eve 0.71.2 → 0.72.0

package/dist/src/harness/{approval-candidates.d.ts → hitl/candidates.d.ts}

@@ -12,7 +12,6 @@ interface ApprovalCandidateAuditRecord {
     readonly status: ApprovalCandidateStatus;
     readonly createdAt: number;
     readonly completedAt?: number;
-    readonly eventEmitted?: boolean;
     readonly expiresAt?: number;
     readonly reason?: string;
 }
@@ -24,11 +23,12 @@ export interface ApprovalResponderIdentity {
 }
 export interface ApprovalSettlementAuditRecord {
     readonly actor: ApprovalResponderIdentity;
+    /** The approver's full auth, which the approved call runs as. Absent for cancellations. */
+    readonly approver?: SessionAuthContext;
     readonly outcome: "allowed" | "cancelled";
     readonly requestId: string;
     readonly settledAt: number;
     readonly candidateId?: string;
-    readonly eventEmitted?: boolean;
 }
 export interface ActiveApprovalCandidate {
     readonly candidateId: string;
@@ -39,7 +39,6 @@ export interface ActiveApprovalCandidate {
     readonly createdAt: number;
     readonly expiresAt: number;
     readonly authorizationChallenges?: readonly AuthorizationChallenge[];
-    readonly pendingEventEmitted?: boolean;
 }
 interface ApprovalStateTransition {
     readonly changed: boolean;
@@ -55,21 +54,6 @@ export declare function createApprovalCandidate(input: {
     readonly responder: SessionAuthContext;
     readonly state: SessionStateMap | undefined;
 }): ApprovalStateTransition;
-/** Marks the pending candidate event as emitted. */
-export declare function markApprovalCandidatePendingEventEmitted(input: {
-    readonly candidateId: string;
-    readonly state: SessionStateMap | undefined;
-}): SessionStateMap | undefined;
-/** Marks a terminal candidate history event as emitted. */
-export declare function markApprovalCandidateHistoryEventEmitted(input: {
-    readonly candidateId: string;
-    readonly state: SessionStateMap | undefined;
-}): SessionStateMap | undefined;
-/** Marks a terminal settlement event as emitted. */
-export declare function markApprovalSettlementEventEmitted(input: {
-    readonly requestId: string;
-    readonly state: SessionStateMap | undefined;
-}): SessionStateMap | undefined;
 /** Marks a candidate as waiting on a private authorization challenge. */
 export declare function markApprovalCandidateAuthorizationRequired(input: {
     readonly authorizationChallenges: readonly AuthorizationChallenge[];
@@ -115,4 +99,15 @@ export declare function getApprovalAuditState(state: SessionStateMap | undefined
     readonly candidateHistory: readonly ApprovalCandidateAuditRecord[];
     readonly settlements: readonly ApprovalSettlementAuditRecord[];
 };
+export declare function sameResponder(a: Pick<SessionAuthContext, "authenticator" | "issuer" | "principalId" | "principalType">, b: Pick<SessionAuthContext, "authenticator" | "issuer" | "principalId" | "principalType">): boolean;
+/**
+ * The held turn moved on, steered or cancelled: the responders still checking one of its
+ * approvals stop, their candidates stale.
+ */
+export declare function retireActiveCandidates(state: SessionStateMap | undefined, input: {
+    readonly completedAt: number;
+    readonly reason: string;
+}): SessionStateMap | undefined;
+/** The full auth of whoever approved `requestId`, when it was allowed. */
+export declare function approverOfRequest(state: SessionStateMap | undefined, requestId: string): SessionAuthContext | undefined;
 export {};

package/dist/src/harness/hitl/candidates.js

@@ -0,0 +1 @@
+const APPROVAL_STATE_KEY=`eve.runtime.hitl.approvalState`;function createApprovalCandidate(e){let t=expireApprovalCandidates({now:e.createdAt,state:e.state}),n=readApprovalState(t);if(n.settlements[e.requestId]!==void 0)return{changed:!1,state:t};let r=e.responder;if(Object.values(n.activeCandidates).find(t=>t.requestId===e.requestId&&t.decision===e.decision&&sameResponder(t.responder,r))!==void 0)return{changed:!1,state:t};let i=n.activeCandidates[e.candidateIdPrefix]!==void 0||n.candidateHistory.some(t=>t.candidateId===e.candidateIdPrefix)?`${e.candidateIdPrefix}.${n.nextCandidateSequence.toString(36)}`:e.candidateIdPrefix;if(n.activeCandidates[i]!==void 0||n.candidateHistory.some(e=>e.candidateId===i))throw Error(`Approval candidate id collision: "${i}".`);let a={candidateId:i,createdAt:e.createdAt,decision:e.decision,expiresAt:e.expiresAt,requestId:e.requestId,responder:r,status:`pending`};return{changed:!0,state:writeApprovalState(t,{...n,activeCandidates:{...n.activeCandidates,[a.candidateId]:a},nextCandidateSequence:n.nextCandidateSequence+1})}}function markApprovalCandidateAuthorizationRequired(e){let t=readApprovalState(e.state),n=t.activeCandidates[e.candidateId];if(n===void 0)return e.state;let r={...n,authorizationChallenges:e.authorizationChallenges,expiresAt:e.expiresAt??n.expiresAt,status:`authorization-required`};return writeApprovalState(e.state,{...t,activeCandidates:{...t.activeCandidates,[e.candidateId]:r}})}function finishApprovalCandidate(e){let t=readApprovalState(e.state),n=t.activeCandidates[e.candidateId];if(n===void 0)return e.state;let r={...t.activeCandidates};return delete r[e.candidateId],writeApprovalState(e.state,{...t,activeCandidates:r,candidateHistory:[...t.candidateHistory,toCandidateAuditRecord({candidate:n,completedAt:e.completedAt,reason:e.reason,status:e.status})]})}function expireApprovalCandidates(e){let t=e.state,n=Object.values(readApprovalState(t).activeCandidates);for(let r of n)r.expiresAt>e.now||(t=finishApprovalCandidate({candidateId:r.candidateId,completedAt:e.now,state:t,status:`timed-out`}));return t}function settleAllowedCandidate(e){let t=expireApprovalCandidates({now:e.settledAt,state:e.state}),n=readApprovalState(t),r=n.activeCandidates[e.candidateId];if(r===void 0){let r=n.candidateHistory.find(t=>t.candidateId===e.candidateId);if((r&&n.settlements[r.requestId])!==void 0)return{changed:!1,state:t};throw Error(`Unknown approval candidate "${e.candidateId}".`)}return settleRequest({candidateId:r.candidateId,outcome:r.decision===`cancel`?`cancelled`:`allowed`,requestId:r.requestId,responder:r.responder,settledAt:e.settledAt,state:t})}function settleDirectApprovalResponse(e){let t=expireApprovalCandidates({now:e.settledAt,state:e.state});return settleRequest({outcome:e.outcome,requestId:e.requestId,responder:e.actor,settledAt:e.settledAt,state:t})}function getActiveApprovalCandidate(e,t){return readApprovalState(e).activeCandidates[t]}function getApprovalAuditState(e){let t=readApprovalState(e);return{activeCandidates:Object.values(t.activeCandidates),candidateHistory:t.candidateHistory,settlements:Object.values(t.settlements)}}function settleRequest(e){let t=readApprovalState(e.state);if(t.settlements[e.requestId]!==void 0)return{changed:!1,state:e.state};let n={actor:projectResponder(e.responder),...e.outcome===`allowed`&&{approver:e.responder},candidateId:e.candidateId,outcome:e.outcome,requestId:e.requestId,settledAt:e.settledAt},r={},i=[...t.candidateHistory];for(let n of Object.values(t.activeCandidates)){if(n.requestId!==e.requestId){r[n.candidateId]=n;continue}i.push(toCandidateAuditRecord({candidate:n,completedAt:e.settledAt,status:n.candidateId===e.candidateId?`allowed`:`stale`}))}let a={activeCandidates:r,candidateHistory:i,nextCandidateSequence:t.nextCandidateSequence,settlements:{...t.settlements,[e.requestId]:n}};return{changed:!0,state:writeApprovalState(e.state,a)}}function toCandidateAuditRecord(e){let{authorizationChallenges:t,responder:n,...r}=e.candidate;return{...r,completedAt:e.completedAt,responder:projectResponder(n),reason:e.reason,status:e.status}}function projectResponder(e){return{authenticator:e.authenticator,issuer:e.issuer,principalId:e.principalId,principalType:e.principalType}}function sameResponder(e,t){return e.authenticator===t.authenticator&&e.issuer===t.issuer&&e.principalId===t.principalId&&e.principalType===t.principalType}function readActiveCandidates(e){return Object.fromEntries(Object.entries(e).map(([e,t])=>[e,{...t,decision:t.decision===`cancel`?`cancel`:`approve`}]))}function readApprovalState(e){let t=e?.[APPROVAL_STATE_KEY];if(typeof t!=`object`||!t)return{activeCandidates:{},candidateHistory:[],nextCandidateSequence:0,settlements:{}};let n=t;return{activeCandidates:typeof n.activeCandidates==`object`&&n.activeCandidates!==null?readActiveCandidates(n.activeCandidates):{},candidateHistory:Array.isArray(n.candidateHistory)?n.candidateHistory:[],nextCandidateSequence:typeof n.nextCandidateSequence==`number`&&Number.isSafeInteger(n.nextCandidateSequence)&&n.nextCandidateSequence>=0?n.nextCandidateSequence:deriveNextCandidateSequence(n),settlements:typeof n.settlements==`object`&&n.settlements!==null?n.settlements:{}}}function deriveNextCandidateSequence(e){return Object.keys(e.activeCandidates??{}).length+(Array.isArray(e.candidateHistory)?e.candidateHistory.length:0)}function writeApprovalState(e,t){return{...e,[APPROVAL_STATE_KEY]:t}}function retireActiveCandidates(e,t){let n=e;for(let e of getApprovalAuditState(n).activeCandidates)n=finishApprovalCandidate({candidateId:e.candidateId,completedAt:t.completedAt,reason:t.reason,state:n,status:`stale`});return n}function approverOfRequest(e,t){return readApprovalState(e).settlements[t]?.approver}export{approverOfRequest,createApprovalCandidate,expireApprovalCandidates,finishApprovalCandidate,getActiveApprovalCandidate,getApprovalAuditState,markApprovalCandidateAuthorizationRequired,retireActiveCandidates,sameResponder,settleAllowedCandidate,settleDirectApprovalResponse};

package/dist/src/harness/hitl/coordinator.js

@@ -0,0 +1 @@
+import{contextStorage}from"#context/container.js";import{AuthKey,SessionKey}from"#context/keys.js";import{buildCallbackContext}from"#context/build-callback-context.js";import{buildApprovalResponseAuth,handleApprovalResponsePolicyError}from"#execution/tool-auth.js";import{suspendedSteps}from"#harness/session-machine/view.js";import{clearPendingAuthorization,getAuthorizationResult,getPendingAuthorization,isAuthorizationSignal}from"#harness/authorization.js";import{isApprovalRequest}from"#harness/input-request-class.js";import{createApprovalCandidate,expireApprovalCandidates,finishApprovalCandidate,getActiveApprovalCandidate,getApprovalAuditState,markApprovalCandidateAuthorizationRequired,sameResponder,settleAllowedCandidate,settleDirectApprovalResponse}from"#harness/hitl/candidates.js";async function coordinateApprovalDelivery(e){let t=e.now??Date.now(),n=getApprovalAuditState(e.session.state).activeCandidates.filter(e=>e.expiresAt<=t),r=n.flatMap(e=>e.authorizationChallenges?.map(e=>e.attemptId??e.candidateId??e.name)??[]),i=expireApprovalCandidates({now:t,state:e.session.state}),a={...e.session,state:clearPendingAuthorization(i,r)},o=getApprovalAuditState(a.state),s=suspendedSteps(a.state).filter(e=>e.requests.length>0),c=new Set(s.flatMap(e=>e.requests.map(e=>e.requestId))),l=o.settlements.filter(e=>c.has(e.requestId)),u=new Set(o.settlements.map(e=>e.requestId)),d=new Set([...u].filter(e=>!c.has(e))),f=hasResponseForRequest(e.stepInput,d),p=f?removeConsumedResponses(e.stepInput,d):e.stepInput;if(f&&l.length===0&&!hasMeaningfulInput(p))return deliveryResult(a,p,`park`);if(s.length===0)return deliveryResult(a,p);let m=p,h=new Set(s.flatMap(e=>e.responseAuthRequiredRequestIds??[])),g=s.flatMap(e=>e.requests),_=new Map(g.map(e=>[e.requestId,e])),v=[],y=[],b=new Set,x=!1,S=getApprovalAuditState(a.state).activeCandidates,C=[...m?.attributedInputResponses??[],...(m?.inputResponses??[]).map(e=>({auth:void 0,response:e}))];for(let{auth:e,response:n}of C){let r=_.get(n.requestId);if(r===void 0||!isApprovalRequest(r))continue;if(!h.has(n.requestId)){let r=contextStorage.getStore(),i=e===void 0?r?.get(AuthKey)??r?.get(SessionKey)?.auth.current??null:e,o=toCandidateDecision(n.optionId),s=suspendedSteps(a.state).find(e=>e.requests.some(e=>e.requestId===n.requestId))?.requester??null;if(i!==null&&o!==void 0&&s!==null&&!sameResponder(s,i)){b.add(n.requestId),y.push(`Only the person who requested this action can respond to this approval.`);continue}if(i!==null&&o!==void 0){let e=settleDirectApprovalResponse({actor:i,outcome:o===`approve`?`allowed`:`cancelled`,requestId:n.requestId,settledAt:t,state:a.state});a={...a,state:e.state},x||=e.changed}continue}b.add(n.requestId);let i=toCandidateDecision(n.optionId);if(i===void 0)continue;let o=e===void 0?buildCallbackContext().session.auth.current:e;if(o===null){y.push(`Authentication is required to respond to this approval.`);continue}let s=createApprovalCandidate({candidateIdPrefix:approvalCandidateIdPrefix(r.requestId,o,i),createdAt:t,decision:i,expiresAt:t+6e5,requestId:r.requestId,responder:o,state:a.state});a={...a,state:s.state},x||=s.changed}let w=removeConsumedResponses(m,b);if(b.size>0)return deliveryResult(a,w,x?`continue-coordination`:`continue`,[],y);if(x)return deliveryResult(a,w,`continue`,[],y);let T=new Set(getPendingAuthorization(a.state)?.challenges.map(e=>e.name)??[]);for(let n of S){if(n.status===`authorization-required`){let e=n.authorizationChallenges??[];if(!e.some(e=>getAuthorizationResult(e.name)!==void 0)){v.push(...e.filter(e=>!T.has(e.name)));continue}}let r=_.get(n.requestId);if(r===void 0||getActiveApprovalCandidate(a.state,n.candidateId)===void 0)continue;let i=await authorizeCandidate({candidateId:n.candidateId,decision:n.decision,now:t,request:r,responder:n.responder,session:a,tools:await e.prepareTools?.(r)??e.tools});a=i.session,x||=i.didCommit,v.push(...i.challenges);let o=getApprovalAuditState(a.state).settlements.find(e=>e.requestId===n.requestId);o!==void 0&&l.push(o)}let E=appendSettledResponses(w,l);return(x||n.length>0)&&getApprovalAuditState(a.state).activeCandidates.length===0?deliveryResult(a,E,`responses-completed`):l.length>0?deliveryResult(a,E,`continue`):x?deliveryResult(a,E,`continue-coordination`):deliveryResult(a,E,v.length>0?`authorization-required`:`continue`,v)}async function authorizeCandidate(e){let t={...e.session,state:expireApprovalCandidates({now:e.now,state:e.session.state})};if(getActiveApprovalCandidate(t.state,e.candidateId)===void 0)return{challenges:[],didCommit:!1,session:t};let n=e.tools.get(e.request.action.toolName)?.approval,r=n!==void 0&&typeof n!=`function`?n.response:void 0;if(r===void 0)return failCandidate({...e,reason:`Approval authorization is temporarily unavailable. Please try again.`,session:t});try{let n=buildCallbackContext(),i=await withAuthorizerTimeout(r({auth:buildApprovalResponseAuth({responder:e.responder,scope:e.candidateId}),request:{callId:e.request.action.callId,requestId:e.request.requestId,principal:suspendedSteps(t.state).find(t=>t.requests.some(t=>t.requestId===e.request.requestId))?.requester??null,toolInput:e.request.action.input,toolName:e.request.action.toolName},response:{decision:e.decision,principal:e.responder},session:{id:n.session.id,initiator:n.session.auth.initiator,parent:n.session.parent,turn:n.session.turn}}));if(i.status===`rejected`)return t={...t,state:finishApprovalCandidate({candidateId:e.candidateId,completedAt:e.now,reason:i.reason,state:t.state,status:`rejected`})},{challenges:[],didCommit:!0,session:t};if(i.status!==`allowed`)return failCandidate({...e,session:t});let a=settleAllowedCandidate({candidateId:e.candidateId,settledAt:e.now,state:t.state});return{challenges:[],didCommit:a.changed,session:{...t,state:a.state}}}catch(n){let r=await handleApprovalResponsePolicyError(n).catch(()=>void 0);if(isAuthorizationSignal(r)){let n=r.challenges.map(e=>Date.parse(e.challenge.expiresAt??``)).filter(Number.isFinite).sort((e,t)=>e-t)[0];return t={...t,state:markApprovalCandidateAuthorizationRequired({authorizationChallenges:r.challenges.map(t=>({...t,candidateId:e.candidateId})),candidateId:e.candidateId,expiresAt:n,state:t.state})},{challenges:r.challenges.map(t=>({...t,candidateId:e.candidateId})),didCommit:!0,session:t}}return failCandidate({...e,session:t})}}function failCandidate(e){let t=e.reason??`We couldn’t verify your response. Please try again.`;return{challenges:[],didCommit:!0,session:{...e.session,state:finishApprovalCandidate({candidateId:e.candidateId,completedAt:e.now,reason:t,state:e.session.state,status:`failed`})}}}function hasResponseForRequest(e,t){return e?.attributedInputResponses?.some(({response:e})=>t.has(e.requestId))===!0||e?.inputResponses?.some(e=>t.has(e.requestId))===!0}function hasMeaningfulInput(e){return e?.message!==void 0||(e?.attributedInputResponses?.length??0)>0||(e?.inputResponses?.length??0)>0||(e?.runtimeActionResults?.length??0)>0}function appendSettledResponses(e,t){if(t.length===0)return e;let n=new Set([...(e?.inputResponses??[]).map(e=>e.requestId),...(e?.attributedInputResponses??[]).map(({response:e})=>e.requestId)]),r=t.filter(e=>!n.has(e.requestId));return r.length===0?e:{...e,inputResponses:[...e?.inputResponses??[],...r.map(e=>({optionId:e.outcome===`allowed`?`approve`:`cancel`,requestId:e.requestId}))]}}function removeConsumedResponses(e,t){if(e===void 0)return;let n=(e.attributedInputResponses??[]).filter(({response:e})=>!t.has(e.requestId)),r=[...(e.inputResponses??[]).filter(e=>!t.has(e.requestId)),...n.map(({response:e})=>e)];return{...e,attributedInputResponses:void 0,inputResponses:r}}function deliveryResult(e,t,n=`continue`,r=[],i=[]){return{challenges:r,feedback:i,kind:n,session:e,stepInput:t}}function toCandidateDecision(e){if(e===`approve`)return`approve`;if(e===`cancel`||e===`deny`)return`cancel`}function approvalCandidateIdPrefix(e,t,n){let r=[t.authenticator,t.issuer??``,t.principalType,t.principalId].join(`:`),i=`${encodeCandidateIdPart(e)}.${encodeCandidateIdPart(r)}`;return n===`approve`?i:`${i}.${encodeCandidateIdPart(n)}`}function encodeCandidateIdPart(e){return Array.from(e,e=>e.codePointAt(0).toString(36)).join(`-`)}async function withAuthorizerTimeout(e){let t;try{return await Promise.race([Promise.resolve(e),new Promise((e,n)=>{t=setTimeout(()=>n(Error(`Approval response authorizer timed out.`)),1e4)})])}finally{t!==void 0&&clearTimeout(t)}}export{coordinateApprovalDelivery};

package/dist/src/harness/hitl/delivery.d.ts

@@ -0,0 +1,37 @@
+import type { UserContent } from "ai";
+import type { StepInput } from "#harness/types.js";
+import type { InputResponse } from "#shared/input.js";
+import type { SuspendedStep } from "#harness/session-machine/state.js";
+import type { SessionView } from "#harness/session-machine/view.js";
+/**
+ * The delivery a step answers with: input queued behind earlier work joins it, unless the
+ * runtime still runs calls (queued input waits for them). A stale answer, to a request that
+ * already closed, never authorizes anything: a stale session-limit answer is dropped, and
+ * any other becomes plain text the model reads.
+ */
+export declare function deliver(view: SessionView, input: StepInput | undefined, options: {
+    readonly takeQueued: boolean;
+}): {
+    readonly input?: StepInput;
+    readonly displayMessage?: string | UserContent;
+    readonly takeQueued: boolean;
+};
+export type ResolvedStepInput = StepInput & {
+    readonly messageConsumed?: boolean;
+};
+/**
+ * Plain text answers the only pending batch, unless a response policy must decide it. A channel
+ * that wraps the typed text in an envelope for the model attaches the text itself to answer with.
+ * The answer consumes the message and the context the channel sent about it: the model reads
+ * neither.
+ */
+export declare function resolveTextInput(batch: Pick<SuspendedStep, "requests" | "responseAuthRequiredRequestIds">, stepInput: StepInput | undefined): ResolvedStepInput | undefined;
+export declare function canonicalize(responses: readonly InputResponse[]): readonly InputResponse[];
+export declare function hasInput(input: StepInput | undefined): boolean;
+export declare function isEmptyInput(input: StepInput): boolean;
+export declare function withoutResponses(input: ResolvedStepInput | undefined): StepInput | undefined;
+/** The turn's own input: the message a plain-text answer didn't consume, and its context. */
+export declare function turnInputOnly(input: ResolvedStepInput | undefined): StepInput | undefined;
+/** What of the input isn't the turn's: answers, and the output the session asks for. */
+export declare function withoutTurnInput(input: ResolvedStepInput | undefined): ResolvedStepInput;
+export declare function compactInput(input: ResolvedStepInput | undefined): ResolvedStepInput;

package/dist/src/harness/hitl/delivery.js

@@ -0,0 +1 @@
+import{attachClientContext,readClientContext}from"#internal/client-context.js";import{readAnswerText}from"#internal/input-text.js";import{resolveTextToResponses}from"#channel/resolve-text.js";import{coalesceTurnInputs}from"#harness/messages.js";import{ownOpenRequestIds}from"#harness/session-machine/transitions.js";import{convertStaleResponsesToUserMessage,dropStaleSessionLimitContinuationResponses}from"#harness/hitl/stale-responses.js";function deliver(e,t,n){let r=n.takeQueued?e.turn.queued:void 0,i=r===void 0?t:t===void 0?r:coalesceTurnInputs(r,t),a=new Set([...ownOpenRequestIds(e),...e.turn.suspended.flatMap(e=>e.requests.map(e=>e.requestId))]),o=new Map(Object.values(e.projection.inputs).map(e=>[e.request.requestId,e.request])),s=convertStaleResponsesToUserMessage({pendingRequestIds:a,requests:o,stepInput:dropStaleSessionLimitContinuationResponses({pendingRequestIds:a,stepInput:i})});return s.kind===`converted`?{displayMessage:s.displayMessage,input:s.stepInput,takeQueued:n.takeQueued}:{input:s.stepInput,takeQueued:n.takeQueued}}function resolveTextInput(e,t){let n=readAnswerText(t);if(t===void 0||n===void 0)return t;let r=new Set(e.requests.map(e=>e.requestId));if(t.inputResponses?.some(e=>r.has(e.requestId)))return t;let i=new Set(e.responseAuthRequiredRequestIds??[]),a=resolveTextToResponses(n,e.requests.filter(e=>!i.has(e.requestId)));return a.length===0?t:compactInput({...t,context:void 0,inputResponses:[...t.inputResponses??[],...a],message:void 0,messageConsumed:!0})}function canonicalize(e){let t=new Map;for(let n of e)t.set(n.requestId,n);return[...t.values()]}function hasInput(e){return e?.message!==void 0||(e?.inputResponses?.length??0)>0}function isEmptyInput(e){return Object.keys(compactInput(e)).length===0}function withoutResponses(e){if(e===void 0)return;let{attributedInputResponses:t,inputResponses:n,messageConsumed:r,...i}=e;return i}function turnInputOnly(e){if(e===void 0)return;let t={};(e.context?.length??0)>0&&(t.context=e.context),e.message!==void 0&&e.messageConsumed!==!0&&(t.message=e.message);let n=attachClientContext(t,readClientContext(e));return isEmptyInput(n)?void 0:n}function withoutTurnInput(e){let t={};return(e?.inputResponses?.length??0)>0&&(t.inputResponses=e.inputResponses),e?.outputSchema!==void 0&&(t.outputSchema=e.outputSchema),t}function compactInput(e){if(e===void 0)return{};let t={};return(e.context?.length??0)>0&&(t.context=e.context),(e.inputResponses?.length??0)>0&&(t.inputResponses=e.inputResponses),e.message!==void 0&&(t.message=e.message),e.messageConsumed===!0&&(t.messageConsumed=!0),e.outputSchema!==void 0&&(t.outputSchema=e.outputSchema),attachClientContext(t,readClientContext(e))}export{canonicalize,compactInput,deliver,hasInput,isEmptyInput,resolveTextInput,turnInputOnly,withoutResponses,withoutTurnInput};

package/dist/src/harness/hitl/index.d.ts

@@ -0,0 +1,49 @@
+import type { ModelMessage } from "ai";
+import type { resolveInlineAuthorizationInterrupt } from "#harness/inline-tool-authorization.js";
+import type { StepCoordinates } from "#harness/session-machine/view.js";
+import type { Step } from "#harness/step/context.js";
+import type { HarnessSessionBase, StepResult } from "#harness/types.js";
+import type { RuntimeWorkflowTaskRequest } from "#shared/action-types.js";
+import type { InputRequest } from "#shared/input.js";
+export { acceptHumanInput, admitApprovedWork, dispatchApprovedWorkflows, hasApprovedWork, runApprovedLocalCalls, type ApprovedWork, type HumanInputIntake, } from "./intake.js";
+export { extractToolApprovalInputRequests } from "#harness/input-extraction.js";
+export { hasRunnableQueue } from "./approvals.js";
+/**
+ * A model step made calls that need a person's approval: the step parks on them beside any
+ * runtime calls it made. The turn holds for the answers, unless the runtime runs calls for it
+ * meanwhile, or queued input can already run.
+ */
+export declare function parkOnApprovals(step: Step, input: {
+    readonly event: StepCoordinates;
+    readonly messages: readonly ModelMessage[];
+    readonly requests: readonly InputRequest[];
+    readonly tasks: readonly RuntimeWorkflowTaskRequest[];
+    /** The step made calls the runtime runs, so the turn waits on them instead. */
+    readonly waitsOnRuntime: boolean;
+}): Promise<StepResult>;
+/**
+ * A call the model step ran needs a sign-in: the calls that need it stop, and the model calls
+ * them again once it completes.
+ */
+export declare function stopForToolSignIn(step: Step, interrupt: NonNullable<ReturnType<typeof resolveInlineAuthorizationInterrupt>>): Promise<StepResult>;
+/**
+ * The gate before each model call. A spent budget asks whether to continue when someone can
+ * answer, and fails the session otherwise. `messages` are the step's: they park with the prompt,
+ * so the input that triggered it survives into the turn that resumes.
+ */
+export declare function enforceBudget(step: Step, messages: readonly ModelMessage[]): Promise<StepResult | undefined>;
+/**
+ * What a model call needs from the approvals: the keys `once()` approvals granted, and a note on
+ * the calls still awaiting approval, which a message may revise.
+ */
+export declare function humanInputContext(step: Step): {
+    readonly approvedTools: ReadonlySet<string>;
+    readonly pendingApprovalsNote?: string;
+};
+/**
+ * Drops what a cleared context owned: sign-in attempts and responders' approval progress. `clear`
+ * reported each close; relay routes for live tasks stay.
+ */
+export declare function discardClearedHumanInput<T extends HarnessSessionBase>(session: T): T;
+/** The cancelled turn's responders stop checking its approvals: their candidates stale. */
+export declare function retireCancelledCandidates<T extends HarnessSessionBase>(session: T): T;

package/dist/src/harness/hitl/index.js

@@ -0,0 +1 @@
+import{renderPendingApprovalsInstruction}from"./approval-prompt.js";import{grantedApprovalKeys,hasRunnableQueue,parkOnApprovals as parkOnApprovals$1,requestLimit,requireSignIn}from"./approvals.js";import{retireActiveCandidates}from"./candidates.js";import{checkSessionUsageLimit}from"./budget.js";import{acceptHumanInput,admitApprovedWork,dispatchApprovedWorkflows,hasApprovedWork,holdForInput,runApprovedLocalCalls}from"./intake.js";import{AuthKey,SessionKey}from"#context/keys.js";import{validateHarnessModelMessages}from"#harness/messages.js";import{fail}from"#harness/session-machine/transitions.js";import{clearPendingAuthorization}from"#harness/authorization.js";import{buildResponseAuthorizationTools}from"#context/build-dynamic-tools.js";import{extractToolApprovalInputRequests}from"#harness/input-extraction.js";async function parkOnApprovals(e,t){let n=parkOnApprovals$1(e.view(),{event:t.event,messages:t.messages,requests:t.requests,tasks:t.tasks,requester:currentRequester(e),responseAuthRequiredRequestIds:responsePolicyRequestIds(e,t.requests)});return await e.apply(n,[...e.session.history,...n.commit??[]]),t.waitsOnRuntime?{next:null,session:e.session}:hasRunnableQueue(e.view())?{next:e.runStep,session:e.session}:holdForInput(e)}async function stopForToolSignIn(e,t){return e.session={...e.session,history:validateHarnessModelMessages(t.history)},await e.apply(requireSignIn(e.view(),{callIdsByName:t.callIdsByName,challenges:t.challenges}),e.session.history),{held:{kind:`request`},next:null,session:e.session}}async function enforceBudget(e,t){let n=checkSessionUsageLimit({canAsk:e.emit!==void 0&&e.config.capabilities?.requestInput===!0,session:e.session,turnSequence:e.position().sequence});if(n.kind!==`within`)return n.kind===`ask`?(e.session={...e.session,history:validateHarnessModelMessages([...t])},await e.apply(requestLimit(e.view(),{request:n.request}),e.session.history),{next:null,session:e.session}):(await e.apply(fail(e.view(),{code:n.code,details:n.details,message:n.message,terminal:{sessionId:e.session.sessionId}})),{next:{done:!0,output:``},session:e.session})}function humanInputContext(e){let t=e.view(),n=responseTools(e);return{approvedTools:grantedApprovalKeys(t,e=>n.get(e.action.toolName)?.approvalKey?.(e.action.input)),pendingApprovalsNote:renderPendingApprovalsInstruction(t.turn.suspended.flatMap(e=>e.requests))}}function discardClearedHumanInput(e){let{"eve.runtime.hitl.approvalState":t,...n}=clearPendingAuthorization(e.state)??{};return{...e,state:Object.keys(n).length>0?n:void 0}}function responseTools(e){return buildResponseAuthorizationTools({authoredTools:e.config.tools,context:e.ctx})}function responsePolicyRequestIds(e,t){let n=responseTools(e);return t.filter(e=>{let t=n.get(e.action.toolName)?.approval;return t!==void 0&&typeof t!=`function`&&t.response!==void 0}).map(e=>e.requestId)}function currentRequester(e){return e.ctx?.get(AuthKey)??e.ctx?.get(SessionKey)?.auth.current??null}function retireCancelledCandidates(e){return{...e,state:retireActiveCandidates(e.state,{completedAt:Date.now(),reason:`Cancelled.`})}}export{acceptHumanInput,admitApprovedWork,discardClearedHumanInput,dispatchApprovedWorkflows,enforceBudget,extractToolApprovalInputRequests,hasApprovedWork,hasRunnableQueue,humanInputContext,parkOnApprovals,retireCancelledCandidates,runApprovedLocalCalls,stopForToolSignIn};

package/dist/src/harness/hitl/intake.d.ts

@@ -0,0 +1,67 @@
+import type { SuspendedStep } from "#harness/session-machine/view.js";
+import { type Step } from "#harness/step/context.js";
+import type { HarnessToolMap, StepInput, StepResult } from "#harness/types.js";
+import type { InstrumentationAttempt } from "#instrumentation/runtime.js";
+/**
+ * What a delivery does before its turn runs. `stop` ends the step: the answers left nothing to
+ * run yet, so the turn holds, or a responder must sign in first. `run` carries the turn's input,
+ * without the answers the session took, and the approved work that runs once the turn opens.
+ */
+export type HumanInputIntake = {
+    readonly kind: "stop";
+    readonly result: StepResult;
+} | {
+    readonly kind: "run";
+    readonly input: StepInput | undefined;
+    /** The message the turn receives, as the stream shows it. */
+    readonly message?: StepInput["message"];
+    /** A plain-text answer consumed the message, so the model never reads it. */
+    readonly consumedMessage: boolean;
+    /** The delivery carries input for a turn. */
+    readonly opensTurn: boolean;
+    readonly approved: ApprovedWork;
+};
+/**
+ * What the delivery decided: the session-limit answer, and the tools approved calls run with.
+ * The calls themselves wait in their suspended steps until the turn's next model step.
+ */
+export interface ApprovedWork {
+    readonly limit?: {
+        readonly granted: boolean;
+    };
+    /** The tools a parked step's calls run with: those of the step that asked. */
+    readonly toolsOf: (step: SuspendedStep | undefined) => HarnessToolMap;
+}
+/**
+ * Answers what the session asked: approval answers pass the response policies, and the session
+ * decides which parked steps they resolve. Runs before the delivery joins its turn, so approved
+ * calls get the tools of the step that asked.
+ */
+export declare function acceptHumanInput(step: Step, input: StepInput | undefined, options: {
+    readonly takeQueued: boolean;
+}): Promise<HumanInputIntake>;
+/**
+ * Admits what a delivery decided before its turn calls the model: a session-limit answer grants a
+ * fresh budget or ends the turn tree, and a step its answers completed, as denials do, commits.
+ */
+export declare function admitApprovedWork(step: Step, work: ApprovedWork): Promise<void>;
+/** Whether a suspended step holds approved calls that haven't run. */
+export declare function hasApprovedWork(step: Step): boolean;
+/**
+ * Approved workflow and agent calls join the runs their steps wait on, once the step has started
+ * and before its budget check. Returns whether any did: the turn then waits on the runtime, and
+ * approved local calls run after their results arrive.
+ */
+export declare function dispatchApprovedWorkflows(step: Step, work: ApprovedWork): Promise<boolean>;
+/**
+ * Runs approved local calls once the step has started and its budget allows it, as the AI SDK ran
+ * them before the model read their results: each with the tools of the step that asked, as
+ * whoever approved it, and in the step's first attempt. A call that needs a sign-in holds the
+ * turn.
+ */
+export declare function runApprovedLocalCalls(step: Step, work: ApprovedWork, setAttemptScope: (scope: InstrumentationAttempt | undefined) => void): Promise<StepResult | undefined>;
+/**
+ * The held turn holds on, for a person to act on its sign-in, approval, or the session-limit
+ * prompt; the session resumes it when they do.
+ */
+export declare function holdForInput(step: Step): Promise<StepResult>;

package/dist/src/harness/hitl/intake.js

@@ -0,0 +1 @@
+import{deliver,turnInputOnly,withoutTurnInput}from"./delivery.js";import{answer,approvingSteps,deferInput,dispatch,grantedApprovalKeys,requireSignIn,stepForRequest,withdrawSignIns}from"./approvals.js";import{getApprovalAuditState,retireActiveCandidates}from"./candidates.js";import{approversOf,setApprovedCallCallers}from"./approved-call-callers.js";import{runApprovedCalls}from"./approved-calls.js";import{coordinateApprovalDelivery}from"./coordinator.js";import{validateHarnessModelMessages}from"#harness/messages.js";import{activeTurnId}from"#harness/session-machine/view.js";import{bumpSessionRuntimeUsageLimits}from"#harness/turn-tag-state.js";import{collectDeferredCalls}from"#harness/coordination.js";import{approvedCalls,completeSignIn,hold,settle}from"#harness/session-machine/transitions.js";import{stepStartedForResolvers}from"#harness/session-machine/resolver-events.js";import{SessionLimitDeclinedError}from"#harness/turn-cancellation.js";import{resolveInlineAuthorizationInterrupt}from"#harness/inline-tool-authorization.js";import{buildResponseAuthorizationTools}from"#context/build-dynamic-tools.js";import{openTurn}from"#harness/step/context.js";import{prepareTurnInput}from"#harness/step/intake.js";import{placeTurnInput}from"#harness/step/prompt.js";async function acceptHumanInput(e,t,n){let{config:r,ctx:i}=e,a=r.signInCompletions??[];a.length>0&&await e.apply(completeSignIn(e.view(),{completions:a}));let o=deliver(e.view(),t,n),s=t?.message!==void 0||o.displayMessage!==void 0?await withdrawSteeredSignIns(e,o.input):o.input,c=new Map,l,restoreTools=async t=>{let n=t?.event??e.position(),a=`${n.turnId}:${n.stepIndex}`,o=c.get(a);if(o!==void 0&&l===a)return o;t!==void 0&&await r.prepareApprovalTurn?.(t.event),i!==void 0&&await r.resolveStepDynamicTools?.({ctx:i,event:stepStartedForResolvers({modelId:e.session.agent.modelReference?.id??`dynamic`,sequence:n.sequence,stepIndex:n.stepIndex,turnId:n.turnId}),messages:e.projectHistory(e.session.history)});let s=buildResponseAuthorizationTools({authoredTools:r.tools,context:i});return c.set(a,s),l=a,s},toolsOf=e=>(e&&c.get(`${e.event.turnId}:${e.event.stepIndex}`))??r.tools,u=e.view().signIns,d=await coordinateApprovalDelivery({session:e.session,stepInput:s,tools:r.tools,prepareTools:t=>restoreTools(stepForRequest(e.view(),t.requestId))});e.session=d.session;for(let t of approvingSteps(e.view(),d.stepInput))await restoreTools(t);let f=answer(e.view(),{approvalKey:t=>toolsOf(stepForRequest(e.view(),t.requestId)).get(t.action.toolName)?.approvalKey?.(t.action.input),delivery:d.stepInput,policy:{...d,audit:getApprovalAuditState(e.session.state),challengesAtStart:u},takeQueued:o.takeQueued});for(let t of f.resolved)await e.instrumentation?.publishInputResolutions({batch:t,sessionId:e.session.sessionId});await e.apply(f);let stop=e=>({kind:`stop`,result:e});switch(f.next){case`park`:return e.view().projection.activeTurnId===void 0?stop({next:null,session:e.session}):stop(await holdForInput(e));case`repeat`:return stop({next:e.runStep,session:e.session});case`sign-in`:return await e.apply(requireSignIn(e.view(),{challenges:d.challenges,queued:d.stepInput})),stop(held(e));case`defer-message`:{if(approvedCalls(e.view().turn).length>0){let t=turnInputOnly(f.input);return t!==void 0&&await e.apply(deferInput(e.view(),t)),stop({next:null,session:e.session})}let t=await prepareTurnInput(e,f.input,{consumedMessage:!1}),n=await openTurn(e,{input:[...t.ephemeral,...t.messages],message:o.displayMessage??f.input?.message});return n===void 0?(e.session={...e.session,history:validateHarnessModelMessages(placeTurnInput(e,e.session.history,t))},stop(await holdForInput(e))):stop(n)}}for(let t of e.view().turn.suspended)(t.approved?.length??0)>0&&await restoreTools(t);let p=f.resolved.some(e=>e.inputs.some(e=>e.request.kind===`tool-approval`)&&e.inputs.every(e=>e.response!==void 0))||approvedCalls(e.view().turn).length>0?turnInputOnly(f.input):void 0;return p!==void 0&&await e.apply(deferInput(e.view(),p)),{approved:{limit:f.limit,toolsOf},consumedMessage:f.consumedMessage===!0,input:p===void 0?f.input:withoutTurnInput(f.input),message:f.consumedMessage===!0||p!==void 0?void 0:o.displayMessage??o.input?.message,kind:`run`,opensTurn:hasTurnInput(o.input)||hasTurnInput(d.stepInput)}}async function admitApprovedWork(e,t){if(t.limit!==void 0){if(!t.limit.granted)throw new SessionLimitDeclinedError;e.session=bumpSessionRuntimeUsageLimits(e.session)}await e.apply(settle(e.view(),{results:[]}))}function hasApprovedWork(e){return approvedCalls(e.view().turn).length>0}async function dispatchApprovedWorkflows(e,t){let n=[],r=[];for(let i of e.view().turn.suspended){let a=t.toolsOf(i),o=(i.approved??[]).filter(e=>a.get(e.action.toolName)?.workflowId!==void 0);if(o.length===0)continue;r.push(...o);let s=collectDeferredCalls({session:e.session,toolCalls:o.map(({action:e})=>({input:e.input,toolCallId:e.callId,toolName:e.toolName})),tools:a,turnId:i.event.turnId});e.session=s.session,n.push(...s.workflowRequests)}return n.length!==0&&(await e.apply(dispatch(e.view(),{approvers:approversOf(r,e.session.state),tasks:n})),!0)}async function runApprovedLocalCalls(e,t,n){let r=e.view().turn.suspended.map(e=>{let n=t.toolsOf(e);return{requests:(e.approved??[]).filter(e=>n.get(e.action.toolName)?.workflowId===void 0),tools:n}}),i=r.flatMap(({requests:e})=>e);e.frameworkToolNames=new Set(r.flatMap(({tools:e})=>[...e].filter(([,e])=>e.frameworkTool===!0).map(([e])=>e))),setApprovedCallCallers(i,e.session.state);let a=e.position(),o=i.length===0?void 0:e.instrumentation?.prepareAttempt({isFrameworkTool:t=>e.frameworkToolNames.has(t),attemptIndex:0,stepIndex:a.stepIndex,turnId:activeTurnId(a)});o!==void 0&&n(o.scope);let s=await Promise.all(r.map(async({requests:t,tools:n})=>t.length===0?{settled:[],toolResults:[]}:await runApprovedCalls({abortSignal:e.config.abortSignal,approvedTools:grantedApprovalKeys(e.view(),e=>n.get(e.action.toolName)?.approvalKey?.(e.action.input)),messages:e.projectHistory(e.session.history),position:e.position(),publish:e.publish,requests:t,telemetry:o?.telemetry,tools:n})));await e.apply(settle(e.view(),{results:s.flatMap(e=>e.settled)}));let c=resolveInlineAuthorizationInterrupt({messages:[],toolResults:s.flatMap(e=>e.toolResults)});if(c!==void 0)return await e.apply(requireSignIn(e.view(),{callIdsByName:c.callIdsByName,challenges:c.challenges}),e.session.history),held(e)}const STEERED_SIGN_IN_REASON=`Cancelled because a new message arrived.`;async function holdForInput(e){return await e.apply(hold(e.view(),{on:`input`})),held(e)}function held(e){return{held:{kind:`request`},next:null,session:e.session}}async function withdrawSteeredSignIns(e,t){let n=[...new Set(e.view().signIns.map(e=>e.name))];return e.session={...e.session,state:retireActiveCandidates(e.session.state,{completedAt:Date.now(),reason:STEERED_SIGN_IN_REASON})},n.length===0?t:(await e.apply(withdrawSignIns(e.view(),STEERED_SIGN_IN_REASON)),{...t,context:[...t?.context??[],`Sign-in to ${n.join(`, `)} was cancelled because the user sent a new message instead. Ask to sign in again only if the new message still needs it.`]})}function hasTurnInput(e){return e===void 0?!1:e.message!==void 0||(e.inputResponses?.length??0)>0}export{acceptHumanInput,admitApprovedWork,dispatchApprovedWorkflows,hasApprovedWork,holdForInput,runApprovedLocalCalls};

package/dist/src/harness/{stale-input-responses.d.ts → hitl/stale-responses.d.ts}

@@ -1,5 +1,6 @@
-import type { ModelMessage, UserContent } from "ai";
+import type { UserContent } from "ai";
 import type { StepInput } from "#harness/types.js";
+import type { InputRequest } from "#shared/input.js";
 type StaleResponseConversion = {
     readonly kind: "unchanged";
     readonly stepInput?: StepInput;
@@ -31,14 +32,15 @@ export declare function dropStaleSessionLimitContinuationResponses(input: {
  * Responses for pending requests stay structured; stale responses become
  * plain user-message text. A stale response never reaches structured HITL
  * processing, so a stale approval cannot authorize an earlier tool call.
- * Request details recovered from history are best-effort model context.
+ * Request details the session still knows are best-effort model context.
  *
  * Assumes {@link dropStaleSessionLimitContinuationResponses} already ran:
  * stale continuation answers must never reach this conversion.
  */
 export declare function convertStaleResponsesToUserMessage(input: {
-    readonly history: readonly ModelMessage[];
     readonly pendingRequestIds: ReadonlySet<string>;
+    /** Requests the session still knows, by id, closed ones included. */
+    readonly requests: ReadonlyMap<string, InputRequest>;
     readonly stepInput?: StepInput;
 }): StaleResponseConversion;
 export {};

package/dist/src/harness/hitl/stale-responses.js

@@ -0,0 +1,3 @@
+import{appendUserContent,normalizeUserContent}from"#harness/messages.js";import{isSessionLimitContinuationRequestId}from"#harness/hitl/budget-request.js";function dropStaleSessionLimitContinuationResponses(e){if(e.stepInput===void 0)return;let t=e.stepInput.inputResponses??[],n=e.stepInput.attributedInputResponses??[],keep=t=>e.pendingRequestIds.has(t)||!isSessionLimitContinuationRequestId(t),r=t.filter(e=>keep(e.requestId)),i=n.filter(({response:e})=>keep(e.requestId));if(r.length===t.length&&i.length===n.length)return e.stepInput;let{attributedInputResponses:a,inputResponses:o,...s}=e.stepInput,c=s;return r.length>0&&(c.inputResponses=r),i.length>0&&(c.attributedInputResponses=i),c}function convertStaleResponsesToUserMessage(e){if(e.stepInput===void 0)return{kind:`unchanged`};let t=e.stepInput.inputResponses??[],n=e.stepInput.attributedInputResponses??[];if(t.length===0&&n.length===0)return{kind:`unchanged`,stepInput:e.stepInput};let r=[],i=[],a=[];for(let n of t)(e.pendingRequestIds.has(n.requestId)?r:a).push(n);for(let t of n)e.pendingRequestIds.has(t.response.requestId)?i.push(t):a.push(t.response);if(a.length===0)return{kind:`unchanged`,stepInput:e.stepInput};let{requests:o}=e,s=appendOptionalUserContent(e.stepInput.message,formatModelMessage(a,o)),c=appendOptionalUserContent(e.stepInput.message,formatDisplayMessage(a,o)),{attributedInputResponses:l,inputResponses:u,...d}=e.stepInput,f={...d,message:s};return r.length>0&&(f.inputResponses=r),i.length>0&&(f.attributedInputResponses=i),{displayMessage:c,kind:`converted`,stepInput:f}}function formatModelMessage(e,t){let n=e.map(e=>{let n=t.get(e.requestId),r=n?.options?.find(t=>t.id===e.optionId),i={};if(e.optionId!==void 0&&(i.optionId=e.optionId),r!==void 0){let e={id:r.id,label:r.label};r.description!==void 0&&(e.description=r.description),i.selectedOption=e}e.text!==void 0&&(i.text=e.text);let a={requestId:e.requestId,response:i};return n?.kind===`tool-approval`&&(a.prompt=n.prompt,a.requestType=`approval`,a.toolName=n.action.toolName),a});return[`The user submitted the following response to an earlier interactive prompt.`,`Treat it as new input at the current point in the conversation and decide whether it is still relevant. This does not authorize an earlier action; request approval again if that action is still needed.`,JSON.stringify(n,null,2)].join(`
+`)}function formatDisplayMessage(e,t){return e.map(e=>e.text!==void 0&&e.text.length>0?e.text:t.get(e.requestId)?.options?.find(t=>t.id===e.optionId)?.label??e.optionId??`Response to an earlier interactive prompt`).join(`
+`)}function appendOptionalUserContent(e,t){let n=normalizeUserContent(e);return n===void 0?t:appendUserContent({appended:t,existing:n})}export{convertStaleResponsesToUserMessage,dropStaleSessionLimitContinuationResponses};

package/dist/src/harness/inline-tool-authorization.d.ts

@@ -11,5 +11,9 @@ export declare function resolveInlineAuthorizationInterrupt(input: {
     readonly toolResults: readonly TypedToolResult<ToolSet>[] | undefined;
 }): {
     readonly challenges: AuthorizationSignal["challenges"];
+    /** The calls each sign-in stopped, by the challenge's connection name. */
+    readonly callIdsByName: ReadonlyMap<string, readonly string[]>;
     readonly history: ModelMessage[];
 } | undefined;
+/** Drops the given calls and their results, keeping protocol-complete sibling calls. */
+export declare function withoutCalls(messages: readonly ModelMessage[], interruptedCallIds: ReadonlySet<string>): ModelMessage[];

package/dist/src/harness/inline-tool-authorization.js

@@ -1 +1 @@
-import{contextStorage}from"#context/container.js";import{isAuthorizationSignal,isPendingAuthorizationToolOutput,resolveActiveAuthorizationChallenges}from"#harness/authorization.js";import{readToolInterrupt}from"#harness/tool-interrupts.js";function isInlineAuthorizationToolResult(e){return isPendingAuthorizationToolOutput(e.output)||readAuthorizationSignal(e)!==void 0}function resolveInlineAuthorizationInterrupt(e){let t=[],n=new Set;for(let r of e.toolResults??[]){let e=readAuthorizationSignal(r);e!==void 0&&(t.push(e),n.add(r.toolCallId))}if(t.length!==0)return{challenges:resolveActiveAuthorizationChallenges(t.flatMap(e=>e.challenges)),history:projectCompletedSiblingCalls(e.messages,n)}}function readAuthorizationSignal(e){let t=contextStorage.getStore(),n=t===void 0?void 0:readToolInterrupt(t,e.toolCallId);return n!==void 0&&isAuthorizationSignal(n)?n:isAuthorizationSignal(e.output)?e.output:void 0}function projectCompletedSiblingCalls(e,t){return e.flatMap(e=>{if(e.role===`assistant`&&Array.isArray(e.content)){let n=e.content.some(e=>e.type===`tool-call`&&t.has(e.toolCallId)),r=e.content.filter(e=>e.type!==`tool-call`||!t.has(e.toolCallId)),i=r.some(e=>e.type===`tool-call`);return r.length===0||n&&!i?[]:[{...e,content:r}]}if(e.role===`tool`){let n=e.content.filter(e=>e.type!==`tool-result`||!t.has(e.toolCallId));return n.length===0?[]:[{...e,content:n}]}return[e]})}export{isInlineAuthorizationToolResult,resolveInlineAuthorizationInterrupt};
+import{contextStorage}from"#context/container.js";import{isAuthorizationSignal,isPendingAuthorizationToolOutput,resolveActiveAuthorizationChallenges}from"#harness/authorization.js";import{readToolInterrupt}from"#harness/tool-interrupts.js";function isInlineAuthorizationToolResult(e){return isPendingAuthorizationToolOutput(e.output)||readAuthorizationSignal(e)!==void 0}function resolveInlineAuthorizationInterrupt(e){let t=[],n=new Set,r=new Map;for(let i of e.toolResults??[]){let e=readAuthorizationSignal(i);if(e!==void 0){t.push(e),n.add(i.toolCallId);for(let t of e.challenges){let e=r.get(t.name)??[];r.set(t.name,[...e,i.toolCallId])}}}if(t.length!==0)return{callIdsByName:r,challenges:resolveActiveAuthorizationChallenges(t.flatMap(e=>e.challenges)),history:withoutCalls(e.messages,n)}}function readAuthorizationSignal(e){let t=contextStorage.getStore(),n=t===void 0?void 0:readToolInterrupt(t,e.toolCallId);return n!==void 0&&isAuthorizationSignal(n)?n:isAuthorizationSignal(e.output)?e.output:void 0}function withoutCalls(e,t){return e.flatMap(e=>{if(e.role===`assistant`&&Array.isArray(e.content)){let n=e.content.some(e=>e.type===`tool-call`&&t.has(e.toolCallId)),r=e.content.filter(e=>e.type!==`tool-call`||!t.has(e.toolCallId)),i=r.some(e=>e.type===`tool-call`);return r.length===0||n&&!i?[]:[{...e,content:r}]}if(e.role===`tool`){let n=e.content.filter(e=>e.type!==`tool-result`||!t.has(e.toolCallId));return n.length===0?[]:[{...e,content:n}]}return[e]})}export{isInlineAuthorizationToolResult,resolveInlineAuthorizationInterrupt,withoutCalls};

package/dist/src/harness/input-extraction.d.ts

@@ -1,4 +1,4 @@
-import type { ContentPart, ModelMessage, ToolSet } from "ai";
+import type { ContentPart, ToolSet } from "ai";
 import type { InputRequest } from "#shared/input.js";
 /**
  * Extracts tool approval input requests from AI SDK content parts that
@@ -8,12 +8,3 @@ export declare function extractToolApprovalInputRequests(input: {
     readonly content: readonly ContentPart<ToolSet>[];
     readonly excludedCallIds?: ReadonlySet<string>;
 }): InputRequest[];
-/**
- * Recovers approval request metadata for submitted input response IDs from
- * model history. The newest occurrence wins so compacted or repeated history
- * does not replace the request that is closest to the current turn.
- */
-export declare function extractHistoricalInputRequests(input: {
-    readonly history: readonly ModelMessage[];
-    readonly requestIds: ReadonlySet<string>;
-}): ReadonlyMap<string, InputRequest>;

package/dist/src/harness/input-extraction.js

@@ -1 +1 @@
-import{z}from"#compiled/zod/index.js";import{toolCallDisplayName}from"#execution/tools/connection-target.js";import{createRuntimeToolCallActionFromToolCall}from"#harness/tool-call-action.js";const ToolCallDescriptorSchema=z.object({input:z.unknown(),toolCallId:z.string(),toolName:z.string()}),PersistedToolCallSchema=ToolCallDescriptorSchema.extend({type:z.literal(`tool-call`)}),ToolApprovalRequestSchema=z.object({approvalId:z.string(),isAutomatic:z.boolean().optional().catch(void 0),toolCall:ToolCallDescriptorSchema.optional().catch(void 0),toolCallId:z.string().optional().catch(void 0),type:z.literal(`tool-approval-request`)});function extractToolApprovalInputRequests(e){return extractApprovalRequests(e)}function extractApprovalRequests(e){let t=[],n=new Map;for(let t of e.content){let e=PersistedToolCallSchema.safeParse(t);e.success&&n.set(e.data.toolCallId,e.data)}for(let r of e.content){let i=ToolApprovalRequestSchema.safeParse(r);if(!i.success)continue;let a=i.data;if(e.includedRequestIds!==void 0&&!e.includedRequestIds.has(a.approvalId)||a.isAutomatic===!0)continue;let o=a.toolCall??(a.toolCallId===void 0?void 0:n.get(a.toolCallId));o!==void 0&&(e.excludedCallIds?.has(o.toolCallId)||t.push({action:createRuntimeToolCallActionFromToolCall({toolCall:o}),allowFreeform:!1,display:`confirmation`,kind:`tool-approval`,options:[{id:`approve`,label:`Approve`},{id:`cancel`,label:`Cancel`}],prompt:`Approve ${toolCallDisplayName(o.toolName,o.input)}?`,requestId:a.approvalId}))}return t}function extractHistoricalInputRequests(e){let t=new Map;for(let n=e.history.length-1;n>=0;--n){let r=e.history[n];if(r?.role!==`assistant`||!Array.isArray(r.content))continue;let i=extractApprovalRequests({content:r.content,includedRequestIds:e.requestIds});for(let n of i)e.requestIds.has(n.requestId)&&!t.has(n.requestId)&&t.set(n.requestId,n);if(t.size===e.requestIds.size)break}return t}export{extractHistoricalInputRequests,extractToolApprovalInputRequests};
+import{z}from"#compiled/zod/index.js";import{toolCallDisplayName}from"#execution/tools/connection-target.js";import{createRuntimeToolCallActionFromToolCall}from"#harness/tool-call-action.js";const ToolCallDescriptorSchema=z.object({input:z.unknown(),toolCallId:z.string(),toolName:z.string()}),PersistedToolCallSchema=ToolCallDescriptorSchema.extend({type:z.literal(`tool-call`)}),ToolApprovalRequestSchema=z.object({approvalId:z.string(),isAutomatic:z.boolean().optional().catch(void 0),toolCall:ToolCallDescriptorSchema.optional().catch(void 0),toolCallId:z.string().optional().catch(void 0),type:z.literal(`tool-approval-request`)});function extractToolApprovalInputRequests(e){return extractApprovalRequests(e)}function extractApprovalRequests(e){let t=[],n=new Map;for(let t of e.content){let e=PersistedToolCallSchema.safeParse(t);e.success&&n.set(e.data.toolCallId,e.data)}for(let r of e.content){let i=ToolApprovalRequestSchema.safeParse(r);if(!i.success)continue;let a=i.data;if(a.isAutomatic===!0)continue;let o=a.toolCall??(a.toolCallId===void 0?void 0:n.get(a.toolCallId));o!==void 0&&(e.excludedCallIds?.has(o.toolCallId)||t.push({action:createRuntimeToolCallActionFromToolCall({toolCall:o}),allowFreeform:!1,display:`confirmation`,kind:`tool-approval`,options:[{id:`approve`,label:`Approve`},{id:`cancel`,label:`Cancel`}],prompt:`Approve ${toolCallDisplayName(o.toolName,o.input)}?`,requestId:a.approvalId}))}return t}export{extractToolApprovalInputRequests};

package/dist/src/harness/input-request-resolution.d.ts

@@ -1,4 +1,4 @@
-import type { PendingInputBatchEvent } from "#harness/pending-input-batches.js";
+import type { StepCoordinates as PendingInputBatchEvent } from "#harness/session-machine/view.js";
 import type { InputRequest, InputResponse } from "#shared/input.js";
 export declare const TOOL_EXECUTION_DENIED_MESSAGE = "Tool execution was denied.";
 type ApprovalTerminalStatus = "approved" | "denied" | "ignored" | "invalid";
@@ -10,10 +10,6 @@ export interface ResolvedInputBatch {
         readonly response?: InputResponse;
     }[];
 }
-export declare function buildResolvedInputBatch(batch: {
-    readonly event?: PendingInputBatchEvent;
-    readonly requests: readonly InputRequest[];
-}, responses: readonly InputResponse[]): ResolvedInputBatch | undefined;
 /**
  * One request's terminal outcome once its batch resolves: an approval's
  * decision, or whether a question received a response.

package/dist/src/harness/input-request-resolution.js

@@ -1 +1 @@
-const TOOL_EXECUTION_DENIED_MESSAGE=`Tool execution was denied.`;function buildResolvedInputBatch(e,t){if(e.event===void 0)return;let n=new Map(t.map(e=>[e.requestId,e]));return{event:e.event,inputs:e.requests.map(e=>{let t=n.get(e.requestId);return{outcome:resolveInputOutcome(e.kind,t),request:e,response:t}})}}function resolveInputOutcome(e,t){return e===`tool-approval`?resolveApprovalOutcome(t).status:t===void 0?`ignored`:`answered`}function resolveApprovalOutcome(e){return e===void 0?{approved:!1,reason:`Ignored because the user continued without responding.`,status:`ignored`}:e.optionId===`approve`?{approved:!0,reason:void 0,status:`approved`}:e.optionId===`cancel`||e.optionId===`deny`?{approved:!1,reason:TOOL_EXECUTION_DENIED_MESSAGE,status:`denied`}:{approved:!1,reason:`Invalid approval response.`,status:`invalid`}}export{TOOL_EXECUTION_DENIED_MESSAGE,buildResolvedInputBatch,resolveApprovalOutcome,resolveInputOutcome};
+const TOOL_EXECUTION_DENIED_MESSAGE=`Tool execution was denied.`;function resolveInputOutcome(e,t){return e===`tool-approval`?resolveApprovalOutcome(t).status:t===void 0?`ignored`:`answered`}function resolveApprovalOutcome(e){return e===void 0?{approved:!1,reason:`Ignored because the user continued without responding.`,status:`ignored`}:e.optionId===`approve`?{approved:!0,reason:void 0,status:`approved`}:e.optionId===`cancel`||e.optionId===`deny`?{approved:!1,reason:TOOL_EXECUTION_DENIED_MESSAGE,status:`denied`}:{approved:!1,reason:`Invalid approval response.`,status:`invalid`}}export{TOOL_EXECUTION_DENIED_MESSAGE,resolveApprovalOutcome,resolveInputOutcome};

package/dist/src/harness/messages.js

@@ -1 +1 @@
-import{attachClientContext,readClientContext}from"#internal/client-context.js";import{attachInputText,readInputText}from"#internal/input-text.js";function createUserMessage(e,t,n){let r={content:t,kind:e,role:`user`};return n===void 0?r:{...r,metadata:n}}function createFrameworkUserMessage(e,t,n){return createUserMessage(e,t,n)}function isUserMessageKind(e){return e===`user`||isFrameworkMessageKind(e)}function isUserModelMessage(e){if(e.role!==`user`)return!1;let{kind:t}=e;return isUserMessageKind(t)}function isFrameworkUserMessage(e){return isUserModelMessage(e)&&e.kind!==`user`}function validateHarnessModelMessages(e){let t=[];for(let n of e){if(n.role!==`user`){t.push(n);continue}if(!isUserModelMessage(n))throw TypeError(`Expected every user-role model message to have a kind.`);t.push(n)}return t}function isFrameworkMessageKind(e){return e===`context.instruction`||e===`context.state`||e===`context.compaction`||e===`memory.load`||e===`execution.continuation`||e===`execution.retry`||e===`task.result`}function markFrameworkStepInput(e,t){return{...e,frameworkMessageKind:t}}function frameworkMessageKindForStepInput(e){if(e===void 0)return;let{frameworkMessageKind:t}=e;return isFrameworkMessageKind(t)?t:void 0}function coalesceTurnInputs(e,t){let n=coalesceInputResponses({a:e.inputResponses,b:t.inputResponses}),r=coalesceMessage({a:e.message,b:t.message}),i=coalesceContext({a:e.context,b:t.context}),a=coalesceFrameworkMessageKind({a:e,b:t}),o=coalesceContext({a:readClientContext(e),b:readClientContext(t)}),s=t.outputSchema??e.outputSchema,c={};n!==void 0&&(c.inputResponses=n),r!==void 0&&(c.message=r),i!==void 0&&(c.context=i),s!==void 0&&(c.outputSchema=s);let l=normalizeUserContent(e.message)===void 0?readInputText(t):normalizeUserContent(t.message)===void 0?readInputText(e):void 0;return attachClientContext(attachInputText(a===void 0?c:markFrameworkStepInput(c,a),l),o)}function normalizeUserContent(e){if(e===void 0)return;if(typeof e==`string`)return e.trim().length>0?e:void 0;let t=e.filter(e=>e.type!==`text`||e.text.trim().length>0);if(t.length!==0)return t.length===e.length?e:t}const TOOL_RESULT_BOUNDARY=`…`;function followsToolResults(e){return e.findLast(e=>e.role!==`user`)?.role===`tool`}function createTurnInputMessages(e){let t=[...readClientContext(e)??[],...e?.context??[]].map(e=>createFrameworkUserMessage(`context.instruction`,e)),n=normalizeUserContent(e?.message);if(n===void 0)return t;let r=frameworkMessageKindForStepInput(e);return[...t,r===void 0?createUserMessage(`user`,n):createFrameworkUserMessage(r,n)]}function normalizeModelMessages(e){return e.flatMap(e=>{if(typeof e.content==`string`)return e.content.trim().length>0?[e]:[];let t=e.content.filter(e=>e.type!==`text`||e.text.trim().length>0);return t.length===0?[]:t.length===e.content.length?[e]:[{...e,content:t}]})}function resolveAssistantStepText(e,t){for(let t=e.length-1;t>=0;--t){let n=e[t];if(n?.role!==`assistant`)continue;let r=extractMessageText(n);if(r.trim().length>0)return r}return t!==void 0&&t.trim().length>0?t:null}function extractMessageText(e){return typeof e.content==`string`?e.content:Array.isArray(e.content)?e.content.flatMap(e=>typeof e==`string`?[e]:`type`in e&&e.type===`text`&&typeof e.text==`string`?[e.text]:[]).join(``):``}function coalesceInputResponses(e){let t=e.a??[],n=e.b??[];if(t.length!==0||n.length!==0)return[...t,...n]}function coalesceContext(e){let t=e.a??[],n=e.b??[];if(t.length!==0||n.length!==0)return[...t,...n]}function coalesceFrameworkMessageKind(e){let t=frameworkMessageKindForStepInput(e.a),n=frameworkMessageKindForStepInput(e.b);return e.a.message===void 0?n:e.b.message===void 0||t===n?t:void 0}function coalesceMessage(e){let t=normalizeUserContent(e.a),n=normalizeUserContent(e.b);return t===void 0?n:n===void 0?t:appendUserContent({appended:n,existing:t})}function appendUserContent(e){return typeof e.existing==`string`&&typeof e.appended==`string`?`${e.existing}\n\n${e.appended}`:[...toUserContentArray(e.existing),...toUserContentArray(e.appended)]}function toUserContentArray(e){return typeof e==`string`?e.length>0?[{type:`text`,text:e}]:[]:Array.isArray(e)?[...e]:[]}function coalesceDeliveries(e){let[t,...n]=e;if(t===void 0)throw Error(`Cannot coalesce an empty delivery batch.`);let r=t.auth,i=t.caller,a=[...t.payloads],o=[...t.deliveryMetadata??[]];for(let e of n){let t=a.length;if(e.auth!==void 0&&(r=e.auth),e.caller!==void 0){if(i!==void 0&&i.callId!==e.caller.callId)throw Error(`Cannot coalesce deliveries from different turns.`);i=e.caller}a.push(...e.payloads),o.push(...(e.deliveryMetadata??[]).map(e=>({...e,payloadIndex:e.payloadIndex+t})))}return{...t,auth:r,caller:i,deliveryMetadata:o.length===0?void 0:o,payloads:a}}export{TOOL_RESULT_BOUNDARY,appendUserContent,coalesceDeliveries,coalesceTurnInputs,createFrameworkUserMessage,createTurnInputMessages,createUserMessage,followsToolResults,frameworkMessageKindForStepInput,isFrameworkMessageKind,isFrameworkUserMessage,isUserMessageKind,isUserModelMessage,markFrameworkStepInput,normalizeModelMessages,normalizeUserContent,resolveAssistantStepText,validateHarnessModelMessages};
+import{attachClientContext,readClientContext}from"#internal/client-context.js";import{attachInputText,readInputText}from"#internal/input-text.js";function createUserMessage(e,t,n){let r={content:t,kind:e,role:`user`};return n===void 0?r:{...r,metadata:n}}function createFrameworkUserMessage(e,t,n){return createUserMessage(e,t,n)}function isUserMessageKind(e){return e===`user`||isFrameworkMessageKind(e)}function isUserModelMessage(e){if(e.role!==`user`)return!1;let{kind:t}=e;return isUserMessageKind(t)}function isFrameworkUserMessage(e){return isUserModelMessage(e)&&e.kind!==`user`}function validateHarnessModelMessages(e){let t=[];for(let n of e){if(n.role!==`user`){t.push(n);continue}if(!isUserModelMessage(n))throw TypeError(`Expected every user-role model message to have a kind.`);t.push(n)}return t}function isFrameworkMessageKind(e){return e===`context.instruction`||e===`context.state`||e===`context.compaction`||e===`memory.load`||e===`execution.continuation`||e===`execution.retry`||e===`task.result`}function markFrameworkStepInput(e,t){return{...e,frameworkMessageKind:t}}function frameworkMessageKindForStepInput(e){if(e===void 0)return;let{frameworkMessageKind:t}=e;return isFrameworkMessageKind(t)?t:void 0}function coalesceTurnInputs(e,t){let n=coalesceInputResponses({a:e.inputResponses,b:t.inputResponses}),r=coalesceMessage({a:e.message,b:t.message}),i=coalesceContext({a:e.context,b:t.context}),a=coalesceFrameworkMessageKind({a:e,b:t}),o=coalesceContext({a:readClientContext(e),b:readClientContext(t)}),s=t.outputSchema??e.outputSchema,c=[...e.attributedInputResponses??[],...t.attributedInputResponses??[]],l={};n!==void 0&&(l.inputResponses=n),c.length>0&&(l.attributedInputResponses=c),r!==void 0&&(l.message=r),i!==void 0&&(l.context=i),s!==void 0&&(l.outputSchema=s);let u=normalizeUserContent(e.message)===void 0?readInputText(t):normalizeUserContent(t.message)===void 0?readInputText(e):void 0;return attachClientContext(attachInputText(a===void 0?l:markFrameworkStepInput(l,a),u),o)}function normalizeUserContent(e){if(e===void 0)return;if(typeof e==`string`)return e.trim().length>0?e:void 0;let t=e.filter(e=>e.type!==`text`||e.text.trim().length>0);if(t.length!==0)return t.length===e.length?e:t}const TOOL_RESULT_BOUNDARY=`…`;function followsToolResults(e){return e.findLast(e=>e.role!==`user`)?.role===`tool`}function createTurnInputMessages(e){let t=[...readClientContext(e)??[],...e?.context??[]].map(e=>createFrameworkUserMessage(`context.instruction`,e)),n=normalizeUserContent(e?.message);if(n===void 0)return t;let r=frameworkMessageKindForStepInput(e);return[...t,r===void 0?createUserMessage(`user`,n):createFrameworkUserMessage(r,n)]}function normalizeModelMessages(e){return e.flatMap(e=>{if(typeof e.content==`string`)return e.content.trim().length>0?[e]:[];let t=e.content.filter(e=>e.type!==`text`||e.text.trim().length>0);return t.length===0?[]:t.length===e.content.length?[e]:[{...e,content:t}]})}function resolveAssistantStepText(e,t){for(let t=e.length-1;t>=0;--t){let n=e[t];if(n?.role!==`assistant`)continue;let r=extractMessageText(n);if(r.trim().length>0)return r}return t!==void 0&&t.trim().length>0?t:null}function extractMessageText(e){return typeof e.content==`string`?e.content:Array.isArray(e.content)?e.content.flatMap(e=>typeof e==`string`?[e]:`type`in e&&e.type===`text`&&typeof e.text==`string`?[e.text]:[]).join(``):``}function coalesceInputResponses(e){let t=e.a??[],n=e.b??[];if(t.length!==0||n.length!==0)return[...t,...n]}function coalesceContext(e){let t=e.a??[],n=e.b??[];if(t.length!==0||n.length!==0)return[...t,...n]}function coalesceFrameworkMessageKind(e){let t=frameworkMessageKindForStepInput(e.a),n=frameworkMessageKindForStepInput(e.b);return e.a.message===void 0?n:e.b.message===void 0||t===n?t:void 0}function coalesceMessage(e){let t=normalizeUserContent(e.a),n=normalizeUserContent(e.b);return t===void 0?n:n===void 0?t:appendUserContent({appended:n,existing:t})}function appendUserContent(e){return typeof e.existing==`string`&&typeof e.appended==`string`?`${e.existing}\n\n${e.appended}`:[...toUserContentArray(e.existing),...toUserContentArray(e.appended)]}function toUserContentArray(e){return typeof e==`string`?e.length>0?[{type:`text`,text:e}]:[]:Array.isArray(e)?[...e]:[]}function coalesceDeliveries(e){let[t,...n]=e;if(t===void 0)throw Error(`Cannot coalesce an empty delivery batch.`);let r=t.auth,i=t.caller,a=[...t.payloads],o=[...t.deliveryMetadata??[]];for(let e of n){let t=a.length;if(e.auth!==void 0&&(r=e.auth),e.caller!==void 0){if(i!==void 0&&i.callId!==e.caller.callId)throw Error(`Cannot coalesce deliveries from different turns.`);i=e.caller}a.push(...e.payloads),o.push(...(e.deliveryMetadata??[]).map(e=>({...e,payloadIndex:e.payloadIndex+t})))}return{...t,auth:r,caller:i,deliveryMetadata:o.length===0?void 0:o,payloads:a}}export{TOOL_RESULT_BOUNDARY,appendUserContent,coalesceDeliveries,coalesceTurnInputs,createFrameworkUserMessage,createTurnInputMessages,createUserMessage,followsToolResults,frameworkMessageKindForStepInput,isFrameworkMessageKind,isFrameworkUserMessage,isUserMessageKind,isUserModelMessage,markFrameworkStepInput,normalizeModelMessages,normalizeUserContent,resolveAssistantStepText,validateHarnessModelMessages};

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

@@ -0,0 +1,78 @@
+import { type LanguageModel, type ModelMessage } from "ai";
+import type { GenerationSteering } from "#harness/generation-steering.js";
+import type { HarnessModelMessage, UserModelMessage } from "#harness/messages.js";
+import { type AnthropicCacheMarker, type PromptCachePath } from "#harness/prompt-cache.js";
+import type { Step } from "#harness/step/context.js";
+import { type Prompt } from "#harness/step/prompt.js";
+import { type HarnessStepResult } from "#harness/step-hooks.js";
+import type { StepResult } from "#harness/types.js";
+import type { InstrumentationAttempt } from "#instrumentation/runtime.js";
+import { type RequestMessages } from "./request.js";
+import { type ModelTools } from "./tools.js";
+/** How one attempt differs from the step's first: what recovery and retries change. */
+export interface ModelCallOptions {
+    readonly disabledProviderTools?: ReadonlySet<string>;
+    readonly extraSystemNote?: string;
+    readonly retryReason?: "empty-response";
+    readonly suppressStepStartedEmission?: boolean;
+    readonly trailingUserNote?: string;
+}
+/** What a model step's caller needs from the step. */
+interface ModelCallerInput {
+    readonly model: LanguageModel;
+    readonly generation: GenerationSteering;
+    readonly hidesHeldText: boolean;
+    readonly turnMessages: readonly UserModelMessage[];
+    readonly approvedTools: ReadonlySet<string>;
+    readonly pendingApprovalsNote: string | undefined;
+    /** The prompt as the step started it, projected for the model. */
+    readonly projectedMessages: HarnessModelMessage[];
+    readonly startStep: (messages: readonly ModelMessage[]) => Promise<void>;
+    readonly setAttemptScope: (scope: InstrumentationAttempt | undefined) => void;
+}
+/**
+ * Calls the model for one step. Each attempt rebuilds the tools, the request, and the agent, so an
+ * earlier stream can't resolve a retry's one-shot step hooks with stale partial output. An attempt
+ * may first compact the prompt, which rewrites the durable messages the step later commits.
+ */
+export declare class ModelCaller {
+    readonly cachePath: PromptCachePath;
+    readonly marker: AnthropicCacheMarker | undefined;
+    private readonly attributionHeaders;
+    /** The prompt as the model reads it, projected for its history view. */
+    private projectedMessages;
+    /** The latest attempt's request; its history is what the step commits. */
+    request: RequestMessages;
+    /** The latest attempt's tools, which decide how the step's calls defer and end the turn. */
+    tools: ModelTools | undefined;
+    /** The messages the latest attempt sent, after attachments were hydrated. */
+    modelMessages: ModelMessage[];
+    requestEnvelopeTokens: number;
+    /** Usage of a call steering interrupted, which the steered step still counts. */
+    private interruptedUsage;
+    private compactionFailure;
+    private attemptIndex;
+    private readonly step;
+    private readonly prompt;
+    private readonly input;
+    constructor(step: Step, prompt: Prompt, input: ModelCallerInput);
+    /** Calls the model, retrying transient failures. */
+    call(options: ModelCallOptions): Promise<HarnessStepResult>;
+    /** A failed compaction fails the step, whatever recovery the call attempted. */
+    throwIfCompactionFailed(): void;
+    /** Steering interrupted the call: the step ends with what it committed, and runs again. */
+    steered(): Promise<StepResult>;
+    /** The step's result counted its usage, so an interrupted call's usage no longer applies. */
+    clearInterruptedUsage(): void;
+    private buildRequest;
+    private instructions;
+    private prepare;
+    /** Compacts the prompt when it's over the threshold, then rebuilds what depends on it. */
+    private compact;
+    private attempt;
+    /** Streams the call, publishing its content as it arrives. */
+    private stream;
+    /** Runs the call without a listener: nothing streams. */
+    private generate;
+}
+export {};

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

@@ -0,0 +1 @@
+import{buildGatewayAttributionHeaders}from"./model.js";import{isEmptyModelResponse,rethrowNoOutputAsEmptyResponse}from"./recovery.js";import{modelInstructions,requestMessages,withTrailingUserNote}from"./request.js";import{answerSkippedToolCalls,appendMissingToolResultMessages,extractToolResultCallIds,withAccumulatedResponseMessages}from"./response.js";import{runModelCallWithRetries}from"./retry.js";import{logToolExecutionError,prepareModelTools}from"./tools.js";import{extractGatewayCostUsd,extractTokenUsageDelta}from"./usage.js";import{ToolLoopAgent,isStepCount}from"ai";import{createLogger,logError}from"#internal/logging.js";import{AuthKey,HistoryStateKey}from"#context/keys.js";import{createStepCompletedEvent}from"#protocol/message.js";import{activeTurnId}from"#harness/session-machine/view.js";import{accumulateTurnUsage,getTurnUsageState,setTurnUsageState}from"#harness/turn-tag-state.js";import{hydrateSandboxAttachments,moveToolResultFilesToUserMessages}from"#harness/attachment-staging.js";import{throwIfTurnAborted}from"#harness/turn-cancellation.js";import{summarizeKnownError}from"#harness/semantic-errors/index.js";import{estimateTokens}from"#harness/token-estimate.js";import{estimateRequestEnvelope}from"#harness/request-envelope.js";import{workingTaskIds}from"#execution/tasks/model-step.js";import{ContentFilteredModelResponseError,EmptyModelResponseError}from"#harness/model-call/errors.js";import{interruptStreamOnFailure}from"#harness/interruptible-stream.js";import{buildToolApproval}from"#harness/tools.js";import{compactPrompt,projectPrompt,withClientContext}from"#harness/step/prompt.js";import{emitStreamContent}from"#harness/emission.js";import{FINAL_OUTPUT_TOOL_NAME}from"#harness/final-output.js";import{detectPromptCachePath,getAnthropicCacheMarker}from"#harness/prompt-cache.js";import{buildStepHooks,emitStepActions,readGatewayGenerationId}from"#harness/step-hooks.js";import{maybeCompact}from"#harness/compaction/step.js";const environment=process.env.NODE_ENV??`unknown`,log=createLogger(`harness.tool-loop`);var ModelCaller=class{cachePath;marker;attributionHeaders;projectedMessages;request;tools;modelMessages=[];requestEnvelopeTokens=0;interruptedUsage;compactionFailure;attemptIndex=0;step;prompt;input;constructor(e,t,n){this.step=e,this.prompt=t,this.input=n,this.cachePath=detectPromptCachePath(n.model),this.marker=this.cachePath.kind===`anthropic-direct`?getAnthropicCacheMarker():void 0,this.attributionHeaders=buildGatewayAttributionHeaders(n.model,e.config.runtimeIdentity),this.projectedMessages=n.projectedMessages,this.request=this.buildRequest(e.config.tools)}async call(e){return await runModelCallWithRetries(t=>this.attempt({...e,suppressStepStartedEmission:t!==1||e.suppressStepStartedEmission}),{canRetry:()=>this.compactionFailure===void 0,sessionId:this.step.session.sessionId,turnId:this.step.position().turnId},this.input.generation.signal)}throwIfCompactionFailed(){if(this.compactionFailure!==void 0)throw this.compactionFailure.error}async steered(){let{step:e}=this;return throwIfTurnAborted(e.config.abortSignal),e.ctx?.set(HistoryStateKey,this.request.historyState),this.interruptedUsage!==void 0&&(e.session=setTurnUsageState(e.session,accumulateTurnUsage({previous:getTurnUsageState(e.session.state),turnId:e.position().turnId,usage:this.interruptedUsage}))),await e.emit?.(createStepCompletedEvent({...e.position(),finishReason:`other`,usage:this.interruptedUsage})),{next:e.runStep,session:{...e.session,history:[...this.request.history]},steered:!0}}clearInterruptedUsage(){this.interruptedUsage=void 0}buildRequest(e){return requestMessages(this.step,{coordinationTools:e,hidesHeldText:this.input.hidesHeldText,messages:this.prompt.messages,pendingApprovalsNote:this.input.pendingApprovalsNote,projectedMessages:this.projectedMessages,turnMessages:this.input.turnMessages})}instructions(e){return modelInstructions({extraSystemNote:e,marker:this.marker,session:this.step.session,systemMessages:this.request.systemMessages})}async prepare(e){let t=await prepareModelTools(this.step,{model:this.input.model,approvedTools:this.input.approvedTools,disabledProviderTools:e.disabledProviderTools,generation:this.input.generation,marker:this.marker});return this.tools=t,this.request=this.buildRequest(t.coordinationTools),this.requestEnvelopeTokens=await estimateRequestEnvelope({history:this.projectedMessages,instructions:this.instructions(e.extraSystemNote),messages:withTrailingUserNote(this.request.nonSystemMessages,e.trailingUserNote),tools:t.effectiveTools}),t}async compact(e,t){let{step:n,prompt:r}=this,{config:i}=n,a;try{a=await maybeCompact({abortSignal:i.abortSignal,auth:n.ctx?.get(AuthKey)??null,emissionState:n.position(),historyProjector:i.historyProjector,messages:[...r.messages],model:this.input.model,promptMessages:withClientContext(r),publish:n.publish,requestEnvelopeTokens:this.requestEnvelopeTokens,resolveModel:i.resolveModel,runtimeIdentity:i.runtimeIdentity,session:n.session,telemetry:n.instrumentation?.telemetry()})}catch(e){throw this.compactionFailure={error:e},e}if(n.session=a.session,!a.compacted)return t;compactPrompt(n,r,a.messages);let{compaction:o}=n.session;return n.session={...n.session,compaction:{recentWindowSize:o.recentWindowSize,threshold:o.threshold,thresholdPercent:o.thresholdPercent}},this.projectedMessages=projectPrompt(n,r),await this.prepare(e)}async attempt(e){let{step:t}=this,{generation:n,model:r}=this.input,i=await this.compact(e,await this.prepare(e));this.requestEnvelopeTokens=Math.max(0,this.requestEnvelopeTokens-(estimateTokens(this.request.history)-estimateTokens(this.prompt.messages))),n.begin(),this.modelMessages=await hydrateSandboxAttachments(this.request.nonSystemMessages),typeof this.input.model!=`string`&&typeof this.input.model.provider==`string`&&this.input.model.provider.endsWith(`.chat`)&&(this.modelMessages=moveToolResultFilesToUserMessages(this.modelMessages));let a=this.instructions(e.extraSystemNote),o=t.instrumentation?.resolveRuntimeContext({emissionState:t.position(),environment,modelInput:{instructions:a,messages:this.modelMessages},session:t.session})??{};e.retryReason&&(o[`eve.retry.reason`]=e.retryReason);let s=withTrailingUserNote(this.modelMessages,e.trailingUserNote),c=t.position(),l=t.instrumentation?.prepareAttempt({isFrameworkTool:e=>t.frameworkToolNames.has(e),attemptIndex:this.attemptIndex++,runtimeContext:o,stepIndex:c.stepIndex,turnId:activeTurnId(c)});this.input.setAttemptScope(l?.scope);let u=buildStepHooks({auth:t.ctx?.get(AuthKey)??null,cachePath:this.cachePath,marker:this.marker,session:t.session,startStep:e.suppressStepStartedEmission===!0?void 0:this.input.startStep}),d={headers:this.attributionHeaders,instructions:a,model:r,onLanguageModelCallEnd:e=>{n.interrupted||(this.interruptedUsage=extractTokenUsageDelta({costUsd:extractGatewayCostUsd(e.providerMetadata),usage:e.usage}))},onToolExecutionEnd:logToolExecutionError,onError(e){n.interrupted||summarizeKnownError(e.error)?.tags.includes(`config`)!==!0&&logError(log,`tool-loop stream error`,e.error)},onStepEnd:u.onStepEnd,onStepStart:u.onStepStart,prepareStep:u.prepareStep,reasoning:t.session.agent.modelReference?.reasoning??t.session.agent.reasoning,runtimeContext:o,stopWhen:isStepCount(1),telemetry:l?.telemetry,toolApproval:buildToolApproval(i.modelTools,n.signal),tools:i.effectiveTools},f=new ToolLoopAgent(d);try{let e=t.emit===void 0?await this.generate(f,s,u.stepResult):await this.stream(f,s,u.stepResult,i);return await l?.complete(),e}catch(e){return n.interrupted?await l?.complete():await l?.fail(e),n.check(),rethrowNoOutputAsEmptyResponse(e)}}async stream(e,t,n,r){let{step:i}=this,{generation:a}=this.input,o=new Set([FINAL_OUTPUT_TOOL_NAME]),s=await e.stream({abortSignal:a.signal,messages:t}),{emittedActionCallIds:c,handledInlineToolResultCallIds:l,invalidInputToolCallIds:u,inlineAuthorizationResults:d,trailingInlineToolResultParts:f}=await emitStreamContent(i.publish,i.position(),interruptStreamOnFailure(s.fullStream,a.signal),{excludedActionToolNames:o,hidesHeldText:this.input.hidesHeldText&&workingTaskIds(i.session).length>0,tools:r.presentationTools});throwIfTurnAborted(i.config.abortSignal),a.check();let[p,m]=await Promise.all([n,s.responseMessages]);assertUsableResponse(p,m,d.length>0||f.length>0);let h=answerSkippedToolCalls(p,r.effectiveTools);await emitStepActions(i.publish,i.position(),h.length===0?p:withAccumulatedResponseMessages({stepResult:p,responseMessages:appendMissingToolResultMessages({append:h,responseMessages:p.response.messages})}),{emittedActionCallIds:c,excludedActionCallIds:u,excludedActionToolNames:o,handledInlineToolResultCallIds:l,tools:r.presentationTools});let g=new Map(p.toolResults.map(e=>[e.toolCallId,e]));for(let e of d)g.set(e.toolCallId,e);return withAccumulatedResponseMessages({invalidInputToolCallIds:u,responseMessages:appendMissingToolResultMessages({append:[...f,...answerSkippedToolCalls(p,r.effectiveTools)],responseMessages:m}),stepResult:p,toolResults:[...g.values()]})}async generate(e,t,n){let{generation:r}=this.input,i=await e.generate({abortSignal:r.signal,messages:t});throwIfTurnAborted(this.step.config.abortSignal),r.check();let a=await n;return assertUsableResponse(a,i.responseMessages,!1),withAccumulatedResponseMessages({responseMessages:appendMissingToolResultMessages({append:answerSkippedToolCalls(a,this.tools.effectiveTools),responseMessages:i.responseMessages}),stepResult:a})}};function assertUsableResponse(e,t,n){if(e.finishReason===`content-filter`)throw new ContentFilteredModelResponseError(readGatewayGenerationId(e.providerMetadata));if(isEmptyModelResponse(e)&&extractToolResultCallIds(t).size===0&&!n)throw new EmptyModelResponseError}export{ModelCaller};

package/dist/src/harness/{model-call-error.d.ts → model-call/errors.d.ts}

@@ -41,7 +41,7 @@ export declare function extractUnsupportedProviderToolTypes(error: unknown): rea
  */
 export declare function extractModelCallErrorDetails(error: unknown): JsonObject;
 /**
- * A model call that produced no content. Raised by tool-loop.ts from
+ * A model call that produced no content. Raised by the model caller from
  * either of its two triggers: `isEmptyModelResponse` (a completed step
  * with finishReason 'other' and no output — AI Gateway HTTP 200 whose
  * stream carries no content, no usage, and no error) or

package/dist/src/harness/model-call/failure.d.ts

@@ -0,0 +1,52 @@
+import type { JsonObject } from "#shared/json.js";
+import type { SemanticErrorSummary } from "#harness/semantic-errors/rule.js";
+import { type UpstreamRejectionSummary } from "#harness/model-call/errors.js";
+import type { Step } from "#harness/step/context.js";
+import type { StepResult } from "#harness/types.js";
+/**
+ * Projects a model-call failure into the `step.failed` / `turn.failed`
+ * `details` payload. Three mutually exclusive branches:
+ *
+ * 1. Catalog match → the rule's curated `name`/`message`/`hint` plus its
+ *    registered `semanticErrorId`, no SDK inspector dump.
+ * 2. Upstream rejection → raw error `message` with the extracted upstream
+ *    identity, no inspector dump. No `semanticErrorId`: the message is
+ *    arbitrary provider prose, not a registered failure shape.
+ * 3. Fallback → full {@link formatError} projection (cause chain via
+ *    `util.inspect`) so unrecognized failures still carry the upstream
+ *    stack to log aggregators.
+ *
+ * All branches merge {@link extractModelCallErrorDetails} on top so the
+ * compact gateway diagnostics (`statusCode`, `upstreamMessage`,
+ * `responseBodySnippet`, ...) always show up next to the message.
+ */
+export declare function buildModelCallFailureDetails(input: {
+    readonly catalogSummary: SemanticErrorSummary | null;
+    readonly error: unknown;
+    readonly errorId: string;
+    readonly modelCallDetails: JsonObject;
+    readonly upstreamRejection: UpstreamRejectionSummary | null;
+}): JsonObject;
+/**
+ * Builds the structured log fields for a model-call failure. When the
+ * failure was recognized (catalog match or extracted upstream rejection),
+ * attach the compact `details` payload and *omit* the raw `error` so the
+ * logger's `util.inspect` of the cause chain (which would render
+ * `[object Object]` for upstream `APICallError` shapes) is bypassed.
+ * Otherwise fall back to the raw error so unrecognized failures keep
+ * their full stack in logs.
+ */
+export declare function buildModelCallFailureLogFields(input: {
+    readonly error: unknown;
+    readonly errorId: string;
+    readonly modelCallDetails: JsonObject;
+    readonly recognized: boolean;
+    readonly sessionId: string;
+    readonly turnId: string;
+}): Record<string, unknown>;
+/**
+ * The model call failed past its retries and recoveries. A failed write to the session's stream is
+ * the workflow's failure, not the model's: the session parks for the user to retry. A terminal
+ * failure ends the session; any other fails the turn recoverably.
+ */
+export declare function reportModelCallFailure(step: Step, error: unknown): Promise<StepResult>;

package/dist/src/harness/model-call/failure.js

@@ -0,0 +1 @@
+import{createErrorId,createLogger,formatError}from"#internal/logging.js";import{toErrorMessage}from"#shared/errors.js";import{fail}from"#harness/session-machine/transitions.js";import{summarizeKnownError}from"#harness/semantic-errors/index.js";import{classifyModelCallError,extractModelCallErrorDetails,extractUpstreamRejectionMessage}from"#harness/model-call/errors.js";import{extractWorkflowStreamWriteErrorDetails}from"#harness/workflow-stream-error.js";const log=createLogger(`harness.tool-loop`);function buildModelCallFailureDetails(e){let{catalogSummary:t,error:n,errorId:r,modelCallDetails:i,upstreamRejection:a}=e;if(t!==null){let e={errorId:r,message:t.message,name:t.name,semanticErrorId:t.id,...i};return t.hint!==void 0&&(e.hint=t.hint),e}return a===null?{...formatError(n,r),...i}:{errorId:r,message:toErrorMessage(n),name:a.name,...i}}function buildModelCallFailureLogFields(e){let t={errorId:e.errorId,sessionId:e.sessionId,turnId:e.turnId};return e.recognized?{...t,details:e.modelCallDetails}:{...t,error:e.error}}async function reportModelCallFailure(e,t){if(e.instrumentation?.recordError(t),e.emit===void 0)throw t;let{sessionId:n}=e.session,{turnId:r}=e.position(),i=extractWorkflowStreamWriteErrorDetails(t);if(i!==null){let a=createErrorId();return log.error(`workflow stream write failed — parking session for retry by the user`,{...i,errorId:a,error:t,sessionId:n,turnId:r}),await e.apply(fail(e.view(),{code:`WORKFLOW_STREAM_WRITE_FAILED`,details:{...i,errorId:a},message:toErrorMessage(t)})),{next:null,session:e.session}}let a=createErrorId(),o=summarizeKnownError(t),s=o===null?extractUpstreamRejectionMessage(t):null,c=o?.message??s?.message??toErrorMessage(t),l=o?.hint===void 0?c:`${c} ${o.hint}`,u=extractModelCallErrorDetails(t),d=buildModelCallFailureDetails({catalogSummary:o,error:t,errorId:a,modelCallDetails:u,upstreamRejection:s}),f=buildModelCallFailureLogFields({error:t,errorId:a,modelCallDetails:u,recognized:o!==null||s!==null,sessionId:n,turnId:r});return classifyModelCallError(t)===`terminal`?(o===null?log.error(s?.message??`model call failed terminally`,f):log.error(`${o.name}: ${o.message}`,{errorId:a,hint:o.hint,sessionId:n,turnId:r}),await e.apply(fail(e.view(),{code:`MODEL_CALL_FAILED`,details:d,message:c,terminal:{sessionId:n}})),{next:e.hasDelegatedCaller?{done:!0,isError:!0,output:l}:{done:!0,output:``},session:e.session}):(log.error(s?.message??`model call failed — parking session for retry by the user`,f),await e.apply(fail(e.view(),{code:`MODEL_CALL_FAILED`,details:d,message:c})),e.session={...e.session,outputSchema:void 0},{next:null,session:e.session,settledTurn:{isError:!0,output:l}})}export{buildModelCallFailureDetails,buildModelCallFailureLogFields,reportModelCallFailure};