eve 0.83.3 → 0.84.1

package/dist/src/harness/call-executor.js

@@ -0,0 +1 @@
+import{asSchema,generateId}from"ai";import{createLogger,logError}from"#internal/logging.js";import{displayTitle}from"#protocol/session-projection/display-name.js";import{toError}from"#shared/errors.js";import{toModelSchema}from"#tools/schema.js";import{isAuthorizationSignal,modelFacingAuthorizationOutput}from"#harness/authorization.js";import{toolCallModelOutput}from"#harness/tool-call-io.js";import{resolveToolCallInputObject}from"#harness/coordination.js";import{callProgress,callSettledFrom,callStarted}from"#harness/call-facts.js";import{decideApproval,invokeTool,isRunnableTool}from"#harness/tools.js";import{isAsyncIterable}from"#shared/async-iterable.js";import{throwIfTurnAborted}from"#harness/turn-cancellation.js";import{createRuntimeToolResultFromValue}from"#harness/action-result-helpers.js";import{TOOL_EXECUTION_DENIED_MESSAGE,failedCall}from"#harness/approval-outcome.js";import{isInvalidToolCall}from"#harness/tool-call-input-errors.js";import{projectToolLabel}from"#harness/action-presentation.js";const log=createLogger(`harness.tool-loop`),APPROVED_CALL_INTERRUPTED_MESSAGE=`The turn was cancelled while this approved tool call was running. It may have partially run; it was not retried.`;async function executeToolCall(e,t,n){let{callId:r,toolName:i,input:a}=e,{definition:o}=t,s=[],c=[],l=!1,u;try{if(throwIfTurnAborted(t.abortSignal),(n===`recheck`?await asSchema(toModelSchema(o.inputSchema,`input`)).validate?.(a):void 0)?.success===!1){let e=`The approved input is no longer valid for tool "${i}". Request a new tool call and approval.`;return await t.beforeResult?.(e),await t.publish(callSettledFrom(createRuntimeToolResultFromValue({callId:r,isError:!0,output:e,toolName:i}),{})),{settled:[{part:{type:`tool-result`,toolCallId:r,toolName:i,output:{type:`error-text`,value:e}}}],toolResults:c}}let e=await decideApproval(o,{abortSignal:t.abortSignal,callId:r,input:a,approvalMemory:t.approvalMemory,recheck:n===`recheck`});if(e.awaitsPerson===!0)return{awaitsApproval:!0,remember:e.remember,settled:s,toolResults:c};if(e.denied){await t.beforeResult?.(e.reason??TOOL_EXECUTION_DENIED_MESSAGE);let n=await publishDenial({callId:r,input:a,toolName:i},{publish:t.publish,reason:e.reason});return{settled:[...s,...n.settled],toolResults:c}}let d=t.clearedBy??e.clearedBy,f,p=!1,m=``,h,g=toolTelemetry(t.telemetry,{callId:t.modelCallId??`approved:${r}`,messages:[...t.messages],toolCall:{input:a,toolCallId:r,toolName:i,type:`tool-call`}});t.beforeExecute?.(),await g.started();let _=0;try{await g.execute(async()=>{let e=performance.now();try{l=!0,await t.publish(callStarted(r,d===void 0?{}:{clearedBy:d}));let e=invokeTool(o,a,{...t.responder!==void 0&&{approval:{responder:t.responder}},abortSignal:t.abortSignal,messages:[...t.messages],toolCallId:r}),i=n===`evaluate`?t.abortSignal:void 0;if(isAsyncIterable(e))for await(let n of untilAborted(e,i)){f=n;let e=withoutSignInSecrets(n);await t.publish(callProgress(r,e,projectToolLabel(o,`delta`,a,e)))}else f=await raceAbort(Promise.resolve(e),i)}finally{_=performance.now()-e}})}catch(e){await g.ended({error:e,input:a,toolCallId:r,toolName:i,type:`tool-error`},_),throwIfTurnAborted(t.abortSignal),logError(log,`tool execution failed`,e,{toolName:i,toolCallId:r}),p=!0,h=toError(e),f=h.message,m=modelFacingError(e)}if(p||await g.ended({input:a,output:withoutSignInSecrets(f),toolCallId:r,toolName:i,type:`tool-result`},_),!p&&isAuthorizationSignal(f)){let e={callId:r,input:a,signal:f,toolName:i};return{settled:s,signIn:o.version===void 0?e:{...e,version:o.version},toolResults:c}}return c.push({input:a,output:f,toolCallId:r,toolName:i,type:`tool-result`}),u={part:{output:p?{type:`error-text`,value:m}:await toolCallModelOutput(o,f,r),toolCallId:r,toolName:i,type:`tool-result`}},s.push(u),await t.beforeResult?.(p?h:void 0),await t.publish(callSettledFrom(createRuntimeToolResultFromValue({callId:r,isError:p,output:h??f,toolName:i}),{label:p?void 0:projectToolLabel(o,`complete`,a,f)})),{settled:s,toolResults:c}}catch(e){if(t.abortSignal?.aborted===!0)return l||throwIfTurnAborted(t.abortSignal),u===void 0?{settled:[interruptedResult(r,i)],toolResults:c}:{settled:[u],toolResults:c};let n=toError(e);logError(log,`tool call failed`,e,{toolName:i,toolCallId:r}),await t.beforeResult?.(n);let a=failedCall({callId:r,toolName:i,message:n});return await t.publish(callSettledFrom(a.result)),{settled:[{part:a.part}],toolResults:c}}}async function raceAbort(e,t){if(t===void 0)return await e;e.catch(()=>{}),t.throwIfAborted();let onAbort,n=new Promise((e,n)=>{onAbort=()=>n(t.reason),t.addEventListener(`abort`,onAbort,{once:!0})});try{return await Promise.race([e,n])}finally{t.removeEventListener(`abort`,onAbort)}}async function*untilAborted(e,t){let n=e[Symbol.asyncIterator](),r=!1;try{for(;;){let e=await raceAbort(n.next(),t);if(e.done===!0){r=!0;return}yield e.value}}finally{!r&&n.return!==void 0&&await raceAbort(Promise.resolve(n.return()),t)}}function modelFacingError(e){return e==null?`unknown error`:typeof e==`string`?e:e instanceof Error?e.toString():JSON.stringify(e)}function withoutSignInSecrets(e){return isAuthorizationSignal(e)?modelFacingAuthorizationOutput(e):e}function interruptedResult(e,t){return{part:{output:{type:`error-text`,value:APPROVED_CALL_INTERRUPTED_MESSAGE},toolCallId:e,toolName:t,type:`tool-result`}}}function toolTelemetry(e,t){let n=e===void 0||e.isEnabled===!1?[]:[e.integrations??[]].flat(),r={...t,functionId:e?.functionId,recordInputs:e?.recordInputs,recordOutputs:e?.recordOutputs,toolContext:void 0},notify=async e=>{await Promise.allSettled(n.map(async t=>await e(t)))};return{started:()=>notify(e=>e.onToolExecutionStart?.(r)),execute:async e=>{let wrapped=e;for(let e of n){if(e.executeTool===void 0)continue;let n=wrapped;wrapped=()=>e.executeTool({...r,execute:n,toolCallId:t.toolCall.toolCallId})}await wrapped()},ended:(e,t)=>notify(n=>n.onToolExecutionEnd?.({...r,toolExecutionMs:t,toolOutput:e}))}}const NO_INLINE_CALLS={approvals:[],parts:[],signIns:[],toolResults:[]};async function executeInlineCalls(e){let t=e.toolCalls.filter(t=>t.providerExecuted!==!0&&!isInvalidToolCall(t)&&!e.excludedCallIds.has(t.toolCallId)&&e.tools.get(t.toolName)!==void 0),n=await Promise.allSettled(t.map(async t=>{let n=e.tools.get(t.toolName),r={callId:t.toolCallId,input:t.input,toolName:t.toolName};try{let i=isRunnableTool(n)?await executeToolCall(r,{...e,definition:n},`evaluate`):await decideDeferredCall(r,n,e);return{call:t,executed:i,approval:i.awaitsApproval===!0?{...approvalRequest(t,n),remember:i.remember}:void 0}}catch(n){throwIfTurnAborted(e.abortSignal);let i=failedCall({...r,message:toError(n)});return await e.publish(callSettledFrom(i.result)),{call:t,executed:{settled:[{part:i.part}],toolResults:[]}}}})),r=n.find(e=>e.status===`rejected`);if(r?.status===`rejected`)throw r.reason;let i=n.flatMap(e=>e.status===`fulfilled`?[e.value]:[]);return{policyApprovals:Object.fromEntries(i.flatMap(({call:e,executed:t})=>t.clearedBy===void 0?[]:[[e.toolCallId,t.clearedBy]])),approvals:i.flatMap(({approval:e})=>e===void 0?[]:[e]),parts:i.flatMap(({executed:e})=>e.settled.map(e=>e.part)),signIns:i.flatMap(({executed:e})=>e.signIn===void 0?[]:[e.signIn]),toolResults:i.flatMap(({executed:e})=>e.toolResults)}}async function decideDeferredCall(e,t,n){if(t.approval===void 0)return{settled:[],toolResults:[]};let r=await decideApproval(t,{abortSignal:n.abortSignal,approvalMemory:n.approvalMemory,callId:e.callId,input:e.input});return r.awaitsPerson===!0?{awaitsApproval:!0,remember:r.remember,settled:[],toolResults:[]}:r.denied?await publishDenial(e,{publish:n.publish,reason:r.reason}):{settled:[],toolResults:[],clearedBy:r.clearedBy}}async function publishDenial(e,t){let n={output:{reason:t.reason,type:`execution-denied`},toolCallId:e.callId,toolName:e.toolName,type:`tool-result`};return await t.publish(callSettledFrom({callId:e.callId,isError:!0,kind:`tool-result`,output:{code:`TOOL_EXECUTION_DENIED`,message:t.reason??TOOL_EXECUTION_DENIED_MESSAGE,tool:{result:`not_run`}},toolName:e.toolName},{rejected:!0})),{settled:[{part:n}],toolResults:[]}}function approvalRequest(e,t){let{toolCallId:n,toolName:r}=e,i=resolveToolCallInputObject(e.input,{callId:n,toolName:r}),a=projectToolLabel(t,`start`,i)??displayTitle(r),o=authoredApprovalPrompt(e,t);return{call:{callId:n,input:i,toolName:r,...t?.version!==void 0&&{version:t.version}},interactionId:generateId(),request:{allowFreeform:!1,display:`confirmation`,kind:`approval`,options:[{id:`approve`,label:`Approve`},{id:`deny`,label:`Deny`}],prompt:o??`Approve ${a}?`}}}function authoredApprovalPrompt(e,t){let n=t?.approval;if(n===void 0||typeof n==`function`)return;let r=typeof e.input==`object`&&e.input!==null&&!Array.isArray(e.input)?e.input:{},i=n.prompt?.({callId:e.toolCallId,input:r,toolName:e.toolName});if(i!==void 0&&typeof i!=`string`)throw Error(`Tool "${e.toolName}" approval prompt must return a string.`);return i}export{APPROVED_CALL_INTERRUPTED_MESSAGE,NO_INLINE_CALLS,executeInlineCalls,executeToolCall,publishDenial};

package/dist/src/harness/call-facts.d.ts

@@ -0,0 +1,31 @@
+import type { SessionEvent } from "#protocol/session-event.js";
+import type { Cause, ErrorInfo, JsonValue } from "#protocol/session-events/envelope.js";
+import type { FactOf } from "#protocol/session-events/facts.js";
+import type { Capability, CapabilityKind, CallOwner, ClearedBy } from "#protocol/session-events/families/call.js";
+import type { HarnessToolDefinition } from "#harness/execute-tool.js";
+import type { RuntimeActionRequest, RuntimeActionResult } from "#shared/action-types.js";
+/** What a call invokes: a tool, an agent, the skill it loads, or a task control, with its label. */
+export declare function capabilityOf(action: RuntimeActionRequest, label?: string): Capability;
+/** What a call to an entry invokes, as its input starts streaming. */
+export declare function announcedKind(name: string, definition: Pick<HarnessToolDefinition, "behavior"> | undefined): CapabilityKind;
+/** The model asked for a call. */
+export declare function callRequested(input: {
+    readonly action: RuntimeActionRequest;
+    readonly owner: CallOwner;
+    readonly label?: string;
+    readonly inputError?: ErrorInfo;
+}): FactOf<"call.requested">;
+/** A call was cleared and began running. */
+export declare function callStarted(callId: string, options?: {
+    readonly clearedBy?: ClearedBy;
+}): FactOf<"call.started">;
+/** A call settled from its runtime result: failed if the result says so, otherwise completed. */
+export declare function callSettledFrom(result: RuntimeActionResult, options?: {
+    readonly label?: string;
+    readonly rejected?: boolean;
+    readonly cause?: Cause;
+}): FactOf<"call.settled">;
+/** A call's partial output, as progress. */
+export declare function callProgress(callId: string, output: unknown, label?: string): SessionEvent;
+/** A value as the wire carries it: JSON, with anything else as its string form. */
+export declare function toJsonValue(value: unknown): JsonValue;

package/dist/src/harness/call-facts.js

@@ -0,0 +1 @@
+import{isTaskControlTool}from"#protocol/task-tools.js";import{SKILL_ENTRY_NAME}from"#protocol/catalog-tools.js";import{defined,fact}from"#harness/facts.js";import{callOutcomeOf}from"#harness/call-outcome.js";function capabilityOf(e,t){let n=e.kind===`load-skill`?{kind:`skill`,name:e.name}:e.kind===`subagent-call`?{kind:`agent`,name:e.name}:e.kind===`remote-agent-call`?{kind:`agent`,name:e.name,remote:!0}:{kind:toolKind(e.toolName),name:e.toolName};return t===void 0?n:{...n,label:t}}function announcedKind(e,t){if(e===SKILL_ENTRY_NAME)return`skill`;let n=t?.behavior?.handling;return n?.kind===`dispatch`&&n.target.kind!==`workflow-tool-call`?`agent`:toolKind(e)}function toolKind(e){return isTaskControlTool(e)?`control`:`tool`}function callRequested(e){let{action:t,inputError:n}=e;return fact(`call.requested`,{callId:t.callId,capability:capabilityOf(t,e.label),input:n===void 0?toJsonValue(t.input):void 0,inputError:n,owner:e.owner})}function callStarted(e,t={}){return fact(`call.started`,{callId:e,clearedBy:t.clearedBy})}function callSettledFrom(e,t={}){let{error:n,outcome:r}=callOutcomeOf(e),i=t.rejected===!0;return fact(`call.settled`,{callId:e.callId,cause:t.cause??(i?{policy:`approval`}:void 0),error:n,outcome:i?`rejected`:r,output:e.output===void 0?void 0:toJsonValue(e.output),label:t.label})}function callProgress(e,t,n){return{data:defined({callId:e,output:toJsonValue(t),label:n}),type:`call.progress`}}function toJsonValue(e){if(e===void 0)return null;if(e===null||typeof e==`string`||typeof e==`boolean`)return e;if(typeof e==`number`)return Number.isFinite(e)?e:String(e);if(typeof e==`bigint`||typeof e==`symbol`||typeof e==`function`)return String(e);try{return JSON.parse(JSON.stringify(e))}catch{return String(e)}}export{announcedKind,callProgress,callRequested,callSettledFrom,callStarted,capabilityOf,toJsonValue};

package/dist/src/harness/call-outcome.d.ts

@@ -0,0 +1,6 @@
+import type { ErrorInfo } from "#protocol/session-events/envelope.js";
+import type { RuntimeActionResult } from "#shared/action-types.js";
+export declare function callOutcomeOf(result: RuntimeActionResult): {
+    readonly outcome: "completed" | "failed";
+    readonly error?: ErrorInfo;
+};

package/dist/src/harness/call-outcome.js

@@ -0,0 +1 @@
+function callOutcomeOf(e){if(e.isError===!0){let t=readOutputError(e.output)??{code:`ACTION_RESULT_FAILED`,message:formatOutput(e.output)};return{error:`errorName`in e&&e.errorName!==void 0?{...t,name:e.errorName}:t,outcome:`failed`}}return{outcome:`completed`}}function readOutputError(e){let t=parseRecord(e);if(t===void 0)return;let n=typeof t.code==`string`&&t.code.length>0?t.code:void 0,r=typeof t.message==`string`&&t.message.length>0?t.message:void 0;return n===void 0||r===void 0?void 0:{code:n,message:r}}function parseRecord(e){if(typeof e==`object`&&e)return e;if(typeof e!=`string`)return;let t=e.trim();if(t.length!==0)try{let e=JSON.parse(t);return typeof e==`object`&&e?e:void 0}catch{return}}function formatOutput(e){if(typeof e==`string`)return e;let t=JSON.stringify(e);return typeof t==`string`&&t.length>0?t:`Action failed.`}export{callOutcomeOf};

package/dist/src/harness/compaction/engine.d.ts

@@ -6,6 +6,17 @@ import type { CompactionConfig } from "#harness/types.js";
  * appended since.
  */
 export declare function getInputTokenCount(messages: readonly ModelMessage[], config: CompactionConfig, requestEnvelopeTokens?: number, previousEnvelopeTokens?: number): number;
+/** The settings without the last measured prompt, which a rewritten history no longer matches. */
+export declare function withoutPromptMeasurement(config: CompactionConfig): CompactionConfig;
+/**
+ * The settings with a model-reported prompt measurement and the envelope around it. Without a
+ * usable measurement, the next count is an estimate, and no envelope baseline applies.
+ */
+export declare function withPromptMeasurement(config: CompactionConfig, measured: {
+    readonly inputTokens?: number;
+    readonly promptMessageCount?: number;
+    readonly requestEnvelopeTokens?: number;
+}): CompactionConfig;
 /**
  * Returns true when the message history and fixed compaction-prompt envelope
  * exceed the compaction threshold.

package/dist/src/harness/compaction/engine.js

@@ -1,3 +1,3 @@
-import{createFrameworkUserMessage,isFrameworkUserMessage}from"#harness/messages.js";import{COMPACTION_CHECKPOINT_MARKER,COMPACTION_PROMPT_ENVELOPE,COMPACTION_RESUMPTION_MESSAGE,TRANSCRIPT_PAYLOAD_LIMIT,clipMiddleUtf16Safe,createCompactionPrompt,serializeToolResultOutput,sliceUtf16Safe,stubContentOutputFileParts}from"#harness/compaction/prompt.js";import{CompactionOverLimitError}from"#harness/compaction/over-limit-error.js";import{estimateTokens}from"#harness/token-estimate.js";const COMPACTION_PROMPT_OVERHEAD_TOKENS=estimateTokens([{content:COMPACTION_PROMPT_ENVELOPE.system,role:`system`},createFrameworkUserMessage(`context.compaction`,COMPACTION_PROMPT_ENVELOPE.prompt)]);function getInputTokenCount(e,t,n=0,r=0){let i=t.lastKnownInputTokens,a=t.lastKnownPromptMessageCount;return i===void 0||a===void 0||!Number.isInteger(a)||a<0||a>e.length?estimateTokens(e)+n:i+estimateTokens(e.slice(a))+Math.max(0,n-r)}function shouldCompact(e,t,n=0,r=0){return e.length>0&&getInputTokenCount(e,t,n,r)+COMPACTION_PROMPT_OVERHEAD_TOKENS>t.threshold}const COMPACTION_HEURISTICS=[toolResultCapHeuristic];function toolResultCapHeuristic(e){let t=withResumptionGuard([...e.previousCheckpoint===void 0?[]:[createFrameworkUserMessage(`context.compaction`,COMPACTION_CHECKPOINT_MARKER),{content:e.previousCheckpoint,role:`assistant`}],...capToolResults(e.older),...e.recent],e.replay,e.config.threshold);return evaluateThreshold(t,e.config,`should-compact`,e.tokenEstimateAdjustment).type===`within-limit`?{messages:t,type:`within-limit`}:{type:`insufficient`}}function evaluateThreshold(e,t,n,r=0){let i=n===`should-compact`?COMPACTION_PROMPT_OVERHEAD_TOKENS:0,a=estimateTokens(e)+i+r;return{estimatedTokens:a,type:a<=t.threshold?`within-limit`:`over-limit`}}async function compactMessages(e,t,n,r=!1,i){let{conversation:a,previousCheckpoint:o}=extractPreviousCheckpoint(e),s=keptTailBudget(t),c=r?{...t,recentWindowSize:1}:t,planTail=e=>{let t=findLastRealUserIndex(e),n=e[t],r=n===void 0?void 0:fitCurrentMessage(n,s),i=r?.replay??n,a=selectRecentWindowSize(e,c,s);return i!==void 0&&t<e.length-a&&(a=selectRecentWindowSize(e,c,Math.max(0,s-estimateTokens([i])))),{currentIndex:t,fit:r,keep:a,replay:i}},l=planTail(a);if(!r&&l.fit===void 0){let{older:n,recent:r}=splitMessagesForCompaction(a,l.keep);if(n.length===0&&o===void 0)return keepNonToolResultMessages(r);let s=Math.max(0,(i??getInputTokenCount(e,t))-estimateTokens(e));for(let e of COMPACTION_HEURISTICS){let i=e({config:t,older:n,previousCheckpoint:o,recent:r,replay:l.replay,tokenEstimateAdjustment:s});if(i.type===`within-limit`)return i.messages}}let u=capToolResults(a),{currentIndex:d,fit:f,keep:p,replay:m}=planTail(u),h=d<0?void 0:u[d],g=h===void 0?void 0:omitFiles(h),_=g===void 0?void 0:fitCurrentMessage(g,1)?.replay,v=g!==void 0&&_!==void 0&&estimateTokens([_])<estimateTokens([g])?_:g;if(evaluateThreshold(withResumptionGuard(checkpointHead(``),v,1/0),t,`estimate`).type===`over-limit`)throw new CompactionOverLimitError(`The agent's system prompt, tools and memory leave only ~${t.threshold} tokens for the conversation, too few to compact it.`);let{older:y,recent:b}=splitMessagesForCompaction(u,p),x=createCompactionPrompt({messages:f===void 0?y:y.map((e,t)=>t===d?f.cut:e),previousCheckpoint:o,transcriptBudgetTokens:t.threshold}),S=await n({messages:[createFrameworkUserMessage(`context.compaction`,x.prompt)],system:x.system}),fits=e=>evaluateThreshold(e,t,`estimate`).type===`within-limit`,C=withResumptionGuard([...checkpointHead(S),...b],m,t.threshold);if(fits(C))return C;let w=keepNonToolResultMessages(b),T=withResumptionGuard([...checkpointHead(S),...w],m,t.threshold);if(fits(T))return T;let E=withResumptionGuard(checkpointHead(S),m,t.threshold);if(fits(E))return E;let withClippedSummary=e=>{let build=n=>withResumptionGuard(checkpointHead(n<MIN_CLIP_CHARS?sliceUtf16Safe(S,n):clipMiddleUtf16Safe(S,n)),e,t.threshold);if(!fits(build(0)))return;let n=0,r=S.length;for(;n<r;){let e=Math.ceil((n+r)/2);fits(build(e))?n=e:r=e-1}return build(n)},D=withClippedSummary(m);if(D!==void 0)return D;let O=t.threshold-estimateTokens(withResumptionGuard(checkpointHead(``),void 0,t.threshold)),k=withClippedSummary(g===void 0?void 0:fitCurrentMessage(g,O)?.replay??g);if(k===void 0)throw new CompactionOverLimitError(`The agent's system prompt, tools and memory leave only ~${t.threshold} tokens for the conversation, too few to compact it.`);return k}function capToolResults(e){return e.map(e=>{if(e.role!==`tool`||typeof e.content==`string`)return e;let t=!1,n=e.content.map(e=>{if(e.type!==`tool-result`)return e;let n=stubContentOutputFileParts(e.output),r=serializeToolResultOutput(n);return r.length<=TRANSCRIPT_PAYLOAD_LIMIT?n===e.output?e:(t=!0,{...e,output:n}):(t=!0,{...e,output:{type:`text`,value:`[Truncated by eve: tool result reduced during context compaction. Re-run the tool if you need the full output.]
+import{createFrameworkUserMessage,isFrameworkUserMessage}from"#harness/messages.js";import{COMPACTION_CHECKPOINT_MARKER,COMPACTION_PROMPT_ENVELOPE,COMPACTION_RESUMPTION_MESSAGE,TRANSCRIPT_PAYLOAD_LIMIT,clipMiddleUtf16Safe,createCompactionPrompt,serializeToolResultOutput,sliceUtf16Safe,stubContentOutputFileParts}from"#harness/compaction/prompt.js";import{CompactionOverLimitError}from"#harness/compaction/over-limit-error.js";import{estimateTokens}from"#harness/token-estimate.js";const COMPACTION_PROMPT_OVERHEAD_TOKENS=estimateTokens([{content:COMPACTION_PROMPT_ENVELOPE.system,role:`system`},createFrameworkUserMessage(`context.compaction`,COMPACTION_PROMPT_ENVELOPE.prompt)]);function getInputTokenCount(e,t,n=0,r=0){let i=t.lastKnownInputTokens,a=t.lastKnownPromptMessageCount;if(i===void 0||a===void 0||!Number.isInteger(a)||a<0||a>e.length)return estimateTokens(e)+n;let o=e.slice(a);return i+(o.length===0?0:estimateTokens(o))+Math.max(0,n-r)}function withoutPromptMeasurement(e){let{lastKnownInputTokens:t,lastKnownPromptMessageCount:n,...r}=e;return r}function withPromptMeasurement(e,t){let{requestEnvelopeTokens:n,...r}=withoutPromptMeasurement(e);return t.inputTokens===void 0||t.promptMessageCount===void 0?r:{...r,lastKnownInputTokens:t.inputTokens,lastKnownPromptMessageCount:t.promptMessageCount,...t.requestEnvelopeTokens!==void 0&&{requestEnvelopeTokens:t.requestEnvelopeTokens}}}function shouldCompact(e,t,n=0,r=0){return e.length>0&&getInputTokenCount(e,t,n,r)+COMPACTION_PROMPT_OVERHEAD_TOKENS>t.threshold}const COMPACTION_HEURISTICS=[toolResultCapHeuristic];function toolResultCapHeuristic(e){let t=withResumptionGuard([...e.previousCheckpoint===void 0?[]:[createFrameworkUserMessage(`context.compaction`,COMPACTION_CHECKPOINT_MARKER),{content:e.previousCheckpoint,role:`assistant`}],...capToolResults(e.older),...e.recent],e.replay,e.config.threshold);return evaluateThreshold(t,e.config,`should-compact`,e.tokenEstimateAdjustment).type===`within-limit`?{messages:t,type:`within-limit`}:{type:`insufficient`}}function evaluateThreshold(e,t,n,r=0){let i=n===`should-compact`?COMPACTION_PROMPT_OVERHEAD_TOKENS:0,a=estimateTokens(e)+i+r;return{estimatedTokens:a,type:a<=t.threshold?`within-limit`:`over-limit`}}async function compactMessages(e,t,n,r=!1,i){let{conversation:a,previousCheckpoint:o}=extractPreviousCheckpoint(e),s=keptTailBudget(t),c=r?{...t,recentWindowSize:1}:t,planTail=e=>{let t=findLastRealUserIndex(e),n=e[t],r=n===void 0?void 0:fitCurrentMessage(n,s),i=r?.replay??n,a=selectRecentWindowSize(e,c,s);return i!==void 0&&t<e.length-a&&(a=selectRecentWindowSize(e,c,Math.max(0,s-estimateTokens([i])))),{currentIndex:t,fit:r,keep:a,replay:i}},l=planTail(a);if(!r&&l.fit===void 0){let{older:n,recent:r}=splitMessagesForCompaction(a,l.keep);if(n.length===0&&o===void 0)return keepNonToolResultMessages(r);let s=Math.max(0,(i??getInputTokenCount(e,t))-estimateTokens(e));for(let e of COMPACTION_HEURISTICS){let i=e({config:t,older:n,previousCheckpoint:o,recent:r,replay:l.replay,tokenEstimateAdjustment:s});if(i.type===`within-limit`)return i.messages}}let u=capToolResults(a),{currentIndex:d,fit:f,keep:p,replay:m}=planTail(u),h=d<0?void 0:u[d],g=h===void 0?void 0:omitFiles(h),_=g===void 0?void 0:fitCurrentMessage(g,1)?.replay,v=g!==void 0&&_!==void 0&&estimateTokens([_])<estimateTokens([g])?_:g;if(evaluateThreshold(withResumptionGuard(checkpointHead(``),v,1/0),t,`estimate`).type===`over-limit`)throw new CompactionOverLimitError(`The agent's system prompt, tools and memory leave only ~${t.threshold} tokens for the conversation, too few to compact it.`);let{older:y,recent:b}=splitMessagesForCompaction(u,p),x=createCompactionPrompt({messages:f===void 0?y:y.map((e,t)=>t===d?f.cut:e),previousCheckpoint:o,transcriptBudgetTokens:t.threshold}),S=await n({messages:[createFrameworkUserMessage(`context.compaction`,x.prompt)],system:x.system}),fits=e=>evaluateThreshold(e,t,`estimate`).type===`within-limit`,C=withResumptionGuard([...checkpointHead(S),...b],m,t.threshold);if(fits(C))return C;let w=keepNonToolResultMessages(b),T=withResumptionGuard([...checkpointHead(S),...w],m,t.threshold);if(fits(T))return T;let E=withResumptionGuard(checkpointHead(S),m,t.threshold);if(fits(E))return E;let withClippedSummary=e=>{let build=n=>withResumptionGuard(checkpointHead(n<MIN_CLIP_CHARS?sliceUtf16Safe(S,n):clipMiddleUtf16Safe(S,n)),e,t.threshold);if(!fits(build(0)))return;let n=0,r=S.length;for(;n<r;){let e=Math.ceil((n+r)/2);fits(build(e))?n=e:r=e-1}return build(n)},D=withClippedSummary(m);if(D!==void 0)return D;let O=t.threshold-estimateTokens(withResumptionGuard(checkpointHead(``),void 0,t.threshold)),k=withClippedSummary(g===void 0?void 0:fitCurrentMessage(g,O)?.replay??g);if(k===void 0)throw new CompactionOverLimitError(`The agent's system prompt, tools and memory leave only ~${t.threshold} tokens for the conversation, too few to compact it.`);return k}function capToolResults(e){return e.map(e=>{if(e.role!==`tool`||typeof e.content==`string`)return e;let t=!1,n=e.content.map(e=>{if(e.type!==`tool-result`)return e;let n=stubContentOutputFileParts(e.output),r=serializeToolResultOutput(n);return r.length<=TRANSCRIPT_PAYLOAD_LIMIT?n===e.output?e:(t=!0,{...e,output:n}):(t=!0,{...e,output:{type:`text`,value:`[Truncated by eve: tool result reduced during context compaction. Re-run the tool if you need the full output.]
 
-${clipMiddleUtf16Safe(r,TRANSCRIPT_PAYLOAD_LIMIT-113)}`}})});return t?{...e,content:n}:e})}function withResumptionGuard(e,t,n){let r=e.at(-1)?.role,i=t!==void 0&&e.some(e=>e.role===`user`&&e.content===t.content);return r!==void 0&&r!==`assistant`?r===`tool`&&t!==void 0&&!i&&estimateTokens([t])<=n?[...e,t]:e:[...e,t!==void 0&&!i?t:createFrameworkUserMessage(`execution.continuation`,COMPACTION_RESUMPTION_MESSAGE)]}function fitCurrentMessage(e,t){let n=messageText(e);if(n===void 0||t<=0||estimateTokens([e])<=t)return;let note=e=>`\n\n[eve: ${e} characters from the middle of this message were cut during context compaction to fit the context window.]\n\n`,split=(t,n)=>splitMessageText(e,t,n,note(n-t)),r=Math.floor(t*4);for(;r>0;){r=Math.min(r,n.length);let e=Math.floor(r*.2),i=sliceUtf16Safe(n,e).length,a=n.length-(r-e),o=n.charCodeAt(a);if(o>=56320&&o<=57343&&(a+=1),i<a){let e=split(i,a);if(estimateTokens([e.replay])<=t)return e}r=Math.floor(r*.9)-1}return split(0,n.length)}const MIN_CLIP_CHARS=40;function checkpointHead(e){return[createFrameworkUserMessage(`context.compaction`,COMPACTION_CHECKPOINT_MARKER),{content:e,role:`assistant`}]}function omitFiles(e){if(typeof e.content==`string`)return e;let t=[],n=!1;for(let r of e.content)r.type===`text`?t.push(r):n||(n=!0,t.push({type:`text`,text:`[eve: a file in this message was omitted during context compaction: it is too large for the context window.]`}));return n?{...e,content:t}:e}function messageText(e){if(typeof e.content==`string`)return e.content;let t=e.content.flatMap(e=>e.type===`text`?[e.text]:[]);return t.length===0?void 0:t.join(``)}function splitMessageText(e,t,n,r){if(typeof e.content==`string`)return{cut:{...e,content:e.content.slice(t,n)},replay:{...e,content:`${e.content.slice(0,t)}${r}${e.content.slice(n)}`}};let i=0,a=!1,o=[],s=[];for(let c of e.content){if(c.type!==`text`){o.push(c),s.push(c);continue}let e=Math.max(0,Math.min(c.text.length,t-i)),l=Math.max(e,Math.min(c.text.length,n-i));i+=c.text.length,l>e&&o.push({...c,text:c.text.slice(e,l)});let u=c.text.slice(0,e),d=c.text.slice(l),f=l>e&&!a;f&&(a=!0);let p=`${u}${f?r:``}${d}`;p.length>0&&s.push({...c,text:p})}return a||s.push({type:`text`,text:r}),{cut:{...e,content:o},replay:{...e,content:s}}}function findLastRealUserIndex(e){for(let t=e.length-1;t>=0;--t){let n=e[t];if(n?.role===`user`&&!isFrameworkUserMessage(n))return t}return-1}function extractPreviousCheckpoint(e){let t=e[0],n=e[1];return t?.role!==`user`||t.content!==COMPACTION_CHECKPOINT_MARKER||n?.role!==`assistant`?{conversation:[...e],previousCheckpoint:void 0}:{conversation:e.slice(2),previousCheckpoint:assistantMessageText(n)}}function keepNonToolResultMessages(e){let t=[];for(let n of e)if(n.role!==`tool`){if(n.role===`assistant`){let e=assistantMessageText(n);e.length>0&&t.push({content:e,role:`assistant`});continue}t.push(n)}return t}function assistantMessageText(e){return typeof e.content==`string`?e.content.trim():e.content.filter(e=>e.type===`text`).map(e=>e.text).join(``).trim()}function keptTailBudget(e){return Math.min(2e4,Math.floor(e.threshold/2))}function selectRecentWindowSize(e,t,n){let r=Math.min(t.recentWindowSize,Math.max(e.length-1,0)),i=0,a=0;for(let t=e.length-1;t>=0&&i<r;--t){let r=e[t];if(r===void 0)continue;let o=estimateTokens([r]);if(a+o>n)break;a+=o,i+=1}return i}function splitMessagesForCompaction(e,t){if(t<=0)return{older:[...e],recent:[]};let n=e.length-t;for(;n<e.length&&e[n]?.role===`tool`;)n+=1;return{older:e.slice(0,n),recent:e.slice(n)}}export{compactMessages,getInputTokenCount,shouldCompact};
+${clipMiddleUtf16Safe(r,TRANSCRIPT_PAYLOAD_LIMIT-113)}`}})});return t?{...e,content:n}:e})}function withResumptionGuard(e,t,n){let r=e.at(-1)?.role,i=t!==void 0&&e.some(e=>e.role===`user`&&e.content===t.content);return r!==void 0&&r!==`assistant`?r===`tool`&&t!==void 0&&!i&&estimateTokens([t])<=n?[...e,t]:e:[...e,t!==void 0&&!i?t:createFrameworkUserMessage(`execution.continuation`,COMPACTION_RESUMPTION_MESSAGE)]}function fitCurrentMessage(e,t){let n=messageText(e);if(n===void 0||t<=0||estimateTokens([e])<=t)return;let note=e=>`\n\n[eve: ${e} characters from the middle of this message were cut during context compaction to fit the context window.]\n\n`,split=(t,n)=>splitMessageText(e,t,n,note(n-t)),r=Math.floor(t*4);for(;r>0;){r=Math.min(r,n.length);let e=Math.floor(r*.2),i=sliceUtf16Safe(n,e).length,a=n.length-(r-e),o=n.charCodeAt(a);if(o>=56320&&o<=57343&&(a+=1),i<a){let e=split(i,a);if(estimateTokens([e.replay])<=t)return e}r=Math.floor(r*.9)-1}return split(0,n.length)}const MIN_CLIP_CHARS=40;function checkpointHead(e){return[createFrameworkUserMessage(`context.compaction`,COMPACTION_CHECKPOINT_MARKER),{content:e,role:`assistant`}]}function omitFiles(e){if(typeof e.content==`string`)return e;let t=[],n=!1;for(let r of e.content)r.type===`text`?t.push(r):n||(n=!0,t.push({type:`text`,text:`[eve: a file in this message was omitted during context compaction: it is too large for the context window.]`}));return n?{...e,content:t}:e}function messageText(e){if(typeof e.content==`string`)return e.content;let t=e.content.flatMap(e=>e.type===`text`?[e.text]:[]);return t.length===0?void 0:t.join(``)}function splitMessageText(e,t,n,r){if(typeof e.content==`string`)return{cut:{...e,content:e.content.slice(t,n)},replay:{...e,content:`${e.content.slice(0,t)}${r}${e.content.slice(n)}`}};let i=0,a=!1,o=[],s=[];for(let c of e.content){if(c.type!==`text`){o.push(c),s.push(c);continue}let e=Math.max(0,Math.min(c.text.length,t-i)),l=Math.max(e,Math.min(c.text.length,n-i));i+=c.text.length,l>e&&o.push({...c,text:c.text.slice(e,l)});let u=c.text.slice(0,e),d=c.text.slice(l),f=l>e&&!a;f&&(a=!0);let p=`${u}${f?r:``}${d}`;p.length>0&&s.push({...c,text:p})}return a||s.push({type:`text`,text:r}),{cut:{...e,content:o},replay:{...e,content:s}}}function findLastRealUserIndex(e){for(let t=e.length-1;t>=0;--t){let n=e[t];if(n?.role===`user`&&!isFrameworkUserMessage(n))return t}return-1}function extractPreviousCheckpoint(e){let t=e[0],n=e[1];return t?.role!==`user`||t.content!==COMPACTION_CHECKPOINT_MARKER||n?.role!==`assistant`?{conversation:[...e],previousCheckpoint:void 0}:{conversation:e.slice(2),previousCheckpoint:assistantMessageText(n)}}function keepNonToolResultMessages(e){let t=[];for(let n of e)if(n.role!==`tool`){if(n.role===`assistant`){let e=assistantMessageText(n);e.length>0&&t.push({content:e,role:`assistant`});continue}t.push(n)}return t}function assistantMessageText(e){return typeof e.content==`string`?e.content.trim():e.content.filter(e=>e.type===`text`).map(e=>e.text).join(``).trim()}function keptTailBudget(e){return Math.min(2e4,Math.floor(e.threshold/2))}function selectRecentWindowSize(e,t,n){let r=Math.min(t.recentWindowSize,Math.max(e.length-1,0)),i=0,a=0;for(let t=e.length-1;t>=0&&i<r;--t){let r=e[t];if(r===void 0)continue;let o=estimateTokens([r]);if(a+o>n)break;a+=o,i+=1}return i}function splitMessagesForCompaction(e,t){if(t<=0)return{older:[...e],recent:[]};let n=e.length-t;for(;n<e.length&&e[n]?.role===`tool`;)n+=1;return{older:e.slice(0,n),recent:e.slice(n)}}export{compactMessages,getInputTokenCount,shouldCompact,withPromptMeasurement,withoutPromptMeasurement};

package/dist/src/harness/compaction/step.d.ts

@@ -1,11 +1,12 @@
 import { type LanguageModel, type ModelMessage, type TelemetryOptions } from "ai";
 import { type HarnessSession, type StepResult, type ToolLoopHarnessConfig } from "#harness/types.js";
 import { type HarnessModelMessage } from "#harness/messages.js";
-import type { HistoryViewProjector } from "#shared/history-view.js";
 import type { Publish } from "#harness/session-machine/commit.js";
 import type { SessionAuthContext } from "#channel/types.js";
 import type { Step } from "#harness/step/context.js";
-import { type TurnPosition } from "#harness/session-machine/view.js";
+import type { TurnPosition } from "#harness/session-machine/view.js";
+import type { Cause } from "#protocol/session-events/envelope.js";
+import type { SessionView } from "#protocol/session-projection/tables.js";
 /**
  * Throws when the fixed part of the request (instructions, tools, memory)
  * leaves no room for any history: compacting the history cannot help then.
@@ -20,7 +21,10 @@ export declare function assertFixedPartFits(input: {
  * park, continue the tool loop, or terminate.
  */
 /** What a step needs from the machine after its model call. */
-/** `session.compact()`: summarizes the history now, then the session waits. */
+/**
+ * `session.compact()`: summarizes the history now, then the session waits. The control is a
+ * delivery: it's admitted with the change, and applied once the change completes.
+ */
 export declare function compactHistory(step: Step): Promise<StepResult>;
 export declare function replaceSessionHistory(session: HarnessSession, history: HarnessModelMessage[]): HarnessSession;
 /**
@@ -32,17 +36,31 @@ export declare function replaceSessionHistory(session: HarnessSession, history:
  * the compacted messages flow through the same `messages` variable the
  * harness uses to rebuild `session.history` after the step.
  */
-export declare function maybeCompact(input: {
+export declare function maybeCompact(input: Parameters<typeof compactOnce>[0]): ReturnType<typeof compactOnce>;
+declare function compactOnce(input: {
     readonly abortSignal?: AbortSignal;
     readonly auth: SessionAuthContext | null;
     /** A manual compaction runs between turns, so no turn's usage reports its summary calls. */
     readonly betweenTurns?: boolean;
+    /**
+     * The change a compaction makes, and the run that writes its summary. `announced` when the
+     * caller already published the change's start and the run's request and start.
+     */
+    readonly change: {
+        readonly changeId: string;
+        readonly summaryRunId: string;
+        readonly announced?: boolean;
+        readonly turnId?: string;
+        readonly cause?: Cause;
+    };
+    /** The step's context, where a memory capture before the compaction reports failing. */
+    readonly ctx?: object;
     readonly emissionState: TurnPosition;
     readonly force?: boolean;
-    readonly historyProjector?: HistoryViewProjector;
     readonly messages: HarnessModelMessage[];
     readonly model: LanguageModel;
     readonly publish: Publish;
+    readonly view: () => SessionView;
     /** Model-visible prompt used only to decide whether durable history needs compaction. */
     readonly promptMessages?: readonly HarnessModelMessage[];
     readonly requestEnvelopeTokens?: number;
@@ -50,6 +68,7 @@ export declare function maybeCompact(input: {
     readonly runtimeIdentity?: ToolLoopHarnessConfig["runtimeIdentity"];
     readonly session: HarnessSession;
     readonly telemetry?: TelemetryOptions;
+    readonly prepareTelemetry?: (runId: string) => TelemetryOptions | undefined;
     readonly traceContextHeaders?: () => Record<string, string> | undefined;
 }): Promise<{
     readonly compacted: boolean;
@@ -66,3 +85,4 @@ export declare function maybeCompact(input: {
  * step.
  */
 export declare function compactedTurnInput<Message extends ModelMessage>(messages: readonly Message[]): Message | undefined;
+export {};

package/dist/src/harness/compaction/step.js

@@ -1 +1 @@
-import{streamText}from"ai";import{createLogger,logError}from"#internal/logging.js";import{contextStorage}from"#context/container.js";import{addStepUsage}from"#harness/session-machine/current.js";import{AuthKey,HistoryStateKey}from"#context/keys.js";import{createCompactionCompletedEvent,createCompactionRequestedEvent}from"#protocol/message.js";import{isUserModelMessage,validateHarnessModelMessages}from"#harness/messages.js";import{drainMemoryCommit,prepareMemoryCompaction}from"#context/memory-lifecycle.js";import{canonicalizeMemoryRecords,shouldCanonicalizeMemory}from"#shared/memory-state.js";import{activeTurnId}from"#harness/session-machine/view.js";import{addUsageOutsideTurns}from"#harness/turn-tag-state.js";import{idle}from"#harness/session-machine/transitions.js";import{stepStartedForResolvers}from"#harness/session-machine/resolver-events.js";import{CompactionOverLimitError}from"#harness/compaction/over-limit-error.js";import{estimateTokens}from"#harness/token-estimate.js";import{requireSessionModelReference}from"#harness/types.js";import{buildGatewayAttributionHeaders,resolveEffectiveRuntimeModel,withGatewayTraceContext}from"#harness/model-call/model.js";import{compactMessages,getInputTokenCount,shouldCompact}from"#harness/compaction/engine.js";import{normalizeModelStreamError}from"#harness/model-call/errors.js";import{runModelCallWithRetries}from"#harness/model-call/retry.js";import{extractGatewayCostUsd,extractTokenUsageDelta,gatewayModelId}from"#harness/model-call/usage.js";import{resolveModelProfile}from"#harness/model-profile.js";import{getRequestEnvelopeTokens}from"#harness/request-envelope.js";import{resolveCallProviderOptions}from"#harness/provider-safety.js";const log=createLogger(`harness.tool-loop`);function assertFixedPartFits(e){let t=e.requestEnvelopeTokens+e.protectedTokens;if(t>=e.threshold)throw new CompactionOverLimitError(`The agent's system prompt, tools, memory and context take ~${Math.round(t)} tokens, over this model's ~${e.threshold}-token compaction limit. Shorten the instructions, remove tools or trim memory.`)}async function compactHistory(e){let{config:t}=e;if(e.session.history.length>0)try{if(e.ctx!==void 0&&t.dispatchDynamicModelEvent!==void 0){let n=e.position();await t.dispatchDynamicModelEvent({ctx:e.ctx,event:stepStartedForResolvers({modelId:e.session.agent.modelReference?.id??`dynamic`,sequence:n.sequence,stepIndex:n.stepIndex,turnId:activeTurnId(n)}),messages:validateHarnessModelMessages(e.projectHistory(e.session.history))})}let n=await resolveEffectiveRuntimeModel({config:t,ctx:e.ctx,session:e.session});e.session=n.session;let r=e.position(),i=await maybeCompact({abortSignal:t.abortSignal,auth:e.ctx?.get(AuthKey)??null,betweenTurns:!0,emissionState:{...r,turnId:activeTurnId(r)},force:!0,historyProjector:t.historyProjector,messages:[...e.session.history],model:n.model,publish:e.publish,requestEnvelopeTokens:getRequestEnvelopeTokens(e.session),resolveModel:t.resolveModel,runtimeIdentity:t.runtimeIdentity,session:e.session,telemetry:e.instrumentation?.telemetry(),traceContextHeaders:e.instrumentation?.gatewayTraceContextHeaders});if(e.session=i.session,i.failure!==void 0)throw i.failure.error}catch(t){logError(log,`manual session compaction failed`,t,{sessionId:e.session.sessionId})}return await e.apply(idle(e.view())),{next:null,session:e.session}}function replaceSessionHistory(e,t){return contextStorage.getStore()?.delete(HistoryStateKey),{...e,history:t,compaction:{recentWindowSize:e.compaction.recentWindowSize,threshold:e.compaction.threshold,thresholdPercent:e.compaction.thresholdPercent}}}async function maybeCompact(e){let{emissionState:t,publish:n}=e,r=e.messages,i=e.session,a=e.promptMessages??r,o=validateHarnessModelMessages(e.historyProjector?.({messages:a,state:i.state})??a),s=e.force===!0||shouldCompact(o,i.compaction,e.requestEnvelopeTokens,getRequestEnvelopeTokens(i)),c=shouldCanonicalizeMemory(r);if(!s&&!c)return{compacted:!1,messages:r,session:i};let l=requireSessionModelReference(i),u=i.agent.compactionModelReference??l,d=u===l?e.model:await e.resolveModel(u),f=resolveModelProfile(d),p=resolveCallProviderOptions({auth:e.auth,profile:f,providerOptions:u.providerOptions,session:i}),m={abortSignal:e.abortSignal,headers:buildGatewayAttributionHeaders(f,e.runtimeIdentity),model:withGatewayTraceContext(d,f,e.traceContextHeaders),providerOptions:p,telemetry:e.telemetry&&{...e.telemetry,functionId:`eve.compaction`}},h=[],summarize=async n=>{let r=await runModelCallWithRetries(()=>streamCompactionSummary({...m,...n}),{sessionId:i.sessionId,turnId:t.turnId},e.abortSignal);if(h.push(r.usage),r.text.trim().length===0)throw Error(`The compaction model returned an empty summary. Finish reason: ${r.finishReason}.`);return r.text};{let a=contextStorage.getStore();a!==void 0&&prepareMemoryCompaction(a,{history:r,state:i.state}),await n(createCompactionRequestedEvent({modelId:gatewayModelId(d)??`unknown`,sequence:t.sequence,sessionId:i.sessionId,stepIndex:t.stepIndex,turnId:t.turnId,usageInputTokens:getInputTokenCount(o,i.compaction,e.requestEnvelopeTokens,getRequestEnvelopeTokens(i))}),o)}let g=canonicalizeMemoryRecords(r),_=validateHarnessModelMessages(e.historyProjector?.({messages:g.ordinary,state:i.state})??g.ordinary),v=e.requestEnvelopeTokens??0,y=Math.max(0,estimateTokens(o)-estimateTokens(_)),b={...i.compaction,threshold:Math.max(1,i.compaction.threshold-v-y),lastKnownInputTokens:i.compaction.lastKnownInputTokens===void 0?void 0:Math.max(0,i.compaction.lastKnownInputTokens-Math.min(getRequestEnvelopeTokens(i)??0,v))},x=[..._],S;try{s&&(assertFixedPartFits({protectedTokens:y,requestEnvelopeTokens:v,threshold:i.compaction.threshold}),x=await compactMessages([..._],b,summarize,e.force===!0,e.requestEnvelopeTokens===void 0?void 0:Math.max(0,getInputTokenCount(o,i.compaction,v,getRequestEnvelopeTokens(i))-v-y)))}catch(e){S={error:e}}for(let n of h)i=e.betweenTurns===!0?addUsageOutsideTurns(i,n):addStepUsage(i,t.turnId,n);if(S!==void 0)return{compacted:!1,failure:S,messages:e.messages,session:i};r=validateHarnessModelMessages([...g.memory,...x]);{let a=contextStorage.getStore();if(a!==void 0&&prepareMemoryCompaction(a,{history:r,state:i.state}),await n(createCompactionCompletedEvent({modelId:gatewayModelId(d)??`unknown`,sequence:t.sequence,sessionId:i.sessionId,stepIndex:t.stepIndex,turnId:t.turnId}),e.historyProjector?.({messages:r,state:i.state})??r),a!==void 0){let t=drainMemoryCommit(a);t!==void 0&&(r=validateHarnessModelMessages(placeRecalledMemory(r,t.recalledMessages,e.betweenTurns===!0)),i={...i,state:t.state})}}return{compacted:!0,messages:r,session:replaceSessionHistory(i,r)}}function compactedTurnInput(e){return e.findLast(e=>isUserModelMessage(e)&&e.kind===`user`)}function placeRecalledMemory(e,t,n){let r=n?void 0:compactedTurnInput(e);if(r===void 0)return[...e,...t];let i=e.indexOf(r);return[...e.slice(0,i),...t,...e.slice(i)]}async function streamCompactionSummary(e){let t=streamText({...e,onError:()=>{}});for await(let e of t.fullStream)if(e.type===`error`)throw normalizeModelStreamError(e.error);let[n,r,i,a]=await Promise.all([t.text,t.finishReason,t.usage,t.providerMetadata]);return{finishReason:r,text:n,usage:extractTokenUsageDelta({costUsd:extractGatewayCostUsd(a),usage:i})}}export{assertFixedPartFits,compactHistory,compactedTurnInput,maybeCompact,replaceSessionHistory};
+import{openWork}from"#protocol/session-projection/selectors.js";import{streamText}from"ai";import{createLogger,logError}from"#internal/logging.js";import{contextStorage}from"#context/container.js";import{AuthKey,HistoryStateKey}from"#context/keys.js";import{toErrorMessage}from"#shared/errors.js";import{takeCompactionCaptureFailure}from"#reactions/kinds/memory.js";import{addTokenUsage}from"#shared/token-usage-math.js";import{isUserModelMessage,validateHarnessModelMessages}from"#harness/messages.js";import{controlDeliveryFor}from"#harness/session-machine/transitions.js";import{contextStarted,controlAdmitted,deliveryApplied,fact,runRequested,runSettledFacts,runStarted,usageOf}from"#harness/facts.js";import{NO_USAGE,markUsageReported}from"#harness/session-usage.js";import{nextChangeId,nextRunId}from"#protocol/session-projection.js";import{CompactionOverLimitError}from"#harness/compaction/over-limit-error.js";import{estimateTokens}from"#harness/token-estimate.js";import{compactMessages,getInputTokenCount,shouldCompact,withoutPromptMeasurement}from"#harness/compaction/engine.js";import{requireSessionModelReference}from"#harness/types.js";import{buildGatewayAttributionHeaders,resolveEffectiveRuntimeModel,withGatewayTraceContext}from"#harness/model-call/model.js";import{normalizeModelStreamError}from"#harness/model-call/errors.js";import{runModelCallWithRetries}from"#harness/model-call/retry.js";import{extractGatewayCostUsd,extractTokenUsageDelta,gatewayModelId}from"#harness/model-call/usage.js";import{resolveModelProfile}from"#harness/model-profile.js";import{resolveCallProviderOptions}from"#harness/provider-safety.js";const log=createLogger(`harness.tool-loop`);function assertFixedPartFits(e){let t=e.requestEnvelopeTokens+e.protectedTokens;if(t>=e.threshold)throw new CompactionOverLimitError(`The agent's system prompt, tools, memory and context take ~${Math.round(t)} tokens, over this model's ~${e.threshold}-token compaction limit. Shorten the instructions, remove tools or trim memory.`)}async function compactHistory(e){let{config:t}=e,{projection:n}=e.view(),r=controlDeliveryFor(e.view(),t.controlDelivery),{deliveryId:i}=r,a=nextChangeId(n);if(e.session.history.length===0)return await e.apply([controlAdmitted(r,`compact`),deliveryApplied(i)]),{next:null,session:e.session};let o=validateHarnessModelMessages(e.session.history);await e.apply([controlAdmitted(r,`compact`),contextStarted({cause:{deliveryId:i},changeId:a,kind:`compaction`})],o);let s=nextRunId(e.view().projection),c=`applied`;try{await e.apply([runRequested({changeId:a},s)],o);let n=await resolveEffectiveRuntimeModel({config:t,ctx:e.ctx,session:e.session});e.session=n.session;let r=e.session.agent.compactionModelReference??requireSessionModelReference(e.session);await e.publish(runStarted({attempt:0,modelId:r.id,runId:s}));let i=await maybeCompact({abortSignal:t.abortSignal,auth:e.ctx?.get(AuthKey)??null,betweenTurns:!0,change:{announced:!0,changeId:a,summaryRunId:s},ctx:e.ctx,emissionState:e.position(),force:!0,messages:[...e.session.history],model:n.model,publish:e.publish,view:()=>e.view().projection.view,requestEnvelopeTokens:e.session.compaction.requestEnvelopeTokens,resolveModel:t.resolveModel,runtimeIdentity:t.runtimeIdentity,session:e.session,prepareTelemetry:t=>e.instrumentation?.prepareAttempt({runId:t,attempt:0}).telemetry,traceContextHeaders:e.instrumentation?.gatewayTraceContextHeaders});if(e.session=i.session,i.failure!==void 0)throw i.failure.error}catch(t){c=`failed`;let n=summaryFailed(e.view().projection.view,{changeId:a,error:t});n.length>0&&await e.publish(n),logError(log,`manual session compaction failed`,t,{sessionId:e.session.sessionId})}let l=c===`failed`?`compaction-failed`:void 0;return await e.apply([fact(`delivery.settled`,{deliveryId:i,outcome:c,reason:l})]),{next:null,session:e.session}}function summaryUsageOf(e){let t=e.filter(e=>e!==void 0);if(t.length!==0)return t.reduce((e,t)=>addTokenUsage(e,usageOf(t)),NO_USAGE)}function summaryFailed(e,t){if(e.changes[t.changeId]?.status!==`running`)return[];let n={code:`COMPACTION_FAILED`,message:toErrorMessage(t.error)},r=openWork(e,{changeId:t.changeId}).runs,{spent:i}=t;return[...r.flatMap(({runId:e})=>runSettledFacts({error:n,outcome:`failed`,runId:e,...i?.runId===e&&{usage:i.usage}})),{data:{changeId:t.changeId,error:n,kind:e.changes[t.changeId]?.kind??`compaction`,outcome:`failed`},type:`context.settled`}]}function replaceSessionHistory(e,t){return contextStorage.getStore()?.delete(HistoryStateKey),{...e,history:t,compaction:withoutPromptMeasurement(e.compaction)}}async function maybeCompact(e){try{return await compactOnce(e)}catch(t){let n=summaryFailed(e.view(),{changeId:e.change.changeId,error:t});throw n.length>0&&await e.publish(n),t}}async function compactOnce(e){let{emissionState:t,publish:n}=e,r=e.messages,i=e.session,a=e.promptMessages??r,o=validateHarnessModelMessages(a),s=e.force===!0||shouldCompact(o,i.compaction,e.requestEnvelopeTokens,i.compaction.requestEnvelopeTokens);if(!s)return{compacted:!1,messages:r,session:i};let c=requireSessionModelReference(i),l=i.agent.compactionModelReference??c,u=l===c?e.model:await e.resolveModel(l),d=resolveModelProfile(u),f=resolveCallProviderOptions({auth:e.auth,profile:d,providerOptions:l.providerOptions,session:i}),p={abortSignal:e.abortSignal,headers:buildGatewayAttributionHeaders(d,e.runtimeIdentity),model:withGatewayTraceContext(u,d,e.traceContextHeaders),providerOptions:f,telemetry:e.telemetry&&{...e.telemetry,functionId:`eve.compaction`}},m,h=[],{change:g}=e,_=g.announced===!0,summarize=async r=>{_||(_=!0,await n([runRequested({changeId:g.changeId},g.summaryRunId),runStarted({attempt:0,modelId:gatewayModelId(u)??`unknown`,runId:g.summaryRunId})]));let a=await runModelCallWithRetries(()=>streamCompactionSummary({...p,telemetry:m??p.telemetry,...r}),{sessionId:i.sessionId,turnId:t.turnId},e.abortSignal);if(h.push(a.usage),a.text.trim().length===0)throw Error(`The compaction model returned an empty summary. Finish reason: ${a.finishReason}.`);return a.text};if(g.announced!==!0){let t=getInputTokenCount(o,i.compaction,e.requestEnvelopeTokens,i.compaction.requestEnvelopeTokens);await n(contextStarted({cause:g.cause,changeId:g.changeId,kind:`compaction`,trigger:e.force===!0||t===null?void 0:{inputTokens:t},turnId:g.turnId}),o)}s&&(m=e.prepareTelemetry?.(g.summaryRunId));let v=validateHarnessModelMessages(r),y=e.requestEnvelopeTokens??0,b=Math.max(0,estimateTokens(o)-estimateTokens(v)),x={...i.compaction,threshold:Math.max(1,i.compaction.threshold-y-b),lastKnownInputTokens:i.compaction.lastKnownInputTokens===void 0?void 0:Math.max(0,i.compaction.lastKnownInputTokens-Math.min(i.compaction.requestEnvelopeTokens??0,y))},S=[...v],C=takeCompactionCaptureFailure(e.ctx),w=C===void 0?void 0:{error:C};try{s&&w===void 0&&(assertFixedPartFits({protectedTokens:b,requestEnvelopeTokens:y,threshold:i.compaction.threshold}),S=await compactMessages([...v],x,summarize,e.force===!0,e.requestEnvelopeTokens===void 0?void 0:Math.max(0,getInputTokenCount(o,i.compaction,y,i.compaction.requestEnvelopeTokens)-y-b)))}catch(e){w={error:e}}let T=s?summaryUsageOf(h):void 0;T!==void 0&&e.betweenTurns===!0&&(i=markUsageReported(i,T));let E=g.summaryRunId;if(w!==void 0)return await n(summaryFailed(e.view(),{changeId:g.changeId,error:w.error,...T!==void 0&&{spent:{runId:E,usage:T}}})),{compacted:!1,failure:w,messages:e.messages,session:i};r=validateHarnessModelMessages(S);{let e=h.length>0?runSettledFacts({finishReason:`stop`,outcome:`completed`,runId:E,usage:T}):_?runSettledFacts({outcome:`abandoned`,runId:E}):[];e.push({data:{changeId:g.changeId,kind:`compaction`,outcome:`completed`},type:`context.settled`}),await n(e,r)}return{compacted:!0,messages:r,session:replaceSessionHistory(i,r)}}function compactedTurnInput(e){return e.findLast(e=>isUserModelMessage(e)&&e.kind===`user`)}async function streamCompactionSummary(e){let t=streamText({...e,onError:()=>{}});for await(let e of t.fullStream)if(e.type===`error`)throw normalizeModelStreamError(e.error);let[n,r,i,a]=await Promise.all([t.text,t.finishReason,t.usage,t.providerMetadata]);return{finishReason:r,text:n,usage:extractTokenUsageDelta({costUsd:extractGatewayCostUsd(a),usage:i})}}export{assertFixedPartFits,compactHistory,compactedTurnInput,maybeCompact,replaceSessionHistory};

package/dist/src/harness/coordination.js

@@ -1 +1 @@
-import{skillTarget}from"#shared/action-request-name.js";import{SKILL_ENTRY_NAME,SKILL_TOOL_NAME}from"#protocol/catalog-tools.js";import{parseJsonObject}from"#shared/json.js";import{findBlockingWorkflowToolRun,removeBlockingWorkflowToolRuns}from"#harness/workflow-tool-runs.js";import{readHitlState}from"#harness/hitl/index.js";import{startsTasks}from"#execution/tasks/tool-entry-point.js";import{commitCallEntry,isTaskTool}from"#execution/tasks/model-step.js";import{normalizeToolModelOutput}from"#harness/tool-model-output.js";import{observeToolOutput}from"#tool-stubs/execute.js";import{markRuntimeWorkflowToolAction}from"#shared/action-types.js";function assertUniqueCoordinationCallIds(e){let t=new Set;for(let n of e){if(t.has(n.callId))throw Error(`Coordination batch contains duplicate callId "${n.callId}".`);t.add(n.callId)}}function forgetFinishedRuns(e,t,n){let r=e,i=[];for(let e of t){if(e.kind!==`tool-result`)continue;let t=findBlockingWorkflowToolRun(r.state,e.callId,n);if(t!==void 0){for(let[e,n]of readHitlState(r.state).relays)n.runId===t.address.runId&&i.push(e);r=removeBlockingWorkflowToolRuns(r,n,t.callId)}}return{requestIds:i,session:r}}async function runtimeResultCalls(e,t){let n=[];for(let r of e){switch(r.kind){case`load-skill-result`:n.push({part:toolResult(r,SKILL_TOOL_NAME,toToolResultOutput(r)),result:r});continue;case`subagent-result`:n.push({part:toolResult(r,r.subagentName,toToolResultOutput(r)),result:r});continue;case`tool-result`:n.push({part:toolResult(r,r.toolName,await projectToolResultOutput(r,t?.get(r.toolName))),result:r});continue}throw Error(`Unsupported runtime action result kind "${String(r)}".`)}return n}function toolResult(e,t,n){return{output:n,toolCallId:e.callId,toolName:t,type:`tool-result`}}function collectWorkflowCalls(e){let{session:t}=e,n=[];for(let r of e.toolCalls){let i=e.tools.get(r.toolName);if(isTaskTool(i))continue;let a=commitCallEntry(t,{callId:r.toolCallId,definition:i,input:resolveToolCallInputObject(r.input,{callId:r.toolCallId,toolName:r.toolName}),toolName:r.toolName,turnId:e.turnId});t=a.session,n.push(createCoordinationRequestFromToolCall({entry:a.entry,input:a.input,toolCall:r,tools:e.tools}))}return{session:t,workflowRequests:n}}function createRuntimeActionRequestFromToolCall(e){let t=e.tools.get(e.toolCall.toolName),n=resolveToolCallInputObject(e.toolCall.input,{callId:e.toolCall.toolCallId,toolName:e.toolCall.toolName}),r=e.toolCall.toolName===SKILL_ENTRY_NAME?skillTarget(n):void 0;if(r!==void 0)return{callId:e.toolCall.toolCallId,input:n,kind:`load-skill`,name:r};let i=t?.behavior?.handling;if(t!==void 0&&i?.kind===`dispatch`&&i.target.kind!==`workflow-tool-call`){let r=i.target,a={callId:e.toolCall.toolCallId,description:t.description,input:n,name:e.toolCall.toolName,nodeId:r.nodeId};return r.kind===`remote-agent-call`?{...a,kind:`remote-agent-call`,remoteAgentName:r.remoteAgentName}:{...a,kind:`subagent-call`,subagentName:r.subagentName}}let a={callId:e.toolCall.toolCallId,input:n,kind:`tool-call`,toolName:e.toolCall.toolName};return t?.workflowId===void 0?a:markRuntimeWorkflowToolAction(a)}function createCoordinationRequestFromToolCall(e){let t=e.tools.get(e.toolCall.toolName);if(t?.workflowId===void 0)throw Error(`Workflow tool "${e.toolCall.toolName}" has no workflow.`);return{callId:e.toolCall.toolCallId,entry:e.entry,executeInput:t.executeInput?.(e.input),input:e.input,kind:`workflow-task`,toolName:e.toolCall.toolName,workflowId:t.workflowId}}function resolveToolCallInputObject(e,t){if(e==null||typeof e==`string`&&e.trim()===``)return{};try{return parseJsonObject(typeof e==`string`?parseJsonStringInput(e):e)}catch(e){let n=e instanceof Error?e.message:String(e);throw TypeError(`Failed to parse tool-call arguments for "${t.toolName}" (${t.callId}): ${n}`,{cause:e})}}function parseJsonStringInput(e){return JSON.parse(e)}async function projectToolResultOutput(e,t){return e.isError===!0||t?.toModelOutput===void 0||startsTasks(t)?toToolResultOutput(e):await observeToolOutput(e.toolName,[{callId:e.callId}],async()=>normalizeToolModelOutput({output:await t.toModelOutput(e.output),toolCallId:e.callId,toolName:e.toolName}))}function toToolResultOutput(e){return typeof e.output==`string`?e.isError===!0?{type:`error-text`,value:e.output}:{type:`text`,value:e.output}:e.isError===!0?{type:`error-json`,value:toMutableJsonValue(e.output)}:{type:`json`,value:toMutableJsonValue(e.output)}}function toMutableJsonValue(e){if(e===null||typeof e==`string`||typeof e==`number`||typeof e==`boolean`)return e;if(Array.isArray(e))return e.map(e=>toMutableJsonValue(e));let t={};for(let[n,r]of Object.entries(e))t[n]=toMutableJsonValue(r);return t}export{assertUniqueCoordinationCallIds,collectWorkflowCalls,createCoordinationRequestFromToolCall,createRuntimeActionRequestFromToolCall,forgetFinishedRuns,resolveToolCallInputObject,runtimeResultCalls};
+import{parseJsonObject}from"#shared/json.js";import{SKILL_ENTRY_NAME,SKILL_TOOL_NAME}from"#protocol/catalog-tools.js";import{readHitlState}from"#harness/hitl/index.js";import{startsTasks}from"#execution/tasks/tool-entry-point.js";import{skillTarget}from"#shared/action-request-name.js";import{commitCallEntry,isTaskTool}from"#execution/tasks/model-step.js";import{normalizeToolModelOutput}from"#harness/tool-model-output.js";import{findBlockingWorkflowToolRun,removeBlockingWorkflowToolRuns}from"#harness/workflow-tool-runs.js";import{observeToolOutput}from"#tool-stubs/execute.js";import{markRuntimeWorkflowToolAction}from"#shared/action-types.js";function assertUniqueCoordinationCallIds(e){let t=new Set;for(let n of e){if(t.has(n.callId))throw Error(`Coordination batch contains duplicate callId "${n.callId}".`);t.add(n.callId)}}function forgetFinishedRuns(e,t,n){let r=e,i=[];for(let e of t){if(e.kind!==`tool-result`)continue;let t=findBlockingWorkflowToolRun(r.state,e.callId,n);if(t!==void 0){for(let[e,n]of readHitlState(r.state).relays)n.runId===t.address.runId&&i.push(e);r=removeBlockingWorkflowToolRuns(r,n,t.callId)}}return{requestIds:i,session:r}}async function runtimeResultCalls(e,t){let n=[];for(let r of e){switch(r.kind){case`load-skill-result`:n.push({part:toolResult(r,SKILL_TOOL_NAME,toToolResultOutput(r)),result:r});continue;case`subagent-result`:n.push({part:toolResult(r,r.subagentName,toToolResultOutput(r)),result:r});continue;case`tool-result`:n.push({part:toolResult(r,r.toolName,await projectToolResultOutput(r,t?.get(r.toolName))),result:r});continue}throw Error(`Unsupported runtime action result kind "${String(r)}".`)}return n}function toolResult(e,t,n){return{output:n,toolCallId:e.callId,toolName:t,type:`tool-result`}}function collectWorkflowCalls(e){let{session:t}=e,n=[];for(let r of e.toolCalls){let i=e.tools.get(r.toolName);if(isTaskTool(i))continue;let a=commitCallEntry(t,{callId:r.toolCallId,definition:i,input:resolveToolCallInputObject(r.input,{callId:r.toolCallId,toolName:r.toolName}),toolName:r.toolName,turnId:e.turnId});t=a.session,n.push(createCoordinationRequestFromToolCall({entry:a.entry,input:a.input,toolCall:r,tools:e.tools}))}return{session:t,workflowRequests:n}}function createRuntimeActionRequestFromToolCall(e){let t=e.tools.get(e.toolCall.toolName),n=resolveToolCallInputObject(e.toolCall.input,{callId:e.toolCall.toolCallId,toolName:e.toolCall.toolName}),r=e.toolCall.toolName===SKILL_ENTRY_NAME?skillTarget(n):void 0;if(r!==void 0)return{callId:e.toolCall.toolCallId,input:n,kind:`load-skill`,name:r};let i=t?.behavior?.handling;if(t!==void 0&&i?.kind===`dispatch`&&i.target.kind!==`workflow-tool-call`){let r=i.target,a={callId:e.toolCall.toolCallId,description:t.description,input:n,name:e.toolCall.toolName,nodeId:r.nodeId};return r.kind===`remote-agent-call`?{...a,kind:`remote-agent-call`,remoteAgentName:r.remoteAgentName}:{...a,kind:`subagent-call`,subagentName:r.subagentName}}let a={callId:e.toolCall.toolCallId,input:n,kind:`tool-call`,toolName:e.toolCall.toolName};return t?.workflowId===void 0?a:markRuntimeWorkflowToolAction(a)}function createCoordinationRequestFromToolCall(e){let t=e.tools.get(e.toolCall.toolName);if(t?.workflowId===void 0)throw Error(`Workflow tool "${e.toolCall.toolName}" has no workflow.`);return{callId:e.toolCall.toolCallId,entry:e.entry,executeInput:t.executeInput?.(e.input),input:e.input,kind:`workflow-task`,toolName:e.toolCall.toolName,workflowId:t.workflowId}}function resolveToolCallInputObject(e,t){if(e==null||typeof e==`string`&&e.trim()===``)return{};try{return parseJsonObject(typeof e==`string`?JSON.parse(e):e)}catch(e){let n=e instanceof Error?e.message:String(e);throw TypeError(`Failed to parse tool-call arguments for "${t.toolName}" (${t.callId}): ${n}`,{cause:e})}}async function projectToolResultOutput(e,t){return e.isError===!0||t?.toModelOutput===void 0||startsTasks(t)?toToolResultOutput(e):await observeToolOutput(e.toolName,[{callId:e.callId}],async()=>normalizeToolModelOutput({output:await t.toModelOutput(e.output),toolCallId:e.callId,toolName:e.toolName}))}function toToolResultOutput(e){return typeof e.output==`string`?{type:e.isError===!0?`error-text`:`text`,value:e.output}:{type:e.isError===!0?`error-json`:`json`,value:structuredClone(e.output)}}export{assertUniqueCoordinationCallIds,collectWorkflowCalls,createCoordinationRequestFromToolCall,createRuntimeActionRequestFromToolCall,forgetFinishedRuns,resolveToolCallInputObject,runtimeResultCalls};

package/dist/src/harness/current-messages.d.ts

@@ -5,6 +5,11 @@ import { type FrameworkMessageKind, type HarnessModelMessage } from "#harness/me
 interface AddCurrentMessageOptions {
     readonly cacheFriendly?: boolean;
 }
+/** Context for a model call that belongs with one turn's input. */
+export interface AnchoredContext {
+    readonly turnId: string;
+    readonly content: string;
+}
 interface CurrentMessagesOptions {
     readonly historyState?: HistoryState;
     readonly currentTurnMessages?: readonly HarnessModelMessage[];
@@ -17,6 +22,11 @@ export declare function createCurrentMessages(history: readonly HarnessModelMess
     readonly nonSystemMessages: readonly HarnessModelMessage[];
     readonly systemMessages: readonly SystemModelMessage[];
     add(message: string, kind: FrameworkMessageKind, options?: AddCurrentMessageOptions): void;
+    /**
+     * Adds context for this call without keeping it in history, each entry just ahead of the input
+     * its turn brought. Add it last: what's added after would land between it and that input.
+     */
+    addAnchored(entries: readonly AnchoredContext[], kind: FrameworkMessageKind): void;
     addAnnouncements(announcements: Readonly<Record<string, Announcement>>): void;
     addSystem(messages: SystemModelMessage | readonly SystemModelMessage[]): void;
 };

package/dist/src/harness/current-messages.js

@@ -1 +1 @@
-import{createFrameworkUserMessage}from"#harness/messages.js";function createCurrentMessages(e,t={}){let n=[...e],r={...t.historyState},i=[],a=[],o=new Set(t.currentTurnMessages),s;for(let n of t.projectedMessages??e)s===void 0&&o.has(n)&&(s=a.length),n.role===`system`?i.push(n):a.push(n);let c=s??a.length,l=e.findIndex(e=>o.has(e)),u=l===-1?e.length:l;function appendUserMessage(e,t){let r=createFrameworkUserMessage(t,e);a.splice(c,0,r),n.splice(u,0,r),c+=1,u+=1}function add(e,t,{cacheFriendly:n=!0}={}){if(n){appendUserMessage(e,t);return}i.push({role:`system`,content:e})}return{add,addAnnouncements(e){for(let t of Object.keys(e).sort()){let n=e[t],i=r.announcements?.[t];i!==n.value&&(appendUserMessage(n.render(i),`context.state`),r.announcements={...r.announcements,[t]:n.value})}},addSystem(e){i.push(...Array.isArray(e)?e:[e])},get nonSystemMessages(){return[...a]},get history(){return[...n]},get historyState(){return{...r}},get systemMessages(){return[...i]}}}export{createCurrentMessages};
+import{createFrameworkUserMessage,turnAnchorOf}from"#harness/messages.js";function createCurrentMessages(e,t={}){let n=[...e],r={...t.historyState},i=[],a=[],o=new Set(t.currentTurnMessages),s;for(let n of t.projectedMessages??e)s===void 0&&o.has(n)&&(s=a.length),n.role===`system`?i.push(n):a.push(n);let c=s??a.length,l=e.findIndex(e=>o.has(e)),u=l===-1?e.length:l;function appendUserMessage(e,t){let r=createFrameworkUserMessage(t,e);a.splice(c,0,r),n.splice(u,0,r),c+=1,u+=1}function add(e,t,{cacheFriendly:n=!0}={}){if(n){appendUserMessage(e,t);return}i.push({role:`system`,content:e})}return{add,addAnchored(e,t){if(e.length===0)return;let n=new Map;a.forEach((e,t)=>{let r=turnAnchorOf(e);r!==void 0&&!n.has(r)&&n.set(r,t)});let r=checkpointLength(a),i=e.map(({turnId:e})=>n.get(e)??r),o=[],s=0;for(let n=0;n<=a.length;n+=1)e.forEach((e,r)=>{i[r]===n&&(o.push(createFrameworkUserMessage(t,e.content)),n<=c&&(s+=1))}),n<a.length&&o.push(a[n]);a.splice(0,a.length,...o),c+=s},addAnnouncements(e){for(let t of Object.keys(e).sort()){let n=e[t],i=r.announcements?.[t];i!==n.value&&(appendUserMessage(n.render(i),`context.state`),r.announcements={...r.announcements,[t]:n.value})}},addSystem(e){i.push(...Array.isArray(e)?e:[e])},get nonSystemMessages(){return[...a]},get history(){return[...n]},get historyState(){return{...r}},get systemMessages(){return[...i]}}}function checkpointLength(e){let[t,n]=e;return t?.role===`user`&&t.kind===`context.compaction`&&n?.role===`assistant`?2:0}export{createCurrentMessages};

package/dist/src/harness/emission.d.ts

@@ -1,4 +1,4 @@
-import type { ModelMessage, TextStreamPart, ToolSet, TypedToolResult } from "ai";
+import type { ModelMessage, TextStreamPart, ToolSet } from "ai";
 type ToolResponsePart = Extract<ModelMessage, {
     role: "tool";
 }>["content"][number];
@@ -7,34 +7,23 @@ type InlineToolResultPart = Extract<ToolResponsePart, {
 }>;
 import type { TurnPosition } from "#harness/session-machine/view.js";
 import type { HarnessEmitFn, HarnessToolLookup } from "#harness/types.js";
-/**
- * Result of consuming one step's `fullStream`.
- *
- * Inline results avoid duplicate post-step events. Approval-resume
- * authorization results also route back to the park detector.
- */
+/** The calls of one step's `fullStream` whose input was invalid, and their results. */
 interface EmittedStreamContent {
-    readonly emittedActionCallIds: ReadonlySet<string>;
-    readonly handledInlineToolResultCallIds: ReadonlySet<string>;
     readonly invalidInputToolCallIds: ReadonlySet<string>;
-    readonly inlineAuthorizationResults: readonly TypedToolResult<ToolSet>[];
     readonly trailingInlineToolResultParts: readonly InlineToolResultPart[];
 }
 interface StreamActionEmissionOptions {
     readonly excludedActionToolNames: ReadonlySet<string>;
     /**
-     * The session is a child or a schedule, whose caller hears only the turn's
-     * real end: text before `eve__task_wait` reports `"tool-calls"`, and
-     * channels don't post it as the reply.
-     */
-    readonly hidesHeldText?: boolean;
-    /**
-     * A task is working, so the turn is held and a text step can't end it. For a
-     * session that `hidesHeldText`, the step reports `"tool-calls"`.
+     * The turn holds while its tasks work and its caller reads only the final reply, so a text step
+     * can't end it: its text narrates, and only the reply after the tasks report is the turn's.
      */
-    readonly tasksWorking?: boolean;
+    readonly held?: boolean;
+    /** A child or schedule's caller reads only the final reply, including at an explicit task wait. */
+    readonly finalReplyOnly?: boolean;
     readonly tools: HarnessToolLookup;
-    readonly unsettledActionToolNames?: Map<string, string>;
+    /** A deliberate turn cancel or steering interrupt, not a provider failure/retry. */
+    readonly interruptSignal?: AbortSignal;
 }
 /**
  * Consumes the AI SDK `fullStream` and emits real-time text and reasoning

package/dist/src/harness/emission.js

@@ -1 +1 @@
-import{TASK_WAIT_TOOL_NAME}from"#protocol/task-tools.js";import{createActionInputAppendedEvent,createActionPartialEvent,createActionResultEvent,createActionsRequestedEvent,createMessageAppendedEvent,createMessageCompletedEvent,createReasoningAppendedEvent,createReasoningCompletedEvent}from"#protocol/message.js";import{createRuntimeToolResultFromStepResult,createRuntimeToolResultFromToolError,createToolResultMessagePartFromToolError,toActionResult}from"#harness/action-result-helpers.js";import{normalizeModelStreamError}from"#harness/model-call/errors.js";import{createInvalidToolCallInputError,isInvalidToolCall,resolveProviderToolCallRequest}from"#harness/tool-call-input-errors.js";import{collectActionPresentation,createPresentedRuntimeActionRequestFromToolCall}from"#harness/action-presentation.js";import{projectDeltaPresentation,projectResultPresentation}from"#harness/tool-presentation.js";import{createProviderStreamActionBatch}from"#harness/stream-actions.js";import{createOrderedStreamEmitter}from"#harness/ordered-stream-emitter.js";import{interruptStreamOnFailure}from"#harness/interruptible-stream.js";import{isInlineAuthorizationToolResult}from"#harness/inline-tool-authorization.js";import{normalizeAssistantStepFinishReason}from"#harness/finish-reason.js";async function emitStreamContent(e,t,n,r){let i=createOrderedStreamEmitter(e),a=createProviderStreamActionBatch({emitFn:i.emit,onActionsEmitted:e=>{for(let{request:t,toolName:n}of e)r?.unsettledActionToolNames?.set(t.action.callId,n)},state:t});try{return await consumeStreamContent(i.emit,t,interruptStreamOnFailure(n,i.failureSignal),a,r)}finally{try{await a.cancel()}finally{await i.closeAndDrain()}}}function reportedFinishReason(e,t){return t&&e===`stop`?`tool-calls`:e}function textFinishReasonBefore(e,t){return e===TASK_WAIT_TOOL_NAME&&!t?`stop`:`tool-calls`}async function consumeStreamContent(e,t,n,r,i){let a=``,o=``,s=`stop`,c,l=new Set,u=new Set,d=new Set,f=new Set,p=new Set,m=new Set,h=[],g=[],_=new Map,v=new Map,flushCurrentMessage=async(n=`tool-calls`)=>{o.length!==0&&(await e(createMessageCompletedEvent({finishReason:n,message:o,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId})),o=``)},emitActionInput=async(n,r,a)=>{await e(createActionInputAppendedEvent({callId:n,inputTextDelta:a,sequence:t.sequence,stepIndex:t.stepIndex,toolName:r,turnId:t.turnId})),i?.unsettledActionToolNames?.set(n,r)},emitActionRequest=async(n,r)=>{let{action:a}=n;u.has(a.callId)||(o.trim().length>0&&await flushCurrentMessage(textFinishReasonBefore(r,i?.hidesHeldText===!0)),u.add(a.callId),_.set(a.callId,a.input),await e(createActionsRequestedEvent({actions:[a],presentation:collectActionPresentation([n]),sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId})),i?.unsettledActionToolNames?.set(a.callId,r))},collectProviderToolCall=async e=>{if(f.has(e.toolCallId)||(f.add(e.toolCallId),u.has(e.toolCallId)))return;u.add(e.toolCallId),o.trim().length>0&&await flushCurrentMessage();let t=resolveProviderToolCallRequest(e,i?.tools??new Map);if(t.toolError!==void 0){m.add(e.toolCallId),await emitActionResult(createRuntimeToolResultFromToolError(t.toolError)),p.add(e.toolCallId),g.push(createToolResultMessagePartFromToolError(t.toolError));return}_.set(t.request.action.callId,t.request.action.input),r.observe(t.request,e.toolName)},emitActionResult=async n=>{if(d.has(n.callId))return;d.add(n.callId);let r=n.isError===!0?void 0:projectResultPresentation(i?.tools.get(n.toolName),n.callId,_.get(n.callId),n.output);await e(createActionResultEvent({presentation:r,result:toActionResult(n,_.get(n.callId)),sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId})),i?.unsettledActionToolNames?.delete(n.callId)},emitActionPartial=async n=>{let r=projectDeltaPresentation(i?.tools.get(n.toolName),n.callId,_.get(n.callId),n.output);await e(createActionPartialEvent({presentation:r,result:n,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId}))},emitToolCall=async e=>{if(isInvalidToolCall(e)){m.add(e.toolCallId);return}if(!(i===void 0||i.excludedActionToolNames.has(e.toolName)))try{await emitActionRequest(createPresentedRuntimeActionRequestFromToolCall({toolCall:e,tools:i.tools}),e.toolName)}catch(t){if(t instanceof TypeError){let n=createInvalidToolCallInputError({error:t,toolCall:e});m.add(e.toolCallId),o.trim().length>0&&await flushCurrentMessage(),await emitActionResult(createRuntimeToolResultFromToolError(n)),p.add(e.toolCallId),g.push(createToolResultMessagePartFromToolError(n));return}throw t}};for await(let d of n)if(c===void 0)switch(d.type){case`reasoning-delta`:await r.flush(),a+=d.text,await e(createReasoningAppendedEvent({reasoningDelta:d.text,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId}));break;case`text-delta`:await r.flush(),a.trim().length>0&&(await e(createReasoningCompletedEvent({reasoning:a,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId})),a=``),o+=d.text,await e(createMessageAppendedEvent({messageDelta:d.text,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId}));break;case`tool-input-start`:if(i===void 0||d.providerExecuted===!0||i.tools.get(d.toolName)?.behavior?.handling?.kind===`provider-tool`||i.excludedActionToolNames.has(d.toolName)){v.delete(d.id);break}await r.flush(),o.trim().length>0&&await flushCurrentMessage(textFinishReasonBefore(d.toolName,i.hidesHeldText===!0)),v.set(d.id,{toolName:d.toolName});break;case`tool-input-delta`:{let e=v.get(d.id);if(e===void 0)break;await r.flush(),await emitActionInput(d.id,e.toolName,d.delta);break}case`tool-input-end`:v.delete(d.id);break;case`tool-call`:{let e=d;v.delete(e.toolCallId),l.add(e.toolCallId),e.providerExecuted===!0?await collectProviderToolCall(e):(await r.flush(),await emitToolCall(e));break}case`tool-result`:{let e=d;if(e.preliminary===!0){e.providerExecuted!==!0&&await emitActionPartial(createRuntimeToolResultFromStepResult(e));break}if(e.providerExecuted===!0){await collectProviderToolCall({input:`input`in e?e.input:void 0,toolCallId:e.toolCallId,toolName:e.toolName}),await r.flush(),await emitActionResult(createRuntimeToolResultFromStepResult(e));break}if(l.has(d.toolCallId)){if(isInlineAuthorizationToolResult(e))break;u.has(d.toolCallId)&&(await emitActionResult(createRuntimeToolResultFromStepResult(e)),p.add(d.toolCallId));break}if(await r.flush(),await flushCurrentMessage(),isInlineAuthorizationToolResult(e)){p.add(d.toolCallId),h.push(e);break}await emitActionResult(createRuntimeToolResultFromStepResult(e)),p.add(d.toolCallId);break}case`tool-error`:{let e=d;e.providerExecuted===!0?(await collectProviderToolCall(e),await r.flush(),await emitActionResult(createRuntimeToolResultFromToolError(e))):u.has(e.toolCallId)&&(await emitActionResult(createRuntimeToolResultFromToolError(e)),p.add(e.toolCallId),g.push(createToolResultMessagePartFromToolError(e)));break}case`finish-step`:s=normalizeAssistantStepFinishReason(d.finishReason),await r.flush();break;case`error`:c=normalizeModelStreamError(d.error);break;case`abort`:throw new DOMException(d.reason??`The model stream was aborted.`,`AbortError`)}if(await r.flush(),c!==void 0)throw c;return a.trim().length>0&&await e(createReasoningCompletedEvent({reasoning:a,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId})),s!==`content-filter`&&o.trim().length>0&&await e(createMessageCompletedEvent({finishReason:reportedFinishReason(s,i?.hidesHeldText===!0&&i.tasksWorking===!0),message:o,sequence:t.sequence,stepIndex:t.stepIndex,turnId:t.turnId})),{emittedActionCallIds:u,handledInlineToolResultCallIds:p,invalidInputToolCallIds:m,inlineAuthorizationResults:h,trailingInlineToolResultParts:g}}export{emitStreamContent};
+import{TASK_WAIT_TOOL_NAME}from"#protocol/task-tools.js";import{announcedKind,callRequested,callSettledFrom}from"#harness/call-facts.js";import{createRuntimeToolResultFromStepResult,createRuntimeToolResultFromToolError,createToolResultMessagePartFromToolError}from"#harness/action-result-helpers.js";import{createInvalidToolCallInputError,isInvalidToolCall,resolveProviderToolCallRequest}from"#harness/tool-call-input-errors.js";import{createPresentedRuntimeActionRequestFromToolCall,projectToolLabel}from"#harness/action-presentation.js";import{normalizeModelStreamError}from"#harness/model-call/errors.js";import{createProviderStreamActionBatch}from"#harness/stream-actions.js";import{createOrderedStreamEmitter}from"#harness/ordered-stream-emitter.js";import{interruptStreamOnFailure}from"#harness/interruptible-stream.js";import{normalizeAssistantStepFinishReason}from"#harness/finish-reason.js";async function emitStreamContent(e,t,n,r){let i=createOrderedStreamEmitter(e),a=createProviderStreamActionBatch({emitFn:i.emit,state:t});try{return await consumeStreamContent(i.emit,t,interruptStreamOnFailure(n,i.failureSignal),a,r)}finally{try{await a.cancel()}finally{await i.closeAndDrain()}}}function finalPhase(e,t){return e===`tool-calls`||t&&e===`stop`?`narration`:`reply`}function textPhaseBefore(e,t){return e===TASK_WAIT_TOOL_NAME&&!t?`reply`:`narration`}async function consumeStreamContent(e,t,n,r,i){let a=t.runId??`${t.turnId}.run`,o={runId:a,turnId:t.turnId},s=0,nextPartId=()=>`${a}.${String(s++)}`,c=``,l,u=``,d,f=`stop`,p,m=new Set,h=new Set,g=new Set,_=new Set,v=[],y=new Map,b=new Map,completePart=async(t,n,r,i)=>{await e({data:{kind:n,partId:t,phase:i,runId:a,value:r},type:`content.completed`})},flushCurrentMessage=async(e=`narration`)=>{u.length!==0&&d!==void 0&&(await completePart(d,`text`,u,e),u=``,d=void 0)},flushCurrentReasoning=async()=>{c.trim().length>0&&l!==void 0&&await completePart(l,`reasoning`,c,`narration`),c=``,l=void 0},emitActionInput=async(t,n)=>{let r=b.get(t);if(r===void 0)return;let a=r.announced?{data:{callId:t,delta:n},type:`call.input`}:{data:{callId:t,delta:n,kind:announcedKind(r.toolName,i?.tools.get(r.toolName)),name:r.toolName},scope:o,type:`call.input`};r.announced=!0,await e(a)},emitActionRequest=async t=>{let{action:n}=t;m.has(n.callId)||(u.trim().length>0&&await flushCurrentMessage(n.kind===`tool-call`?textPhaseBefore(n.toolName,i?.finalReplyOnly===!0):`narration`),m.add(n.callId),y.set(n.callId,n.input),await e(callRequested({action:n,owner:{runId:a},label:t.presentationLabel})))},collectProviderToolCall=async e=>{if(g.has(e.toolCallId)||(g.add(e.toolCallId),m.has(e.toolCallId)))return;m.add(e.toolCallId),u.trim().length>0&&await flushCurrentMessage();let t=resolveProviderToolCallRequest(e,i?.tools??new Map);if(t.toolError!==void 0){_.add(e.toolCallId),await emitInvalidCall({callId:e.toolCallId,input:{},kind:`tool-call`,toolName:e.toolName},createRuntimeToolResultFromToolError(t.toolError)),v.push(createToolResultMessagePartFromToolError(t.toolError));return}y.set(t.request.action.callId,t.request.action.input),r.observe(t.request)},emitInvalidCall=async(t,n)=>{m.add(t.callId);let r=typeof n.output==`string`?n.output:JSON.stringify(n.output??null);await e(callRequested({action:t,inputError:{code:`INVALID_TOOL_INPUT`,message:r},owner:{runId:a}})),await emitActionResult(n)},emitActionResult=async t=>{if(h.has(t.callId))return;h.add(t.callId);let n=t.isError===!0?void 0:projectToolLabel(i?.tools.get(t.toolName),`complete`,y.get(t.callId),t.output);await e(callSettledFrom(t,{label:n}))},emitToolCall=async e=>{if(isInvalidToolCall(e)){_.add(e.toolCallId);return}if(!(i===void 0||i.excludedActionToolNames.has(e.toolName)))try{await emitActionRequest(createPresentedRuntimeActionRequestFromToolCall({toolCall:e,tools:i.tools}))}catch(t){if(t instanceof TypeError){let n=createInvalidToolCallInputError({error:t,toolCall:e});_.add(e.toolCallId),u.trim().length>0&&await flushCurrentMessage(),await emitInvalidCall({callId:e.toolCallId,input:{},kind:`tool-call`,toolName:e.toolName},createRuntimeToolResultFromToolError(n)),v.push(createToolResultMessagePartFromToolError(n));return}throw t}};try{for await(let t of n)if(p===void 0)switch(t.type){case`reasoning-delta`:{await r.flush(),c+=t.text;let n=l===void 0;l??=nextPartId(),await e(n?{data:{delta:t.text,kind:`reasoning`,partId:l},scope:o,type:`content.delta`}:{data:{delta:t.text,partId:l},type:`content.delta`});break}case`text-delta`:{await r.flush(),await flushCurrentReasoning(),u+=t.text;let n=d===void 0;d??=nextPartId(),await e(n?{data:{delta:t.text,kind:`text`,partId:d},scope:o,type:`content.delta`}:{data:{delta:t.text,partId:d},type:`content.delta`});break}case`tool-input-start`:if(i===void 0||t.providerExecuted===!0||i.tools.get(t.toolName)?.behavior?.handling?.kind===`provider-tool`||i.excludedActionToolNames.has(t.toolName)){b.delete(t.id);break}await r.flush(),u.trim().length>0&&await flushCurrentMessage(textPhaseBefore(t.toolName,i.finalReplyOnly===!0)),b.set(t.id,{announced:!1,toolName:t.toolName});break;case`tool-input-delta`:if(!b.has(t.id))break;await r.flush(),await emitActionInput(t.id,t.delta);break;case`tool-input-end`:b.delete(t.id);break;case`tool-call`:{let e=t;b.delete(e.toolCallId),e.providerExecuted===!0?await collectProviderToolCall(e):(await r.flush(),await emitToolCall(e));break}case`tool-result`:{let e=t;if(e.providerExecuted!==!0||e.preliminary===!0)break;await collectProviderToolCall({input:`input`in e?e.input:void 0,toolCallId:e.toolCallId,toolName:e.toolName}),await r.flush(),await emitActionResult(createRuntimeToolResultFromStepResult(e));break}case`tool-error`:{let e=t;if(e.providerExecuted!==!0)break;await collectProviderToolCall(e),await r.flush(),await emitActionResult(createRuntimeToolResultFromToolError(e));break}case`finish-step`:f=normalizeAssistantStepFinishReason(t.finishReason),await r.flush();break;case`error`:p=normalizeModelStreamError(t.error);break;case`abort`:throw new DOMException(t.reason??`The model stream was aborted.`,`AbortError`)}}catch(t){if(i?.interruptSignal?.aborted===!0){let t=[];l!==void 0&&c.length>0&&t.push({type:`content.completed`,data:{partId:l,runId:a,kind:`reasoning`,phase:`narration`,value:c,interrupted:!0}}),d!==void 0&&u.length>0&&t.push({type:`content.completed`,data:{partId:d,runId:a,kind:`text`,phase:`narration`,value:u,interrupted:!0}}),t.length>0&&await e(t)}throw t}if(await r.flush(),p!==void 0)throw p;return await flushCurrentReasoning(),f!==`content-filter`&&u.trim().length>0&&d!==void 0&&await completePart(d,`text`,u,finalPhase(f,i?.held===!0)),{invalidInputToolCallIds:_,trailingInlineToolResultParts:v}}export{emitStreamContent};

package/dist/src/harness/execute-call.d.ts

@@ -3,10 +3,8 @@
  * seam reads an `eve__tool` or `eve__skill` call as the call to the entry it
  * names. Model history keeps the model's own calls.
  */
-import type { ModelMessage, TelemetryOptions, TextStreamPart, ToolSet } from "ai";
-import type { HarnessToolDefinition } from "#harness/execute-tool.js";
+import type { ModelMessage, TextStreamPart, ToolSet } from "ai";
 import type { HarnessStepResult } from "#harness/step-hooks.js";
-import { toolCallModelOutput } from "#harness/tool-call-io.js";
 import type { CallResolver, ResolvedCall, ToolCallLike } from "#harness/types.js";
 import type { ToolExecuteOptions } from "#tools/definition.js";
 /**
@@ -15,23 +13,14 @@ import type { ToolExecuteOptions } from "#tools/definition.js";
  * call made through one.
  */
 export declare function historyCallNames(messages: readonly ModelMessage[]): ReadonlyMap<string, string>;
-/** Whether a call runs as a workflow tool or agent the harness dispatches after the step. */
-export declare function dispatchesAfterStep(resolved: ResolvedCall<ToolCallLike> | undefined): boolean;
-/** Runs a resolved catalog call, or stands in for one the harness dispatches. */
-export declare function runEntryCall(resolved: ResolvedCall<ToolCallLike>, options: ToolExecuteOptions): unknown;
 /**
- * The call a tool stub matches. An `eve__tool` call matches as the call to its
- * entry, with the entry's input, so a stub applies to it exactly as to a direct
- * call. A call the harness dispatches after the step has none here: its
- * dispatch applies the stub, as it does for a direct call.
+ * Runs a resolved catalog call. eve resolves every call before it runs one, so this runs only
+ * for a catalog tool's own `execute`, which no resolved call reaches.
  */
-export declare function stubbedCall(call: ToolCallLike, resolved: ResolvedCall<ToolCallLike> | undefined): ToolCallLike | undefined;
-/** The model output of a call's result; a stand-in's output is dropped with it. */
-export declare function entryModelOutput(definition: Pick<HarnessToolDefinition, "name" | "toModelOutput">, output: unknown, toolCallId: string | undefined): ReturnType<typeof toolCallModelOutput>;
+export declare function runEntryCall(resolved: ResolvedCall<ToolCallLike>, options: ToolExecuteOptions): unknown;
 /**
- * The step as the harness reads it: catalog calls become calls to
- * their entries and stand-in results are dropped. The response messages, which
- * become history, keep the model's own calls.
+ * The step as the harness reads it: catalog calls become calls to their entries. The response
+ * messages, which become history, keep the model's own calls.
  */
 export declare function toEntryStep<T extends ModelMessage>(step: HarnessStepResult, responseMessages: readonly T[], resolve: CallResolver): readonly [HarnessStepResult, T[]];
 /**
@@ -40,9 +29,3 @@ export declare function toEntryStep<T extends ModelMessage>(step: HarnessStepRes
  * input is complete.
  */
 export declare function toEntryStream(stream: AsyncIterable<TextStreamPart<ToolSet>>, resolve: CallResolver): AsyncIterable<TextStreamPart<ToolSet>>;
-/**
- * Telemetry as the harness reports it. Integrations see each catalog call as
- * the call to its entry, and never see a stand-in run: a direct
- * call to the same entry never runs in the SDK.
- */
-export declare function toEntryTelemetry(telemetry: TelemetryOptions | undefined, resolve: CallResolver): TelemetryOptions | undefined;

package/dist/src/harness/execute-call.js

@@ -1 +1 @@
-import{CALL_TOOL_NAME,SKILL_TOOL_NAME}from"#protocol/catalog-tools.js";import{isWorkflowTool}from"#execution/tasks/model-step.js";import{toolCallModelOutput}from"#harness/tool-call-io.js";function historyCallNames(e){let t=new Map;for(let n of e)if(n.role===`assistant`&&Array.isArray(n.content))for(let e of n.content)e.type===`tool-call`&&t.set(e.toolCallId,e.toolName);return t}const DISPATCHED=Object.freeze({dispatched:!0});function dispatchesAfterStep(e){return e!==void 0&&isWorkflowTool(e.definition)}function runEntryCall(e,t){if(dispatchesAfterStep(e))return DISPATCHED;let{call:n,definition:r}=e;if(r.execute===void 0)throw Error(`"${n.toolName}" cannot run.`);return r.execute(n.input,t)}function stubbedCall(e,t){return t===void 0?e:dispatchesAfterStep(t)?void 0:t.call}function entryModelOutput(e,t,n){return t===DISPATCHED?Promise.resolve({type:`json`,value:null}):toolCallModelOutput(e,t,n)}function toEntryStep(e,t,n){let r=new Set((e.toolResults??[]).filter(e=>e.output===DISPATCHED).map(e=>e.toolCallId)),kept=e=>e.type!==`tool-result`||!r.has(e.toolCallId);return[{content:(e.content??[]).filter(kept).map(e=>toEntryPart(e,n)),finishReason:e.finishReason,invalidInputToolCallIds:e.invalidInputToolCallIds,providerMetadata:e.providerMetadata,response:{...e.response,messages:withoutResults(e.response.messages,kept)},text:e.text,toolCalls:(e.toolCalls??[]).map(e=>entryCallOf(e,n)),toolResults:(e.toolResults??[]).filter(kept).map(e=>entryCallOf(e,n)),usage:e.usage},withoutResults(t,kept)]}async function*toEntryStream(e,t){let n=new Set;for await(let r of e){if(r.type===`tool-input-start`&&(r.toolName===CALL_TOOL_NAME||r.toolName===SKILL_TOOL_NAME)){n.add(r.id);continue}r.type===`tool-input-delta`&&n.has(r.id)||r.type===`tool-input-end`&&n.delete(r.id)||(r.type!==`tool-result`||r.output!==DISPATCHED)&&(yield toEntryPart(r,t))}}function toEntryTelemetry(e,t){let n=e?.integrations;if(n===void 0)return e;let r=Array.isArray(n)?n:[n],i=new Set;return{...e,integrations:r.map(e=>entryToolExecutions(e,t,i))}}function entryToolExecutions(e,t,n){let{executeTool:r,onToolExecutionEnd:i,onToolExecutionStart:a}=e;if(!r&&!i&&!a)return e;let reportAsEntry=r=>i=>{let a=t(i.toolCall);if(dispatchesAfterStep(a)){n.add(i.toolCall.toolCallId);return}return r?.call(e,{...i,toolCall:a?.call??i.toolCall})};return Object.create(e,{executeTool:{value(t){return n.has(t.toolCallId)||r===void 0?t.execute():r.call(e,t)}},onToolExecutionEnd:{value:reportAsEntry(i)},onToolExecutionStart:{value:reportAsEntry(a)}})}function entryCallOf(e,t){return t(e)?.call??e}function toEntryPart(e,t){switch(e.type){case`tool-call`:case`tool-result`:case`tool-error`:return entryCallOf(e,t);case`tool-approval-request`:{let n=e;return n.toolCall===void 0?e:{...n,toolCall:entryCallOf(n.toolCall,t)}}default:return e}}function withoutResults(e,t){return e.flatMap(e=>{if(e.role!==`tool`)return[e];let n=e.content.filter(t);return n.length===0?[]:[{...e,content:n}]})}export{dispatchesAfterStep,entryModelOutput,historyCallNames,runEntryCall,stubbedCall,toEntryStep,toEntryStream,toEntryTelemetry};
+import{CALL_TOOL_NAME,SKILL_TOOL_NAME}from"#protocol/catalog-tools.js";function historyCallNames(e){let t=new Map;for(let n of e)if(n.role===`assistant`&&Array.isArray(n.content))for(let e of n.content)e.type===`tool-call`&&t.set(e.toolCallId,e.toolName);return t}function runEntryCall(e,t){let{call:n,definition:r}=e;if(r.execute===void 0)throw Error(`"${n.toolName}" cannot run.`);return r.execute(n.input,t)}function toEntryStep(e,t,n){return[{approvalRequests:e.approvalRequests,policyApprovals:e.policyApprovals,callId:e.callId,content:(e.content??[]).map(e=>toEntryPart(e,n)),finishReason:e.finishReason,invalidInputToolCallIds:e.invalidInputToolCallIds,providerMetadata:e.providerMetadata,response:e.response,signIns:e.signIns,text:e.text,toolCalls:(e.toolCalls??[]).map(e=>entryCallOf(e,n)),toolResults:(e.toolResults??[]).map(e=>entryCallOf(e,n)),usage:e.usage},[...t]]}async function*toEntryStream(e,t){let n=new Set;for await(let r of e){if(r.type===`tool-input-start`&&(r.toolName===CALL_TOOL_NAME||r.toolName===SKILL_TOOL_NAME)){n.add(r.id);continue}r.type===`tool-input-delta`&&n.has(r.id)||r.type===`tool-input-end`&&n.delete(r.id)||(yield toEntryPart(r,t))}}function entryCallOf(e,t){return t(e)?.call??e}function toEntryPart(e,t){switch(e.type){case`tool-call`:case`tool-result`:case`tool-error`:return entryCallOf(e,t);case`tool-approval-request`:{let n=e;return n.toolCall===void 0?e:{...n,toolCall:entryCallOf(n.toolCall,t)}}default:return e}}export{historyCallNames,runEntryCall,toEntryStep,toEntryStream};

package/dist/src/harness/execute-tool.d.ts

@@ -1,5 +1,5 @@
 import type { FlexibleSchema } from "ai";
-import type { ToolApproval } from "#approval/definition.js";
+import type { Approval } from "#approval/definition.js";
 import type { InternalToolDefinition, InternalToolLabelDefinition, ToolExecuteOptions } from "#tools/definition.js";
 import type { JsonValue } from "#shared/json.js";
 import type { PreparedToolBehavior } from "#tools/behavior.js";
@@ -9,7 +9,6 @@ import type { PreparedToolBehavior } from "#tools/behavior.js";
 export interface HarnessToolDefinition {
     readonly availableInSubagents?: boolean;
     readonly label?: InternalToolLabelDefinition;
-    readonly approvalKey?: (toolInput: Readonly<Record<string, unknown>>) => string;
     readonly behavior?: PreparedToolBehavior;
     /** Kept out of the model's tool list; the model reaches it through `eve__search` and `eve__tool`. */
     readonly deferred?: boolean;
@@ -23,7 +22,7 @@ export interface HarnessToolDefinition {
     readonly frameworkAction?: "task-cancel" | "task-wait";
     readonly inputSchema: FlexibleSchema;
     readonly name: string;
-    readonly approval?: ToolApproval;
+    readonly approval?: Approval;
     readonly outputSchema?: FlexibleSchema;
     /**
      * Advertise this tool only to the root session, hiding it from subagent
@@ -32,6 +31,11 @@ export interface HarnessToolDefinition {
      */
     readonly rootOnly?: boolean;
     readonly toModelOutput?: (output: unknown) => unknown;
+    /**
+     * The version of a dynamically sourced entry the model was offered. A call made from one
+     * version runs only while the entry still has it.
+     */
+    readonly version?: string;
     /** Present when this tool starts an associated durable workflow. */
     readonly workflowId?: string;
 }

package/dist/src/harness/facts.d.ts

@@ -0,0 +1,82 @@
+import type { SessionEvent } from "#protocol/session-event.js";
+import type { Cause, ErrorInfo, Usage } from "#protocol/session-events/envelope.js";
+import type { Fact, FactOf } from "#protocol/session-events/facts.js";
+import type { ModelOwner } from "#protocol/session-events/families/model.js";
+import type { SessionProjection } from "#protocol/session-projection.js";
+import type { ControlDelivery } from "#harness/types.js";
+import type { JsonObject } from "#shared/json.js";
+/** `value` without the fields it leaves undefined. */
+export declare function defined<T extends object>(value: T): T;
+/** A fact of `type`, without the fields its data leaves undefined. */
+export declare function fact<T extends Fact["type"]>(type: T, data: FactOf<T>["data"]): FactOf<T>;
+/** Spend as `usage.recorded` carries it: every count, zero when unreported, and a known cost. */
+export declare function usageOf(counts: Partial<Usage>): Usage;
+/** A failure as the stream reports it: its code and message, with its support id and remedy. */
+export declare function errorInfoOf(failure: {
+    readonly code: string;
+    readonly details?: JsonObject;
+    readonly message: string;
+}): ErrorInfo;
+/**
+ * The facts that end a session, in one commit, `session.ended` last. The writer closes what the
+ * session leaves open before it (`cascade`): an open turn ends cancelled, since its session ended,
+ * not it; `failure(view, turnId)` reads the session's error. Every reader stops there.
+ */
+export declare function sessionEndedFacts(projection: SessionProjection | undefined, ending: {
+    readonly outcome: "completed" | "failed";
+    readonly cause?: Cause;
+    /** The reset control that ends the session: admitted first, applied before the end. */
+    readonly control?: ControlDelivery;
+    /** Controls that arrived before the session started, which settle as it starts. */
+    readonly settles?: readonly ControlSettlement[];
+    readonly error?: ErrorInfo;
+}): SessionEvent[];
+/** How a turn ends. */
+export interface TurnEnding {
+    readonly outcome: FactOf<"turn.settled">["data"]["outcome"];
+    readonly cause?: Cause;
+    readonly error?: ErrorInfo;
+}
+/**
+ * The fact that ends a turn, the open one unless `turnId` names another, with what it replied.
+ * The writer closes what it leaves open before it, and settles the deliveries it answered after
+ * it (`cascade`).
+ */
+export declare function turnSettled(projection: SessionProjection, ending: TurnEnding, turnId?: string | undefined): SessionEvent[];
+/** A model run is asked for: a turn's, or a context change's summary run. */
+export declare function runRequested(owner: ModelOwner, runId: string): SessionEvent;
+/** The run's model was chosen and an attempt at its provider call begins. */
+export declare function runStarted(input: {
+    readonly runId: string;
+    readonly modelId: string;
+    readonly attempt: number;
+}): SessionEvent;
+/** How a run ended, and what it spent. */
+export interface RunSettlement {
+    readonly runId: string;
+    readonly outcome: FactOf<"model.settled">["data"]["outcome"];
+    readonly finishReason?: string;
+    readonly generationId?: string;
+    readonly error?: ErrorInfo;
+    readonly usage?: Usage;
+}
+/** The facts that settle a run: its end, then what it spent. */
+export declare function runSettledFacts(input: RunSettlement): SessionEvent[];
+/**
+ * What the agent behind a call spent, recorded against the call; without one, spend that arrived
+ * after its call settled, which no open entity owns (`delegated-late`).
+ */
+export declare function delegatedUsageFact(callId: string | undefined, usage: Usage): SessionEvent;
+/** A context change starts. */
+export declare function contextStarted(data: FactOf<"context.started">["data"]): SessionEvent;
+/** A control's admission: the delivery it arrived as, its kind, and who sent it. */
+export declare function controlAdmitted(delivery: ControlDelivery, control: string): SessionEvent;
+export interface ControlSettlement {
+    readonly control: ControlDelivery;
+    readonly source: "cancel" | "clear" | "compact";
+    readonly outcome: "ignored" | "applied";
+}
+/** A control that finishes without a step of its own, admitted and settled together. */
+export declare function controlSettled(settlement: ControlSettlement): SessionEvent[];
+/** A control or callback delivery's settlement once what it asked has happened. */
+export declare function deliveryApplied(deliveryId: string): SessionEvent;