@rulemetric/cli 0.14.2 → 0.15.0

package/dist/chunk-V576KKH6.js

@@ -0,0 +1 @@
+import{a as m,c as Ds}from"./chunk-ZLLER3HP.js";import{boolean as sn,date as js,index as Qt,integer as zs,jsonb as Rs,pgTable as Yt,text as D,timestamp as G,uniqueIndex as ks,uuid as me}from"drizzle-orm/pg-core";var an=Yt("announcements",{id:me("id").primaryKey().defaultRandom(),version:D("version").notNull().unique(),releaseDate:js("release_date").notNull(),subject:D("subject").notNull(),htmlBody:D("html_body").notNull(),textBody:D("text_body").notNull(),sections:Rs("sections").notNull().default({}),status:D("status").notNull().default("draft"),recipientCount:zs("recipient_count"),sentAt:G("sent_at",{withTimezone:!0}),sentBy:me("sent_by"),error:D("error"),createdAt:G("created_at",{withTimezone:!0}).notNull().defaultNow(),updatedAt:G("updated_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Qt("idx_announcements_status").on(e.status),Qt("idx_announcements_created").on(e.createdAt)]),dn=Yt("announcement_deliveries",{id:me("id").primaryKey().defaultRandom(),announcementId:me("announcement_id").notNull(),userId:me("user_id").notNull(),email:D("email").notNull(),status:D("status").notNull().default("pending"),providerId:D("provider_id"),error:D("error"),sentAt:G("sent_at",{withTimezone:!0}),createdAt:G("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[ks("uniq_announcement_deliveries_announcement_user").on(e.announcementId,e.userId),Qt("idx_announcement_deliveries_user").on(e.userId,e.createdAt)]),un=Yt("notification_preferences",{userId:me("user_id").primaryKey(),productAnnouncements:sn("product_announcements").notNull().default(!0),weeklyDigest:sn("weekly_digest").notNull().default(!0),unsubscribeToken:D("unsubscribe_token").notNull().unique(),unsubscribedAt:G("unsubscribed_at",{withTimezone:!0}),createdAt:G("created_at",{withTimezone:!0}).notNull().defaultNow(),updatedAt:G("updated_at",{withTimezone:!0}).notNull().defaultNow()});import{boolean as lt,integer as qe,jsonb as Es,pgTable as Ss,real as pn,text as Oe,timestamp as Wt,uuid as _n}from"drizzle-orm/pg-core";import{pgTable as ut,text as fe,timestamp as Ce,jsonb as ln,uuid as ne,primaryKey as cn}from"drizzle-orm/pg-core";var l=ut("organizations",{id:ne("id").primaryKey().defaultRandom(),name:fe("name").unique().notNull(),slug:fe("slug").unique().notNull(),metadata:ln("metadata"),createdAt:Ce("created_at",{withTimezone:!0}).defaultNow(),updatedAt:Ce("updated_at",{withTimezone:!0}).defaultNow()}),v=ut("teams",{id:ne("id").primaryKey().defaultRandom(),orgId:ne("org_id").references(()=>l.id,{onDelete:"cascade"}),name:fe("name").notNull(),slug:fe("slug").notNull(),metadata:ln("metadata"),createdAt:Ce("created_at",{withTimezone:!0}).defaultNow()}),Ne=ut("user_organizations",{userId:ne("user_id").notNull(),orgId:ne("org_id").references(()=>l.id,{onDelete:"cascade"}).notNull(),role:fe("role").notNull().default("member"),createdAt:Ce("created_at",{withTimezone:!0}).defaultNow()},e=>[cn({columns:[e.userId,e.orgId]})]),xe=ut("user_teams",{userId:ne("user_id").notNull(),teamId:ne("team_id").references(()=>v.id,{onDelete:"cascade"}).notNull(),role:fe("role").notNull().default("member"),createdAt:Ce("created_at",{withTimezone:!0}).defaultNow()},e=>[cn({columns:[e.userId,e.teamId]})]);var o=Ss("profiles",{id:_n("id").primaryKey(),email:Oe("email"),measurementPreferences:Es("measurement_preferences"),displayName:Oe("display_name"),activeOrgId:_n("active_org_id").references(()=>l.id,{onDelete:"set null"}),isSuperadmin:lt("is_superadmin").notNull().default(!1),claudeCodePlan:Oe("claude_code_plan"),planSource:Oe("plan_source").notNull().default("detected"),setupCompletedAt:Wt("setup_completed_at",{withTimezone:!0}),liveSendSkipped:lt("live_send_skipped").notNull().default(!1),autoAcceptSuggestions:lt("auto_accept_suggestions").notNull().default(!0),evolutionPaused:lt("evolution_paused").notNull().default(!1),suggestionConfidenceThreshold:pn("suggestion_confidence_threshold").notNull().default(.7),autoAcceptMinScore:pn("auto_accept_min_score"),autoAcceptMaxPerRun:qe("auto_accept_max_per_run"),autoAcceptMaxPerWindow:qe("auto_accept_max_per_window"),autoAcceptWindowDays:qe("auto_accept_window_days"),trialMaxConcurrent:qe("trial_max_concurrent"),trialMaxPerRun:qe("trial_max_per_run"),harborAbCadence:Oe("harbor_ab_cadence").notNull().default("off"),createdAt:Wt("created_at",{withTimezone:!0}).defaultNow(),updatedAt:Wt("updated_at",{withTimezone:!0}).defaultNow()});import{pgTable as bs,text as Xt,timestamp as ct,uuid as pt,index as mn}from"drizzle-orm/pg-core";var P=bs("org_invitations",{id:pt("id").primaryKey().defaultRandom(),orgId:pt("org_id").notNull().references(()=>l.id,{onDelete:"cascade"}),email:Xt("email").notNull(),role:Xt("role").notNull().default("member"),teamId:pt("team_id").references(()=>v.id,{onDelete:"set null"}),invitedBy:pt("invited_by").notNull().references(()=>o.id,{onDelete:"cascade"}),token:Xt("token").notNull().unique(),expiresAt:ct("expires_at",{withTimezone:!0}).notNull(),acceptedAt:ct("accepted_at",{withTimezone:!0}),revokedAt:ct("revoked_at",{withTimezone:!0}),createdAt:ct("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[mn("idx_org_invitations_token").on(e.token),mn("idx_org_invitations_org").on(e.orgId)]);import{pgTable as Zt,text as N,timestamp as V,jsonb as _t,uuid as z,integer as mt,boolean as Ps,uniqueIndex as ft,index as Me}from"drizzle-orm/pg-core";var n=Zt("instructions",{id:z("id").primaryKey().defaultRandom(),name:N("name").notNull(),description:N("description").notNull(),type:N("type").notNull().default("instruction"),scope:N("scope").notNull(),scopeRef:z("scope_ref"),version:N("version").notNull(),content:N("content").notNull(),frontmatter:_t("frontmatter").notNull(),slots:_t("slots"),tags:N("tags").array(),compatibility:N("compatibility").array(),createdBy:z("created_by"),createdAt:V("created_at",{withTimezone:!0}).defaultNow(),updatedAt:V("updated_at",{withTimezone:!0}).defaultNow(),archived:Ps("archived").default(!1),archivedAt:V("archived_at",{withTimezone:!0}),archivedBy:z("archived_by"),forkedFromId:z("forked_from_id").references(()=>n.id,{onDelete:"set null"}),upstreamVersion:N("upstream_version"),badge:N("badge"),contentHash:N("content_hash"),contextIdentity:N("context_identity"),scanScore:mt("scan_score"),scanFindings:_t("scan_findings"),popularityScore:mt("popularity_score"),sourceUpdatedAt:V("source_updated_at",{withTimezone:!0})},e=>[ft("instructions_name_scope_ref_idx").on(e.name,e.scope,e.scopeRef),Me("idx_instructions_scope").on(e.scope,e.scopeRef),Me("idx_instructions_created_by").on(e.createdBy),Me("idx_instructions_context_identity").on(e.createdBy,e.contextIdentity),Me("idx_instructions_public_recency").on(e.scope,e.sourceUpdatedAt.desc())]),fn=Zt("instruction_version_identities",{id:z("id").primaryKey().defaultRandom(),instructionId:z("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),contentHash:N("content_hash").notNull(),versionOrdinal:mt("version_ordinal").notNull(),content:N("content"),firstSeenAt:V("first_seen_at",{withTimezone:!0}).notNull().defaultNow(),lastSeenAt:V("last_seen_at",{withTimezone:!0}).notNull().defaultNow(),createdAt:V("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[ft("instruction_version_identities_hash_unique").on(e.instructionId,e.contentHash),ft("instruction_version_identities_ordinal_unique").on(e.instructionId,e.versionOrdinal),Me("idx_ivi_instruction").on(e.instructionId)]),he=Zt("instruction_versions",{id:z("id").primaryKey().defaultRandom(),instructionId:z("instruction_id").references(()=>n.id,{onDelete:"cascade"}),versionNumber:mt("version_number").notNull(),content:N("content").notNull(),frontmatter:_t("frontmatter").notNull(),changeSummary:N("change_summary"),createdBy:z("created_by"),createdAt:V("created_at",{withTimezone:!0}).defaultNow()},e=>[ft("instruction_versions_instruction_version_idx").on(e.instructionId,e.versionNumber)]);import{pgTable as Ks,text as Cs,timestamp as qs,jsonb as Os,uuid as eo,index as Nn}from"drizzle-orm/pg-core";var ge=Ks("instruction_events",{id:eo("id").primaryKey().defaultRandom(),instructionId:eo("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),eventType:Cs("event_type").notNull(),actorId:eo("actor_id").notNull(),metadata:Os("metadata"),createdAt:qs("created_at",{withTimezone:!0}).defaultNow()},e=>[Nn("idx_instruction_events_instruction").on(e.instructionId),Nn("idx_instruction_events_type").on(e.eventType)]);import{boolean as Ms,jsonb as Hs,pgTable as oo,primaryKey as xn,text as re,timestamp as He,uniqueIndex as to,uuid as se}from"drizzle-orm/pg-core";import{sql as Ls}from"drizzle-orm";var s=oo("projects",{id:se("id").primaryKey().defaultRandom(),orgId:se("org_id").references(()=>l.id,{onDelete:"cascade"}),name:re("name").notNull(),description:re("description"),githubRepo:re("github_repo"),gitRemoteUrl:re("git_remote_url"),gitRootCommit:re("git_root_commit"),gitRemoteAliases:re("git_remote_aliases").array().notNull().default([]),metadata:Hs("metadata"),createdBy:se("created_by"),createdAt:He("created_at",{withTimezone:!0}).defaultNow(),updatedAt:He("updated_at",{withTimezone:!0}).defaultNow()},e=>[to("projects_org_name_idx").on(e.orgId,e.name),to("projects_git_remote_url_unique").on(e.gitRemoteUrl).where(Ls`${e.gitRemoteUrl} IS NOT NULL`)]),F=oo("project_instructions",{projectId:se("project_id").references(()=>s.id,{onDelete:"cascade"}).notNull(),instructionId:se("instruction_id").references(()=>n.id,{onDelete:"cascade"}).notNull(),addedAt:He("added_at",{withTimezone:!0}).defaultNow(),addedBy:se("added_by")},e=>[xn({columns:[e.projectId,e.instructionId]})]),hn=oo("project_checkouts",{projectId:se("project_id").references(()=>s.id,{onDelete:"cascade"}).notNull(),path:re("path").notNull(),isWorktree:Ms("is_worktree").notNull().default(!1),firstSeen:He("first_seen",{withTimezone:!0}).notNull().defaultNow(),lastSeen:He("last_seen",{withTimezone:!0}).notNull().defaultNow()},e=>[xn({columns:[e.projectId,e.path]}),to("project_checkouts_path_unique").on(e.path)]);import{pgTable as so,text as I,timestamp as Le,jsonb as ie,uuid as K,integer as xt,uniqueIndex as vn,index as R}from"drizzle-orm/pg-core";import{pgTable as Us,text as ve,timestamp as no,jsonb as gn,uuid as Nt,integer as Bs,boolean as Gs,index as ro}from"drizzle-orm/pg-core";var h=Us("agent_runs",{id:Nt("id").primaryKey().defaultRandom(),userId:Nt("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),orgId:Nt("org_id").references(()=>l.id,{onDelete:"set null"}),projectId:Nt("project_id").references(()=>s.id,{onDelete:"set null"}),projectPath:ve("project_path"),gitRemote:ve("git_remote"),baseBranch:ve("base_branch"),title:ve("title"),prompt:ve("prompt"),toolSpec:gn("tool_spec").notNull().default([]),autoWorktree:Gs("auto_worktree").notNull().default(!0),status:ve("status").notNull().default("pending"),sessionCount:Bs("session_count").notNull().default(0),metadata:gn("metadata").notNull().default({}),createdAt:no("created_at",{withTimezone:!0}).notNull().defaultNow(),updatedAt:no("updated_at",{withTimezone:!0}).notNull().defaultNow(),completedAt:no("completed_at",{withTimezone:!0})},e=>[ro("idx_agent_runs_user").on(e.userId),ro("idx_agent_runs_org_created").on(e.orgId,e.createdAt),ro("idx_agent_runs_project").on(e.projectId)]);var a=so("sessions",{id:K("id").primaryKey().defaultRandom(),userId:K("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),externalSessionId:I("external_session_id"),projectId:K("project_id").references(()=>s.id,{onDelete:"set null"}),orgId:K("org_id").references(()=>l.id,{onDelete:"set null"}),tool:I("tool").notNull(),projectPath:I("project_path"),gitBranch:I("git_branch"),startedAt:Le("started_at",{withTimezone:!0}).notNull(),endedAt:Le("ended_at",{withTimezone:!0}),eventCount:xt("event_count").default(0),tokenUsage:ie("token_usage"),analysis:ie("analysis"),analyzedAt:Le("analyzed_at",{withTimezone:!0}),analyzedEventCount:xt("analyzed_event_count"),analyzedTurnCount:xt("analyzed_turn_count"),metadata:ie("metadata").default({}),origin:I("origin").notNull().default("unknown"),runId:K("run_id").references(()=>h.id,{onDelete:"set null"}),model:I("model"),harnessVersion:I("harness_version"),createdAt:Le("created_at",{withTimezone:!0}).defaultNow()},e=>[R("idx_sessions_user").on(e.userId),R("idx_sessions_tool").on(e.tool),R("idx_sessions_started_at").on(e.startedAt),R("idx_sessions_project").on(e.projectId),R("idx_sessions_org_started_at").on(e.orgId,e.startedAt),R("idx_sessions_run").on(e.runId),vn("idx_sessions_user_external").on(e.userId,e.externalSessionId)]),Ie=so("session_events",{id:K("id").primaryKey().defaultRandom(),sessionId:K("session_id").notNull().references(()=>a.id,{onDelete:"cascade"}),sequence:xt("sequence").notNull(),eventType:I("event_type").notNull(),content:I("content"),toolName:I("tool_name"),toolInput:ie("tool_input"),toolOutput:ie("tool_output"),tokenUsage:ie("token_usage"),model:I("model"),timestamp:Le("timestamp",{withTimezone:!0}).notNull(),metadata:ie("metadata").default({})},e=>[vn("session_events_session_sequence_idx").on(e.sessionId,e.sequence),R("idx_session_events_session_id").on(e.sessionId),R("idx_session_events_event_type").on(e.eventType),R("idx_session_events_model").on(e.model)]),$=so("session_instructions",{sessionId:K("session_id").notNull().references(()=>a.id,{onDelete:"cascade"}),instructionId:K("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),linkMethod:I("link_method").notNull(),contentHash:I("content_hash")},e=>[R("idx_session_instructions_instruction_id").on(e.instructionId)]);import{pgTable as In,text as ht,timestamp as yn,uuid as k,integer as wn,index as ye,unique as Vs}from"drizzle-orm/pg-core";var C=In("treatment_assignments",{id:k("id").primaryKey().defaultRandom(),userId:k("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),treatmentType:ht("treatment_type").notNull(),treatmentId:k("treatment_id").notNull(),treatmentVersion:wn("treatment_version").notNull().default(1),projectId:k("project_id").references(()=>s.id,{onDelete:"set null"}),assignedBy:k("assigned_by").references(()=>o.id,{onDelete:"set null"}),method:ht("method").notNull(),assignedAt:yn("assigned_at",{withTimezone:!0}).notNull().defaultNow()},e=>[ye("idx_treatment_assignments_user").on(e.userId),ye("idx_treatment_assignments_treatment").on(e.treatmentType,e.treatmentId),ye("idx_treatment_assignments_project").on(e.projectId)]),q=In("treatment_exposures",{id:k("id").primaryKey().defaultRandom(),sessionId:k("session_id").notNull().references(()=>a.id,{onDelete:"cascade"}),userId:k("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),treatmentType:ht("treatment_type").notNull(),treatmentId:k("treatment_id").notNull(),treatmentVersion:wn("treatment_version").notNull().default(1),assignmentId:k("assignment_id").references(()=>C.id,{onDelete:"set null"}),method:ht("method").notNull(),exposedAt:yn("exposed_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Vs("treatment_exposures_unique").on(e.sessionId,e.treatmentType,e.treatmentId,e.treatmentVersion),ye("idx_treatment_exposures_treatment").on(e.treatmentType,e.treatmentId),ye("idx_treatment_exposures_user").on(e.userId),ye("idx_treatment_exposures_session").on(e.sessionId)]);import{pgTable as Fs,text as io,timestamp as $s,uuid as gt,integer as Js,index as Tn,uniqueIndex as Qs}from"drizzle-orm/pg-core";var An=Fs("pending_treatment_exposures",{id:gt("id").primaryKey().defaultRandom(),userId:gt("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),externalSessionId:io("external_session_id").notNull(),treatmentType:io("treatment_type").notNull(),treatmentId:gt("treatment_id").notNull(),treatmentVersion:Js("treatment_version").notNull().default(1),method:io("method").notNull().default("injected"),projectId:gt("project_id").references(()=>s.id,{onDelete:"set null"}),createdAt:$s("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Qs("pending_treatment_exposures_unique").on(e.externalSessionId,e.treatmentType,e.treatmentId,e.treatmentVersion),Tn("idx_pending_treatment_exposures_ext").on(e.externalSessionId),Tn("idx_pending_treatment_exposures_user").on(e.userId)]);import{pgTable as Dn,text as J,timestamp as jn,uuid as we,numeric as Ys,index as ao,unique as Ws}from"drizzle-orm/pg-core";var Ue=Dn("instruction_experiments",{id:we("id").primaryKey().defaultRandom(),userId:we("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),instructionId:we("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),projectPath:J("project_path").notNull(),status:J("status").notNull().default("running"),treatedShare:Ys("treated_share").notNull().default("0.5"),salt:J("salt").notNull(),contentHash:J("content_hash"),source:J("source").notNull().default("manual"),stoppedReason:J("stopped_reason"),createdAt:jn("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[ao("idx_instruction_experiments_lookup").on(e.userId,e.projectPath)]),zn=Dn("instruction_experiment_arms",{id:we("id").primaryKey().defaultRandom(),experimentId:we("experiment_id").notNull().references(()=>Ue.id,{onDelete:"cascade"}),userId:we("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),externalSessionId:J("external_session_id").notNull(),arm:J("arm").notNull(),assignedAt:jn("assigned_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Ws("instruction_experiment_arms_unique").on(e.experimentId,e.externalSessionId),ao("idx_instruction_experiment_arms_experiment").on(e.experimentId,e.arm),ao("idx_instruction_experiment_arms_session").on(e.externalSessionId)]);import{pgTable as Xs,text as uo,timestamp as Zs,uuid as vt,numeric as lo,integer as Rn,jsonb as ei,index as ti,unique as oi}from"drizzle-orm/pg-core";var kn=Xs("instruction_verdicts",{id:vt("id").primaryKey().defaultRandom(),userId:vt("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),instructionId:vt("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),contentHash:uo("content_hash").notNull(),projectPath:uo("project_path").notNull(),segment:ei("segment").notNull().default({}),decision:uo("decision").notNull(),delta:lo("delta").notNull(),ciLower:lo("ci_lower").notNull(),ciUpper:lo("ci_upper").notNull(),nTreated:Rn("n_treated").notNull(),nControl:Rn("n_control").notNull(),experimentId:vt("experiment_id").references(()=>Ue.id,{onDelete:"set null"}),decidedAt:Zs("decided_at",{withTimezone:!0}).notNull().defaultNow()},e=>[oi("instruction_verdicts_unique").on(e.instructionId,e.contentHash,e.projectPath),ti("idx_instruction_verdicts_render").on(e.userId,e.projectPath)]);import{pgTable as ni,text as It,timestamp as co,uuid as po,integer as ri,unique as si,index as ii}from"drizzle-orm/pg-core";var En=ni("insight_recommendation_log",{id:po("id").primaryKey().defaultRandom(),userId:po("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),projectPath:It("project_path").notNull().default(""),category:It("category").notNull(),contentHash:It("content_hash").notNull(),label:It("label").notNull().default(""),firstSeenAt:co("first_seen_at",{withTimezone:!0}).notNull().defaultNow(),lastSeenAt:co("last_seen_at",{withTimezone:!0}).notNull().defaultNow(),timesGenerated:ri("times_generated").notNull().default(1),lastJobId:po("last_job_id"),dismissedAt:co("dismissed_at",{withTimezone:!0})},e=>[si("insight_recommendation_log_unique").on(e.userId,e.projectPath,e.contentHash),ii("idx_insight_recommendation_log_read").on(e.userId,e.projectPath)]);import{pgTable as ai,text as Te,timestamp as _o,uuid as Be,integer as Sn,index as yt}from"drizzle-orm/pg-core";var w=ai("memories",{id:Be("id").primaryKey().defaultRandom(),userId:Be("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),orgId:Be("org_id"),projectId:Be("project_id").references(()=>s.id,{onDelete:"set null"}),version:Sn("version").notNull().default(1),parentId:Be("parent_id").references(()=>w.id,{onDelete:"set null"}),kind:Te("kind").notNull(),title:Te("title").notNull(),body:Te("body").notNull(),source:Te("source").notNull(),status:Te("status").notNull().default("active"),trustClass:Te("trust_class"),trustVersion:Sn("trust_version"),trustComputedAt:_o("trust_computed_at",{withTimezone:!0}),createdAt:_o("created_at",{withTimezone:!0}).notNull().defaultNow(),updatedAt:_o("updated_at",{withTimezone:!0}).notNull().defaultNow()},e=>[yt("idx_memories_user").on(e.userId),yt("idx_memories_project").on(e.projectId),yt("idx_memories_parent").on(e.parentId),yt("idx_memories_user_status").on(e.userId,e.status)]);import{pgTable as di,text as ae,timestamp as mo,uuid as fo,integer as bn,jsonb as ui,index as Pn}from"drizzle-orm/pg-core";var Kn=di("memory_experiments",{id:fo("id").primaryKey().defaultRandom(),memoryId:fo("memory_id").notNull().references(()=>w.id,{onDelete:"cascade"}),userId:fo("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),memoryVersion:bn("memory_version").notNull().default(1),memoryBody:ae("memory_body"),taskPrompt:ae("task_prompt").notNull(),expectations:ae("expectations").array().notNull().default([]),trials:bn("trials").notNull().default(10),model:ae("model"),engine:ae("engine").$type().notNull().default("claude"),status:ae("status").notNull().default("pending"),result:ui("result").$type(),error:ae("error"),createdAt:mo("created_at",{withTimezone:!0}).notNull().defaultNow(),startedAt:mo("started_at",{withTimezone:!0}),completedAt:mo("completed_at",{withTimezone:!0})},e=>[Pn("idx_memory_experiments_memory").on(e.memoryId),Pn("idx_memory_experiments_user").on(e.userId)]);import{pgTable as li,text as Ae,timestamp as No,uuid as xo,jsonb as ci,uniqueIndex as pi,index as Cn}from"drizzle-orm/pg-core";var wt=li("session_terminal_targets",{id:xo("id").primaryKey().defaultRandom(),userId:xo("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),sessionId:xo("session_id").references(()=>a.id,{onDelete:"set null"}),externalSessionId:Ae("external_session_id"),kind:Ae("kind").notNull(),target:Ae("target").notNull(),status:Ae("status").notNull().default("active"),attachCommand:Ae("attach_command").notNull(),projectPath:Ae("project_path"),metadata:ci("metadata").notNull().default({}),lastSeenAt:No("last_seen_at",{withTimezone:!0}).defaultNow(),createdAt:No("created_at",{withTimezone:!0}).defaultNow(),updatedAt:No("updated_at",{withTimezone:!0}).defaultNow()},e=>[pi("idx_session_terminal_targets_user_target").on(e.userId,e.target),Cn("idx_session_terminal_targets_session").on(e.sessionId),Cn("idx_session_terminal_targets_user_status").on(e.userId,e.status)]);import{boolean as _i,index as Tt,jsonb as qn,pgTable as mi,text as de,timestamp as Ge,uuid as At,integer as ho}from"drizzle-orm/pg-core";var On=mi("session_send_jobs",{id:At("id").primaryKey().defaultRandom(),userId:At("user_id").notNull(),sessionId:At("session_id").notNull().references(()=>a.id,{onDelete:"cascade"}),terminalTargetId:At("terminal_target_id").references(()=>wt.id,{onDelete:"set null"}),target:de("target").notNull(),message:de("message").notNull(),submit:_i("submit").notNull().default(!0),delivery:de("delivery").notNull().default("live"),status:de("status").notNull().default("pending"),result:qn("result"),error:de("error"),attempts:ho("attempts").notNull().default(0),taskKind:de("task_kind"),payload:qn("payload"),claimedBy:de("claimed_by"),claimedAt:Ge("claimed_at",{withTimezone:!0}),leaseExpiresAt:Ge("lease_expires_at",{withTimezone:!0}),leaseRecoveries:ho("lease_recoveries").notNull().default(0),hostSuspendRecoveries:ho("host_suspend_recoveries").notNull().default(0),createdAt:Ge("created_at",{withTimezone:!0}).defaultNow(),updatedAt:Ge("updated_at",{withTimezone:!0}).defaultNow(),completedAt:Ge("completed_at",{withTimezone:!0})},e=>[Tt("idx_session_send_jobs_user").on(e.userId),Tt("idx_session_send_jobs_session").on(e.sessionId),Tt("idx_session_send_jobs_status").on(e.status),Tt("idx_session_send_jobs_pending").on(e.userId,e.status)]);import{pgTable as Hn,text as ue,timestamp as Dt,jsonb as Ln,uuid as Q,integer as fi,boolean as Ni,uniqueIndex as xi,index as Mn,primaryKey as hi}from"drizzle-orm/pg-core";var j=Hn("packs",{id:Q("id").primaryKey().defaultRandom(),name:ue("name").notNull(),description:ue("description").notNull(),scope:ue("scope").notNull(),scopeRef:Q("scope_ref"),version:ue("version").notNull().default("1.0.0"),tags:ue("tags").array(),badge:ue("badge"),metadata:Ln("metadata"),createdBy:Q("created_by"),createdAt:Dt("created_at",{withTimezone:!0}).defaultNow(),updatedAt:Dt("updated_at",{withTimezone:!0}).defaultNow(),archived:Ni("archived").default(!1),archivedAt:Dt("archived_at",{withTimezone:!0}),archivedBy:Q("archived_by"),forkedFromId:Q("forked_from_id").references(()=>j.id,{onDelete:"set null"}),upstreamVersion:ue("upstream_version")},e=>[xi("packs_name_scope_ref_idx").on(e.name,e.scope,e.scopeRef),Mn("idx_packs_scope").on(e.scope,e.scopeRef),Mn("idx_packs_created_by").on(e.createdBy)]),Y=Hn("pack_instructions",{packId:Q("pack_id").references(()=>j.id,{onDelete:"cascade"}).notNull(),instructionId:Q("instruction_id").references(()=>n.id,{onDelete:"cascade"}).notNull(),position:fi("position").notNull().default(0),slotOverrides:Ln("slot_overrides"),addedAt:Dt("added_at",{withTimezone:!0}).defaultNow(),addedBy:Q("added_by")},e=>[hi({columns:[e.packId,e.instructionId]})]);import{pgTable as gi,text as go,timestamp as vi,jsonb as Ii,uuid as vo,index as De}from"drizzle-orm/pg-core";var je=gi("instruction_usage_events",{id:vo("id").primaryKey().defaultRandom(),instructionId:vo("instruction_id").references(()=>n.id,{onDelete:"set null"}),userId:vo("user_id").notNull(),surface:go("surface").notNull(),action:go("action").notNull(),metadata:Ii("metadata"),traceId:go("trace_id"),createdAt:vi("created_at",{withTimezone:!0}).defaultNow()},e=>[De("idx_usage_events_instruction").on(e.instructionId),De("idx_usage_events_user").on(e.userId),De("idx_usage_events_surface").on(e.surface),De("idx_usage_events_action").on(e.action),De("idx_usage_events_created_at").on(e.createdAt),De("idx_usage_events_user_created").on(e.userId,e.createdAt)]);import{pgTable as yi,text as jt,timestamp as zt,jsonb as wi,uuid as Un,index as Io}from"drizzle-orm/pg-core";var Bn=yi("worker_connections",{id:Un("id").primaryKey().defaultRandom(),userId:Un("user_id").notNull(),workerType:jt("worker_type").notNull().default("mcp"),clientName:jt("client_name").notNull(),clientVersion:jt("client_version"),workerVersion:jt("worker_version"),lastHeartbeat:zt("last_heartbeat",{withTimezone:!0}).notNull().defaultNow(),presenceGapStartedAt:zt("presence_gap_started_at",{withTimezone:!0}),presenceGapEndedAt:zt("presence_gap_ended_at",{withTimezone:!0}),connectedAt:zt("connected_at",{withTimezone:!0}).notNull().defaultNow(),metadata:wi("metadata")},e=>[Io("idx_worker_connections_user").on(e.userId),Io("idx_worker_connections_heartbeat").on(e.userId,e.lastHeartbeat),Io("idx_worker_connections_type").on(e.userId,e.workerType)]);import{pgTable as Ti,text as Rt,timestamp as Ai,jsonb as Di,uuid as Gn,index as Vn}from"drizzle-orm/pg-core";var Fn=Ti("cli_setup_events",{id:Gn("id").primaryKey().defaultRandom(),userId:Gn("user_id").notNull(),deviceId:Rt("device_id").notNull(),event:Rt("event").notNull(),cliVersion:Rt("cli_version"),os:Rt("os"),props:Di("props").notNull().default({}),createdAt:Ai("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Vn("idx_cli_setup_events_user").on(e.userId,e.createdAt),Vn("idx_cli_setup_events_device").on(e.deviceId,e.createdAt)]);import{pgTable as ji,text as le,integer as zi,boolean as Ri,timestamp as ki,jsonb as Ei,uuid as Si,index as $n}from"drizzle-orm/pg-core";var Jn=ji("cli_error_events",{id:Si("id").primaryKey().defaultRandom(),deviceId:le("device_id"),event:le("event").notNull(),command:le("command"),statusCode:zi("status_code"),message:le("message"),cliVersion:le("cli_version"),os:le("os"),nodeVersion:le("node_version"),proxied:Ri("proxied"),props:Ei("props").notNull().default({}),createdAt:ki("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[$n("idx_cli_error_events_created").on(e.createdAt),$n("idx_cli_error_events_device").on(e.deviceId,e.createdAt)]);import{sql as yo}from"drizzle-orm";import{pgTable as ce,text as p,timestamp as g,jsonb as T,uuid as c,integer as O,boolean as Ve,index as f,unique as Qn}from"drizzle-orm/pg-core";var r=ce("eval_targets",{id:c("id").primaryKey().defaultRandom(),userId:c("user_id").notNull(),name:p("name").notNull(),type:p("type").notNull(),content:p("content").notNull(),description:p("description").notNull().default(""),metadata:T("metadata").notNull().default({}),version:O("version").notNull().default(1),parentId:c("parent_id").references(()=>r.id,{onDelete:"set null"}),instructionId:c("instruction_id").references(()=>n.id,{onDelete:"set null"}),projectPath:p("project_path"),createdAt:g("created_at",{withTimezone:!0}).defaultNow(),updatedAt:g("updated_at",{withTimezone:!0}).defaultNow()},e=>[f("eval_targets_name_user_idx").on(e.name,e.userId),f("idx_eval_targets_user").on(e.userId),f("idx_eval_targets_type").on(e.type),f("idx_eval_targets_parent").on(e.parentId),f("idx_eval_targets_instruction").on(e.instructionId).where(yo`${e.instructionId} IS NOT NULL`),f("idx_eval_targets_project").on(e.projectPath).where(yo`${e.projectPath} IS NOT NULL`)]),M=ce("evals",{id:c("id").primaryKey().defaultRandom(),evalTargetId:c("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),name:p("name").notNull(),prompt:p("prompt").notNull(),expectedOutput:p("expected_output").notNull().default(""),files:T("files").notNull().default([]),expectations:T("expectations").notNull().default([]),retiredAt:g("retired_at",{withTimezone:!0}),retiredReason:p("retired_reason"),createdAt:g("created_at",{withTimezone:!0}).defaultNow(),updatedAt:g("updated_at",{withTimezone:!0}).defaultNow(),contentUpdatedAt:g("content_updated_at",{withTimezone:!0}).notNull().defaultNow()},e=>[f("idx_evals_target").on(e.evalTargetId),f("idx_evals_target_active").on(e.evalTargetId).where(yo`${e.retiredAt} IS NULL`)]),E=ce("eval_runs",{id:c("id").primaryKey().defaultRandom(),jobId:c("job_id"),evalId:c("eval_id").notNull().references(()=>M.id,{onDelete:"cascade"}),evalTargetId:c("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),evalTargetVersion:O("eval_target_version").notNull(),configuration:p("configuration").notNull(),runNumber:O("run_number").notNull().default(1),iteration:O("iteration").notNull().default(1),status:p("status").notNull().default("completed"),outputs:T("outputs").notNull().default({}),timing:T("timing").notNull().default({}),grading:T("grading").notNull().default({}),comparison:T("comparison"),createdAt:g("created_at",{withTimezone:!0}).defaultNow()},e=>[f("idx_eval_runs_eval").on(e.evalId),f("idx_eval_runs_target").on(e.evalTargetId),f("idx_eval_runs_iteration").on(e.evalTargetId,e.iteration)]),Yn=ce("eval_comparisons",{id:c("id").primaryKey().defaultRandom(),userId:c("user_id").notNull(),targetAId:c("target_a_id").notNull().references(()=>r.id,{onDelete:"cascade"}),targetBId:c("target_b_id").notNull().references(()=>r.id,{onDelete:"cascade"}),targetAVersion:O("target_a_version").notNull(),targetBVersion:O("target_b_version").notNull(),isSwapped:Ve("is_swapped").notNull().default(!1),winnerTargetId:c("winner_target_id").references(()=>r.id,{onDelete:"set null"}),perEval:T("per_eval").notNull().default([]),overall:T("overall").notNull().default({}),rawOutput:p("raw_output"),model:p("model"),status:p("status").notNull().default("completed"),createdAt:g("created_at",{withTimezone:!0}).defaultNow()},e=>[f("idx_eval_comparisons_target_a").on(e.targetAId),f("idx_eval_comparisons_target_b").on(e.targetBId),f("idx_eval_comparisons_user").on(e.userId)]),Wn=ce("eval_grade_annotations",{id:c("id").primaryKey().defaultRandom(),userId:c("user_id").notNull(),evalRunId:c("eval_run_id").notNull().references(()=>E.id,{onDelete:"cascade"}),evalTargetId:c("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),expectationIndex:O("expectation_index").notNull(),expectationText:p("expectation_text").notNull(),judgePassed:Ve("judge_passed").notNull(),judgeMethod:p("judge_method").notNull(),judgeModel:p("judge_model"),verdict:p("verdict").notNull(),flag:p("flag"),note:p("note"),createdAt:g("created_at",{withTimezone:!0}).defaultNow(),updatedAt:g("updated_at",{withTimezone:!0}).defaultNow()},e=>[f("idx_eval_grade_annotations_target").on(e.evalTargetId),f("idx_eval_grade_annotations_run").on(e.evalRunId),Qn("eval_grade_annotations_eval_run_id_expectation_index_key").on(e.evalRunId,e.expectationIndex)]),Xn=ce("judge_reviewer_verdicts",{id:c("id").primaryKey().defaultRandom(),userId:c("user_id").notNull(),evalTargetId:c("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),evalRunId:c("eval_run_id").notNull().references(()=>E.id,{onDelete:"cascade"}),expectationIndex:O("expectation_index").notNull(),expectationText:p("expectation_text").notNull(),judgePassed:Ve("judge_passed").notNull(),humanPassed:Ve("human_passed").notNull(),reviewerEngine:p("reviewer_engine").notNull(),reviewerModel:p("reviewer_model").notNull(),reviewerPassed:Ve("reviewer_passed").notNull(),resolvedAt:g("resolved_at",{withTimezone:!0}),resolution:p("resolution"),createdAt:g("created_at",{withTimezone:!0}).defaultNow(),updatedAt:g("updated_at",{withTimezone:!0}).defaultNow()},e=>[f("idx_jrv_target").on(e.evalTargetId),Qn("judge_reviewer_verdicts_eval_run_id_expectation_index_review_key").on(e.evalRunId,e.expectationIndex,e.reviewerModel)]),ze=ce("benchmarks",{id:c("id").primaryKey().defaultRandom(),evalTargetId:c("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),iteration:O("iteration").notNull().default(1),runSummary:T("run_summary").notNull().default({}),notes:T("notes").notNull().default([]),feedback:T("feedback"),createdAt:g("created_at",{withTimezone:!0}).defaultNow()},e=>[f("idx_benchmarks_target").on(e.evalTargetId)]);import{sql as wo}from"drizzle-orm";import{pgTable as bi,text as H,timestamp as Pi,jsonb as kt,uuid as Et,boolean as Ki,index as Fe,uniqueIndex as Ci}from"drizzle-orm/pg-core";var Zn=bi("instruction_promotions",{id:Et("id").primaryKey().defaultRandom(),userId:Et("user_id").notNull(),evalTargetId:Et("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),instructionId:Et("instruction_id").references(()=>n.id,{onDelete:"set null"}),fromContentHash:H("from_content_hash").notNull(),toContentHash:H("to_content_hash").notNull(),fromContent:H("from_content").notNull(),toContent:H("to_content").notNull(),proposalSource:H("proposal_source").notNull(),proposalRationale:H("proposal_rationale").notNull().default(""),decision:H("decision").notNull(),applied:Ki("applied").notNull().default(!1),refusals:kt("refusals").notNull().default([]),rationale:H("rationale").notNull().default(""),effectClass:H("effect_class").notNull(),verdict:kt("verdict").notNull().default({}),gates:kt("gates").notNull().default({}),touchedSections:kt("touched_sections").notNull().default([]),createdAt:Pi("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Fe("idx_instruction_promotions_budget").on(e.evalTargetId,e.createdAt.desc()).where(wo`${e.applied}`),Fe("idx_instruction_promotions_target").on(e.evalTargetId,e.createdAt.desc()),Fe("idx_instruction_promotions_user").on(e.userId,e.createdAt.desc()),Fe("idx_instruction_promotions_to_hash").on(e.toContentHash),Ci("uq_instruction_promotions_applied").on(e.evalTargetId,e.fromContentHash,e.toContentHash).where(wo`${e.applied}`),Fe("idx_instruction_promotions_applied_hash").on(e.toContentHash,e.createdAt.desc()).where(wo`${e.applied}`)]);import{index as $e,integer as qi,jsonb as Oi,pgTable as Mi,text as y,timestamp as Re,uniqueIndex as Hi,uuid as Je}from"drizzle-orm/pg-core";var er=Mi("recommendation_applications",{id:Je("id").primaryKey().defaultRandom(),userId:Je("user_id").notNull(),instructionId:Je("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),projectId:Je("project_id").references(()=>s.id,{onDelete:"set null"}),projectPath:y("project_path").notNull(),harness:y("harness").notNull(),targetPath:y("target_path").notNull(),status:y("status").notNull().default("proposed"),applyMode:y("apply_mode").notNull(),beforeHash:y("before_hash"),appliedHash:y("applied_hash"),appliedAt:Re("applied_at",{withTimezone:!0}),measurementStartedAt:Re("measurement_started_at",{withTimezone:!0}),measurementEndedAt:Re("measurement_ended_at",{withTimezone:!0}),verdict:y("verdict"),verdictDetail:y("verdict_detail"),revertedAt:Re("reverted_at",{withTimezone:!0}),boundaryClass:y("boundary_class"),targets:y("targets").array().notNull().default([]),inverse:Oi("inverse"),requires:y("requires").array().notNull().default([]),provides:y("provides").array().notNull().default([]),applyBatchId:Je("apply_batch_id"),applySeq:qi("apply_seq"),createdAt:Re("created_at",{withTimezone:!0}).notNull().defaultNow(),updatedAt:Re("updated_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Hi("uq_recommendation_applications_target").on(e.userId,e.instructionId,e.projectPath,e.harness,e.targetPath),$e("idx_recommendation_applications_user_status").on(e.userId,e.status,e.updatedAt.desc()),$e("idx_recommendation_applications_user_harness").on(e.userId,e.harness,e.updatedAt.desc()),$e("idx_recommendation_applications_project").on(e.projectId,e.updatedAt.desc()),$e("idx_recommendation_applications_batch").on(e.userId,e.applyBatchId,e.applySeq.desc()),$e("idx_recommendation_applications_requires").using("gin",e.requires)]);import{pgTable as Li,text as S,timestamp as tr,jsonb as Qe,uuid as To,integer as ke,real as Ui,index as W,uniqueIndex as or}from"drizzle-orm/pg-core";import{sql as Ao}from"drizzle-orm";var X=Li("context_snapshots",{id:To("id").primaryKey().defaultRandom(),sessionId:To("session_id").references(()=>a.id,{onDelete:"cascade"}),userId:To("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),sequence:ke("sequence").notNull().default(0),provider:S("provider").notNull(),model:S("model").notNull(),systemPrompt:S("system_prompt").notNull(),injectedInstructions:Qe("injected_instructions").notNull().default([]),messagesSummary:Qe("messages_summary").notNull(),tools:Qe("tools").notNull().default([]),systemHash:S("system_hash").notNull(),requestHash:S("request_hash").notNull().default(""),toolsHash:S("tools_hash"),injectedHash:S("injected_hash"),messagesDelta:Qe("messages_delta").default([]),messagesDeltaRef:S("messages_delta_ref"),messagesBaseLength:ke("messages_base_length"),payloadPrunedAt:tr("payload_pruned_at",{withTimezone:!0}),deltaOffloadState:S("delta_offload_state").notNull().default("inline"),deltaGzipBytes:ke("delta_gzip_bytes"),deltaOffloadError:S("delta_offload_error"),deltaOffloadAttempts:ke("delta_offload_attempts").notNull().default(0),maxTokens:ke("max_tokens"),temperature:Ui("temperature"),estimatedTokens:ke("estimated_tokens"),capturedAt:tr("captured_at",{withTimezone:!0}).notNull().defaultNow(),metadata:Qe("metadata").default({})},e=>[W("idx_context_snapshots_session").on(e.sessionId),W("idx_context_snapshots_user").on(e.userId),W("idx_context_snapshots_hash").on(e.systemHash),W("idx_context_snapshots_request_hash").on(e.requestHash),or("idx_snapshots_session_request").on(e.sessionId,e.requestHash),or("idx_snapshots_session_sequence").on(e.sessionId,e.sequence),W("idx_context_snapshots_user_session_captured").on(e.userId,e.sessionId,e.capturedAt.desc()),W("idx_snapshots_unpruned_captured").on(e.capturedAt,e.id).where(Ao`${e.payloadPrunedAt} is null`),W("idx_snapshots_pending_offload").on(e.deltaOffloadState,e.capturedAt,e.id).where(Ao`${e.deltaOffloadState} = 'pending'`),W("idx_snapshots_delta_ref").on(e.messagesDeltaRef).where(Ao`${e.messagesDeltaRef} is not null`)]);import{pgTable as Bi,text as nr,timestamp as rr,jsonb as Gi,integer as Vi,index as Fi}from"drizzle-orm/pg-core";var sr=Bi("content_blobs",{hash:nr("hash").primaryKey(),contentType:nr("content_type").notNull(),content:Gi("content").notNull(),sizeBytes:Vi("size_bytes").notNull().default(0),createdAt:rr("created_at",{withTimezone:!0}).defaultNow(),lastReferencedAt:rr("last_referenced_at",{withTimezone:!0}).notNull().defaultNow()},e=>[Fi("idx_content_blobs_type").on(e.contentType)]);import{integer as ir,jsonb as $i,pgTable as Ji,timestamp as Qi,uuid as Yi}from"drizzle-orm/pg-core";var ar=Ji("session_capture_heads",{sessionId:Yi("session_id").primaryKey().references(()=>a.id,{onDelete:"cascade"}),sequence:ir("sequence").notNull(),rawMessageHashes:$i("raw_message_hashes").notNull().$type(),version:ir("version").notNull(),updatedAt:Qi("updated_at",{withTimezone:!0}).notNull().defaultNow()});import{index as Wi,integer as dr,pgTable as Xi,text as Do,timestamp as ur,uniqueIndex as Zi,uuid as ea}from"drizzle-orm/pg-core";import{sql as ta}from"drizzle-orm";var lr=Xi("snapshot_delta_objects",{ref:Do("ref").primaryKey(),snapshotId:ea("snapshot_id"),compressedBytes:dr("compressed_bytes").notNull(),state:Do("state").notNull(),createdAt:ur("created_at",{withTimezone:!0}).notNull().defaultNow(),lastAttemptAt:ur("last_attempt_at",{withTimezone:!0}),attempts:dr("attempts").notNull().default(0),lastError:Do("last_error")},e=>[Wi("idx_snapshot_delta_objects_state_created").on(e.state,e.createdAt),Zi("snapshot_delta_objects_snapshot_id_unique").on(e.snapshotId).where(ta`${e.snapshotId} is not null`)]);import{pgTable as oa,text as St,timestamp as Ye,jsonb as bt,uuid as jo,index as Pt,integer as zo}from"drizzle-orm/pg-core";var b=oa("eval_jobs",{id:jo("id").primaryKey().defaultRandom(),userId:jo("user_id").notNull(),evalTargetId:jo("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),status:St("status").notNull().default("pending"),config:bt("config").notNull().default({}),progress:bt("progress").notNull().default({}),result:bt("result"),error:St("error"),attempts:zo("attempts").notNull().default(0),taskKind:St("task_kind"),payload:bt("payload"),claimedBy:St("claimed_by"),claimedAt:Ye("claimed_at",{withTimezone:!0}),leaseExpiresAt:Ye("lease_expires_at",{withTimezone:!0}),leaseRecoveries:zo("lease_recoveries").notNull().default(0),hostSuspendRecoveries:zo("host_suspend_recoveries").notNull().default(0),createdAt:Ye("created_at",{withTimezone:!0}).defaultNow(),updatedAt:Ye("updated_at",{withTimezone:!0}).defaultNow(),completedAt:Ye("completed_at",{withTimezone:!0})},e=>[Pt("idx_eval_jobs_user").on(e.userId),Pt("idx_eval_jobs_target").on(e.evalTargetId),Pt("idx_eval_jobs_status").on(e.status),Pt("idx_eval_jobs_pending").on(e.userId,e.status)]);import{pgTable as na,text as We,timestamp as Xe,jsonb as cr,uuid as pr,integer as Ro,index as _r}from"drizzle-orm/pg-core";var mr=na("agent_jobs",{id:pr("id").primaryKey().defaultRandom(),userId:pr("user_id").notNull(),taskKind:We("task_kind").notNull(),status:We("status").notNull().default("pending"),payload:cr("payload").notNull(),dedupeKey:We("dedupe_key"),result:cr("result"),error:We("error"),attempts:Ro("attempts").notNull().default(0),claimedBy:We("claimed_by"),claimedAt:Xe("claimed_at",{withTimezone:!0}),leaseExpiresAt:Xe("lease_expires_at",{withTimezone:!0}),leaseRecoveries:Ro("lease_recoveries").notNull().default(0),hostSuspendRecoveries:Ro("host_suspend_recoveries").notNull().default(0),createdAt:Xe("created_at",{withTimezone:!0}).notNull().defaultNow(),updatedAt:Xe("updated_at",{withTimezone:!0}).notNull().defaultNow(),completedAt:Xe("completed_at",{withTimezone:!0})},e=>[_r("idx_agent_jobs_claim").on(e.userId,e.status,e.taskKind,e.createdAt),_r("idx_agent_jobs_dedupe").on(e.userId,e.taskKind,e.dedupeKey)]);import{pgTable as ra,text as Ze,timestamp as et,jsonb as Kt,uuid as ko,index as Ct,integer as Eo}from"drizzle-orm/pg-core";var fr=ra("insights_jobs",{id:ko("id").primaryKey().defaultRandom(),userId:ko("user_id").notNull(),status:Ze("status").notNull().default("pending"),projectPath:Ze("project_path"),projectId:ko("project_id").references(()=>s.id,{onDelete:"set null"}),config:Kt("config").notNull().default({}),progress:Kt("progress").notNull().default({}),result:Kt("result"),error:Ze("error"),attempts:Eo("attempts").notNull().default(0),taskKind:Ze("task_kind"),payload:Kt("payload"),claimedBy:Ze("claimed_by"),claimedAt:et("claimed_at",{withTimezone:!0}),leaseExpiresAt:et("lease_expires_at",{withTimezone:!0}),leaseRecoveries:Eo("lease_recoveries").notNull().default(0),hostSuspendRecoveries:Eo("host_suspend_recoveries").notNull().default(0),createdAt:et("created_at",{withTimezone:!0}).defaultNow(),updatedAt:et("updated_at",{withTimezone:!0}).defaultNow(),completedAt:et("completed_at",{withTimezone:!0})},e=>[Ct("idx_insights_jobs_user").on(e.userId),Ct("idx_insights_jobs_status").on(e.status),Ct("idx_insights_jobs_pending").on(e.userId,e.status),Ct("idx_insights_jobs_project").on(e.projectId)]);import{pgTable as sa,text as So,timestamp as qt,uuid as Nr,index as ia,uniqueIndex as aa}from"drizzle-orm/pg-core";var Ee=sa("api_keys",{id:Nr("id").primaryKey().defaultRandom(),userId:Nr("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),name:So("name").notNull(),keyHash:So("key_hash").notNull(),keyPrefix:So("key_prefix").notNull(),lastUsedAt:qt("last_used_at",{withTimezone:!0}),expiresAt:qt("expires_at",{withTimezone:!0}),revokedAt:qt("revoked_at",{withTimezone:!0}),createdAt:qt("created_at",{withTimezone:!0}).defaultNow()},e=>[ia("idx_api_keys_user").on(e.userId),aa("idx_api_keys_hash").on(e.keyHash)]);import{pgTable as da,text as bo,timestamp as xr,uuid as ua,index as la}from"drizzle-orm/pg-core";var hr=da("extension_auth_codes",{code:bo("code").primaryKey(),userId:ua("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),editor:bo("editor"),email:bo("email"),expiresAt:xr("expires_at",{withTimezone:!0}).notNull(),createdAt:xr("created_at",{withTimezone:!0}).defaultNow()},e=>[la("idx_ext_auth_codes_expiry").on(e.expiresAt)]);import{pgTable as ca,text as pa,timestamp as gr,jsonb as _a,uuid as Po,index as vr}from"drizzle-orm/pg-core";var Se=ca("eval_conversations",{id:Po("id").primaryKey().defaultRandom(),evalTargetId:Po("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),userId:Po("user_id").notNull(),messages:_a("messages").notNull().default([]),status:pa("status").notNull().default("active"),createdAt:gr("created_at",{withTimezone:!0}).defaultNow(),updatedAt:gr("updated_at",{withTimezone:!0}).defaultNow()},e=>[vr("idx_eval_conversations_target").on(e.evalTargetId),vr("idx_eval_conversations_user").on(e.userId)]);import{pgTable as ma,text as L,timestamp as Ir,jsonb as fa,uuid as Ko}from"drizzle-orm/pg-core";var yr=ma("skill_submissions",{id:Ko("id").primaryKey().defaultRandom(),skillId:L("skill_id").notNull(),name:L("name").notNull(),description:L("description").notNull(),category:L("category").notNull(),tools:L("tools").array().notNull(),tags:L("tags").array().default([]),content:L("content").notNull(),status:L("status").notNull().default("pending"),submittedBy:Ko("submitted_by").notNull(),reviewedBy:Ko("reviewed_by"),reviewNote:L("review_note"),metadata:fa("metadata"),createdAt:Ir("created_at",{withTimezone:!0}).defaultNow(),updatedAt:Ir("updated_at",{withTimezone:!0}).defaultNow()});import{pgTable as Na,text as tt,timestamp as wr,jsonb as xa,uuid as ha,uniqueIndex as ga,index as Tr}from"drizzle-orm/pg-core";var Ar=Na("harness_releases",{id:ha("id").primaryKey().defaultRandom(),harness:tt("harness").notNull(),version:tt("version").notNull(),publishedAt:wr("published_at",{withTimezone:!0}).notNull(),title:tt("title"),body:tt("body"),categories:xa("categories").default({}),url:tt("url"),createdAt:wr("created_at",{withTimezone:!0}).defaultNow()},e=>[ga("idx_harness_releases_unique").on(e.harness,e.version),Tr("idx_harness_releases_harness").on(e.harness),Tr("idx_harness_releases_published").on(e.publishedAt)]);import{pgTable as va,text as be,timestamp as Dr,jsonb as Ia,uuid as jr,index as zr,uniqueIndex as ya}from"drizzle-orm/pg-core";var Co=va("notifications",{id:jr("id").primaryKey().defaultRandom(),userId:jr("user_id").notNull(),type:be("type").notNull(),title:be("title").notNull(),body:be("body"),url:be("url"),sourceTable:be("source_table"),sourceId:be("source_id"),metadata:Ia("metadata").notNull().default({}),readAt:Dr("read_at",{withTimezone:!0}),createdAt:Dr("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[zr("idx_notifications_user_created").on(e.userId,e.createdAt),zr("idx_notifications_user_unread").on(e.userId),ya("idx_notifications_source_dedup").on(e.userId,e.sourceTable,e.sourceId)]);import{pgTable as wa,text as ot,timestamp as nt,jsonb as qo,uuid as Oo,index as Ot,integer as Mo}from"drizzle-orm/pg-core";var Pe=wa("launch_jobs",{id:Oo("id").primaryKey().defaultRandom(),userId:Oo("user_id").notNull(),tool:ot("tool").notNull(),status:ot("status").notNull().default("pending"),config:qo("config").notNull().default({}),result:qo("result"),error:ot("error"),attempts:Mo("attempts").notNull().default(0),taskKind:ot("task_kind"),payload:qo("payload"),claimedBy:ot("claimed_by"),claimedAt:nt("claimed_at",{withTimezone:!0}),leaseExpiresAt:nt("lease_expires_at",{withTimezone:!0}),leaseRecoveries:Mo("lease_recoveries").notNull().default(0),hostSuspendRecoveries:Mo("host_suspend_recoveries").notNull().default(0),createdAt:nt("created_at",{withTimezone:!0}).defaultNow(),updatedAt:nt("updated_at",{withTimezone:!0}).defaultNow(),completedAt:nt("completed_at",{withTimezone:!0}),runId:Oo("run_id").references(()=>h.id,{onDelete:"set null"})},e=>[Ot("idx_launch_jobs_user").on(e.userId),Ot("idx_launch_jobs_status").on(e.status),Ot("idx_launch_jobs_pending").on(e.userId,e.status),Ot("idx_launch_jobs_run").on(e.runId)]);import{pgTable as Rr,uuid as U,text as Z,real as Ta,timestamp as rt,boolean as Aa,uniqueIndex as Da,index as st}from"drizzle-orm/pg-core";var kr=["add","remove","trim"],Ho=Rr("instruction_suggestions",{id:U("id").primaryKey().defaultRandom(),userId:U("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),instructionId:U("instruction_id").notNull().references(()=>n.id,{onDelete:"cascade"}),projectPath:Z("project_path").notNull(),projectId:U("project_id").references(()=>s.id,{onDelete:"set null"}),score:Ta("score").notNull(),reason:Z("reason").notNull(),kind:Z("kind").$type().notNull().default("add"),generatedAt:rt("generated_at",{withTimezone:!0}).notNull().defaultNow(),isStale:Aa("is_stale").notNull().default(!1),dismissedAt:rt("dismissed_at",{withTimezone:!0}),acceptedAt:rt("accepted_at",{withTimezone:!0}),declinedAt:rt("declined_at",{withTimezone:!0}),declinedReason:Z("declined_reason"),promotedToInstructionId:U("promoted_to_instruction_id").references(()=>n.id,{onDelete:"set null"})},e=>[Da("idx_instruction_suggestions_user_project_instruction_kind").on(e.userId,e.projectPath,e.instructionId,e.kind),st("idx_instruction_suggestions_user_project").on(e.userId,e.projectPath),st("idx_instruction_suggestions_user_proj_id").on(e.userId,e.projectId),st("idx_instruction_suggestions_generated_at").on(e.generatedAt)]),Er=Rr("instruction_suggestion_events",{id:U("id").primaryKey().defaultRandom(),suggestionId:U("suggestion_id").notNull().references(()=>Ho.id,{onDelete:"cascade"}),userId:U("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),eventType:Z("event_type").notNull(),tool:Z("tool"),sessionId:U("session_id"),projectPath:Z("project_path").notNull(),actor:Z("actor"),createdAt:rt("created_at",{withTimezone:!0}).notNull().defaultNow()},e=>[st("idx_instruction_suggestion_events_suggestion_id").on(e.suggestionId),st("idx_instruction_suggestion_events_user_id").on(e.userId)]);import{pgTable as ja,text as Lo,integer as za,numeric as it,timestamp as Uo,index as Ra,primaryKey as ka}from"drizzle-orm/pg-core";var Sr=ja("model_pricing",{provider:Lo("provider").notNull(),modelId:Lo("model_id").notNull(),inputPerM:it("input_per_m").notNull(),outputPerM:it("output_per_m").notNull(),cacheReadPerM:it("cache_read_per_m").notNull(),cacheWrite5mPerM:it("cache_write_5m_per_m").notNull(),cacheWrite1hPerM:it("cache_write_1h_per_m").notNull(),contextWindow:za("context_window"),effectiveFrom:Uo("effective_from",{withTimezone:!0}).notNull(),effectiveUntil:Uo("effective_until",{withTimezone:!0}),source:Lo("source").notNull(),verifiedAt:Uo("verified_at",{withTimezone:!0})},e=>[ka({columns:[e.provider,e.modelId,e.effectiveFrom]}),Ra("idx_model_pricing_active").on(e.provider,e.modelId,e.effectiveFrom.desc())]);import{pgTable as Ea,text as pe,numeric as br,timestamp as Pr,uuid as Kr,jsonb as Sa,index as Cr}from"drizzle-orm/pg-core";var qr=Ea("limit_observations",{id:Kr("id").primaryKey().defaultRandom(),observedAt:Pr("observed_at",{withTimezone:!0}).notNull().defaultNow(),provider:pe("provider").notNull(),plan:pe("plan").notNull(),limitType:pe("limit_type").notNull(),value:br("value").notNull(),unit:pe("unit").notNull(),comparator:pe("comparator").notNull().default("="),resetsAt:Pr("resets_at",{withTimezone:!0}),sourceType:pe("source_type").notNull(),sourceUrl:pe("source_url"),sourceEvidence:Sa("source_evidence"),confidence:br("confidence").notNull(),userId:Kr("user_id")},e=>[Cr("idx_limit_obs_lookup").on(e.provider,e.plan,e.limitType,e.observedAt.desc()),Cr("idx_limit_obs_user").on(e.userId,e.observedAt.desc())]);import{pgTable as ba,text as Mt,timestamp as Pa,uuid as Ka,jsonb as Ca,boolean as qa,uniqueIndex as Oa}from"drizzle-orm/pg-core";var Or=ba("provider_docs_snapshots",{id:Ka("id").primaryKey().defaultRandom(),fetchedAt:Pa("fetched_at",{withTimezone:!0}).notNull().defaultNow(),provider:Mt("provider").notNull(),docUrl:Mt("doc_url").notNull(),contentHash:Mt("content_hash").notNull(),content:Mt("content").notNull(),parseSuccess:qa("parse_success").notNull(),parseErrors:Ca("parse_errors")},e=>[Oa("idx_docs_dedup").on(e.provider,e.docUrl,e.contentHash)]);import{pgTable as Ma,text as Ht,numeric as Mr,timestamp as Hr,uuid as Ha,primaryKey as La,index as Ua}from"drizzle-orm/pg-core";var Lr=Ma("plan_limits_active",{provider:Ht("provider").notNull(),plan:Ht("plan").notNull(),limitType:Ht("limit_type").notNull(),value:Mr("value").notNull(),unit:Ht("unit").notNull(),effectiveFrom:Hr("effective_from",{withTimezone:!0}).notNull(),effectiveUntil:Hr("effective_until",{withTimezone:!0}),confidence:Mr("confidence").notNull(),sourceObservationIds:Ha("source_observation_ids").array().notNull()},e=>[La({columns:[e.provider,e.plan,e.limitType,e.effectiveFrom]}),Ua("idx_plan_limits_lookup").on(e.provider,e.plan,e.limitType,e.effectiveFrom.desc())]);import{pgTable as Ba,text as Bo,timestamp as Ga,uuid as Ur,jsonb as Va,index as Fa}from"drizzle-orm/pg-core";var Br=Ba("alert_dispatches",{id:Ur("id").primaryKey().defaultRandom(),userId:Ur("user_id").notNull(),alertKey:Bo("alert_key").notNull(),severity:Bo("severity").notNull(),lastDispatchedAt:Ga("last_dispatched_at",{withTimezone:!0}).notNull().defaultNow(),lastState:Va("last_state").notNull(),channel:Bo("channel").notNull()},e=>[Fa("idx_alert_user_key").on(e.userId,e.alertKey,e.lastDispatchedAt.desc())]);import{date as Gr,integer as Go,pgTable as Vr,primaryKey as $a,text as Ja,timestamp as Vo,uuid as Qa}from"drizzle-orm/pg-core";var Fr=Vr("weekly_digest_deliveries",{userId:Qa("user_id").notNull().references(()=>o.id,{onDelete:"cascade"}),weekStart:Gr("week_start").notNull(),sentAt:Vo("sent_at",{withTimezone:!0}).notNull().defaultNow(),providerMessageId:Ja("provider_message_id")},e=>[$a({columns:[e.userId,e.weekStart]})]),$r=Vr("weekly_digest_runs",{weekStart:Gr("week_start").primaryKey(),startedAt:Vo("started_at",{withTimezone:!0}).notNull().defaultNow(),finishedAt:Vo("finished_at",{withTimezone:!0}),sent:Go("sent"),skipped:Go("skipped"),failed:Go("failed")});import{pgTable as Ya,text as Fo,timestamp as Jr,uuid as $o,integer as Wa,index as Xa,unique as Za}from"drizzle-orm/pg-core";var Qr=Ya("scaffolding_fingerprints",{id:$o("id").primaryKey().defaultRandom(),userId:$o("user_id").notNull(),harness:Fo("harness").notNull(),model:Fo("model").notNull(),toolsHash:Fo("tools_hash").notNull(),firstSeenAt:Jr("first_seen_at",{withTimezone:!0}).notNull().defaultNow(),lastSeenAt:Jr("last_seen_at",{withTimezone:!0}).notNull().defaultNow(),snapshotCount:Wa("snapshot_count").notNull().default(1),firstSessionId:$o("first_session_id")},e=>[Za("scaffolding_fingerprints_user_id_harness_model_tools_hash_key").on(e.userId,e.harness,e.model,e.toolsHash),Xa("idx_scaffolding_fp_segment").on(e.userId,e.harness,e.model)]);import{pgTable as ed,text as Jo,timestamp as td,jsonb as od,uuid as Yr,integer as nd,unique as rd}from"drizzle-orm/pg-core";var Wr=ed("scaffolding_segments",{id:Yr("id").primaryKey().defaultRandom(),userId:Yr("user_id").notNull(),harness:Jo("harness").notNull(),model:Jo("model").notNull(),coreNames:od("core_names").notNull(),coreHash:Jo("core_hash").notNull(),activeFingerprints:nd("active_fingerprints").notNull(),computedAt:td("computed_at",{withTimezone:!0}).notNull().defaultNow()},e=>[rd("scaffolding_segments_user_id_harness_model_key").on(e.userId,e.harness,e.model)]);import{pgTable as sd,text as ee,timestamp as Xr,uuid as Qo,uniqueIndex as id,index as ad}from"drizzle-orm/pg-core";var Zr=["blog","docs","changelog"],Yo=sd("vendor_guidance_snapshots",{id:Qo("id").primaryKey().defaultRandom(),fetchedAt:Xr("fetched_at",{withTimezone:!0}).notNull().defaultNow(),sourceId:ee("source_id").notNull(),harness:ee("harness").notNull(),vendor:ee("vendor").notNull(),kind:ee("kind").$type().notNull(),url:ee("url").notNull(),title:ee("title").notNull(),contentHash:ee("content_hash").notNull(),text:ee("text").notNull(),previousSnapshotId:Qo("previous_snapshot_id").references(()=>Yo.id,{onDelete:"set null"}),pitchedAt:Xr("pitched_at",{withTimezone:!0}),pitchInstructionId:Qo("pitch_instruction_id").references(()=>n.id,{onDelete:"set null"})},e=>[id("idx_vendor_guidance_dedup").on(e.sourceId,e.contentHash),ad("idx_vendor_guidance_source_fetched").on(e.sourceId,e.fetchedAt)]);import{pgTable as ts,uuid as Ke,text as Lt,jsonb as es,timestamp as Wo,uniqueIndex as dd,integer as ud,unique as ld}from"drizzle-orm/pg-core";var te=ts("performance_studies",{id:Ke("id").primaryKey().defaultRandom(),userId:Ke("user_id").notNull(),evalTargetId:Ke("eval_target_id").notNull().references(()=>r.id,{onDelete:"cascade"}),projectPath:Lt("project_path").notNull(),protocol:es("protocol").notNull(),protocolHash:Lt("protocol_hash").notNull(),result:es("result"),finalizedAt:Wo("finalized_at",{withTimezone:!0}),createdAt:Wo("created_at",{withTimezone:!0}).defaultNow().notNull()},e=>[dd("performance_studies_protocol_unique").on(e.userId,e.evalTargetId,e.protocolHash)]),_e=ts("performance_study_calls",{id:Ke("id").primaryKey().defaultRandom(),studyId:Ke("study_id").notNull().references(()=>te.id,{onDelete:"cascade"}),jobId:Ke("job_id").notNull().references(()=>b.id,{onDelete:"cascade"}),taskId:Lt("task_id").notNull(),arm:Lt("arm").notNull(),repetition:ud("repetition").notNull(),reservedAt:Wo("reserved_at",{withTimezone:!0}).defaultNow().notNull()},e=>[ld("performance_study_calls_study_id_task_id_arm_repetition_key").on(e.studyId,e.taskId,e.arm,e.repetition)]);import{drizzle as is}from"drizzle-orm/postgres-js";import as from"postgres";var Ut={};Ds(Ut,{GUIDANCE_KINDS:()=>Zr,SUGGESTION_KINDS:()=>kr,agentJobs:()=>mr,agentRuns:()=>h,agentRunsRelations:()=>Ad,alertDispatches:()=>Br,announcementDeliveries:()=>dn,announcements:()=>an,apiKeys:()=>Ee,apiKeysRelations:()=>pd,benchmarks:()=>ze,benchmarksRelations:()=>qd,cliErrorEvents:()=>Jn,cliSetupEvents:()=>Fn,contentBlobs:()=>sr,contextSnapshots:()=>X,contextSnapshotsRelations:()=>Hd,evalComparisons:()=>Yn,evalConversations:()=>Se,evalConversationsRelations:()=>Md,evalGradeAnnotations:()=>Wn,evalJobs:()=>b,evalJobsRelations:()=>Od,evalRuns:()=>E,evalRunsRelations:()=>Cd,evalTargets:()=>r,evalTargetsRelations:()=>Pd,evals:()=>M,evalsRelations:()=>Kd,extensionAuthCodes:()=>hr,harnessReleases:()=>Ar,insightRecommendationLog:()=>En,insightsJobs:()=>fr,instructionEvents:()=>ge,instructionEventsRelations:()=>gd,instructionExperimentArms:()=>zn,instructionExperiments:()=>Ue,instructionPromotions:()=>Zn,instructionSuggestionEvents:()=>Er,instructionSuggestions:()=>Ho,instructionUsageEvents:()=>je,instructionUsageEventsRelations:()=>bd,instructionVerdicts:()=>kn,instructionVersionIdentities:()=>fn,instructionVersions:()=>he,instructionVersionsRelations:()=>Id,instructions:()=>n,instructionsRelations:()=>hd,judgeReviewerVerdicts:()=>Xn,launchJobs:()=>Pe,launchJobsRelations:()=>Dd,limitObservations:()=>qr,memories:()=>w,memoriesRelations:()=>vd,memoryExperiments:()=>Kn,modelPricing:()=>Sr,notificationPreferences:()=>un,notifications:()=>Co,orgInvitations:()=>P,orgInvitationsRelations:()=>fd,organizations:()=>l,organizationsRelations:()=>_d,packInstructions:()=>Y,packInstructionsRelations:()=>Sd,packs:()=>j,packsRelations:()=>Ed,pendingTreatmentExposures:()=>An,performanceStudies:()=>te,performanceStudiesRelations:()=>Ld,performanceStudyCalls:()=>_e,performanceStudyCallsRelations:()=>Ud,planLimitsActive:()=>Lr,profiles:()=>o,profilesRelations:()=>cd,projectCheckouts:()=>hn,projectInstructions:()=>F,projectInstructionsRelations:()=>wd,projects:()=>s,projectsRelations:()=>yd,providerDocsSnapshots:()=>Or,recommendationApplications:()=>er,scaffoldingFingerprints:()=>Qr,scaffoldingSegments:()=>Wr,sessionCaptureHeads:()=>ar,sessionEvents:()=>Ie,sessionEventsRelations:()=>jd,sessionInstructions:()=>$,sessionInstructionsRelations:()=>zd,sessionSendJobs:()=>On,sessionTerminalTargets:()=>wt,sessions:()=>a,sessionsRelations:()=>Td,skillSubmissions:()=>yr,snapshotDeltaObjects:()=>lr,teams:()=>v,teamsRelations:()=>md,treatmentAssignments:()=>C,treatmentAssignmentsRelations:()=>Rd,treatmentExposures:()=>q,treatmentExposuresRelations:()=>kd,userOrganizations:()=>Ne,userOrganizationsRelations:()=>Nd,userTeams:()=>xe,userTeamsRelations:()=>xd,vendorGuidanceSnapshots:()=>Yo,weeklyDigestDeliveries:()=>Fr,weeklyDigestRuns:()=>$r,workerConnections:()=>Bn});import{relations as i}from"drizzle-orm";var cd=i(o,({many:e})=>({sessions:e(a),contextSnapshots:e(X),apiKeys:e(Ee)})),pd=i(Ee,({one:e})=>({profile:e(o,{fields:[Ee.userId],references:[o.id]})})),_d=i(l,({many:e})=>({teams:e(v),userOrganizations:e(Ne),projects:e(s),invitations:e(P)})),md=i(v,({one:e,many:t})=>({organization:e(l,{fields:[v.orgId],references:[l.id]}),userTeams:t(xe),invitations:t(P)})),fd=i(P,({one:e})=>({organization:e(l,{fields:[P.orgId],references:[l.id]}),team:e(v,{fields:[P.teamId],references:[v.id]}),inviter:e(o,{fields:[P.invitedBy],references:[o.id]})})),Nd=i(Ne,({one:e})=>({organization:e(l,{fields:[Ne.orgId],references:[l.id]})})),xd=i(xe,({one:e})=>({team:e(v,{fields:[xe.teamId],references:[v.id]})})),hd=i(n,({one:e,many:t})=>({versions:t(he),projectInstructions:t(F),events:t(ge),usageEvents:t(je),sessionInstructions:t($),packInstructions:t(Y),forkedFrom:e(n,{fields:[n.forkedFromId],references:[n.id]})})),gd=i(ge,({one:e})=>({instruction:e(n,{fields:[ge.instructionId],references:[n.id]})})),vd=i(w,({one:e,many:t})=>({project:e(s,{fields:[w.projectId],references:[s.id]}),parent:e(w,{fields:[w.parentId],references:[w.id]}),versions:t(w)})),Id=i(he,({one:e})=>({instruction:e(n,{fields:[he.instructionId],references:[n.id]})})),yd=i(s,({one:e,many:t})=>({organization:e(l,{fields:[s.orgId],references:[l.id]}),projectInstructions:t(F),sessions:t(a)})),wd=i(F,({one:e})=>({project:e(s,{fields:[F.projectId],references:[s.id]}),instruction:e(n,{fields:[F.instructionId],references:[n.id]})})),Td=i(a,({one:e,many:t})=>({profile:e(o,{fields:[a.userId],references:[o.id]}),project:e(s,{fields:[a.projectId],references:[s.id]}),run:e(h,{fields:[a.runId],references:[h.id]}),events:t(Ie),sessionInstructions:t($),treatmentExposures:t(q),contextSnapshots:t(X)})),Ad=i(h,({one:e,many:t})=>({profile:e(o,{fields:[h.userId],references:[o.id]}),organization:e(l,{fields:[h.orgId],references:[l.id]}),project:e(s,{fields:[h.projectId],references:[s.id]}),sessions:t(a),launchJobs:t(Pe)})),Dd=i(Pe,({one:e})=>({run:e(h,{fields:[Pe.runId],references:[h.id]})})),jd=i(Ie,({one:e})=>({session:e(a,{fields:[Ie.sessionId],references:[a.id]})})),zd=i($,({one:e})=>({session:e(a,{fields:[$.sessionId],references:[a.id]}),instruction:e(n,{fields:[$.instructionId],references:[n.id]})})),Rd=i(C,({one:e,many:t})=>({owner:e(o,{fields:[C.userId],references:[o.id]}),project:e(s,{fields:[C.projectId],references:[s.id]}),exposures:t(q)})),kd=i(q,({one:e})=>({session:e(a,{fields:[q.sessionId],references:[a.id]}),owner:e(o,{fields:[q.userId],references:[o.id]}),assignment:e(C,{fields:[q.assignmentId],references:[C.id]})})),Ed=i(j,({one:e,many:t})=>({packInstructions:t(Y),forkedFrom:e(j,{fields:[j.forkedFromId],references:[j.id]})})),Sd=i(Y,({one:e})=>({pack:e(j,{fields:[Y.packId],references:[j.id]}),instruction:e(n,{fields:[Y.instructionId],references:[n.id]})})),bd=i(je,({one:e})=>({instruction:e(n,{fields:[je.instructionId],references:[n.id]})})),Pd=i(r,({one:e,many:t})=>({parent:e(r,{fields:[r.parentId],references:[r.id]}),evals:t(M),evalRuns:t(E),benchmarks:t(ze),evalJobs:t(b),evalConversations:t(Se)})),Kd=i(M,({one:e,many:t})=>({evalTarget:e(r,{fields:[M.evalTargetId],references:[r.id]}),runs:t(E)})),Cd=i(E,({one:e})=>({eval:e(M,{fields:[E.evalId],references:[M.id]}),evalTarget:e(r,{fields:[E.evalTargetId],references:[r.id]})})),qd=i(ze,({one:e})=>({evalTarget:e(r,{fields:[ze.evalTargetId],references:[r.id]})})),Od=i(b,({one:e})=>({evalTarget:e(r,{fields:[b.evalTargetId],references:[r.id]})})),Md=i(Se,({one:e})=>({evalTarget:e(r,{fields:[Se.evalTargetId],references:[r.id]})})),Hd=i(X,({one:e})=>({session:e(a,{fields:[X.sessionId],references:[a.id]}),profile:e(o,{fields:[X.userId],references:[o.id]})})),Ld=i(te,({one:e,many:t})=>({target:e(r,{fields:[te.evalTargetId],references:[r.id]}),calls:t(_e)})),Ud=i(_e,({one:e})=>({study:e(te,{fields:[_e.studyId],references:[te.id]}),job:e(b,{fields:[_e.jobId],references:[b.id]})}));var at=0,os=0,dt=0,ns=0;var Bd=0;function Xo(){at++,at>os&&(os=at)}m(Xo,"trackDbOpStart");function Zo(){at>0&&at--}m(Zo,"trackDbOpEnd");function en(){dt++,dt>ns&&(ns=dt)}m(en,"trackHeavyDbOpStart");function tn(){dt>0&&dt--}m(tn,"trackHeavyDbOpEnd");function rs(){Bd++}m(rs,"recordQueryDeadline");var Gt,Ft,on=!1,Bt;var ss,$t,Vt;function ds(e){let t=process.env[e];if(!t)throw new Error(`Missing required env var: ${e}`);return t}m(ds,"getEnv");function us(){if(on)throw new Error("getDb() called after closeDb(): the connection pool is closed for shutdown. Stop issuing queries (timers/requests) before draining the pool.");if(!Gt){let e=process.env.PG_APPLICATION_NAME||"rulemetric",t=Number(process.env.PG_POOL_MAX)||3,d=Number(process.env.PG_CONNECT_TIMEOUT_S)||10,u=Number(process.env.PG_STATEMENT_TIMEOUT_MS)||0,_=Number(process.env.PG_IDLE_TIMEOUT)||5,x=Number(process.env.PG_KEEP_ALIVE_S)||60;Ft=as(ds("DATABASE_URL"),{prepare:!1,max:t,connect_timeout:d,idle_timeout:_,keep_alive:x,max_lifetime:300,connection:{application_name:e,...u>0?{statement_timeout:u}:{}}}),Gt=is(Ft,{schema:Ut})}return Gt}m(us,"getDb");function ls(){if(on)throw new Error("getHeavyDb() called after closeDb(): the connection pool is closed for shutdown. Stop issuing queries (timers/requests) before draining the pool.");if(!Vt){let e=(process.env.PG_APPLICATION_NAME||"rulemetric")+":heavy",t=Number(process.env.PG_HEAVY_POOL_MAX)||2;$t=as(ds("DATABASE_URL"),{prepare:!1,max:t,connect_timeout:Number(process.env.PG_CONNECT_TIMEOUT_S)||10,idle_timeout:Number(process.env.PG_IDLE_TIMEOUT)||5,keep_alive:Number(process.env.PG_KEEP_ALIVE_S)||60,max_lifetime:300,connection:{application_name:e}}),Vt=is($t,{schema:Ut})}return Vt}m(ls,"getHeavyDb");async function Gd(){if(Bt)return Bt;on=!0;let e=Ft,t=ss,d=$t;return Ft=void 0,Gt=void 0,ss=void 0,$t=void 0,Vt=void 0,Bt=Promise.allSettled([e?e.end({timeout:5}):Promise.resolve(),t?t.end({timeout:5}):Promise.resolve(),d?d.end({timeout:5}):Promise.resolve()]).then(()=>{}),Bt}m(Gd,"closeDb");var Vd=new Set(["CONNECT_TIMEOUT","ETIMEDOUT","ENOTFOUND","ECONNREFUSED","EAI_AGAIN"]);function cs(e){let t=e,d=t?.code??t?.cause?.code??t?.errno;return d!==void 0&&Vd.has(d)}m(cs,"isRetryableConnectError");var Fd=new Set(["CONNECTION_CLOSED","ECONNRESET","EPIPE"]);function ps(e){let t=e,d=t?.code??t?.cause?.code??t?.errno;return d!==void 0&&Fd.has(d)}m(ps,"isRetryableResetError");async function _s(e,t={}){let d=t.attempts??3,u=t.baseMs??100,_;for(let x=0;x<d;x++)try{return await e()}catch(A){_=A;let B=cs(A)||t.retryResets===!0&&ps(A);if(x===d-1||!B)throw A;await new Promise(oe=>setTimeout(oe,u*2**x))}throw _}m(_s,"withDbRetry");var Jt=class extends Error{static{m(this,"QueryDeadlineError")}code="QUERY_DEADLINE";constructor(t){super(`Query exceeded the client-side deadline of ${t}ms \u2014 the DB pool is likely saturated (acquire-wait) or holding a dead socket; failing fast to avoid pinning the caller. See client.ts.`),this.name="QueryDeadlineError"}};function ms(){let e=Number(process.env.PG_QUERY_DEADLINE_MS);return Number.isFinite(e)&&e>=0?e:2e4}m(ms,"queryDeadlineMs");function fs(){let e=Number(process.env.PG_HEAVY_QUERY_DEADLINE_MS);return Number.isFinite(e)&&e>=0?e:6e4}m(fs,"heavyQueryDeadlineMs");function $d(e,t,d=!0){let u,_=new Promise((x,A)=>{u=setTimeout(()=>{d&&rs(),A(new Jt(t))},t);let B=u;typeof B.unref=="function"&&B.unref()});return Promise.race([e,_]).finally(()=>clearTimeout(u))}m($d,"raceWithDeadline");var Ns=new Set(["select","selectDistinct","selectDistinctOn","insert","update","delete","execute","with","$count","refreshMaterializedView"]),xs=new Set(["select","selectDistinct","selectDistinctOn","$count"]);function nn(e,t=!1,d=0,u=!1){return e===null||typeof e!="object"&&typeof e!="function"?e:new Proxy(e,{get(_,x,A){if(x==="then"){let oe=Reflect.get(_,x,A);return typeof oe!="function"?oe:(vs,Is)=>{let ys=u?en:Xo,ws=u?tn:Zo;ys();let rn=_s(()=>new Promise((Ts,As)=>oe.call(_,Ts,As)),{retryResets:t});return(d>0?$d(rn,d,!u):rn).finally(ws).then(vs,Is)}}let B=Reflect.get(_,x,A);return typeof B=="function"?(...oe)=>nn(B.apply(_,oe),t,d,u):B}})}m(nn,"wrapBuilder");function hs(e,t=!1){let d=t?en:Xo,u=t?tn:Zo;d();try{return Promise.resolve(e()).finally(u)}catch(_){throw u(),_}}m(hs,"gaugeTransaction");var gs=new Proxy({},{get(e,t){let d=us(),u=d[t];if(typeof t=="string"&&Ns.has(t)&&typeof u=="function"){let _=xs.has(t),x=ms();return(...A)=>nn(u.apply(d,A),_,x)}return t==="transaction"&&typeof u=="function"?(..._)=>hs(()=>u.apply(d,_)):u}}),Jd=new Proxy({},{get(e,t){let d=ls(),u=d[t];if(typeof t=="string"&&Ns.has(t)&&typeof u=="function"){let _=xs.has(t);return(...x)=>nn(u.apply(d,x),_,fs(),!0)}return t==="transaction"&&typeof u=="function"?(..._)=>hs(()=>u.apply(d,_),!0):u}});export{an as a,us as b,Gd as c};

package/dist/{chunk-VXNICAVG.js → chunk-VUNJGZPP.js}

@@ -1 +1 @@
-import{a as E}from"./chunk-DQYYHX6K.js";import{a as I}from"./chunk-RBSEE2Z6.js";import{a as w}from"./chunk-M7QSC64D.js";import{c as v,e as h}from"./chunk-T47SFA5X.js";import{b as R}from"./chunk-SWSOEYC7.js";import{e as m}from"./chunk-GIXHG5FK.js";import{a as d}from"./chunk-ZLLER3HP.js";var T=6e5;async function P(e,$){let i=d(n=>$?.logger?.info?.(n)??console.log(n),"log"),{evalTargetId:r}=e,k=e.measurement?v(e):{},o=await m(`/api/eval-targets/${r}/judge-calibration`),l=!e.measurement||o.methodology?.measurement?.engine===e.measurement.engine&&o.methodology.measurement.model===e.measurement.model&&o.methodology.current===R(e.measurement.engine),a=I({powered:l&&o.powerCheck.powered,mde:l?o.powerCheck.mde??Number.POSITIVE_INFINITY:Number.POSITIVE_INFINITY},(l?o.caseDiscrimination:void 0)??{total:0,discriminating:0,nonDiscriminating:0,unpaired:0},{runsCompleted:0,maxRuns:e.maxRuns??24,batchesRun:0,maxBatches:1});if(a.action==="stop")return i(`[cron_eval_autorun] ${r}: ${a.reason}`),{skipped:a.code,reason:a.reason};let t=await m(`/api/eval-targets/${r}`);if(t.metadata?.autoRunEnabled===!1)return{skipped:"paused",reason:"Nightly measurement is off for this target."};let c=w(t);if(c)return i(`[cron_eval_autorun] ${r}: ${c.reason}`),c;let p=(await m(`/api/eval-targets/${r}/evals`)).filter(n=>!n.retiredAt);if(p.length===0)return{skipped:"no-live-cases"};let{completed:f,failed:b,results:N}=await h({cases:E(p,e.maxRuns??12),targetContent:t.content,concurrency:2,runNumber:Math.floor(Date.now()/1e3),opts:{targetId:r,targetVersion:t.version??1,iteration:1,timeout:e.timeout??T,model:e.model,graderModel:e.graderModel,graderEngine:e.graderEngine,...k,skipGrading:!1,targetType:t.type,targetName:t.name},onLog:i}),g={};for(let n of N){if(n.status!=="failed")continue;let s=n.failure?.kind??"unknown";g[s]=(g[s]??0)+1}let u=Object.entries(g).sort(([,n],[,s])=>s-n).map(([n,s])=>`${n}x${s}`).join(", ");return i(`[cron_eval_autorun] ${t.name}: ${f} completed, ${b} failed`+(u===""?"":` (${u})`)),{ranRuns:f,failedRuns:b,...u===""?{}:{failures:u},reason:a.reason}}d(P,"cronEvalAutorun");export{P as a};
+import{a as E}from"./chunk-DQYYHX6K.js";import{a as I}from"./chunk-RBSEE2Z6.js";import{a as w}from"./chunk-M7QSC64D.js";import{c as v,e as h}from"./chunk-FADINV4H.js";import{b as R}from"./chunk-SWSOEYC7.js";import{e as m}from"./chunk-GIXHG5FK.js";import{a as d}from"./chunk-ZLLER3HP.js";var T=6e5;async function P(e,$){let i=d(n=>$?.logger?.info?.(n)??console.log(n),"log"),{evalTargetId:r}=e,k=e.measurement?v(e):{},o=await m(`/api/eval-targets/${r}/judge-calibration`),l=!e.measurement||o.methodology?.measurement?.engine===e.measurement.engine&&o.methodology.measurement.model===e.measurement.model&&o.methodology.current===R(e.measurement.engine),a=I({powered:l&&o.powerCheck.powered,mde:l?o.powerCheck.mde??Number.POSITIVE_INFINITY:Number.POSITIVE_INFINITY},(l?o.caseDiscrimination:void 0)??{total:0,discriminating:0,nonDiscriminating:0,unpaired:0},{runsCompleted:0,maxRuns:e.maxRuns??24,batchesRun:0,maxBatches:1});if(a.action==="stop")return i(`[cron_eval_autorun] ${r}: ${a.reason}`),{skipped:a.code,reason:a.reason};let t=await m(`/api/eval-targets/${r}`);if(t.metadata?.autoRunEnabled===!1)return{skipped:"paused",reason:"Nightly measurement is off for this target."};let c=w(t);if(c)return i(`[cron_eval_autorun] ${r}: ${c.reason}`),c;let p=(await m(`/api/eval-targets/${r}/evals`)).filter(n=>!n.retiredAt);if(p.length===0)return{skipped:"no-live-cases"};let{completed:f,failed:b,results:N}=await h({cases:E(p,e.maxRuns??12),targetContent:t.content,concurrency:2,runNumber:Math.floor(Date.now()/1e3),opts:{targetId:r,targetVersion:t.version??1,iteration:1,timeout:e.timeout??T,model:e.model,graderModel:e.graderModel,graderEngine:e.graderEngine,...k,skipGrading:!1,targetType:t.type,targetName:t.name},onLog:i}),g={};for(let n of N){if(n.status!=="failed")continue;let s=n.failure?.kind??"unknown";g[s]=(g[s]??0)+1}let u=Object.entries(g).sort(([,n],[,s])=>s-n).map(([n,s])=>`${n}x${s}`).join(", ");return i(`[cron_eval_autorun] ${t.name}: ${f} completed, ${b} failed`+(u===""?"":` (${u})`)),{ranRuns:f,failedRuns:b,...u===""?{}:{failures:u},reason:a.reason}}d(P,"cronEvalAutorun");export{P as a};

package/dist/{chunk-J6RQ5AH2.js → chunk-W6MMJRFZ.js}

@@ -1,4 +1,4 @@
-import{e as b,g as y}from"./chunk-7BCZDVIU.js";import{e as M,f as w}from"./chunk-GIXHG5FK.js";import{a as u}from"./chunk-ZLLER3HP.js";var $={anthropic:["https://platform.claude.com/docs/en/about-claude/pricing.md"],openai:["https://platform.openai.com/docs/pricing"]},R=.7,v=16e3,x=["You extract LLM API pricing from a provider pricing page.","You will be given the page text in <data> tags and, in this system prompt,","a list of model ids to find. For EACH requested model id, return its","published USD price per 1,000,000 tokens.","","Rules:","- Only report a model if you can find its rate ON THE PAGE. If it is not on","  the page, set found=false and omit the numbers \u2014 never guess from memory.","- inputPerM and outputPerM are the standard (non-batch, non-priority) rates.","- cacheReadPerM: the cached/discounted-input rate if published, else omit.","- For Anthropic models ONLY, if cache-write rates are not explicitly listed,","  compute them as cacheWrite5mPerM = inputPerM * 1.25 and","  cacheWrite1hPerM = inputPerM * 2.0. For all other providers, omit them.","- confidence is 0..1: how sure you are the numbers are correct for THIS id.","- contextWindow in tokens if stated, else omit.","","Respond with ONLY a JSON object of the form:",'{"rates":[{"modelId":"...","found":true,"confidence":0.9,"inputPerM":5,"outputPerM":25,"cacheReadPerM":0.5,"contextWindow":1000000}]}'].join(`
+import{e as b,g as y}from"./chunk-XCH4HPEZ.js";import{e as M,f as w}from"./chunk-GIXHG5FK.js";import{a as u}from"./chunk-ZLLER3HP.js";var $={anthropic:["https://platform.claude.com/docs/en/about-claude/pricing.md"],openai:["https://platform.openai.com/docs/pricing"]},R=.7,v=16e3,x=["You extract LLM API pricing from a provider pricing page.","You will be given the page text in <data> tags and, in this system prompt,","a list of model ids to find. For EACH requested model id, return its","published USD price per 1,000,000 tokens.","","Rules:","- Only report a model if you can find its rate ON THE PAGE. If it is not on","  the page, set found=false and omit the numbers \u2014 never guess from memory.","- inputPerM and outputPerM are the standard (non-batch, non-priority) rates.","- cacheReadPerM: the cached/discounted-input rate if published, else omit.","- For Anthropic models ONLY, if cache-write rates are not explicitly listed,","  compute them as cacheWrite5mPerM = inputPerM * 1.25 and","  cacheWrite1hPerM = inputPerM * 2.0. For all other providers, omit them.","- confidence is 0..1: how sure you are the numbers are correct for THIS id.","- contextWindow in tokens if stated, else omit.","","Respond with ONLY a JSON object of the form:",'{"rates":[{"modelId":"...","found":true,"confidence":0.9,"inputPerM":5,"outputPerM":25,"cacheReadPerM":0.5,"contextWindow":1000000}]}'].join(`
 `);async function C(f){try{let n=await fetch(f,{headers:{accept:"text/plain, text/markdown, text/html"}});if(!n.ok)return null;let i=await n.text();return i.length>v?i.slice(0,v):i}catch{return null}}u(C,"fetchPage");var L=u(async(f,n)=>{let i=await M("/api/pricing/coverage"),I=[...i.unpriced,...i.stale],m=new Map;for(let r of I)m.set(`${r.provider}::${r.modelId}`,r);let a=[...m.values()];if(a.length===0)return n.logger.info("[cron-refresh-pricing] coverage clean \u2014 nothing unpriced or stale"),{status:"clean",checked:i.unpriced.length+i.stale.length};if(b()==="none")return n.logger.warn(`[cron-refresh-pricing] ${a.length} model(s) need pricing but no LLM backend (install the claude CLI or set ANTHROPIC_API_KEY) \u2014 skipping`),{status:"no-llm",pending:a.length};let d=new Map;for(let r of a){let o=d.get(r.provider)??[];o.push(r.modelId),d.set(r.provider,o)}let t={written:0,unchanged:0,flagged:0,rejected:0,skipped:0},c=[];for(let[r,o]of d){let h=$[r];if(!h){n.logger.warn(`[cron-refresh-pricing] no pricing source configured for provider '${r}' (${o.length} model(s)) \u2014 skipping`),t.skipped+=o.length;continue}let p="",l="";for(let e of h){let s=await C(e);s&&(p+=`
 
 === ${e} ===

package/dist/{chunk-XZW5ZHFC.js → chunk-WQFR2LOY.js}

@@ -1,4 +1,4 @@
-import{a as M}from"./chunk-5RBLIVBF.js";import{c as q}from"./chunk-GLBWNB2N.js";import{a as K,b as Y}from"./chunk-CQD2R3BA.js";import{a as A,d as N}from"./chunk-3UKZJAFS.js";import{a as j,b as k,c as V,d as D}from"./chunk-77FH6UZP.js";import{a as Z}from"./chunk-KE5K3F7J.js";import{b as W,e as B,g as T}from"./chunk-7BCZDVIU.js";import{e as h,f as U,h as m}from"./chunk-GIXHG5FK.js";import{a as L,b as H}from"./chunk-X6CF5KTW.js";import{a as f}from"./chunk-ZLLER3HP.js";var Q=f((l,{limit:a,sort:n,eventTypes:s})=>h(`/api/sessions/${l}/events?limit=${a}&sort=${n}&eventType=${encodeURIComponent(s.join(","))}&hasContent=true`),"fetchTranscriptTurns");async function ms(l,a){let{mirrorId:n,projectPath:s,projectId:r,config:u}=l,e=f(t=>a.logger.info(t),"log");if(u?.sessionId)return es({mirrorId:n,sessionId:u.sessionId},e);e(`Processing insights job ${n}...`),await m(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:t}=await U("/api/sessions/repair-analyzed-at",{});t>0&&e(`  Repaired analyzed_at on ${t} orphan session(s).`)}catch(t){e(`  Warning: analyzed_at repair failed: ${t instanceof Error?t.message:String(t)}`)}let p=20,C=200,z=15*6e4,F=3,E=Date.now()+z,I=B()==="none"?"no-llm-backend":null;if(I&&a.logger.warn?.("[process-insights] no LLM backend (no claude/codex CLI and no ANTHROPIC_API_KEY) \u2014 facet grading SKIPPED for this run; the summary below is computed from already-graded sessions only"),!I){let t=0,i=0;try{for(;t<C&&Date.now()<E;){let c=await h(`/api/sessions?needsFacets=true&graderVersion=${2}&limit=${p}&offset=${i}`);if(c.length===0)break;e(`  Generating facets for ${c.length} un-analyzed sessions (drained ${t} so far)...`);for(let b=0;b<c.length&&!(t>=C||Date.now()>=E);b+=F){let R=c.slice(b,b+F);await Promise.all(R.map(async g=>{if(!(t>=C||Date.now()>=E))try{let d=await N(ss=>Q(g.id,ss));if(d.turnCount<2){t++,i++;return}let O=await T(D,d.text,4096);O&&typeof O.outcome=="string"?(await m(`/api/sessions/${g.id}/enrich`,{analysis:{source:"generated_facets",...O,graderVersion:2},...g.eventCount!=null?{analyzedEventCount:g.eventCount}:{},analyzedTurnCount:d.turnCount}),e(`    Generated facets for session ${g.id} (${g.tool}, ${d.turnCount} turns)`)):i++,t++}catch(d){e(`    Failed to generate facets for session ${g.id}: ${d instanceof Error?d.message:String(d)}`),t++,i++}}))}if(c.length<p)break}}catch(c){e(`  Warning: facet generation failed after ${t} sessions: ${c instanceof Error?c.message:String(c)}`)}if(t>0){let c=Date.now()>=E?`time budget (${Math.round(z/6e4)}m)`:t>=C?`hard cap (${C})`:"queue empty";e(`  Facet drain stopped on ${c} after ${t} session(s); the rest resume next run.`)}}let _=await ns(s,r);if(!_)throw new M("No session data found \u2014 ensure sessions are captured and enriched.");e(`  Found ${_.stats.sessions} sessions, generating recommendations...`);let y=await Y({projectId:r,projectPath:s},h);_.summary+=`
+import{a as M}from"./chunk-5RBLIVBF.js";import{c as q}from"./chunk-GLBWNB2N.js";import{a as K,b as Y}from"./chunk-CQD2R3BA.js";import{a as A,d as N}from"./chunk-3UKZJAFS.js";import{a as j,b as k,c as V,d as D}from"./chunk-77FH6UZP.js";import{a as Z}from"./chunk-KE5K3F7J.js";import{b as W,e as B,g as T}from"./chunk-XCH4HPEZ.js";import{e as h,f as U,h as m}from"./chunk-GIXHG5FK.js";import{a as L,b as H}from"./chunk-X6CF5KTW.js";import{a as f}from"./chunk-ZLLER3HP.js";var Q=f((l,{limit:a,sort:n,eventTypes:s})=>h(`/api/sessions/${l}/events?limit=${a}&sort=${n}&eventType=${encodeURIComponent(s.join(","))}&hasContent=true`),"fetchTranscriptTurns");async function ms(l,a){let{mirrorId:n,projectPath:s,projectId:r,config:u}=l,e=f(t=>a.logger.info(t),"log");if(u?.sessionId)return es({mirrorId:n,sessionId:u.sessionId},e);e(`Processing insights job ${n}...`),await m(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:"reading_data"}});try{let{repaired:t}=await U("/api/sessions/repair-analyzed-at",{});t>0&&e(`  Repaired analyzed_at on ${t} orphan session(s).`)}catch(t){e(`  Warning: analyzed_at repair failed: ${t instanceof Error?t.message:String(t)}`)}let p=20,C=200,z=15*6e4,F=3,E=Date.now()+z,I=B()==="none"?"no-llm-backend":null;if(I&&a.logger.warn?.("[process-insights] no LLM backend (no claude/codex CLI and no ANTHROPIC_API_KEY) \u2014 facet grading SKIPPED for this run; the summary below is computed from already-graded sessions only"),!I){let t=0,i=0;try{for(;t<C&&Date.now()<E;){let c=await h(`/api/sessions?needsFacets=true&graderVersion=${2}&limit=${p}&offset=${i}`);if(c.length===0)break;e(`  Generating facets for ${c.length} un-analyzed sessions (drained ${t} so far)...`);for(let b=0;b<c.length&&!(t>=C||Date.now()>=E);b+=F){let R=c.slice(b,b+F);await Promise.all(R.map(async g=>{if(!(t>=C||Date.now()>=E))try{let d=await N(ss=>Q(g.id,ss));if(d.turnCount<2){t++,i++;return}let O=await T(D,d.text,4096);O&&typeof O.outcome=="string"?(await m(`/api/sessions/${g.id}/enrich`,{analysis:{source:"generated_facets",...O,graderVersion:2},...g.eventCount!=null?{analyzedEventCount:g.eventCount}:{},analyzedTurnCount:d.turnCount}),e(`    Generated facets for session ${g.id} (${g.tool}, ${d.turnCount} turns)`)):i++,t++}catch(d){e(`    Failed to generate facets for session ${g.id}: ${d instanceof Error?d.message:String(d)}`),t++,i++}}))}if(c.length<p)break}}catch(c){e(`  Warning: facet generation failed after ${t} sessions: ${c instanceof Error?c.message:String(c)}`)}if(t>0){let c=Date.now()>=E?`time budget (${Math.round(z/6e4)}m)`:t>=C?`hard cap (${C})`:"queue empty";e(`  Facet drain stopped on ${c} after ${t} session(s); the rest resume next run.`)}}let _=await ns(s,r);if(!_)throw new M("No session data found \u2014 ensure sessions are captured and enriched.");e(`  Found ${_.stats.sessions} sessions, generating recommendations...`);let y=await Y({projectId:r,projectPath:s},h);_.summary+=`
 
 ${y.text}`,e(`  Session evidence: ${y.sessionIds.length} usable, ${y.failedSessionIds.length} unavailable`);let S=await ts(s,e);S&&e(`  Generator memory: ${S.split(`
 `).length} lines of prior-advice context`);let o={},$=[],w=k.length+1;for(let t=0;t<k.length;t++){let i=k[t];e(`  [${t+1}/${w}] Generating ${i.name}...`);try{await m(`/api/insights-jobs/${n}/progress`,{status:"running",progress:{step:i.name,current:t+1,total:w,partial:{...o}}})}catch(d){e(`  Warning: progress update failed for ${i.name}: ${d instanceof Error?d.message:String(d)}`)}e(`  Running LLM prompt for ${i.name}...`);let c=i.usesGeneratorMemory&&S?`${_.summary}

package/dist/{chunk-7BCZDVIU.js → chunk-XCH4HPEZ.js}

@@ -1,4 +1,4 @@
-import{a as S,b as P}from"./chunk-PPAHIVG5.js";import{a as E,b as C,c as _}from"./chunk-CT7FWSIU.js";import{a as V,b as v,c as T}from"./chunk-5MXZJWQS.js";import{a as b}from"./chunk-N7ODX6PO.js";import{d as K}from"./chunk-SG4UDDHB.js";import{c as z,d as H}from"./chunk-GPBN4ROE.js";import{a as g,b as U}from"./chunk-6XTLP2G7.js";import{a as c}from"./chunk-ZLLER3HP.js";import{execSync as M,spawn as ee}from"node:child_process";import{tmpdir as B}from"node:os";import{mkdtempSync as te,rmSync as ne,writeFileSync as re}from"node:fs";import{join as Y}from"node:path";function oe(t,o){let{available:r,backend:n}=o??{available:c(l=>g(l)!==null,"available"),backend:A()},e=P(t);if(r(e.bin))return{bin:e.bin,args:e.args,delivery:e.delivery??"stdin",substitutedFrom:null};let i=n==="codex-cli"?S.codex:n==="claude-cli"?S.claude_code:n==="copilot-cli"?S.copilot:null;return!i||!r(i.bin)?null:{bin:i.bin,args:i.args,delivery:i.delivery??"stdin",substitutedFrom:e.bin}}c(oe,"planToolExecution");function X(t,o=300){let r=t.trim().replace(/\s+/g," ");return r.length>o?`${r.slice(0,o)}\u2026`:r}c(X,"excerpt");async function ye(t,o,r){let n=oe(t);if(!n){let s=`no engine can run ${t}'s prompt: ${P(t).bin} is not installed and no substitutable backend was detected`;return console.warn(`[llm-client] ${s}`),{ok:!1,reason:s}}let{bin:e,args:i,delivery:l}=n;n.substitutedFrom&&console.warn(`[llm-client] ${n.substitutedFrom} is not installed \u2014 running ${t}'s analysis prompt on ${e} instead`);let a=g(e);if(!a){let s=`${e} not found on PATH, well-known dirs, or the node bin dir`;return console.warn(`[llm-client] ${s}`),{ok:!1,reason:s}}let u=`${o}
+import{a as S,b as P}from"./chunk-PPAHIVG5.js";import{a as E,b as C,c as _}from"./chunk-CT7FWSIU.js";import{a as V,b as v,c as T}from"./chunk-5MXZJWQS.js";import{a as b}from"./chunk-N7ODX6PO.js";import{d as K}from"./chunk-JNLDISBA.js";import{c as z,d as H}from"./chunk-GPBN4ROE.js";import{a as g,b as U}from"./chunk-6XTLP2G7.js";import{a as c}from"./chunk-ZLLER3HP.js";import{execSync as M,spawn as ee}from"node:child_process";import{tmpdir as B}from"node:os";import{mkdtempSync as te,rmSync as ne,writeFileSync as re}from"node:fs";import{join as Y}from"node:path";function oe(t,o){let{available:r,backend:n}=o??{available:c(l=>g(l)!==null,"available"),backend:A()},e=P(t);if(r(e.bin))return{bin:e.bin,args:e.args,delivery:e.delivery??"stdin",substitutedFrom:null};let i=n==="codex-cli"?S.codex:n==="claude-cli"?S.claude_code:n==="copilot-cli"?S.copilot:null;return!i||!r(i.bin)?null:{bin:i.bin,args:i.args,delivery:i.delivery??"stdin",substitutedFrom:e.bin}}c(oe,"planToolExecution");function X(t,o=300){let r=t.trim().replace(/\s+/g," ");return r.length>o?`${r.slice(0,o)}\u2026`:r}c(X,"excerpt");async function ye(t,o,r){let n=oe(t);if(!n){let s=`no engine can run ${t}'s prompt: ${P(t).bin} is not installed and no substitutable backend was detected`;return console.warn(`[llm-client] ${s}`),{ok:!1,reason:s}}let{bin:e,args:i,delivery:l}=n;n.substitutedFrom&&console.warn(`[llm-client] ${n.substitutedFrom} is not installed \u2014 running ${t}'s analysis prompt on ${e} instead`);let a=g(e);if(!a){let s=`${e} not found on PATH, well-known dirs, or the node bin dir`;return console.warn(`[llm-client] ${s}`),{ok:!1,reason:s}}let u=`${o}
 
 SESSION DATA:
 ${r}`,m={...b()};delete m.HTTPS_PROXY;let p=null,y=i,$=u;if(l==="argv-file"){p=te(Y(B(),"rulemetric-tool-prompt-"));let s=Y(p,"session-data.txt");re(s,r),y=[...i,`@${s}`,`${o}

package/dist/{chunk-RSVNQ2XI.js → chunk-YXMAGC53.js}

@@ -1 +1 @@
-import{d as E}from"./chunk-T47SFA5X.js";import{b as F,c as x}from"./chunk-KGURGIZ2.js";import{f as R}from"./chunk-GIXHG5FK.js";import{a}from"./chunk-ZLLER3HP.js";var _="with_candidate",w="with_prior_version";function S(e){return JSON.stringify([e.outputs.engine??null,e.outputs.model??null,e.outputs.methodology??null,e.outputs.deliveryRegime??null,e.grading.judge_engine??null,e.grading.judge_model??null])}a(S,"observedExecutionKey");function I(e){let r=e.filter(n=>n.status==="completed"),i=new Set(r.map(n=>JSON.stringify([n.outputs.engine??null,n.outputs.model??null,n.outputs.methodology??null,n.outputs.deliveryRegime??null]))),t=new Set(r.filter(n=>n.grading.judge_model!=null).map(n=>JSON.stringify([n.grading.judge_engine??"claude",n.grading.judge_model])));return i.size===1&&t.size<=1?JSON.stringify([[...i][0],[...t][0]??null]):null}a(I,"measurementBatchKey");function A(e){let r=e.filter(t=>t.configuration===_),i=[];for(let t of r){let n=e.find(s=>s.configuration===w&&s.evalId===t.evalId&&s.runNumber===t.runNumber);if(!n||S(n)!==S(t))continue;let l=h(t),m=h(n);l==null||m==null||i.push({withScore:l,withoutScore:m})}return i}a(A,"pairCandidateRuns");function h(e){return e.status!=="completed"?null:e.grading?.summary?.pass_rate??null}a(h,"evidenceOf");function $(e){let r=[],i=0;for(let t of e){if(t.status==="failed"){r.push(t.failure??{kind:"unknown",class:"infrastructure"});continue}h(t)==null&&(i+=1)}return{infrastructureFailed:r.length,ungradeable:i,summary:F(r)}}a($,"accountForFailures");function Q(e){let{pairs:r,minPairs:i,accounting:t}=e;return r>=i?{kind:"proceed"}:t.infrastructureFailed<=0?{kind:"proceed"}:{kind:"infrastructure-incomplete",reason:`${r}/${i} pair(s) \u2014 ${t.infrastructureFailed} run(s) never executed (${t.summary.describe()}). Retry when the environment recovers; this is not a statement about the target.`}}a(Q,"attributeShortfall");function k(e,r,i,t){let n=[];for(let l of e)n.push({eval_:l,configuration:_,targetContent:r,runNumber:t}),n.push({eval_:l,configuration:w,targetContent:i,runNumber:t});return n}a(k,"buildCandidateQueue");async function j(e){let{cases:r,candidateContent:i,priorVersionContent:t,concurrency:n,runNumber:l,opts:m,onLog:s=a(()=>{},"onLog")}=e,g=k(r,i,t,l),o=[],y=Math.max(1,n),p=null;for(let u=0;u<g.length;u+=y){let d=g.slice(u,u+y);o.push(...await Promise.all(d.map(c=>E(c,m,s))));let C=o.find(c=>c.failure&&x(c.failure.kind));if(C?.failure){p=C.failure.kind;let c=g.length-o.length;s(`  \u26D4 aborting batch \u2014 ${p}: the account is out of quota. Skipped ${c} remaining run(s) that would certainly have failed.`);break}}for(let u of o)try{await R("/api/eval-runs",u)}catch(d){s(`  \u26A0 failed to post a run result: ${d instanceof Error?d.message:String(d)}`)}let b=A(o),f=$(o),v=r.length-b.length;if(v>0){let u=f.infrastructureFailed?` \u2014 ${f.infrastructureFailed} run(s) never executed (${f.summary.describe()})`:"";s(`  \u26A0 ${v}/${r.length} case(s) produced no usable pair (failed or ungradeable arm)${u}`)}return{completed:o.filter(u=>u.status==="completed").length,failed:o.filter(u=>u.status==="failed").length,results:o,pairs:b,accounting:f,abortedForQuota:p}}a(j,"runCandidateBatchAndPost");export{_ as a,w as b,S as c,I as d,A as e,$ as f,Q as g,k as h,j as i};
+import{d as E}from"./chunk-FADINV4H.js";import{b as F,c as x}from"./chunk-KGURGIZ2.js";import{f as R}from"./chunk-GIXHG5FK.js";import{a}from"./chunk-ZLLER3HP.js";var _="with_candidate",w="with_prior_version";function S(e){return JSON.stringify([e.outputs.engine??null,e.outputs.model??null,e.outputs.methodology??null,e.outputs.deliveryRegime??null,e.grading.judge_engine??null,e.grading.judge_model??null])}a(S,"observedExecutionKey");function I(e){let r=e.filter(n=>n.status==="completed"),i=new Set(r.map(n=>JSON.stringify([n.outputs.engine??null,n.outputs.model??null,n.outputs.methodology??null,n.outputs.deliveryRegime??null]))),t=new Set(r.filter(n=>n.grading.judge_model!=null).map(n=>JSON.stringify([n.grading.judge_engine??"claude",n.grading.judge_model])));return i.size===1&&t.size<=1?JSON.stringify([[...i][0],[...t][0]??null]):null}a(I,"measurementBatchKey");function A(e){let r=e.filter(t=>t.configuration===_),i=[];for(let t of r){let n=e.find(s=>s.configuration===w&&s.evalId===t.evalId&&s.runNumber===t.runNumber);if(!n||S(n)!==S(t))continue;let l=h(t),m=h(n);l==null||m==null||i.push({withScore:l,withoutScore:m})}return i}a(A,"pairCandidateRuns");function h(e){return e.status!=="completed"?null:e.grading?.summary?.pass_rate??null}a(h,"evidenceOf");function $(e){let r=[],i=0;for(let t of e){if(t.status==="failed"){r.push(t.failure??{kind:"unknown",class:"infrastructure"});continue}h(t)==null&&(i+=1)}return{infrastructureFailed:r.length,ungradeable:i,summary:F(r)}}a($,"accountForFailures");function Q(e){let{pairs:r,minPairs:i,accounting:t}=e;return r>=i?{kind:"proceed"}:t.infrastructureFailed<=0?{kind:"proceed"}:{kind:"infrastructure-incomplete",reason:`${r}/${i} pair(s) \u2014 ${t.infrastructureFailed} run(s) never executed (${t.summary.describe()}). Retry when the environment recovers; this is not a statement about the target.`}}a(Q,"attributeShortfall");function k(e,r,i,t){let n=[];for(let l of e)n.push({eval_:l,configuration:_,targetContent:r,runNumber:t}),n.push({eval_:l,configuration:w,targetContent:i,runNumber:t});return n}a(k,"buildCandidateQueue");async function j(e){let{cases:r,candidateContent:i,priorVersionContent:t,concurrency:n,runNumber:l,opts:m,onLog:s=a(()=>{},"onLog")}=e,g=k(r,i,t,l),o=[],y=Math.max(1,n),p=null;for(let u=0;u<g.length;u+=y){let d=g.slice(u,u+y);o.push(...await Promise.all(d.map(c=>E(c,m,s))));let C=o.find(c=>c.failure&&x(c.failure.kind));if(C?.failure){p=C.failure.kind;let c=g.length-o.length;s(`  \u26D4 aborting batch \u2014 ${p}: the account is out of quota. Skipped ${c} remaining run(s) that would certainly have failed.`);break}}for(let u of o)try{await R("/api/eval-runs",u)}catch(d){s(`  \u26A0 failed to post a run result: ${d instanceof Error?d.message:String(d)}`)}let b=A(o),f=$(o),v=r.length-b.length;if(v>0){let u=f.infrastructureFailed?` \u2014 ${f.infrastructureFailed} run(s) never executed (${f.summary.describe()})`:"";s(`  \u26A0 ${v}/${r.length} case(s) produced no usable pair (failed or ungradeable arm)${u}`)}return{completed:o.filter(u=>u.status==="completed").length,failed:o.filter(u=>u.status==="failed").length,results:o,pairs:b,accounting:f,abortedForQuota:p}}a(j,"runCandidateBatchAndPost");export{_ as a,w as b,S as c,I as d,A as e,$ as f,Q as g,k as h,j as i};

package/dist/commands/diagnostics.js

@@ -1 +1 @@
-import{c as q,d as K}from"../chunk-TYFEWLYZ.js";import{b as L,c as W}from"../chunk-V5U3E4BZ.js";import{a as J}from"../chunk-2TJ7Q2OU.js";import{e as O,f as _,g as M}from"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import{d as oe,e as ne}from"../chunk-3NWJ7CZ2.js";import"../chunk-M7QSC64D.js";import"../chunk-6TGFU6RU.js";import{i as z,j as Q,k as Z,m as ee,n as te}from"../chunk-LB24UKHZ.js";import"../chunk-G7V4ZZZ3.js";import"../chunk-O4K4M5G6.js";import"../chunk-AHLVY52I.js";import"../chunk-RSVNQ2XI.js";import"../chunk-ZK3ACE4T.js";import"../chunk-7BCZDVIU.js";import"../chunk-PPAHIVG5.js";import"../chunk-CT7FWSIU.js";import"../chunk-5MXZJWQS.js";import"../chunk-N7ODX6PO.js";import{a as V}from"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-T47SFA5X.js";import"../chunk-SWSOEYC7.js";import"../chunk-U2NT5OSQ.js";import"../chunk-KGURGIZ2.js";import"../chunk-GYNZ3OZF.js";import"../chunk-UPFH72PF.js";import"../chunk-DVU5PHPM.js";import"../chunk-CDY7TQFZ.js";import"../chunk-VBXL7XO7.js";import"../chunk-FBSZ6KCM.js";import"../chunk-JP6ULOEX.js";import"../chunk-SG4UDDHB.js";import{d as F}from"../chunk-GPBN4ROE.js";import"../chunk-6XTLP2G7.js";import"../chunk-PJ3WHQ67.js";import{a as X,b as D,c as Y,g as j,h as G}from"../chunk-QI3FJKGK.js";import{a as B}from"../chunk-WUCACPGC.js";import"../chunk-IZTIYR5Q.js";import"../chunk-WL5ZWQAS.js";import"../chunk-KEVDLLHT.js";import"../chunk-DX6IPDED.js";import"../chunk-3AV3GTMP.js";import{a as U,e as y}from"../chunk-GIXHG5FK.js";import"../chunk-X6CF5KTW.js";import"../chunk-XQYAVSUQ.js";import{a as N,g as v,l as x}from"../chunk-PZNWH4VQ.js";import{a as S}from"../chunk-ZLLER3HP.js";import{spawnSync as le}from"node:child_process";import{existsSync as ce,readFileSync as ue}from"node:fs";import{join as se}from"node:path";var I=class h extends B{static{S(this,"Diagnostics")}static description='Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';static examples=["<%= config.bin %> diagnostics"];baseUrl(){return process.env.RULEMETRIC_API_URL??N().apiUrl??U}async checkDirect(){let r=process.env.HTTPS_PROXY,i=process.env.HTTP_PROXY;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let s=new AbortController,l=setTimeout(()=>s.abort(),8e3);try{return(await globalThis.fetch(`${this.baseUrl()}/health`,{signal:s.signal})).ok}catch{return!1}finally{clearTimeout(l),r!==void 0&&(process.env.HTTPS_PROXY=r),i!==void 0&&(process.env.HTTP_PROXY=i)}}configuredProxy(){if(process.env.HTTPS_PROXY)return process.env.HTTPS_PROXY;for(let r of["settings.local.json","settings.json"]){let i=se(process.cwd(),".claude",r);if(ce(i))try{let s=JSON.parse(ue(i,"utf8"));if(s.env?.HTTPS_PROXY)return s.env.HTTPS_PROXY}catch{}}return null}checkThroughProxy(r){let i=le("curl",["-sS","-x",r,"--max-time","8","-o","/dev/null","-w","%{http_code}",`${this.baseUrl()}/health`],{encoding:"utf8",stdio:["ignore","pipe","ignore"]});return i.error||typeof i.stdout!="string"?null:i.stdout.trim()==="200"}async run(){await this.parse(h);let r=process.version,i=this.configuredProxy(),s=i!=null,[l,m,d]=await Promise.all([this.checkDirect(),M(O),M(_)]),p=s&&i?this.checkThroughProxy(i):null,t=S((e,c,T)=>{let u=c===null?"\u2013":c?"[OK]":"[!!]";this.log(`  ${u} ${e.padEnd(22)} ${T}`)},"line");this.log("RuleMetric diagnostics"),this.log(""),t("Node",!L(r),`${r}${L(r)?" (< 20.11 \u2014 breaks the bundled proxy)":""}`),t("API (direct)",l,`${this.baseUrl()}/health ${l?"reachable":"UNREACHABLE"}`),t("Authenticated",v(),v()?"yes":"no \u2014 run `rulemetric auth login`"),t("Proxy configured",s?!0:null,s?i:"no HTTPS_PROXY set"),s&&(t("Gateway :"+O,m,m?"listening":"NOT listening"),t("mitmproxy :"+_,d,d?"listening":"not listening (gateway should route direct)"),t("API (via proxy)",p,p===null?"inconclusive (curl unavailable)":p?"reachable":"FAILED \u2014 502/blocked through your proxy"));let a=F();t("claude CLI",a.claudeBin!=null,a.claudeBin??"not found \u2014 LLM features (grading, distill) will no-op"),t("codex CLI",a.codexBin!=null?!0:null,a.codexBin?`${a.codexBin}${a.codexConfiguredModel?` (model ${a.codexConfiguredModel})`:""}`:"not found \u2014 optional; enables the cross-vendor reviewer (evals validate-judge)");let o=null;if(v()&&l)try{o=await y("/api/workers/health")}catch{o=null}if(o){let e=o.worker;t("Worker",!!e.active,e.present?e.active?`online (last heartbeat ${e.ageSeconds??"?"}s ago${e.clientVersion?`, v${e.clientVersion}`:""})`:`NOT RUNNING \u2014 last heartbeat ${A(e.ageSeconds)} ago`:"never connected \u2014 no worker has ever reported in");let c=o.worker.engines;if(c!=null){let u=[c.claude&&"claude",c.codex&&"codex"].filter(Boolean);t("LLM engine",u.length>0,u.length>0?`${u.join(" + ")} available to the worker`:"NONE \u2014 the worker found no `claude` and no `codex`. LLM jobs will complete with empty output rather than failing.")}let T=o.pending.reduce((u,E)=>u+E.count,0);if(T>0){let u=o.verdict.stuck.reduce(($,k)=>$+k.count,0),E=o.verdict.unclaimable.reduce(($,k)=>$+k.count,0),R=(o.verdict.engineBlocked??[]).reduce(($,k)=>$+k.count,0);t("Job queue",u===0&&R===0,`${T} pending`+(u>0?` \u2014 ${u} waiting on this worker`:"")+(R>0?` \u2014 ${R} need a local LLM engine`:"")+(E>0?` \u2014 ${E} need a worker with DATABASE_URL`:""))}else t("Job queue",!0,"empty")}else v()&&l&&t("Worker",null,"could not read /workers/health (older API?)");let n=o?.worker.machine;if(n){let e=n.suspends+n.stalls,c=n.lastPause;t("Worker pauses",e===0?!0:null,e===0?`none since ${n.since} \u2014 no sleep or stall seen by the worker`:`${n.suspends} suspend${n.suspends===1?"":"s"}`+(n.stalls>0?` + ${n.stalls} event-loop stall${n.stalls===1?"":"s"}`:"")+` since ${n.since}, ${A(Math.round(n.pausedMs/1e3))} paused total`+(c?`; last: ${c.kind} of ${A(Math.round(c.wallGapMs/1e3))} ending ${c.detectedAt}`:"")+" \u2014 jobs in flight during a suspend lose their lease and are redone on wake")}let f=await q();if(f){t("Machine sleep (pmset)",f.sleeps===0?!0:null,K(f));for(let e of f.recent.slice(-3)){let c=e.seconds!=null?` (${A(e.seconds)})`:"";this.log(`        \u21B3 ${e.at.toISOString()} ${e.kind}${c}: ${e.reason.slice(0,80)}`)}}let C=V(process.cwd()),g=await D({configDir:x()}),re=await Y({configDir:x()}),w=` \u2014 ${G(re)}`,b=J(process.cwd()),H=[b.user&&`user hooks (${b.user})`,b.workspace&&`workspace hooks (${b.workspace})`].filter(Boolean),ie={claude_code:`session hooks (+ proxy for instruction linking)${w}`,codex:`worker transcript sync of ~/.codex/sessions \u2014 ${j(g)}`,cursor:`user hooks (~/.cursor/hooks.json)${w}`,copilot_agent:`workspace hooks${w}`,antigravity:H.length>0?`${H.join(" + ")}${w}`:"NOT CAPTURED \u2014 no rulemetric block in ~/.gemini/config/hooks.json or .agents/hooks.json; run `rulemetric hooks install` (user hooks are on by default)"};this.log(""),this.log("  Capture coverage (agents found on this machine):");for(let e of C)this.log(`    \u2022 ${e.padEnd(15)} ${ie[e]??"captured"}`);if(C.includes("codex"))if(g.error)this.log(`    ! ${g.error}`);else if(g.mode==="allowlist")for(let e of g.roots)this.log(`        \u21B3 ${e}`);else g.mode==="all"&&this.log(`        \u21B3 scope it with ${se(x(),X)} ({"codex":{"roots":["/path/to/repo"]}}) or RULEMETRIC_CODEX_CAPTURE=off`);o&&!o.worker.active&&C.includes("codex")&&g.mode!=="off"&&this.log("    ! Codex capture runs INSIDE the worker \u2014 it is stopped while the worker is down."),v()&&l&&await this.reportEvolutionSpend();let P=W({nodeVersion:r,directApiOk:l,httpsProxySet:s,gatewayAlive:m,mitmproxyAlive:d,proxyPathOk:p});this.log("");let ae=P.severity==="ok"?"[OK]":P.severity==="warn"?"[!!]":"[XX]";this.log(`${ae} ${P.headline}`);for(let e of P.hints)this.log(`   \u2022 ${e}`);if(o&&o.verdict.severity!=="ok"){let e=o.verdict.severity==="warn"?"[!!]":"[XX]";this.log(""),this.log(`${e} ${o.verdict.headline}`);for(let c of o.verdict.hints)this.log(`   \u2022 ${c}`)}}async reportEvolutionSpend(){let r;try{r=await y("/api/eval-targets?type=instruction&limit=200")}catch{return}let i=r.filter(s=>{let l=s.metadata??{};return l.autoEvolveEnabled===!0||l.autoEvolveEnabled==="true"});if(i.length!==0){this.log(""),this.log("  Nightly instruction evolution (cron_instruction_evolution, 07:30 UTC):");for(let s of i){let l=s.metadata??{},m=Q(l),d=z(l);if(m){this.log(`    \u2022 ${s.name} \u2014 PAUSED (metadata.autoEvolvePaused). Opt-in preserved.`),this.log("        \u21B3 resume: set autoEvolvePaused=false");continue}let p=null;try{p=(await y(`/api/eval-targets/${s.id}/evals`)).filter(a=>!a.retiredAt).length}catch{}if(this.log(`    \u2022 ${s.name}`),p!=null){let t=ne(p,d.minPairs,oe(d,p));if(!t.feasible)this.log(`        \u21B3 cannot reach a verdict \u2014 ${t.reason}`);else{let a=Z({liveCases:p,passes:t.passes});this.log(`        \u21B3 each night: ${te(a)}`)}}this.log(`        \u21B3 pairs budget ${d.minPairs}`+(d.configured?" (metadata.autoEvolvePairs)":" (default)")+` \xB7 writes ${l.autoEvolveApply===!0||l.autoEvolveApply==="true"?"ENABLED":"off (dry run)"}`),d.note&&this.log(`        ! ${d.note}`);try{let t=await y(`/api/eval-targets/${s.id}/promotions?limit=1`),a=t.promotions[0];if(a){let n=a.gates?.cost,f=n?.modelCalls?`${n.modelCalls} model calls, ${n.usablePairs??"?"} usable pair(s)`+(n.mde!=null?`, resolved +/-${ee(n.mde)}`:"")+(n.wallMs?`, ${(n.wallMs/6e4).toFixed(0)} min`:""):"cost not recorded (decision predates cost tracking)";this.log(`        \u21B3 last run ${new Date(a.createdAt).toISOString().slice(0,10)}: ${a.decision}${a.applied?" (APPLIED)":""} \u2014 ${f}`)}let o=t.counts??{};this.log(`        \u21B3 lifetime: ${o.promote??0} promote / ${o.revert??0} revert / ${o.refuse??0} refuse`)}catch{}}this.log("        \u21B3 pause the spend: metadata.autoEvolvePaused=true (keeps the opt-in)")}}};function A(h){if(h===null)return"never";if(h<90)return`${Math.round(h)}s`;let r=Math.round(h/60);return r<90?`${r}m`:`${Math.round(r/60)}h`}S(A,"formatAge");export{I as default};
+import{c as q,d as K}from"../chunk-TYFEWLYZ.js";import{b as L,c as W}from"../chunk-V5U3E4BZ.js";import{a as J}from"../chunk-2TJ7Q2OU.js";import{e as O,f as _,g as M}from"../chunk-QWCEZEDO.js";import"../chunk-KKDHD4XA.js";import"../chunk-SZ62CP55.js";import{d as oe,e as ne}from"../chunk-DYET4EYU.js";import"../chunk-M7QSC64D.js";import"../chunk-7LFCZI64.js";import"../chunk-CMDOLJ23.js";import{i as z,j as Q,k as Z,m as ee,n as te}from"../chunk-TWPD3ICV.js";import"../chunk-G7V4ZZZ3.js";import"../chunk-LJE5J3EW.js";import"../chunk-AHLVY52I.js";import"../chunk-YXMAGC53.js";import"../chunk-T7HZZ4DJ.js";import"../chunk-XCH4HPEZ.js";import"../chunk-PPAHIVG5.js";import"../chunk-CT7FWSIU.js";import"../chunk-5MXZJWQS.js";import"../chunk-N7ODX6PO.js";import{a as V}from"../chunk-OZQNSELV.js";import"../chunk-HFGO5762.js";import"../chunk-VS5ROGDP.js";import"../chunk-FADINV4H.js";import"../chunk-SWSOEYC7.js";import"../chunk-U2NT5OSQ.js";import"../chunk-KGURGIZ2.js";import"../chunk-GYNZ3OZF.js";import"../chunk-UPFH72PF.js";import"../chunk-FG343JM3.js";import"../chunk-CDY7TQFZ.js";import"../chunk-HT6R6KR3.js";import"../chunk-4GOXEEXZ.js";import"../chunk-JP6ULOEX.js";import"../chunk-JNLDISBA.js";import{d as F}from"../chunk-GPBN4ROE.js";import"../chunk-6XTLP2G7.js";import"../chunk-PJ3WHQ67.js";import{a as X,b as D,c as Y,g as j,h as G}from"../chunk-QI3FJKGK.js";import{a as B}from"../chunk-WUCACPGC.js";import"../chunk-IZTIYR5Q.js";import"../chunk-WL5ZWQAS.js";import"../chunk-KEVDLLHT.js";import"../chunk-DX6IPDED.js";import"../chunk-3AV3GTMP.js";import{a as U,e as y}from"../chunk-GIXHG5FK.js";import"../chunk-X6CF5KTW.js";import"../chunk-XQYAVSUQ.js";import{a as N,g as v,l as x}from"../chunk-PZNWH4VQ.js";import{a as S}from"../chunk-ZLLER3HP.js";import{spawnSync as le}from"node:child_process";import{existsSync as ce,readFileSync as ue}from"node:fs";import{join as se}from"node:path";var I=class h extends B{static{S(this,"Diagnostics")}static description='Diagnose capture / proxy / connectivity problems (e.g. a client-side "502 Bad Gateway")';static examples=["<%= config.bin %> diagnostics"];baseUrl(){return process.env.RULEMETRIC_API_URL??N().apiUrl??U}async checkDirect(){let r=process.env.HTTPS_PROXY,i=process.env.HTTP_PROXY;delete process.env.HTTPS_PROXY,delete process.env.HTTP_PROXY;let s=new AbortController,l=setTimeout(()=>s.abort(),8e3);try{return(await globalThis.fetch(`${this.baseUrl()}/health`,{signal:s.signal})).ok}catch{return!1}finally{clearTimeout(l),r!==void 0&&(process.env.HTTPS_PROXY=r),i!==void 0&&(process.env.HTTP_PROXY=i)}}configuredProxy(){if(process.env.HTTPS_PROXY)return process.env.HTTPS_PROXY;for(let r of["settings.local.json","settings.json"]){let i=se(process.cwd(),".claude",r);if(ce(i))try{let s=JSON.parse(ue(i,"utf8"));if(s.env?.HTTPS_PROXY)return s.env.HTTPS_PROXY}catch{}}return null}checkThroughProxy(r){let i=le("curl",["-sS","-x",r,"--max-time","8","-o","/dev/null","-w","%{http_code}",`${this.baseUrl()}/health`],{encoding:"utf8",stdio:["ignore","pipe","ignore"]});return i.error||typeof i.stdout!="string"?null:i.stdout.trim()==="200"}async run(){await this.parse(h);let r=process.version,i=this.configuredProxy(),s=i!=null,[l,m,d]=await Promise.all([this.checkDirect(),M(O),M(_)]),p=s&&i?this.checkThroughProxy(i):null,t=S((e,c,T)=>{let u=c===null?"\u2013":c?"[OK]":"[!!]";this.log(`  ${u} ${e.padEnd(22)} ${T}`)},"line");this.log("RuleMetric diagnostics"),this.log(""),t("Node",!L(r),`${r}${L(r)?" (< 20.11 \u2014 breaks the bundled proxy)":""}`),t("API (direct)",l,`${this.baseUrl()}/health ${l?"reachable":"UNREACHABLE"}`),t("Authenticated",v(),v()?"yes":"no \u2014 run `rulemetric auth login`"),t("Proxy configured",s?!0:null,s?i:"no HTTPS_PROXY set"),s&&(t("Gateway :"+O,m,m?"listening":"NOT listening"),t("mitmproxy :"+_,d,d?"listening":"not listening (gateway should route direct)"),t("API (via proxy)",p,p===null?"inconclusive (curl unavailable)":p?"reachable":"FAILED \u2014 502/blocked through your proxy"));let a=F();t("claude CLI",a.claudeBin!=null,a.claudeBin??"not found \u2014 LLM features (grading, distill) will no-op"),t("codex CLI",a.codexBin!=null?!0:null,a.codexBin?`${a.codexBin}${a.codexConfiguredModel?` (model ${a.codexConfiguredModel})`:""}`:"not found \u2014 optional; enables the cross-vendor reviewer (evals validate-judge)");let o=null;if(v()&&l)try{o=await y("/api/workers/health")}catch{o=null}if(o){let e=o.worker;t("Worker",!!e.active,e.present?e.active?`online (last heartbeat ${e.ageSeconds??"?"}s ago${e.clientVersion?`, v${e.clientVersion}`:""})`:`NOT RUNNING \u2014 last heartbeat ${A(e.ageSeconds)} ago`:"never connected \u2014 no worker has ever reported in");let c=o.worker.engines;if(c!=null){let u=[c.claude&&"claude",c.codex&&"codex"].filter(Boolean);t("LLM engine",u.length>0,u.length>0?`${u.join(" + ")} available to the worker`:"NONE \u2014 the worker found no `claude` and no `codex`. LLM jobs will complete with empty output rather than failing.")}let T=o.pending.reduce((u,E)=>u+E.count,0);if(T>0){let u=o.verdict.stuck.reduce(($,k)=>$+k.count,0),E=o.verdict.unclaimable.reduce(($,k)=>$+k.count,0),R=(o.verdict.engineBlocked??[]).reduce(($,k)=>$+k.count,0);t("Job queue",u===0&&R===0,`${T} pending`+(u>0?` \u2014 ${u} waiting on this worker`:"")+(R>0?` \u2014 ${R} need a local LLM engine`:"")+(E>0?` \u2014 ${E} need a worker with DATABASE_URL`:""))}else t("Job queue",!0,"empty")}else v()&&l&&t("Worker",null,"could not read /workers/health (older API?)");let n=o?.worker.machine;if(n){let e=n.suspends+n.stalls,c=n.lastPause;t("Worker pauses",e===0?!0:null,e===0?`none since ${n.since} \u2014 no sleep or stall seen by the worker`:`${n.suspends} suspend${n.suspends===1?"":"s"}`+(n.stalls>0?` + ${n.stalls} event-loop stall${n.stalls===1?"":"s"}`:"")+` since ${n.since}, ${A(Math.round(n.pausedMs/1e3))} paused total`+(c?`; last: ${c.kind} of ${A(Math.round(c.wallGapMs/1e3))} ending ${c.detectedAt}`:"")+" \u2014 jobs in flight during a suspend lose their lease and are redone on wake")}let f=await q();if(f){t("Machine sleep (pmset)",f.sleeps===0?!0:null,K(f));for(let e of f.recent.slice(-3)){let c=e.seconds!=null?` (${A(e.seconds)})`:"";this.log(`        \u21B3 ${e.at.toISOString()} ${e.kind}${c}: ${e.reason.slice(0,80)}`)}}let C=V(process.cwd()),g=await D({configDir:x()}),re=await Y({configDir:x()}),w=` \u2014 ${G(re)}`,b=J(process.cwd()),H=[b.user&&`user hooks (${b.user})`,b.workspace&&`workspace hooks (${b.workspace})`].filter(Boolean),ie={claude_code:`session hooks (+ proxy for instruction linking)${w}`,codex:`worker transcript sync of ~/.codex/sessions \u2014 ${j(g)}`,cursor:`user hooks (~/.cursor/hooks.json)${w}`,copilot_agent:`workspace hooks${w}`,antigravity:H.length>0?`${H.join(" + ")}${w}`:"NOT CAPTURED \u2014 no rulemetric block in ~/.gemini/config/hooks.json or .agents/hooks.json; run `rulemetric hooks install` (user hooks are on by default)"};this.log(""),this.log("  Capture coverage (agents found on this machine):");for(let e of C)this.log(`    \u2022 ${e.padEnd(15)} ${ie[e]??"captured"}`);if(C.includes("codex"))if(g.error)this.log(`    ! ${g.error}`);else if(g.mode==="allowlist")for(let e of g.roots)this.log(`        \u21B3 ${e}`);else g.mode==="all"&&this.log(`        \u21B3 scope it with ${se(x(),X)} ({"codex":{"roots":["/path/to/repo"]}}) or RULEMETRIC_CODEX_CAPTURE=off`);o&&!o.worker.active&&C.includes("codex")&&g.mode!=="off"&&this.log("    ! Codex capture runs INSIDE the worker \u2014 it is stopped while the worker is down."),v()&&l&&await this.reportEvolutionSpend();let P=W({nodeVersion:r,directApiOk:l,httpsProxySet:s,gatewayAlive:m,mitmproxyAlive:d,proxyPathOk:p});this.log("");let ae=P.severity==="ok"?"[OK]":P.severity==="warn"?"[!!]":"[XX]";this.log(`${ae} ${P.headline}`);for(let e of P.hints)this.log(`   \u2022 ${e}`);if(o&&o.verdict.severity!=="ok"){let e=o.verdict.severity==="warn"?"[!!]":"[XX]";this.log(""),this.log(`${e} ${o.verdict.headline}`);for(let c of o.verdict.hints)this.log(`   \u2022 ${c}`)}}async reportEvolutionSpend(){let r;try{r=await y("/api/eval-targets?type=instruction&limit=200")}catch{return}let i=r.filter(s=>{let l=s.metadata??{};return l.autoEvolveEnabled===!0||l.autoEvolveEnabled==="true"});if(i.length!==0){this.log(""),this.log("  Nightly instruction evolution (cron_instruction_evolution, 07:30 UTC):");for(let s of i){let l=s.metadata??{},m=Q(l),d=z(l);if(m){this.log(`    \u2022 ${s.name} \u2014 PAUSED (metadata.autoEvolvePaused). Opt-in preserved.`),this.log("        \u21B3 resume: set autoEvolvePaused=false");continue}let p=null;try{p=(await y(`/api/eval-targets/${s.id}/evals`)).filter(a=>!a.retiredAt).length}catch{}if(this.log(`    \u2022 ${s.name}`),p!=null){let t=ne(p,d.minPairs,oe(d,p));if(!t.feasible)this.log(`        \u21B3 cannot reach a verdict \u2014 ${t.reason}`);else{let a=Z({liveCases:p,passes:t.passes});this.log(`        \u21B3 each night: ${te(a)}`)}}this.log(`        \u21B3 pairs budget ${d.minPairs}`+(d.configured?" (metadata.autoEvolvePairs)":" (default)")+` \xB7 writes ${l.autoEvolveApply===!0||l.autoEvolveApply==="true"?"ENABLED":"off (dry run)"}`),d.note&&this.log(`        ! ${d.note}`);try{let t=await y(`/api/eval-targets/${s.id}/promotions?limit=1`),a=t.promotions[0];if(a){let n=a.gates?.cost,f=n?.modelCalls?`${n.modelCalls} model calls, ${n.usablePairs??"?"} usable pair(s)`+(n.mde!=null?`, resolved +/-${ee(n.mde)}`:"")+(n.wallMs?`, ${(n.wallMs/6e4).toFixed(0)} min`:""):"cost not recorded (decision predates cost tracking)";this.log(`        \u21B3 last run ${new Date(a.createdAt).toISOString().slice(0,10)}: ${a.decision}${a.applied?" (APPLIED)":""} \u2014 ${f}`)}let o=t.counts??{};this.log(`        \u21B3 lifetime: ${o.promote??0} promote / ${o.revert??0} revert / ${o.refuse??0} refuse`)}catch{}}this.log("        \u21B3 pause the spend: metadata.autoEvolvePaused=true (keeps the opt-in)")}}};function A(h){if(h===null)return"never";if(h<90)return`${Math.round(h)}s`;let r=Math.round(h/60);return r<90?`${r}m`:`${Math.round(r/60)}h`}S(A,"formatAge");export{I as default};

package/dist/commands/evals/agent.js

@@ -1,3 +1,3 @@
-import{a as G,b as V,c as j,d as q}from"../../chunk-J4W6GNCF.js";import{b as M}from"../../chunk-GH4EZK2O.js";import{c as g}from"../../chunk-Q776YUBD.js";import{e as D}from"../../chunk-H33XYL7O.js";import"../../chunk-XXPWPKTD.js";import"../../chunk-5TM7AOH2.js";import"../../chunk-IEQDLURA.js";import"../../chunk-KINIFUZQ.js";import"../../chunk-ZYGBABMZ.js";import"../../chunk-HM63AB2C.js";import"../../chunk-PXYK4MBX.js";import"../../chunk-ZTQUGAZG.js";import"../../chunk-XZW5ZHFC.js";import"../../chunk-F5PJN2EY.js";import"../../chunk-XIOYV5A7.js";import"../../chunk-5VNEHPGC.js";import"../../chunk-5PMTNNA6.js";import"../../chunk-ISFPPDDB.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-H4QPAEEA.js";import"../../chunk-J6RQ5AH2.js";import"../../chunk-PZS55PW2.js";import"../../chunk-QPAOX7P2.js";import"../../chunk-3ZHVP3KE.js";import"../../chunk-QENSCK77.js";import"../../chunk-HOQPVOMF.js";import"../../chunk-VCUW2NAP.js";import"../../chunk-FZMO7XSN.js";import"../../chunk-MOJY6S5J.js";import"../../chunk-CYCH2EQC.js";import"../../chunk-VXNICAVG.js";import"../../chunk-L562Y4RN.js";import"../../chunk-CZBO3A6B.js";import"../../chunk-CVPAKEEZ.js";import"../../chunk-GLBWNB2N.js";import"../../chunk-CQD2R3BA.js";import"../../chunk-3UKZJAFS.js";import"../../chunk-77FH6UZP.js";import"../../chunk-KE5K3F7J.js";import"../../chunk-PIVEJTNF.js";import"../../chunk-7MTYUTUN.js";import"../../chunk-FS3TAQAE.js";import"../../chunk-GIRPEMTW.js";import"../../chunk-SIGZEQ3K.js";import"../../chunk-DQYYHX6K.js";import"../../chunk-RE7RAZQ5.js";import"../../chunk-6T6YFPBW.js";import"../../chunk-ISYU5BMU.js";import"../../chunk-DANMUHNW.js";import"../../chunk-T75KAEJO.js";import"../../chunk-FXQS7AEZ.js";import"../../chunk-LJRE3LZ2.js";import"../../chunk-RR7V5CK5.js";import"../../chunk-WIKCYUZC.js";import"../../chunk-DXNUCWUX.js";import"../../chunk-6AAAHV2B.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-MHVHGKDK.js";import"../../chunk-RBSEE2Z6.js";import"../../chunk-KOYHJFB3.js";import"../../chunk-3NWJ7CZ2.js";import"../../chunk-M7QSC64D.js";import"../../chunk-6TGFU6RU.js";import"../../chunk-LB24UKHZ.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-O4K4M5G6.js";import"../../chunk-AHLVY52I.js";import"../../chunk-RSVNQ2XI.js";import"../../chunk-ZK3ACE4T.js";import"../../chunk-7BCZDVIU.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-T47SFA5X.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-DVU5PHPM.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-VBXL7XO7.js";import"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{b as A,d as C,g as S,i as E,j as L,k as $}from"../../chunk-QI3FJKGK.js";import{a as O}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as m}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{e as I,l as p}from"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as de,context as me}from"@opentelemetry/api";import{NodeSDK as W}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as Y}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as J,ConsoleSpanExporter as z,SimpleSpanProcessor as Q}from"@opentelemetry/sdk-trace-node";import{Resource as Z}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as ee,ATTR_SERVICE_VERSION as te}from"@opentelemetry/semantic-conventions";function F(r){if(!r)return{};let n={};for(let o of r.split(",")){let s=o.indexOf("=");if(s===-1)continue;let i=o.slice(0,s).trim(),c=o.slice(s+1).trim();i&&(n[i]=c)}return n}e(F,"parseOtlpHeaders");var u=null;function T(){return process.env.TELEMETRY_ENABLED!=="false"}e(T,"isTelemetryEnabled");function x(r){if(!T()||u!==null)return"off";let n=new Z({[ee]:r.serviceName,[te]:r.serviceVersion??"0.1.0"}),o=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(o){let c=F(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new Y({url:`${o}/v1/traces`,headers:c});i=new J(l),s="otlp"}else if(process.env.OTEL_SPAN_EXPORT==="none")i=void 0,s="silent";else{let c=new z;i=new Q(c),s="console"}return u=new W({resource:n,spanProcessors:i?[i]:[]}),u.start(),s}e(x,"initTelemetry");async function h(){u&&(await u.shutdown(),u=null)}e(h,"shutdownTelemetry");import{SpanKind as Pe,SpanStatusCode as Re}from"@opentelemetry/api";import{trace as Ie,context as Oe}from"@opentelemetry/api";import*as f from"@sentry/node";var v=!1;function w(r){if(v)return!1;let n=process.env.SENTRY_DSN;return n?(f.init({dsn:n,serverName:r.serviceName,environment:process.env.NODE_ENV??"development",tracesSampleRate:0}),v=!0,!0):!1}e(w,"initSentry");async function y(r){v&&await f.flush(r)}e(y,"flushSentry");import{existsSync as ne,mkdirSync as oe,readFileSync as k,unlinkSync as X,writeFileSync as U}from"node:fs";import{dirname as se,join as R}from"node:path";import{execFileSync as ie}from"node:child_process";var a=R(p(),"evals-agent.lock");function ce(r){try{return process.kill(r,0),!0}catch(n){return n.code==="EPERM"}}e(ce,"pidAlive");function ae(r){if(!ce(r))return!1;try{let n=ie("ps",["-p",String(r),"-o","command="],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return/run\.js\s+evals\s+agent|evals\s+agent/.test(n)}catch{return!0}}e(ae,"isLiveAgent");function le(){try{oe(se(a),{recursive:!0})}catch{}for(let r=0;r<2;r++)try{return U(a,String(process.pid),{flag:"wx"}),{ok:!0}}catch(n){if(n.code!=="EEXIST")throw n;let o=0;try{o=Number(k(a,"utf-8").trim())}catch{}if(o&&o!==process.pid&&ae(o))return{ok:!1,holder:o};try{X(a)}catch{}}try{U(a,String(process.pid))}catch{}return{ok:!0}}e(le,"acquireAgentLock");function P(){try{ne(a)&&Number(k(a,"utf-8").trim())===process.pid&&X(a)}catch{}}e(P,"releaseAgentLock");var N=class r extends O{static{e(this,"EvalAgentCommand")}static description="The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.";static examples=["<%= config.bin %> evals agent"];async run(){await this.parse(r),this.requireAuth();let n=le();if(!n.ok){this.log(`Another rulemetric evals agent is already running (pid ${n.holder}). Refusing to start a second instance.`);return}process.env.PG_APPLICATION_NAME||="rulemetric-agent",w({serviceName:"rulemetric-agent"})&&this.log("[sentry] error capture enabled");let s=x({serviceName:"rulemetric-agent"});s==="otlp"?this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`):s==="console"?this.log("[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)"):this.log("[otel] disabled (TELEMETRY_ENABLED=false)"),!I()&&!process.env.RULEMETRIC_API_KEY&&this.error("No auth token found. Run `rulemetric auth login` first."),this.log("Starting thin agent (Phase 4b parallel-run)..."),this.log(`Long-polling /api/work/next. Press Ctrl+C to stop.
+import{a as G,b as V,c as j,d as q}from"../../chunk-J4W6GNCF.js";import{b as M}from"../../chunk-GH4EZK2O.js";import{c as g}from"../../chunk-V576KKH6.js";import{e as D}from"../../chunk-N2PP2A7M.js";import"../../chunk-XXPWPKTD.js";import"../../chunk-5TM7AOH2.js";import"../../chunk-IEQDLURA.js";import"../../chunk-KINIFUZQ.js";import"../../chunk-5PMTNNA6.js";import"../../chunk-ZYGBABMZ.js";import"../../chunk-IFIWDL7G.js";import"../../chunk-FZMO7XSN.js";import"../../chunk-LLXC5SRC.js";import"../../chunk-4DA6L7KL.js";import"../../chunk-JU5ZNXPK.js";import"../../chunk-WQFR2LOY.js";import"../../chunk-Q7IDLOSP.js";import"../../chunk-XIOYV5A7.js";import"../../chunk-ISFPPDDB.js";import"../../chunk-5RBLIVBF.js";import"../../chunk-JSGXFD42.js";import"../../chunk-CZBO3A6B.js";import"../../chunk-JUBXWTJA.js";import"../../chunk-W6MMJRFZ.js";import"../../chunk-PZS55PW2.js";import"../../chunk-QPAOX7P2.js";import"../../chunk-3ZHVP3KE.js";import"../../chunk-CIKQXFQ2.js";import"../../chunk-HOQPVOMF.js";import"../../chunk-5Z6XXOJL.js";import"../../chunk-MOJY6S5J.js";import"../../chunk-CYCH2EQC.js";import"../../chunk-CVPAKEEZ.js";import"../../chunk-VUNJGZPP.js";import"../../chunk-L562Y4RN.js";import"../../chunk-GLBWNB2N.js";import"../../chunk-CQD2R3BA.js";import"../../chunk-3UKZJAFS.js";import"../../chunk-77FH6UZP.js";import"../../chunk-KE5K3F7J.js";import"../../chunk-PIVEJTNF.js";import"../../chunk-KR36ZDGN.js";import"../../chunk-JOXTWLLK.js";import"../../chunk-GIRPEMTW.js";import"../../chunk-SIGZEQ3K.js";import"../../chunk-DQYYHX6K.js";import"../../chunk-RE7RAZQ5.js";import"../../chunk-6T6YFPBW.js";import"../../chunk-ISYU5BMU.js";import"../../chunk-DANMUHNW.js";import"../../chunk-T75KAEJO.js";import"../../chunk-FXQS7AEZ.js";import"../../chunk-LJRE3LZ2.js";import"../../chunk-EJY5GK6A.js";import"../../chunk-74KE4VOX.js";import"../../chunk-HXOQUYG3.js";import"../../chunk-3H44B6PB.js";import"../../chunk-6UEFCMR2.js";import"../../chunk-MODXEGCU.js";import"../../chunk-MHVHGKDK.js";import"../../chunk-RBSEE2Z6.js";import"../../chunk-PDXHRF4E.js";import"../../chunk-DYET4EYU.js";import"../../chunk-M7QSC64D.js";import"../../chunk-7LFCZI64.js";import"../../chunk-CMDOLJ23.js";import"../../chunk-TWPD3ICV.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-LJE5J3EW.js";import"../../chunk-AHLVY52I.js";import"../../chunk-YXMAGC53.js";import"../../chunk-T7HZZ4DJ.js";import"../../chunk-XCH4HPEZ.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-OZQNSELV.js";import"../../chunk-HFGO5762.js";import"../../chunk-VS5ROGDP.js";import"../../chunk-FADINV4H.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-HT6R6KR3.js";import"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{b as A,d as C,g as S,i as E,j as L,k as $}from"../../chunk-QI3FJKGK.js";import{a as O}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{f as m}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import{e as I,l as p}from"../../chunk-PZNWH4VQ.js";import{a as e}from"../../chunk-ZLLER3HP.js";import{trace as de,context as me}from"@opentelemetry/api";import{NodeSDK as W}from"@opentelemetry/sdk-node";import{OTLPTraceExporter as Y}from"@opentelemetry/exporter-trace-otlp-http";import{BatchSpanProcessor as J,ConsoleSpanExporter as z,SimpleSpanProcessor as Q}from"@opentelemetry/sdk-trace-node";import{Resource as Z}from"@opentelemetry/resources";import{ATTR_SERVICE_NAME as ee,ATTR_SERVICE_VERSION as te}from"@opentelemetry/semantic-conventions";function F(r){if(!r)return{};let n={};for(let o of r.split(",")){let s=o.indexOf("=");if(s===-1)continue;let i=o.slice(0,s).trim(),c=o.slice(s+1).trim();i&&(n[i]=c)}return n}e(F,"parseOtlpHeaders");var u=null;function T(){return process.env.TELEMETRY_ENABLED!=="false"}e(T,"isTelemetryEnabled");function x(r){if(!T()||u!==null)return"off";let n=new Z({[ee]:r.serviceName,[te]:r.serviceVersion??"0.1.0"}),o=process.env.OTEL_EXPORTER_OTLP_ENDPOINT,s,i;if(o){let c=F(process.env.OTEL_EXPORTER_OTLP_HEADERS),l=new Y({url:`${o}/v1/traces`,headers:c});i=new J(l),s="otlp"}else if(process.env.OTEL_SPAN_EXPORT==="none")i=void 0,s="silent";else{let c=new z;i=new Q(c),s="console"}return u=new W({resource:n,spanProcessors:i?[i]:[]}),u.start(),s}e(x,"initTelemetry");async function h(){u&&(await u.shutdown(),u=null)}e(h,"shutdownTelemetry");import{SpanKind as Pe,SpanStatusCode as Re}from"@opentelemetry/api";import{trace as Ie,context as Oe}from"@opentelemetry/api";import*as f from"@sentry/node";var v=!1;function w(r){if(v)return!1;let n=process.env.SENTRY_DSN;return n?(f.init({dsn:n,serverName:r.serviceName,environment:process.env.NODE_ENV??"development",tracesSampleRate:0}),v=!0,!0):!1}e(w,"initSentry");async function y(r){v&&await f.flush(r)}e(y,"flushSentry");import{existsSync as ne,mkdirSync as oe,readFileSync as k,unlinkSync as X,writeFileSync as U}from"node:fs";import{dirname as se,join as R}from"node:path";import{execFileSync as ie}from"node:child_process";var a=R(p(),"evals-agent.lock");function ce(r){try{return process.kill(r,0),!0}catch(n){return n.code==="EPERM"}}e(ce,"pidAlive");function ae(r){if(!ce(r))return!1;try{let n=ie("ps",["-p",String(r),"-o","command="],{encoding:"utf-8",stdio:["ignore","pipe","ignore"]});return/run\.js\s+evals\s+agent|evals\s+agent/.test(n)}catch{return!0}}e(ae,"isLiveAgent");function le(){try{oe(se(a),{recursive:!0})}catch{}for(let r=0;r<2;r++)try{return U(a,String(process.pid),{flag:"wx"}),{ok:!0}}catch(n){if(n.code!=="EEXIST")throw n;let o=0;try{o=Number(k(a,"utf-8").trim())}catch{}if(o&&o!==process.pid&&ae(o))return{ok:!1,holder:o};try{X(a)}catch{}}try{U(a,String(process.pid))}catch{}return{ok:!0}}e(le,"acquireAgentLock");function P(){try{ne(a)&&Number(k(a,"utf-8").trim())===process.pid&&X(a)}catch{}}e(P,"releaseAgentLock");var N=class r extends O{static{e(this,"EvalAgentCommand")}static description="The worker daemon: long-polls /api/work/next, claims jobs, and runs them locally. Installed as a launchd service by `rulemetric service install`.";static examples=["<%= config.bin %> evals agent"];async run(){await this.parse(r),this.requireAuth();let n=le();if(!n.ok){this.log(`Another rulemetric evals agent is already running (pid ${n.holder}). Refusing to start a second instance.`);return}process.env.PG_APPLICATION_NAME||="rulemetric-agent",w({serviceName:"rulemetric-agent"})&&this.log("[sentry] error capture enabled");let s=x({serviceName:"rulemetric-agent"});s==="otlp"?this.log(`[otel] traces exporting to ${process.env.OTEL_EXPORTER_OTLP_ENDPOINT}`):s==="console"?this.log("[otel] cron spans printing to stdout (set OTEL_EXPORTER_OTLP_ENDPOINT to ship to a backend)"):this.log("[otel] disabled (TELEMETRY_ENABLED=false)"),!I()&&!process.env.RULEMETRIC_API_KEY&&this.error("No auth token found. Run `rulemetric auth login` first."),this.log("Starting thin agent (Phase 4b parallel-run)..."),this.log(`Long-polling /api/work/next. Press Ctrl+C to stop.
 `);let c=new AbortController,l=await A({configDir:p()});if(this.log(`[transcript-sync] Codex capture: ${S(l)}`),l.error)this.log(`[transcript-sync] WARNING: ${l.error}`);else if(l.mode==="allowlist")for(let t of l.roots)this.log(`[transcript-sync]   \u2022 ${t}`);let d=await C({configDir:p()});if(this.log(`[transcript-sync] Pi capture: ${E(d)}`),d.error)this.log(`[transcript-sync] WARNING: ${d.error}`);else if(d.mode==="allowlist")for(let t of d.roots)this.log(`[transcript-sync]   \u2022 ${t}`);let B=new M({postSnapshot:e((t,K)=>m(`/api/sessions/${t}/context-snapshots`,K),"postSnapshot"),linkInstructions:e(t=>m("/api/sessions/link-instructions",{sessionId:t}),"linkInstructions"),log:e(t=>this.log(`[transcript-sync] ${t}`),"log")}),H=new G({providers:[V(void 0,void 0,L({configDir:p(),onChange:e(t=>this.log(`[transcript-sync] Codex capture changed: ${S(t)}`),"onChange")})),j(void 0,void 0,$({configDir:p(),onChange:e(t=>this.log(`[transcript-sync] Pi capture changed: ${E(t)}`),"onChange")}))],statePath:R(p(),"transcript-sync-state.json"),importTranscript:e(t=>m("/api/sessions/import",t),"importTranscript"),emitInstructions:e(t=>B.emit(t),"emitInstructions"),log:e(t=>this.log(`[transcript-sync] ${t}`),"log")});q(H,{signal:c.signal,log:e(t=>this.log(`[transcript-sync] ${t}`),"log")});let _=e(async t=>{this.log(`
 [${t}] shutting down agent...`),c.abort(),await Promise.allSettled([y(2e3),h(),g()]),P(),process.exit(0)},"shutdown");process.on("SIGINT",()=>{_("SIGINT")}),process.on("SIGTERM",()=>{_("SIGTERM")});let b=!1;try{await D({signal:c.signal,log:e(t=>this.log(t),"log"),version:this.config.version,readInstalledVersion:e(()=>{try{return JSON.parse(k(R(this.config.root,"package.json"),"utf-8")).version??null}catch{return null}},"readInstalledVersion"),onVersionDrift:e(()=>{b=!0},"onVersionDrift")}),b&&(this.log("[upgrade] Installed CLI version changed \u2014 restarting worker onto the new code."),await Promise.allSettled([y(2e3),h(),g()]),P(),process.exit(1))}finally{P(),await g()}}};export{N as default};

package/dist/commands/evals/auto-run.js

@@ -1,4 +1,4 @@
-import{a as k}from"../../chunk-RBSEE2Z6.js";import{e as $}from"../../chunk-T47SFA5X.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-DVU5PHPM.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as w}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as A,Flags as r}from"@oclif/core";var m=class y extends w{static{h(this,"EvalAutoRun")}static description="Run evals in a loop until the effect verdict is statistically powered (\xA76f auto-run-to-power)";static examples=["<%= config.bin %> evals auto-run <target-id>            # preview (cost + plan)","<%= config.bin %> evals auto-run <target-id> --yes      # run until powered / budget","<%= config.bin %> evals auto-run <target-id> --yes --max-runs 120 --batch-parallel 4"];static args={targetId:A.string({description:"Eval target ID",required:!0})};static flags={yes:r.boolean({char:"y",description:"Actually run. Without this, only previews the plan + cost."}),"max-runs":r.integer({description:"Hard ceiling on completed runs",default:200}),"batch-parallel":r.integer({description:"Concurrent executions within a batch",default:3}),timeout:r.integer({description:"Timeout per eval in ms",default:12e4}),model:r.string({description:"Model for eval execution"}),"grader-model":r.string({description:"Model for grading"}),"skip-baseline":r.boolean({description:"Skip without_target runs (no delta \u2014 not recommended)",default:!1})};async run(){let{args:C,flags:e}=await this.parse(y);this.requireAuth();let i=C.targetId,c=20,a=await n(`/api/eval-targets/${i}`),l=await n(`/api/eval-targets/${i}/evals`);l.length===0&&this.error(`No active cases on "${a.name}". Generate some (\`evals create\` / the web Generate button) or run \`evals refine-cases\` first.`);let b=l.length*(e["skip-baseline"]?1:2),f=Math.ceil(e["max-runs"]/b);if(this.log(`Auto-run "${a.name}" v${a.version} toward a powered \xB1${c}pp verdict.`),this.log(`Plan: up to ${e["max-runs"]} runs (${l.length} active case(s) \xD7 ${e["skip-baseline"]?1:2}/batch = ${b}/batch, \u2264${f} batches). Each run is a synthetic \`claude -p\` execution + grade.`),this.log('Note: this is a controlled synthetic A/B. For a DEPLOYED instruction with real usage, the Insights observational verdict is cheaper and uses real sessions \u2014 auto-run is for pre-deploy, catalog skills, or targets with no real "without" cohort.'),!e.yes){this.log(`
+import{a as k}from"../../chunk-RBSEE2Z6.js";import{e as $}from"../../chunk-FADINV4H.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as w}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as n}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as A,Flags as r}from"@oclif/core";var m=class y extends w{static{h(this,"EvalAutoRun")}static description="Run evals in a loop until the effect verdict is statistically powered (\xA76f auto-run-to-power)";static examples=["<%= config.bin %> evals auto-run <target-id>            # preview (cost + plan)","<%= config.bin %> evals auto-run <target-id> --yes      # run until powered / budget","<%= config.bin %> evals auto-run <target-id> --yes --max-runs 120 --batch-parallel 4"];static args={targetId:A.string({description:"Eval target ID",required:!0})};static flags={yes:r.boolean({char:"y",description:"Actually run. Without this, only previews the plan + cost."}),"max-runs":r.integer({description:"Hard ceiling on completed runs",default:200}),"batch-parallel":r.integer({description:"Concurrent executions within a batch",default:3}),timeout:r.integer({description:"Timeout per eval in ms",default:12e4}),model:r.string({description:"Model for eval execution"}),"grader-model":r.string({description:"Model for grading"}),"skip-baseline":r.boolean({description:"Skip without_target runs (no delta \u2014 not recommended)",default:!1})};async run(){let{args:C,flags:e}=await this.parse(y);this.requireAuth();let i=C.targetId,c=20,a=await n(`/api/eval-targets/${i}`),l=await n(`/api/eval-targets/${i}/evals`);l.length===0&&this.error(`No active cases on "${a.name}". Generate some (\`evals create\` / the web Generate button) or run \`evals refine-cases\` first.`);let b=l.length*(e["skip-baseline"]?1:2),f=Math.ceil(e["max-runs"]/b);if(this.log(`Auto-run "${a.name}" v${a.version} toward a powered \xB1${c}pp verdict.`),this.log(`Plan: up to ${e["max-runs"]} runs (${l.length} active case(s) \xD7 ${e["skip-baseline"]?1:2}/batch = ${b}/batch, \u2264${f} batches). Each run is a synthetic \`claude -p\` execution + grade.`),this.log('Note: this is a controlled synthetic A/B. For a DEPLOYED instruction with real usage, the Insights observational verdict is cheaper and uses real sessions \u2014 auto-run is for pre-deploy, catalog skills, or targets with no real "without" cohort.'),!e.yes){this.log(`
 Preview only. Re-run with --yes to start.`);return}let t={runsCompleted:0,maxRuns:e["max-runs"],batchesRun:0,maxBatches:f},x={targetId:i,targetVersion:a.version,iteration:1,timeout:e.timeout,model:e.model,graderModel:e["grader-model"],skipGrading:!1,targetType:a.type,targetName:a.name};for(;;){let s;try{s=await n(`/api/eval-targets/${i}/judge-calibration`)}catch(o){this.log(`
 \u25A0 stopping \u2014 could not read the power check: ${o instanceof Error?o.message:String(o)}`);break}let E={powered:s.powerCheck?.powered??!1,mde:s.powerCheck?.mde==null?1/0:s.powerCheck.mde},B=s.caseDiscrimination??{total:0,discriminating:0,nonDiscriminating:0,unpaired:0},p=k(E,B,t,{targetEffectPp:c});if(this.log(`
 ${p.action==="stop"?"\u25A0":"\u25B8"} ${p.reason}`),p.action==="stop")break;let v=await n(`/api/eval-targets/${i}/evals`);if(v.length===0){this.log("\u25A0 no active cases left \u2014 stopping.");break}let{completed:g,failed:D}=await $({cases:v,targetContent:a.content,concurrency:e["batch-parallel"],runNumber:t.batchesRun+1,opts:x,skipBaseline:e["skip-baseline"],onLog:h(o=>this.log(o),"onLog")});if(t.runsCompleted+=g,t.batchesRun+=1,this.log(`Batch ${t.batchesRun}: ${g} completed, ${D} failed \u2014 ${t.runsCompleted}/${t.maxRuns} runs, ${t.batchesRun}/${t.maxBatches} batches.`),g===0){this.log("\u25A0 batch produced 0 completed runs \u2014 the executor appears broken (check the model / `claude -p`). Aborting.");break}}let d=await n(`/api/eval-targets/${i}/judge-calibration`),u=d.powerCheck?.mde,R=u==null||!Number.isFinite(u)?"unknown":`\xB1${Math.round(u*100)}pp`;this.log(`

package/dist/commands/evals/benchmark.js

@@ -1,4 +1,4 @@
-import{b as h,c as f}from"../../chunk-H2FMAJUJ.js";import"../../chunk-VBXL7XO7.js";import{a as m}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as c}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as v}from"@oclif/core";var d=class b extends m{static{u(this,"EvalBenchmarkCommand")}static description="Aggregate eval runs into a benchmark for a target";static examples=["<%= config.bin %> evals benchmark <target-id>","<%= config.bin %> evals benchmark <target-id> --iteration 2"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={iteration:v.integer({description:"Filter runs by iteration",default:1})};async run(){let{args:o,flags:l}=await this.parse(b);this.requireAuth();let g=await c(`/api/eval-targets/${o.targetId}`);this.log(`Aggregating benchmark for "${g.name}" v${g.version}, iteration ${l.iteration}`);let a=await h(o.targetId,l.iteration);a.totalRuns===0&&this.error("No completed runs found. Run evals first with `rulemetric evals run`."),this.log(`
+import{b as h,c as f}from"../../chunk-QFTAU33Y.js";import"../../chunk-HT6R6KR3.js";import{a as m}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as c}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as u}from"../../chunk-ZLLER3HP.js";import{Args as p,Flags as v}from"@oclif/core";var d=class b extends m{static{u(this,"EvalBenchmarkCommand")}static description="Aggregate eval runs into a benchmark for a target";static examples=["<%= config.bin %> evals benchmark <target-id>","<%= config.bin %> evals benchmark <target-id> --iteration 2"];static args={targetId:p.string({description:"Eval target ID",required:!0})};static flags={iteration:v.integer({description:"Filter runs by iteration",default:1})};async run(){let{args:o,flags:l}=await this.parse(b);this.requireAuth();let g=await c(`/api/eval-targets/${o.targetId}`);this.log(`Aggregating benchmark for "${g.name}" v${g.version}, iteration ${l.iteration}`);let a=await h(o.targetId,l.iteration);a.totalRuns===0&&this.error("No completed runs found. Run evals first with `rulemetric evals run`."),this.log(`
 Aggregated ${a.totalRuns} runs
 `);for(let[e,n]of Object.entries(a.runSummary)){this.log(`${e}:`);for(let[r,i]of Object.entries(n)){let s=i;this.log(`  ${r}: mean=${s.mean} stddev=${s.stddev} [${s.min}, ${s.max}]`)}}let t=a.verdict;if(t)if(this.log(`
 Verdict (${t.mode}, baseline=${t.baseline}):`),!t.sufficient)this.log(`  ${t.insufficientReason??"insufficient data"} \u2014 no effect claimed`);else{let e=t.delta>=0?"+":"";this.log(`  pass-rate delta: ${e}${(t.delta*100).toFixed(0)}% (n=${t.nWith} paired)`),this.log(`  95% CI: ${(t.ci.lower*100).toFixed(0)}\u2013${(t.ci.upper*100).toFixed(0)}%`),t.pValue!=null&&this.log(`  McNemar p: ${t.pValue.toFixed(4)}`),this.log(`  significant: ${t.significant?"YES":"no"}`)}if(a.notes.length>0){this.log(`

package/dist/commands/evals/compare.js

@@ -1,4 +1,4 @@
-import{a as $,b as A}from"../../chunk-647M2I2Y.js";import{f as p}from"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as l,f as w}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as B,Flags as b}from"@oclif/core";var c=class I extends f{static{h(this,"EvalCompareCommand")}static description="Blind A/B comparison between two eval target versions";static examples=["<%= config.bin %> evals compare <target-a> <target-b>"];static args={targetA:B.string({description:"First eval target ID",required:!0}),targetB:B.string({description:"Second eval target ID",required:!0})};static flags={model:b.string({description:"Model for execution and comparison"}),timeout:b.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:g,flags:o}=await this.parse(I);this.requireAuth();let[e,r]=await Promise.all([l(`/api/eval-targets/${g.targetA}`),l(`/api/eval-targets/${g.targetB}`)]),m=await l(`/api/eval-targets/${g.targetA}/evals`);m.length===0&&this.error("No evals found for target A. Create some first."),this.log(`Comparing "${e.name}" v${e.version} vs "${r.name}" v${r.version}`),this.log(`Running ${m.length} eval(s)...
+import{a as $,b as A}from"../../chunk-647M2I2Y.js";import{f as p}from"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as l,f as w}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as h}from"../../chunk-ZLLER3HP.js";import{Args as B,Flags as b}from"@oclif/core";var c=class I extends f{static{h(this,"EvalCompareCommand")}static description="Blind A/B comparison between two eval target versions";static examples=["<%= config.bin %> evals compare <target-a> <target-b>"];static args={targetA:B.string({description:"First eval target ID",required:!0}),targetB:B.string({description:"Second eval target ID",required:!0})};static flags={model:b.string({description:"Model for execution and comparison"}),timeout:b.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:g,flags:o}=await this.parse(I);this.requireAuth();let[e,r]=await Promise.all([l(`/api/eval-targets/${g.targetA}`),l(`/api/eval-targets/${g.targetB}`)]),m=await l(`/api/eval-targets/${g.targetA}/evals`);m.length===0&&this.error("No evals found for target A. Create some first."),this.log(`Comparing "${e.name}" v${e.version} vs "${r.name}" v${r.version}`),this.log(`Running ${m.length} eval(s)...
 `);let a=Math.random()>.5,E=a?r.content:e.content,T=a?e.content:r.content,C=a?r.name:e.name,x=a?e.name:r.name,v=[];for(let t of m){this.log(`[${t.name}] Running A...`);let s=await p({prompt:t.prompt,targetContent:E,timeout:o.timeout,model:o.model});this.log(`[${t.name}] Running B...`);let d=await p({prompt:t.prompt,targetContent:T,timeout:o.timeout,model:o.model});(s.error||d.error)&&this.error(`[${t.name}] execution failed (A: ${s.error??"ok"} | B: ${d.error??"ok"}) \u2014 aborting before the blind comparison`),v.push({evalName:t.name,outputA:s.output,outputB:d.output})}this.log(`
 Running blind comparison...`);let N=v.map(t=>`
 ## Eval: ${t.evalName}

package/dist/commands/evals/optimize-description.js

@@ -1,4 +1,4 @@
-import{a as P,b as S}from"../../chunk-Q36QSFSB.js";import{f as b}from"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as N}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as y,f as q,g as w}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as E}from"../../chunk-ZLLER3HP.js";import{Args as Q,Flags as v}from"@oclif/core";var C=process.env.RULEMETRIC_WEB_URL??"https://rulemetric.com",_=30*6e4,R=class k extends N{static{E(this,"EvalOptimizeDescriptionCommand")}static description="Optimize eval target description for better triggering accuracy";static examples=["<%= config.bin %> evals optimize-description <target-id>","<%= config.bin %> evals optimize-description <target-id> --queries 10"];static args={targetId:Q.string({description:"Eval target ID",required:!0})};static flags={"max-iterations":v.integer({description:"Maximum optimization iterations",default:5}),model:v.string({description:"Model for trigger testing and improvement"}),queries:v.integer({description:"Number of trigger queries to generate",default:20}),"runs-per-query":v.integer({description:"Runs per query for reliability",default:3}),"poll-interval":v.integer({description:"Ms between approval polls",default:5e3}),timeout:v.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:i,flags:t}=await this.parse(k);this.requireAuth();let e=await y(`/api/eval-targets/${i.targetId}`);this.log(`Optimizing description for "${e.name}" v${e.version}`),this.log(`Current description: "${e.description}"`),this.log("Scores measure description-gated invocation choice (catalog decision protocol)."),this.log(""),this.log("Generating trigger queries...");let r=await this.generateTriggerQueries(e,t.queries,t),a=await q(`/api/eval-targets/${i.targetId}/trigger-evals`,{queries:r}),g=`${C}/eval-targets/${i.targetId}`;this.log(`
+import{a as P,b as S}from"../../chunk-6QMIDH6X.js";import{f as b}from"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as N}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as y,f as q,g as w}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as E}from"../../chunk-ZLLER3HP.js";import{Args as Q,Flags as v}from"@oclif/core";var C=process.env.RULEMETRIC_WEB_URL??"https://rulemetric.com",_=30*6e4,R=class k extends N{static{E(this,"EvalOptimizeDescriptionCommand")}static description="Optimize eval target description for better triggering accuracy";static examples=["<%= config.bin %> evals optimize-description <target-id>","<%= config.bin %> evals optimize-description <target-id> --queries 10"];static args={targetId:Q.string({description:"Eval target ID",required:!0})};static flags={"max-iterations":v.integer({description:"Maximum optimization iterations",default:5}),model:v.string({description:"Model for trigger testing and improvement"}),queries:v.integer({description:"Number of trigger queries to generate",default:20}),"runs-per-query":v.integer({description:"Runs per query for reliability",default:3}),"poll-interval":v.integer({description:"Ms between approval polls",default:5e3}),timeout:v.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:i,flags:t}=await this.parse(k);this.requireAuth();let e=await y(`/api/eval-targets/${i.targetId}`);this.log(`Optimizing description for "${e.name}" v${e.version}`),this.log(`Current description: "${e.description}"`),this.log("Scores measure description-gated invocation choice (catalog decision protocol)."),this.log(""),this.log("Generating trigger queries...");let r=await this.generateTriggerQueries(e,t.queries,t),a=await q(`/api/eval-targets/${i.targetId}/trigger-evals`,{queries:r}),g=`${C}/eval-targets/${i.targetId}`;this.log(`
 Generated ${r.length} trigger queries.`),this.log("Review and approve them at:"),this.log(`  ${g}`),this.log("  (Triggers tab \u2192 click to review)"),this.log(`
 Waiting for approval...`);let c=Date.now(),d=!1;for(;!d;){Date.now()-c>_&&this.error(`Timed out after ${_/6e4} minutes waiting for approval. Approve the trigger set at ${g} (Triggers tab) and re-run this command.`),await new Promise(h=>setTimeout(h,t["poll-interval"]));let p=(await y(`/api/eval-targets/${i.targetId}/trigger-evals?status=approved`)).find(h=>h.id===a.id);p&&(d=!0,r.length=0,r.push(...p.queries),this.log("Trigger queries approved!"))}let n=r.map((u,p)=>p).sort(()=>Math.random()-.5),s=Math.ceil(n.length*.6),o=n.slice(0,s),I=n.slice(s);await w(`/api/trigger-evals/${a.id}`,{trainIndices:o,testIndices:I});let x=o.map(u=>r[u]),D=I.map(u=>r[u]);this.log(`
 Split: ${x.length} train, ${D.length} test`);let m=e.description,T=m,$=0;for(let u=1;u<=t["max-iterations"];u++){this.log(`

package/dist/commands/evals/optimize.js

@@ -1,4 +1,4 @@
-import{b as P,c as C}from"../../chunk-7H4HJ2RV.js";import{a as V}from"../../chunk-KOYHJFB3.js";import{a as N}from"../../chunk-DVU5PHPM.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-VBXL7XO7.js";import{f as F}from"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as A}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as w,f as m,g as S}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as k}from"../../chunk-ZLLER3HP.js";import{createInterface as _}from"node:readline/promises";import{Args as q,Flags as g}from"@oclif/core";var $=class O extends A{static{k(this,"EvalOptimizeCommand")}static description="Iteratively optimize eval target content using eval results as feedback";static examples=["<%= config.bin %> evals optimize <target-id>","<%= config.bin %> evals optimize <target-id> --auto --max-iterations 3"];static args={targetId:q.string({description:"Eval target ID",required:!0})};static flags={auto:g.boolean({description:"Skip interactive review, apply suggestions automatically",default:!1}),"max-iterations":g.integer({description:"Maximum optimization iterations",default:5}),threshold:g.string({description:"Stop if pass rate reaches this threshold (0-1)",default:"0.95"}),model:g.string({description:"Model for eval execution"}),"analyzer-model":g.string({description:"Model for analysis/suggestions"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:n,flags:e}=await this.parse(O);this.requireAuth();let t=await w(`/api/eval-targets/${n.targetId}`),i=await w(`/api/eval-targets/${n.targetId}/evals`);i.length===0&&this.error("No evals found. Create some first with `rulemetric evals create`.");let s=[...i].sort(()=>Math.random()-.5),c=Math.ceil(s.length*.6),u=s.slice(0,c),d=s.slice(c);this.log(`Optimizing "${t.name}" v${t.version}`),this.log(`Evals: ${u.length} train, ${d.length} test`);let h=parseFloat(e.threshold);this.log(`Config: max ${e["max-iterations"]} iterations, threshold ${(h*100).toFixed(0)}%`),e.auto&&this.log("Mode: auto (no interactive review)"),this.log("");let T=t.version,f=0,x=0;for(let l=1;l<=e["max-iterations"];l++){this.log(`
+import{b as P,c as C}from"../../chunk-AFEJYDUK.js";import{a as V}from"../../chunk-PDXHRF4E.js";import{a as N}from"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-HT6R6KR3.js";import{f as F}from"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as A}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as w,f as m,g as S}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as k}from"../../chunk-ZLLER3HP.js";import{createInterface as _}from"node:readline/promises";import{Args as q,Flags as g}from"@oclif/core";var $=class O extends A{static{k(this,"EvalOptimizeCommand")}static description="Iteratively optimize eval target content using eval results as feedback";static examples=["<%= config.bin %> evals optimize <target-id>","<%= config.bin %> evals optimize <target-id> --auto --max-iterations 3"];static args={targetId:q.string({description:"Eval target ID",required:!0})};static flags={auto:g.boolean({description:"Skip interactive review, apply suggestions automatically",default:!1}),"max-iterations":g.integer({description:"Maximum optimization iterations",default:5}),threshold:g.string({description:"Stop if pass rate reaches this threshold (0-1)",default:"0.95"}),model:g.string({description:"Model for eval execution"}),"analyzer-model":g.string({description:"Model for analysis/suggestions"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:n,flags:e}=await this.parse(O);this.requireAuth();let t=await w(`/api/eval-targets/${n.targetId}`),i=await w(`/api/eval-targets/${n.targetId}/evals`);i.length===0&&this.error("No evals found. Create some first with `rulemetric evals create`.");let s=[...i].sort(()=>Math.random()-.5),c=Math.ceil(s.length*.6),u=s.slice(0,c),d=s.slice(c);this.log(`Optimizing "${t.name}" v${t.version}`),this.log(`Evals: ${u.length} train, ${d.length} test`);let h=parseFloat(e.threshold);this.log(`Config: max ${e["max-iterations"]} iterations, threshold ${(h*100).toFixed(0)}%`),e.auto&&this.log("Mode: auto (no interactive review)"),this.log("");let T=t.version,f=0,x=0;for(let l=1;l<=e["max-iterations"];l++){this.log(`
 === Iteration ${l} ===`),this.log("Running train evals...");let b=await this.runEvalsOnTarget(t,u,e),p=this.computePassRate(b);if(this.log(`Train pass rate: ${(p*100).toFixed(1)}%`),p>=h){this.log(`Train score ${(p*100).toFixed(1)}% >= threshold ${(h*100).toFixed(0)}%. Stopping.`);let o=await this.runEvalsOnTarget(t,d,e),z=this.computePassRate(o);this.log(`Final test score: ${(z*100).toFixed(1)}%`),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,trainScore:p,testScore:z,analysis:{reason:"threshold_reached"},status:"completed"});break}this.log("Analyzing results...");let M=b.map(o=>({evalName:o.evalName,prompt:o.prompt,expectations:o.expectations,grading:o.grading})),a=await V({content:t.content,evalResults:M,model:e["analyzer-model"]??e.model,timeout:e.timeout});this.log(`
 Assessment: ${a.overallAssessment}`),this.log(`Suggestions: ${a.suggestions.length}`);for(let o of a.suggestions)this.log(`  - [${o.section}] ${o.reasoning}`);if(a.proposedContent===t.content){this.log("No changes proposed. Stopping."),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,trainScore:p,analysis:{reason:"no_changes"},status:"completed"});break}if(!e.auto&&!await this.confirm(`
 Apply the proposed changes as a new version? [y/N] `)){this.log("Changes not applied. Stopping."),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,trainScore:p,analysis:{reason:"user_declined",suggestions:a.suggestions},status:"completed"});break}let r=await S(`/api/eval-targets/${t.id}`,{content:a.proposedContent});this.log(`Created v${r.version}`),this.log(`Running test evals on v${t.version} (baseline) and v${r.version}...`);let E=await this.runEvalsOnTarget(t,d,e),R=await this.runEvalsOnTarget(r,d,e),v=this.computePassRate(R),I=this.computePassRate(E);this.log(`Test pass rate: v${t.version} ${(I*100).toFixed(1)}% \u2192 v${r.version} ${(v*100).toFixed(1)}%`);let y=P(E,R);if(y&&this.log(C(y,t.version,r.version)),await m("/api/optimization-runs",{evalTargetId:t.id,type:"content",iteration:l,inputVersion:t.version,outputVersion:r.version,trainScore:p,testScore:v,analysis:{suggestions:a.suggestions,overallAssessment:a.overallAssessment,priorTestScore:I,verdict:y},status:"completed"}),v>f)f=v,T=r.version,x=0;else if(x++,x>=2){this.log("No improvement for 2 consecutive iterations. Stopping.");break}t=r}this.log(`

package/dist/commands/evals/refine-cases.js

@@ -1,4 +1,4 @@
-import{b as w,c as $}from"../../chunk-RR7V5CK5.js";import"../../chunk-WIKCYUZC.js";import{b as v}from"../../chunk-DXNUCWUX.js";import"../../chunk-3NWJ7CZ2.js";import"../../chunk-M7QSC64D.js";import"../../chunk-6TGFU6RU.js";import"../../chunk-LB24UKHZ.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-O4K4M5G6.js";import"../../chunk-AHLVY52I.js";import"../../chunk-RSVNQ2XI.js";import"../../chunk-ZK3ACE4T.js";import"../../chunk-7BCZDVIU.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-T47SFA5X.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-DVU5PHPM.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-VBXL7XO7.js";import"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as d,f as p,h}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Args as P,Flags as y}from"@oclif/core";var c=a(s=>s==null?"n/a":`${Math.round(s*100)}%`,"pct"),g=class s extends f{static{a(this,"EvalRefineCases")}static description="Retire non-discriminating eval cases and regenerate discriminating replacements (\xA76e)";static examples=["<%= config.bin %> evals refine-cases <target-id>            # preview","<%= config.bin %> evals refine-cases <target-id> --yes      # apply: retire + regenerate","<%= config.bin %> evals refine-cases <target-id> --yes --no-generate"];static args={targetId:P.string({description:"Eval target ID",required:!0})};static flags={yes:y.boolean({char:"y",description:"Apply (retire + regenerate). Without this, only previews."}),"no-generate":y.boolean({description:"Retire dead cases but do not regenerate replacements."})};async run(){let{args:b,flags:l}=await this.parse(s);this.requireAuth();let n=b.targetId,o=await d(`/api/eval-targets/${n}`),m=await d(`/api/eval-targets/${n}/judge-calibration`),i=m.caseDiscrimination,t=w(i);if(this.log(`Target "${o.name}" v${o.version} \u2014 judge ${m.judge.currentModel??"(unstamped)"}`),t.refused==="no-discrimination-data"){this.log("No case-discrimination data yet \u2014 run some evals with AND without the target first (`evals run` / the web Run Evals button), then refine.");return}if(this.log(`${i.total} diagnosable case(s): ${i.discriminating} discriminate, ${i.nonDiscriminating} flat/wrong-sign, ${i.unpaired} unpaired.`),t.refused==="all-discriminating"){this.log("Every case discriminates \u2014 nothing to retire. \u{1F389}");return}let C=i.cases.filter(e=>!e.discriminating),u=new Set(t.retire.map(e=>e.evalId));this.log(`
+import{b as w,c as $}from"../../chunk-EJY5GK6A.js";import"../../chunk-74KE4VOX.js";import{b as v}from"../../chunk-HXOQUYG3.js";import"../../chunk-DYET4EYU.js";import"../../chunk-M7QSC64D.js";import"../../chunk-7LFCZI64.js";import"../../chunk-CMDOLJ23.js";import"../../chunk-TWPD3ICV.js";import"../../chunk-G7V4ZZZ3.js";import"../../chunk-LJE5J3EW.js";import"../../chunk-AHLVY52I.js";import"../../chunk-YXMAGC53.js";import"../../chunk-T7HZZ4DJ.js";import"../../chunk-XCH4HPEZ.js";import"../../chunk-PPAHIVG5.js";import"../../chunk-CT7FWSIU.js";import"../../chunk-5MXZJWQS.js";import"../../chunk-N7ODX6PO.js";import"../../chunk-FADINV4H.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-HT6R6KR3.js";import"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as f}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as d,f as p,h}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a}from"../../chunk-ZLLER3HP.js";import{Args as P,Flags as y}from"@oclif/core";var c=a(s=>s==null?"n/a":`${Math.round(s*100)}%`,"pct"),g=class s extends f{static{a(this,"EvalRefineCases")}static description="Retire non-discriminating eval cases and regenerate discriminating replacements (\xA76e)";static examples=["<%= config.bin %> evals refine-cases <target-id>            # preview","<%= config.bin %> evals refine-cases <target-id> --yes      # apply: retire + regenerate","<%= config.bin %> evals refine-cases <target-id> --yes --no-generate"];static args={targetId:P.string({description:"Eval target ID",required:!0})};static flags={yes:y.boolean({char:"y",description:"Apply (retire + regenerate). Without this, only previews."}),"no-generate":y.boolean({description:"Retire dead cases but do not regenerate replacements."})};async run(){let{args:b,flags:l}=await this.parse(s);this.requireAuth();let n=b.targetId,o=await d(`/api/eval-targets/${n}`),m=await d(`/api/eval-targets/${n}/judge-calibration`),i=m.caseDiscrimination,t=w(i);if(this.log(`Target "${o.name}" v${o.version} \u2014 judge ${m.judge.currentModel??"(unstamped)"}`),t.refused==="no-discrimination-data"){this.log("No case-discrimination data yet \u2014 run some evals with AND without the target first (`evals run` / the web Run Evals button), then refine.");return}if(this.log(`${i.total} diagnosable case(s): ${i.discriminating} discriminate, ${i.nonDiscriminating} flat/wrong-sign, ${i.unpaired} unpaired.`),t.refused==="all-discriminating"){this.log("Every case discriminates \u2014 nothing to retire. \u{1F389}");return}let C=i.cases.filter(e=>!e.discriminating),u=new Set(t.retire.map(e=>e.evalId));this.log(`
 Would retire ${t.retire.length} non-discriminating case(s):`);for(let e of C)u.has(e.evalId)&&this.log(`  \u2717 ${e.name} \u2014 ${c(e.withPassRate)} with vs ${c(e.withoutPassRate)} without (\u0394${(e.separation*100).toFixed(0)}pp) \u2014 ${e.note}`);for(let e of i.unpairedCases)u.has(e.evalId)&&this.log(`  \u2717 ${e.name} \u2014 ${c(e.withPassRate)} with vs ${c(e.withoutPassRate)} without \u2014 unpaired`);if(t.replacementsRequired&&this.log(`
 ${t.summary}. Replacements are generated and created BEFORE anything is retired \u2014 a target left under the floor drops out of the nightly evolution rotation with no error anywhere.`),!l.yes){this.log(`
 Preview only. Re-run with --yes to retire ${t.retire.length} case(s)${l["no-generate"]?"":" and generate replacements"}.`);return}let r=await $(t,{generate:a(e=>l["no-generate"]?Promise.resolve([]):v({content:o.content,name:o.name,feedback:e}),"generate"),createCase:a(e=>p(`/api/eval-targets/${n}/evals`,{name:e.name,prompt:e.prompt,expectations:e.expectations}),"createCase"),retireCase:a(async(e,R)=>{await h(`/api/eval-targets/${n}/evals/${e}`,{retire:!0,retiredReason:R})},"retireCase"),log:a(e=>this.log(e),"log")});if(r.skipped==="no-llm-backend"){this.warn(r.summary+" Install the `claude` CLI (or set ANTHROPIC_API_KEY), then run `evals refine-cases` again.");return}if(r.skipped==="no-replacements-generated"){this.warn(r.summary);return}this.log(`

package/dist/commands/evals/run.js

@@ -1,3 +1,3 @@
-import{d as k}from"../../chunk-T47SFA5X.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-DVU5PHPM.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-FBSZ6KCM.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-SG4UDDHB.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as $}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as v,f as h}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as R}from"../../chunk-ZLLER3HP.js";import{Args as I,Flags as g}from"@oclif/core";var b=class _ extends ${static{R(this,"EvalRunCommand")}static description="Run evals against a target, optionally grade results, and post to API";static examples=["<%= config.bin %> evals run <target-id>","<%= config.bin %> evals run <target-id> --parallel 3 --runs-per-eval 2","<%= config.bin %> evals run <target-id> --skip-baseline --skip-grading"];static args={targetId:I.string({description:"Eval target ID",required:!0})};static flags={parallel:g.integer({description:"Number of parallel eval executions",default:1}),"runs-per-eval":g.integer({description:"Number of runs per eval for variance analysis",default:1}),iteration:g.integer({description:"Iteration number for tracking",default:1}),"skip-baseline":g.boolean({description:"Skip baseline (without_target) runs",default:!1}),"skip-grading":g.boolean({description:"Skip LLM grading step",default:!1}),model:g.string({description:"Model for eval execution (default: user configured)"}),"grader-model":g.string({description:"Model for grading (default: same as execution model)"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:c,flags:r}=await this.parse(_);this.requireAuth();let i=await v(`/api/eval-targets/${c.targetId}`),l=await v(`/api/eval-targets/${c.targetId}/evals`);l.length===0&&this.error(`No evals found for target "${i.name}". Create some with \`rulemetric evals create\`.`),this.log(`Running ${l.length} eval(s) for "${i.name}" v${i.version}`),this.log(`Config: ${r["runs-per-eval"]} run(s) per eval, parallel=${r.parallel}, iteration=${r.iteration}`),r["skip-baseline"]&&this.log("Skipping baseline runs"),r["skip-grading"]&&this.log("Skipping grading"),this.log("");let m=[{name:"with_target",content:i.content}];r["skip-baseline"]||m.push({name:"without_target"});let p=[];for(let a of l)for(let n of m)for(let t=1;t<=r["runs-per-eval"];t++)p.push({eval_:a,configuration:n.name,targetContent:n.content,runNumber:t});let u=[],y=r.parallel;for(let a=0;a<p.length;a+=y){let n=p.slice(a,a+y),t=await Promise.all(n.map(d=>this.executeAndGrade(d,i,r)));u.push(...t)}this.log(""),this.log("Posting results to API...");for(let a of u)await h("/api/eval-runs",a);try{await h(`/api/eval-targets/${c.targetId}/label-state`,{})}catch{}this.log(""),this.log("=== Results Summary ===");for(let a of m){let n=u.filter(e=>e.configuration===a.name),t=n.filter(e=>e.status==="completed"),d=t.filter(e=>Object.keys(e.grading).length>0);if(this.log(`
+import{d as k}from"../../chunk-FADINV4H.js";import"../../chunk-SWSOEYC7.js";import"../../chunk-U2NT5OSQ.js";import"../../chunk-KGURGIZ2.js";import"../../chunk-GYNZ3OZF.js";import"../../chunk-UPFH72PF.js";import"../../chunk-FG343JM3.js";import"../../chunk-CDY7TQFZ.js";import"../../chunk-4GOXEEXZ.js";import"../../chunk-JP6ULOEX.js";import"../../chunk-JNLDISBA.js";import"../../chunk-GPBN4ROE.js";import"../../chunk-6XTLP2G7.js";import"../../chunk-PJ3WHQ67.js";import{a as $}from"../../chunk-WUCACPGC.js";import"../../chunk-IZTIYR5Q.js";import"../../chunk-WL5ZWQAS.js";import"../../chunk-KEVDLLHT.js";import"../../chunk-DX6IPDED.js";import"../../chunk-3AV3GTMP.js";import{e as v,f as h}from"../../chunk-GIXHG5FK.js";import"../../chunk-X6CF5KTW.js";import"../../chunk-XQYAVSUQ.js";import"../../chunk-PZNWH4VQ.js";import{a as R}from"../../chunk-ZLLER3HP.js";import{Args as I,Flags as g}from"@oclif/core";var b=class _ extends ${static{R(this,"EvalRunCommand")}static description="Run evals against a target, optionally grade results, and post to API";static examples=["<%= config.bin %> evals run <target-id>","<%= config.bin %> evals run <target-id> --parallel 3 --runs-per-eval 2","<%= config.bin %> evals run <target-id> --skip-baseline --skip-grading"];static args={targetId:I.string({description:"Eval target ID",required:!0})};static flags={parallel:g.integer({description:"Number of parallel eval executions",default:1}),"runs-per-eval":g.integer({description:"Number of runs per eval for variance analysis",default:1}),iteration:g.integer({description:"Iteration number for tracking",default:1}),"skip-baseline":g.boolean({description:"Skip baseline (without_target) runs",default:!1}),"skip-grading":g.boolean({description:"Skip LLM grading step",default:!1}),model:g.string({description:"Model for eval execution (default: user configured)"}),"grader-model":g.string({description:"Model for grading (default: same as execution model)"}),timeout:g.integer({description:"Timeout per eval in ms",default:12e4})};async run(){let{args:c,flags:r}=await this.parse(_);this.requireAuth();let i=await v(`/api/eval-targets/${c.targetId}`),l=await v(`/api/eval-targets/${c.targetId}/evals`);l.length===0&&this.error(`No evals found for target "${i.name}". Create some with \`rulemetric evals create\`.`),this.log(`Running ${l.length} eval(s) for "${i.name}" v${i.version}`),this.log(`Config: ${r["runs-per-eval"]} run(s) per eval, parallel=${r.parallel}, iteration=${r.iteration}`),r["skip-baseline"]&&this.log("Skipping baseline runs"),r["skip-grading"]&&this.log("Skipping grading"),this.log("");let m=[{name:"with_target",content:i.content}];r["skip-baseline"]||m.push({name:"without_target"});let p=[];for(let a of l)for(let n of m)for(let t=1;t<=r["runs-per-eval"];t++)p.push({eval_:a,configuration:n.name,targetContent:n.content,runNumber:t});let u=[],y=r.parallel;for(let a=0;a<p.length;a+=y){let n=p.slice(a,a+y),t=await Promise.all(n.map(d=>this.executeAndGrade(d,i,r)));u.push(...t)}this.log(""),this.log("Posting results to API...");for(let a of u)await h("/api/eval-runs",a);try{await h(`/api/eval-targets/${c.targetId}/label-state`,{})}catch{}this.log(""),this.log("=== Results Summary ===");for(let a of m){let n=u.filter(e=>e.configuration===a.name),t=n.filter(e=>e.status==="completed"),d=t.filter(e=>Object.keys(e.grading).length>0);if(this.log(`
 ${a.name}:`),this.log(`  Runs: ${t.length}/${n.length} completed`),d.length>0){let e=d.map(o=>o.grading.summary?.pass_rate??0),s=e.reduce((o,f)=>o+f,0)/e.length;this.log(`  Avg pass_rate: ${(s*100).toFixed(1)}%`)}for(let e of n){let s=l.find(x=>x.id===e.evalId)?.name??e.evalId,o=e.grading.summary?.pass_rate,f=o!==void 0?` pass_rate=${(o*100).toFixed(0)}%`:"",w=e.timing.duration_ms?` ${(e.timing.duration_ms/1e3).toFixed(1)}s`:"";this.log(`  [${e.status}] ${s} run#${e.runNumber}${f}${w}`)}}if(!r["skip-baseline"]&&u.length>0){let a=u.filter(t=>t.configuration==="with_target"&&Object.keys(t.grading).length>0),n=u.filter(t=>t.configuration==="without_target"&&Object.keys(t.grading).length>0);if(a.length>0&&n.length>0){let t=a.map(s=>s.grading.summary?.pass_rate??0).reduce((s,o)=>s+o,0)/a.length,d=n.map(s=>s.grading.summary?.pass_rate??0).reduce((s,o)=>s+o,0)/n.length,e=t-d;this.log(`
 Delta: ${e>0?"+":""}${(e*100).toFixed(1)}% (with_target vs without_target)`)}}}async executeAndGrade(c,r,i){return k(c,{targetId:r.id,targetVersion:r.version,iteration:i.iteration,timeout:i.timeout,model:i.model,graderModel:i["grader-model"],skipGrading:i["skip-grading"],targetType:r.type,targetName:r.name},l=>this.log(l))}};export{b as default};