@@ -165,11 +165,14 @@ interface FakePi {
165165 readonly deferNextState : ( ) => void ;
166166 readonly resolveDeferredState : ( data : unknown ) => Effect . Effect < void > ;
167167 readonly failNextState : ( ) => void ;
168+ readonly failNextThinking : ( ) => void ;
168169 readonly deferNextLifecycle : ( type : "switch_session" | "new_session" | "fork" | "abort" ) => void ;
169170 readonly deferNextModelSelection : ( ) => void ;
170171 readonly queueModels : ( models : ReadonlyArray < unknown > ) => void ;
171172 readonly vetoNextNewSession : ( ) => void ;
172173 readonly allRequests : ( ) => ReadonlyArray < PiRpcRecord > ;
174+ readonly promptThinkingLevels : ReadonlyArray < string > ;
175+ readonly promptModels : ReadonlyArray < string > ;
173176 readonly queueStats : ( data : unknown ) => void ;
174177 /** Pi before 0.84.4 answers clear_queue as an unknown command. */
175178 readonly setClearQueue : (
@@ -213,6 +216,7 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
213216 let deferState = false ;
214217 let deferredStateRequest : PiRpcRecord | undefined ;
215218 let failState = false ;
219+ let failThinking = false ;
216220 let vetoSwitch = false ;
217221 let vetoNewSession = false ;
218222 let deferredLifecycle : string | undefined ;
@@ -223,6 +227,10 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
223227 let isCompacting = false ;
224228 let clearQueue : Parameters < FakePi [ "setClearQueue" ] > [ 0 ] = { steering : [ ] , followUp : [ ] } ;
225229 let stdinBuffer = "" ;
230+ let thinkingLevel = "high" ;
231+ const promptThinkingLevels : Array < string > = [ ] ;
232+ const promptModels : Array < string > = [ ] ;
233+ let currentModel = "default" ;
226234
227235 const emit = ( record : PiRpcRecord , chunkBytes ?: number ) =>
228236 Effect . gen ( function * ( ) {
@@ -259,6 +267,18 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
259267 ...stateQueue . shift ( ) ,
260268 } ,
261269 } ;
270+ case "set_model" :
271+ // Pi reapplies the target model's configured thinking default.
272+ thinkingLevel = "low" ;
273+ currentModel = `${ String ( record [ "provider" ] ) } /${ String ( record [ "modelId" ] ) } ` ;
274+ return base ;
275+ case "set_thinking_level" :
276+ if ( failThinking ) {
277+ failThinking = false ;
278+ return { ...base , success : false , error : "thinking unavailable" } ;
279+ }
280+ thinkingLevel = String ( record [ "level" ] ) ;
281+ return base ;
262282 case "prompt" :
263283 if ( isCompacting ) {
264284 return { ...base , success : false , error : PI_COMPACTING_PROMPT_ERROR } ;
@@ -312,6 +332,10 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
312332 stdinBuffer = stdinBuffer . slice ( newline + 1 ) ;
313333 if ( line . length === 0 ) continue ;
314334 const record = decodeJsonLine ( line ) as PiRpcRecord ;
335+ if ( record [ "type" ] === "prompt" ) {
336+ promptThinkingLevels . push ( thinkingLevel ) ;
337+ promptModels . push ( currentModel ) ;
338+ }
315339 allRequests . push ( record ) ;
316340 yield * Queue . offer ( requests , record ) ;
317341 if ( record [ "type" ] === "get_state" && deferState ) {
@@ -366,6 +390,11 @@ const makeFakePi: Effect.Effect<FakePi> = Effect.gen(function* () {
366390
367391 return {
368392 spawner,
393+ failNextThinking : ( ) => {
394+ failThinking = true ;
395+ } ,
396+ promptThinkingLevels,
397+ promptModels,
369398 emit,
370399 takeRequest,
371400 queueEntries : ( data ) => entriesQueue . push ( data ) ,
@@ -2214,55 +2243,149 @@ describe("PiAdapterV2", () => {
22142243 } ) . pipe ( Effect . scoped , Effect . provide ( layerTest ) ) ,
22152244 ) ;
22162245
2217- it . effect ( "resets applied thinking when returning to Pi default" , ( ) =>
2218- Effect . gen ( function * ( ) {
2219- const fake = yield * makeFakePi ;
2220- fake . queueState ( {
2221- model : { provider : "xai" , id : "grok-4.6" } ,
2222- thinkingLevel : "medium" ,
2223- sessionFile : FAKE_SESSION_FILE ,
2224- sessionId : "abc" ,
2225- } ) ;
2226- const { runtime, takeEvent } = yield * openRuntime ( fake ) ;
2227- const providerThread = yield * runtime . ensureThread ( {
2228- threadId : THREAD_ID ,
2229- modelSelection : modelSelection ( "default" ) ,
2230- runtimePolicy,
2231- } ) ;
2246+ it . effect . each ( [ "anthropic/model-b" , "default" ] ) (
2247+ "preserves explicit thinking when switching models to %s in the same runtime" ,
2248+ ( targetModel ) =>
2249+ Effect . gen ( function * ( ) {
2250+ const fake = yield * makeFakePi ;
2251+ fake . queueState ( {
2252+ model : { provider : "anthropic" , id : "model-b" } ,
2253+ thinkingLevel : "high" ,
2254+ sessionFile : FAKE_SESSION_FILE ,
2255+ } ) ;
2256+ const { runtime, takeEvent } = yield * openRuntime ( fake ) ;
2257+ const providerThread = yield * runtime . ensureThread ( {
2258+ threadId : THREAD_ID ,
2259+ modelSelection : modelSelection ( "default" ) ,
2260+ runtimePolicy,
2261+ } ) ;
2262+ const selection = ( model : string ) : ModelSelection => ( {
2263+ instanceId : PI_INSTANCE_ID ,
2264+ model,
2265+ options : [ { id : "thinking" , value : "high" } ] ,
2266+ } ) ;
22322267
2233- yield * startTurn ( runtime , providerThread , "default" , [ ] , "Hello pi" , {
2234- instanceId : PI_INSTANCE_ID ,
2235- model : "xai/grok-4.6" ,
2236- options : [ { id : "thinking" , value : "high" } ] ,
2237- } ) ;
2238- const modelRequest = yield * fake . takeRequest ( "set_model" ) ;
2239- assert . equal ( modelRequest [ "provider" ] , "xai" ) ;
2240- assert . equal ( modelRequest [ "modelId" ] , "grok-4.6" ) ;
2241- const levelRequest = yield * fake . takeRequest ( "set_thinking_level" ) ;
2242- assert . equal ( levelRequest [ "level" ] , "high" ) ;
2243- yield * fake . emit ( { type : "agent_start" } ) ;
2244- yield * fake . emit ( { type : "agent_end" , messages : [ ] , willRetry : false } ) ;
2245- yield * fake . emit ( { type : "agent_settled" } ) ;
2246- yield * takeEvent ( ( event ) => event . type === "turn.terminal" ) ;
2268+ yield * startTurn (
2269+ runtime ,
2270+ providerThread ,
2271+ "default" ,
2272+ [ ] ,
2273+ "First" ,
2274+ selection ( "anthropic/model-a" ) ,
2275+ ) ;
2276+ yield * fake . takeRequest ( "prompt" ) ;
2277+ yield * fake . emit ( { type : "agent_start" } ) ;
2278+ yield * fake . emit ( { type : "agent_end" , messages : [ ] , willRetry : false } ) ;
2279+ yield * fake . emit ( { type : "agent_settled" } ) ;
2280+ yield * takeEvent ( ( event ) => event . type === "turn.terminal" ) ;
22472281
2248- yield * startTurn (
2249- runtime ,
2250- providerThread ,
2251- "default" ,
2252- [ ] ,
2253- "Hello pi" ,
2254- {
2282+ yield * startTurn (
2283+ runtime ,
2284+ providerThread ,
2285+ "default" ,
2286+ [ ] ,
2287+ "Second" ,
2288+ selection ( targetModel ) ,
2289+ 2 ,
2290+ ) ;
2291+ yield * fake . takeRequest ( "prompt" ) ;
2292+ assert . deepEqual ( fake . promptThinkingLevels , [ "high" , "high" ] ) ;
2293+ } ) . pipe ( Effect . scoped , Effect . provide ( layerTest ) ) ,
2294+ ) ;
2295+
2296+ it . effect . each ( [ false , true , "previous" ] as const ) (
2297+ "resets applied thinking when returning to Pi default (failed first restore: %s)" ,
2298+ ( failFirstRestore ) =>
2299+ Effect . gen ( function * ( ) {
2300+ const fake = yield * makeFakePi ;
2301+ fake . queueState ( {
2302+ model : { provider : "xai" , id : "grok-4.6" } ,
2303+ thinkingLevel : "medium" ,
2304+ sessionFile : FAKE_SESSION_FILE ,
2305+ sessionId : "abc" ,
2306+ } ) ;
2307+ const { runtime, takeEvent } = yield * openRuntime ( fake ) ;
2308+ const providerThread = yield * runtime . ensureThread ( {
2309+ threadId : THREAD_ID ,
2310+ modelSelection : modelSelection ( "default" ) ,
2311+ runtimePolicy,
2312+ } ) ;
2313+
2314+ yield * startTurn ( runtime , providerThread , "default" , [ ] , "Hello pi" , {
22552315 instanceId : PI_INSTANCE_ID ,
2256- model : "default" ,
2257- } ,
2258- 2 ,
2259- ) ;
2260- const replayModel = yield * fake . takeRequest ( "set_model" ) ;
2261- assert . equal ( replayModel [ "provider" ] , "xai" ) ;
2262- assert . equal ( replayModel [ "modelId" ] , "grok-4.6" ) ;
2263- const resetLevel = yield * fake . takeRequest ( "set_thinking_level" ) ;
2264- assert . equal ( resetLevel [ "level" ] , "medium" ) ;
2265- } ) . pipe ( Effect . scoped , Effect . provide ( layerTest ) ) ,
2316+ model : "anthropic/model-a" ,
2317+ options : [ { id : "thinking" , value : "high" } ] ,
2318+ } ) ;
2319+ const modelRequest = yield * fake . takeRequest ( "set_model" ) ;
2320+ assert . equal ( modelRequest [ "provider" ] , "anthropic" ) ;
2321+ assert . equal ( modelRequest [ "modelId" ] , "model-a" ) ;
2322+ const levelRequest = yield * fake . takeRequest ( "set_thinking_level" ) ;
2323+ assert . equal ( levelRequest [ "level" ] , "high" ) ;
2324+ yield * fake . takeRequest ( "prompt" ) ;
2325+ yield * fake . emit ( { type : "agent_start" } ) ;
2326+ yield * fake . emit ( { type : "agent_end" , messages : [ ] , willRetry : false } ) ;
2327+ yield * fake . emit ( { type : "agent_settled" } ) ;
2328+ yield * takeEvent ( ( event ) => event . type === "turn.terminal" ) ;
2329+
2330+ if ( failFirstRestore ) {
2331+ fake . failNextThinking ( ) ;
2332+ const result = yield * startTurn (
2333+ runtime ,
2334+ providerThread ,
2335+ "default" ,
2336+ [ ] ,
2337+ "Hello pi" ,
2338+ {
2339+ instanceId : PI_INSTANCE_ID ,
2340+ model : "default" ,
2341+ } ,
2342+ 2 ,
2343+ ) . pipe ( Effect . exit ) ;
2344+ assert . equal ( result . _tag , "Failure" ) ;
2345+ yield * fake . takeRequest ( "set_model" ) ;
2346+ yield * fake . takeRequest ( "set_thinking_level" ) ;
2347+ }
2348+
2349+ yield * startTurn (
2350+ runtime ,
2351+ providerThread ,
2352+ "default" ,
2353+ [ ] ,
2354+ "Hello pi" ,
2355+ {
2356+ instanceId : PI_INSTANCE_ID ,
2357+ model : failFirstRestore === "previous" ? "anthropic/model-a" : "default" ,
2358+ ...( failFirstRestore === "previous"
2359+ ? { options : [ { id : "thinking" , value : "high" } ] }
2360+ : { } ) ,
2361+ } ,
2362+ 2 ,
2363+ ) ;
2364+ yield * fake . takeRequest ( "prompt" ) ;
2365+ assert . deepEqual ( fake . promptThinkingLevels , [
2366+ "high" ,
2367+ failFirstRestore === "previous" ? "high" : "medium" ,
2368+ ] ) ;
2369+ assert . deepEqual ( fake . promptModels , [
2370+ "anthropic/model-a" ,
2371+ failFirstRestore === "previous" ? "anthropic/model-a" : "xai/grok-4.6" ,
2372+ ] ) ;
2373+ const replayModel = fake
2374+ . allRequests ( )
2375+ . findLast ( ( request ) => request [ "type" ] === "set_model" ) ! ;
2376+ assert . equal (
2377+ replayModel [ "provider" ] ,
2378+ failFirstRestore === "previous" ? "anthropic" : "xai" ,
2379+ ) ;
2380+ assert . equal (
2381+ replayModel [ "modelId" ] ,
2382+ failFirstRestore === "previous" ? "model-a" : "grok-4.6" ,
2383+ ) ;
2384+ const resetLevel = fake
2385+ . allRequests ( )
2386+ . findLast ( ( request ) => request [ "type" ] === "set_thinking_level" ) ! ;
2387+ assert . equal ( resetLevel [ "level" ] , failFirstRestore === "previous" ? "high" : "medium" ) ;
2388+ } ) . pipe ( Effect . scoped , Effect . provide ( layerTest ) ) ,
22662389 ) ;
22672390
22682391 it . effect . each ( [
0 commit comments