@tangle-network/agent-runtime 0.349.0 → 0.350.2

package/dist/{activation-wBipBkCf.d.ts → activation-DyiHrTlS.d.ts}

@@ -1,11 +1,10 @@
-import { Ja as AgentTurnUsage, Jp as ExecutorConfig, Sd as WorkerSlots, W_ as OtelExportConfig, Ya as CollectedAgentTurn, _m as SharedBoxPlacement, af as BudgetPool, vm as SharedBoxPlacementOptions } from "./types-C8BGOCBn.js";
+import { Fm as SharedBoxPlacement, Im as SharedBoxPlacementOptions, Sf as BudgetPool, fm as ExecutorConfig, lo as AgentTurnUsage, lv as OtelExportConfig, uo as CollectedAgentTurn, zd as WorkerSlots } from "./types-Fowuwdz2.js";
 import { AgentCandidateBundle, AgentCandidateExperiment, AgentImprovementActivation, AgentImprovementActivationIntent, AgentImprovementActivationOutcome, AgentImprovementActivationResult, AgentImprovementActivationTarget, AgentImprovementEvaluation, AgentImprovementProposal, AgentImprovementReview, AgentImprovementReviewDecision, AgentImprovementSurface, AgentProfile, AgentProfileDiff, AgentProfileImprovementChange, AgentProfileImprovementExecutionRef, AgentProfileImprovementExperiment, AgentTrainingReceipt, HarnessType, ReasoningEffort, Sha256Digest } from "@tangle-network/agent-interface";
 import { CostLedgerHandle, CostReceiptInput, MaximumCharge, PairedPromotionDecision, PendingCostCallView, ProposalFinding, RunRecord } from "@tangle-network/agent-eval";
 import { ProposalFinding as ProposalFinding$1 } from "@tangle-network/agent-eval/analyst";
 import { CampaignCellResult, CampaignScenarioIdentity, CampaignStorage, CompareOptimizationMethodsOptions, OptimizationMethod, OptimizationMethodComparison, Scenario, SearchAllocator, SearchCellResult, SearchCellStage, SearchClaim, SearchClaimVerification, SearchHistoryReceipt, SearchLane, SearchModelIdentity, SearchPolicy, SearchSourceRef, SearchSplit, SearchStateView, SurfaceProposer, WorktreeAdapter } from "@tangle-network/agent-eval/campaign";
 import { DispatchContext, MutableSurface as MutableSurface$1, Scenario as Scenario$1, SelfImproveBudget, SelfImproveOptions, SelfImproveProposerResult } from "@tangle-network/agent-eval/contract";
 import { EvaluationClaim, SealedExperiment } from "@tangle-network/agent-eval/experiment";
-//#region src/improvement/improvement-driver.d.ts
 /** The byte-producing path that differs between the cheap
  *  reflective path and the full agentic path. A generator makes (uncommitted)
  *  changes inside `worktreePath`; the driver commits them via the worktree
@@ -48,8 +47,6 @@ interface CandidateGenerator {
     rationale?: string;
   }>;
 }
-//#endregion
-//#region src/improvement/agentic-generator.d.ts
 /**
  * Outcome of verifying a candidate worktree.
  *
@@ -204,13 +201,9 @@ declare function agenticGenerator(opts: AgenticGeneratorOptions): CandidateGener
  *  binary or spawn fault throws (a setup bug, not a failed candidate — no
  *  silent fallback). */
 declare function commandVerifier(command: string, args?: string[], timeoutMs?: number): Verifier;
-//#endregion
-//#region src/improvement/profile-types.d.ts
 type DeepReadonly<T> = T extends ((...args: never[]) => unknown) ? T : T extends readonly (infer TItem)[] ? readonly DeepReadonly<TItem>[] : T extends object ? { readonly [TKey in keyof T]: DeepReadonly<T[TKey]>; } : T;
 /** Complete immutable profile value used during measured execution. */
 type ReadonlyAgentProfile = DeepReadonly<AgentProfile>;
-//#endregion
-//#region src/improvement/search-executor.d.ts
 /** What a lane's slots are. */
 type SearchLaneKind = 'shared-box' | 'dedicated' | 'subscription' | 'router';
 /** An execution lane: slots that share one cost rule, and the profiles they cannot run. */
@@ -312,8 +305,6 @@ interface SearchCellContext {
 type SearchTraceOptions = 'off' | {
   readonly otlp: OtelExportConfig;
 };
-//#endregion
-//#region src/improvement/improve-types.d.ts
 /** The executable agent lever `improve` optimizes — every surface a proposal can name
  * (`AgentImprovementSurface`) except `knowledge`, which the corpus lane owns and `improve`
  * does not produce. Deriving it means every surface `improve` produces can also be reported, which
@@ -694,8 +685,6 @@ interface ImproveCodeResult<TScenario extends Scenario$1, TArtifact> extends Imp
   raw: SelfImproveProposerResult<TScenario, TArtifact>;
 }
 type ImproveResult<TScenario extends Scenario$1, TArtifact> = ImproveMethodResult | ImproveSearchResult | ImproveCodeResult<TScenario, TArtifact>;
-//#endregion
-//#region src/improvement/training.d.ts
 interface ProfileTrainerRequest {
   version: 1;
   invocationId: string;
@@ -791,8 +780,6 @@ interface ControlledTrainingCommand {
 }
 /** Execute one pinned command without a shell, in the runtime-owned job directory. POSIX only. */
 declare function createCommandProfileTrainer(input: ControlledTrainingCommand): ProfileTrainer;
-//#endregion
-//#region src/improvement/search-decision.d.ts
 /** What Runtime decided about a closed search. */
 interface SearchShipDecision {
   decision: 'ship' | 'hold';
@@ -838,8 +825,6 @@ interface SearchImprovementDecision extends SearchShipDecision {
  * a registry promotion, or to re-decide one that an earlier rule held.
  */
 declare function decideSearchImprovement(options: DecideSearchImprovementOptions): Promise<SearchImprovementDecision>;
-//#endregion
-//#region src/improvement/search-method.d.ts
 interface SearchMethodOptions {
   /** Writes each child surface from the parents the policy chose. It reads the
    * train split only (`ProposeContext.train` and `summary`); `history` is empty. */
@@ -879,8 +864,6 @@ interface SearchMethodOptions {
 }
 /** Build Runtime's native search for `improve(profile, { method })`. */
 declare function searchMethod(options: SearchMethodOptions): ImproveSearchMethod;
-//#endregion
-//#region src/improvement/improve.d.ts
 /** Train and serve a checkpoint without implying that it improved held-out quality. */
 declare function improve(profile: ReadonlyAgentProfile, opts: ImproveTrainingOptions): Promise<ImproveTrainingResult>;
 /**
@@ -899,8 +882,6 @@ declare function improve<TScenario extends Scenario$1, TArtifact>(profile: Reado
  * Optimize repository code through Runtime's isolated worktree path.
  */
 declare function improve<TScenario extends Scenario$1, TArtifact>(opts: ImproveCodeRunOptions<TScenario, TArtifact>): Promise<ImproveCodeResult<TScenario, TArtifact>>;
-//#endregion
-//#region src/intelligence/improvement-surfaces.d.ts
 /** Agent improvement surfaces delivered as exact `AgentProfileDiff` replacements. */
 declare const AGENT_IMPROVEMENT_PROFILE_SURFACES: readonly ["prompt", "skills", "tools", "mcp", "hooks", "subagents"];
 type AgentImprovementProfileSurface = (typeof AGENT_IMPROVEMENT_PROFILE_SURFACES)[number];
@@ -944,8 +925,6 @@ declare function agentImprovementTargetProfileDiffs(target: {
  * another, then prove the patch preserves the complete candidate state.
  */
 declare function agentImprovementProfileDiffs(baselineInput: AgentProfile, candidateInput: AgentProfile, options: AgentImprovementTargetProfileDiffOptions): [AgentProfileDiff, ...AgentProfileDiff[]];
-//#endregion
-//#region src/intelligence/activation.d.ts
 interface CreateAgentImprovementActivationResultOptions {
   completedAt: string;
   outcome: AgentImprovementActivationOutcome;
@@ -1079,6 +1058,5 @@ declare function verifyAgentImprovementActivation(input: {
   review: unknown;
   activation: unknown;
 }): AgentImprovementActivation;
-//#endregion
 export { ImproveCodeBaseOptions as $, CandidateGenerator as $t, AgentProfileMeasuredSurface as A, SearchPromotion as At, DecideSearchImprovementOptions as B, sharedBoxLane as Bt, verifyAgentImprovementProposal as C, ImproveSearchOptions as Ct, AgentImprovementActivationTargetIdentity as D, ImprovementCodeCandidate as Dt, AGENT_PROFILE_MEASURED_SURFACES as E, ImproveSurface as Et, isAgentImprovementProfileSurface as F, SearchLaneOptions as Ft, ImproveTrainingOptions as G, AgenticGeneratorOptions as Gt, decideSearchImprovement as H, DeepReadonly as Ht, isAgentProfileMeasuredSurface as I, SearchTraceOptions as It, ProfileTrainerRequest as J, AgenticGeneratorShotReceipt as Jt, ImproveTrainingResult as K, AgenticGeneratorShotDisposition as Kt, improve as L, SharedBoxSearchLane as Lt, agentImprovementProfileSurfaceDigest as M, SearchEnvironmentFault as Mt, agentImprovementProfileSurfaceInput as N, SearchLane$1 as Nt, AgentImprovementProfileSurface as O, ImprovementProfileCandidate as Ot, agentImprovementTargetProfileDiffs as P, SearchLaneKind as Pt, ImproveCandidateValidator as Q, commandVerifier as Qt, SearchMethodOptions as R, dedicatedLane as Rt, verifyAgentImprovementActivationResult as S, ImproveSearchMethod as St, AGENT_IMPROVEMENT_PROFILE_SURFACES as T, ImproveSkillsOptions as Tt, CheckpointServingPort as U, ReadonlyAgentProfile as Ut, SearchImprovementDecision as V, subscriptionLane as Vt, ControlledTrainingCommand as W, AgenticGeneratorExecutorForWorktree as Wt, createCommandProfileTrainer as X, VerifyResult as Xt, TrainingBoundaryResult as Y, Verifier as Yt, ImproveCandidateValidationInput as Z, agenticGenerator as Zt, createAgentImprovementActivationResult as _, ImproveResult as _t, AgentImprovementActivationTransitionInput as a, ImproveLineage as at, reviewAgentImprovementProposal as b, ImproveSearchAgent as bt, CreateAgentImprovementActivationOptions as c, ImproveMethodLineage as ct, ExecuteAgentImprovementActivationInput as d, ImproveMethodSource as dt, ImproveCodeOptions as et, ExecuteAgentImprovementActivationOptions as f, ImproveOptimizationRunOptions as ft, createAgentImprovementActivation as g, ImproveProfileSurface as gt, SealedCandidateActivationTransitionInput as h, ImproveProfileComponents as ht, AgentImprovementActivationTransition as i, ImproveCustomCodeGeneratorOptions as it, agentImprovementProfileDiffs as j, SearchCellContext as jt, AgentImprovementTargetProfileDiffOptions as k, SearchDispatchContext as kt, CreateAgentImprovementActivationResultOptions as l, ImproveMethodOptions as lt, ReviewAgentImprovementInput as m, ImproveProfileAgent as mt, AgentImprovementActivationResultStore as n, ImproveCodeRunOptions as nt, AgentProfileImprovementActivationOperation as o, ImproveMethodContext as ot, ProfileImprovementActivationTransitionInput as p, ImproveOptions as pt, ProfileTrainer as q, AgenticGeneratorShotExecution as qt, AgentImprovementActivationTargetPlan as r, ImproveCost as rt, AgentProfileImprovementActivationTargetPlan as s, ImproveMethodFactory as st, AgentImprovementActivationReconciliation as t, ImproveCodeResult as tt, CreateAgentImprovementProposalOptions as u, ImproveMethodResult as ut, createAgentImprovementProposal as v, ImproveRuntimeCodeGeneratorOptions as vt, verifyAgentImprovementReview as w, ImproveSearchResult as wt, verifyAgentImprovementActivation as x, ImproveSearchIdentity as xt, executeAgentImprovementActivation as y, ImproveScenarioPartitions as yt, searchMethod as z, routerLane as zt };
-//# sourceMappingURL=activation-wBipBkCf.d.ts.map
+//# sourceMappingURL=activation-DyiHrTlS.d.ts.map

package/dist/agent.d.ts

@@ -1,10 +1,9 @@
-import { S as RuntimeStreamEvent, SS as SandboxClient, Zx as AgentRunSpec, bS as OutputAdapter } from "./types-C8BGOCBn.js";
-import { a as ProfileMaterializationContract, c as assertProfileMaterializes, n as AssertProfileMaterializationOptions, o as ProfileMaterializationIssue, r as DefineProfileMaterializationContractOptions, s as ValidateProfileMaterializationOptions, t as AgentProfileMaterializationAxis } from "./profile-materialization-bPm0gGLO.js";
-import { a as createSurfaceImprovementProposer, c as SurfaceValidationIssue, i as SurfaceImprovementEdit, l as renderSurfaceIssues, n as DraftPatchInput, o as AgentSurfaces, r as DraftPatchOutput, s as ResolvedSurface, t as CreateSurfaceImprovementProposerOptions, u as validateSurfaces } from "./improvement-adapter-sIthWkcN.js";
+import { S as RuntimeStreamEvent, VS as SandboxClient, gS as AgentRunSpec, zS as OutputAdapter } from "./types-Fowuwdz2.js";
+import { a as ProfileMaterializationContract, c as assertProfileMaterializes, n as AssertProfileMaterializationOptions, o as ProfileMaterializationIssue, r as DefineProfileMaterializationContractOptions, s as ValidateProfileMaterializationOptions, t as AgentProfileMaterializationAxis } from "./profile-materialization-BRY22Oh8.js";
+import { a as createSurfaceImprovementProposer, c as SurfaceValidationIssue, i as SurfaceImprovementEdit, l as renderSurfaceIssues, n as DraftPatchInput, o as AgentSurfaces, r as DraftPatchOutput, s as ResolvedSurface, t as CreateSurfaceImprovementProposerOptions, u as validateSurfaces } from "./improvement-adapter-D5GRU1eY.js";
 import { AgentProfile, AgentProfileFileMount, AgentProfileMcpServer } from "@tangle-network/agent-interface";
 import { TraceAnalystDefinition } from "@tangle-network/agent-eval";
 import { SandboxEvent } from "@tangle-network/sandbox";
-//#region src/agent/define-agent.d.ts
 /**
  * The full agent manifest. Each agent ships ONE of these.
  *
@@ -166,7 +165,7 @@ interface AgentRunInvocation<TRunOutput> {
  * the agent's real streaming runtime (`runChatTurn` or equivalent) once
  * the eval path consumes the manifest end-to-end.
  */
-declare function unimplementedAgentRun<TRunOutput = unknown>(reason?: string): AgentRunInvocation<TRunOutput>;
+export declare function unimplementedAgentRun<TRunOutput = unknown>(reason?: string): AgentRunInvocation<TRunOutput>;
 interface AgentRunContext {
   /** Substrate-managed trace emitter. */
   emitter: import('@tangle-network/agent-eval').TraceEmitter;
@@ -192,7 +191,7 @@ interface AnalystConfig {
   };
 }
 /** Thrown when `defineAgent` finds a required surface missing on disk. */
-declare class AgentManifestError extends Error {
+export declare class AgentManifestError extends Error {
   readonly agentId: string;
   readonly issues: ReadonlyArray<unknown>;
   constructor(message: string, agentId: string, issues?: ReadonlyArray<unknown>);
@@ -210,9 +209,7 @@ declare class AgentManifestError extends Error {
  * `runtime.act` see the typed shapes via TypeScript inference at
  * the call site.
  */
-declare function defineAgent<TPersona = unknown, TRunOutput = unknown>(manifest: AgentManifest<TPersona, TRunOutput>): AgentManifest<TPersona, TRunOutput>;
-//#endregion
-//#region src/agent/sandbox-act.d.ts
+export declare function defineAgent<TPersona = unknown, TRunOutput = unknown>(manifest: AgentManifest<TPersona, TRunOutput>): AgentManifest<TPersona, TRunOutput>;
 /** Per-persona profile-merge slots applied over the base profile (§1.5: the caller authors the
  *  per-persona profile). Each slot overlays the base; an absent slot leaves the base untouched. */
 interface SandboxActComposeOverrides {
@@ -258,7 +255,6 @@ interface CreateSandboxActOptions<TPersona, TRunOutput> {
  * synchronously with a live `events` iterator and an `output` promise that
  * resolves only after the iterator drains.
  */
-declare function createSandboxAct<TPersona, TRunOutput>(options: CreateSandboxActOptions<TPersona, TRunOutput>): (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput>;
-//#endregion
-export { type AgentManifest, AgentManifestError, type AgentRubric, type AgentRunContext, type AgentRunInvocation, type AgentRuntime, type AgentSurfaces, type AnalystConfig, type AssertProfileMaterializationOptions, type CreateSandboxActOptions, type CreateSurfaceImprovementProposerOptions, type DefineProfileMaterializationContractOptions, type DraftPatchInput, type DraftPatchOutput, type JudgeConfig, type ProfileMaterializationContract, type ProfileMaterializationIssue, type ResolvedSurface, type RubricDimension, type SandboxActComposeOverrides, type SurfaceImprovementEdit, type SurfaceValidationIssue, type ValidateProfileMaterializationOptions, assertProfileMaterializes, createSandboxAct, createSurfaceImprovementProposer, defineAgent, renderSurfaceIssues, unimplementedAgentRun, validateSurfaces };
+export declare function createSandboxAct<TPersona, TRunOutput>(options: CreateSandboxActOptions<TPersona, TRunOutput>): (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput>;
+export { type AgentManifest, type AgentRubric, type AgentRunContext, type AgentRunInvocation, type AgentRuntime, type AgentSurfaces, type AnalystConfig, type AssertProfileMaterializationOptions, type CreateSandboxActOptions, type CreateSurfaceImprovementProposerOptions, type DefineProfileMaterializationContractOptions, type DraftPatchInput, type DraftPatchOutput, type JudgeConfig, type ProfileMaterializationContract, type ProfileMaterializationIssue, type ResolvedSurface, type RubricDimension, type SandboxActComposeOverrides, type SurfaceImprovementEdit, type SurfaceValidationIssue, type ValidateProfileMaterializationOptions, assertProfileMaterializes, createSurfaceImprovementProposer, renderSurfaceIssues, validateSurfaces };
 //# sourceMappingURL=agent.d.ts.map

package/dist/agent.js

@@ -1,10 +1,10 @@
-import { p as sha256 } from "./record-events-D93DkpAx.js";
-import { Lt as createSandboxForSpec, Qi as mapSandboxEvent, Qr as sandboxActProfileMaterialization, Ur as assertProfileMaterialization, Wr as assertProfileMaterializes } from "./supervisor-wS9Ehz_x.js";
-import "./runtime-CsvGCQpl.js";
+import { p as sha256 } from "./record-events-B2GUgYdz.js";
+import { $r as sandboxActProfileMaterialization, Gr as assertProfileMaterializes, Lt as createSandboxForSpec, Wr as assertProfileMaterialization, ea as mapSandboxEvent } from "./supervisor-C1OePYE-.js";
+import "./runtime-6Ec9sHRe.js";
 import { existsSync, readFileSync, statSync } from "node:fs";
 import { isAbsolute, join, relative, resolve, sep } from "node:path";
 import { parseFindingSubject } from "@tangle-network/agent-eval/analyst";
-//#region src/agent/surfaces.ts
+// src/agent/surfaces.ts
 /**
 * `AgentSurfaces` — declarative map of the mutable file/directory paths
 * the self-improvement loop can edit on behalf of an agent.
@@ -124,7 +124,8 @@ function candidatePathsForSubject(subject, surfaces) {
 function safeJoin(root, ...children) {
 	if (children.some((child) => child.includes("\0") || isAbsolute(child))) return null;
 	const rootAbsolute = resolve(root);
-	const escaped = relative(rootAbsolute, resolve(rootAbsolute, ...children));
+	const targetAbsolute = resolve(rootAbsolute, ...children);
+	const escaped = relative(rootAbsolute, targetAbsolute);
 	if (escaped === ".." || escaped.startsWith(`..${sep}`) || isAbsolute(escaped)) return null;
 	return join(root, ...children);
 }
@@ -245,8 +246,7 @@ function renderSurfaceIssues(issues, repoRoot) {
 		"surfaces rather than fabricating a path."
 	].join("\n");
 }
-//#endregion
-//#region src/agent/define-agent.ts
+// src/agent/define-agent.ts
 /**
 * Stub for agents whose `runtime.act` is not yet wired to the substrate's
 * eval path. Preserves the streaming contract (empty event stream + a
@@ -295,8 +295,7 @@ function defineAgent(manifest) {
 	if (manifest.rubric.dimensions.length > 0 && (total < .5 || total > 1.5)) throw new AgentManifestError(`defineAgent(${manifest.id}): rubric dimension weights sum to ${total.toFixed(3)} — should be ~1.0`, manifest.id);
 	return manifest;
 }
-//#endregion
-//#region src/agent/improvement-adapter.ts
+// src/agent/improvement-adapter.ts
 /**
 * Surface improvement proposer — resolves analyst findings into LLM-drafted
 * candidate patches without changing the caller's repository.
@@ -410,8 +409,7 @@ function createSurfaceImprovementProposer(opts) {
 		};
 	} };
 }
-//#endregion
-//#region src/agent/sandbox-act.ts
+// src/agent/sandbox-act.ts
 /**
 * Build an `AgentRuntime.act` implementation backed by a single prod-profile
 * sandbox run. The returned function honours the `act` contract: it returns
@@ -494,7 +492,6 @@ function applyComposeOverrides(base, overrides) {
 		}
 	};
 }
-//#endregion
 export { AgentManifestError, assertProfileMaterializes, createSandboxAct, createSurfaceImprovementProposer, defineAgent, renderSurfaceIssues, unimplementedAgentRun, validateSurfaces };
 
 //# sourceMappingURL=agent.js.map

package/dist/agent.js.map

@@ -1 +1 @@
-{"version":3,"file":"agent.js","names":[],"sources":["../src/agent/surfaces.ts","../src/agent/define-agent.ts","../src/agent/improvement-adapter.ts","../src/agent/sandbox-act.ts"],"sourcesContent":["/**\n * `AgentSurfaces` — declarative map of the mutable file/directory paths\n * the self-improvement loop can edit on behalf of an agent.\n *\n * The substrate uses this map to resolve every parsed `FindingSubject`\n * (from agent-eval) to a real on-disk path. No per-vertical glue;\n * no fabricated paths; no silent `existsSync(...)` skips that hide\n * misconfiguration from the operator.\n *\n * Surfaces are validated at `defineAgent` time — missing paths fail\n * loud with a list of every offender. A surface that's not needed\n * (e.g. an agent with no RAG corpora) is simply omitted; the loop\n * refuses to route those subjects rather than fabricating a target.\n */\n\nimport { existsSync, statSync } from 'node:fs'\nimport { isAbsolute, join, relative, resolve, sep } from 'node:path'\nimport type { FindingSubject } from '@tangle-network/agent-eval'\n\n/**\n * Surface declarations. Every path is repo-relative (or absolute) at\n * `defineAgent` time. At resolution time, paths are joined against the\n * agent's `repoRoot`.\n *\n * `systemPrompt`, `tools`, `personas` are DIRECTORIES; the loop appends\n * `<section>.md`, `<tool>/README.md`, `<persona-id>.yaml` etc.\n * `rubric`, `outputSchema` are SINGLE FILES; the loop edits them in\n * place.\n *\n * `knowledge` is the agent-knowledge root (typically `.agent-knowledge`);\n * `applyKnowledgeWriteBlocks` writes pages relative to it.\n *\n * Optional surfaces (`scaffolding`, `memory`, `rag`, `outputSchema`)\n * can be omitted — the loop will reject findings targeting them with a\n * clear log message instead of fabricating a path.\n */\nexport interface AgentSurfaces {\n  /** Directory containing one markdown file per system-prompt section. */\n  systemPrompt: string\n  /** Directory containing one subdir per tool (`<tool>/README.md`). */\n  tools: string\n  /** Single file (TypeScript module) defining the rubric weights + dimensions. */\n  rubric: string\n  /** Knowledge-base root; typically `.agent-knowledge`. */\n  knowledge: string\n  /** Directory containing one YAML/JSON file per persona. */\n  personas: string\n  /** Optional: directory containing scaffolding rules (precondition checks, retry policies). */\n  scaffolding?: string\n  /** Optional: memory store path (JSONL / SQLite / DB). */\n  memory?: string\n  /** Optional: directory containing RAG corpora (`<corpus>/<doc-id>.md`). */\n  rag?: string\n  /** Optional: single file defining the output schema (Zod / JSON Schema). */\n  outputSchema?: string\n  /** Optional: directory containing Agent Skill packages. */\n  skills?: string\n  /** Optional: directory containing MCP server/tool configuration. */\n  mcp?: string\n  /** Optional: directory containing hook definitions. */\n  hooks?: string\n  /** Optional: directory containing subagent definitions. */\n  subagents?: string\n  /** Optional: directory containing orchestration/workflow policies. */\n  workflows?: string\n  /** Optional: single file containing rollout-policy settings. */\n  rolloutPolicy?: string\n  /** Optional: single canonical AgentProfile file. */\n  agentProfile?: string\n  /** Optional: source root for code findings. */\n  code?: string\n}\n\nexport interface ResolvedSurface {\n  /** Absolute filesystem path the operator can `cat` / `vim`. */\n  absolutePath: string\n  /** Repo-relative path for PR descriptions, diffs, audit logs. */\n  repoRelativePath: string\n  /** Whether the path currently exists on disk. */\n  exists: boolean\n  /** The substrate's intent: edit an existing file or create a new one. */\n  intent: 'edit-existing' | 'create-new'\n}\n\n/**\n * Resolve a parsed `FindingSubject` to the file path the substrate\n * should edit (or create) on disk.\n *\n * Returns `null` when:\n *   - the subject targets a surface the agent didn't declare\n *     (e.g. `rag:*` when `surfaces.rag` is undefined), OR\n *   - the subject is a `cluster` (failure-mode emits these as evidence,\n *     not actionable mutations — they don't route to a file).\n *\n * Returns a `ResolvedSurface` with `intent: 'create-new'` when the\n * subject names a path that doesn't yet exist (e.g. a new wiki page).\n * The caller chooses whether to honour the create — for tightly-managed\n * surfaces like `systemPrompt` it's usually a contract violation\n * (the analyst named a section that doesn't exist); for `knowledge`\n * it's the whole point.\n */\nexport function resolveSubjectPath(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ResolvedSurface | null {\n  const candidates = candidatePathsForSubject(subject, surfaces)\n  if (candidates.length === 0) return null\n\n  // Probe candidates in order, preferring the first one that exists on disk.\n  // Lets the substrate accept both the flat `<section>.md` convention and\n  // the skill-dir `<section>/SKILL.md` convention without forcing one layout.\n  // When none exists, fall back to the first candidate (canonical create-new).\n  for (const rel of candidates) {\n    const abs = isAbsolute(rel) ? rel : join(repoRoot, rel)\n    if (existsSync(abs)) {\n      return { absolutePath: abs, repoRelativePath: rel, exists: true, intent: 'edit-existing' }\n    }\n  }\n  const fallback = candidates[0]!\n  const fallbackAbs = isAbsolute(fallback) ? fallback : join(repoRoot, fallback)\n  return {\n    absolutePath: fallbackAbs,\n    repoRelativePath: fallback,\n    exists: false,\n    intent: 'create-new',\n  }\n}\n\nfunction candidatePathsForSubject(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n): ReadonlyArray<string> {\n  switch (subject.kind) {\n    case 'knowledge.wiki':\n    case 'knowledge.stale':\n      return optionalPath(safeJoin(surfaces.knowledge, `${subject.slug}.md`))\n    case 'knowledge.claim':\n      // Claims land in a per-topic claims directory under the knowledge root.\n      return optionalPath(safeJoin(surfaces.knowledge, 'claims', `${slugify(subject.topic)}.md`))\n    case 'knowledge.raw':\n      return optionalPath(safeJoin(surfaces.knowledge, 'raw', `${subject.sourceId}.md`))\n    case 'system-prompt': {\n      const slug = slugify(subject.section)\n      // Prefer flat layout for create-new (canonical); probe skill-dir layout\n      // in case the existing repo (tax/legal/gtm/creative) uses\n      // `<section>/SKILL.md` already.\n      return [\n        safeJoin(surfaces.systemPrompt, `${slug}.md`),\n        safeJoin(surfaces.systemPrompt, slug, 'SKILL.md'),\n        safeJoin(surfaces.systemPrompt, slug, 'index.md'),\n      ].filter((path): path is string => path !== null)\n    }\n    case 'skill': {\n      if (!surfaces.skills) return []\n      return [\n        safeJoin(surfaces.skills, subject.name, 'SKILL.md'),\n        safeJoin(surfaces.skills, `${subject.name}.md`),\n      ].filter((path): path is string => path !== null)\n    }\n    case 'tool-doc':\n      if (subject.aspect) {\n        return optionalPath(safeJoin(surfaces.tools, subject.tool, `${slugify(subject.aspect)}.md`))\n      }\n      // tool-doc default: `<tool>/README.md`; also probe `<tool>.md` for flat\n      // tool-list repos.\n      return [\n        safeJoin(surfaces.tools, subject.tool, 'README.md'),\n        safeJoin(surfaces.tools, `${subject.tool}.md`),\n      ].filter((path): path is string => path !== null)\n    case 'new-tool':\n      return optionalPath(safeJoin(surfaces.tools, subject.name, 'README.md'))\n    case 'mcp':\n      if (!surfaces.mcp) return []\n      return subject.tool\n        ? optionalPath(safeJoin(surfaces.mcp, subject.server, `${subject.tool}.md`))\n        : [\n            safeJoin(surfaces.mcp, `${subject.server}.json`),\n            safeJoin(surfaces.mcp, subject.server, 'README.md'),\n          ].filter((path): path is string => path !== null)\n    case 'hook':\n      if (!surfaces.hooks) return []\n      return [\n        safeJoin(surfaces.hooks, `${subject.name}.md`),\n        safeJoin(surfaces.hooks, `${subject.name}.json`),\n      ].filter((path): path is string => path !== null)\n    case 'subagent':\n      if (!surfaces.subagents) return []\n      return [\n        safeJoin(surfaces.subagents, `${subject.name}.md`),\n        safeJoin(surfaces.subagents, `${subject.name}.yaml`),\n        safeJoin(surfaces.subagents, `${subject.name}.json`),\n      ].filter((path): path is string => path !== null)\n    case 'workflow':\n      if (!surfaces.workflows) return []\n      return [\n        safeJoin(surfaces.workflows, `${subject.name}.md`),\n        safeJoin(surfaces.workflows, `${subject.name}.yaml`),\n        safeJoin(surfaces.workflows, `${subject.name}.json`),\n      ].filter((path): path is string => path !== null)\n    case 'rollout-policy':\n      return surfaces.rolloutPolicy ? [surfaces.rolloutPolicy] : []\n    case 'agent-profile':\n      return surfaces.agentProfile ? [surfaces.agentProfile] : []\n    case 'code': {\n      if (!surfaces.code) return []\n      const path = safeJoin(surfaces.code, subject.path)\n      return path ? [path] : []\n    }\n    case 'rag':\n      if (!surfaces.rag) return []\n      return optionalPath(safeJoin(surfaces.rag, subject.corpus, `${subject.docId}.md`))\n    case 'memory':\n      if (!surfaces.memory) return []\n      return optionalPath(safeJoin(surfaces.memory, `${slugify(subject.key)}.json`))\n    case 'scaffolding':\n      if (!surfaces.scaffolding) return []\n      return optionalPath(safeJoin(surfaces.scaffolding, `${slugify(subject.concern)}.md`))\n    case 'output-schema':\n      if (!surfaces.outputSchema) return []\n      return [surfaces.outputSchema]\n    case 'websearch.outdated':\n    case 'prior-run-summary':\n      // Stale signals don't map to a single file — handled by the knowledge\n      // adapter as `agent-knowledge:stale:*` after operator review.\n      return []\n    case 'cluster':\n      // failure-mode cluster labels are evidence, not mutations.\n      return []\n  }\n}\n\nfunction safeJoin(root: string, ...children: string[]): string | null {\n  if (children.some((child) => child.includes('\\0') || isAbsolute(child))) return null\n  const rootAbsolute = resolve(root)\n  const targetAbsolute = resolve(rootAbsolute, ...children)\n  const escaped = relative(rootAbsolute, targetAbsolute)\n  if (escaped === '..' || escaped.startsWith(`..${sep}`) || isAbsolute(escaped)) return null\n  return join(root, ...children)\n}\n\nfunction optionalPath(path: string | null): string[] {\n  return path ? [path] : []\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9-]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 200) || 'untitled'\n  )\n}\n\n/**\n * Validate that every declared surface exists on disk under `repoRoot`.\n *\n * Returns an array of `SurfaceValidationIssue` — empty when all required\n * surfaces resolve. `defineAgent` throws with the issues rendered, so\n * a misconfigured manifest fails at startup (not at the first finding\n * the loop produces 20 minutes later).\n */\nexport interface SurfaceValidationIssue {\n  surface: keyof AgentSurfaces\n  path: string\n  reason: 'missing' | 'not-directory' | 'not-file'\n}\n\n/** Validate an `AgentSurfaces` map on disk — missing paths fail loud at `defineAgent` time instead of silently skipping self-improvement edits. */\nexport function validateSurfaces(\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ReadonlyArray<SurfaceValidationIssue> {\n  const issues: SurfaceValidationIssue[] = []\n  const dirSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'systemPrompt',\n    'tools',\n    'personas',\n    'knowledge',\n  ]\n  const fileSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['rubric']\n  const optionalDirSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'scaffolding',\n    'memory',\n    'rag',\n    'skills',\n    'mcp',\n    'hooks',\n    'subagents',\n    'workflows',\n    'code',\n  ]\n  const optionalFileSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'outputSchema',\n    'rolloutPolicy',\n    'agentProfile',\n  ]\n\n  for (const key of dirSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    } else if (!statSync(abs).isDirectory()) {\n      issues.push({ surface: key, path: p, reason: 'not-directory' })\n    }\n  }\n  for (const key of fileSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    } else if (!statSync(abs).isFile()) {\n      issues.push({ surface: key, path: p, reason: 'not-file' })\n    }\n  }\n  for (const key of [...optionalDirSurfaces, ...optionalFileSurfaces]) {\n    const p = surfaces[key] as string | undefined\n    if (p === undefined) continue\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n      continue\n    }\n    const expectedDirectory = optionalDirSurfaces.includes(key)\n    if (expectedDirectory && !statSync(abs).isDirectory()) {\n      issues.push({ surface: key, path: p, reason: 'not-directory' })\n    } else if (!expectedDirectory && !statSync(abs).isFile()) {\n      issues.push({ surface: key, path: p, reason: 'not-file' })\n    }\n  }\n  return issues\n}\n\n/** Format a list of surface validation issues into a human-readable error string. */\nexport function renderSurfaceIssues(\n  issues: ReadonlyArray<SurfaceValidationIssue>,\n  repoRoot: string,\n): string {\n  if (issues.length === 0) return ''\n  const lines = issues.map(\n    (i) => `  - ${i.surface}: ${i.path ? `\"${i.path}\"` : '<not set>'} (${i.reason})`,\n  )\n  return [\n    `Agent surface validation failed against repoRoot=${repoRoot}:`,\n    ...lines,\n    '',\n    'Fix the manifest: every required surface must point at an existing',\n    'directory (systemPrompt / tools / personas / knowledge) or file',\n    '(rubric). Optional surfaces (scaffolding / memory / rag / outputSchema)',\n    'may be omitted; the loop will reject findings targeting omitted',\n    'surfaces rather than fabricating a path.',\n  ].join('\\n')\n}\n","/**\n * `defineAgent` — typed, validated manifest API for declarative agent\n * configuration. The substrate consumes this manifest to wire the\n * canonical eval pattern + analyst self-improvement loop without any\n * per-vertical glue.\n *\n * Design goal: scale to 1000s of vertical agents. Every agent declares\n * its surfaces, rubric, runtime, and analyst configuration in ~50 lines.\n * No per-vertical `ImprovementProposalSource`. No per-vertical CLI. No\n * fabricated paths.\n *\n * Validation: `defineAgent` runs `validateSurfaces` synchronously and\n * throws a structured error if any required surface is missing on\n * disk. The cost is one filesystem stat per surface (cheap); the\n * benefit is a manifest that can't ship broken.\n *\n * @experimental Manifests validate and load, but `runtime.act` is not yet wired\n * end-to-end into the substrate's eval path (`unimplementedAgentRun` is the\n * shipped default).\n */\n\nimport type { TraceAnalystDefinition } from '@tangle-network/agent-eval'\nimport type { RuntimeStreamEvent } from '../types'\nimport { type AgentSurfaces, renderSurfaceIssues, validateSurfaces } from './surfaces'\n\n// ── manifest ─────────────────────────────────────────────────────────\n\n/**\n * The full agent manifest. Each agent ships ONE of these.\n *\n * Generics:\n *   `TPersona` — the agent's persona shape (loaded from\n *     `surfaces.personas`). Defaults to `unknown` so the substrate's\n *     persona discovery (`loadPersonas`) can accept anything; per-agent\n *     code re-narrows when it matters.\n *   `TRunOutput` — the shape `runtime.act` returns. Used by the rubric\n *     scorers and emitted into the trace.\n */\nexport interface AgentManifest<TPersona = unknown, TRunOutput = unknown> {\n  /**\n   * Stable identifier — used as `projectId` in traces, as the analyst\n   * loop's `runId` prefix, and as the namespace under which findings\n   * are persisted. MUST match the agent's repo name to keep\n   * cross-repo telemetry joinable.\n   */\n  id: string\n\n  /**\n   * Filesystem root the substrate resolves surface paths against.\n   * Typically `process.cwd()` or a fixed absolute path. Use an\n   * absolute path when the agent's tests may run from subdirectories\n   * (vitest sometimes shifts cwd).\n   */\n  repoRoot: string\n\n  /**\n   * Map of mutable surfaces the self-improvement loop can edit. See\n   * `AgentSurfaces` — required: `systemPrompt`, `tools`, `rubric`,\n   * `knowledge`, `personas`. Optional: `scaffolding`, `memory`, `rag`,\n   * `outputSchema`.\n   *\n   * Every required path is validated at `defineAgent` time. Missing\n   * paths throw with the full list of offenders.\n   */\n  surfaces: AgentSurfaces\n\n  /**\n   * Rubric the substrate uses to score each run. Dimensions × weights\n   * × judges. The substrate computes the weighted composite and\n   * stamps it into the RunRecord.\n   */\n  rubric: AgentRubric<TRunOutput>\n\n  /**\n   * Runtime adapter — how the substrate INVOKES the agent against a\n   * persona. The `act` function takes a persona + a context (with the\n   * tracer the substrate threads through for span emission) and\n   * returns the run output the rubric will score.\n   *\n   * The agent's existing production runtime goes in here; the\n   * substrate is intentionally thin around it.\n   */\n  runtime: AgentRuntime<TPersona, TRunOutput>\n\n  /**\n   * Persona discovery — the substrate loads personas via this function\n   * at eval start. Can read from `surfaces.personas`, an API, or be\n   * hardcoded. The substrate calls it once per `runAgentEval` call;\n   * persona ordering is preserved.\n   */\n  personas: () => Promise<ReadonlyArray<TPersona>>\n\n  /**\n   * Analyst kinds the substrate runs against each persona's trace.\n   * Defaults to `DEFAULT_TRACE_ANALYST_KINDS` from agent-eval. Per-agent\n   * authors can prune (e.g. skip `knowledge-poisoning` when there's no\n   * knowledge base) or extend (custom domain kinds).\n   *\n   * Empty array disables the loop — useful for `pnpm eval --no-analyst`.\n   */\n  analystKinds: ReadonlyArray<TraceAnalystDefinition>\n\n  /**\n   * Analyst LLM configuration. The substrate uses these for all four\n   * kinds (override per-kind via `analystKinds` if needed).\n   */\n  analyst: AnalystConfig\n}\n\nexport interface AgentRubric<TRunOutput> {\n  /** Dimensions composing the weighted score. Weights sum to 1.0 by convention. */\n  dimensions: ReadonlyArray<RubricDimension<TRunOutput>>\n  /**\n   * Optional judges layered on top of deterministic dimensions. Each\n   * judge returns a score per dimension; the substrate averages judges\n   * (mean by default) for the LLM contribution.\n   */\n  judges?: ReadonlyArray<JudgeConfig<TRunOutput>>\n}\n\nexport interface RubricDimension<TRunOutput> {\n  /** Unique identifier — appears in finding subjects (`rubric:<id>`). */\n  id: string\n  /** 0..1 — weight in the composite. */\n  weight: number\n  /**\n   * Deterministic scorer: given the persona + run output, returns a\n   * 0..1 score. The substrate sums weight × score across dimensions\n   * for the deterministic composite; judges supplement subjective dims.\n   */\n  score: (input: { persona: unknown; output: TRunOutput }) => number\n  /** Optional human-readable label for reports. */\n  label?: string\n}\n\nexport interface JudgeConfig<TRunOutput> {\n  /** Judge identifier — appears in trace spans + manifest. */\n  id: string\n  /** Model snapshot to invoke. Pin the snapshot (`claude-sonnet-4-6@2025-04-15`); the validator rejects bare aliases. */\n  model: string\n  /** Dimensions this judge scores. */\n  dimensions: ReadonlyArray<string>\n  /**\n   * Optional rubric anchors — text examples the judge sees as a\n   * few-shot prompt to calibrate. STRONGLY recommended for subjective\n   * dimensions; required by the calibration gate (Pearson ≥0.7).\n   */\n  anchors?: ReadonlyArray<{ input: string; output: TRunOutput; expected: Record<string, number> }>\n}\n\nexport interface AgentRuntime<TPersona, TRunOutput> {\n  /**\n   * Invoke the agent against one persona. Returns BOTH:\n   *   - `events`: an `AsyncIterable<RuntimeStreamEvent>` the chat-centric\n   *     product consumes verbatim (SSE / WebSocket / inline render).\n   *     **Streaming is mandatory — never collapse this to a single Promise.**\n   *     The agent's existing `runChatTurn` (or equivalent async generator)\n   *     plugs in here directly.\n   *   - `output`: a `Promise<TRunOutput>` resolved AFTER the event stream\n   *     drains. The eval substrate awaits this for rubric scoring; chat\n   *     products usually ignore it (they already rendered incrementally).\n   *\n   * Implementation contract:\n   *   1. `act` MUST return immediately (synchronous construction of the\n   *      `events` iterator + the `output` promise).\n   *   2. Iterating `events` drives the underlying LLM/tool calls — the\n   *      caller chooses when to consume.\n   *   3. `output` resolves only after the iterator yields its terminal\n   *      event (typically `task_end`); see `collectAgentRun` helper.\n   *\n   * `ctx.emitter` is the substrate-threaded `TraceEmitter` — runtimes\n   * SHOULD record LLM/tool spans through it for capture integrity.\n   * `ctx.deadlineMs` is wall-clock; the runtime SHOULD honour for graceful\n   * cancel. `ctx.signal` is the standard abort signal.\n   */\n  act: (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput>\n}\n\nexport interface AgentRunInvocation<TRunOutput> {\n  /** Live stream of typed runtime events. Consumed by chat UX directly. */\n  events: AsyncIterable<RuntimeStreamEvent>\n  /** Final structured output the rubric scores. Resolves after `events` drains. */\n  output: Promise<TRunOutput>\n}\n\n/**\n * Stub for agents whose `runtime.act` is not yet wired to the substrate's\n * eval path. Preserves the streaming contract (empty event stream + a\n * rejected `output` promise that tells the caller exactly what to fix).\n *\n * Per-vertical manifests usually start with this stub and replace it with\n * the agent's real streaming runtime (`runChatTurn` or equivalent) once\n * the eval path consumes the manifest end-to-end.\n */\nexport function unimplementedAgentRun<TRunOutput = unknown>(\n  reason = 'AgentRuntime.act is not yet wired for this manifest',\n): AgentRunInvocation<TRunOutput> {\n  return {\n    events: (async function* empty(): AsyncIterable<RuntimeStreamEvent> {})(),\n    output: Promise.reject(new Error(reason)),\n  }\n}\n\n/**\n * Drain `act`'s `events` into an array AND await its `output`. Useful for\n * eval / outcome-measurement code paths that don't care about live\n * rendering. The events array is preserved so the substrate can inspect\n * tool calls / readiness / questions retrospectively.\n *\n * IMPORTANT: chat-centric UX MUST NOT call this — it defeats streaming\n * (no incremental render). Use `for await (const ev of invocation.events)`\n * directly in the chat surface.\n */\nexport async function collectAgentRun<TRunOutput>(\n  invocation: AgentRunInvocation<TRunOutput>,\n): Promise<{ events: ReadonlyArray<RuntimeStreamEvent>; output: TRunOutput }> {\n  const events: RuntimeStreamEvent[] = []\n  for await (const ev of invocation.events) events.push(ev)\n  const output = await invocation.output\n  return { events, output }\n}\n\nexport interface AgentRunContext {\n  /** Substrate-managed trace emitter. */\n  emitter: import('@tangle-network/agent-eval').TraceEmitter\n  /** Stable run id for this persona × variant cell. */\n  runId: string\n  /** Variant the runtime is exercising (e.g. `'baseline'`, `'source-grounded'`). */\n  variantId?: string\n  /** Wall-clock deadline (epoch ms). The runtime SHOULD honour for graceful cancel. */\n  deadlineMs?: number\n  /** Optional abort signal. */\n  signal?: AbortSignal\n}\n\nexport interface AnalystConfig {\n  /** Model the analyst kinds use. Override per-kind via `analystKinds[i].cost.models`. */\n  model: string\n  /** Optional total budget across all kinds for one run. Substrate enforces via `BudgetGuard`. */\n  budgetUsd?: number\n  /** Backend hint for the AxAIService factory — same shape every kind uses. */\n  backend?: {\n    name?: 'openai' | 'router'\n    apiKey?: string\n    baseUrl?: string\n  }\n}\n\n// ── factory + validation ─────────────────────────────────────────────\n\n/** Thrown when `defineAgent` finds a required surface missing on disk. */\nexport class AgentManifestError extends Error {\n  constructor(\n    message: string,\n    public readonly agentId: string,\n    public readonly issues: ReadonlyArray<unknown> = [],\n  ) {\n    super(message)\n    this.name = 'AgentManifestError'\n  }\n}\n\n/**\n * Construct a validated agent manifest. Throws `AgentManifestError`\n * if any required surface is missing on disk.\n *\n * Generics: pass your persona / output types if you want narrowed\n * `runtime.act` signatures:\n *   `defineAgent<TaxPersona, TaxRunOutput>({ ... })`\n *\n * Most callers don't need the generics — the substrate operates on\n * `unknown` payloads internally and the manifest's `score` /\n * `runtime.act` see the typed shapes via TypeScript inference at\n * the call site.\n */\nexport function defineAgent<TPersona = unknown, TRunOutput = unknown>(\n  manifest: AgentManifest<TPersona, TRunOutput>,\n): AgentManifest<TPersona, TRunOutput> {\n  if (!manifest.id || manifest.id.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `id` is required', manifest.id ?? '')\n  }\n  if (!manifest.repoRoot || manifest.repoRoot.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `repoRoot` is required', manifest.id)\n  }\n  const issues = validateSurfaces(manifest.surfaces, manifest.repoRoot)\n  if (issues.length > 0) {\n    throw new AgentManifestError(\n      renderSurfaceIssues(issues, manifest.repoRoot),\n      manifest.id,\n      issues,\n    )\n  }\n  // Lightweight rubric sanity: weights sum to ~1.0 (no hard requirement —\n  // the substrate normalizes — but flag wildly miscalibrated weights).\n  const total = manifest.rubric.dimensions.reduce((acc, d) => acc + d.weight, 0)\n  if (manifest.rubric.dimensions.length > 0 && (total < 0.5 || total > 1.5)) {\n    throw new AgentManifestError(\n      `defineAgent(${manifest.id}): rubric dimension weights sum to ${total.toFixed(3)} — should be ~1.0`,\n      manifest.id,\n    )\n  }\n  return manifest\n}\n","/**\n * Surface improvement proposer — resolves analyst findings into LLM-drafted\n * candidate patches without changing the caller's repository.\n *\n * The proposer parses each finding's `subject` via\n * `parseFindingSubject` (agent-eval), resolves it to a real file path\n * via the agent's `AgentSurfaces`, reads the current content, and asks\n * an LLM to draft a unified-diff patch given the finding + current\n * content + per-kind editing-discipline rules.\n *\n * Fail-loud rules:\n *   - Findings whose subject doesn't parse → counted in `errors`.\n *   - Findings whose subject targets an undeclared surface → counted in\n *     `errors` with the offending kind in the message.\n *   - Findings whose target path doesn't exist AND the kind isn't a\n *     create-new variant (`new-tool`, `knowledge.wiki`) → counted in\n *     `errors` with the resolved path in the message.\n *   - LLM drafts that fail JSON-schema validation → counted in\n *     `errors` with the schema issue.\n *\n * No silent skips. Every dropped finding has a recorded reason the\n * loop's report surfaces.\n */\n\nimport { readFileSync } from 'node:fs'\nimport type { AnalystFinding, FindingSubject } from '@tangle-network/agent-eval'\nimport { parseFindingSubject } from '@tangle-network/agent-eval/analyst'\nimport type { ImprovementProposalSource } from '../analyst-loop/types'\nimport { sha256 } from '../content-digest'\nimport type { AgentSurfaces, ResolvedSurface } from './surfaces'\nimport { resolveSubjectPath } from './surfaces'\n\n// ── proposal shape ───────────────────────────────────────────────────\n\nexport interface SurfaceImprovementEdit {\n  /** Stable id derived from the source finding so re-proposals are idempotent. */\n  id: string\n  /** The finding that produced this edit — for revert + audit trail. */\n  sourceFindingId: string\n  /** Parsed subject; included so the apply step doesn't re-parse. */\n  subject: FindingSubject\n  /** Resolved on-disk target. */\n  target: ResolvedSurface\n  /** SHA-256 of the current file content the patch was drafted against. */\n  baseSha256: string\n  /** Unified-diff patch the LLM drafted (relative to `target.absolutePath`). */\n  patch: string\n  /** One-line summary the operator sees in the report / PR title. */\n  summary: string\n  /** Multi-line rationale for the PR body — finding context + LLM reasoning. */\n  rationale: string\n  /** Carry-forward from the finding so the apply gate can check the threshold. */\n  confidence: number\n  /** Carry-forward severity for prioritization. */\n  severity: AnalystFinding['severity']\n}\n\nexport interface CreateSurfaceImprovementProposerOptions {\n  surfaces: AgentSurfaces\n  repoRoot: string\n  /**\n   * LLM-draft callback. Given a finding + current file content + the\n   * resolved target, returns a unified-diff patch + summary + rationale.\n   *\n   * Required — the substrate doesn't ship a hardcoded prompt; the agent\n   * author picks the model (Haiku for cheap routine drafts, Sonnet for\n   * substantive prompt rewrites, etc.) via this callback.\n   */\n  draftPatch: (input: DraftPatchInput) => Promise<DraftPatchOutput>\n  /**\n   * When the resolved target doesn't exist, allow the substrate to\n   * CREATE the file (for `knowledge.wiki`, `new-tool` subjects). Default\n   * true for those kinds, false for `system-prompt` / `rubric` / etc.\n   * (named sections that don't exist are a contract violation, not a\n   * scaffolding opportunity).\n   */\n  allowCreateForKinds?: ReadonlyArray<FindingSubject['kind']>\n}\n\nexport interface DraftPatchInput {\n  finding: AnalystFinding\n  subject: FindingSubject\n  target: ResolvedSurface\n  /** Current file content (empty string when `intent === 'create-new'`). */\n  currentContent: string\n}\n\nexport interface DraftPatchOutput {\n  /** Unified diff against the current file content. Empty string skips this finding. */\n  patch: string\n  /** One-line summary for the operator. */\n  summary: string\n  /** Multi-line rationale for the PR body. */\n  rationale: string\n}\n\n// ── factory ──────────────────────────────────────────────────────────\n\nconst DEFAULT_CREATE_KINDS: ReadonlyArray<FindingSubject['kind']> = [\n  'knowledge.wiki',\n  'knowledge.claim',\n  'knowledge.raw',\n  'new-tool',\n]\n\n/** Resolve each finding to a real surface and draft a detached patch candidate. */\nexport function createSurfaceImprovementProposer(\n  opts: CreateSurfaceImprovementProposerOptions,\n): ImprovementProposalSource<SurfaceImprovementEdit> {\n  const allowCreate = opts.allowCreateForKinds ?? DEFAULT_CREATE_KINDS\n\n  return {\n    async proposeFromFindings(findings) {\n      const edits: SurfaceImprovementEdit[] = []\n      const errors: Array<{ findingId: string; subject: string; message: string }> = []\n      let skipped = 0\n\n      for (const f of findings) {\n        const subject = parseFindingSubject(f.subject)\n        if (subject === null) {\n          if (f.subject !== undefined) {\n            errors.push({\n              findingId: f.finding_id,\n              subject: f.subject,\n              message: 'subject does not parse against the finding-subject grammar',\n            })\n          } else {\n            // Subject-less findings are descriptive, not actionable —\n            // legitimate; count in `skipped` not `errors`.\n            skipped += 1\n          }\n          continue\n        }\n\n        // `cluster` findings (failure-mode) are evidence, not mutations.\n        if (subject.kind === 'cluster') {\n          skipped += 1\n          continue\n        }\n\n        // Knowledge findings flow to the knowledge proposal source so they do not double-route.\n        if (subject.kind.startsWith('knowledge.')) {\n          skipped += 1\n          continue\n        }\n\n        const target = resolveSubjectPath(subject, opts.surfaces, opts.repoRoot)\n        if (target === null) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `subject kind \"${subject.kind}\" targets an undeclared surface; declare it in AgentSurfaces or stop emitting this subject`,\n          })\n          continue\n        }\n\n        if (target.intent === 'create-new' && !allowCreate.includes(subject.kind)) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `target ${target.repoRelativePath} does not exist; the kind \"${subject.kind}\" requires an existing target (analyst named a section that isn't in the codebase)`,\n          })\n          continue\n        }\n\n        const currentContent = target.exists ? readFileSync(target.absolutePath, 'utf-8') : ''\n\n        let draft: DraftPatchOutput\n        try {\n          draft = await opts.draftPatch({ finding: f, subject, target, currentContent })\n        } catch (err) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `draftPatch threw: ${err instanceof Error ? err.message : String(err)}`,\n          })\n          continue\n        }\n\n        if (draft.patch.trim().length === 0) {\n          skipped += 1\n          continue\n        }\n\n        edits.push({\n          id: `imp-${f.finding_id}`,\n          sourceFindingId: f.finding_id,\n          subject,\n          target,\n          baseSha256: sha256(currentContent),\n          patch: draft.patch,\n          summary: draft.summary,\n          rationale: draft.rationale,\n          confidence: f.confidence,\n          severity: f.severity,\n        })\n      }\n\n      return { edits, skipped, errors }\n    },\n  }\n}\n","/**\n * Sandbox bridge for `AgentRuntime.act` — prod-faithful eval execution.\n *\n * The point of this adapter is parity: the eval substrate must run the agent\n * through the SAME profile the production chat turn uses, or scorecard numbers\n * grade a profile that never ships. `createSandboxAct` boots a sandbox with the\n * agent's profile (the caller's `baseProfile`, with optional per-persona\n * overrides) through the loop kernel's own {@link createSandboxForSpec},\n * streams the `streamPrompt` events mapped to the `RuntimeStreamEvent`\n * vocabulary, and resolves the `OutputAdapter`-parsed output for rubric scoring\n * — satisfying the `act` streaming contract with one code path shared by chat\n * and eval.\n *\n * Agents with a bespoke streaming chat turn should wire THAT into `act`\n * directly (the contract is designed for it); this adapter is the default for\n * agents whose turn is a plain prod-profile sandbox dispatch — notably the\n * agents agent-builder generates.\n */\n\nimport type {\n  AgentProfile,\n  AgentProfileFileMount,\n  AgentProfileMcpServer,\n} from '@tangle-network/agent-interface'\nimport type { SandboxEvent } from '@tangle-network/sandbox'\nimport type { AgentRunSpec, OutputAdapter, SandboxClient } from '../runtime'\nimport { mapSandboxEvent } from '../runtime'\nimport { createSandboxForSpec } from '../runtime/iteration-sandboxes'\nimport type { RuntimeStreamEvent } from '../types'\nimport type { AgentRunContext, AgentRunInvocation } from './define-agent'\nimport {\n  type AgentProfileMaterializationAxis,\n  assertProfileMaterialization,\n  sandboxActProfileMaterialization,\n} from './profile-materialization'\n\n/** Per-persona profile-merge slots applied over the base profile (§1.5: the caller authors the\n *  per-persona profile). Each slot overlays the base; an absent slot leaves the base untouched. */\nexport interface SandboxActComposeOverrides {\n  /** Replace the base profile's system prompt (e.g. a workspace-augmented prompt). */\n  systemPrompt?: string\n  /** Extra file mounts layered after the base profile's `resources.files`. */\n  extraFiles?: AgentProfileFileMount[]\n  /** Override the profile `name`. Defaults to the base profile's name. */\n  name?: string\n  /** Box built-in tool ON/OFF flags merged over the base profile's `tools` (overlay wins per key). */\n  tools?: Record<string, boolean>\n  /** MCP connections merged over the base profile's `mcp` (overlay wins per key). */\n  mcpConnections?: Record<string, AgentProfileMcpServer>\n}\n\nexport interface CreateSandboxActOptions<TPersona, TRunOutput> {\n  /** Canonical agent profile — the same one the prod chat turn uses. */\n  baseProfile: AgentProfile\n  /** Sandbox client used to boot the per-run sandbox. */\n  sandboxClient: SandboxClient\n  /** Persona → prompt. Pure; the eval cell's input. */\n  buildPrompt: (persona: TPersona) => string\n  /** Sandbox event stream → typed output the rubric scores. */\n  output: OutputAdapter<TRunOutput>\n  /**\n   * Per-persona profile overrides (workspace-augmented system prompt, extra\n   * file mounts, tool flags, MCP connections). Overlaid onto `baseProfile`.\n   */\n  compose?: (persona: TPersona) => SandboxActComposeOverrides\n  /** Sandbox-SDK overrides forwarded to `createSandboxForSpec`. */\n  sandboxOverrides?: AgentRunSpec<unknown>['sandboxOverrides']\n  /** Optional changed axes the caller expects this path to carry. */\n  requiredProfileAxes?: readonly AgentProfileMaterializationAxis[]\n  /** Stable run name surfaced in mapped `llm_call` events. */\n  name?: string\n  /** Override the `SandboxEvent → RuntimeStreamEvent` mapper. */\n  mapEvent?: (\n    event: SandboxEvent,\n    opts: { agentRunName?: string },\n  ) => RuntimeStreamEvent | undefined\n}\n\n/**\n * Build an `AgentRuntime.act` implementation backed by a single prod-profile\n * sandbox run. The returned function honours the `act` contract: it returns\n * synchronously with a live `events` iterator and an `output` promise that\n * resolves only after the iterator drains.\n */\nexport function createSandboxAct<TPersona, TRunOutput>(\n  options: CreateSandboxActOptions<TPersona, TRunOutput>,\n): (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput> {\n  assertProfileMaterialization({\n    contract: sandboxActProfileMaterialization,\n    changedAxes: options.requiredProfileAxes ?? [],\n    context: 'createSandboxAct',\n  })\n  const mapEvent = options.mapEvent ?? mapSandboxEvent\n\n  return (persona: TPersona, ctx: AgentRunContext): AgentRunInvocation<TRunOutput> => {\n    const profile = applyComposeOverrides(options.baseProfile, options.compose?.(persona))\n    const agentRunName = options.name ?? profile.name ?? 'agent'\n    const message = options.buildPrompt(persona)\n    const signal = ctx.signal ?? new AbortController().signal\n\n    const raw: SandboxEvent[] = []\n    let settle!: (value: TRunOutput) => void\n    let fail!: (err: unknown) => void\n    const output = new Promise<TRunOutput>((resolve, reject) => {\n      settle = resolve\n      fail = reject\n    })\n    // The output promise rejects when the stream errors; if the caller ignores\n    // `output` (chat UX) the rejection is still observed by the events iterator\n    // throwing. Attach a no-op catch so an ignored rejection is never \"unhandled\".\n    output.catch(() => {})\n\n    const spec: AgentRunSpec<unknown> = {\n      profile,\n      taskToPrompt: () => message,\n      name: agentRunName,\n      ...(options.sandboxOverrides ? { sandboxOverrides: options.sandboxOverrides } : {}),\n    }\n\n    async function* events(): AsyncIterable<RuntimeStreamEvent> {\n      try {\n        const box = await createSandboxForSpec(options.sandboxClient, spec, signal)\n        for await (const event of box.streamPrompt(message, { signal })) {\n          raw.push(event)\n          const mapped = mapEvent(event, { agentRunName })\n          if (mapped) yield mapped\n        }\n        settle(options.output.parse(raw))\n      } catch (err) {\n        fail(err)\n        throw err\n      }\n    }\n\n    return { events: events(), output }\n  }\n}\n\n/** Overlay the per-persona overrides onto the base profile. Each slot merges over the base; an\n *  absent override leaves the base profile untouched. */\nfunction applyComposeOverrides(\n  base: AgentProfile,\n  overrides: SandboxActComposeOverrides | undefined,\n): AgentProfile {\n  if (!overrides) return base\n  const prompt = overrides.systemPrompt\n    ? { ...base.prompt, systemPrompt: overrides.systemPrompt }\n    : base.prompt\n  const mergedTools = overrides.tools ? { ...(base.tools ?? {}), ...overrides.tools } : base.tools\n  const mergedMcp = overrides.mcpConnections\n    ? { ...(base.mcp ?? {}), ...overrides.mcpConnections }\n    : base.mcp\n  const baseFiles = base.resources?.files ?? []\n  const mergedFiles: AgentProfileFileMount[] = overrides.extraFiles?.length\n    ? [...baseFiles, ...overrides.extraFiles]\n    : [...baseFiles]\n  return {\n    ...base,\n    name: overrides.name ?? base.name,\n    prompt,\n    ...(mergedTools ? { tools: mergedTools } : {}),\n    ...(mergedMcp ? { mcp: mergedMcp } : {}),\n    resources: { ...base.resources, files: mergedFiles },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqGA,SAAgB,mBACd,SACA,UACA,UACwB;CACxB,MAAM,aAAa,yBAAyB,SAAS,QAAQ;CAC7D,IAAI,WAAW,WAAW,GAAG,OAAO;CAMpC,KAAK,MAAM,OAAO,YAAY;EAC5B,MAAM,MAAM,WAAW,GAAG,IAAI,MAAM,KAAK,UAAU,GAAG;EACtD,IAAI,WAAW,GAAG,GAChB,OAAO;GAAE,cAAc;GAAK,kBAAkB;GAAK,QAAQ;GAAM,QAAQ;EAAgB;CAE7F;CACA,MAAM,WAAW,WAAW;CAE5B,OAAO;EACL,cAFkB,WAAW,QAAQ,IAAI,WAAW,KAAK,UAAU,QAAQ;EAG3E,kBAAkB;EAClB,QAAQ;EACR,QAAQ;CACV;AACF;AAEA,SAAS,yBACP,SACA,UACuB;CACvB,QAAQ,QAAQ,MAAhB;EACE,KAAK;EACL,KAAK,mBACH,OAAO,aAAa,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,IAAI,CAAC;EACxE,KAAK,mBAEH,OAAO,aAAa,SAAS,SAAS,WAAW,UAAU,GAAG,QAAQ,QAAQ,KAAK,EAAE,IAAI,CAAC;EAC5F,KAAK,iBACH,OAAO,aAAa,SAAS,SAAS,WAAW,OAAO,GAAG,QAAQ,SAAS,IAAI,CAAC;EACnF,KAAK,iBAAiB;GACpB,MAAM,OAAO,QAAQ,QAAQ,OAAO;GAIpC,OAAO;IACL,SAAS,SAAS,cAAc,GAAG,KAAK,IAAI;IAC5C,SAAS,SAAS,cAAc,MAAM,UAAU;IAChD,SAAS,SAAS,cAAc,MAAM,UAAU;GAClD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD;EACA,KAAK;GACH,IAAI,CAAC,SAAS,QAAQ,OAAO,CAAC;GAC9B,OAAO,CACL,SAAS,SAAS,QAAQ,QAAQ,MAAM,UAAU,GAClD,SAAS,SAAS,QAAQ,GAAG,QAAQ,KAAK,IAAI,CAChD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAElD,KAAK;GACH,IAAI,QAAQ,QACV,OAAO,aAAa,SAAS,SAAS,OAAO,QAAQ,MAAM,GAAG,QAAQ,QAAQ,MAAM,EAAE,IAAI,CAAC;GAI7F,OAAO,CACL,SAAS,SAAS,OAAO,QAAQ,MAAM,WAAW,GAClD,SAAS,SAAS,OAAO,GAAG,QAAQ,KAAK,IAAI,CAC/C,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK,YACH,OAAO,aAAa,SAAS,SAAS,OAAO,QAAQ,MAAM,WAAW,CAAC;EACzE,KAAK;GACH,IAAI,CAAC,SAAS,KAAK,OAAO,CAAC;GAC3B,OAAO,QAAQ,OACX,aAAa,SAAS,SAAS,KAAK,QAAQ,QAAQ,GAAG,QAAQ,KAAK,IAAI,CAAC,IACzE,CACE,SAAS,SAAS,KAAK,GAAG,QAAQ,OAAO,MAAM,GAC/C,SAAS,SAAS,KAAK,QAAQ,QAAQ,WAAW,CACpD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EACtD,KAAK;GACH,IAAI,CAAC,SAAS,OAAO,OAAO,CAAC;GAC7B,OAAO,CACL,SAAS,SAAS,OAAO,GAAG,QAAQ,KAAK,IAAI,GAC7C,SAAS,SAAS,OAAO,GAAG,QAAQ,KAAK,MAAM,CACjD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK;GACH,IAAI,CAAC,SAAS,WAAW,OAAO,CAAC;GACjC,OAAO;IACL,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,IAAI;IACjD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;IACnD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;GACrD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK;GACH,IAAI,CAAC,SAAS,WAAW,OAAO,CAAC;GACjC,OAAO;IACL,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,IAAI;IACjD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;IACnD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;GACrD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK,kBACH,OAAO,SAAS,gBAAgB,CAAC,SAAS,aAAa,IAAI,CAAC;EAC9D,KAAK,iBACH,OAAO,SAAS,eAAe,CAAC,SAAS,YAAY,IAAI,CAAC;EAC5D,KAAK,QAAQ;GACX,IAAI,CAAC,SAAS,MAAM,OAAO,CAAC;GAC5B,MAAM,OAAO,SAAS,SAAS,MAAM,QAAQ,IAAI;GACjD,OAAO,OAAO,CAAC,IAAI,IAAI,CAAC;EAC1B;EACA,KAAK;GACH,IAAI,CAAC,SAAS,KAAK,OAAO,CAAC;GAC3B,OAAO,aAAa,SAAS,SAAS,KAAK,QAAQ,QAAQ,GAAG,QAAQ,MAAM,IAAI,CAAC;EACnF,KAAK;GACH,IAAI,CAAC,SAAS,QAAQ,OAAO,CAAC;GAC9B,OAAO,aAAa,SAAS,SAAS,QAAQ,GAAG,QAAQ,QAAQ,GAAG,EAAE,MAAM,CAAC;EAC/E,KAAK;GACH,IAAI,CAAC,SAAS,aAAa,OAAO,CAAC;GACnC,OAAO,aAAa,SAAS,SAAS,aAAa,GAAG,QAAQ,QAAQ,OAAO,EAAE,IAAI,CAAC;EACtF,KAAK;GACH,IAAI,CAAC,SAAS,cAAc,OAAO,CAAC;GACpC,OAAO,CAAC,SAAS,YAAY;EAC/B,KAAK;EACL,KAAK,qBAGH,OAAO,CAAC;EACV,KAAK,WAEH,OAAO,CAAC;CACZ;AACF;AAEA,SAAS,SAAS,MAAc,GAAG,UAAmC;CACpE,IAAI,SAAS,MAAM,UAAU,MAAM,SAAS,IAAI,KAAK,WAAW,KAAK,CAAC,GAAG,OAAO;CAChF,MAAM,eAAe,QAAQ,IAAI;CAEjC,MAAM,UAAU,SAAS,cADF,QAAQ,cAAc,GAAG,QACI,CAAC;CACrD,IAAI,YAAY,QAAQ,QAAQ,WAAW,KAAK,KAAK,KAAK,WAAW,OAAO,GAAG,OAAO;CACtF,OAAO,KAAK,MAAM,GAAG,QAAQ;AAC/B;AAEA,SAAS,aAAa,MAA+B;CACnD,OAAO,OAAO,CAAC,IAAI,IAAI,CAAC;AAC1B;AAEA,SAAS,QAAQ,GAAmB;CAClC,OACE,EACG,YAAY,CAAC,CACb,QAAQ,gBAAgB,GAAG,CAAC,CAC5B,QAAQ,YAAY,EAAE,CAAC,CACvB,MAAM,GAAG,GAAG,KAAK;AAExB;;AAiBA,SAAgB,iBACd,UACA,UACuC;CACvC,MAAM,SAAmC,CAAC;CAC1C,MAAM,cAAkD;EACtD;EACA;EACA;EACA;CACF;CACA,MAAM,eAAmD,CAAC,QAAQ;CAClE,MAAM,sBAA0D;EAC9D;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;CACA,MAAM,uBAA2D;EAC/D;EACA;EACA;CACF;CAEA,KAAK,MAAM,OAAO,aAAa;EAC7B,MAAM,IAAI,SAAS;EACnB,IAAI,CAAC,GAAG;GACN,OAAO,KAAK;IAAE,SAAS;IAAK,MAAM;IAAI,QAAQ;GAAU,CAAC;GACzD;EACF;EACA,MAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;EAChD,IAAI,CAAC,WAAW,GAAG,GACjB,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAU,CAAC;OACnD,IAAI,CAAC,SAAS,GAAG,CAAC,CAAC,YAAY,GACpC,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAgB,CAAC;CAElE;CACA,KAAK,MAAM,OAAO,cAAc;EAC9B,MAAM,IAAI,SAAS;EACnB,IAAI,CAAC,GAAG;GACN,OAAO,KAAK;IAAE,SAAS;IAAK,MAAM;IAAI,QAAQ;GAAU,CAAC;GACzD;EACF;EACA,MAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;EAChD,IAAI,CAAC,WAAW,GAAG,GACjB,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAU,CAAC;OACnD,IAAI,CAAC,SAAS,GAAG,CAAC,CAAC,OAAO,GAC/B,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAW,CAAC;CAE7D;CACA,KAAK,MAAM,OAAO,CAAC,GAAG,qBAAqB,GAAG,oBAAoB,GAAG;EACnE,MAAM,IAAI,SAAS;EACnB,IAAI,MAAM,KAAA,GAAW;EACrB,MAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;EAChD,IAAI,CAAC,WAAW,GAAG,GAAG;GACpB,OAAO,KAAK;IAAE,SAAS;IAAK,MAAM;IAAG,QAAQ;GAAU,CAAC;GACxD;EACF;EACA,MAAM,oBAAoB,oBAAoB,SAAS,GAAG;EAC1D,IAAI,qBAAqB,CAAC,SAAS,GAAG,CAAC,CAAC,YAAY,GAClD,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAgB,CAAC;OACzD,IAAI,CAAC,qBAAqB,CAAC,SAAS,GAAG,CAAC,CAAC,OAAO,GACrD,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAW,CAAC;CAE7D;CACA,OAAO;AACT;;AAGA,SAAgB,oBACd,QACA,UACQ;CACR,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,QAAQ,OAAO,KAClB,MAAM,OAAO,EAAE,QAAQ,IAAI,EAAE,OAAO,IAAI,EAAE,KAAK,KAAK,YAAY,IAAI,EAAE,OAAO,EAChF;CACA,OAAO;EACL,oDAAoD,SAAS;EAC7D,GAAG;EACH;EACA;EACA;EACA;EACA;EACA;CACF,CAAC,CAAC,KAAK,IAAI;AACb;;;;;;;;;;;;ACxKA,SAAgB,sBACd,SAAS,uDACuB;CAChC,OAAO;EACL,SAAS,gBAAgB,QAA2C,CAAC,EAAA,CAAG;EACxE,QAAQ,QAAQ,OAAO,IAAI,MAAM,MAAM,CAAC;CAC1C;AACF;;AAkDA,IAAa,qBAAb,cAAwC,MAAM;CAG1B;CACA;CAHlB,YACE,SACA,SACA,SAAiD,CAAC,GAClD;EACA,MAAM,OAAO;EAHG,KAAA,UAAA;EACA,KAAA,SAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;;;;;;;;;;;;AAeA,SAAgB,YACd,UACqC;CACrC,IAAI,CAAC,SAAS,MAAM,SAAS,GAAG,KAAK,CAAC,CAAC,WAAW,GAChD,MAAM,IAAI,mBAAmB,iCAAiC,SAAS,MAAM,EAAE;CAEjF,IAAI,CAAC,SAAS,YAAY,SAAS,SAAS,KAAK,CAAC,CAAC,WAAW,GAC5D,MAAM,IAAI,mBAAmB,uCAAuC,SAAS,EAAE;CAEjF,MAAM,SAAS,iBAAiB,SAAS,UAAU,SAAS,QAAQ;CACpE,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,mBACR,oBAAoB,QAAQ,SAAS,QAAQ,GAC7C,SAAS,IACT,MACF;CAIF,MAAM,QAAQ,SAAS,OAAO,WAAW,QAAQ,KAAK,MAAM,MAAM,EAAE,QAAQ,CAAC;CAC7E,IAAI,SAAS,OAAO,WAAW,SAAS,MAAM,QAAQ,MAAO,QAAQ,MACnE,MAAM,IAAI,mBACR,eAAe,SAAS,GAAG,qCAAqC,MAAM,QAAQ,CAAC,EAAE,oBACjF,SAAS,EACX;CAEF,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;;AC5MA,MAAM,uBAA8D;CAClE;CACA;CACA;CACA;AACF;;AAGA,SAAgB,iCACd,MACmD;CACnD,MAAM,cAAc,KAAK,uBAAuB;CAEhD,OAAO,EACL,MAAM,oBAAoB,UAAU;EAClC,MAAM,QAAkC,CAAC;EACzC,MAAM,SAAyE,CAAC;EAChF,IAAI,UAAU;EAEd,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,UAAU,oBAAoB,EAAE,OAAO;GAC7C,IAAI,YAAY,MAAM;IACpB,IAAI,EAAE,YAAY,KAAA,GAChB,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE;KACX,SAAS;IACX,CAAC;SAID,WAAW;IAEb;GACF;GAGA,IAAI,QAAQ,SAAS,WAAW;IAC9B,WAAW;IACX;GACF;GAGA,IAAI,QAAQ,KAAK,WAAW,YAAY,GAAG;IACzC,WAAW;IACX;GACF;GAEA,MAAM,SAAS,mBAAmB,SAAS,KAAK,UAAU,KAAK,QAAQ;GACvE,IAAI,WAAW,MAAM;IACnB,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE,WAAW;KACtB,SAAS,iBAAiB,QAAQ,KAAK;IACzC,CAAC;IACD;GACF;GAEA,IAAI,OAAO,WAAW,gBAAgB,CAAC,YAAY,SAAS,QAAQ,IAAI,GAAG;IACzE,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE,WAAW;KACtB,SAAS,UAAU,OAAO,iBAAiB,6BAA6B,QAAQ,KAAK;IACvF,CAAC;IACD;GACF;GAEA,MAAM,iBAAiB,OAAO,SAAS,aAAa,OAAO,cAAc,OAAO,IAAI;GAEpF,IAAI;GACJ,IAAI;IACF,QAAQ,MAAM,KAAK,WAAW;KAAE,SAAS;KAAG;KAAS;KAAQ;IAAe,CAAC;GAC/E,SAAS,KAAK;IACZ,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE,WAAW;KACtB,SAAS,qBAAqB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC/E,CAAC;IACD;GACF;GAEA,IAAI,MAAM,MAAM,KAAK,CAAC,CAAC,WAAW,GAAG;IACnC,WAAW;IACX;GACF;GAEA,MAAM,KAAK;IACT,IAAI,OAAO,EAAE;IACb,iBAAiB,EAAE;IACnB;IACA;IACA,YAAY,OAAO,cAAc;IACjC,OAAO,MAAM;IACb,SAAS,MAAM;IACf,WAAW,MAAM;IACjB,YAAY,EAAE;IACd,UAAU,EAAE;GACd,CAAC;EACH;EAEA,OAAO;GAAE;GAAO;GAAS;EAAO;CAClC,EACF;AACF;;;;;;;;;ACrHA,SAAgB,iBACd,SAC6E;CAC7E,6BAA6B;EAC3B,UAAU;EACV,aAAa,QAAQ,uBAAuB,CAAC;EAC7C,SAAS;CACX,CAAC;CACD,MAAM,WAAW,QAAQ,YAAY;CAErC,QAAQ,SAAmB,QAAyD;EAClF,MAAM,UAAU,sBAAsB,QAAQ,aAAa,QAAQ,UAAU,OAAO,CAAC;EACrF,MAAM,eAAe,QAAQ,QAAQ,QAAQ,QAAQ;EACrD,MAAM,UAAU,QAAQ,YAAY,OAAO;EAC3C,MAAM,SAAS,IAAI,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAEnD,MAAM,MAAsB,CAAC;EAC7B,IAAI;EACJ,IAAI;EACJ,MAAM,SAAS,IAAI,SAAqB,SAAS,WAAW;GAC1D,SAAS;GACT,OAAO;EACT,CAAC;EAID,OAAO,YAAY,CAAC,CAAC;EAErB,MAAM,OAA8B;GAClC;GACA,oBAAoB;GACpB,MAAM;GACN,GAAI,QAAQ,mBAAmB,EAAE,kBAAkB,QAAQ,iBAAiB,IAAI,CAAC;EACnF;EAEA,gBAAgB,SAA4C;GAC1D,IAAI;IACF,MAAM,MAAM,MAAM,qBAAqB,QAAQ,eAAe,MAAM,MAAM;IAC1E,WAAW,MAAM,SAAS,IAAI,aAAa,SAAS,EAAE,OAAO,CAAC,GAAG;KAC/D,IAAI,KAAK,KAAK;KACd,MAAM,SAAS,SAAS,OAAO,EAAE,aAAa,CAAC;KAC/C,IAAI,QAAQ,MAAM;IACpB;IACA,OAAO,QAAQ,OAAO,MAAM,GAAG,CAAC;GAClC,SAAS,KAAK;IACZ,KAAK,GAAG;IACR,MAAM;GACR;EACF;EAEA,OAAO;GAAE,QAAQ,OAAO;GAAG;EAAO;CACpC;AACF;;;AAIA,SAAS,sBACP,MACA,WACc;CACd,IAAI,CAAC,WAAW,OAAO;CACvB,MAAM,SAAS,UAAU,eACrB;EAAE,GAAG,KAAK;EAAQ,cAAc,UAAU;CAAa,IACvD,KAAK;CACT,MAAM,cAAc,UAAU,QAAQ;EAAE,GAAI,KAAK,SAAS,CAAC;EAAI,GAAG,UAAU;CAAM,IAAI,KAAK;CAC3F,MAAM,YAAY,UAAU,iBACxB;EAAE,GAAI,KAAK,OAAO,CAAC;EAAI,GAAG,UAAU;CAAe,IACnD,KAAK;CACT,MAAM,YAAY,KAAK,WAAW,SAAS,CAAC;CAC5C,MAAM,cAAuC,UAAU,YAAY,SAC/D,CAAC,GAAG,WAAW,GAAG,UAAU,UAAU,IACtC,CAAC,GAAG,SAAS;CACjB,OAAO;EACL,GAAG;EACH,MAAM,UAAU,QAAQ,KAAK;EAC7B;EACA,GAAI,cAAc,EAAE,OAAO,YAAY,IAAI,CAAC;EAC5C,GAAI,YAAY,EAAE,KAAK,UAAU,IAAI,CAAC;EACtC,WAAW;GAAE,GAAG,KAAK;GAAW,OAAO;EAAY;CACrD;AACF"}
+{"version":3,"file":"agent.js","names":[],"sources":["../src/agent/surfaces.ts","../src/agent/define-agent.ts","../src/agent/improvement-adapter.ts","../src/agent/sandbox-act.ts"],"sourcesContent":["/**\n * `AgentSurfaces` — declarative map of the mutable file/directory paths\n * the self-improvement loop can edit on behalf of an agent.\n *\n * The substrate uses this map to resolve every parsed `FindingSubject`\n * (from agent-eval) to a real on-disk path. No per-vertical glue;\n * no fabricated paths; no silent `existsSync(...)` skips that hide\n * misconfiguration from the operator.\n *\n * Surfaces are validated at `defineAgent` time — missing paths fail\n * loud with a list of every offender. A surface that's not needed\n * (e.g. an agent with no RAG corpora) is simply omitted; the loop\n * refuses to route those subjects rather than fabricating a target.\n */\n\nimport { existsSync, statSync } from 'node:fs'\nimport { isAbsolute, join, relative, resolve, sep } from 'node:path'\nimport type { FindingSubject } from '@tangle-network/agent-eval'\n\n/**\n * Surface declarations. Every path is repo-relative (or absolute) at\n * `defineAgent` time. At resolution time, paths are joined against the\n * agent's `repoRoot`.\n *\n * `systemPrompt`, `tools`, `personas` are DIRECTORIES; the loop appends\n * `<section>.md`, `<tool>/README.md`, `<persona-id>.yaml` etc.\n * `rubric`, `outputSchema` are SINGLE FILES; the loop edits them in\n * place.\n *\n * `knowledge` is the agent-knowledge root (typically `.agent-knowledge`);\n * `applyKnowledgeWriteBlocks` writes pages relative to it.\n *\n * Optional surfaces (`scaffolding`, `memory`, `rag`, `outputSchema`)\n * can be omitted — the loop will reject findings targeting them with a\n * clear log message instead of fabricating a path.\n */\nexport interface AgentSurfaces {\n  /** Directory containing one markdown file per system-prompt section. */\n  systemPrompt: string\n  /** Directory containing one subdir per tool (`<tool>/README.md`). */\n  tools: string\n  /** Single file (TypeScript module) defining the rubric weights + dimensions. */\n  rubric: string\n  /** Knowledge-base root; typically `.agent-knowledge`. */\n  knowledge: string\n  /** Directory containing one YAML/JSON file per persona. */\n  personas: string\n  /** Optional: directory containing scaffolding rules (precondition checks, retry policies). */\n  scaffolding?: string\n  /** Optional: memory store path (JSONL / SQLite / DB). */\n  memory?: string\n  /** Optional: directory containing RAG corpora (`<corpus>/<doc-id>.md`). */\n  rag?: string\n  /** Optional: single file defining the output schema (Zod / JSON Schema). */\n  outputSchema?: string\n  /** Optional: directory containing Agent Skill packages. */\n  skills?: string\n  /** Optional: directory containing MCP server/tool configuration. */\n  mcp?: string\n  /** Optional: directory containing hook definitions. */\n  hooks?: string\n  /** Optional: directory containing subagent definitions. */\n  subagents?: string\n  /** Optional: directory containing orchestration/workflow policies. */\n  workflows?: string\n  /** Optional: single file containing rollout-policy settings. */\n  rolloutPolicy?: string\n  /** Optional: single canonical AgentProfile file. */\n  agentProfile?: string\n  /** Optional: source root for code findings. */\n  code?: string\n}\n\nexport interface ResolvedSurface {\n  /** Absolute filesystem path the operator can `cat` / `vim`. */\n  absolutePath: string\n  /** Repo-relative path for PR descriptions, diffs, audit logs. */\n  repoRelativePath: string\n  /** Whether the path currently exists on disk. */\n  exists: boolean\n  /** The substrate's intent: edit an existing file or create a new one. */\n  intent: 'edit-existing' | 'create-new'\n}\n\n/**\n * Resolve a parsed `FindingSubject` to the file path the substrate\n * should edit (or create) on disk.\n *\n * Returns `null` when:\n *   - the subject targets a surface the agent didn't declare\n *     (e.g. `rag:*` when `surfaces.rag` is undefined), OR\n *   - the subject is a `cluster` (failure-mode emits these as evidence,\n *     not actionable mutations — they don't route to a file).\n *\n * Returns a `ResolvedSurface` with `intent: 'create-new'` when the\n * subject names a path that doesn't yet exist (e.g. a new wiki page).\n * The caller chooses whether to honour the create — for tightly-managed\n * surfaces like `systemPrompt` it's usually a contract violation\n * (the analyst named a section that doesn't exist); for `knowledge`\n * it's the whole point.\n */\nexport function resolveSubjectPath(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ResolvedSurface | null {\n  const candidates = candidatePathsForSubject(subject, surfaces)\n  if (candidates.length === 0) return null\n\n  // Probe candidates in order, preferring the first one that exists on disk.\n  // Lets the substrate accept both the flat `<section>.md` convention and\n  // the skill-dir `<section>/SKILL.md` convention without forcing one layout.\n  // When none exists, fall back to the first candidate (canonical create-new).\n  for (const rel of candidates) {\n    const abs = isAbsolute(rel) ? rel : join(repoRoot, rel)\n    if (existsSync(abs)) {\n      return { absolutePath: abs, repoRelativePath: rel, exists: true, intent: 'edit-existing' }\n    }\n  }\n  const fallback = candidates[0]!\n  const fallbackAbs = isAbsolute(fallback) ? fallback : join(repoRoot, fallback)\n  return {\n    absolutePath: fallbackAbs,\n    repoRelativePath: fallback,\n    exists: false,\n    intent: 'create-new',\n  }\n}\n\nfunction candidatePathsForSubject(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n): ReadonlyArray<string> {\n  switch (subject.kind) {\n    case 'knowledge.wiki':\n    case 'knowledge.stale':\n      return optionalPath(safeJoin(surfaces.knowledge, `${subject.slug}.md`))\n    case 'knowledge.claim':\n      // Claims land in a per-topic claims directory under the knowledge root.\n      return optionalPath(safeJoin(surfaces.knowledge, 'claims', `${slugify(subject.topic)}.md`))\n    case 'knowledge.raw':\n      return optionalPath(safeJoin(surfaces.knowledge, 'raw', `${subject.sourceId}.md`))\n    case 'system-prompt': {\n      const slug = slugify(subject.section)\n      // Prefer flat layout for create-new (canonical); probe skill-dir layout\n      // in case the existing repo (tax/legal/gtm/creative) uses\n      // `<section>/SKILL.md` already.\n      return [\n        safeJoin(surfaces.systemPrompt, `${slug}.md`),\n        safeJoin(surfaces.systemPrompt, slug, 'SKILL.md'),\n        safeJoin(surfaces.systemPrompt, slug, 'index.md'),\n      ].filter((path): path is string => path !== null)\n    }\n    case 'skill': {\n      if (!surfaces.skills) return []\n      return [\n        safeJoin(surfaces.skills, subject.name, 'SKILL.md'),\n        safeJoin(surfaces.skills, `${subject.name}.md`),\n      ].filter((path): path is string => path !== null)\n    }\n    case 'tool-doc':\n      if (subject.aspect) {\n        return optionalPath(safeJoin(surfaces.tools, subject.tool, `${slugify(subject.aspect)}.md`))\n      }\n      // tool-doc default: `<tool>/README.md`; also probe `<tool>.md` for flat\n      // tool-list repos.\n      return [\n        safeJoin(surfaces.tools, subject.tool, 'README.md'),\n        safeJoin(surfaces.tools, `${subject.tool}.md`),\n      ].filter((path): path is string => path !== null)\n    case 'new-tool':\n      return optionalPath(safeJoin(surfaces.tools, subject.name, 'README.md'))\n    case 'mcp':\n      if (!surfaces.mcp) return []\n      return subject.tool\n        ? optionalPath(safeJoin(surfaces.mcp, subject.server, `${subject.tool}.md`))\n        : [\n            safeJoin(surfaces.mcp, `${subject.server}.json`),\n            safeJoin(surfaces.mcp, subject.server, 'README.md'),\n          ].filter((path): path is string => path !== null)\n    case 'hook':\n      if (!surfaces.hooks) return []\n      return [\n        safeJoin(surfaces.hooks, `${subject.name}.md`),\n        safeJoin(surfaces.hooks, `${subject.name}.json`),\n      ].filter((path): path is string => path !== null)\n    case 'subagent':\n      if (!surfaces.subagents) return []\n      return [\n        safeJoin(surfaces.subagents, `${subject.name}.md`),\n        safeJoin(surfaces.subagents, `${subject.name}.yaml`),\n        safeJoin(surfaces.subagents, `${subject.name}.json`),\n      ].filter((path): path is string => path !== null)\n    case 'workflow':\n      if (!surfaces.workflows) return []\n      return [\n        safeJoin(surfaces.workflows, `${subject.name}.md`),\n        safeJoin(surfaces.workflows, `${subject.name}.yaml`),\n        safeJoin(surfaces.workflows, `${subject.name}.json`),\n      ].filter((path): path is string => path !== null)\n    case 'rollout-policy':\n      return surfaces.rolloutPolicy ? [surfaces.rolloutPolicy] : []\n    case 'agent-profile':\n      return surfaces.agentProfile ? [surfaces.agentProfile] : []\n    case 'code': {\n      if (!surfaces.code) return []\n      const path = safeJoin(surfaces.code, subject.path)\n      return path ? [path] : []\n    }\n    case 'rag':\n      if (!surfaces.rag) return []\n      return optionalPath(safeJoin(surfaces.rag, subject.corpus, `${subject.docId}.md`))\n    case 'memory':\n      if (!surfaces.memory) return []\n      return optionalPath(safeJoin(surfaces.memory, `${slugify(subject.key)}.json`))\n    case 'scaffolding':\n      if (!surfaces.scaffolding) return []\n      return optionalPath(safeJoin(surfaces.scaffolding, `${slugify(subject.concern)}.md`))\n    case 'output-schema':\n      if (!surfaces.outputSchema) return []\n      return [surfaces.outputSchema]\n    case 'websearch.outdated':\n    case 'prior-run-summary':\n      // Stale signals don't map to a single file — handled by the knowledge\n      // adapter as `agent-knowledge:stale:*` after operator review.\n      return []\n    case 'cluster':\n      // failure-mode cluster labels are evidence, not mutations.\n      return []\n  }\n}\n\nfunction safeJoin(root: string, ...children: string[]): string | null {\n  if (children.some((child) => child.includes('\\0') || isAbsolute(child))) return null\n  const rootAbsolute = resolve(root)\n  const targetAbsolute = resolve(rootAbsolute, ...children)\n  const escaped = relative(rootAbsolute, targetAbsolute)\n  if (escaped === '..' || escaped.startsWith(`..${sep}`) || isAbsolute(escaped)) return null\n  return join(root, ...children)\n}\n\nfunction optionalPath(path: string | null): string[] {\n  return path ? [path] : []\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9-]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 200) || 'untitled'\n  )\n}\n\n/**\n * Validate that every declared surface exists on disk under `repoRoot`.\n *\n * Returns an array of `SurfaceValidationIssue` — empty when all required\n * surfaces resolve. `defineAgent` throws with the issues rendered, so\n * a misconfigured manifest fails at startup (not at the first finding\n * the loop produces 20 minutes later).\n */\nexport interface SurfaceValidationIssue {\n  surface: keyof AgentSurfaces\n  path: string\n  reason: 'missing' | 'not-directory' | 'not-file'\n}\n\n/** Validate an `AgentSurfaces` map on disk — missing paths fail loud at `defineAgent` time instead of silently skipping self-improvement edits. */\nexport function validateSurfaces(\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ReadonlyArray<SurfaceValidationIssue> {\n  const issues: SurfaceValidationIssue[] = []\n  const dirSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'systemPrompt',\n    'tools',\n    'personas',\n    'knowledge',\n  ]\n  const fileSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['rubric']\n  const optionalDirSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'scaffolding',\n    'memory',\n    'rag',\n    'skills',\n    'mcp',\n    'hooks',\n    'subagents',\n    'workflows',\n    'code',\n  ]\n  const optionalFileSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'outputSchema',\n    'rolloutPolicy',\n    'agentProfile',\n  ]\n\n  for (const key of dirSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    } else if (!statSync(abs).isDirectory()) {\n      issues.push({ surface: key, path: p, reason: 'not-directory' })\n    }\n  }\n  for (const key of fileSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    } else if (!statSync(abs).isFile()) {\n      issues.push({ surface: key, path: p, reason: 'not-file' })\n    }\n  }\n  for (const key of [...optionalDirSurfaces, ...optionalFileSurfaces]) {\n    const p = surfaces[key] as string | undefined\n    if (p === undefined) continue\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n      continue\n    }\n    const expectedDirectory = optionalDirSurfaces.includes(key)\n    if (expectedDirectory && !statSync(abs).isDirectory()) {\n      issues.push({ surface: key, path: p, reason: 'not-directory' })\n    } else if (!expectedDirectory && !statSync(abs).isFile()) {\n      issues.push({ surface: key, path: p, reason: 'not-file' })\n    }\n  }\n  return issues\n}\n\n/** Format a list of surface validation issues into a human-readable error string. */\nexport function renderSurfaceIssues(\n  issues: ReadonlyArray<SurfaceValidationIssue>,\n  repoRoot: string,\n): string {\n  if (issues.length === 0) return ''\n  const lines = issues.map(\n    (i) => `  - ${i.surface}: ${i.path ? `\"${i.path}\"` : '<not set>'} (${i.reason})`,\n  )\n  return [\n    `Agent surface validation failed against repoRoot=${repoRoot}:`,\n    ...lines,\n    '',\n    'Fix the manifest: every required surface must point at an existing',\n    'directory (systemPrompt / tools / personas / knowledge) or file',\n    '(rubric). Optional surfaces (scaffolding / memory / rag / outputSchema)',\n    'may be omitted; the loop will reject findings targeting omitted',\n    'surfaces rather than fabricating a path.',\n  ].join('\\n')\n}\n","/**\n * `defineAgent` — typed, validated manifest API for declarative agent\n * configuration. The substrate consumes this manifest to wire the\n * canonical eval pattern + analyst self-improvement loop without any\n * per-vertical glue.\n *\n * Design goal: scale to 1000s of vertical agents. Every agent declares\n * its surfaces, rubric, runtime, and analyst configuration in ~50 lines.\n * No per-vertical `ImprovementProposalSource`. No per-vertical CLI. No\n * fabricated paths.\n *\n * Validation: `defineAgent` runs `validateSurfaces` synchronously and\n * throws a structured error if any required surface is missing on\n * disk. The cost is one filesystem stat per surface (cheap); the\n * benefit is a manifest that can't ship broken.\n *\n * @experimental Manifests validate and load, but `runtime.act` is not yet wired\n * end-to-end into the substrate's eval path (`unimplementedAgentRun` is the\n * shipped default).\n */\n\nimport type { TraceAnalystDefinition } from '@tangle-network/agent-eval'\nimport type { RuntimeStreamEvent } from '../types'\nimport { type AgentSurfaces, renderSurfaceIssues, validateSurfaces } from './surfaces'\n\n// ── manifest ─────────────────────────────────────────────────────────\n\n/**\n * The full agent manifest. Each agent ships ONE of these.\n *\n * Generics:\n *   `TPersona` — the agent's persona shape (loaded from\n *     `surfaces.personas`). Defaults to `unknown` so the substrate's\n *     persona discovery (`loadPersonas`) can accept anything; per-agent\n *     code re-narrows when it matters.\n *   `TRunOutput` — the shape `runtime.act` returns. Used by the rubric\n *     scorers and emitted into the trace.\n */\nexport interface AgentManifest<TPersona = unknown, TRunOutput = unknown> {\n  /**\n   * Stable identifier — used as `projectId` in traces, as the analyst\n   * loop's `runId` prefix, and as the namespace under which findings\n   * are persisted. MUST match the agent's repo name to keep\n   * cross-repo telemetry joinable.\n   */\n  id: string\n\n  /**\n   * Filesystem root the substrate resolves surface paths against.\n   * Typically `process.cwd()` or a fixed absolute path. Use an\n   * absolute path when the agent's tests may run from subdirectories\n   * (vitest sometimes shifts cwd).\n   */\n  repoRoot: string\n\n  /**\n   * Map of mutable surfaces the self-improvement loop can edit. See\n   * `AgentSurfaces` — required: `systemPrompt`, `tools`, `rubric`,\n   * `knowledge`, `personas`. Optional: `scaffolding`, `memory`, `rag`,\n   * `outputSchema`.\n   *\n   * Every required path is validated at `defineAgent` time. Missing\n   * paths throw with the full list of offenders.\n   */\n  surfaces: AgentSurfaces\n\n  /**\n   * Rubric the substrate uses to score each run. Dimensions × weights\n   * × judges. The substrate computes the weighted composite and\n   * stamps it into the RunRecord.\n   */\n  rubric: AgentRubric<TRunOutput>\n\n  /**\n   * Runtime adapter — how the substrate INVOKES the agent against a\n   * persona. The `act` function takes a persona + a context (with the\n   * tracer the substrate threads through for span emission) and\n   * returns the run output the rubric will score.\n   *\n   * The agent's existing production runtime goes in here; the\n   * substrate is intentionally thin around it.\n   */\n  runtime: AgentRuntime<TPersona, TRunOutput>\n\n  /**\n   * Persona discovery — the substrate loads personas via this function\n   * at eval start. Can read from `surfaces.personas`, an API, or be\n   * hardcoded. The substrate calls it once per `runAgentEval` call;\n   * persona ordering is preserved.\n   */\n  personas: () => Promise<ReadonlyArray<TPersona>>\n\n  /**\n   * Analyst kinds the substrate runs against each persona's trace.\n   * Defaults to `DEFAULT_TRACE_ANALYST_KINDS` from agent-eval. Per-agent\n   * authors can prune (e.g. skip `knowledge-poisoning` when there's no\n   * knowledge base) or extend (custom domain kinds).\n   *\n   * Empty array disables the loop — useful for `pnpm eval --no-analyst`.\n   */\n  analystKinds: ReadonlyArray<TraceAnalystDefinition>\n\n  /**\n   * Analyst LLM configuration. The substrate uses these for all four\n   * kinds (override per-kind via `analystKinds` if needed).\n   */\n  analyst: AnalystConfig\n}\n\nexport interface AgentRubric<TRunOutput> {\n  /** Dimensions composing the weighted score. Weights sum to 1.0 by convention. */\n  dimensions: ReadonlyArray<RubricDimension<TRunOutput>>\n  /**\n   * Optional judges layered on top of deterministic dimensions. Each\n   * judge returns a score per dimension; the substrate averages judges\n   * (mean by default) for the LLM contribution.\n   */\n  judges?: ReadonlyArray<JudgeConfig<TRunOutput>>\n}\n\nexport interface RubricDimension<TRunOutput> {\n  /** Unique identifier — appears in finding subjects (`rubric:<id>`). */\n  id: string\n  /** 0..1 — weight in the composite. */\n  weight: number\n  /**\n   * Deterministic scorer: given the persona + run output, returns a\n   * 0..1 score. The substrate sums weight × score across dimensions\n   * for the deterministic composite; judges supplement subjective dims.\n   */\n  score: (input: { persona: unknown; output: TRunOutput }) => number\n  /** Optional human-readable label for reports. */\n  label?: string\n}\n\nexport interface JudgeConfig<TRunOutput> {\n  /** Judge identifier — appears in trace spans + manifest. */\n  id: string\n  /** Model snapshot to invoke. Pin the snapshot (`claude-sonnet-4-6@2025-04-15`); the validator rejects bare aliases. */\n  model: string\n  /** Dimensions this judge scores. */\n  dimensions: ReadonlyArray<string>\n  /**\n   * Optional rubric anchors — text examples the judge sees as a\n   * few-shot prompt to calibrate. STRONGLY recommended for subjective\n   * dimensions; required by the calibration gate (Pearson ≥0.7).\n   */\n  anchors?: ReadonlyArray<{ input: string; output: TRunOutput; expected: Record<string, number> }>\n}\n\nexport interface AgentRuntime<TPersona, TRunOutput> {\n  /**\n   * Invoke the agent against one persona. Returns BOTH:\n   *   - `events`: an `AsyncIterable<RuntimeStreamEvent>` the chat-centric\n   *     product consumes verbatim (SSE / WebSocket / inline render).\n   *     **Streaming is mandatory — never collapse this to a single Promise.**\n   *     The agent's existing `runChatTurn` (or equivalent async generator)\n   *     plugs in here directly.\n   *   - `output`: a `Promise<TRunOutput>` resolved AFTER the event stream\n   *     drains. The eval substrate awaits this for rubric scoring; chat\n   *     products usually ignore it (they already rendered incrementally).\n   *\n   * Implementation contract:\n   *   1. `act` MUST return immediately (synchronous construction of the\n   *      `events` iterator + the `output` promise).\n   *   2. Iterating `events` drives the underlying LLM/tool calls — the\n   *      caller chooses when to consume.\n   *   3. `output` resolves only after the iterator yields its terminal\n   *      event (typically `task_end`); see `collectAgentRun` helper.\n   *\n   * `ctx.emitter` is the substrate-threaded `TraceEmitter` — runtimes\n   * SHOULD record LLM/tool spans through it for capture integrity.\n   * `ctx.deadlineMs` is wall-clock; the runtime SHOULD honour for graceful\n   * cancel. `ctx.signal` is the standard abort signal.\n   */\n  act: (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput>\n}\n\nexport interface AgentRunInvocation<TRunOutput> {\n  /** Live stream of typed runtime events. Consumed by chat UX directly. */\n  events: AsyncIterable<RuntimeStreamEvent>\n  /** Final structured output the rubric scores. Resolves after `events` drains. */\n  output: Promise<TRunOutput>\n}\n\n/**\n * Stub for agents whose `runtime.act` is not yet wired to the substrate's\n * eval path. Preserves the streaming contract (empty event stream + a\n * rejected `output` promise that tells the caller exactly what to fix).\n *\n * Per-vertical manifests usually start with this stub and replace it with\n * the agent's real streaming runtime (`runChatTurn` or equivalent) once\n * the eval path consumes the manifest end-to-end.\n */\nexport function unimplementedAgentRun<TRunOutput = unknown>(\n  reason = 'AgentRuntime.act is not yet wired for this manifest',\n): AgentRunInvocation<TRunOutput> {\n  return {\n    events: (async function* empty(): AsyncIterable<RuntimeStreamEvent> {})(),\n    output: Promise.reject(new Error(reason)),\n  }\n}\n\n/**\n * Drain `act`'s `events` into an array AND await its `output`. Useful for\n * eval / outcome-measurement code paths that don't care about live\n * rendering. The events array is preserved so the substrate can inspect\n * tool calls / readiness / questions retrospectively.\n *\n * IMPORTANT: chat-centric UX MUST NOT call this — it defeats streaming\n * (no incremental render). Use `for await (const ev of invocation.events)`\n * directly in the chat surface.\n */\nexport async function collectAgentRun<TRunOutput>(\n  invocation: AgentRunInvocation<TRunOutput>,\n): Promise<{ events: ReadonlyArray<RuntimeStreamEvent>; output: TRunOutput }> {\n  const events: RuntimeStreamEvent[] = []\n  for await (const ev of invocation.events) events.push(ev)\n  const output = await invocation.output\n  return { events, output }\n}\n\nexport interface AgentRunContext {\n  /** Substrate-managed trace emitter. */\n  emitter: import('@tangle-network/agent-eval').TraceEmitter\n  /** Stable run id for this persona × variant cell. */\n  runId: string\n  /** Variant the runtime is exercising (e.g. `'baseline'`, `'source-grounded'`). */\n  variantId?: string\n  /** Wall-clock deadline (epoch ms). The runtime SHOULD honour for graceful cancel. */\n  deadlineMs?: number\n  /** Optional abort signal. */\n  signal?: AbortSignal\n}\n\nexport interface AnalystConfig {\n  /** Model the analyst kinds use. Override per-kind via `analystKinds[i].cost.models`. */\n  model: string\n  /** Optional total budget across all kinds for one run. Substrate enforces via `BudgetGuard`. */\n  budgetUsd?: number\n  /** Backend hint for the AxAIService factory — same shape every kind uses. */\n  backend?: {\n    name?: 'openai' | 'router'\n    apiKey?: string\n    baseUrl?: string\n  }\n}\n\n// ── factory + validation ─────────────────────────────────────────────\n\n/** Thrown when `defineAgent` finds a required surface missing on disk. */\nexport class AgentManifestError extends Error {\n  constructor(\n    message: string,\n    public readonly agentId: string,\n    public readonly issues: ReadonlyArray<unknown> = [],\n  ) {\n    super(message)\n    this.name = 'AgentManifestError'\n  }\n}\n\n/**\n * Construct a validated agent manifest. Throws `AgentManifestError`\n * if any required surface is missing on disk.\n *\n * Generics: pass your persona / output types if you want narrowed\n * `runtime.act` signatures:\n *   `defineAgent<TaxPersona, TaxRunOutput>({ ... })`\n *\n * Most callers don't need the generics — the substrate operates on\n * `unknown` payloads internally and the manifest's `score` /\n * `runtime.act` see the typed shapes via TypeScript inference at\n * the call site.\n */\nexport function defineAgent<TPersona = unknown, TRunOutput = unknown>(\n  manifest: AgentManifest<TPersona, TRunOutput>,\n): AgentManifest<TPersona, TRunOutput> {\n  if (!manifest.id || manifest.id.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `id` is required', manifest.id ?? '')\n  }\n  if (!manifest.repoRoot || manifest.repoRoot.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `repoRoot` is required', manifest.id)\n  }\n  const issues = validateSurfaces(manifest.surfaces, manifest.repoRoot)\n  if (issues.length > 0) {\n    throw new AgentManifestError(\n      renderSurfaceIssues(issues, manifest.repoRoot),\n      manifest.id,\n      issues,\n    )\n  }\n  // Lightweight rubric sanity: weights sum to ~1.0 (no hard requirement —\n  // the substrate normalizes — but flag wildly miscalibrated weights).\n  const total = manifest.rubric.dimensions.reduce((acc, d) => acc + d.weight, 0)\n  if (manifest.rubric.dimensions.length > 0 && (total < 0.5 || total > 1.5)) {\n    throw new AgentManifestError(\n      `defineAgent(${manifest.id}): rubric dimension weights sum to ${total.toFixed(3)} — should be ~1.0`,\n      manifest.id,\n    )\n  }\n  return manifest\n}\n","/**\n * Surface improvement proposer — resolves analyst findings into LLM-drafted\n * candidate patches without changing the caller's repository.\n *\n * The proposer parses each finding's `subject` via\n * `parseFindingSubject` (agent-eval), resolves it to a real file path\n * via the agent's `AgentSurfaces`, reads the current content, and asks\n * an LLM to draft a unified-diff patch given the finding + current\n * content + per-kind editing-discipline rules.\n *\n * Fail-loud rules:\n *   - Findings whose subject doesn't parse → counted in `errors`.\n *   - Findings whose subject targets an undeclared surface → counted in\n *     `errors` with the offending kind in the message.\n *   - Findings whose target path doesn't exist AND the kind isn't a\n *     create-new variant (`new-tool`, `knowledge.wiki`) → counted in\n *     `errors` with the resolved path in the message.\n *   - LLM drafts that fail JSON-schema validation → counted in\n *     `errors` with the schema issue.\n *\n * No silent skips. Every dropped finding has a recorded reason the\n * loop's report surfaces.\n */\n\nimport { readFileSync } from 'node:fs'\nimport type { AnalystFinding, FindingSubject } from '@tangle-network/agent-eval'\nimport { parseFindingSubject } from '@tangle-network/agent-eval/analyst'\nimport type { ImprovementProposalSource } from '../analyst-loop/types'\nimport { sha256 } from '../content-digest'\nimport type { AgentSurfaces, ResolvedSurface } from './surfaces'\nimport { resolveSubjectPath } from './surfaces'\n\n// ── proposal shape ───────────────────────────────────────────────────\n\nexport interface SurfaceImprovementEdit {\n  /** Stable id derived from the source finding so re-proposals are idempotent. */\n  id: string\n  /** The finding that produced this edit — for revert + audit trail. */\n  sourceFindingId: string\n  /** Parsed subject; included so the apply step doesn't re-parse. */\n  subject: FindingSubject\n  /** Resolved on-disk target. */\n  target: ResolvedSurface\n  /** SHA-256 of the current file content the patch was drafted against. */\n  baseSha256: string\n  /** Unified-diff patch the LLM drafted (relative to `target.absolutePath`). */\n  patch: string\n  /** One-line summary the operator sees in the report / PR title. */\n  summary: string\n  /** Multi-line rationale for the PR body — finding context + LLM reasoning. */\n  rationale: string\n  /** Carry-forward from the finding so the apply gate can check the threshold. */\n  confidence: number\n  /** Carry-forward severity for prioritization. */\n  severity: AnalystFinding['severity']\n}\n\nexport interface CreateSurfaceImprovementProposerOptions {\n  surfaces: AgentSurfaces\n  repoRoot: string\n  /**\n   * LLM-draft callback. Given a finding + current file content + the\n   * resolved target, returns a unified-diff patch + summary + rationale.\n   *\n   * Required — the substrate doesn't ship a hardcoded prompt; the agent\n   * author picks the model (Haiku for cheap routine drafts, Sonnet for\n   * substantive prompt rewrites, etc.) via this callback.\n   */\n  draftPatch: (input: DraftPatchInput) => Promise<DraftPatchOutput>\n  /**\n   * When the resolved target doesn't exist, allow the substrate to\n   * CREATE the file (for `knowledge.wiki`, `new-tool` subjects). Default\n   * true for those kinds, false for `system-prompt` / `rubric` / etc.\n   * (named sections that don't exist are a contract violation, not a\n   * scaffolding opportunity).\n   */\n  allowCreateForKinds?: ReadonlyArray<FindingSubject['kind']>\n}\n\nexport interface DraftPatchInput {\n  finding: AnalystFinding\n  subject: FindingSubject\n  target: ResolvedSurface\n  /** Current file content (empty string when `intent === 'create-new'`). */\n  currentContent: string\n}\n\nexport interface DraftPatchOutput {\n  /** Unified diff against the current file content. Empty string skips this finding. */\n  patch: string\n  /** One-line summary for the operator. */\n  summary: string\n  /** Multi-line rationale for the PR body. */\n  rationale: string\n}\n\n// ── factory ──────────────────────────────────────────────────────────\n\nconst DEFAULT_CREATE_KINDS: ReadonlyArray<FindingSubject['kind']> = [\n  'knowledge.wiki',\n  'knowledge.claim',\n  'knowledge.raw',\n  'new-tool',\n]\n\n/** Resolve each finding to a real surface and draft a detached patch candidate. */\nexport function createSurfaceImprovementProposer(\n  opts: CreateSurfaceImprovementProposerOptions,\n): ImprovementProposalSource<SurfaceImprovementEdit> {\n  const allowCreate = opts.allowCreateForKinds ?? DEFAULT_CREATE_KINDS\n\n  return {\n    async proposeFromFindings(findings) {\n      const edits: SurfaceImprovementEdit[] = []\n      const errors: Array<{ findingId: string; subject: string; message: string }> = []\n      let skipped = 0\n\n      for (const f of findings) {\n        const subject = parseFindingSubject(f.subject)\n        if (subject === null) {\n          if (f.subject !== undefined) {\n            errors.push({\n              findingId: f.finding_id,\n              subject: f.subject,\n              message: 'subject does not parse against the finding-subject grammar',\n            })\n          } else {\n            // Subject-less findings are descriptive, not actionable —\n            // legitimate; count in `skipped` not `errors`.\n            skipped += 1\n          }\n          continue\n        }\n\n        // `cluster` findings (failure-mode) are evidence, not mutations.\n        if (subject.kind === 'cluster') {\n          skipped += 1\n          continue\n        }\n\n        // Knowledge findings flow to the knowledge proposal source so they do not double-route.\n        if (subject.kind.startsWith('knowledge.')) {\n          skipped += 1\n          continue\n        }\n\n        const target = resolveSubjectPath(subject, opts.surfaces, opts.repoRoot)\n        if (target === null) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `subject kind \"${subject.kind}\" targets an undeclared surface; declare it in AgentSurfaces or stop emitting this subject`,\n          })\n          continue\n        }\n\n        if (target.intent === 'create-new' && !allowCreate.includes(subject.kind)) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `target ${target.repoRelativePath} does not exist; the kind \"${subject.kind}\" requires an existing target (analyst named a section that isn't in the codebase)`,\n          })\n          continue\n        }\n\n        const currentContent = target.exists ? readFileSync(target.absolutePath, 'utf-8') : ''\n\n        let draft: DraftPatchOutput\n        try {\n          draft = await opts.draftPatch({ finding: f, subject, target, currentContent })\n        } catch (err) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `draftPatch threw: ${err instanceof Error ? err.message : String(err)}`,\n          })\n          continue\n        }\n\n        if (draft.patch.trim().length === 0) {\n          skipped += 1\n          continue\n        }\n\n        edits.push({\n          id: `imp-${f.finding_id}`,\n          sourceFindingId: f.finding_id,\n          subject,\n          target,\n          baseSha256: sha256(currentContent),\n          patch: draft.patch,\n          summary: draft.summary,\n          rationale: draft.rationale,\n          confidence: f.confidence,\n          severity: f.severity,\n        })\n      }\n\n      return { edits, skipped, errors }\n    },\n  }\n}\n","/**\n * Sandbox bridge for `AgentRuntime.act` — prod-faithful eval execution.\n *\n * The point of this adapter is parity: the eval substrate must run the agent\n * through the SAME profile the production chat turn uses, or scorecard numbers\n * grade a profile that never ships. `createSandboxAct` boots a sandbox with the\n * agent's profile (the caller's `baseProfile`, with optional per-persona\n * overrides) through the loop kernel's own {@link createSandboxForSpec},\n * streams the `streamPrompt` events mapped to the `RuntimeStreamEvent`\n * vocabulary, and resolves the `OutputAdapter`-parsed output for rubric scoring\n * — satisfying the `act` streaming contract with one code path shared by chat\n * and eval.\n *\n * Agents with a bespoke streaming chat turn should wire THAT into `act`\n * directly (the contract is designed for it); this adapter is the default for\n * agents whose turn is a plain prod-profile sandbox dispatch — notably the\n * agents agent-builder generates.\n */\n\nimport type {\n  AgentProfile,\n  AgentProfileFileMount,\n  AgentProfileMcpServer,\n} from '@tangle-network/agent-interface'\nimport type { SandboxEvent } from '@tangle-network/sandbox'\nimport type { AgentRunSpec, OutputAdapter, SandboxClient } from '../runtime'\nimport { mapSandboxEvent } from '../runtime'\nimport { createSandboxForSpec } from '../runtime/iteration-sandboxes'\nimport type { RuntimeStreamEvent } from '../types'\nimport type { AgentRunContext, AgentRunInvocation } from './define-agent'\nimport {\n  type AgentProfileMaterializationAxis,\n  assertProfileMaterialization,\n  sandboxActProfileMaterialization,\n} from './profile-materialization'\n\n/** Per-persona profile-merge slots applied over the base profile (§1.5: the caller authors the\n *  per-persona profile). Each slot overlays the base; an absent slot leaves the base untouched. */\nexport interface SandboxActComposeOverrides {\n  /** Replace the base profile's system prompt (e.g. a workspace-augmented prompt). */\n  systemPrompt?: string\n  /** Extra file mounts layered after the base profile's `resources.files`. */\n  extraFiles?: AgentProfileFileMount[]\n  /** Override the profile `name`. Defaults to the base profile's name. */\n  name?: string\n  /** Box built-in tool ON/OFF flags merged over the base profile's `tools` (overlay wins per key). */\n  tools?: Record<string, boolean>\n  /** MCP connections merged over the base profile's `mcp` (overlay wins per key). */\n  mcpConnections?: Record<string, AgentProfileMcpServer>\n}\n\nexport interface CreateSandboxActOptions<TPersona, TRunOutput> {\n  /** Canonical agent profile — the same one the prod chat turn uses. */\n  baseProfile: AgentProfile\n  /** Sandbox client used to boot the per-run sandbox. */\n  sandboxClient: SandboxClient\n  /** Persona → prompt. Pure; the eval cell's input. */\n  buildPrompt: (persona: TPersona) => string\n  /** Sandbox event stream → typed output the rubric scores. */\n  output: OutputAdapter<TRunOutput>\n  /**\n   * Per-persona profile overrides (workspace-augmented system prompt, extra\n   * file mounts, tool flags, MCP connections). Overlaid onto `baseProfile`.\n   */\n  compose?: (persona: TPersona) => SandboxActComposeOverrides\n  /** Sandbox-SDK overrides forwarded to `createSandboxForSpec`. */\n  sandboxOverrides?: AgentRunSpec<unknown>['sandboxOverrides']\n  /** Optional changed axes the caller expects this path to carry. */\n  requiredProfileAxes?: readonly AgentProfileMaterializationAxis[]\n  /** Stable run name surfaced in mapped `llm_call` events. */\n  name?: string\n  /** Override the `SandboxEvent → RuntimeStreamEvent` mapper. */\n  mapEvent?: (\n    event: SandboxEvent,\n    opts: { agentRunName?: string },\n  ) => RuntimeStreamEvent | undefined\n}\n\n/**\n * Build an `AgentRuntime.act` implementation backed by a single prod-profile\n * sandbox run. The returned function honours the `act` contract: it returns\n * synchronously with a live `events` iterator and an `output` promise that\n * resolves only after the iterator drains.\n */\nexport function createSandboxAct<TPersona, TRunOutput>(\n  options: CreateSandboxActOptions<TPersona, TRunOutput>,\n): (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput> {\n  assertProfileMaterialization({\n    contract: sandboxActProfileMaterialization,\n    changedAxes: options.requiredProfileAxes ?? [],\n    context: 'createSandboxAct',\n  })\n  const mapEvent = options.mapEvent ?? mapSandboxEvent\n\n  return (persona: TPersona, ctx: AgentRunContext): AgentRunInvocation<TRunOutput> => {\n    const profile = applyComposeOverrides(options.baseProfile, options.compose?.(persona))\n    const agentRunName = options.name ?? profile.name ?? 'agent'\n    const message = options.buildPrompt(persona)\n    const signal = ctx.signal ?? new AbortController().signal\n\n    const raw: SandboxEvent[] = []\n    let settle!: (value: TRunOutput) => void\n    let fail!: (err: unknown) => void\n    const output = new Promise<TRunOutput>((resolve, reject) => {\n      settle = resolve\n      fail = reject\n    })\n    // The output promise rejects when the stream errors; if the caller ignores\n    // `output` (chat UX) the rejection is still observed by the events iterator\n    // throwing. Attach a no-op catch so an ignored rejection is never \"unhandled\".\n    output.catch(() => {})\n\n    const spec: AgentRunSpec<unknown> = {\n      profile,\n      taskToPrompt: () => message,\n      name: agentRunName,\n      ...(options.sandboxOverrides ? { sandboxOverrides: options.sandboxOverrides } : {}),\n    }\n\n    async function* events(): AsyncIterable<RuntimeStreamEvent> {\n      try {\n        const box = await createSandboxForSpec(options.sandboxClient, spec, signal)\n        for await (const event of box.streamPrompt(message, { signal })) {\n          raw.push(event)\n          const mapped = mapEvent(event, { agentRunName })\n          if (mapped) yield mapped\n        }\n        settle(options.output.parse(raw))\n      } catch (err) {\n        fail(err)\n        throw err\n      }\n    }\n\n    return { events: events(), output }\n  }\n}\n\n/** Overlay the per-persona overrides onto the base profile. Each slot merges over the base; an\n *  absent override leaves the base profile untouched. */\nfunction applyComposeOverrides(\n  base: AgentProfile,\n  overrides: SandboxActComposeOverrides | undefined,\n): AgentProfile {\n  if (!overrides) return base\n  const prompt = overrides.systemPrompt\n    ? { ...base.prompt, systemPrompt: overrides.systemPrompt }\n    : base.prompt\n  const mergedTools = overrides.tools ? { ...(base.tools ?? {}), ...overrides.tools } : base.tools\n  const mergedMcp = overrides.mcpConnections\n    ? { ...(base.mcp ?? {}), ...overrides.mcpConnections }\n    : base.mcp\n  const baseFiles = base.resources?.files ?? []\n  const mergedFiles: AgentProfileFileMount[] = overrides.extraFiles?.length\n    ? [...baseFiles, ...overrides.extraFiles]\n    : [...baseFiles]\n  return {\n    ...base,\n    name: overrides.name ?? base.name,\n    prompt,\n    ...(mergedTools ? { tools: mergedTools } : {}),\n    ...(mergedMcp ? { mcp: mergedMcp } : {}),\n    resources: { ...base.resources, files: mergedFiles },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAqGA,SAAgB,mBACd,SACA,UACA,UACwB;CACxB,MAAM,aAAa,yBAAyB,SAAS,QAAQ;CAC7D,IAAI,WAAW,WAAW,GAAG,OAAO;CAMpC,KAAK,MAAM,OAAO,YAAY;EAC5B,MAAM,MAAM,WAAW,GAAG,IAAI,MAAM,KAAK,UAAU,GAAG;EACtD,IAAI,WAAW,GAAG,GAChB,OAAO;GAAE,cAAc;GAAK,kBAAkB;GAAK,QAAQ;GAAM,QAAQ;EAAgB;CAE7F;CACA,MAAM,WAAW,WAAW;CAE5B,OAAO;EACL,cAFkB,WAAW,QAAQ,IAAI,WAAW,KAAK,UAAU,QAAQ;EAG3E,kBAAkB;EAClB,QAAQ;EACR,QAAQ;CACV;AACF;AAEA,SAAS,yBACP,SACA,UACuB;CACvB,QAAQ,QAAQ,MAAhB;EACE,KAAK;EACL,KAAK,mBACH,OAAO,aAAa,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,IAAI,CAAC;EACxE,KAAK,mBAEH,OAAO,aAAa,SAAS,SAAS,WAAW,UAAU,GAAG,QAAQ,QAAQ,KAAK,EAAE,IAAI,CAAC;EAC5F,KAAK,iBACH,OAAO,aAAa,SAAS,SAAS,WAAW,OAAO,GAAG,QAAQ,SAAS,IAAI,CAAC;EACnF,KAAK,iBAAiB;GACpB,MAAM,OAAO,QAAQ,QAAQ,OAAO;GAIpC,OAAO;IACL,SAAS,SAAS,cAAc,GAAG,KAAK,IAAI;IAC5C,SAAS,SAAS,cAAc,MAAM,UAAU;IAChD,SAAS,SAAS,cAAc,MAAM,UAAU;GAClD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD;EACA,KAAK;GACH,IAAI,CAAC,SAAS,QAAQ,OAAO,CAAC;GAC9B,OAAO,CACL,SAAS,SAAS,QAAQ,QAAQ,MAAM,UAAU,GAClD,SAAS,SAAS,QAAQ,GAAG,QAAQ,KAAK,IAAI,CAChD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAElD,KAAK;GACH,IAAI,QAAQ,QACV,OAAO,aAAa,SAAS,SAAS,OAAO,QAAQ,MAAM,GAAG,QAAQ,QAAQ,MAAM,EAAE,IAAI,CAAC;GAI7F,OAAO,CACL,SAAS,SAAS,OAAO,QAAQ,MAAM,WAAW,GAClD,SAAS,SAAS,OAAO,GAAG,QAAQ,KAAK,IAAI,CAC/C,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK,YACH,OAAO,aAAa,SAAS,SAAS,OAAO,QAAQ,MAAM,WAAW,CAAC;EACzE,KAAK;GACH,IAAI,CAAC,SAAS,KAAK,OAAO,CAAC;GAC3B,OAAO,QAAQ,OACX,aAAa,SAAS,SAAS,KAAK,QAAQ,QAAQ,GAAG,QAAQ,KAAK,IAAI,CAAC,IACzE,CACE,SAAS,SAAS,KAAK,GAAG,QAAQ,OAAO,MAAM,GAC/C,SAAS,SAAS,KAAK,QAAQ,QAAQ,WAAW,CACpD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EACtD,KAAK;GACH,IAAI,CAAC,SAAS,OAAO,OAAO,CAAC;GAC7B,OAAO,CACL,SAAS,SAAS,OAAO,GAAG,QAAQ,KAAK,IAAI,GAC7C,SAAS,SAAS,OAAO,GAAG,QAAQ,KAAK,MAAM,CACjD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK;GACH,IAAI,CAAC,SAAS,WAAW,OAAO,CAAC;GACjC,OAAO;IACL,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,IAAI;IACjD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;IACnD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;GACrD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK;GACH,IAAI,CAAC,SAAS,WAAW,OAAO,CAAC;GACjC,OAAO;IACL,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,IAAI;IACjD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;IACnD,SAAS,SAAS,WAAW,GAAG,QAAQ,KAAK,MAAM;GACrD,CAAC,CAAC,QAAQ,SAAyB,SAAS,IAAI;EAClD,KAAK,kBACH,OAAO,SAAS,gBAAgB,CAAC,SAAS,aAAa,IAAI,CAAC;EAC9D,KAAK,iBACH,OAAO,SAAS,eAAe,CAAC,SAAS,YAAY,IAAI,CAAC;EAC5D,KAAK,QAAQ;GACX,IAAI,CAAC,SAAS,MAAM,OAAO,CAAC;GAC5B,MAAM,OAAO,SAAS,SAAS,MAAM,QAAQ,IAAI;GACjD,OAAO,OAAO,CAAC,IAAI,IAAI,CAAC;EAC1B;EACA,KAAK;GACH,IAAI,CAAC,SAAS,KAAK,OAAO,CAAC;GAC3B,OAAO,aAAa,SAAS,SAAS,KAAK,QAAQ,QAAQ,GAAG,QAAQ,MAAM,IAAI,CAAC;EACnF,KAAK;GACH,IAAI,CAAC,SAAS,QAAQ,OAAO,CAAC;GAC9B,OAAO,aAAa,SAAS,SAAS,QAAQ,GAAG,QAAQ,QAAQ,GAAG,EAAE,MAAM,CAAC;EAC/E,KAAK;GACH,IAAI,CAAC,SAAS,aAAa,OAAO,CAAC;GACnC,OAAO,aAAa,SAAS,SAAS,aAAa,GAAG,QAAQ,QAAQ,OAAO,EAAE,IAAI,CAAC;EACtF,KAAK;GACH,IAAI,CAAC,SAAS,cAAc,OAAO,CAAC;GACpC,OAAO,CAAC,SAAS,YAAY;EAC/B,KAAK;EACL,KAAK,qBAGH,OAAO,CAAC;EACV,KAAK,WAEH,OAAO,CAAC;CACZ;AACF;AAEA,SAAS,SAAS,MAAc,GAAG,UAAmC;CACpE,IAAI,SAAS,MAAM,UAAU,MAAM,SAAS,IAAI,KAAK,WAAW,KAAK,CAAC,GAAG,OAAO;CAChF,MAAM,eAAe,QAAQ,IAAI;CACjC,MAAM,iBAAiB,QAAQ,cAAc,GAAG,QAAQ;CACxD,MAAM,UAAU,SAAS,cAAc,cAAc;CACrD,IAAI,YAAY,QAAQ,QAAQ,WAAW,KAAK,KAAK,KAAK,WAAW,OAAO,GAAG,OAAO;CACtF,OAAO,KAAK,MAAM,GAAG,QAAQ;AAC/B;AAEA,SAAS,aAAa,MAA+B;CACnD,OAAO,OAAO,CAAC,IAAI,IAAI,CAAC;AAC1B;AAEA,SAAS,QAAQ,GAAmB;CAClC,OACE,EACG,YAAY,CAAC,CACb,QAAQ,gBAAgB,GAAG,CAAC,CAC5B,QAAQ,YAAY,EAAE,CAAC,CACvB,MAAM,GAAG,GAAG,KAAK;AAExB;;AAiBA,SAAgB,iBACd,UACA,UACuC;CACvC,MAAM,SAAmC,CAAC;CAC1C,MAAM,cAAkD;EACtD;EACA;EACA;EACA;CACF;CACA,MAAM,eAAmD,CAAC,QAAQ;CAClE,MAAM,sBAA0D;EAC9D;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;CACF;CACA,MAAM,uBAA2D;EAC/D;EACA;EACA;CACF;CAEA,KAAK,MAAM,OAAO,aAAa;EAC7B,MAAM,IAAI,SAAS;EACnB,IAAI,CAAC,GAAG;GACN,OAAO,KAAK;IAAE,SAAS;IAAK,MAAM;IAAI,QAAQ;GAAU,CAAC;GACzD;EACF;EACA,MAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;EAChD,IAAI,CAAC,WAAW,GAAG,GACjB,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAU,CAAC;OACnD,IAAI,CAAC,SAAS,GAAG,CAAC,CAAC,YAAY,GACpC,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAgB,CAAC;CAElE;CACA,KAAK,MAAM,OAAO,cAAc;EAC9B,MAAM,IAAI,SAAS;EACnB,IAAI,CAAC,GAAG;GACN,OAAO,KAAK;IAAE,SAAS;IAAK,MAAM;IAAI,QAAQ;GAAU,CAAC;GACzD;EACF;EACA,MAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;EAChD,IAAI,CAAC,WAAW,GAAG,GACjB,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAU,CAAC;OACnD,IAAI,CAAC,SAAS,GAAG,CAAC,CAAC,OAAO,GAC/B,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAW,CAAC;CAE7D;CACA,KAAK,MAAM,OAAO,CAAC,GAAG,qBAAqB,GAAG,oBAAoB,GAAG;EACnE,MAAM,IAAI,SAAS;EACnB,IAAI,MAAM,KAAA,GAAW;EACrB,MAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;EAChD,IAAI,CAAC,WAAW,GAAG,GAAG;GACpB,OAAO,KAAK;IAAE,SAAS;IAAK,MAAM;IAAG,QAAQ;GAAU,CAAC;GACxD;EACF;EACA,MAAM,oBAAoB,oBAAoB,SAAS,GAAG;EAC1D,IAAI,qBAAqB,CAAC,SAAS,GAAG,CAAC,CAAC,YAAY,GAClD,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAgB,CAAC;OACzD,IAAI,CAAC,qBAAqB,CAAC,SAAS,GAAG,CAAC,CAAC,OAAO,GACrD,OAAO,KAAK;GAAE,SAAS;GAAK,MAAM;GAAG,QAAQ;EAAW,CAAC;CAE7D;CACA,OAAO;AACT;;AAGA,SAAgB,oBACd,QACA,UACQ;CACR,IAAI,OAAO,WAAW,GAAG,OAAO;CAChC,MAAM,QAAQ,OAAO,KAClB,MAAM,OAAO,EAAE,QAAQ,IAAI,EAAE,OAAO,IAAI,EAAE,KAAK,KAAK,YAAY,IAAI,EAAE,OAAO,EAChF;CACA,OAAO;EACL,oDAAoD,SAAS;EAC7D,GAAG;EACH;EACA;EACA;EACA;EACA;EACA;CACF,CAAC,CAAC,KAAK,IAAI;AACb;;;;;;;;;;;ACxKA,SAAgB,sBACd,SAAS,uDACuB;CAChC,OAAO;EACL,SAAS,gBAAgB,QAA2C,CAAC,EAAA,CAAG;EACxE,QAAQ,QAAQ,OAAO,IAAI,MAAM,MAAM,CAAC;CAC1C;AACF;;AAkDA,IAAa,qBAAb,cAAwC,MAAM;CAG1B;CACA;CAHlB,YACE,SACA,SACA,SAAiD,CAAC,GAClD;EACA,MAAM,OAAO;EAHG,KAAA,UAAA;EACA,KAAA,SAAA;EAGhB,KAAK,OAAO;CACd;AACF;;;;;;;;;;;;;;AAeA,SAAgB,YACd,UACqC;CACrC,IAAI,CAAC,SAAS,MAAM,SAAS,GAAG,KAAK,CAAC,CAAC,WAAW,GAChD,MAAM,IAAI,mBAAmB,iCAAiC,SAAS,MAAM,EAAE;CAEjF,IAAI,CAAC,SAAS,YAAY,SAAS,SAAS,KAAK,CAAC,CAAC,WAAW,GAC5D,MAAM,IAAI,mBAAmB,uCAAuC,SAAS,EAAE;CAEjF,MAAM,SAAS,iBAAiB,SAAS,UAAU,SAAS,QAAQ;CACpE,IAAI,OAAO,SAAS,GAClB,MAAM,IAAI,mBACR,oBAAoB,QAAQ,SAAS,QAAQ,GAC7C,SAAS,IACT,MACF;CAIF,MAAM,QAAQ,SAAS,OAAO,WAAW,QAAQ,KAAK,MAAM,MAAM,EAAE,QAAQ,CAAC;CAC7E,IAAI,SAAS,OAAO,WAAW,SAAS,MAAM,QAAQ,MAAO,QAAQ,MACnE,MAAM,IAAI,mBACR,eAAe,SAAS,GAAG,qCAAqC,MAAM,QAAQ,CAAC,EAAE,oBACjF,SAAS,EACX;CAEF,OAAO;AACT;;;;;;;;;;;;;;;;;;;;;;;;;AC5MA,MAAM,uBAA8D;CAClE;CACA;CACA;CACA;AACF;;AAGA,SAAgB,iCACd,MACmD;CACnD,MAAM,cAAc,KAAK,uBAAuB;CAEhD,OAAO,EACL,MAAM,oBAAoB,UAAU;EAClC,MAAM,QAAkC,CAAC;EACzC,MAAM,SAAyE,CAAC;EAChF,IAAI,UAAU;EAEd,KAAK,MAAM,KAAK,UAAU;GACxB,MAAM,UAAU,oBAAoB,EAAE,OAAO;GAC7C,IAAI,YAAY,MAAM;IACpB,IAAI,EAAE,YAAY,KAAA,GAChB,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE;KACX,SAAS;IACX,CAAC;SAID,WAAW;IAEb;GACF;GAGA,IAAI,QAAQ,SAAS,WAAW;IAC9B,WAAW;IACX;GACF;GAGA,IAAI,QAAQ,KAAK,WAAW,YAAY,GAAG;IACzC,WAAW;IACX;GACF;GAEA,MAAM,SAAS,mBAAmB,SAAS,KAAK,UAAU,KAAK,QAAQ;GACvE,IAAI,WAAW,MAAM;IACnB,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE,WAAW;KACtB,SAAS,iBAAiB,QAAQ,KAAK;IACzC,CAAC;IACD;GACF;GAEA,IAAI,OAAO,WAAW,gBAAgB,CAAC,YAAY,SAAS,QAAQ,IAAI,GAAG;IACzE,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE,WAAW;KACtB,SAAS,UAAU,OAAO,iBAAiB,6BAA6B,QAAQ,KAAK;IACvF,CAAC;IACD;GACF;GAEA,MAAM,iBAAiB,OAAO,SAAS,aAAa,OAAO,cAAc,OAAO,IAAI;GAEpF,IAAI;GACJ,IAAI;IACF,QAAQ,MAAM,KAAK,WAAW;KAAE,SAAS;KAAG;KAAS;KAAQ;IAAe,CAAC;GAC/E,SAAS,KAAK;IACZ,OAAO,KAAK;KACV,WAAW,EAAE;KACb,SAAS,EAAE,WAAW;KACtB,SAAS,qBAAqB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;IAC/E,CAAC;IACD;GACF;GAEA,IAAI,MAAM,MAAM,KAAK,CAAC,CAAC,WAAW,GAAG;IACnC,WAAW;IACX;GACF;GAEA,MAAM,KAAK;IACT,IAAI,OAAO,EAAE;IACb,iBAAiB,EAAE;IACnB;IACA;IACA,YAAY,OAAO,cAAc;IACjC,OAAO,MAAM;IACb,SAAS,MAAM;IACf,WAAW,MAAM;IACjB,YAAY,EAAE;IACd,UAAU,EAAE;GACd,CAAC;EACH;EAEA,OAAO;GAAE;GAAO;GAAS;EAAO;CAClC,EACF;AACF;;;;;;;;ACrHA,SAAgB,iBACd,SAC6E;CAC7E,6BAA6B;EAC3B,UAAU;EACV,aAAa,QAAQ,uBAAuB,CAAC;EAC7C,SAAS;CACX,CAAC;CACD,MAAM,WAAW,QAAQ,YAAY;CAErC,QAAQ,SAAmB,QAAyD;EAClF,MAAM,UAAU,sBAAsB,QAAQ,aAAa,QAAQ,UAAU,OAAO,CAAC;EACrF,MAAM,eAAe,QAAQ,QAAQ,QAAQ,QAAQ;EACrD,MAAM,UAAU,QAAQ,YAAY,OAAO;EAC3C,MAAM,SAAS,IAAI,UAAU,IAAI,gBAAgB,CAAC,CAAC;EAEnD,MAAM,MAAsB,CAAC;EAC7B,IAAI;EACJ,IAAI;EACJ,MAAM,SAAS,IAAI,SAAqB,SAAS,WAAW;GAC1D,SAAS;GACT,OAAO;EACT,CAAC;EAID,OAAO,YAAY,CAAC,CAAC;EAErB,MAAM,OAA8B;GAClC;GACA,oBAAoB;GACpB,MAAM;GACN,GAAI,QAAQ,mBAAmB,EAAE,kBAAkB,QAAQ,iBAAiB,IAAI,CAAC;EACnF;EAEA,gBAAgB,SAA4C;GAC1D,IAAI;IACF,MAAM,MAAM,MAAM,qBAAqB,QAAQ,eAAe,MAAM,MAAM;IAC1E,WAAW,MAAM,SAAS,IAAI,aAAa,SAAS,EAAE,OAAO,CAAC,GAAG;KAC/D,IAAI,KAAK,KAAK;KACd,MAAM,SAAS,SAAS,OAAO,EAAE,aAAa,CAAC;KAC/C,IAAI,QAAQ,MAAM;IACpB;IACA,OAAO,QAAQ,OAAO,MAAM,GAAG,CAAC;GAClC,SAAS,KAAK;IACZ,KAAK,GAAG;IACR,MAAM;GACR;EACF;EAEA,OAAO;GAAE,QAAQ,OAAO;GAAG;EAAO;CACpC;AACF;;;AAIA,SAAS,sBACP,MACA,WACc;CACd,IAAI,CAAC,WAAW,OAAO;CACvB,MAAM,SAAS,UAAU,eACrB;EAAE,GAAG,KAAK;EAAQ,cAAc,UAAU;CAAa,IACvD,KAAK;CACT,MAAM,cAAc,UAAU,QAAQ;EAAE,GAAI,KAAK,SAAS,CAAC;EAAI,GAAG,UAAU;CAAM,IAAI,KAAK;CAC3F,MAAM,YAAY,UAAU,iBACxB;EAAE,GAAI,KAAK,OAAO,CAAC;EAAI,GAAG,UAAU;CAAe,IACnD,KAAK;CACT,MAAM,YAAY,KAAK,WAAW,SAAS,CAAC;CAC5C,MAAM,cAAuC,UAAU,YAAY,SAC/D,CAAC,GAAG,WAAW,GAAG,UAAU,UAAU,IACtC,CAAC,GAAG,SAAS;CACjB,OAAO;EACL,GAAG;EACH,MAAM,UAAU,QAAQ,KAAK;EAC7B;EACA,GAAI,cAAc,EAAE,OAAO,YAAY,IAAI,CAAC;EAC5C,GAAI,YAAY,EAAE,KAAK,UAAU,IAAI,CAAC;EACtC,WAAW;GAAE,GAAG,KAAK;GAAW,OAAO;EAAY;CACrD;AACF"}

package/dist/{analyst-loop-DZw5QWT9.js → analyst-loop-BZvXCDKy.js}

@@ -1,5 +1,5 @@
 import { diffFindings } from "@tangle-network/agent-eval";
-//#region src/analyst-loop/run-analyst-loop.ts
+// src/analyst-loop/run-analyst-loop.ts
 /** Analyze a run and apply accepted knowledge and agent-surface proposals. */
 async function runAnalystLoop(opts) {
 	const log = opts.log ?? defaultLog;
@@ -170,7 +170,6 @@ function defaultLog(msg, fields) {
 	if (fields) console.log(`[analyst-loop] ${msg}`, fields);
 	else console.log(`[analyst-loop] ${msg}`);
 }
-//#endregion
 export { runAnalystLoop as t };
 
-//# sourceMappingURL=analyst-loop-DZw5QWT9.js.map
+//# sourceMappingURL=analyst-loop-BZvXCDKy.js.map

package/dist/{analyst-loop-DZw5QWT9.js.map → analyst-loop-BZvXCDKy.js.map}

@@ -1 +1 @@
-{"version":3,"file":"analyst-loop-DZw5QWT9.js","names":[],"sources":["../src/analyst-loop/run-analyst-loop.ts"],"sourcesContent":["/**\n * `runAnalystLoop` — the one call agent apps reach for to close the\n * recursive-self-improvement loop.\n *\n *   1. Load baseline findings (last run, or the slice the caller specifies)\n *   2. Run the analyst registry with priorFindings injected\n *   3. Persist the new run's findings to the ledger\n *   4. Diff the new run against the baseline\n *   5. Hand the findings to the knowledge proposal source\n *   6. Hand the findings to the agent-surface proposal source\n *   7. Return a single report the consumer renders / persists / acts on.\n *\n * Proposal sources are optional: the loop also works as a plain\n * \"run + diff + report\" primitive.\n */\n\nimport type { AnalystFinding, AnalystRunResult, FindingsDiff } from '@tangle-network/agent-eval'\nimport { diffFindings } from '@tangle-network/agent-eval'\n\nimport type {\n  AnalystLoopEvent,\n  AnalystRegistryStreamingLike,\n  ImprovementReport,\n  KnowledgeReport,\n  RunAnalystLoopOpts,\n  RunAnalystLoopResult,\n} from './types'\n\n/** Analyze a run and apply accepted knowledge and agent-surface proposals. */\nexport async function runAnalystLoop<TProposal = unknown, TEdit = unknown>(\n  opts: RunAnalystLoopOpts,\n): Promise<RunAnalystLoopResult<TProposal, TEdit>> {\n  const log = opts.log ?? defaultLog\n  const strategy = opts.priorFindingsStrategy ?? 'per-kind'\n  const emit = makeEmitter(opts.onEvent)\n  const startedAt = Date.now()\n\n  // 1. Resolve baseline + load prior findings.\n  const baselineRunId = resolveBaselineRunId(opts)\n  const priorAll: ReadonlyArray<AnalystFinding & { run_id: string }> = baselineRunId\n    ? (opts.findingsStore?.loadRun(baselineRunId) ?? [])\n    : []\n  log('baseline resolved', { baselineRunId, prior_findings: priorAll.length })\n  await emit({\n    type: 'baseline-resolved',\n    runId: opts.runId,\n    baselineRunId,\n    priorFindingCount: priorAll.length,\n  })\n\n  // 2. Run the registry. Strategy controls how analysts see priors.\n  //    When the registry exposes runStream, forward each event verbatim\n  //    so subscribers see per-analyst progress in real time.\n  const priorFindings = buildPriorFindingsInput(priorAll, strategy, opts.registry.list())\n  const analystResult = await runRegistry(opts, priorFindings, emit)\n  log('analyst run complete', {\n    findings: analystResult.findings.length,\n    cost_usd: analystResult.total_cost_usd,\n    per_analyst: analystResult.per_analyst.map((s) => ({\n      id: s.analyst_id,\n      status: s.status,\n      n: s.findings_count,\n    })),\n  })\n\n  // 3. Persist the new run before proposal generation so the ledger remains\n  //    the source of truth if a proposal source throws.\n  if (opts.findingsStore && analystResult.findings.length > 0) {\n    await opts.findingsStore.append(opts.runId, analystResult.findings)\n    await emit({\n      type: 'findings-persisted',\n      runId: opts.runId,\n      count: analystResult.findings.length,\n    })\n  }\n\n  // 4. Diff vs baseline.\n  let diff: FindingsDiff | null = null\n  if (baselineRunId && analystResult.findings.length > 0) {\n    diff = diffFindings(\n      priorAll.map((f) => ({ ...f })),\n      analystResult.findings.map((f) => ({ ...f, run_id: opts.runId })),\n    )\n    log('diff vs baseline', {\n      appeared: diff.appeared.length,\n      disappeared: diff.disappeared.length,\n      persisted: diff.persisted.length,\n      changed: diff.changed.length,\n    })\n    await emit({\n      type: 'diff-computed',\n      runId: opts.runId,\n      baselineRunId,\n      appeared: diff.appeared.length,\n      disappeared: diff.disappeared.length,\n      persisted: diff.persisted.length,\n      changed: diff.changed.length,\n    })\n  }\n\n  // 5. Knowledge proposals. This loop never writes live knowledge.\n  let knowledge: KnowledgeReport<TProposal> | null = null\n  if (opts.knowledgeProposalSource) {\n    knowledge = await runKnowledgeProposalSource(opts, analystResult.findings, log, emit)\n  }\n\n  // 6. Agent-surface proposals. This loop never writes live agent state.\n  let improvement: ImprovementReport<TEdit> | null = null\n  if (opts.improvementProposalSource) {\n    improvement = await runImprovementProposalSource(opts, analystResult.findings, log, emit)\n  }\n\n  const durationMs = Math.max(0, Date.now() - startedAt)\n  await emit({\n    type: 'loop-completed',\n    runId: opts.runId,\n    durationMs,\n  })\n\n  return {\n    runId: opts.runId,\n    baselineRunId,\n    durationMs,\n    analystResult,\n    diff,\n    knowledge,\n    improvement,\n  }\n}\n\ntype Emitter = (event: AnalystLoopEvent) => Promise<void>\n\nfunction makeEmitter(onEvent: RunAnalystLoopOpts['onEvent']): Emitter {\n  if (!onEvent) return async () => {}\n  return async (event) => {\n    await onEvent(event)\n  }\n}\n\nasync function runRegistry(\n  opts: RunAnalystLoopOpts,\n  priorFindings: ReturnType<typeof buildPriorFindingsInput>,\n  emit: Emitter,\n): Promise<AnalystRunResult> {\n  const reg = opts.registry as AnalystRegistryStreamingLike\n  const registryOptions = {\n    ...(priorFindings ? { priorFindings } : {}),\n    ...(opts.chainFindings !== undefined ? { chainFindings: opts.chainFindings } : {}),\n    ...(opts.costLedger ? { costLedger: opts.costLedger } : {}),\n    ...(opts.costPhase ? { costPhase: opts.costPhase } : {}),\n    ...(opts.signal ? { signal: opts.signal } : {}),\n  }\n  if (typeof reg.runStream === 'function' && opts.onEvent) {\n    let final: AnalystRunResult | null = null\n    for await (const ev of reg.runStream(opts.runId, opts.inputs, registryOptions)) {\n      await emit({ type: 'analyst', runId: opts.runId, event: ev })\n      if (ev.type === 'run-completed') final = ev.result\n    }\n    if (!final) {\n      throw new Error('runAnalystLoop: registry.runStream ended without run-completed event')\n    }\n    return final\n  }\n  return opts.registry.run(opts.runId, opts.inputs, registryOptions)\n}\n\nfunction resolveBaselineRunId(opts: RunAnalystLoopOpts): string | null {\n  if (opts.baselineRunId === null) return null\n  if (typeof opts.baselineRunId === 'string') return opts.baselineRunId\n  if (!opts.findingsStore) return null\n  const all = opts.findingsStore.loadAll()\n  let last: string | null = null\n  for (const row of all) {\n    if (row.run_id === opts.runId) continue\n    last = row.run_id\n  }\n  return last\n}\n\nfunction buildPriorFindingsInput(\n  prior: ReadonlyArray<AnalystFinding & { run_id: string }>,\n  strategy: 'per-kind' | 'wildcard' | 'none',\n  registry: ReadonlyArray<{ id: string }>,\n): ReadonlyArray<AnalystFinding> | Record<string, ReadonlyArray<AnalystFinding>> | undefined {\n  if (strategy === 'none' || prior.length === 0) return undefined\n  const stripped = prior.map(({ run_id: _run_id, ...rest }) => rest as AnalystFinding)\n  if (strategy === 'wildcard') {\n    return { '*': stripped }\n  }\n  void registry\n  return stripped\n}\n\nasync function runKnowledgeProposalSource<TProposal>(\n  opts: RunAnalystLoopOpts,\n  findings: ReadonlyArray<AnalystFinding>,\n  log: NonNullable<RunAnalystLoopOpts['log']>,\n  emit: Emitter,\n): Promise<KnowledgeReport<TProposal>> {\n  const source = opts.knowledgeProposalSource!\n  const batch = await source.proposeFromFindings(findings)\n  log('knowledge.proposeFromFindings', {\n    proposals: batch.proposals.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n  await emit({\n    type: 'knowledge-proposed',\n    runId: opts.runId,\n    proposalCount: batch.proposals.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n\n  return {\n    proposals: batch.proposals as TProposal[],\n    skipped: batch.skipped,\n    errors: batch.errors,\n  }\n}\n\nasync function runImprovementProposalSource<TEdit>(\n  opts: RunAnalystLoopOpts,\n  findings: ReadonlyArray<AnalystFinding>,\n  log: NonNullable<RunAnalystLoopOpts['log']>,\n  emit: Emitter,\n): Promise<ImprovementReport<TEdit>> {\n  const source = opts.improvementProposalSource!\n  const batch = await source.proposeFromFindings(findings)\n  log('improvement.proposeFromFindings', {\n    edits: batch.edits.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n  await emit({\n    type: 'improvement-proposed',\n    runId: opts.runId,\n    editCount: batch.edits.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n\n  return {\n    edits: batch.edits as TEdit[],\n    skipped: batch.skipped,\n    errors: batch.errors,\n  }\n}\n\nfunction defaultLog(msg: string, fields?: Record<string, unknown>): void {\n  if (fields) console.log(`[analyst-loop] ${msg}`, fields)\n  else console.log(`[analyst-loop] ${msg}`)\n}\n"],"mappings":";;;AA6BA,eAAsB,eACpB,MACiD;CACjD,MAAM,MAAM,KAAK,OAAO;CACxB,MAAM,WAAW,KAAK,yBAAyB;CAC/C,MAAM,OAAO,YAAY,KAAK,OAAO;CACrC,MAAM,YAAY,KAAK,IAAI;CAG3B,MAAM,gBAAgB,qBAAqB,IAAI;CAC/C,MAAM,WAA+D,gBAChE,KAAK,eAAe,QAAQ,aAAa,KAAK,CAAC,IAChD,CAAC;CACL,IAAI,qBAAqB;EAAE;EAAe,gBAAgB,SAAS;CAAO,CAAC;CAC3E,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ;EACA,mBAAmB,SAAS;CAC9B,CAAC;CAMD,MAAM,gBAAgB,MAAM,YAAY,MADlB,wBAAwB,UAAU,UAAU,KAAK,SAAS,KAAK,CAC3B,GAAG,IAAI;CACjE,IAAI,wBAAwB;EAC1B,UAAU,cAAc,SAAS;EACjC,UAAU,cAAc;EACxB,aAAa,cAAc,YAAY,KAAK,OAAO;GACjD,IAAI,EAAE;GACN,QAAQ,EAAE;GACV,GAAG,EAAE;EACP,EAAE;CACJ,CAAC;CAID,IAAI,KAAK,iBAAiB,cAAc,SAAS,SAAS,GAAG;EAC3D,MAAM,KAAK,cAAc,OAAO,KAAK,OAAO,cAAc,QAAQ;EAClE,MAAM,KAAK;GACT,MAAM;GACN,OAAO,KAAK;GACZ,OAAO,cAAc,SAAS;EAChC,CAAC;CACH;CAGA,IAAI,OAA4B;CAChC,IAAI,iBAAiB,cAAc,SAAS,SAAS,GAAG;EACtD,OAAO,aACL,SAAS,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE,GAC9B,cAAc,SAAS,KAAK,OAAO;GAAE,GAAG;GAAG,QAAQ,KAAK;EAAM,EAAE,CAClE;EACA,IAAI,oBAAoB;GACtB,UAAU,KAAK,SAAS;GACxB,aAAa,KAAK,YAAY;GAC9B,WAAW,KAAK,UAAU;GAC1B,SAAS,KAAK,QAAQ;EACxB,CAAC;EACD,MAAM,KAAK;GACT,MAAM;GACN,OAAO,KAAK;GACZ;GACA,UAAU,KAAK,SAAS;GACxB,aAAa,KAAK,YAAY;GAC9B,WAAW,KAAK,UAAU;GAC1B,SAAS,KAAK,QAAQ;EACxB,CAAC;CACH;CAGA,IAAI,YAA+C;CACnD,IAAI,KAAK,yBACP,YAAY,MAAM,2BAA2B,MAAM,cAAc,UAAU,KAAK,IAAI;CAItF,IAAI,cAA+C;CACnD,IAAI,KAAK,2BACP,cAAc,MAAM,6BAA6B,MAAM,cAAc,UAAU,KAAK,IAAI;CAG1F,MAAM,aAAa,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,SAAS;CACrD,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ;CACF,CAAC;CAED,OAAO;EACL,OAAO,KAAK;EACZ;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAIA,SAAS,YAAY,SAAiD;CACpE,IAAI,CAAC,SAAS,OAAO,YAAY,CAAC;CAClC,OAAO,OAAO,UAAU;EACtB,MAAM,QAAQ,KAAK;CACrB;AACF;AAEA,eAAe,YACb,MACA,eACA,MAC2B;CAC3B,MAAM,MAAM,KAAK;CACjB,MAAM,kBAAkB;EACtB,GAAI,gBAAgB,EAAE,cAAc,IAAI,CAAC;EACzC,GAAI,KAAK,kBAAkB,KAAA,IAAY,EAAE,eAAe,KAAK,cAAc,IAAI,CAAC;EAChF,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EACzD,GAAI,KAAK,YAAY,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;EACtD,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;CAC/C;CACA,IAAI,OAAO,IAAI,cAAc,cAAc,KAAK,SAAS;EACvD,IAAI,QAAiC;EACrC,WAAW,MAAM,MAAM,IAAI,UAAU,KAAK,OAAO,KAAK,QAAQ,eAAe,GAAG;GAC9E,MAAM,KAAK;IAAE,MAAM;IAAW,OAAO,KAAK;IAAO,OAAO;GAAG,CAAC;GAC5D,IAAI,GAAG,SAAS,iBAAiB,QAAQ,GAAG;EAC9C;EACA,IAAI,CAAC,OACH,MAAM,IAAI,MAAM,sEAAsE;EAExF,OAAO;CACT;CACA,OAAO,KAAK,SAAS,IAAI,KAAK,OAAO,KAAK,QAAQ,eAAe;AACnE;AAEA,SAAS,qBAAqB,MAAyC;CACrE,IAAI,KAAK,kBAAkB,MAAM,OAAO;CACxC,IAAI,OAAO,KAAK,kBAAkB,UAAU,OAAO,KAAK;CACxD,IAAI,CAAC,KAAK,eAAe,OAAO;CAChC,MAAM,MAAM,KAAK,cAAc,QAAQ;CACvC,IAAI,OAAsB;CAC1B,KAAK,MAAM,OAAO,KAAK;EACrB,IAAI,IAAI,WAAW,KAAK,OAAO;EAC/B,OAAO,IAAI;CACb;CACA,OAAO;AACT;AAEA,SAAS,wBACP,OACA,UACA,UAC2F;CAC3F,IAAI,aAAa,UAAU,MAAM,WAAW,GAAG,OAAO,KAAA;CACtD,MAAM,WAAW,MAAM,KAAK,EAAE,QAAQ,SAAS,GAAG,WAAW,IAAsB;CACnF,IAAI,aAAa,YACf,OAAO,EAAE,KAAK,SAAS;CAGzB,OAAO;AACT;AAEA,eAAe,2BACb,MACA,UACA,KACA,MACqC;CAErC,MAAM,QAAQ,MADC,KAAK,wBACO,oBAAoB,QAAQ;CACvD,IAAI,iCAAiC;EACnC,WAAW,MAAM,UAAU;EAC3B,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CACD,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ,eAAe,MAAM,UAAU;EAC/B,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CAED,OAAO;EACL,WAAW,MAAM;EACjB,SAAS,MAAM;EACf,QAAQ,MAAM;CAChB;AACF;AAEA,eAAe,6BACb,MACA,UACA,KACA,MACmC;CAEnC,MAAM,QAAQ,MADC,KAAK,0BACO,oBAAoB,QAAQ;CACvD,IAAI,mCAAmC;EACrC,OAAO,MAAM,MAAM;EACnB,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CACD,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ,WAAW,MAAM,MAAM;EACvB,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CAED,OAAO;EACL,OAAO,MAAM;EACb,SAAS,MAAM;EACf,QAAQ,MAAM;CAChB;AACF;AAEA,SAAS,WAAW,KAAa,QAAwC;CACvE,IAAI,QAAQ,QAAQ,IAAI,kBAAkB,OAAO,MAAM;MAClD,QAAQ,IAAI,kBAAkB,KAAK;AAC1C"}
+{"version":3,"file":"analyst-loop-BZvXCDKy.js","names":[],"sources":["../src/analyst-loop/run-analyst-loop.ts"],"sourcesContent":["/**\n * `runAnalystLoop` — the one call agent apps reach for to close the\n * recursive-self-improvement loop.\n *\n *   1. Load baseline findings (last run, or the slice the caller specifies)\n *   2. Run the analyst registry with priorFindings injected\n *   3. Persist the new run's findings to the ledger\n *   4. Diff the new run against the baseline\n *   5. Hand the findings to the knowledge proposal source\n *   6. Hand the findings to the agent-surface proposal source\n *   7. Return a single report the consumer renders / persists / acts on.\n *\n * Proposal sources are optional: the loop also works as a plain\n * \"run + diff + report\" primitive.\n */\n\nimport type { AnalystFinding, AnalystRunResult, FindingsDiff } from '@tangle-network/agent-eval'\nimport { diffFindings } from '@tangle-network/agent-eval'\n\nimport type {\n  AnalystLoopEvent,\n  AnalystRegistryStreamingLike,\n  ImprovementReport,\n  KnowledgeReport,\n  RunAnalystLoopOpts,\n  RunAnalystLoopResult,\n} from './types'\n\n/** Analyze a run and apply accepted knowledge and agent-surface proposals. */\nexport async function runAnalystLoop<TProposal = unknown, TEdit = unknown>(\n  opts: RunAnalystLoopOpts,\n): Promise<RunAnalystLoopResult<TProposal, TEdit>> {\n  const log = opts.log ?? defaultLog\n  const strategy = opts.priorFindingsStrategy ?? 'per-kind'\n  const emit = makeEmitter(opts.onEvent)\n  const startedAt = Date.now()\n\n  // 1. Resolve baseline + load prior findings.\n  const baselineRunId = resolveBaselineRunId(opts)\n  const priorAll: ReadonlyArray<AnalystFinding & { run_id: string }> = baselineRunId\n    ? (opts.findingsStore?.loadRun(baselineRunId) ?? [])\n    : []\n  log('baseline resolved', { baselineRunId, prior_findings: priorAll.length })\n  await emit({\n    type: 'baseline-resolved',\n    runId: opts.runId,\n    baselineRunId,\n    priorFindingCount: priorAll.length,\n  })\n\n  // 2. Run the registry. Strategy controls how analysts see priors.\n  //    When the registry exposes runStream, forward each event verbatim\n  //    so subscribers see per-analyst progress in real time.\n  const priorFindings = buildPriorFindingsInput(priorAll, strategy, opts.registry.list())\n  const analystResult = await runRegistry(opts, priorFindings, emit)\n  log('analyst run complete', {\n    findings: analystResult.findings.length,\n    cost_usd: analystResult.total_cost_usd,\n    per_analyst: analystResult.per_analyst.map((s) => ({\n      id: s.analyst_id,\n      status: s.status,\n      n: s.findings_count,\n    })),\n  })\n\n  // 3. Persist the new run before proposal generation so the ledger remains\n  //    the source of truth if a proposal source throws.\n  if (opts.findingsStore && analystResult.findings.length > 0) {\n    await opts.findingsStore.append(opts.runId, analystResult.findings)\n    await emit({\n      type: 'findings-persisted',\n      runId: opts.runId,\n      count: analystResult.findings.length,\n    })\n  }\n\n  // 4. Diff vs baseline.\n  let diff: FindingsDiff | null = null\n  if (baselineRunId && analystResult.findings.length > 0) {\n    diff = diffFindings(\n      priorAll.map((f) => ({ ...f })),\n      analystResult.findings.map((f) => ({ ...f, run_id: opts.runId })),\n    )\n    log('diff vs baseline', {\n      appeared: diff.appeared.length,\n      disappeared: diff.disappeared.length,\n      persisted: diff.persisted.length,\n      changed: diff.changed.length,\n    })\n    await emit({\n      type: 'diff-computed',\n      runId: opts.runId,\n      baselineRunId,\n      appeared: diff.appeared.length,\n      disappeared: diff.disappeared.length,\n      persisted: diff.persisted.length,\n      changed: diff.changed.length,\n    })\n  }\n\n  // 5. Knowledge proposals. This loop never writes live knowledge.\n  let knowledge: KnowledgeReport<TProposal> | null = null\n  if (opts.knowledgeProposalSource) {\n    knowledge = await runKnowledgeProposalSource(opts, analystResult.findings, log, emit)\n  }\n\n  // 6. Agent-surface proposals. This loop never writes live agent state.\n  let improvement: ImprovementReport<TEdit> | null = null\n  if (opts.improvementProposalSource) {\n    improvement = await runImprovementProposalSource(opts, analystResult.findings, log, emit)\n  }\n\n  const durationMs = Math.max(0, Date.now() - startedAt)\n  await emit({\n    type: 'loop-completed',\n    runId: opts.runId,\n    durationMs,\n  })\n\n  return {\n    runId: opts.runId,\n    baselineRunId,\n    durationMs,\n    analystResult,\n    diff,\n    knowledge,\n    improvement,\n  }\n}\n\ntype Emitter = (event: AnalystLoopEvent) => Promise<void>\n\nfunction makeEmitter(onEvent: RunAnalystLoopOpts['onEvent']): Emitter {\n  if (!onEvent) return async () => {}\n  return async (event) => {\n    await onEvent(event)\n  }\n}\n\nasync function runRegistry(\n  opts: RunAnalystLoopOpts,\n  priorFindings: ReturnType<typeof buildPriorFindingsInput>,\n  emit: Emitter,\n): Promise<AnalystRunResult> {\n  const reg = opts.registry as AnalystRegistryStreamingLike\n  const registryOptions = {\n    ...(priorFindings ? { priorFindings } : {}),\n    ...(opts.chainFindings !== undefined ? { chainFindings: opts.chainFindings } : {}),\n    ...(opts.costLedger ? { costLedger: opts.costLedger } : {}),\n    ...(opts.costPhase ? { costPhase: opts.costPhase } : {}),\n    ...(opts.signal ? { signal: opts.signal } : {}),\n  }\n  if (typeof reg.runStream === 'function' && opts.onEvent) {\n    let final: AnalystRunResult | null = null\n    for await (const ev of reg.runStream(opts.runId, opts.inputs, registryOptions)) {\n      await emit({ type: 'analyst', runId: opts.runId, event: ev })\n      if (ev.type === 'run-completed') final = ev.result\n    }\n    if (!final) {\n      throw new Error('runAnalystLoop: registry.runStream ended without run-completed event')\n    }\n    return final\n  }\n  return opts.registry.run(opts.runId, opts.inputs, registryOptions)\n}\n\nfunction resolveBaselineRunId(opts: RunAnalystLoopOpts): string | null {\n  if (opts.baselineRunId === null) return null\n  if (typeof opts.baselineRunId === 'string') return opts.baselineRunId\n  if (!opts.findingsStore) return null\n  const all = opts.findingsStore.loadAll()\n  let last: string | null = null\n  for (const row of all) {\n    if (row.run_id === opts.runId) continue\n    last = row.run_id\n  }\n  return last\n}\n\nfunction buildPriorFindingsInput(\n  prior: ReadonlyArray<AnalystFinding & { run_id: string }>,\n  strategy: 'per-kind' | 'wildcard' | 'none',\n  registry: ReadonlyArray<{ id: string }>,\n): ReadonlyArray<AnalystFinding> | Record<string, ReadonlyArray<AnalystFinding>> | undefined {\n  if (strategy === 'none' || prior.length === 0) return undefined\n  const stripped = prior.map(({ run_id: _run_id, ...rest }) => rest as AnalystFinding)\n  if (strategy === 'wildcard') {\n    return { '*': stripped }\n  }\n  void registry\n  return stripped\n}\n\nasync function runKnowledgeProposalSource<TProposal>(\n  opts: RunAnalystLoopOpts,\n  findings: ReadonlyArray<AnalystFinding>,\n  log: NonNullable<RunAnalystLoopOpts['log']>,\n  emit: Emitter,\n): Promise<KnowledgeReport<TProposal>> {\n  const source = opts.knowledgeProposalSource!\n  const batch = await source.proposeFromFindings(findings)\n  log('knowledge.proposeFromFindings', {\n    proposals: batch.proposals.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n  await emit({\n    type: 'knowledge-proposed',\n    runId: opts.runId,\n    proposalCount: batch.proposals.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n\n  return {\n    proposals: batch.proposals as TProposal[],\n    skipped: batch.skipped,\n    errors: batch.errors,\n  }\n}\n\nasync function runImprovementProposalSource<TEdit>(\n  opts: RunAnalystLoopOpts,\n  findings: ReadonlyArray<AnalystFinding>,\n  log: NonNullable<RunAnalystLoopOpts['log']>,\n  emit: Emitter,\n): Promise<ImprovementReport<TEdit>> {\n  const source = opts.improvementProposalSource!\n  const batch = await source.proposeFromFindings(findings)\n  log('improvement.proposeFromFindings', {\n    edits: batch.edits.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n  await emit({\n    type: 'improvement-proposed',\n    runId: opts.runId,\n    editCount: batch.edits.length,\n    skipped: batch.skipped,\n    errors: batch.errors.length,\n  })\n\n  return {\n    edits: batch.edits as TEdit[],\n    skipped: batch.skipped,\n    errors: batch.errors,\n  }\n}\n\nfunction defaultLog(msg: string, fields?: Record<string, unknown>): void {\n  if (fields) console.log(`[analyst-loop] ${msg}`, fields)\n  else console.log(`[analyst-loop] ${msg}`)\n}\n"],"mappings":";;;AA6BA,eAAsB,eACpB,MACiD;CACjD,MAAM,MAAM,KAAK,OAAO;CACxB,MAAM,WAAW,KAAK,yBAAyB;CAC/C,MAAM,OAAO,YAAY,KAAK,OAAO;CACrC,MAAM,YAAY,KAAK,IAAI;CAG3B,MAAM,gBAAgB,qBAAqB,IAAI;CAC/C,MAAM,WAA+D,gBAChE,KAAK,eAAe,QAAQ,aAAa,KAAK,CAAC,IAChD,CAAC;CACL,IAAI,qBAAqB;EAAE;EAAe,gBAAgB,SAAS;CAAO,CAAC;CAC3E,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ;EACA,mBAAmB,SAAS;CAC9B,CAAC;CAMD,MAAM,gBAAgB,MAAM,YAAY,MADlB,wBAAwB,UAAU,UAAU,KAAK,SAAS,KAAK,CAC3B,GAAG,IAAI;CACjE,IAAI,wBAAwB;EAC1B,UAAU,cAAc,SAAS;EACjC,UAAU,cAAc;EACxB,aAAa,cAAc,YAAY,KAAK,OAAO;GACjD,IAAI,EAAE;GACN,QAAQ,EAAE;GACV,GAAG,EAAE;EACP,EAAE;CACJ,CAAC;CAID,IAAI,KAAK,iBAAiB,cAAc,SAAS,SAAS,GAAG;EAC3D,MAAM,KAAK,cAAc,OAAO,KAAK,OAAO,cAAc,QAAQ;EAClE,MAAM,KAAK;GACT,MAAM;GACN,OAAO,KAAK;GACZ,OAAO,cAAc,SAAS;EAChC,CAAC;CACH;CAGA,IAAI,OAA4B;CAChC,IAAI,iBAAiB,cAAc,SAAS,SAAS,GAAG;EACtD,OAAO,aACL,SAAS,KAAK,OAAO,EAAE,GAAG,EAAE,EAAE,GAC9B,cAAc,SAAS,KAAK,OAAO;GAAE,GAAG;GAAG,QAAQ,KAAK;EAAM,EAAE,CAClE;EACA,IAAI,oBAAoB;GACtB,UAAU,KAAK,SAAS;GACxB,aAAa,KAAK,YAAY;GAC9B,WAAW,KAAK,UAAU;GAC1B,SAAS,KAAK,QAAQ;EACxB,CAAC;EACD,MAAM,KAAK;GACT,MAAM;GACN,OAAO,KAAK;GACZ;GACA,UAAU,KAAK,SAAS;GACxB,aAAa,KAAK,YAAY;GAC9B,WAAW,KAAK,UAAU;GAC1B,SAAS,KAAK,QAAQ;EACxB,CAAC;CACH;CAGA,IAAI,YAA+C;CACnD,IAAI,KAAK,yBACP,YAAY,MAAM,2BAA2B,MAAM,cAAc,UAAU,KAAK,IAAI;CAItF,IAAI,cAA+C;CACnD,IAAI,KAAK,2BACP,cAAc,MAAM,6BAA6B,MAAM,cAAc,UAAU,KAAK,IAAI;CAG1F,MAAM,aAAa,KAAK,IAAI,GAAG,KAAK,IAAI,IAAI,SAAS;CACrD,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ;CACF,CAAC;CAED,OAAO;EACL,OAAO,KAAK;EACZ;EACA;EACA;EACA;EACA;EACA;CACF;AACF;AAIA,SAAS,YAAY,SAAiD;CACpE,IAAI,CAAC,SAAS,OAAO,YAAY,CAAC;CAClC,OAAO,OAAO,UAAU;EACtB,MAAM,QAAQ,KAAK;CACrB;AACF;AAEA,eAAe,YACb,MACA,eACA,MAC2B;CAC3B,MAAM,MAAM,KAAK;CACjB,MAAM,kBAAkB;EACtB,GAAI,gBAAgB,EAAE,cAAc,IAAI,CAAC;EACzC,GAAI,KAAK,kBAAkB,KAAA,IAAY,EAAE,eAAe,KAAK,cAAc,IAAI,CAAC;EAChF,GAAI,KAAK,aAAa,EAAE,YAAY,KAAK,WAAW,IAAI,CAAC;EACzD,GAAI,KAAK,YAAY,EAAE,WAAW,KAAK,UAAU,IAAI,CAAC;EACtD,GAAI,KAAK,SAAS,EAAE,QAAQ,KAAK,OAAO,IAAI,CAAC;CAC/C;CACA,IAAI,OAAO,IAAI,cAAc,cAAc,KAAK,SAAS;EACvD,IAAI,QAAiC;EACrC,WAAW,MAAM,MAAM,IAAI,UAAU,KAAK,OAAO,KAAK,QAAQ,eAAe,GAAG;GAC9E,MAAM,KAAK;IAAE,MAAM;IAAW,OAAO,KAAK;IAAO,OAAO;GAAG,CAAC;GAC5D,IAAI,GAAG,SAAS,iBAAiB,QAAQ,GAAG;EAC9C;EACA,IAAI,CAAC,OACH,MAAM,IAAI,MAAM,sEAAsE;EAExF,OAAO;CACT;CACA,OAAO,KAAK,SAAS,IAAI,KAAK,OAAO,KAAK,QAAQ,eAAe;AACnE;AAEA,SAAS,qBAAqB,MAAyC;CACrE,IAAI,KAAK,kBAAkB,MAAM,OAAO;CACxC,IAAI,OAAO,KAAK,kBAAkB,UAAU,OAAO,KAAK;CACxD,IAAI,CAAC,KAAK,eAAe,OAAO;CAChC,MAAM,MAAM,KAAK,cAAc,QAAQ;CACvC,IAAI,OAAsB;CAC1B,KAAK,MAAM,OAAO,KAAK;EACrB,IAAI,IAAI,WAAW,KAAK,OAAO;EAC/B,OAAO,IAAI;CACb;CACA,OAAO;AACT;AAEA,SAAS,wBACP,OACA,UACA,UAC2F;CAC3F,IAAI,aAAa,UAAU,MAAM,WAAW,GAAG,OAAO,KAAA;CACtD,MAAM,WAAW,MAAM,KAAK,EAAE,QAAQ,SAAS,GAAG,WAAW,IAAsB;CACnF,IAAI,aAAa,YACf,OAAO,EAAE,KAAK,SAAS;CAGzB,OAAO;AACT;AAEA,eAAe,2BACb,MACA,UACA,KACA,MACqC;CAErC,MAAM,QAAQ,MADC,KAAK,wBACO,oBAAoB,QAAQ;CACvD,IAAI,iCAAiC;EACnC,WAAW,MAAM,UAAU;EAC3B,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CACD,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ,eAAe,MAAM,UAAU;EAC/B,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CAED,OAAO;EACL,WAAW,MAAM;EACjB,SAAS,MAAM;EACf,QAAQ,MAAM;CAChB;AACF;AAEA,eAAe,6BACb,MACA,UACA,KACA,MACmC;CAEnC,MAAM,QAAQ,MADC,KAAK,0BACO,oBAAoB,QAAQ;CACvD,IAAI,mCAAmC;EACrC,OAAO,MAAM,MAAM;EACnB,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CACD,MAAM,KAAK;EACT,MAAM;EACN,OAAO,KAAK;EACZ,WAAW,MAAM,MAAM;EACvB,SAAS,MAAM;EACf,QAAQ,MAAM,OAAO;CACvB,CAAC;CAED,OAAO;EACL,OAAO,MAAM;EACb,SAAS,MAAM;EACf,QAAQ,MAAM;CAChB;AACF;AAEA,SAAS,WAAW,KAAa,QAAwC;CACvE,IAAI,QAAQ,QAAQ,IAAI,kBAAkB,OAAO,MAAM;MAClD,QAAQ,IAAI,kBAAkB,KAAK;AAC1C"}

package/dist/analyst-loop.d.ts

@@ -1,7 +1,5 @@
-import { a as ImprovementProposalSource, c as KnowledgeProposalSource, d as RunAnalystLoopResult, i as ImprovementEditBatch, l as KnowledgeReport, n as AnalystRegistryLike, o as ImprovementReport, r as FindingsStoreLike, s as KnowledgeProposalBatch, t as AnalystLoopEvent, u as RunAnalystLoopOpts } from "./types-Cwrea-bU.js";
-//#region src/analyst-loop/run-analyst-loop.d.ts
+import { a as ImprovementProposalSource, c as KnowledgeProposalSource, d as RunAnalystLoopResult, i as ImprovementEditBatch, l as KnowledgeReport, n as AnalystRegistryLike, o as ImprovementReport, r as FindingsStoreLike, s as KnowledgeProposalBatch, t as AnalystLoopEvent, u as RunAnalystLoopOpts } from "./types-qsqfyQoY.js";
 /** Analyze a run and apply accepted knowledge and agent-surface proposals. */
-declare function runAnalystLoop<TProposal = unknown, TEdit = unknown>(opts: RunAnalystLoopOpts): Promise<RunAnalystLoopResult<TProposal, TEdit>>;
-//#endregion
-export { type AnalystLoopEvent, type AnalystRegistryLike, type FindingsStoreLike, type ImprovementEditBatch, type ImprovementProposalSource, type ImprovementReport, type KnowledgeProposalBatch, type KnowledgeProposalSource, type KnowledgeReport, type RunAnalystLoopOpts, type RunAnalystLoopResult, runAnalystLoop };
+export declare function runAnalystLoop<TProposal = unknown, TEdit = unknown>(opts: RunAnalystLoopOpts): Promise<RunAnalystLoopResult<TProposal, TEdit>>;
+export type { AnalystLoopEvent, AnalystRegistryLike, FindingsStoreLike, ImprovementEditBatch, ImprovementProposalSource, ImprovementReport, KnowledgeProposalBatch, KnowledgeProposalSource, KnowledgeReport, RunAnalystLoopOpts, RunAnalystLoopResult };
 //# sourceMappingURL=analyst-loop.d.ts.map

package/dist/analyst-loop.js

@@ -1,2 +1 @@
-import { t as runAnalystLoop } from "./analyst-loop-DZw5QWT9.js";
-export { runAnalystLoop };
+export { t as runAnalystLoop } from "./analyst-loop-BZvXCDKy.js";

package/dist/{bundle-1xaJEgjq.js → bundle-Ze8pd07a.js}

@@ -1,6 +1,6 @@
-import { P as canonicalCandidateDigest$1, R as immutableCandidateValue, z as omitTopLevelDigest } from "./sandbox-profile-compat-DtuJWFrM.js";
+import { P as canonicalCandidateDigest$1, R as immutableCandidateValue, z as omitTopLevelDigest } from "./sandbox-profile-compat-BxMQ1VXv.js";
 import { agentCandidateBundleSchema } from "@tangle-network/agent-interface";
-//#region src/candidate-execution/bundle.ts
+// src/candidate-execution/bundle.ts
 /** Validate and content-address a candidate bundle before it crosses an approval boundary. */
 function sealAgentCandidateBundle(input) {
 	const digest = canonicalCandidateDigest$1(input);
@@ -11,7 +11,6 @@ function sealAgentCandidateBundle(input) {
 	if (canonicalCandidateDigest$1(omitTopLevelDigest(parsed)) !== digest) throw new Error("candidate bundle changed while validating its canonical wire shape");
 	return immutableCandidateValue(parsed);
 }
-//#endregion
 export { sealAgentCandidateBundle as t };
 
-//# sourceMappingURL=bundle-1xaJEgjq.js.map
+//# sourceMappingURL=bundle-Ze8pd07a.js.map