@tangle-network/agent-runtime 0.104.0 → 0.105.0

package/dist/{completion-gate-Bv47cOnn.d.ts → completion-gate-BMy5LGoP.d.ts}

@@ -1,5 +1,5 @@
 import { WorkspacePlanReceipt } from '@tangle-network/agent-profile-materialize';
-import { L as LocalHarness, C as CodexTokenUsage, c as CodexExecutionPolicy } from './local-harness-CtK9dSny.js';
+import { L as LocalHarness, C as CodexTokenUsage, a as CodexExecutionPolicy } from './local-harness-BDNcl6jI.js';
 import { f as Executor } from './types-CvNwMzQt.js';
 
 /**

package/dist/{coordination-CxtzMxSM.d.ts → coordination-BZZSVYpZ.d.ts}

@@ -6,8 +6,8 @@ import { a as SandboxClient, E as ExecCtx, t as LoopTraceEmitter, c as LoopTrace
 import { AgentProfile, BackendType, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
 import { b as ToolSpec, c as RuntimeTelemetryOptions, R as RouterConfig } from './sanitize-DEbPNtyI.js';
 import { R as RuntimeStreamEvent } from './types-BwoZWq-i.js';
-import { G as GitRunner, a as WorktreeCheckRunner, D as DeliverableSpec } from './completion-gate-Bv47cOnn.js';
-import { L as LocalHarness } from './local-harness-CtK9dSny.js';
+import { G as GitRunner, a as WorktreeCheckRunner, D as DeliverableSpec } from './completion-gate-BMy5LGoP.js';
+import { L as LocalHarness } from './local-harness-BDNcl6jI.js';
 import { ProviderExecutorOptions, AgentEnvironmentProviderRegistry } from './environment-provider.js';
 import { AgentEnvironmentProvider } from '@tangle-network/agent-interface/environment-provider';
 

package/dist/index.d.ts

@@ -1,41 +1,41 @@
-import { AgentEvalError, AnalystFinding, KnowledgeReadinessReport, RunRecord, ControlEvalResult } from '@tangle-network/agent-eval';
+import { AgentEvalError, AnalystFinding, ConfigError, KnowledgeReadinessReport, RunRecord, ControlEvalResult } 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 { c as createOpenAICompatibleBackend } from './backends-BO4Jqoe2.js';
 export { a as createIterableBackend, b as createSandboxPromptBackend } from './backends-BO4Jqoe2.js';
 export { AGENT_CANDIDATE_EXECUTION_SUPPORT, AgentCandidateBundleInput, AgentCandidateCodeSource, AgentCandidateCodeSurfaceSource, AgentCandidateProfileSource, AgentCandidateWorkspaceArchiveLimits, BuildAgentCandidateBundleInput, CANDIDATE_KNOWLEDGE_RETRIEVAL_CONFIG_ENV, CANDIDATE_KNOWLEDGE_ROOT_ENV, CaptureAgentCandidateWorkspaceOptions, CapturedAgentCandidateWorkspace, CreateAgentCandidateWorkspacePortOptions, DisposePreparedAgentCandidateOptions, ExactProcessCandidateExecutorOptions, FileAgentCandidateExecutionClaimStore, FileAgentCandidateExecutionClaimStoreOptions, RecoverExpiredAgentCandidateOptions, buildAgentCandidateBundle, candidateExecutionClaim, candidateKnowledgeExecutionPaths, captureAgentCandidateWorkspace, captureAgentCandidateWorkspaceFiles, createAgentCandidateWorkspacePort, disposePreparedAgentCandidateExecution, exactProcessProviderAsCandidateExecutor, persistCandidateOutputArtifact, recoverExpiredAgentCandidateExecution, sealAgentCandidateBundle, verifyAgentCandidateBundle } from './candidate-execution/index.js';
-export { a as AgentCandidateExecutionAttemptRecord, b as AgentCandidateExecutionAttemptRef, c as AgentCandidateExecutionClaim, d as AgentCandidateExecutionClaimResult, A as AgentCandidateExecutionClaimStore, e as AgentCandidateExecutionCleanupHandles, f as AgentCandidateExecutionFailureClass, g as AgentCandidateExecutionFinishResult, h as AgentCandidateExecutionLease, i as AgentCandidateExecutionPhase, j as AgentCandidateExecutionPhaseResult, k as AgentCandidateExecutionRecoveryEvidence, l as AgentCandidateExecutionStageResult, m as AgentCandidateExecutionTerminalRecord, n as AgentCandidateExecutionTerminalResult, o as AgentCandidateModelGrantActivateInput, q as AgentCandidateModelGrantClient, r as AgentCandidateModelGrantReservation, s as AgentCandidateModelGrantReserveInput, t as AgentCandidateModelGrantSettleInput, u as AgentCandidatePreparationEvidence, v as AgentCandidateRetryRejection, C as CreateProtectedAgentCandidateModelPortOptions, E as ExecutePreparedAgentCandidateOptions, I as InMemoryAgentCandidateExecutionClaimStore, P as PrepareAgentCandidateExecutionOptions, w as applyExactAgentProfileDiff, x as assertCandidateProfileBinding, y as createProtectedAgentCandidateModelPort, z as executePreparedAgentCandidate, B as parseExactAgentProfile, D as parseExactAgentProfileDiff, F as prepareAgentCandidateExecution } from './protected-model-port-DzkUY3HQ.js';
+export { a as AgentCandidateExecutionAttemptRecord, b as AgentCandidateExecutionAttemptRef, c as AgentCandidateExecutionClaim, d as AgentCandidateExecutionClaimResult, A as AgentCandidateExecutionClaimStore, e as AgentCandidateExecutionCleanupHandles, f as AgentCandidateExecutionFailureClass, g as AgentCandidateExecutionFinishResult, h as AgentCandidateExecutionLease, i as AgentCandidateExecutionPhase, j as AgentCandidateExecutionPhaseResult, k as AgentCandidateExecutionRecoveryEvidence, l as AgentCandidateExecutionStageResult, m as AgentCandidateExecutionTerminalRecord, n as AgentCandidateExecutionTerminalResult, o as AgentCandidateModelGrantActivateInput, q as AgentCandidateModelGrantClient, r as AgentCandidateModelGrantReservation, s as AgentCandidateModelGrantReserveInput, t as AgentCandidateModelGrantSettleInput, u as AgentCandidatePreparationEvidence, v as AgentCandidateRetryRejection, C as CreateProtectedAgentCandidateModelPortOptions, E as ExecutePreparedAgentCandidateOptions, I as InMemoryAgentCandidateExecutionClaimStore, P as PrepareAgentCandidateExecutionOptions, w as applyExactAgentProfileDiff, x as assertCandidateProfileBinding, y as createProtectedAgentCandidateModelPort, z as executePreparedAgentCandidate, B as parseExactAgentProfile, D as parseExactAgentProfileDiff, F as prepareAgentCandidateExecution } from './protected-model-port-B4t-OQjL.js';
 export { j as AgentCandidateArtifactPort, a as AgentCandidateBenchmarkGraderPort, k as AgentCandidateContainerPort, e as AgentCandidateExecutionPorts, l as AgentCandidateExecutorFinalCapture, m as AgentCandidateExecutorMemoryCapture, A as AgentCandidateExecutorPort, n as AgentCandidateExecutorProfileFile, o as AgentCandidateExecutorRequest, p as AgentCandidateExecutorStopRequest, q as AgentCandidateExecutorTaskOutcomeCapture, r as AgentCandidateExecutorWorkspaceFile, s as AgentCandidateExecutorWorkspaceInput, t as AgentCandidateMemoryPort, u as AgentCandidateMemoryResetResult, v as AgentCandidateModelLimits, f as AgentCandidateModelPort, b as AgentCandidateOutputArtifactPort, w as AgentCandidateOutputPurpose, h as AgentCandidateProtectedModelActivation, g as AgentCandidateProtectedModelReservation, i as AgentCandidateProtectedModelSettlement, x as AgentCandidateProtectedRunCapture, y as AgentCandidateRepositoryPort, c as AgentCandidateRunFinalization, d as AgentCandidateTaskExecution, z as AgentCandidateVerificationPorts, B as AgentCandidateWorkspacePort, C as CANDIDATE_TRACE_ENV, D as CANDIDATE_TRACE_TAGS, E as CanonicalCandidateDocument, P as PreparedAgentCandidateExecution, F as PreparedAgentCandidateInstruction, G as PreparedAgentCandidateKnowledge, H as PreparedAgentCandidateLaunch, I as PreparedAgentCandidateTrace, R as ResolvedAgentCandidateContainer, V as VerifiedAgentCandidate, J as VerifiedAgentCandidateTaskOutcome } from './types-Cbp72qyu.js';
 export { AuthSource, BackendCallPolicy, CircuitBreakerConfig, CircuitBreakerState, CircuitOpenError, Conversation, ConversationDriveState, ConversationJournal, ConversationJournalEntry, ConversationParticipant, ConversationPolicy, ConversationResult, ConversationStreamEvent, ConversationTurn, D1DatabaseLike, D1StmtLike, DEFAULT_MAX_DEPTH, DeadlineExceededError, FORWARD_HEADERS, FileConversationJournal, ForwardHeaderName, HaltContext, HaltPredicate, HaltReason, HaltSignal, InMemoryConversationJournal, PersonaConversationResult, PersonaDriver, PropagatedHeaders, RetryBackoff, RetryableErrorPredicate, RunConversationOptions, RunPersonaConfig, RunPersonaConversationOptions, SqlAdapter, SqlConversationJournal, TurnOrder, buildForwardHeaders, computeBackoff, createConversationBackend, d1ToSqlAdapter, defaultIsRetryable, defineConversation, isDepthExceeded, makePerAttemptSignal, readDepth, runConversation, runConversationStream, runPersonaConversation, runPersonaDispatch, sleep, slugifySpeaker, turnId } from './conversation.js';
-import { V as Verifier, C as CandidateGenerator } from './improve--2A0gK1-.js';
-export { A as AGENTIC_PROFILE_RESOURCE_ROOT, b as AgenticGeneratorOptions, c as AgenticGeneratorShotDisposition, d as AgenticGeneratorShotExecution, e as AgenticGeneratorShotReceipt, f as ImproveCodeOptions, I as ImproveOptions, a as ImproveResult, g as ImproveSkillsOptions, h as ImproveSurface, i as ImprovementCandidate, j as ImprovementDriverOptions, M as ManagedImprovementDriver, k as VerifyResult, l as agenticGenerator, m as commandVerifier, n as defaultBuildPrompt, o as improve, p as improvementDriver } from './improve--2A0gK1-.js';
-import { ProposeContext, SurfaceProposer, MutableSurface } from '@tangle-network/agent-eval/campaign';
-import { L as LocalHarness, r as runLocalHarness } from './local-harness-CtK9dSny.js';
+import { V as Verifier, C as CandidateGenerator, c as ImproveMethodOptions, b as ImproveMethodResult, e as ImproveCodeRunOptions, f as ImproveCodeResult, R as Redactor, g as ImproveCandidateValidationInput, h as ImproveMethodFactory, i as ReadonlyAgentProfile } from './redact-BuFjxBUJ.js';
+export { A as AGENTIC_PROFILE_RESOURCE_ROOT, j as AgenticGeneratorOptions, k as AgenticGeneratorShotDisposition, l as AgenticGeneratorShotExecution, m as AgenticGeneratorShotReceipt, D as DeepReadonly, n as ImproveCandidateValidator, o as ImproveCodeOptions, p as ImproveCost, q as ImproveLineage, s as ImproveMethodContext, t as ImproveMethodSource, u as ImproveOptimizationRunOptions, I as ImproveOptions, v as ImproveProfileAgent, w as ImproveProfileComponents, x as ImproveProfileSurface, a as ImproveResult, y as ImproveSkillsOptions, z as ImproveSurface, B as ImprovementCandidate, E as ImprovementCodeCandidate, F as ImprovementProfileCandidate, G as VerifyResult, H as agenticGenerator, J as commandVerifier, K as defaultBuildPrompt } from './redact-BuFjxBUJ.js';
+import { L as LocalHarness, r as runLocalHarness } from './local-harness-BDNcl6jI.js';
 import { T as ToolLoopChat, c as RuntimeTelemetryOptions } from './sanitize-DEbPNtyI.js';
 export { d as RuntimeEventCollector, e as RuntimeStreamEventCollector, S as SanitizedKnowledgeReadinessReport, f as createRuntimeEventCollector, g as createRuntimeStreamEventCollector, s as sanitizeAgentRuntimeEvent, h as sanitizeKnowledgeReadinessReport, i as sanitizeRuntimeStreamEvent } from './sanitize-DEbPNtyI.js';
-import { AgentProfile, AgentProfileDiff } from '@tangle-network/agent-interface';
+import { Scenario, SelfImproveOptions, MutableSurface } from '@tangle-network/agent-eval/contract';
+import { AgentProfile } from '@tangle-network/agent-interface';
 export { AgentCandidateBenchmarkGraderIdentity } from '@tangle-network/agent-interface';
-import { Scenario, SelfImproveOptions } from '@tangle-network/agent-eval/contract';
+import { GepaOptimizationMethodConfig, SkillOptOptimizationMethodConfig } from '@tangle-network/agent-eval/campaign';
 import { S as SurfaceImprovementEdit } from './improvement-adapter-HAZz-7vK.js';
 import { I as ImprovementProposalSource } from './types-CmYCMbFT.js';
-import { S as StructuralRolloutPolicy } from './structural-rollout-C4mabf87.js';
+import { S as StructuralRolloutPolicy } from './structural-rollout-DEh6CEsa.js';
 export { AgentKnowledgeReadinessCheckOptions, CreateKnowledgeImprovementActivationExecutorOptions, KnowledgeImprovementActivationExecutor, KnowledgeImprovementCandidatePair, KnowledgeImprovementExperimentBundles, KnowledgeImprovementJobMeasurement, KnowledgeImprovementJobResult, KnowledgeReadinessCheck, KnowledgeReadinessCheckInput, KnowledgeReadinessCheckResult, RESEARCH_SUPERVISOR_SYSTEM_PROMPT, RunKnowledgeImprovementJobOptions, SupervisedKnowledgeUpdateInput, SupervisedKnowledgeUpdateOptions, SupervisedKnowledgeUpdateResult, SupervisedKnowledgeUpdater, buildKnowledgeImprovementExperimentBundles, createAgentKnowledgeReadinessCheck, createKnowledgeImprovementActivationExecutor, 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-DUOGtvOl.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, w as worktreeLoopRunner } from './loop-runner-bin-68aoM9-7.js';
 export { m as mcpToolsForRuntimeMcp, a as mcpToolsForRuntimeMcpSubset } from './openai-tools-fnj6SRVg.js';
-export { b1 as EvalRunEvent, b2 as EvalRunGeneration, b3 as EvalRunsExportConfig, b4 as EvalRunsExportResult, b5 as INTELLIGENCE_WIRE_VERSION, b6 as LoopSpanNode, b7 as OtelAttribute, b8 as OtelExportConfig, b9 as OtelExporter, ba as OtelSpan, bb as RuntimeEventOtelOptions, bc as buildLoopOtelSpans, bd as buildLoopSpanNodes, be as buildRuntimeEventOtelSpans, bf as createOtelExporter, bg as exportEvalRuns, bh as loopEventToOtelSpan } from './coordination-CxtzMxSM.js';
+export { b1 as EvalRunEvent, b2 as EvalRunGeneration, b3 as EvalRunsExportConfig, b4 as EvalRunsExportResult, b5 as INTELLIGENCE_WIRE_VERSION, b6 as LoopSpanNode, b7 as OtelAttribute, b8 as OtelExportConfig, b9 as OtelExporter, ba as OtelSpan, bb as RuntimeEventOtelOptions, bc as buildLoopOtelSpans, bd as buildLoopSpanNodes, be as buildRuntimeEventOtelSpans, bf as createOtelExporter, bg as exportEvalRuns, bh as loopEventToOtelSpan } from './coordination-BZZSVYpZ.js';
 import { K as KnowledgeReadinessDecision, A as AgentBackendInput, c as AgentExecutionBackend, h as RunAgentTaskOptions, i as AgentTaskRunResult, j as RunAgentTaskStreamOptions, R as RuntimeStreamEvent, k as RuntimeSessionStore, l as RuntimeSession } from './types-BwoZWq-i.js';
 export { m as AgentAdapter, d as AgentBackendContext, n as AgentKnowledgeProvider, e as AgentRuntimeEvent, o as AgentRuntimeEventSink, p as AgentTaskContext, g as AgentTaskSpec, f as AgentTaskStatus, B as BackendErrorDetail, b as OpenAIChatResponseFormat, O as OpenAIChatTool, a as OpenAIChatToolChoice } from './types-BwoZWq-i.js';
 import { R as RuntimeHooks } from './types-BCoemcXU.js';
 export { d as RuntimeDecisionEvidenceRef, e as RuntimeDecisionKind, f as RuntimeDecisionPoint, g as RuntimeHookContext, h as RuntimeHookErrorContext, i as RuntimeHookEvent, j as RuntimeHookPhase, k as RuntimeHookTarget, l as RuntimeRunHandle, m as RuntimeRunPersistenceAdapter, n as RuntimeRunRow, o as composeRuntimeHooks, p as defineRuntimeHooks, q as notifyRuntimeDecisionPoint, r as notifyRuntimeHookEvent, s as startRuntimeRun } from './types-BCoemcXU.js';
 import '@tangle-network/agent-interface/environment-provider';
 import 'node:child_process';
-import './worktree-fanout-WsOzt8qu.js';
+import './worktree-fanout-DxhAWr5Z.js';
 import './types-CvNwMzQt.js';
 import '@tangle-network/sandbox';
-import './completion-gate-Bv47cOnn.js';
+import './completion-gate-BMy5LGoP.js';
 import '@tangle-network/agent-profile-materialize';
 import '@tangle-network/agent-knowledge';
 import './activation-Cm-9MMX1.js';
-import './supervise-7q9F5brx.js';
+import './supervise-4_48wwvy.js';
 import './kb-gate-CwHO0vz6.js';
 import './substrate-Dyiqmd3o.js';
 import './environment-provider.js';
@@ -267,75 +267,6 @@ declare function toolBuildPrompt(args: FindingsArg): string;
 /** Build the starting instruction for a coder agent tasked with implementing a new MCP server. */
 declare function mcpBuildPrompt(args: FindingsArg): string;
 
-/**
- * Campaign `spans.jsonl` → OTLP-flat JSONL — the missing converter between
- * what the substrate RECORDS and what its trace analysts READ.
- *
- * agent-eval's `runCampaign` durably records one `spans.jsonl` per cell
- * (`defaultBuildTraceWriter`): flat records
- * `{ name, cellId, startMs, durationMs?, ...attributes }` with no trace/span
- * ids. Its trace consumers (`OtlpFileTraceStore`, the trace-analyst registry,
- * `haloProposer`/`traceAnalystProposer` via `resolveTraces`) read OTLP-flat
- * JSONL (`trace_id`/`span_id`/ISO times/status/attributes — the shape
- * `projectOtlpFlatLine` parses). Nothing shipped converts between the two, so
- * the traces real optimization runs write could never reach the trace-native
- * proposers. This module is that wire adapter:
- *
- *   - {@link campaignCellSpansToOtlp} — one cell's `spans.jsonl` content →
- *     OTLP lines (a per-cell root AGENT anchor span + one child per record).
- *   - {@link convertCampaignDirToOtlp} — walk any campaign run/generation dir
- *     for `spans.jsonl` files and concatenate their OTLP lines.
- *   - {@link campaignTraceResolver} — the `resolveTraces` implementation for
- *     `traceAnalystProposer`/`haloProposer`: proposing generation g reads the
- *     traces the loop just recorded (`gen-<g-1>`, or `baseline` for g = 0)
- *     under the same `runDir` handed to `improve()`/`selfImprove()`.
- *
- * Trace identity: one trace per CELL, keyed on the cell's on-disk path — the
- * same sanitized `cellId` recurs across the baseline and every candidate
- * campaign, so folding the id alone would merge distinct runs into one trace.
- * Ids are deterministic FNV-1a folds to OTLP's 32/16-hex width, so re-converts
- * are stable and byte-identical.
- */
-
-interface CampaignOtlpOptions {
-    /** OTLP `service.name` on every emitted span. Default `'campaign'`. */
-    serviceName?: string;
-}
-/**
- * Convert ONE cell's `spans.jsonl` content to OTLP-flat JSONL lines.
- * `cellKey` is the identity the trace id folds from — pass the cell's on-disk
- * path (unique per campaign); `cellId` is the display/attribute label.
- * Returns `[]` for empty/recordless content (a dispatch that never touched
- * `ctx.trace`/`ctx.cost` writes an empty file — that is data, not an error).
- */
-declare function campaignCellSpansToOtlp(content: string, cell: {
-    cellId: string;
-    cellKey?: string;
-}, opts?: CampaignOtlpOptions): string[];
-/**
- * Walk `dir` (a campaign run dir, a generation dir, or a whole `selfImprove`
- * run root) for `spans.jsonl` files and return their concatenated OTLP-flat
- * JSONL — the exact string the `resolveTraces` contract expects. `''` when no
- * spans exist (the proposers fail loud on empty by design).
- */
-declare function convertCampaignDirToOtlp(dir: string, opts?: CampaignOtlpOptions): string;
-interface CampaignTraceResolverOptions extends CampaignOtlpOptions {
-    /** The `selfImprove`/`improve()` run root — the SAME `runDir` the loop
-     *  records under (`<runDir>/baseline/...`, `<runDir>/gen-<g>/candidate-<i>/...`).
-     *  Must be a real path; a `mem://` run records nothing to resolve. */
-    runDir: string;
-}
-/**
- * Build the `resolveTraces` function `traceAnalystProposer`/`haloProposer`
- * take: proposing generation g reads the traces of the campaigns the loop just
- * scored — `gen-<g-1>` (or `baseline` when g = 0), falling back to every trace
- * under the run root when that directory has none (e.g. a caller pointing at a
- * single campaign dir rather than a loop root).
- *
- *   traceAnalystProposer({ ..., resolveTraces: campaignTraceResolver({ runDir }) })
- */
-declare function campaignTraceResolver(opts: CampaignTraceResolverOptions): (ctx: Pick<ProposeContext, 'generation'>) => string;
-
 /**
  * `driverLoopGenerator` — the driver→worker `CandidateGenerator`: the build
  * loop run by the ATOM instead of the canned respawn.
@@ -442,6 +373,26 @@ interface ToAnalystFindingsOptions {
  */
 declare function toAnalystFindings(findings: readonly unknown[], opts?: ToAnalystFindingsOptions): AnalystFinding[];
 
+/**
+ * `improve` runs one complete optimization method against an exact profile
+ * surface. Runtime extracts and materializes the profile value; agent-eval owns
+ * optimization, disjoint data partitions, final-test scoring, and uncertainty.
+ *
+ * Code is the sole exception. It uses Runtime's isolated git worktrees because
+ * checkout ownership and cleanup cannot cross a generic optimizer boundary.
+ *
+ * @experimental
+ */
+
+/**
+ * Optimize one exact profile surface with a complete method.
+ */
+declare function improve<TScenario extends Scenario, TArtifact>(profile: AgentProfile, opts: ImproveMethodOptions<TScenario, TArtifact>): Promise<ImproveMethodResult>;
+/**
+ * Optimize repository code through Runtime's isolated worktree path.
+ */
+declare function improve<TScenario extends Scenario, TArtifact>(opts: ImproveCodeRunOptions<TScenario, TArtifact>): Promise<ImproveCodeResult<TScenario, TArtifact>>;
+
 /**
  * `mcpServeVerifier` — the intrinsic verifier for a built MCP server: the
  * boot-and-probe checker named in docs/artifact-lifecycle-frontier.md. A
@@ -475,6 +426,58 @@ interface McpServeSpec {
 /** Build a `Verifier` that boots a generated MCP server over stdio and checks it exposes tools. */
 declare function mcpServeVerifier(spec: McpServeSpec): Verifier;
 
+/** Runtime context appended to an official optimizer's own configuration. */
+interface OfficialOptimizerContextOptions {
+    /** Context supplied to the optimizer before Runtime appends the profile surface and findings. */
+    background?: string;
+    /** Include current trace or analyst findings in the optimizer background. Default true. */
+    includeFindings?: boolean;
+    /** Reject oversized serialized findings before starting Python. Default 50,000 characters. */
+    maxFindingsChars?: number;
+    /**
+     * Redact caller-supplied context and descriptors before they leave Runtime.
+     * The built-in redactor is the default. Pass `false` only for public data
+     * that has already been reviewed.
+     */
+    redact?: Redactor | false;
+    /** Authorize one exact candidate containing structurally sensitive fields.
+     * The callback must return true for every accepted baseline and candidate. */
+    authorizeSensitiveCandidate?: (input: OfficialSensitiveCandidateInput) => boolean;
+}
+interface OfficialSensitiveCandidateInput extends ImproveCandidateValidationInput {
+    sensitivePaths: readonly string[];
+}
+/** Official GEPA configuration plus bounded Runtime findings context. */
+type OfficialGepaOptions<TScenario extends {
+    id: string;
+    kind: string;
+}, TArtifact = unknown> = Omit<GepaOptimizationMethodConfig<TScenario, TArtifact>, 'background' | 'evaluationId'> & OfficialOptimizerContextOptions;
+/** Official SkillOpt configuration plus bounded Runtime findings context. */
+type OfficialSkillOptOptions<TScenario extends {
+    id: string;
+    kind: string;
+}, TArtifact = unknown> = Omit<SkillOptOptimizationMethodConfig<TScenario, TArtifact>, 'background' | 'evaluationId'> & OfficialOptimizerContextOptions;
+/** Missing optional Python dependencies for an official optimizer. */
+declare class OfficialOptimizerUnavailableError extends ConfigError {
+    readonly optimizer: 'gepa' | 'skillopt';
+    constructor(optimizer: 'gepa' | 'skillopt', cause: unknown);
+}
+/**
+ * Build a complete method backed by GEPA's official Optimize Anything API.
+ *
+ * The recipe is passed through unchanged. Use `engine`, `sequential`,
+ * `adaptive-sequential`, `best-of`, `vote`, or `omni` explicitly.
+ */
+declare function officialGepa<TScenario extends {
+    id: string;
+    kind: string;
+}, TArtifact = unknown>(options: OfficialGepaOptions<TScenario, TArtifact>): ImproveMethodFactory<TScenario, TArtifact>;
+/** Build a complete method backed by Microsoft's official SkillOpt trainer. */
+declare function officialSkillOpt<TScenario extends {
+    id: string;
+    kind: string;
+}, TArtifact = unknown>(options: OfficialSkillOptOptions<TScenario, TArtifact>): ImproveMethodFactory<TScenario, TArtifact>;
+
 /**
  * The senior scientific-method optimizer doctrine — the ONE substantial prompt
  * core shared by every builder/author surface (tool build, MCP build, codebase
@@ -515,19 +518,6 @@ declare const researchDriverNote: string;
  */
 declare const strategyAuthorMethod: string;
 
-type ProfileDiffProposerContext<TFindings = unknown> = ProposeContext<TFindings> & {
-    profile: AgentProfile;
-};
-interface ProfileDiffProposerOptions<TFindings = unknown> {
-    proposeDiffs(context: ProfileDiffProposerContext<TFindings>): Promise<readonly AgentProfileDiff[]> | readonly AgentProfileDiff[];
-}
-/**
- * Turn exact AgentProfileDiffs from any source into full profile candidates for
- * the shared optimization loop. Research, catalogs, humans, and trace miners
- * differ only in `proposeDiffs`; measurement and promotion stay identical.
- */
-declare function profileDiffProposer<TFindings = unknown>(options: ProfileDiffProposerOptions<TFindings>): SurfaceProposer<TFindings>;
-
 /**
  *
  * `rawTraceDistiller` — the meta-harness `analyzeGeneration` producer.
@@ -585,10 +575,11 @@ interface RawTraceDistillerOptions {
  * FILESYSTEM CONTEXT — paths into the prior generation's real run traces plus a
  * grep/cat-to-diagnose instruction — instead of a pre-summarized digest.
  *
- * Drop-in for `opts.analyzeGeneration` on `improve()` / `selfImprove()`:
+ * Drop-in for `analyzeGeneration` on `improve({ surface: 'code' })`:
  *
- *   await improve(profile, seedFindings, {
+ *   await improve({
  *     surface: 'code',
+ *     findings: seedFindings,
  *     code: { repoRoot },
  *     runDir: '/abs/run',                 // MUST be a real path — the traces live here
  *     analyzeGeneration: rawTraceDistiller(),
@@ -605,9 +596,8 @@ declare function rawTraceDistiller<TScenario extends Scenario = Scenario, TArtif
  * the candidate worktree. `maxShots` is ignored — reflection is single-shot by
  * construction (the patches are already drafted).
  *
- * This is the `shots=1, sandbox=off` setting of the one improvement driver.
- * The `agenticGenerator` (a multi-shot verify-in-session loop) is the
- * `shots=N` setting — both plug into the same `improvementDriver`.
+ * This is the `shots=1, sandbox=off` code-candidate setting.
+ * `agenticGenerator` supplies the multi-shot verify-in-session setting.
  *
  * @experimental
  */
@@ -618,60 +608,13 @@ 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
- */
+/** Serialize and apply Runtime's structural rollout profile coordinate. */
 
 /** 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). */
+/** Parse a serialized policy surface. Returns `undefined` for non-strings,
+ * malformed JSON, or values outside the policy invariants. Unknown fields are
+ * dropped; supported optional fields are preserved. */
 declare function parseRolloutPolicy(surface: MutableSurface): StructuralRolloutPolicy | undefined;
 /** Normalize an untyped policy bag (a parsed surface or a profile extension) into
  *  a full `StructuralRolloutPolicy`, defaults merged. Returns `undefined` when any
@@ -679,32 +622,13 @@ declare function parseRolloutPolicy(surface: MutableSurface): StructuralRolloutP
  *  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). */
+/** Stable serialization with fixed field order. */
 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): 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, policy: StructuralRolloutPolicy): AgentProfile;
-/** 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;
+ *  not opt into structural rollout. */
+declare function structuralRolloutPolicyFromProfile(profile: ReadonlyAgentProfile): StructuralRolloutPolicy | undefined;
+/** Persist a detached policy under the profile extension without mutating the input. */
+declare function applyRolloutPolicyToProfile(profile: ReadonlyAgentProfile, policy: StructuralRolloutPolicy): AgentProfile;
 
 /**
  *
@@ -1131,4 +1055,4 @@ 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 { AgentBackendInput, type AgentBackendKind, AgentExecutionBackend, AgentTaskRunResult, BackendTransportError, type CampaignOtlpOptions, type CampaignTraceResolverOptions, CandidateGenerator, type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, DEFAULT_ROUTER_BASE_URL, type DriverLoopGeneratorOptions, InMemoryRuntimeSessionStore, LIFTED_FINDING_ANALYST_ID, type McpServeSpec, type ModelInfo, PlannerError, type ProfileDiffProposerContext, type ProfileDiffProposerOptions, ROLLOUT_POLICY_BOUNDS, ROLLOUT_POLICY_EXTENSION, type RawTraceDistillerOptions, type ReflectiveGeneratorOptions, type ResolveAgentBackendOptions, type ResolvedChatModel, type RouterEnv, type RunChatTurnInput, type RunToolLoopOptions, RuntimeHooks, RuntimeRunStateError, RuntimeSessionStore, RuntimeStreamEvent, RuntimeTelemetryOptions, type StreamToolLoopOptions, type StreamToolLoopYield, type ToAnalystFindingsOptions, type ToolCallOutcome, type ToolLoopAssistantToolCall, type ToolLoopCall, type ToolLoopEvent, type ToolLoopMessage, type ToolLoopResult, type ToolLoopStopReason, Verifier, applyRolloutPolicyToProfile, applyRunRecordDefaults, buildDriverSystem, campaignCellSpansToOtlp, campaignTraceResolver, cleanModelId, convertCampaignDirToOtlp, createOpenAICompatibleBackend, decideKnowledgeReadiness, deriveExecutionId, driverLoopGenerator, enumerateNeighborPolicies, findingLines, getModels, handleChatTurn, isAnalystFinding, mcpBuildPrompt, mcpServeVerifier, normalizeRolloutPolicy, optimizerMethod, parseRolloutPolicy, profileDiffProposer, rawTraceDistiller, readinessServerSentEvent, reflectiveGenerator, researchDriverNote, resolveAgentBackend, resolveChatModel, resolveRouterBaseUrl, rolloutPolicyProposer, runAgentTask, runAgentTaskStream, runToolLoop, runtimeStreamServerSentEvent, serializeRolloutPolicy, strategyAuthorMethod, streamToolLoop, structuralRolloutPolicyFromProfile, toAnalystFindings, toolBuildPrompt, validateChatModelId };
+export { AgentBackendInput, type AgentBackendKind, AgentExecutionBackend, AgentTaskRunResult, BackendTransportError, CandidateGenerator, type ChatStreamEvent, type ChatTurnHooks, type ChatTurnIdentity, type ChatTurnProducer, type ChatTurnResult, DEFAULT_ROUTER_BASE_URL, type DriverLoopGeneratorOptions, ImproveCandidateValidationInput, ImproveCodeResult, ImproveCodeRunOptions, ImproveMethodFactory, ImproveMethodOptions, ImproveMethodResult, InMemoryRuntimeSessionStore, LIFTED_FINDING_ANALYST_ID, type McpServeSpec, type ModelInfo, type OfficialGepaOptions, type OfficialOptimizerContextOptions, OfficialOptimizerUnavailableError, type OfficialSensitiveCandidateInput, type OfficialSkillOptOptions, PlannerError, ROLLOUT_POLICY_EXTENSION, type RawTraceDistillerOptions, ReadonlyAgentProfile, type ReflectiveGeneratorOptions, type ResolveAgentBackendOptions, type ResolvedChatModel, type RouterEnv, type RunChatTurnInput, type RunToolLoopOptions, RuntimeHooks, RuntimeRunStateError, RuntimeSessionStore, RuntimeStreamEvent, RuntimeTelemetryOptions, type StreamToolLoopOptions, type StreamToolLoopYield, type ToAnalystFindingsOptions, type ToolCallOutcome, type ToolLoopAssistantToolCall, type ToolLoopCall, type ToolLoopEvent, type ToolLoopMessage, type ToolLoopResult, type ToolLoopStopReason, Verifier, applyRolloutPolicyToProfile, applyRunRecordDefaults, buildDriverSystem, cleanModelId, createOpenAICompatibleBackend, decideKnowledgeReadiness, deriveExecutionId, driverLoopGenerator, findingLines, getModels, handleChatTurn, improve, isAnalystFinding, mcpBuildPrompt, mcpServeVerifier, normalizeRolloutPolicy, officialGepa, officialSkillOpt, optimizerMethod, parseRolloutPolicy, rawTraceDistiller, readinessServerSentEvent, reflectiveGenerator, researchDriverNote, resolveAgentBackend, resolveChatModel, resolveRouterBaseUrl, runAgentTask, runAgentTaskStream, runToolLoop, runtimeStreamServerSentEvent, serializeRolloutPolicy, strategyAuthorMethod, streamToolLoop, structuralRolloutPolicyFromProfile, toAnalystFindings, toolBuildPrompt, validateChatModelId };