? content [ 0 ] . content
222+ : content
223+ expect ( parts [ 0 ] ) . toMatchObject ( { value : "look" } )
224+ if ( supportsImages ) {
225+ expect ( parts [ 1 ] ) . toMatchObject ( {
226+ mimeType : "image/png" ,
227+ data : Buffer . from ( "iVBORw==" , "base64" ) ,
179228 } )
180- const countTokens = vi . fn < vscode . LanguageModelChat [ "countTokens" ] > ( ) . mockResolvedValue ( 37 )
181- handler [ "client" ] = {
182- ...mockLanguageModelChat ,
183- ...{ capabilities : { supportsImageToText : supportsImages } } ,
184- sendRequest,
185- countTokens,
186- }
187- const messages : Anthropic . Messages . MessageParam [ ] =
188- path === "message"
189- ? [ { role : "user" , content : [ { type : "text" , text : "look" } , image ] } ]
190- : [
191- {
192- role : "assistant" ,
193- content : [ { type : "tool_use" , id : "image-tool" , name : "read_file" , input : { } } ] ,
194- } ,
195- {
196- role : "user" ,
197- content : [
198- {
199- type : "tool_result" ,
200- tool_use_id : "image-tool" ,
201- content : [ { type : "text" , text : "look" } , image ] ,
202- } ,
203- ] ,
204- } ,
205- ]
206- const chunks = await collectStream ( handler . createMessage ( "system" , messages ) )
207- const sent = sendRequest . mock . calls [ 0 ] [ 0 ]
208- expect ( chunks ) . toContainEqual ( { type : "usage" , inputTokens : sent . length * 37 , outputTokens : 37 } )
209- if ( supportsImages ) {
210- for ( const message of sent ) {
211- expect ( countTokens ) . toHaveBeenCalledWith ( message , expect . anything ( ) )
212- }
213- } else {
214- expect ( countTokens . mock . calls . every ( ( [ content ] ) => typeof content === "string" ) ) . toBe ( true )
215- }
216- const content = sent [ sent . length - 1 ] . content
217- const parts =
218- path === "tool_result" && content [ 0 ] instanceof vscode . LanguageModelToolResultPart
219- ? content [ 0 ] . content
220- : content
221- expect ( parts [ 0 ] ) . toMatchObject ( { value : "look" } )
222- if ( supportsImages ) {
223- expect ( parts [ 1 ] ) . toMatchObject ( {
224- mimeType : "image/png" ,
225- data : Buffer . from ( "iVBORw==" , "base64" ) ,
226- } )
227- } else {
228- expect ( parts [ 1 ] ) . toBeInstanceOf ( vscode . LanguageModelTextPart )
229- expect ( parts [ 1 ] ) . toMatchObject ( { value : expect . stringContaining ( "not supported" ) } )
230- }
229+ } else {
230+ expect ( parts [ 1 ] ) . toBeInstanceOf ( vscode . LanguageModelTextPart )
231+ expect ( parts [ 1 ] ) . toMatchObject ( { value : expect . stringContaining ( "not supported" ) } )
231232 }
232233 } ,
233234 )
@@ -278,12 +279,13 @@ describe("VsCodeLmHandler", () => {
278279 ...{ capabilities : fixture . capabilities } ,
279280 maxInputTokens : fixture . liveWindow ,
280281 sendRequest,
281- // Text-only models count extracted strings; every one exceeds the window on its own.
282- countTokens : vi
283- . fn < vscode . LanguageModelChat [ "countTokens" ] > ( )
284- . mockImplementation ( async ( content ) =>
285- typeof content === "string" || content . role === userRole ? fixture . inputTokens : 0 ,
286- ) ,
282+ // Image models count whole messages; text-only models count extracted strings.
283+ countTokens : vi . fn < vscode . LanguageModelChat [ "countTokens" ] > ( ) . mockImplementation ( async ( content ) => {
284+ if ( typeof content === "string" ) {
285+ return fixture . capabilities === imageCapable ? 0 : fixture . inputTokens
286+ }
287+ return content . role === userRole ? fixture . inputTokens : 0
288+ } ) ,
287289 }
288290 const windowSpy = vi . spyOn ( handler , "getCondenseContextWindow" ) . mockReturnValue ( fixture . curatedWindow )
289291 try {
0 commit comments