@tangle-network/agent-runtime 0.90.0 → 0.90.1

package/dist/{completion-gate-BDaBP8dd.d.ts → completion-gate-DkAnUmpb.d.ts}

@@ -1,5 +1,5 @@
-import { L as LocalHarness } from './local-harness-sI0S_XNA.js';
-import { c as Executor } from './types-DAJQRIUD.js';
+import { L as LocalHarness } from './local-harness-dcD5WTTr.js';
+import { c as Executor } from './types-eMNgWgFi.js';
 
 /**
  *

package/dist/{coordination-DxHduZg7.d.ts → coordination-rRj5hjJK.d.ts}

@@ -1,12 +1,12 @@
-import { E as ExecutorFactory, e as ExecutorRegistry, i as Spend, A as Agent, S as Scope, a as ResultBlobStore, B as Budget } from './types-DAJQRIUD.js';
+import { E as ExecutorFactory, h as ExecutorRegistry, e as Spend, A as Agent, S as Scope, a as ResultBlobStore, B as Budget } from './types-eMNgWgFi.js';
 import { AgentProfile } from '@tangle-network/agent-interface';
 import { U as UiLens, a as UiFinding, C as CoderTask } from './substrate-DO2GHNg2.js';
-import { S as SandboxClient, E as ExecCtx, e as LoopTraceEmitter, c as LoopTraceEvent, A as AgentRunSpec } from './types-Dnk189QA.js';
+import { S as SandboxClient, E as ExecCtx, f as LoopTraceEmitter, d as LoopTraceEvent, A as AgentRunSpec } from './types-SyuwunY_.js';
 import { BackendType, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
 import { AgentEvalError } from '@tangle-network/agent-eval';
 import { b as ToolSpec, R as RouterConfig } from './router-client-DJImUDlm.js';
-import { G as GitRunner, a as WorktreeCheckRunner, D as DeliverableSpec } from './completion-gate-BDaBP8dd.js';
-import { L as LocalHarness } from './local-harness-sI0S_XNA.js';
+import { G as GitRunner, a as WorktreeCheckRunner, D as DeliverableSpec } from './completion-gate-DkAnUmpb.js';
+import { L as LocalHarness } from './local-harness-dcD5WTTr.js';
 import { ProviderExecutorOptions, AgentEnvironmentProviderRegistry } from './environment-provider.js';
 import { AgentEnvironmentProvider } from '@tangle-network/agent-interface/environment-provider';
 
@@ -717,7 +717,7 @@ interface EvalRunEvent {
 interface EvalRunsExportConfig {
     /** Bearer key — tenant is resolved server-side from it. Reads TANGLE_API_KEY. */
     apiKey?: string;
-    /** Intelligence base. Reads INTELLIGENCE_BASE env, else prod. */
+    /** Intelligence base. Reads TANGLE_INTELLIGENCE_URL env, else prod. */
     base?: string;
     /** Idempotency-Key header (e.g. the runId) — safe retries + upsert. */
     idempotencyKey?: string;

package/dist/environment-provider.d.ts

@@ -2,8 +2,8 @@ import { AgentProfile, AgentProfileValidationResult } from '@tangle-network/agen
 import { CreateAgentEnvironmentInput, AgentTurnInput, AgentEnvironmentProvider, AgentEnvironmentCapabilities, AgentProfileRef } from '@tangle-network/agent-interface/environment-provider';
 export { AgentEnvironment, AgentEnvironmentCapabilities, AgentEnvironmentEvent, AgentEnvironmentProvider, AgentEnvironmentQuery, AgentEnvironmentStatus, AgentEnvironmentSummary, AgentProfileRef, AgentSession, AgentSessionRef, AgentSessionStatus, AgentTurnInput, AgentTurnResult, CheckpointRef, CheckpointRequest, CreateAgentEnvironmentInput, ExecRequest, ExecResult, ForkRequest, PlacementInfo, ResourceRequest, WorkspaceRequest } from '@tangle-network/agent-interface/environment-provider';
 import { CreateSandboxOptions, BackendType } from '@tangle-network/sandbox';
-import { R as Runtime, E as ExecutorFactory } from './types-DAJQRIUD.js';
-import { S as SandboxClient } from './types-Dnk189QA.js';
+import { R as Runtime, E as ExecutorFactory } from './types-eMNgWgFi.js';
+import { S as SandboxClient } from './types-SyuwunY_.js';
 import '@tangle-network/agent-eval';
 
 /** Provider object or registry name accepted by runtime provider adapters.

package/dist/index.d.ts

@@ -1,28 +1,29 @@
 import { AgentProfile, AgentEvalError, AnalystFinding, KnowledgeReadinessReport, RunRecord, ControlEvalResult, KnowledgeRequirement } from '@tangle-network/agent-eval';
 export { AgentEvalError, AgentEvalErrorCode, ConfigError, ControlBudget, ControlDecision, ControlEvalResult, ControlRunResult, ControlStep, DataAcquisitionPlan, JudgeError, KnowledgeReadinessReport, KnowledgeRequirement, NotFoundError, RunRecord, ValidationError } from '@tangle-network/agent-eval';
-import { f as AgentBackendInput, O as OpenAIChatTool, g as OpenAIChatToolChoice, h as OpenAIChatResponseFormat, i as AgentExecutionBackend, j as AgentBackendContext, a as RuntimeStreamEvent, K as KnowledgeReadinessDecision, k as RunAgentTaskOptions, l as AgentTaskRunResult, m as RunAgentTaskStreamOptions, n as AgentRuntimeEvent, o as AgentTaskStatus, p as RuntimeSessionStore, q as RuntimeSession, R as RuntimeHooks } from './types-Dnk189QA.js';
-export { r as AgentAdapter, s as AgentKnowledgeProvider, t as AgentRuntimeEventSink, u as AgentTaskContext, v as AgentTaskSpec, B as BackendErrorDetail, w as RuntimeDecisionEvidenceRef, x as RuntimeDecisionKind, y as RuntimeDecisionPoint, z as RuntimeHookContext, C as RuntimeHookErrorContext, F as RuntimeHookEvent, G as RuntimeHookPhase, H as RuntimeHookTarget, J as RuntimeRunHandle, M as RuntimeRunPersistenceAdapter, N as RuntimeRunRow, P as composeRuntimeHooks, Q as defineRuntimeHooks, T as notifyRuntimeDecisionPoint, U as notifyRuntimeHookEvent, W as startRuntimeRun } from './types-Dnk189QA.js';
-import { Scenario, ProfileDispatchFn } from '@tangle-network/agent-eval/campaign';
-import { V as Verifier, C as CandidateGenerator } from './mcp-serve-verifier-FL7-ZEb_.js';
-export { A as AgenticGeneratorOptions, I as ImprovementDriverOptions, M as McpServeSpec, a as VerifyResult, b as agenticGenerator, c as commandVerifier, i as improvementDriver, m as mcpServeVerifier } from './mcp-serve-verifier-FL7-ZEb_.js';
+import { g as AgentBackendInput, O as OpenAIChatTool, h as OpenAIChatToolChoice, i as OpenAIChatResponseFormat, j as AgentExecutionBackend, k as AgentBackendContext, b as RuntimeStreamEvent, K as KnowledgeReadinessDecision, l as RunAgentTaskOptions, m as AgentTaskRunResult, n as RunAgentTaskStreamOptions, o as AgentRuntimeEvent, p as AgentTaskStatus, q as RuntimeSessionStore, r as RuntimeSession, R as RuntimeHooks } from './types-SyuwunY_.js';
+export { s as AgentAdapter, t as AgentKnowledgeProvider, u as AgentRuntimeEventSink, v as AgentTaskContext, w as AgentTaskSpec, B as BackendErrorDetail, x as RuntimeDecisionEvidenceRef, y as RuntimeDecisionKind, z as RuntimeDecisionPoint, C as RuntimeHookContext, F as RuntimeHookErrorContext, G as RuntimeHookEvent, H as RuntimeHookPhase, J as RuntimeHookTarget, M as RuntimeRunHandle, N as RuntimeRunPersistenceAdapter, P as RuntimeRunRow, Q as composeRuntimeHooks, T as defineRuntimeHooks, U as notifyRuntimeDecisionPoint, W as notifyRuntimeHookEvent, X as startRuntimeRun } from './types-SyuwunY_.js';
+import { Scenario, ProfileDispatchFn, MutableSurface as MutableSurface$1, SurfaceProposer as SurfaceProposer$1 } from '@tangle-network/agent-eval/campaign';
+import { V as Verifier, C as CandidateGenerator } from './mcp-serve-verifier-XsX8rkB9.js';
+export { A as AgenticGeneratorOptions, I as ImprovementDriverOptions, M as McpServeSpec, a as VerifyResult, b as agenticGenerator, c as commandVerifier, i as improvementDriver, m as mcpServeVerifier } from './mcp-serve-verifier-XsX8rkB9.js';
 import { Scenario as Scenario$1, SurfaceProposer, JudgeConfig, MutableSurface, DispatchContext, SelfImproveBudget, SelfImproveLlm, SelfImproveOptions, SelfImproveResult } from '@tangle-network/agent-eval/contract';
 import { AgentProfile as AgentProfile$1 } from '@tangle-network/agent-interface';
-import { L as LocalHarness } from './local-harness-sI0S_XNA.js';
+import { L as LocalHarness } from './local-harness-dcD5WTTr.js';
 import { S as SurfaceImprovementEdit } from './improvement-adapter-CDR8QNVM.js';
 import { I as ImprovementAdapter } from './types-BC3bZpH0.js';
+import { S as StructuralRolloutPolicy } from './structural-rollout-MwlpgQ-6.js';
 export { AgentKnowledgeReadinessCheckOptions, KnowledgeImprovementJobMeasurement, KnowledgeImprovementJobResult, KnowledgeReadinessCheck, KnowledgeReadinessCheckInput, KnowledgeReadinessCheckResult, RESEARCH_SUPERVISOR_SYSTEM_PROMPT, RunKnowledgeImprovementJobOptions, SupervisedKnowledgeUpdateInput, SupervisedKnowledgeUpdateOptions, SupervisedKnowledgeUpdateResult, SupervisedKnowledgeUpdater, createAgentKnowledgeReadinessCheck, createSupervisedKnowledgeUpdater, formatSupervisedKnowledgeTask, knowledgeReadinessDeliverable, runKnowledgeImprovementJob, runSupervisedKnowledgeUpdate } from './knowledge.js';
-export { D as DELEGATED_LOOP_MODES, a as DelegatedLoopMode, b as DelegatedLoopRegistry, c as DelegatedLoopResult, d as DelegatedLoopRunner, L as LoopRunnerCliArgs, e as LoopRunnerCliResult, R as ResearchLoopResult, f as ResearchLoopRunnerOptions, g as RunDelegatedLoopOptions, V as VetoedFact, W as WorktreeLoopRunnerOptions, h as auditLoopRunner, i as isDelegatedLoopMode, p as parseLoopRunnerArgv, r as researchLoopRunner, j as runDelegatedLoop, k as runLoopRunnerCli, s as selfImproveLoopRunner, w as worktreeLoopRunner } from './loop-runner-bin-D3GmB2Ru.js';
-export { m as mcpToolsForRuntimeMcp, a as mcpToolsForRuntimeMcpSubset } from './openai-tools-D0xmzo0y.js';
-export { b0 as EvalRunEvent, b1 as EvalRunGeneration, b2 as EvalRunsExportConfig, b3 as EvalRunsExportResult, b4 as INTELLIGENCE_WIRE_VERSION, b5 as LoopSpanNode, b6 as OtelAttribute, b7 as OtelExportConfig, b8 as OtelExporter, b9 as OtelSpan, ba as buildLoopOtelSpans, bb as buildLoopSpanNodes, bc as createOtelExporter, bd as exportEvalRuns, be as loopEventToOtelSpan } from './coordination-DxHduZg7.js';
+export { D as DELEGATED_LOOP_MODES, a as DelegatedLoopMode, b as DelegatedLoopRegistry, c as DelegatedLoopResult, d as DelegatedLoopRunner, L as LoopRunnerCliArgs, e as LoopRunnerCliResult, R as ResearchLoopResult, f as ResearchLoopRunnerOptions, g as RunDelegatedLoopOptions, V as VetoedFact, W as WorktreeLoopRunnerOptions, h as auditLoopRunner, i as isDelegatedLoopMode, p as parseLoopRunnerArgv, r as researchLoopRunner, j as runDelegatedLoop, k as runLoopRunnerCli, s as selfImproveLoopRunner, w as worktreeLoopRunner } from './loop-runner-bin-DTbZVGfM.js';
+export { m as mcpToolsForRuntimeMcp, a as mcpToolsForRuntimeMcpSubset } from './openai-tools-C4ZfUD4L.js';
+export { b0 as EvalRunEvent, b1 as EvalRunGeneration, b2 as EvalRunsExportConfig, b3 as EvalRunsExportResult, b4 as INTELLIGENCE_WIRE_VERSION, b5 as LoopSpanNode, b6 as OtelAttribute, b7 as OtelExportConfig, b8 as OtelExporter, b9 as OtelSpan, ba as buildLoopOtelSpans, bb as buildLoopSpanNodes, bc as createOtelExporter, bd as exportEvalRuns, be as loopEventToOtelSpan } from './coordination-rRj5hjJK.js';
 import '@tangle-network/sandbox';
 import 'node:child_process';
+import './worktree-fanout-BDFQIO-Y.js';
+import './types-eMNgWgFi.js';
+import './completion-gate-DkAnUmpb.js';
 import '@tangle-network/agent-knowledge';
-import './supervise-BpCdssu0.js';
-import './types-DAJQRIUD.js';
+import './supervise-DPmYPk0j.js';
 import './router-client-DJImUDlm.js';
-import './completion-gate-BDaBP8dd.js';
 import './kb-gate-CwHO0vz6.js';
-import './worktree-fanout-CfRXYmgV.js';
 import './substrate-DO2GHNg2.js';
 import './environment-provider.js';
 import '@tangle-network/agent-interface/environment-provider';
@@ -1182,6 +1183,11 @@ declare function mcpBuildPrompt(args: FindingsArg): string;
  *
  *   - `surface: 'prompt'` → `gepaProposer` mutates `profile.prompt.systemPrompt`.
  *   - `surface: 'skills'` → `skillOptProposer` mutates a skills document string.
+ *   - `surface: 'rollout-policy'` → `rolloutPolicyProposer` mutates the
+ *     inference-time `StructuralRolloutPolicy` dials ({ k, repairRounds, testgen })
+ *     persisted in `profile.extensions['structural-rollout']` — deterministic
+ *     bounded neighbor enumeration; the held-out gate does the deciding. No-op
+ *     (nothing proposed, nothing shipped) when the profile has no such extension.
  *   - `surface` ∈ {`tools`, `mcp`, `hooks`, `code`} → no zero-config default
  *     proposer exists (a code/config proposer needs caller-supplied wiring — a
  *     worktree repo root, a candidate generator, a serializer). The facade
@@ -1196,8 +1202,10 @@ declare function mcpBuildPrompt(args: FindingsArg): string;
  */
 
 /** The agent-profile lever `improve` optimizes. Mirrors the AgentProfile-law
- *  profile levers; `code` is the implementation-tier surface. */
-type ImproveSurface = 'prompt' | 'skills' | 'tools' | 'mcp' | 'hooks' | 'code';
+ *  profile levers; `code` is the implementation-tier surface, `rollout-policy`
+ *  the inference-time structuralRollout dials
+ *  (`profile.extensions['structural-rollout']`). */
+type ImproveSurface = 'prompt' | 'skills' | 'tools' | 'mcp' | 'hooks' | 'code' | 'rollout-policy';
 interface ImproveOptions<TScenario extends Scenario$1, TArtifact> {
     /** Which profile lever to optimize. Default `'prompt'`. Selects the default
      *  generator + the baseline-surface extraction shape. */
@@ -1411,6 +1419,94 @@ interface ReflectiveGeneratorOptions {
 /** Cheap no-sandbox `CandidateGenerator` (the `shots=1` setting): draft surface edits via the improvement adapter and apply them as one coherent candidate. */
 declare function reflectiveGenerator(opts: ReflectiveGeneratorOptions): CandidateGenerator;
 
+/**
+ * `rolloutPolicyProposer` — the `'rollout-policy'` surface for `improve()`: the
+ * inference-time `StructuralRolloutPolicy` dials { k, repairRounds, testgen } as a
+ * held-out-gated optimizable surface.
+ *
+ * Why this seam: agent-eval's loop contract is already generic — `MutableSurface`
+ * admits any string, documented as "serialized tool config" — so the policy rides
+ * the SAME serialize→propose→gate→parse-back cycle the tools/mcp/hooks surfaces
+ * use. No agent-eval changes; the only net-new piece is this proposer.
+ *
+ * Why deterministic: prompt-wording proposals are a measured zero on this stack,
+ * and the policy space is tiny and fully enumerable. The proposer emits bounded
+ * single-dial neighbors (k±2 in [1,10], repairRounds±1 in [0,3], testgen±3 in
+ * [0,10], ≤4 per generation) and lets the held-out gate do ALL the deciding — an
+ * LLM proposer would add cost and nondeterminism with nothing to reason about.
+ *
+ * Persistence: the policy lives in `profile.extensions['structural-rollout']`
+ * (AgentProfile's designed slot for runtime-specific config). A gated winner is
+ * written back there by `improve()`, the same profile-field write-back every other
+ * config surface gets; `structuralRolloutPolicyFromProfile` is the read side a
+ * runtime caller feeds to `structuralRollout({ policy })`.
+ *
+ * @experimental
+ */
+
+/** The profile extensions namespace the policy persists under. */
+declare const ROLLOUT_POLICY_EXTENSION = "structural-rollout";
+/** Proposal bounds per dial. These are the SEARCH bounds (what the proposer may
+ *  explore), chosen so every reachable value is a measured-sane recipe: k=1 is the
+ *  low-compute preset, testgen=0 disables check authoring, repairRounds caps where
+ *  the measured increment flattens (+1–3pp beyond round 2). */
+declare const ROLLOUT_POLICY_BOUNDS: {
+    readonly k: {
+        readonly min: 1;
+        readonly max: 10;
+        readonly step: 2;
+    };
+    readonly repairRounds: {
+        readonly min: 0;
+        readonly max: 3;
+        readonly step: 1;
+    };
+    readonly testgen: {
+        readonly min: 0;
+        readonly max: 10;
+        readonly step: 3;
+    };
+};
+/** Parse a serialized policy surface. Defensive by design — the proposer reads
+ *  `ctx.currentSurface`, which the loop types as `string | CodeSurface`. Returns
+ *  `undefined` (never throws) for non-strings, malformed JSON, or a shape that
+ *  violates the policy's own invariants: the no-op signal. Unknown dials are
+ *  dropped; `diverse`/`temperature` ride through untouched (the proposer never
+ *  mutates them — `diverse` is a measured paired null). */
+declare function parseRolloutPolicy(surface: MutableSurface$1): StructuralRolloutPolicy | undefined;
+/** Normalize an untyped policy bag (a parsed surface or a profile extension) into
+ *  a full `StructuralRolloutPolicy`, defaults merged. Returns `undefined` when any
+ *  present dial violates the policy invariants (mirrors `resolvePolicy`: integer
+ *  k ≥ 1, repairRounds ≥ 0, testgen ≥ 0) — a corrupt config must read as "not
+ *  configured", never as a fabricated recipe. */
+declare function normalizeRolloutPolicy(raw: unknown): StructuralRolloutPolicy | undefined;
+/** Stable serialization — dial order is fixed so identical policies produce
+ *  identical surfaces (the loop dedupes/hashes candidates by surface content). */
+declare function serializeRolloutPolicy(policy: StructuralRolloutPolicy): string;
+/** Read the persisted policy off the profile. `undefined` when the profile does
+ *  not opt into structural rollout — the improve() surface no-ops then, because
+ *  tuning dials nothing consumes would ship dead config. */
+declare function structuralRolloutPolicyFromProfile(profile: AgentProfile$1): StructuralRolloutPolicy | undefined;
+/** Persist a policy into the profile's extensions namespace. Shallow copy; never
+ *  mutates the input profile (the applyWinnerToProfile contract). */
+declare function applyRolloutPolicyToProfile(profile: AgentProfile$1, policy: StructuralRolloutPolicy): AgentProfile$1;
+/** All bounded single-dial neighbors of `policy`, in a fixed priority order: k
+ *  first (selection breadth carries 85–92% of the measured effect), then
+ *  repairRounds, then testgen. Steps clamp to the dial's bounds; clamped-to-no-op
+ *  and duplicate policies are dropped. */
+declare function enumerateNeighborPolicies(policy: StructuralRolloutPolicy): StructuralRolloutPolicy[];
+/**
+ * The deterministic `SurfaceProposer` for the `'rollout-policy'` surface.
+ *
+ * Each generation: parse the current policy surface, enumerate its bounded
+ * single-dial neighbors, and return at most `min(populationSize, 4)` of them,
+ * rotating the enumeration window by generation so successive generations explore
+ * different neighbors when nothing promoted. Proposes NOTHING when the surface
+ * carries no policy (the profile never opted in) — an empty proposal is the
+ * loop-native no-op, mirroring `improvementDriver`'s no-findings behavior.
+ */
+declare function rolloutPolicyProposer(): SurfaceProposer$1;
+
 /**
  *
  * Chat-model resolution + catalog validation — the shared primitive every
@@ -1945,4 +2041,4 @@ interface StreamToolLoopOptions<Raw> {
  *  `capped` if it stops for any non-completed reason with calls still pending. */
 declare function streamToolLoop<Raw>(opts: StreamToolLoopOptions<Raw>): AsyncGenerator<StreamToolLoopYield<Raw>, void, unknown>;
 
-export { AgentBackendContext, AgentBackendInput, type AgentBackendKind, AgentExecutionBackend, AgentRuntimeEvent, AgentTaskRunResult, AgentTaskStatus, type AuthSource, type BackendCallPolicy, BackendTransportError, CandidateGenerator, type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, type CircuitBreakerConfig, CircuitBreakerState, CircuitOpenError, type Conversation, type ConversationDriveState, type ConversationJournal, type ConversationJournalEntry, type ConversationParticipant, type ConversationPolicy, type ConversationResult, type ConversationStreamEvent, type ConversationTurn, type D1DatabaseLike, type D1StmtLike, DEFAULT_MAX_DEPTH, DEFAULT_ROUTER_BASE_URL, DeadlineExceededError, FORWARD_HEADERS, FileConversationJournal, type ForwardHeaderName, type HaltContext, type HaltPredicate, type HaltReason, type HaltSignal, type ImproveOptions, type ImproveResult, type ImproveSurface, InMemoryConversationJournal, InMemoryRuntimeSessionStore, type ModelInfo, OpenAIChatResponseFormat, OpenAIChatTool, OpenAIChatToolChoice, type PersonaConversationResult, type PersonaDriver, PlannerError, type PropagatedHeaders, type RawTraceDistillerOptions, type ReflectiveGeneratorOptions, type ResolveAgentBackendOptions, type ResolvedChatModel, type RetryBackoff, type RetryableErrorPredicate, type RouterEnv, type RunChatTurnInput, type RunConversationOptions, type RunPersonaConfig, type RunPersonaConversationOptions, type RunToolLoopOptions, type RuntimeEventCollector, RuntimeHooks, RuntimeRunStateError, RuntimeSessionStore, RuntimeStreamEvent, type RuntimeStreamEventCollector, type RuntimeTelemetryOptions, type SanitizedKnowledgeReadinessReport, type SqlAdapter, SqlConversationJournal, type StreamToolLoopOptions, type StreamToolLoopYield, type ToolCallOutcome, type ToolLoopAssistantToolCall, type ToolLoopCall, type ToolLoopEvent, type ToolLoopMessage, type ToolLoopResult, type ToolLoopStopReason, type TurnOrder, Verifier, applyRunRecordDefaults, buildForwardHeaders, cleanModelId, computeBackoff, createConversationBackend, createIterableBackend, createOpenAICompatibleBackend, createRuntimeEventCollector, createRuntimeStreamEventCollector, createSandboxPromptBackend, d1ToSqlAdapter, decideKnowledgeReadiness, defaultIsRetryable, defineConversation, deriveExecutionId, getModels, handleChatTurn, improve, isDepthExceeded, makePerAttemptSignal, mcpBuildPrompt, rawTraceDistiller, readDepth, readinessServerSentEvent, reflectiveGenerator, resolveAgentBackend, resolveChatModel, resolveRouterBaseUrl, runAgentTask, runAgentTaskStream, runConversation, runConversationStream, runPersonaConversation, runPersonaDispatch, runToolLoop, runtimeStreamServerSentEvent, sanitizeAgentRuntimeEvent, sanitizeKnowledgeReadinessReport, sanitizeRuntimeStreamEvent, sleep, slugifySpeaker, streamToolLoop, toolBuildPrompt, turnId, validateChatModelId };
+export { AgentBackendContext, AgentBackendInput, type AgentBackendKind, AgentExecutionBackend, AgentRuntimeEvent, AgentTaskRunResult, AgentTaskStatus, type AuthSource, type BackendCallPolicy, BackendTransportError, CandidateGenerator, type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, type CircuitBreakerConfig, CircuitBreakerState, CircuitOpenError, type Conversation, type ConversationDriveState, type ConversationJournal, type ConversationJournalEntry, type ConversationParticipant, type ConversationPolicy, type ConversationResult, type ConversationStreamEvent, type ConversationTurn, type D1DatabaseLike, type D1StmtLike, DEFAULT_MAX_DEPTH, DEFAULT_ROUTER_BASE_URL, DeadlineExceededError, FORWARD_HEADERS, FileConversationJournal, type ForwardHeaderName, type HaltContext, type HaltPredicate, type HaltReason, type HaltSignal, type ImproveOptions, type ImproveResult, type ImproveSurface, InMemoryConversationJournal, InMemoryRuntimeSessionStore, type ModelInfo, OpenAIChatResponseFormat, OpenAIChatTool, OpenAIChatToolChoice, type PersonaConversationResult, type PersonaDriver, PlannerError, type PropagatedHeaders, ROLLOUT_POLICY_BOUNDS, ROLLOUT_POLICY_EXTENSION, type RawTraceDistillerOptions, type ReflectiveGeneratorOptions, type ResolveAgentBackendOptions, type ResolvedChatModel, type RetryBackoff, type RetryableErrorPredicate, type RouterEnv, type RunChatTurnInput, type RunConversationOptions, type RunPersonaConfig, type RunPersonaConversationOptions, type RunToolLoopOptions, type RuntimeEventCollector, RuntimeHooks, RuntimeRunStateError, RuntimeSessionStore, RuntimeStreamEvent, type RuntimeStreamEventCollector, type RuntimeTelemetryOptions, type SanitizedKnowledgeReadinessReport, type SqlAdapter, SqlConversationJournal, type StreamToolLoopOptions, type StreamToolLoopYield, type ToolCallOutcome, type ToolLoopAssistantToolCall, type ToolLoopCall, type ToolLoopEvent, type ToolLoopMessage, type ToolLoopResult, type ToolLoopStopReason, type TurnOrder, Verifier, applyRolloutPolicyToProfile, applyRunRecordDefaults, buildForwardHeaders, cleanModelId, computeBackoff, createConversationBackend, createIterableBackend, createOpenAICompatibleBackend, createRuntimeEventCollector, createRuntimeStreamEventCollector, createSandboxPromptBackend, d1ToSqlAdapter, decideKnowledgeReadiness, defaultIsRetryable, defineConversation, deriveExecutionId, enumerateNeighborPolicies, getModels, handleChatTurn, improve, isDepthExceeded, makePerAttemptSignal, mcpBuildPrompt, normalizeRolloutPolicy, parseRolloutPolicy, rawTraceDistiller, readDepth, readinessServerSentEvent, reflectiveGenerator, resolveAgentBackend, resolveChatModel, resolveRouterBaseUrl, rolloutPolicyProposer, runAgentTask, runAgentTaskStream, runConversation, runConversationStream, runPersonaConversation, runPersonaDispatch, runToolLoop, runtimeStreamServerSentEvent, sanitizeAgentRuntimeEvent, sanitizeKnowledgeReadinessReport, sanitizeRuntimeStreamEvent, serializeRolloutPolicy, sleep, slugifySpeaker, streamToolLoop, structuralRolloutPolicyFromProfile, toolBuildPrompt, turnId, validateChatModelId };

package/dist/index.js

@@ -1,7 +1,7 @@
 import {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
-} from "./chunk-XQEISTK2.js";
+} from "./chunk-QK4DV5PR.js";
 import {
   DEFAULT_ROUTER_BASE_URL,
   cleanModelId,
@@ -20,18 +20,19 @@ import {
   runLoopRunnerCli,
   selfImproveLoopRunner,
   worktreeLoopRunner
-} from "./chunk-MHK62APK.js";
+} from "./chunk-WRUSWK4F.js";
 import "./chunk-SGKPNBXE.js";
 import {
   InMemoryRuntimeSessionStore,
   createIterableBackend,
   createOpenAICompatibleBackend,
   createSandboxPromptBackend,
+  defaultStructuralRolloutPolicy,
   newRuntimeSession,
   normalizeBackendStreamEvent,
   nowIso,
   touchSession
-} from "./chunk-4IBAMGBE.js";
+} from "./chunk-ZV4LXYCJ.js";
 import "./chunk-ZQZX77MM.js";
 import {
   agenticGenerator,
@@ -39,7 +40,7 @@ import {
   mcpBuildPrompt,
   mcpServeVerifier,
   toolBuildPrompt
-} from "./chunk-XMOU4WEY.js";
+} from "./chunk-OOL3675H.js";
 import {
   RESEARCH_SUPERVISOR_SYSTEM_PROMPT,
   createAgentKnowledgeReadinessCheck,
@@ -48,7 +49,7 @@ import {
   knowledgeReadinessDeliverable,
   runKnowledgeImprovementJob,
   runSupervisedKnowledgeUpdate
-} from "./chunk-74SBMDTO.js";
+} from "./chunk-R2VAJGR3.js";
 import "./chunk-5QOB7H74.js";
 import {
   assertModelAllowed,
@@ -56,7 +57,7 @@ import {
   defineRuntimeHooks,
   notifyRuntimeDecisionPoint,
   notifyRuntimeHookEvent
-} from "./chunk-QSNSMJSZ.js";
+} from "./chunk-RYBVU4M3.js";
 import "./chunk-DPEUKJRO.js";
 import {
   INTELLIGENCE_WIRE_VERSION,
@@ -65,10 +66,10 @@ import {
   createOtelExporter,
   exportEvalRuns,
   loopEventToOtelSpan
-} from "./chunk-UD4BHQMI.js";
+} from "./chunk-IVGYLCFH.js";
 import "./chunk-BZF3KQ6G.js";
 import "./chunk-FVJ7M3DA.js";
-import "./chunk-7LO5GMAO.js";
+import "./chunk-CMYMTRGA.js";
 import {
   AgentEvalError,
   BackendTransportError,
@@ -1298,7 +1299,9 @@ function improvementDriver(opts) {
     kind: `improvement:${opts.generator.kind}`,
     async propose(ctx) {
       const findings = resolveFindings(ctx);
-      if (findings.length === 0 && ctx.report === void 0) return [];
+      if (findings.length === 0 && ctx.report === void 0 && !opts.generator.proposesWithoutFindings) {
+        return [];
+      }
       const surfaces = [];
       for (let i = 0; i < ctx.populationSize; i++) {
         if (ctx.signal.aborted) break;
@@ -1360,7 +1363,10 @@ function rawTraceDistiller(options = {}) {
     })).sort((a, b) => a.composite - b.composite).slice(0, maxCandidates);
     const totalFailingCells = ranked.reduce((n, c) => n + c.cells.length, 0);
     if (totalFailingCells === 0) {
-      return options.fallbackFindings ?? [
+      if (options.fallbackFindings && options.fallbackFindings.length > 0) {
+        return options.fallbackFindings;
+      }
+      return [
         makeFinding({
           analyst_id: ANALYST_ID,
           severity: "info",
@@ -1508,6 +1514,121 @@ function truncate(s, n) {
   return s.length <= n ? s : `${s.slice(0, n - 1)}\u2026`;
 }
 
+// src/improvement/rollout-policy.ts
+var ROLLOUT_POLICY_EXTENSION = "structural-rollout";
+var ROLLOUT_POLICY_BOUNDS = {
+  k: { min: 1, max: 10, step: 2 },
+  repairRounds: { min: 0, max: 3, step: 1 },
+  testgen: { min: 0, max: 10, step: 3 }
+};
+var MAX_CANDIDATES_PER_GENERATION = 4;
+var clamp = (v, min, max) => Math.min(max, Math.max(min, v));
+var isBoundedInt = (v, min) => typeof v === "number" && Number.isInteger(v) && v >= min;
+function parseRolloutPolicy(surface) {
+  if (typeof surface !== "string" || surface.trim().length === 0) return void 0;
+  let raw;
+  try {
+    raw = JSON.parse(surface);
+  } catch {
+    return void 0;
+  }
+  return normalizeRolloutPolicy(raw);
+}
+function normalizeRolloutPolicy(raw) {
+  if (typeof raw !== "object" || raw === null || Array.isArray(raw)) return void 0;
+  const bag = raw;
+  const k = bag.k ?? defaultStructuralRolloutPolicy.k;
+  const repairRounds = bag.repairRounds ?? defaultStructuralRolloutPolicy.repairRounds;
+  const testgen = bag.testgen ?? defaultStructuralRolloutPolicy.testgen;
+  if (!isBoundedInt(k, 1) || !isBoundedInt(repairRounds, 0) || !isBoundedInt(testgen, 0)) {
+    return void 0;
+  }
+  return {
+    k,
+    repairRounds,
+    testgen,
+    ...typeof bag.diverse === "boolean" ? { diverse: bag.diverse } : {},
+    ...typeof bag.temperature === "number" ? { temperature: bag.temperature } : {}
+  };
+}
+function serializeRolloutPolicy(policy) {
+  return JSON.stringify({
+    k: policy.k,
+    repairRounds: policy.repairRounds,
+    testgen: policy.testgen,
+    ...policy.diverse !== void 0 ? { diverse: policy.diverse } : {},
+    ...policy.temperature !== void 0 ? { temperature: policy.temperature } : {}
+  });
+}
+function structuralRolloutPolicyFromProfile(profile) {
+  const bag = profile.extensions?.[ROLLOUT_POLICY_EXTENSION];
+  if (bag === void 0) return void 0;
+  return normalizeRolloutPolicy(bag);
+}
+function applyRolloutPolicyToProfile(profile, policy) {
+  const bag = {
+    k: policy.k,
+    repairRounds: policy.repairRounds,
+    testgen: policy.testgen,
+    ...policy.diverse !== void 0 ? { diverse: policy.diverse } : {},
+    ...policy.temperature !== void 0 ? { temperature: policy.temperature } : {}
+  };
+  return {
+    ...profile,
+    extensions: { ...profile.extensions, [ROLLOUT_POLICY_EXTENSION]: bag }
+  };
+}
+function enumerateNeighborPolicies(policy) {
+  const moves = [
+    { dial: "k", delta: 1 },
+    { dial: "k", delta: -1 },
+    { dial: "repairRounds", delta: 1 },
+    { dial: "repairRounds", delta: -1 },
+    { dial: "testgen", delta: 1 },
+    { dial: "testgen", delta: -1 }
+  ];
+  const seen = /* @__PURE__ */ new Set([serializeRolloutPolicy(policy)]);
+  const neighbors = [];
+  for (const move of moves) {
+    const bounds = ROLLOUT_POLICY_BOUNDS[move.dial];
+    const next = clamp(policy[move.dial] + move.delta * bounds.step, bounds.min, bounds.max);
+    const candidate = { ...policy, [move.dial]: next };
+    const key = serializeRolloutPolicy(candidate);
+    if (seen.has(key)) continue;
+    seen.add(key);
+    neighbors.push(candidate);
+  }
+  return neighbors;
+}
+function candidateLabel(base, next) {
+  for (const dial of ["k", "repairRounds", "testgen"]) {
+    if (next[dial] !== base[dial]) return `${dial} ${base[dial]}\u2192${next[dial]}`;
+  }
+  return "unchanged";
+}
+function rolloutPolicyProposer() {
+  return {
+    kind: "rollout-policy",
+    async propose(ctx) {
+      const policy = parseRolloutPolicy(ctx.currentSurface);
+      if (!policy) return [];
+      const neighbors = enumerateNeighborPolicies(policy);
+      if (neighbors.length === 0) return [];
+      const cap = Math.max(1, Math.min(ctx.populationSize, MAX_CANDIDATES_PER_GENERATION));
+      const start = ctx.generation * cap % neighbors.length;
+      const window = [];
+      for (let i = 0; i < Math.min(cap, neighbors.length); i += 1) {
+        window.push(neighbors[(start + i) % neighbors.length]);
+      }
+      return window.map((candidate) => ({
+        surface: serializeRolloutPolicy(candidate),
+        label: candidateLabel(policy, candidate),
+        rationale: "bounded single-dial neighbor of the current structuralRollout policy; the held-out gate decides (deterministic enumeration \u2014 the dial space is tiny and prompt-style reflective proposals are a measured zero here)"
+      }));
+    }
+  };
+}
+
 // src/improvement/improve.ts
 var defaultReflectionModel = "deepseek-v4-flash";
 function llmClientOptions(llm) {
@@ -1520,6 +1641,8 @@ function defaultGeneratorFor(surface, llm) {
       return gepaProposer({ llm: llmClientOptions(llm), model, target: "agent system prompt" });
     case "skills":
       return skillOptProposer({ llm: llmClientOptions(llm), model, target: "agent skill document" });
+    case "rollout-policy":
+      return rolloutPolicyProposer();
     default:
       return void 0;
   }
@@ -1536,6 +1659,10 @@ function baselineSurfaceFor(profile, surface, skills) {
       return JSON.stringify(profile.mcp ?? {});
     case "hooks":
       return JSON.stringify(profile.hooks ?? {});
+    case "rollout-policy": {
+      const policy = structuralRolloutPolicyFromProfile(profile);
+      return policy ? serializeRolloutPolicy(policy) : "";
+    }
     case "code":
       return "";
   }
@@ -1609,6 +1736,15 @@ function applyWinnerToProfile(profile, surface, winner) {
       return { ...profile, mcp: parseWinnerJson(winner, surface) };
     case "hooks":
       return { ...profile, hooks: parseWinnerJson(winner, surface) };
+    case "rollout-policy": {
+      const policy = normalizeRolloutPolicy(parseWinnerJson(winner, surface));
+      if (!policy) {
+        throw new ConfigError(
+          `improve(): the shipped 'rollout-policy' winner is not a valid StructuralRolloutPolicy (integer k >= 1, repairRounds >= 0, testgen >= 0), so it cannot be applied: ${winner}`
+        );
+      }
+      return applyRolloutPolicyToProfile(profile, policy);
+    }
     case "code":
       return profile;
   }
@@ -3127,10 +3263,13 @@ export {
   NotFoundError,
   PlannerError,
   RESEARCH_SUPERVISOR_SYSTEM_PROMPT,
+  ROLLOUT_POLICY_BOUNDS,
+  ROLLOUT_POLICY_EXTENSION,
   RuntimeRunStateError,
   SqlConversationJournal,
   ValidationError,
   agenticGenerator,
+  applyRolloutPolicyToProfile,
   applyRunRecordDefaults,
   auditLoopRunner,
   buildForwardHeaders,
@@ -3155,6 +3294,7 @@ export {
   defineConversation,
   defineRuntimeHooks,
   deriveExecutionId,
+  enumerateNeighborPolicies,
   exportEvalRuns,
   formatSupervisedKnowledgeTask,
   getModels,
@@ -3170,9 +3310,11 @@ export {
   mcpServeVerifier,
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset,
+  normalizeRolloutPolicy,
   notifyRuntimeDecisionPoint,
   notifyRuntimeHookEvent,
   parseLoopRunnerArgv,
+  parseRolloutPolicy,
   rawTraceDistiller,
   readDepth,
   readinessServerSentEvent,
@@ -3181,6 +3323,7 @@ export {
   resolveAgentBackend,
   resolveChatModel,
   resolveRouterBaseUrl,
+  rolloutPolicyProposer,
   runAgentTask,
   runAgentTaskStream,
   runConversation,
@@ -3197,10 +3340,12 @@ export {
   sanitizeKnowledgeReadinessReport,
   sanitizeRuntimeStreamEvent,
   selfImproveLoopRunner,
+  serializeRolloutPolicy,
   sleep,
   slugifySpeaker,
   startRuntimeRun,
   streamToolLoop,
+  structuralRolloutPolicyFromProfile,
   toolBuildPrompt,
   turnId,
   validateChatModelId,