@lunora/testing 1.0.0-alpha.140 → 1.0.0-alpha.142

package/README.md

@@ -167,8 +167,14 @@ The virtual clock is **per-harness** — advancing one harness's clock does not
 affect other harnesses, so tests running in parallel are isolated.
 
 Provide a `functions` map (`{ "path:name": registeredFn }`) so the scheduler
-can dispatch jobs. Paths not in the map produce a `console.warn` and are
-silently dropped (matching production behaviour for unknown paths).
+can dispatch jobs. A path not in the map fails the job with `FUNCTION_NOT_FOUND`,
+which then walks the retry budget into `t.scheduler.failures()` — matching
+production, where the DO fires the job at the Worker and dead-letters it.
+
+A retryable failure is re-enqueued silently, so `failures()` only fills once the
+whole backoff schedule (~930s of virtual clock) has elapsed — and at that point
+`advance()`/`runPending()` re-throw the failure by default. Pass
+`{ throwOnError: false }` to inspect `failures()` instead.
 
 #### Subscription testing
 

package/dist/index.d.mts

@@ -592,7 +592,9 @@ interface RecordedWideEvent {
  * stays `undefined`, matching the optional `ctx.env?` field.
  * - `ctx.fetch` (actions): inject a custom `fetch` via `options.fetch`.
  * - `ctx.scheduler` (mutations + actions): fully functional fake with virtual clock;
- * control via `harness.scheduler.advance(ms)` / `runPending()` / `list()`.
+ * control via `harness.scheduler.advance(ms)` / `runPending()` / `list()`. As in
+ * production it is transactional inside a mutation: a job scheduled by a mutation
+ * that then throws never becomes pending.
  * - `harness.subscribe(query, args)`: async iterable that re-emits after mutations.
  *
  * **v1 stubs (still throwing):** `ctx.storage`, `ctx.vectors`, `ctx.workflows`.

package/dist/index.d.ts

@@ -592,7 +592,9 @@ interface RecordedWideEvent {
  * stays `undefined`, matching the optional `ctx.env?` field.
  * - `ctx.fetch` (actions): inject a custom `fetch` via `options.fetch`.
  * - `ctx.scheduler` (mutations + actions): fully functional fake with virtual clock;
- * control via `harness.scheduler.advance(ms)` / `runPending()` / `list()`.
+ * control via `harness.scheduler.advance(ms)` / `runPending()` / `list()`. As in
+ * production it is transactional inside a mutation: a job scheduled by a mutation
+ * that then throws never becomes pending.
  * - `harness.subscribe(query, args)`: async iterable that re-emits after mutations.
  *
  * **v1 stubs (still throwing):** `ctx.storage`, `ctx.vectors`, `ctx.workflows`.

package/dist/index.mjs

@@ -1 +1 @@
-import{agentHarness as o,finalTurn as t,toolCallTurn as a}from"./packem_shared/agentHarness-DA9onGVO.mjs";import{evaluationAttributes as l,recordEvaluation as n}from"./packem_shared/evaluationAttributes-CYLgTP1P.mjs";import{lunoraTest as s}from"./packem_shared/lunoraTest-BOqU8mY2.mjs";import{groundednessScorer as p,mrrScorer as u,ndcgAtK as x,precisionAtK as S,recallAtK as f}from"./packem_shared/groundednessScorer-CznJBIjn.mjs";import{containsScorer as g,evaluate as A,exactMatchScorer as v,keywordScorer as K,llmScorer as M,regexScorer as T,scoreSample as k}from"./packem_shared/containsScorer-gv0Sdd86.mjs";import{extractLink as C,listCapturedMail as b,waitForMail as h}from"@lunora/mail/testing";export{o as agentHarness,g as containsScorer,A as evaluate,l as evaluationAttributes,v as exactMatchScorer,C as extractLink,t as finalTurn,p as groundednessScorer,K as keywordScorer,b as listCapturedMail,M as llmScorer,s as lunoraTest,u as mrrScorer,x as ndcgAtK,S as precisionAtK,f as recallAtK,n as recordEvaluation,T as regexScorer,k as scoreSample,a as toolCallTurn,h as waitForMail};
+import{agentHarness as o,finalTurn as t,toolCallTurn as a}from"./packem_shared/agentHarness-DA9onGVO.mjs";import{evaluationAttributes as l,recordEvaluation as n}from"./packem_shared/evaluationAttributes-CYLgTP1P.mjs";import{lunoraTest as s}from"./packem_shared/lunoraTest-PUpvwIBo.mjs";import{groundednessScorer as p,mrrScorer as u,ndcgAtK as x,precisionAtK as S,recallAtK as f}from"./packem_shared/groundednessScorer-kgXJCZDg.mjs";import{containsScorer as g,evaluate as A,exactMatchScorer as v,keywordScorer as K,llmScorer as M,regexScorer as T,scoreSample as k}from"./packem_shared/containsScorer-DoLfFOs6.mjs";import{extractLink as C,listCapturedMail as b,waitForMail as h}from"@lunora/mail/testing";export{o as agentHarness,g as containsScorer,A as evaluate,l as evaluationAttributes,v as exactMatchScorer,C as extractLink,t as finalTurn,p as groundednessScorer,K as keywordScorer,b as listCapturedMail,M as llmScorer,s as lunoraTest,u as mrrScorer,x as ndcgAtK,S as precisionAtK,f as recallAtK,n as recordEvaluation,T as regexScorer,k as scoreSample,a as toolCallTurn,h as waitForMail};

package/dist/packem_shared/containsScorer-DoLfFOs6.mjs

@@ -0,0 +1,2 @@
+import{LunoraError as i}from"@lunora/errors";const p=/^\s*(-?\d+(?:\.\d+)?)(?=$|[\s\-–—])/u,u=e=>Number.isFinite(e)?Math.min(1,Math.max(0,e)):0,h=e=>typeof e=="number"?{score:u(e)}:{score:u(e.score),...e.reason===void 0?{}:{reason:e.reason}},m=e=>e.length===0?0:e.reduce((t,n)=>t+n,0)/e.length,x=(e,t={})=>{if(e.length===0)throw new i("BAD_REQUEST","@lunora/testing: containsScorer requires a non-empty needle");return{name:`contains:${e}`,score:({output:n})=>{const a=t.caseSensitive?n:n.toLowerCase(),r=t.caseSensitive?e:e.toLowerCase();return a.includes(r)?1:0}}},v=(e,t="regex")=>{const n=new RegExp(e.source,e.flags.replaceAll(/[gy]/gu,""));return{name:t,score:({output:a})=>n.test(a)?1:0}},$=()=>({name:"exact-match",score:({expected:e,output:t})=>t.trim()===e?.trim()?1:0}),E=e=>{if(e.length===0)throw new i("BAD_REQUEST","@lunora/testing: keywordScorer requires at least one keyword");return{name:"keyword-coverage",score:({output:t})=>{const n=t.toLowerCase(),a=e.filter(r=>n.includes(r.toLowerCase())).length;return{reason:`${String(a)}/${String(e.length)} keywords present`,score:a/e.length}}}},w=(e,t)=>[`Rate the ASSISTANT OUTPUT against this criterion: ${e}`,"Respond with a single number from 0 (fails) to 1 (fully meets), then a dash and a one-line reason.",...t.input===void 0?[]:["",`Input: ${t.input}`],...t.expected===void 0?[]:["",`Reference answer: ${t.expected}`],"",`Assistant output: ${t.output}`].join(`
+`),S=e=>{const t=p.exec(e),n=t===null?Number.NaN:Number(t[1]);if(!Number.isFinite(n)||n<0||n>1)throw new i("BAD_REQUEST",`@lunora/testing: the LLM judge did not answer with a score in [0, 1] — got ${JSON.stringify(e.trim().slice(0,200))}`);return{reason:e.trim(),score:n}},N=e=>({name:e.name??"llm-judge",score:async t=>S(await e.judge(w(e.criteria,t)))}),f=async(e,t)=>{const n=await Promise.all(t.map(async o=>({name:o.name,result:h(await o.score(e))}))),a={},r=new Map;for(const{name:o,result:c}of n){const s=r.get(o)??0;r.set(o,s+1),a[s===0?o:`${o}#${String(s+1)}`]=c}return{average:m(n.map(({result:o})=>o.score)),scores:a}},A=async(e,t,n)=>{const a=await Promise.all(e.map(async r=>{const o=await t(r.input),c=typeof o=="string"?o:o.output,s=typeof o=="string"||o.metadata===void 0?r.metadata:{...r.metadata,...o.metadata},g={input:r.input,output:c,...r.expected===void 0?{}:{expected:r.expected},...s===void 0?{}:{metadata:s}},{average:d,scores:l}=await f(g,n);return{average:d,input:r.input,output:c,scores:l}}));return{average:m(a.map(r=>r.average)),items:a}};export{x as containsScorer,A as evaluate,$ as exactMatchScorer,E as keywordScorer,N as llmScorer,S as parseJudgeScore,v as regexScorer,f as scoreSample};

package/dist/packem_shared/{groundednessScorer-CznJBIjn.mjs → groundednessScorer-kgXJCZDg.mjs}

@@ -1,3 +1,3 @@
-import{LunoraError as g}from"@lunora/errors";import{parseJudgeScore as h}from"./containsScorer-gv0Sdd86.mjs";const w="retrieved",v="relevant",l=(e,t)=>{const n=e.metadata?.[t];return Array.isArray(n)?n.filter(r=>typeof r=="string"&&r.length>0):[]},s=(e,t)=>{const n=l(e,t?.relevantKey??v);if(n.length!==0)return{relevant:new Set(n),retrieved:l(e,t?.retrievedKey??w)}},a=e=>({reason:`no gold ids under metadata.${e?.relevantKey??v} — cannot score retrieval`,score:0}),c=(e,t)=>{if(e!==void 0&&(!Number.isInteger(e)||e<1))throw new g("BAD_REQUEST",`@lunora/testing: ${t} \`k\` must be a positive integer`)},u=(e,t)=>t===void 0?e.retrieved:e.retrieved.slice(0,t),f=(e,t)=>{const n=new Set;for(const r of e)t.has(r)&&n.add(r);return n.size},S=(e,t)=>{const n=new Set;let r=0;for(const[o,i]of e.entries())t.has(i)&&!n.has(i)&&(n.add(i),r+=1/Math.log2(o+2));return r},m=e=>{let t=0;for(let n=0;n<e;n+=1)t+=1/Math.log2(n+2);return t},p=(e,t)=>(c(e,"recallAtK"),{name:e===void 0?"recall":`recall@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=f(u(r,e),r.relevant);return{reason:`${String(o)}/${String(r.relevant.size)} gold ids retrieved`,score:o/r.relevant.size}}}),x=(e,t)=>(c(e,"precisionAtK"),{name:e===void 0?"precision":`precision@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=u(r,e);if(o.length===0)return{reason:"nothing retrieved",score:0};const i=f(o,r.relevant);return{reason:`${String(i)}/${String(o.length)} retrieved ids are gold`,score:i/o.length}}}),y=e=>({name:"mrr",score:t=>{const n=s(t,e);if(n===void 0)return a(e);const r=n.retrieved.findIndex(o=>n.relevant.has(o));return r===-1?{reason:"no gold id retrieved",score:0}:{reason:`first gold id at rank ${String(r+1)}`,score:1/(r+1)}}}),$=(e,t)=>(c(e,"ndcgAtK"),{name:e===void 0?"ndcg":`ndcg@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=u(r,e);if(o.length===0)return{reason:"nothing retrieved",score:0};const i=S(o,r.relevant),d=m(e===void 0?r.relevant.size:Math.min(r.relevant.size,e));return{reason:`dcg ${i.toFixed(3)} / ideal ${d.toFixed(3)}`,score:d===0?0:i/d}}}),K=e=>{if(typeof e.judge!="function")throw new g("BAD_REQUEST","@lunora/testing: groundednessScorer requires an injected `judge` function");const t=e.contextKey??"context";return{name:e.name??"groundedness",score:async n=>{const r=n.metadata?.[t];if(typeof r!="string"||r.trim().length===0)return{reason:`no retrieved context under metadata.${t}`,score:0};const o=await e.judge(["Rate how well the ASSISTANT ANSWER is supported by the RETRIEVED CONTEXT below.","Score 1 if every claim in the answer is supported by the context, 0 if the answer asserts","anything the context does not support. Judge support only — do NOT reward correctness","the context does not contain.","Respond with a single number from 0 to 1, then a dash and a one-line reason.","",`Retrieved context:
+import{LunoraError as g}from"@lunora/errors";import{parseJudgeScore as h}from"./containsScorer-DoLfFOs6.mjs";const w="retrieved",v="relevant",l=(e,t)=>{const n=e.metadata?.[t];return Array.isArray(n)?n.filter(r=>typeof r=="string"&&r.length>0):[]},s=(e,t)=>{const n=l(e,t?.relevantKey??v);if(n.length!==0)return{relevant:new Set(n),retrieved:l(e,t?.retrievedKey??w)}},a=e=>({reason:`no gold ids under metadata.${e?.relevantKey??v} — cannot score retrieval`,score:0}),c=(e,t)=>{if(e!==void 0&&(!Number.isInteger(e)||e<1))throw new g("BAD_REQUEST",`@lunora/testing: ${t} \`k\` must be a positive integer`)},u=(e,t)=>t===void 0?e.retrieved:e.retrieved.slice(0,t),f=(e,t)=>{const n=new Set;for(const r of e)t.has(r)&&n.add(r);return n.size},S=(e,t)=>{const n=new Set;let r=0;for(const[o,i]of e.entries())t.has(i)&&!n.has(i)&&(n.add(i),r+=1/Math.log2(o+2));return r},m=e=>{let t=0;for(let n=0;n<e;n+=1)t+=1/Math.log2(n+2);return t},p=(e,t)=>(c(e,"recallAtK"),{name:e===void 0?"recall":`recall@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=f(u(r,e),r.relevant);return{reason:`${String(o)}/${String(r.relevant.size)} gold ids retrieved`,score:o/r.relevant.size}}}),x=(e,t)=>(c(e,"precisionAtK"),{name:e===void 0?"precision":`precision@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=u(r,e);if(o.length===0)return{reason:"nothing retrieved",score:0};const i=f(o,r.relevant);return{reason:`${String(i)}/${String(o.length)} retrieved ids are gold`,score:i/o.length}}}),y=e=>({name:"mrr",score:t=>{const n=s(t,e);if(n===void 0)return a(e);const r=n.retrieved.findIndex(o=>n.relevant.has(o));return r===-1?{reason:"no gold id retrieved",score:0}:{reason:`first gold id at rank ${String(r+1)}`,score:1/(r+1)}}}),$=(e,t)=>(c(e,"ndcgAtK"),{name:e===void 0?"ndcg":`ndcg@${String(e)}`,score:n=>{const r=s(n,t);if(r===void 0)return a(t);const o=u(r,e);if(o.length===0)return{reason:"nothing retrieved",score:0};const i=S(o,r.relevant),d=m(e===void 0?r.relevant.size:Math.min(r.relevant.size,e));return{reason:`dcg ${i.toFixed(3)} / ideal ${d.toFixed(3)}`,score:d===0?0:i/d}}}),K=e=>{if(typeof e.judge!="function")throw new g("BAD_REQUEST","@lunora/testing: groundednessScorer requires an injected `judge` function");const t=e.contextKey??"context";return{name:e.name??"groundedness",score:async n=>{const r=n.metadata?.[t];if(typeof r!="string"||r.trim().length===0)return{reason:`no retrieved context under metadata.${t}`,score:0};const o=await e.judge(["Rate how well the ASSISTANT ANSWER is supported by the RETRIEVED CONTEXT below.","Score 1 if every claim in the answer is supported by the context, 0 if the answer asserts","anything the context does not support. Judge support only — do NOT reward correctness","the context does not contain.","Respond with a single number from 0 to 1, then a dash and a one-line reason.","",`Retrieved context:
 ${r}`,"",`Assistant answer: ${n.output}`].join(`
 `));return h(o)}}};export{K as groundednessScorer,y as mrrScorer,$ as ndcgAtK,x as precisionAtK,p as recallAtK};

package/dist/packem_shared/lunoraTest-PUpvwIBo.mjs

@@ -0,0 +1 @@
+import{LunoraError as k}from"@lunora/errors";import{withDeferredSchedules as rt,beginDeferredSchedules as nt}from"@lunora/server";import{runShardMigrations as ot,TransactionHeadroomTracker as Y,createShardCtxDb as st,RLS_UNWRAP_SYMBOL as it}from"@lunora/shard-engine";import{evaluationAttributes as ct}from"./evaluationAttributes-CYLgTP1P.mjs";import{assertScheduleDelay as at,MAX_RETRY_ATTEMPTS as ut,RETRY_BASE_DELAY_MS as dt}from"@lunora/scheduler";import{DatabaseSync as lt}from"node:sqlite";const ht=(o,s,i,u,g)=>{let v=g,b=1;const d=new Map,m=[],h=(r,l,a={},t)=>{JSON.stringify(a);const e=t??`fake-job-${String(b)}`;return b+=1,d.set(e,{args:a,enqueuedAt:v,functionPath:l,id:e,scheduledFor:r}),e},S=r=>typeof r=="string"?r:"__lunoraRef"in r?r.__lunoraRef:r.name??r.binding??"",x={cancel:r=>{const l=d.has(r);return d.delete(r),Promise.resolve({cancelled:l})},get:r=>Promise.resolve(d.get(r)??null),list:()=>Promise.resolve([...d.values()]),runAfter:(r,l,a,t)=>{at(r,"ctx.scheduler.runAfter");const e=h(v+r,S(l),a,t?.id);return Promise.resolve(e)},runAt:(r,l,a,t)=>{const e=h(r,S(l),a,t?.id);return Promise.resolve(e)}},A=async r=>{d.delete(r.id);const l=u(),a=r.functionPath===void 0?void 0:l.get(r.functionPath);if(a===void 0)throw new k("FUNCTION_NOT_FOUND",`unknown functionPath "${r.functionPath??"(workflow-targeted)"}"`,{status:404});if(a.kind!=="mutation"&&a.kind!=="action")throw new k("FUNCTION_NOT_FOUND",`functionPath "${r.functionPath??"(workflow-targeted)"}" is a ${a.kind} — only mutations and actions can be scheduled`,{status:404});const t=o(),e=a.kind==="action"?i():s();await t(a.kind,a,{...e,now:v},r.args)},I=async r=>{const l=[...d.values()].filter(e=>e.scheduledFor<=r).toSorted((e,w)=>e.scheduledFor-w.scheduledFor),a=[];let t=0;for(const e of l)if(d.has(e.id)){t+=1;try{await A(e)}catch(w){const R=(e.attempts??0)+1;if(R>ut){const _={args:e.args,error:w,functionPath:e.functionPath??"",id:e.id};a.push(_),m.push(_)}else{const _=dt*2**(R-1);d.set(e.id,{...e,attempts:R,scheduledFor:v+_})}}}return{executed:t,failed:a}},T=async(r,l)=>{const{executed:a,failed:t}=await I(r);if(t.length>0&&(l?.throwOnError??!0)){const[e]=t;throw t.length===1&&e!==void 0?e.error:new AggregateError(t.map(w=>w.error),`${String(t.length)} scheduled jobs failed: ${t.map(w=>w.functionPath).join(", ")}`)}return a};return{controls:{advance:(r,l)=>(v+=r,T(v,l)),failures:()=>[...m],list:()=>[...d.values()],runPending:r=>T(Number.POSITIVE_INFINITY,r)},scheduler:x}},ft=()=>{const o=new lt(":memory:"),s=u=>({one(){if(u.length!==1)throw new k("INTERNAL",`expected exactly one row, received ${String(u.length)}`);const[g]=u;return g},[Symbol.iterator](){return u[Symbol.iterator]()},toArray(){return u}});return{close:()=>{o.close()},sql:{exec:(u,...g)=>{const b=o.prepare(u).all(...g);return s(b)}}}},vt=o=>{if(o===null||typeof o.userId!="string"||o.userId.length===0)return{claims:null,userId:null};const{userId:s,...i}=o;return{claims:Object.keys(i).length>0?i:null,userId:s}};class mt extends Y{current=new Y;reset(){this.current=new Y}headroom(){return this.current.headroom()}recordRead(s){this.current.recordRead(s)}recordWrite(s){this.current.recordWrite(s)}}const O=o=>{if(typeof o!="object"||o===null)return;const{kind:s}=o;if(s==="query"||s==="mutation"||s==="action")return s},wt=o=>typeof o=="object"&&o!==null&&o.visibility==="internal"?"internal":"public",B=o=>{throw new k("INTERNAL",`ctx.${o} is not available in the in-memory @lunora/testing harness (v1)`)},y=o=>new Proxy((...s)=>B(o),{apply:()=>B(o),get:()=>B(o)}),V={spanId:"0000000000000001",traceId:"00000000000000000000000000000001"},pt={addEvent:()=>{},addLink:()=>{},recordEvaluation:()=>{},recordException:()=>{},setAttribute:()=>{},setAttributes:()=>{},spanContext:()=>V},gt=()=>{const o={attributes:{},events:[],links:[]},s={addEvent:(i,u)=>{o.events.push({...u===void 0?{}:{attributes:{...u}},name:i})},addLink:i=>{o.links.push({spanId:i.spanId,traceId:i.traceId})},recordEvaluation:i=>{Object.assign(o.attributes,ct(i))},recordException:i=>{const u={"exception.message":i instanceof Error?i.message:String(i),"exception.type":i instanceof Error?i.constructor.name:"Error"};i instanceof Error&&i.stack!==void 0&&(u["exception.stacktrace"]=i.stack),s.addEvent("exception",u)},setAttribute:(i,u)=>{o.attributes[i]=u},setAttributes:i=>{Object.assign(o.attributes,i)},spanContext:()=>V};return{handle:s,recorded:o}},U=async(o,s)=>await s(U,pt),H={count:()=>{},gauge:()=>{},record:()=>{}},Q={debug:()=>{},error:()=>{},event:()=>{},fatal:()=>{},info:()=>{},log:()=>{},trace:()=>{},warn:()=>{},with:()=>Q},yt=(o,s,i)=>(g,v)=>{let b=!1;const d=[];let m,h,S=0,x=0;const A=()=>O(g)?o("query",g,s,v,!1):Promise.resolve(g(s)),I=(t,e)=>{if(t<x)return;x=t;const w={done:!1,value:e};if(d.length===0)m=w,h=void 0;else{m=void 0,h=void 0;for(const R of d.splice(0))R.resolve(w)}},T=(t,e)=>{if(!(t<x))if(x=t,d.length===0)h={error:e},m=void 0;else{m=void 0,h=void 0;for(const w of d.splice(0))w.reject(e)}},P=t=>e=>{I(t,e)},r=t=>e=>{T(t,e)},l=()=>{if(b)return;S+=1;const t=S;A().then(P(t)).catch(r(t))};i.add(l);const a={[Symbol.asyncIterator](){return a},next:()=>{if(b)return Promise.resolve({done:!0,value:void 0});if(x===S){if(h!==void 0){const{error:t}=h;return h=void 0,Promise.reject(t)}if(m!==void 0){const t=m;return m=void 0,Promise.resolve(t)}}return x<S?new Promise((t,e)=>{d.push({reject:e,resolve:t})}):A().then(t=>{if(h!==void 0){const{error:e}=h;throw h=void 0,e}if(m!==void 0){const e=m;return m=void 0,e}return{done:!1,value:t}})},return:()=>{b=!0,i.delete(l);for(const t of d.splice(0))t.resolve({done:!0,value:void 0});return Promise.resolve({done:!0,value:void 0})}};return A().then(P(0)).catch(r(0)),a},Nt=(o,s)=>{const{close:i,sql:u}=ft(),g=o;ot(u,g);const v=new mt,b=f=>{const p=st({auth:{identity:f.claims,userId:f.userId},enforceRls:s?.enforceRls??!0,headroom:v,schema:g,sql:u});return{database:p,rawDatabase:p[it]??p}},d=f=>{u.exec.call(u,f)};let m=Promise.resolve();const h=f=>{const p=async()=>{const M=nt({scheduler:W});v.reset(),d("BEGIN");try{const F=await f();return d("COMMIT"),await M(!0),F}catch(F){try{d("ROLLBACK")}catch{}throw await M(!1),F}},C=m.then(p);return m=C.then(()=>{},()=>{}),C};let S=!1;const x=()=>{S||(S=!0,i())},A=new Map(Object.entries(s?.functions??{}).map(([f,p])=>[f,p])),I=new Set,T=()=>{for(const f of I)f()},P=f=>(T(),f);let r,l,a;const t=s?.now??Date.now(),e=gt(),w=(f,p)=>{if(f===void 0)throw new k("INTERNAL",`[fake-scheduler] ${p} not yet available — scheduler.advance called before harness construction completed`);return f},{controls:R,scheduler:_}=ht(()=>w(r,"dispatch"),()=>w(l,"mutationContext"),()=>w(a,"actionContext"),()=>A,t),W=rt(_),J=f=>{const p=vt(f),C={getIdentity:()=>Promise.resolve(p.claims),userId:p.userId},{database:M,rawDatabase:F}=b(p),D={auth:C,db:M,env:s?.env,log:Q,metrics:H,now:t,span:e.handle,trace:U,runQuery:((n,c)=>N("query",n,D,c)),secrets:y("secrets"),storage:y("storage"),vectors:y("vectors")},L={auth:C,db:M,env:s?.env,log:Q,metrics:H,now:t,span:e.handle,trace:U,runMutation:((n,c)=>N("mutation",n,L,c)),runQuery:((n,c)=>N("query",n,D,c)),scheduler:W,secrets:y("secrets"),storage:y("storage"),vectors:y("vectors"),workflows:y("workflows")};l??=L;const X={...L,db:F},j={auth:C,db:M,env:s?.env,fetch:s?.fetch??y("fetch"),log:Q,metrics:H,now:t,span:e.handle,trace:U,runAction:((n,c)=>N("action",n,j,c)),runMutation:((n,c)=>h(()=>N("mutation",n,L,c)).then(P)),runQuery:((n,c)=>N("query",n,D,c)),scheduler:W,secrets:y("secrets"),storage:y("storage"),vectors:y("vectors"),workflows:y("workflows")};a??=j;const q=(n,c,E,$,et)=>{const K=O(c);if(K!==n)throw new k("INTERNAL",`expected a registered ${n}, received a ${K??"non-function"} reference`);if(!et&&wt(c)==="internal")throw new k("INTERNAL",`This ${n} is an internal function — it is unreachable from the external RPC boundary in production. Call it through ctx.run${n.charAt(0).toUpperCase()}${n.slice(1)} from another function instead.`);return Promise.resolve(c.handler(E,$??{}))},N=(n,c,E,$)=>q(n,c,E,$,!0);r??=(n,c,E,$)=>n==="mutation"?h(()=>q("mutation",c,E,$,!0)).then(P):(v.reset(),N("action",c,E,$));const G=((n,c)=>(v.reset(),O(n)?q("query",n,D,c,!1):Promise.resolve(n(D)))),z=((n,c)=>{const E=O(n)?()=>q("mutation",n,L,c,!1):()=>n(L);return h(E).then(P)}),Z=((n,c)=>(v.reset(),O(n)?q("action",n,j,c,!1):Promise.resolve(n(j)))),tt=yt((...n)=>(v.reset(),q(...n)),D,I);return{action:Z,close:x,mutation:z,query:G,run:n=>h(()=>n(X)).then(P),scheduler:R,subscribe:tt,wideEvent:()=>e.recorded,withIdentity:n=>J(n)}};return J(null)};export{Nt as lunoraTest};

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lunora/testing",
-  "version": "1.0.0-alpha.140",
+  "version": "1.0.0-alpha.142",
   "description": "Testing toolkit for Lunora: an in-memory harness for queries, mutations, and actions",
   "keywords": [
     "cloudflare",
@@ -50,12 +50,12 @@
     "access": "public"
   },
   "dependencies": {
-    "@lunora/agent": "1.0.0-alpha.81",
-    "@lunora/errors": "1.0.0-alpha.29",
-    "@lunora/mail": "1.0.0-alpha.59",
-    "@lunora/scheduler": "1.0.0-alpha.49",
-    "@lunora/server": "1.0.0-alpha.99",
-    "@lunora/shard-engine": "1.0.0-alpha.52"
+    "@lunora/agent": "1.0.0-alpha.83",
+    "@lunora/errors": "1.0.0-alpha.30",
+    "@lunora/mail": "1.0.0-alpha.61",
+    "@lunora/scheduler": "1.0.0-alpha.50",
+    "@lunora/server": "1.0.0-alpha.100",
+    "@lunora/shard-engine": "1.0.0-alpha.53"
   },
   "peerDependencies": {
     "@playwright/test": "^1.61.1"

package/dist/packem_shared/containsScorer-gv0Sdd86.mjs

@@ -1,2 +0,0 @@
-import{LunoraError as g}from"@lunora/errors";const l=/^\s*(-?\d+(?:\.\d+)?)/u,i=e=>Number.isFinite(e)?Math.min(1,Math.max(0,e)):0,h=e=>typeof e=="number"?{score:i(e)}:{score:i(e.score),...e.reason===void 0?{}:{reason:e.reason}},u=e=>e.length===0?0:e.reduce((t,o)=>t+o,0)/e.length,y=(e,t={})=>({name:`contains:${e}`,score:({output:o})=>{const a=t.caseSensitive?o:o.toLowerCase(),r=t.caseSensitive?e:e.toLowerCase();return a.includes(r)?1:0}}),v=(e,t="regex")=>({name:t,score:({output:o})=>e.test(o)?1:0}),$=()=>({name:"exact-match",score:({expected:e,output:t})=>t.trim()===e?.trim()?1:0}),L=e=>{if(e.length===0)throw new g("BAD_REQUEST","@lunora/testing: keywordScorer requires at least one keyword");return{name:"keyword-coverage",score:({output:t})=>{const o=t.toLowerCase(),a=e.filter(r=>o.includes(r.toLowerCase())).length;return{reason:`${String(a)}/${String(e.length)} keywords present`,score:a/e.length}}}},S=(e,t)=>[`Rate the ASSISTANT OUTPUT against this criterion: ${e}`,"Respond with a single number from 0 (fails) to 1 (fully meets), then a dash and a one-line reason.",...t.input===void 0?[]:["",`Input: ${t.input}`],...t.expected===void 0?[]:["",`Reference answer: ${t.expected}`],"",`Assistant output: ${t.output}`].join(`
-`),w=e=>{const t=l.exec(e);return{reason:e.trim(),score:t?i(Number(t[1])):0}},b=e=>({name:e.name??"llm-judge",score:async t=>w(await e.judge(S(e.criteria,t)))}),f=async(e,t)=>{const o=await Promise.all(t.map(async n=>({name:n.name,result:h(await n.score(e))}))),a={},r=new Map;for(const{name:n,result:s}of o){const c=r.get(n)??0;r.set(n,c+1),a[c===0?n:`${n}#${String(c+1)}`]=s}return{average:u(o.map(({result:n})=>n.score)),scores:a}},A=async(e,t,o)=>{const a=await Promise.all(e.map(async r=>{const n=await t(r.input),s=typeof n=="string"?n:n.output,c=typeof n=="string"||n.metadata===void 0?r.metadata:{...r.metadata,...n.metadata},m={input:r.input,output:s,...r.expected===void 0?{}:{expected:r.expected},...c===void 0?{}:{metadata:c}},{average:d,scores:p}=await f(m,o);return{average:d,input:r.input,output:s,scores:p}}));return{average:u(a.map(r=>r.average)),items:a}};export{y as containsScorer,A as evaluate,$ as exactMatchScorer,L as keywordScorer,b as llmScorer,w as parseJudgeScore,v as regexScorer,f as scoreSample};

package/dist/packem_shared/lunoraTest-BOqU8mY2.mjs

@@ -1 +0,0 @@
-import{LunoraError as _}from"@lunora/errors";import{runShardMigrations as O,createShardCtxDb as z,RLS_UNWRAP_SYMBOL as Z}from"@lunora/shard-engine";import{evaluationAttributes as tt}from"./evaluationAttributes-CYLgTP1P.mjs";import{MAX_RETRY_ATTEMPTS as et,RETRY_BASE_DELAY_MS as nt}from"@lunora/scheduler";import{DatabaseSync as ot}from"node:sqlite";const rt=(r,c,s,u,g)=>{let p=g,y=1;const d=new Map,f=[],h=(e,l,a={})=>{const t=`fake-job-${String(y)}`;return y+=1,d.set(t,{args:a,enqueuedAt:p,functionPath:l,id:t,scheduledFor:e}),t},b=e=>typeof e=="string"?e:e.name??e.binding??"",P={cancel:e=>{const l=d.has(e);return d.delete(e),Promise.resolve({cancelled:l})},get:e=>Promise.resolve(d.get(e)??null),list:()=>Promise.resolve([...d.values()]),runAfter:(e,l,a)=>{const t=h(p+e,b(l),a);return Promise.resolve(t)},runAt:(e,l,a)=>{const t=h(e,b(l),a);return Promise.resolve(t)}},k=async e=>{d.delete(e.id);const l=u(),a=e.functionPath===void 0?void 0:l.get(e.functionPath);if(a===void 0){console.warn(`[fake-scheduler] unknown functionPath "${e.functionPath??"(workflow-targeted)"}" — job ${e.id} dropped`);return}if(a.kind==="mutation"||a.kind==="action"){const t=r(),n=a.kind==="action"?s():c();await t(a.kind,a,n,e.args)}else console.warn(`[fake-scheduler] functionPath "${e.functionPath??"(workflow-targeted)"}" is a ${a.kind} — only mutations and actions can be scheduled; job ${e.id} dropped`)},M=async e=>{const l=[...d.values()].filter(n=>n.scheduledFor<=e).toSorted((n,v)=>n.scheduledFor-v.scheduledFor),a=[];let t=0;for(const n of l)if(d.has(n.id)){t+=1;try{await k(n)}catch(v){const T=(n.attempts??0)+1;if(T>et){const N={args:n.args,error:v,functionPath:n.functionPath??"",id:n.id};a.push(N),f.push(N)}else{const N=nt*2**(T-1);d.set(n.id,{...n,attempts:T,scheduledFor:p+N})}}}return{executed:t,failed:a}},C=async(e,l)=>{const{executed:a,failed:t}=await M(e);if(t.length>0&&(l?.throwOnError??!0)){const[n]=t;throw t.length===1&&n!==void 0?n.error:new AggregateError(t.map(v=>v.error),`${String(t.length)} scheduled jobs failed: ${t.map(v=>v.functionPath).join(", ")}`)}return a};return{controls:{advance:(e,l)=>(p+=e,C(p,l)),failures:()=>[...f],list:()=>[...d.values()],runPending:e=>C(Number.POSITIVE_INFINITY,e)},scheduler:P}},st=()=>{const r=new ot(":memory:"),c=u=>({one(){if(u.length!==1)throw new _("INTERNAL",`expected exactly one row, received ${String(u.length)}`);const[g]=u;return g},[Symbol.iterator](){return u[Symbol.iterator]()},toArray(){return u}});return{close:()=>{r.close()},sql:{exec:(u,...g)=>{const y=r.prepare(u).all(...g);return c(y)}}}},q=r=>{if(typeof r!="object"||r===null)return;const{kind:c}=r;if(c==="query"||c==="mutation"||c==="action")return c},it=r=>typeof r=="object"&&r!==null&&r.visibility==="internal"?"internal":"public",Q=r=>{throw new _("INTERNAL",`ctx.${r} is not available in the in-memory @lunora/testing harness (v1)`)},w=r=>new Proxy((...c)=>Q(r),{apply:()=>Q(r),get:()=>Q(r)}),K={spanId:"0000000000000001",traceId:"00000000000000000000000000000001"},ct={addEvent:()=>{},addLink:()=>{},recordEvaluation:()=>{},recordException:()=>{},setAttribute:()=>{},setAttributes:()=>{},spanContext:()=>K},at=()=>{const r={attributes:{},events:[],links:[]},c={addEvent:(s,u)=>{r.events.push({...u===void 0?{}:{attributes:{...u}},name:s})},addLink:s=>{r.links.push({spanId:s.spanId,traceId:s.traceId})},recordEvaluation:s=>{Object.assign(r.attributes,tt(s))},recordException:s=>{const u={"exception.message":s instanceof Error?s.message:String(s),"exception.type":s instanceof Error?s.constructor.name:"Error"};s instanceof Error&&s.stack!==void 0&&(u["exception.stacktrace"]=s.stack),c.addEvent("exception",u)},setAttribute:(s,u)=>{r.attributes[s]=u},setAttributes:s=>{Object.assign(r.attributes,s)},spanContext:()=>K};return{handle:c,recorded:r}},D=async(r,c)=>await c(D,ct),Y={count:()=>{},gauge:()=>{},record:()=>{}},j={debug:()=>{},error:()=>{},event:()=>{},fatal:()=>{},info:()=>{},log:()=>{},trace:()=>{},warn:()=>{},with:()=>j},ut=(r,c,s)=>(g,p)=>{let y=!1;const d=[];let f,h,b=0,P=0;const k=()=>q(g)?r("query",g,c,p,!1):Promise.resolve(g(c)),M=(t,n)=>{if(t<P)return;P=t;const v={done:!1,value:n};if(d.length===0)f=v,h=void 0;else{f=void 0,h=void 0;for(const T of d.splice(0))T.resolve(v)}},C=(t,n)=>{if(!(t<P))if(P=t,d.length===0)h={error:n},f=void 0;else{f=void 0,h=void 0;for(const v of d.splice(0))v.reject(n)}},R=t=>n=>{M(t,n)},e=t=>n=>{C(t,n)},l=()=>{if(y)return;b+=1;const t=b;k().then(R(t)).catch(e(t))};s.add(l);const a={[Symbol.asyncIterator](){return a},next:()=>{if(y)return Promise.resolve({done:!0,value:void 0});if(P===b){if(h!==void 0){const{error:t}=h;return h=void 0,Promise.reject(t)}if(f!==void 0){const t=f;return f=void 0,Promise.resolve(t)}}return P<b?new Promise((t,n)=>{d.push({reject:n,resolve:t})}):k().then(t=>{if(h!==void 0){const{error:n}=h;throw h=void 0,n}if(f!==void 0){const n=f;return f=void 0,n}return{done:!1,value:t}})},return:()=>{y=!0,s.delete(l);for(const t of d.splice(0))t.resolve({done:!0,value:void 0});return Promise.resolve({done:!0,value:void 0})}};return k().then(R(0)).catch(e(0)),a},pt=(r,c)=>{const{close:s,sql:u}=st(),g=r;O(u,g);const p=z({enforceRls:c?.enforceRls??!0,schema:g,sql:u}),y=p[Z]??p,d=m=>{u.exec.call(u,m)};let f=Promise.resolve();const h=m=>{const E=async()=>{d("BEGIN");try{const S=await m();return d("COMMIT"),S}catch(S){try{d("ROLLBACK")}catch{}throw S}},x=f.then(E);return f=x.then(()=>{},()=>{}),x};let b=!1;const P=()=>{b||(b=!0,s())},k=new Map(Object.entries(c?.functions??{}).map(([m,E])=>[m,E])),M=new Set,C=()=>{for(const m of M)m()},R=m=>(C(),m);let e,l,a;const t=c?.now??Date.now(),n=at(),v=(m,E)=>{if(m===void 0)throw new _("INTERNAL",`[fake-scheduler] ${E} not yet available — scheduler.advance called before harness construction completed`);return m},{controls:T,scheduler:N}=rt(()=>v(e,"dispatch"),()=>v(l,"mutationContext"),()=>v(a,"actionContext"),()=>k,t),B=m=>{const E={getIdentity:()=>Promise.resolve(m??null),userId:m?.userId??null},x={auth:E,db:p,env:c?.env,log:j,metrics:Y,now:t,span:n.handle,trace:D,runQuery:((o,i)=>I("query",o,x,i)),secrets:w("secrets"),storage:w("storage"),vectors:w("vectors")},S={auth:E,db:p,env:c?.env,log:j,metrics:Y,now:t,span:n.handle,trace:D,runMutation:((o,i)=>I("mutation",o,S,i)),runQuery:((o,i)=>I("query",o,x,i)),scheduler:N,secrets:w("secrets"),storage:w("storage"),vectors:w("vectors"),workflows:w("workflows")};l??=S;const U={...S,db:y},F={auth:E,db:p,env:c?.env,fetch:c?.fetch??w("fetch"),log:j,metrics:Y,now:t,span:n.handle,trace:D,runAction:((o,i)=>I("action",o,F,i)),runMutation:((o,i)=>I("mutation",o,S,i)),runQuery:((o,i)=>I("query",o,x,i)),scheduler:N,secrets:w("secrets"),storage:w("storage"),vectors:w("vectors"),workflows:w("workflows")};a??=F;const $=(o,i,A,L,J)=>{const H=q(i);if(H!==o)throw new _("INTERNAL",`expected a registered ${o}, received a ${H??"non-function"} reference`);if(!J&&it(i)==="internal")throw new _("INTERNAL",`This ${o} is an internal function — it is unreachable from the external RPC boundary in production. Call it through ctx.run${o.charAt(0).toUpperCase()}${o.slice(1)} from another function instead.`);return Promise.resolve(i.handler(A,L??{}))},I=(o,i,A,L)=>$(o,i,A,L,!0);e??=(o,i,A,L)=>o==="mutation"?h(()=>$("mutation",i,A,L,!0)).then(R):I("action",i,A,L);const V=((o,i)=>q(o)?$("query",o,x,i,!1):Promise.resolve(o(x))),W=((o,i)=>{const A=q(o)?()=>$("mutation",o,S,i,!1):()=>o(S);return h(A).then(R)}),X=((o,i)=>q(o)?$("action",o,F,i,!1):Promise.resolve(o(F))),G=ut($,x,M);return{action:X,close:P,mutation:W,query:V,run:o=>h(()=>o(U)).then(R),scheduler:T,subscribe:G,wideEvent:()=>n.recorded,withIdentity:o=>B(o)}};return B(null)};export{pt as lunoraTest};