@tangle-network/agent-runtime 0.89.0 → 0.90.1

package/dist/loops.d.ts

@@ -2,27 +2,30 @@ import { ChatClient, RunRecord, HarnessType, AgentProfile, AnalystFinding, Analy
 export { AnalystFinding, DefaultVerdict, computeFindingId, makeFinding } from '@tangle-network/agent-eval';
 import { SandboxEvent, SandboxInstance, CreateSandboxOptions, PromptOptions, TaskOptions } from '@tangle-network/sandbox';
 export { AgentProfile, CreateSandboxOptions, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
-import { d as ResultBlobStore, S as SpawnJournal, N as NodeId, j as SpawnEvent, E as ExecutorFactory, c as Agent, B as Budget, i as Scope, g as Settled, f as SupervisedResult, h as Spend, U as UsageEvent, b as ExecutorRegistry, k as Supervisor } from './types-Driepl87.js';
-export { A as AgentSpec, a as Executor, l as ExecutorContext, m as ExecutorResult, R as Runtime, n as SupervisorOpts, T as TreeView, W as WidenGate } from './types-Driepl87.js';
-import { ak as MakeWorkerAgent, A as AnalystRegistry, o as CoordinationTools, n as CoordinationEvent, ar as QuestionPolicy, a$ as ExecutorConfig } from './coordination-CuDLO8wj.js';
-export { b0 as BusEvent, b1 as BusRecord, b2 as BusStats, b3 as EventBus, b4 as ProviderSeam, b5 as PublishOptions, b6 as cliWorktreeExecutor, b7 as createEventBus, b8 as createExecutor, b9 as createExecutorRegistry } from './coordination-CuDLO8wj.js';
-import { R as RuntimeHooks, I as Iteration, S as SandboxClient, W as Driver, A as AgentRunSpec, b as OutputAdapter, V as Validator, E as ExecCtx, X as LoopWinner, Y as LoopLineageOptions, Z as LoopResult, L as LoopTokenUsage, a as RuntimeStreamEvent, _ as MountRecorder, i as AgentExecutionBackend, o as AgentTaskStatus, B as BackendErrorDetail } from './types-ESeMOj94.js';
-export { $ as LoopDecisionPayload, a0 as LoopEndedPayload, a1 as LoopIterationDispatchPayload, a2 as LoopIterationEndedPayload, a3 as LoopIterationStartedPayload, a4 as LoopPlanDescription, a5 as LoopPlanPayload, d as LoopSandboxPlacement, a6 as LoopStartedPayload, a7 as LoopTeardownFailedPayload, e as LoopTraceEmitter, c as LoopTraceEvent, a8 as MountManifestEntry, a9 as RunProvenance, aa as SelectionReceipt, ab as ValidationCtx } from './types-ESeMOj94.js';
+import { a as ResultBlobStore, g as SpawnJournal, N as NodeId, j as SpawnEvent, E as ExecutorFactory, A as Agent, B as Budget, d as Settled, S as Scope, b as SupervisedResult, e as Spend, U as UsageEvent, h as ExecutorRegistry, k as Supervisor } from './types-eMNgWgFi.js';
+export { f as AgentSpec, c as Executor, l as ExecutorContext, m as ExecutorResult, R as Runtime, n as SupervisorOpts, T as TreeView, W as WidenGate } from './types-eMNgWgFi.js';
+import { M as MakeWorkerAgent, A as AnalystRegistry, o as CoordinationTools, n as CoordinationEvent, as as QuestionPolicy, E as ExecutorConfig } from './coordination-rRj5hjJK.js';
+export { bf as BusEvent, bg as BusRecord, bh as BusStats, bi as EventBus, bj as ProviderSeam, bk as PublishOptions, bl as cliWorktreeExecutor, bm as createEventBus, bn as createExecutor, bo as createExecutorRegistry } from './coordination-rRj5hjJK.js';
+import { R as RuntimeHooks, I as Iteration, S as SandboxClient, D as Driver, A as AgentRunSpec, c as OutputAdapter, V as Validator, E as ExecCtx, Y as LoopWinner, Z as LoopLineageOptions, _ as LoopResult, L as LoopTokenUsage, b as RuntimeStreamEvent, $ as MountRecorder, j as AgentExecutionBackend, p as AgentTaskStatus, B as BackendErrorDetail } from './types-SyuwunY_.js';
+export { a0 as LoopDecisionPayload, a1 as LoopEndedPayload, a2 as LoopIterationDispatchPayload, a3 as LoopIterationEndedPayload, a4 as LoopIterationStartedPayload, a5 as LoopPlanDescription, a6 as LoopPlanPayload, e as LoopSandboxPlacement, a7 as LoopStartedPayload, a8 as LoopTeardownFailedPayload, f as LoopTraceEmitter, d as LoopTraceEvent, a9 as MountManifestEntry, aa as RunProvenance, a as SelectionReceipt, ab as ValidationCtx } from './types-SyuwunY_.js';
 import { RunProfileMatrixResult, Scenario, ProfileDispatchFn, JudgeConfig, RunProfileMatrixOptions, DispatchFn } from '@tangle-network/agent-eval/campaign';
 export { AgentEnvironmentProviderRef, AgentEnvironmentProviderRegistry, ProviderAsSandboxClientOptions, ProviderExecutorOptions, SandboxClientProviderOptions, createAgentEnvironmentProviderRegistry, providerAsExecutor, providerAsSandboxClient, resolveAgentEnvironmentProvider, sandboxClientAsProvider } from './environment-provider.js';
-import { C as CorpusRecord, c as Corpus, O as Outcome, S as ScopeAnalyzeInput, d as AssertTraceDerivedFindings, e as SteerContext, f as ScopeAnalyst, F as FanoutOptions, g as CombinatorShape, h as ScopeWidenGate, L as LoopUntilSpec, P as PanelSpec, i as PipelineStage, W as WinnerStrategy, j as FanoutWinnerSelector, V as VerifySpec, k as WidenSpec, l as CorpusFilter, R as RenderCorpusToInstructionsOptions, D as DefinePersonaInput, m as Persona, n as RunPersonifiedOptions, o as ShapeRegistry, p as LoopShape, E as EqualKArm, q as EqualKOnCostOptions, r as EqualKVerdict, T as TrajectoryReportOptions, s as TrajectoryReport } from './worktree-fanout-D6xR2CIA.js';
-export { A as AuthoredHarness, t as DefinePersona, u as EqualKOnCost, v as Fanout, w as FanoutSynthesis, x as FlatWidenGate, y as LoopUntil, z as LoopUntilState, B as Panel, G as PanelJudge, H as PanelVerdict, I as PatchDeliverableOptions, J as PersonaContext, K as PersonaExecutors, M as Pipeline, N as RenderCorpusToInstructions, Q as RunPersonified, U as ShapeBudget, X as ShapeContext, Y as TrajectoryNode, Z as TrajectoryReportFn, _ as Verify, $ as Widen, a0 as WidenDecision, a1 as WidenLineage, a2 as WorktreeCliExecutorOptions, a as WorktreeFanoutOptions, b as WorktreePatchArtifact, a3 as createWorktreeCliExecutor, a4 as patchDelivered, a5 as worktreeFanout } from './worktree-fanout-D6xR2CIA.js';
+import { c as CorpusRecord, C as Corpus, S as ScopeAnalyzeInput, d as AssertTraceDerivedFindings, O as Outcome, e as SteerContext, f as ScopeAnalyst, F as FanoutOptions, g as CombinatorShape, h as ScopeWidenGate, L as LoopUntilSpec, P as PanelSpec, i as PipelineStage, W as WinnerStrategy, j as FanoutWinnerSelector, V as VerifySpec, k as WidenSpec, l as CorpusFilter, R as RenderCorpusToInstructionsOptions, D as DefinePersonaInput, m as Persona, n as RunPersonifiedOptions, o as ShapeRegistry, p as LoopShape, E as EqualKArm, q as EqualKOnCostOptions, r as EqualKVerdict, T as TrajectoryReportOptions, s as TrajectoryReport } from './worktree-fanout-BDFQIO-Y.js';
+export { A as AuthoredHarness, t as DefinePersona, u as EqualKOnCost, v as Fanout, w as FanoutSynthesis, x as FlatWidenGate, y as LoopUntil, z as LoopUntilState, B as Panel, G as PanelJudge, H as PanelVerdict, I as PatchDeliverableOptions, J as PersonaContext, K as PersonaExecutors, M as Pipeline, N as RenderCorpusToInstructions, Q as RunPersonified, U as ShapeBudget, X as ShapeContext, Y as TrajectoryNode, Z as TrajectoryReportFn, _ as Verify, $ as Widen, a0 as WidenDecision, a1 as WidenLineage, a2 as WorktreeCliExecutorOptions, a as WorktreeFanoutOptions, b as WorktreePatchArtifact, a3 as createWorktreeCliExecutor, a4 as patchDelivered, a5 as worktreeFanout } from './worktree-fanout-BDFQIO-Y.js';
+import { A as AgenticSurface, a as AgenticTask, b as AgenticOptions, c as Strategy, d as ArtifactHandle, e as SurfaceScore, f as AgenticTool } from './structural-rollout-MwlpgQ-6.js';
+export { g as AgenticRunResult, C as CheckExecChannel, h as CheckOutcome, i as CheckRunContext, j as CheckRunner, k as CheckSource, l as CheckSourceCtx, m as CorpusReadbackOptions, R as RepairStop, n as RunAgenticOptions, o as ShotPersona, p as ShotSpec, q as StrategyCtx, r as StrategyResult, s as StructuralRolloutConfig, S as StructuralRolloutPolicy, t as StructuralRolloutResult, V as VisibleCheck, u as adaptiveRefine, v as breadthStrategy, w as canDisplace, x as compareCheckOutcomes, y as composeCheckSources, z as defaultExtractCandidate, B as defaultStructuralRolloutPolicy, D as defineStrategy, E as depthStrategy, F as filterAuthoredAsserts, G as modelAuthoredChecks, H as officialChecksFromMeta, I as refine, J as resolveEntrySymbol, K as runAgentic, L as sample, M as sampleThenRefine, N as sandboxCheckRunner, O as selectBestIndex, P as structuralRollout, Q as visibleCheckScore } from './structural-rollout-MwlpgQ-6.js';
 import { b as AnalystRegistryLike } from './types-BC3bZpH0.js';
 import { AgentProfile as AgentProfile$1 } from '@tangle-network/agent-interface';
-import { R as RouterConfig, a as ToolLoopChat, b as ToolLoopCompactionOptions } from './router-client-r8y_VFVM.js';
-export { c as RouterChatResult, d as RouterChatToolsResult, e as RouterToolCall, f as RouterToolLoopResult, g as ToolLoopCompaction, T as ToolSpec, r as routerBrain, h as routerChatWithTools, i as routerChatWithUsage, j as routerToolLoop } from './router-client-r8y_VFVM.js';
-import { D as DeliverableSpec } from './worktree-harness-CX_McRLp.js';
-export { g as WorktreeCommandResult, h as gateOnDeliverable } from './worktree-harness-CX_McRLp.js';
+import { R as RouterConfig, T as ToolLoopChat, a as ToolLoopCompactionOptions } from './router-client-DJImUDlm.js';
+export { c as RouterChatResult, d as RouterChatToolsResult, e as RouterToolCall, f as RouterToolLoopResult, g as ToolLoopCompaction, b as ToolSpec, r as routerBrain, h as routerChatWithTools, i as routerChatWithUsage, j as routerToolLoop } from './router-client-DJImUDlm.js';
+import { D as DeliverableSpec } from './completion-gate-DkAnUmpb.js';
+export { g as WorktreeCommandResult, h as gateOnDeliverable } from './completion-gate-DkAnUmpb.js';
+import { S as SupervisorProfile } from './supervise-DPmYPk0j.js';
+export { D as DriveHarness, a as SuperviseOptions, b as SupervisorAgentDeps, s as supervise, c as supervisorAgent, w as workerFromBackend } from './supervise-DPmYPk0j.js';
 import { stuckLoopView, toolWasteView } from '@tangle-network/agent-eval/pipelines';
 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 './substrate-DO2GHNg2.js';
-import './otel-export-BKmNwiCb.js';
-import './local-harness-sI0S_XNA.js';
+import './local-harness-dcD5WTTr.js';
 import 'node:child_process';
 
 /**
@@ -1059,235 +1062,6 @@ declare function loopCampaignDispatch<Task, Output, Decision, TScenario extends
  */
 declare function loopDispatch<Task, Output, Decision, TScenario extends Scenario, TArtifact>(opts: LoopDispatchOptions<Task, Output, Decision, TScenario, TArtifact>): ProfileDispatchFn<TScenario, TArtifact>;
 
-/**
- * The general agentic primitive — sequential (depth) and parallel (breadth) over a shared,
- * checkable artifact, driven through the keystone Supervisor as one recursive `Agent.act`.
- *
- * The domain lives behind ONE seam — `AgenticSurface` (open an artifact, list tools, call a tool,
- * score the artifact, close it). EnterpriseOps implements it (seed a gym DB, MCP tools, SQL
- * verifier); Commit0/AppWorld/terminal-bench implement it the same way (a repo workspace, shell
- * tools, the test suite). The drivers below are domain-blind: they run over any surface.
- *
- * Two shapes, the agent's POMDP rollout as the unit:
- *  - DEPTH   one persistent artifact carried across shots. Each shot the agent works the tool loop;
- *            between shots a trace-analyst (selector≠judge: reads the trajectory, never the score)
- *            steers the resumed session toward what's unfinished. shot n stands on shot n-1's
- *            artifact state + history. This is continuation — long-horizon, same artifact.
- *  - BREADTH K independent artifacts, each a fresh rollout, the deployable verifier picks the best.
- *
- * Both are an `Agent` whose `act` spawns leaf shots through `scope.spawn` and reacts via
- * `scope.next()` — so the conserved budget pool meters them (equal-k by construction), the journal
- * records the tree, and the same primitive nests. `runAgentic` runs the chosen driver through
- * `createSupervisor().run`. The leaf (one shot over a handle) is resolved per-spawn from a
- * surface-closed registry — the open `Executor` seam, not bespoke per-benchmark glue.
- */
-
-interface AgenticTask {
-    readonly id: string;
-    readonly systemPrompt: string;
-    readonly userPrompt: string;
-    /** Opaque domain payload the surface reads (EOPS: servers/verifiers/tools). Drivers never read it. */
-    readonly meta?: Record<string, unknown>;
-}
-interface ArtifactHandle {
-    readonly id: string;
-    readonly surface: string;
-    /** Opaque per-artifact context the surface stashes (EOPS: the seeded gym server + db id). */
-    readonly ctx?: unknown;
-}
-interface AgenticTool {
-    readonly type: 'function';
-    readonly function: {
-        name: string;
-        description?: string;
-        parameters: Record<string, unknown>;
-    };
-}
-interface SurfaceScore {
-    passes: number;
-    total: number;
-    /** Checks excluded as malformed (data defect, not the agent). `total === 0` ⇒ unscoreable. */
-    errored: number;
-}
-/** A stateful, checkable environment an agent operates over with tools. Open behind one interface. */
-interface AgenticSurface {
-    readonly name: string;
-    open(task: AgenticTask): Promise<ArtifactHandle>;
-    tools(task: AgenticTask, handle: ArtifactHandle): Promise<AgenticTool[]>;
-    call(handle: ArtifactHandle, name: string, args: Record<string, unknown>): Promise<string>;
-    score(task: AgenticTask, handle: ArtifactHandle): Promise<SurfaceScore>;
-    close(handle: ArtifactHandle): Promise<void>;
-}
-interface AgenticOptions {
-    routerBaseUrl: string;
-    routerKey: string;
-    model: string;
-    /** Optional completion transport (see `RouterConfig.complete`): when set, BOTH legs of an
-     *  offline run use it instead of `fetch`-ing the router — the worker's tool loop (threaded into
-     *  its `routerToolLoop` cfg) AND the analyst's critic (its `ChatClient` is bound to this same
-     *  transport). One injected responder serves both, as a localhost mock endpoint would. Absent ⇒
-     *  the live router fetch path (the default). */
-    complete?: (body: Record<string, unknown>) => Promise<unknown>;
-    temperature?: number;
-    /** Completion cap per worker turn — REQUIRED for thinking models (they burn unbounded
-     *  budgets on reasoning and return empty content without it). Omitted ⇒ provider default. */
-    maxTokens?: number;
-    /** Turns the agent may take within ONE shot before the driver intervenes. */
-    innerTurns?: number;
-    /** The depth STEERER's analyst instruction (observe()'s system prompt). The knob a
-     *  prompt optimizer (GEPA) tunes — the analyst IS the steerer. Omitted ⇒ the default. */
-    analystInstruction?: string;
-    /** The critic's model — lets the analyst be a stronger (or cheaper) model than the
-     *  worker. Omitted ⇒ the worker's `model`. */
-    analystModel?: string;
-    /** Across-run learning: when set, the analyst's observe() pass appends trace-derived
-     *  facts here (the flywheel write side). Priming (the read side) is the caller's move —
-     *  query the corpus and fold facts into the task's systemPrompt before runAgentic. */
-    corpus?: Corpus;
-    /** Tags written onto learned facts (and used by the caller's priming query). */
-    corpusTags?: string[];
-}
-type Msg = Record<string, unknown>;
-interface ShotResult {
-    messages: Msg[];
-    score: number;
-    passes: number;
-    total: number;
-    completions: number;
-    toolErrors: number;
-}
-interface AgenticRunResult {
-    /** The strategy name (built-in 'depth'/'breadth' or a custom strategy's name). */
-    mode: string;
-    score: number;
-    resolved: boolean;
-    completions: number;
-    /** DEPTH: score after each shot — the progress-over-rounds curve. BREADTH: best-so-far per rollout. */
-    progression: number[];
-    shots: number;
-    /** The cost vector, stamped by `runAgentic` from the Supervisor's conserved pool: real
-     *  router tokens, priced usd (0 when the model is unpriced — never fabricated), wall ms. */
-    usd: number;
-    ms: number;
-    tokens: {
-        input: number;
-        output: number;
-    };
-}
-/** DEPTH: one persistent artifact, carried across analyst-steered shots. */
-declare function depthStrategy(surface: AgenticSurface, task: AgenticTask, opts: AgenticOptions, cfg: {
-    maxShots: number;
-}): Agent<unknown, Outcome<unknown>>;
-/** BREADTH: K independent rollouts (each own artifact), verifier picks the best. */
-declare function breadthStrategy(_surface: AgenticSurface, task: AgenticTask, opts: AgenticOptions, cfg: {
-    width: number;
-}): Agent<unknown, Outcome<unknown>>;
-/**
- * A Strategy is HOW you spend the compute budget to beat the Environment's check — it
- * builds the driver `Agent` the Supervisor runs. This is the OPEN extension point: a dev
- * authors their own by implementing `driver()` to return an Agent whose `act()` spawns
- * shots/analysts via `scope.spawn` / `scope.next` / `scope.send`. The two built-ins are
- * the reference implementations to copy:
- *   sample — K INDEPENDENT attempts, keep the best-verifying (best-of-N / resample).
- *   refine — attempt → observe() reads the trace → steer the next → repeat (iterate).
- * (A multi-agent "team" is just a Strategy whose driver spawns several different agents.)
- */
-interface Strategy {
-    readonly name: string;
-    driver(surface: AgenticSurface, task: AgenticTask, opts: AgenticOptions, budget: number): Agent<unknown, Outcome<unknown>>;
-}
-/** Built-in `Strategy`: K independent attempts, keep the best-verifying (best-of-N / resample). */
-declare const sample: Strategy;
-/** Built-in `Strategy`: attempt → `observe()` reads the trace → steer the next attempt → repeat (deepen one lineage). */
-declare const refine: Strategy;
-/** A role for one shot — multi-agent loops (researcher + engineer, a panel of k
- *  researchers) give each shot its own system prompt and optionally its own model. */
-interface ShotPersona {
-    /** Replaces the task's systemPrompt for a FRESH shot; on a carried conversation it is
-     *  injected as a hand-off message (the transcript's earlier roles stay intact). */
-    systemPrompt?: string;
-    /** Per-shot model override (e.g. a stronger model for the engineer shot). */
-    model?: string;
-}
-interface ShotSpec {
-    /** present ⇒ continue this artifact (depth); absent ⇒ the shot opens a fresh one (sample/restart). */
-    handle?: ArtifactHandle;
-    messages?: Msg[];
-    steer?: string;
-    persona?: ShotPersona;
-    /** Restrict THIS shot to a subset of the domain's tools (by name) — focus a shot on
-     *  the relevant capabilities. Restriction-only; unknown names throw. Omitted ⇒ all. */
-    tools?: string[];
-}
-interface StrategyResult {
-    score: number;
-    resolved: boolean;
-    completions: number;
-    progression: number[];
-    shots: number;
-}
-/** Artifact lifecycle a strategy may manage itself — open/close ONLY. Raw `call`/`score`
- *  are withheld: scores reach the body solely through `shot()`'s ShotResult (the
- *  harness-verified channel), so a body cannot peek the check or fabricate around it. */
-interface StrategyArtifacts {
-    readonly name: string;
-    open(task: AgenticTask): Promise<ArtifactHandle>;
-    close(handle: ArtifactHandle): Promise<void>;
-}
-/** What a strategy body composes with: the artifact lifecycle, the budget, and the two steps. */
-interface StrategyCtx {
-    /** Open/close artifacts the body manages itself (e.g. one persistent handle for depth). */
-    readonly surface: StrategyArtifacts;
-    readonly task: AgenticTask;
-    readonly opts: AgenticOptions;
-    readonly budget: number;
-    readonly scope: Scope<Outcome<unknown>>;
-    /** Run ONE worker shot; its harness-scored result, or null if it went down. */
-    shot(spec?: ShotSpec): Promise<ShotResult | null>;
-    /** The firewalled critic reads the trajectory → a steer string, or null on COMPLETE/down. */
-    critique(messages: Msg[]): Promise<string | null>;
-    /** The RAW analyst channel: the firewalled critic answers `instruction` over the
-     *  trajectory verbatim — no findings extraction, so verdict-shaped formats
-     *  (CONTINUE/STOP decisions, calibrated predictions) survive. Same firewall:
-     *  trajectory in, never scores. Null when the analyst went down. */
-    consult(messages: Msg[], instruction: string): Promise<string | null>;
-    /** The tools THIS artifact's task actually offers (names + descriptions only — never
-     *  the implementations). Tool sets vary per task on heterogeneous domains; a strategy
-     *  that restricts shots MUST select from this list, never from hardcoded names. */
-    listTools(handle: ArtifactHandle): Promise<Array<{
-        name: string;
-        description?: string;
-    }>>;
-}
-/** Author a Strategy from the composable steps — the open, compact way. */
-declare function defineStrategy(name: string, run: (ctx: StrategyCtx) => Promise<StrategyResult>): Strategy;
-/** A NEW strategy, authored from the steps (~20 lines): refine, but when a steered shot
- *  fails to improve the score it ABANDONS that line and restarts fresh (branch-when-stuck)
- *  — the widen/MCTS idea the depth-stuck failure motivated. Scored keep-best (the best
- *  checkpoint across all lines), the deployable metric. This is the "experts build BETTER
- *  optimizations" path: a new technique, compact, with zero Supervisor ceremony. */
-declare const adaptiveRefine: Strategy;
-/** The explore-then-exploit MIX: spend ⌈budget/2⌉ on independent samples (kept open),
- *  then refine the best-verifying line with the remaining budget. Sample's basin escape +
- *  refine's accumulation — the third built-in, authored from the public steps. */
-declare const sampleThenRefine: Strategy;
-interface RunAgenticOptions extends AgenticOptions {
-    surface: AgenticSurface;
-    task: AgenticTask;
-    /** Lifecycle observability — every spawn/settle (shots, analysts) streams here live.
-     *  The seam online watchdogs/route-auditors subscribe to. */
-    hooks?: RuntimeHooks;
-    /** A Strategy (the open way) — author/pass your own. Overrides `mode` when present. */
-    strategy?: Strategy;
-    /** Built-in shorthand: 'depth'→refine, 'breadth'→sample. Default 'depth'. */
-    mode?: 'depth' | 'breadth';
-    /** budget: refine→max shots; sample→rollout width. */
-    budget: number;
-    rootBudget?: Budget;
-}
-/** Run a Strategy through the keystone Supervisor — `Agent.act` over a conserved-budget Scope. */
-declare function runAgentic(opts: RunAgenticOptions): Promise<AgenticRunResult>;
-
 /**
  * runBenchmark — the packaged optimization suite. Define a domain by implementing an
  * `Environment` (open / tools / call / score / close); get the optimization strategies
@@ -2432,6 +2206,14 @@ interface SandboxRun<Out> {
  * @experimental
  */
 type OpenSandboxRunPromptOptions = Omit<PromptOptions, 'signal' | 'sessionId'>;
+/** Context available after the box/session exists and before the first prompt is
+ * drained. Intended for benchmark-owned workspace setup such as cloning a repo
+ * into a fixed path. */
+interface OpenSandboxRunBeforeStartContext {
+    readonly box: SandboxInstance;
+    readonly sessionId: string;
+    readonly signal: AbortSignal;
+}
 /** @experimental */
 interface OpenSandboxRunOptions {
     /** Profile + sandbox env/overrides. `sandboxOverrides.backend.type` is the harness. */
@@ -2446,6 +2228,10 @@ interface OpenSandboxRunOptions {
     /** Per-prompt sandbox SDK options forwarded to both `start()` and `resume()`.
      *  The runtime still owns the session id and abort signal for each turn. */
     promptOptions?: OpenSandboxRunPromptOptions;
+    /** Optional pre-start workspace setup. Runs after `lineage.start()` creates the
+     * box/session and before the first prompt stream is consumed. A thrown error
+     * fails the turn before the agent spends tokens. */
+    beforeStart?: (ctx: OpenSandboxRunBeforeStartContext) => Promise<void> | void;
     /** Test seam for deterministic hook timestamps. Defaults to `Date.now`. */
     now?: () => number;
     /** Bounds box-creation bursts inside lineage fanout. Default from lineage. */
@@ -3392,6 +3178,10 @@ declare function serveCoordinationMcp(opts: {
     /** Hard cap on simultaneously-LIVE workers — `spawn_agent` fails closed once this many are in
      *  flight (a concurrency fence on top of the conserved-pool fence). Omit/`<= 0` = no cap. */
     maxLiveWorkers?: number;
+    /** Max wall-clock ms a single `await_event` may block before returning a re-pollable
+     *  `{ pending, live }` snapshot instead of erroring on the client's request timeout. Omit =
+     *  {@link DEFAULT_AWAIT_EVENT_TIMEOUT_MS}; `<= 0` = prior unbounded block (in-process only). */
+    awaitTimeoutMs?: number;
     port?: number;
     host?: string;
     /** Trace-analyst lenses the driver can run (`run_analyst`) or auto-fire on settle. */
@@ -3403,66 +3193,6 @@ declare function serveCoordinationMcp(opts: {
     questionPolicy?: QuestionPolicy;
 }): Promise<CoordinationMcpHandle>;
 
-/** The supervisor's profile — the subset of an `AgentProfile` that selects + shapes its brain.
- *  `harness` is the backend-as-data discriminant; `systemPrompt` is the standing instruction. */
-interface SupervisorProfile {
-    readonly name?: string;
-    /** null/undefined → router brain (in-process tool-loop); a coding-CLI harness → sandboxed brain. */
-    readonly harness?: string | null;
-    /** The router model when the brain is router-driven (falls back to the deps router config). */
-    readonly model?: string;
-    /** The standing instructions ("you delegate, you do not solve"). */
-    readonly systemPrompt?: string;
-}
-/** How to run a sandboxed harness as the DRIVER, with the coordination verbs mounted — the substrate
- *  seam the caller supplies (mirrors `makeWorkerAgent` for spawned children). It runs `profile` on
- *  `task` in its backend (sandbox / cli-bridge) with `coordinationMcpUrl` mounted as an MCP server,
- *  so the harness calls spawn_agent / await_event / stop as native tools over the live scope. */
-type DriveHarness = (args: {
-    readonly profile: SupervisorProfile;
-    readonly task: unknown;
-    readonly scope: Scope<unknown>;
-    readonly coordinationMcpUrl: string;
-}) => Promise<void>;
-interface SupervisorAgentDeps {
-    readonly blobs: ResultBlobStore;
-    /** Resolve a spawned worker `profile` to a leaf agent — the recursion seam (same for both arms). */
-    readonly makeWorkerAgent: MakeWorkerAgent;
-    /** Per-child budget reserved from the conserved pool on each spawn. */
-    readonly perWorker: Budget;
-    /** Hard cap on simultaneously-LIVE workers across both arms — `spawn_agent` fails closed once
-     *  this many are in flight (a concurrency fence on top of the conserved-pool fence; bounds live
-     *  boxes/sandboxes, not total work). Omit/`<= 0` = no cap. */
-    readonly maxLiveWorkers?: number;
-    /** Router substrate for a router-brained supervisor (`harness` null). The profile's model wins. */
-    readonly router?: RouterConfig;
-    /** Inject the brain directly (tests / advanced) instead of resolving `routerBrain` from the profile. */
-    readonly brain?: ToolLoopChat;
-    /** Required for a sandboxed-harness supervisor (`harness` set): runs the harness as the driver. */
-    readonly driveHarness?: DriveHarness;
-    /** WORK tools the supervisor may call DIRECTLY (router arm) — so it can do simple work ITSELF and
-     *  only delegate when it needs parallelism. Pair with `executeExtraTool`. */
-    readonly extraTools?: ReadonlyArray<{
-        readonly name: string;
-        readonly description?: string;
-        readonly parameters: Record<string, unknown>;
-    }>;
-    /** Runs an `extraTools` call; null/undefined falls through to the coordination dispatch. */
-    readonly executeExtraTool?: (name: string, args: Record<string, unknown>) => Promise<string | null | undefined>;
-    /** Analyst lenses available to the driver (both arms). Required for `analyzeOnSettle`. */
-    readonly analysts?: AnalystRegistry;
-    /** Analyst kinds run on each worker-settle → a `finding` the driver composes its next steer from
-     *  (the self-improving UP-leg). Unset/empty = status quo (no analyst feed). Requires `analysts`. */
-    readonly analyzeOnSettle?: ReadonlyArray<string>;
-    readonly maxTurns?: number;
-    /** Give the supervisor brain a chapter-lifecycle on its OWN context window (router arm only) — it
-     *  distills its coordination transcript to a compact progress note once it exceeds the threshold,
-     *  instead of re-billing the whole thing every turn. See `DriverAgentOptions.compaction`. */
-    readonly compaction?: ToolLoopCompactionOptions;
-}
-/** Build a supervisor `Agent` from its profile: the brain resolves from `profile.harness` (backend-as-data), the same resolution rule as every worker. */
-declare function supervisorAgent(profile: SupervisorProfile, deps: SupervisorAgentDeps): Agent<unknown, unknown>;
-
 /**
  *
  * `delegate` — the one generic delegation verb. You hand it an INTENT (what you want done) and it
@@ -3803,71 +3533,6 @@ declare function createScope<Out>(args: ScopeArgs): Scope<Out>;
  */
 declare function settledToIteration<Out>(settled: Settled<Out>): Iteration<unknown, Out>;
 
-/** Build the worker seam from a backend (WHERE workers run) + an optional completion oracle (the
- *  deliverable check that makes "settled ⟺ delivered" true — the guard against "ran but didn't
- *  deliver"). The ONE place a backend becomes a spawnable worker. */
-declare function workerFromBackend(backend: ExecutorConfig, deliverable?: DeliverableSpec<unknown>): MakeWorkerAgent;
-interface SuperviseOptions {
-    /** The conserved compute pool for the whole run. */
-    readonly budget: Budget;
-    /** WHERE workers run — derives the worker seam. Provide this OR an explicit `makeWorkerAgent`. */
-    readonly backend?: ExecutorConfig;
-    /** The completion oracle for backend-derived workers (settled ⟺ delivered). Strongly recommended:
-     *  without it the supervisor trusts a worker's self-report — exactly the "ran but didn't deliver"
-     *  failure mode of a static orchestrator. */
-    readonly deliverable?: DeliverableSpec<unknown>;
-    /** Override the worker seam directly (tests / advanced) instead of deriving it from `backend`. */
-    readonly makeWorkerAgent?: MakeWorkerAgent;
-    /** The supervisor's router substrate (`harness` null). The profile's model wins. */
-    readonly router?: RouterConfig;
-    /** Inject the supervisor brain directly (tests / advanced). */
-    readonly brain?: ToolLoopChat;
-    /** Run a sandboxed-harness supervisor (`harness` set). */
-    readonly driveHarness?: DriveHarness;
-    /** WORK tools the supervisor may call DIRECTLY — so a recursive atom can ACT (do simple work
-     *  itself) OR SPAWN (delegate when it needs parallelism), not be a pure manager. Pair with
-     *  `executeExtraTool`. Router arm only (`harness` null). */
-    readonly extraTools?: ReadonlyArray<{
-        readonly name: string;
-        readonly description?: string;
-        readonly parameters: Record<string, unknown>;
-    }>;
-    /** Runs an `extraTools` call; null/undefined falls through to the coordination dispatch. */
-    readonly executeExtraTool?: (name: string, args: Record<string, unknown>) => Promise<string | null | undefined>;
-    /** Per-child budget reserved on each spawn. Defaults to a quarter of the pool's tokens. */
-    readonly perWorker?: Budget;
-    /** Hard cap on simultaneously-LIVE workers — `spawn_agent` fails closed once this many are in
-     *  flight. The conserved pool bounds TOTAL work; this bounds SIMULTANEOUS work (live boxes/
-     *  sandboxes a real fleet runs at once). Omit/`<= 0` = no cap (the pool stays the only fence). */
-    readonly maxLiveWorkers?: number;
-    /** Analyst lenses available to the driver. Required for `analyzeOnSettle`. Unset → status quo
-     *  (the driver receives settled worker outputs, no analyst findings). */
-    readonly analysts?: AnalystRegistry;
-    /** Analyst kind ids run AUTOMATICALLY when a worker settles `done` — each re-enters as a `finding`
-     *  the driver pulls (`await_event`) and composes its next steer from. The self-improving UP-leg,
-     *  threaded to the driver at this level (propagate to sub-drivers via a recursive `makeWorkerAgent`).
-     *  Omit/empty = status quo (no analyst feed). Requires `analysts`. */
-    readonly analyzeOnSettle?: ReadonlyArray<string>;
-    /** Worker output store. Defaults to in-memory. */
-    readonly blobs?: ResultBlobStore;
-    readonly maxDepth?: number;
-    readonly maxTurns?: number;
-    /** Give the supervisor brain a chapter-lifecycle on its OWN context window (router arm only): once
-     *  its coordination transcript exceeds `thresholdTokens` it distills to a compact progress note and
-     *  continues, instead of re-billing the whole transcript every turn (the cost that makes the LLM-brain
-     *  front door lose to a dumb-Ralph respawn). The live `Scope` roster is the durable state across
-     *  chapters. Default off. `distill` defaults to a brain self-summary + the settled-worker roster. */
-    readonly compaction?: ToolLoopCompactionOptions;
-    readonly runId?: string;
-    readonly now?: () => number;
-    /** Restrict the run to this subset of models. When set, every configured model — the
-     *  supervisor router model, the profile's model, and the backend's model — must be a member,
-     *  or `supervise()` throws a `ConfigError` before any compute is spent. Unset = unrestricted. */
-    readonly allowedModels?: readonly string[];
-}
-/** One-call supervisor: build + run a supervisor from its profile with sensible defaults; the raw `supervisorAgent` + `createSupervisor().run` seams stay available for power use. */
-declare function supervise(profile: SupervisorProfile, task: unknown, opts: SuperviseOptions): Promise<SupervisedResult<unknown>>;
-
 /**
  *
  * The `Supervisor` impl (KEYSTONE, build step 5).
@@ -4089,4 +3754,4 @@ declare function runInWorkspace<T>(ws: Workspace, body: (cwd: string) => Promise
     commitOnInvalid?: boolean;
 }): Promise<WorkspaceRun<T>>;
 
-export { Agent, AgentRunSpec, type AgentTurnBackend, type AgentTurnUsage, type AgenticOptions, type AgenticRunResult, type AgenticSurface, type AgenticTask, type AgenticTool, AnalystRegistry, type AnytimeReport, type AnytimeStrategySummary, type AnytimeTaskCurve, type ApplyContinuation, type ArtifactHandle, AssertTraceDerivedFindings, type AuditIntentInput, type AuditIntentOptions, type AuthorStrategyOptions, type AuthoredProfile, type AuthoredStrategy, type BenchmarkCell, type BenchmarkConfig, type BenchmarkLift, type BenchmarkReport, type BenchmarkStrategySummary, type BenchmarkTaskRow, Budget, type BudgetPool, type BudgetReadout, type ChampionPick, type ChampionPolicy, type CheckpointCapableBox, type CollectedAgentTurn, CombinatorShape, type CompletionAnalyst, type CompletionEvidence, type CompletionPolicy, type CompletionVerdict, CoordinationEvent, type CoordinationMcpHandle, Corpus, CorpusFilter, CorpusRecord, type CreateScopeAnalystOptions, type CriuCapableClient, DefinePersonaInput, type DefinedLeaderboard, type DelegateOptions, type Deliverable, DeliverableSpec, type DriveHarness, Driver, type DriverAgentOptions, type DumbDriverOptions, type Environment, EqualKArm, EqualKOnCostOptions, EqualKVerdict, type EvolutionArchiveNode, type EvolutionAuthor, type EvolutionBandInfo, type EvolutionCandidate, type EvolutionGeneration, type EvolutionReport, ExecCtx, ExecutorConfig, ExecutorFactory, ExecutorRegistry, FanoutOptions, FanoutWinnerSelector, FileCorpus, type ForkCapableBox, type GitWorkspaceOptions, type HarvestCorpusOptions, type HarvestFailure, type HarvestReport, InMemoryCorpus, InMemoryResultBlobStore, type InMemoryRunContext, type InMemoryRunContextOptions, InMemorySpawnJournal, type InProcessOnPrompt, type InProcessPromptCtx, type InProcessSandboxClientOptions, type Inbox, type InboxMessage, type IntentAudit, type Interval, Iteration, type Leaderboard, type LeaderboardBenchScore, type LeaderboardBenchTask, type LeaderboardBenchmarkAdapter, type LeaderboardFlagSpec, type LeaderboardIterationInfo, type LeaderboardOptions, type LeaderboardRow, type LeaderboardRunContext, type LeaderboardScenario, type LeaderboardScore, type LeaderboardSpec, type LoopCampaignDispatchOptions, type LoopDispatchOptions, LoopLineageOptions, type LoopOptionsForDispatch, LoopResult, LoopShape, LoopTokenUsage, LoopUntilSpec, LoopWinner, MakeWorkerAgent, type McpEndpoint, type McpEnvironmentOptions, MountRecorder, type NaiveDriverOptions, type Observation, type ObserveInput, type ObserveOptions, type OpenSandboxRunOptions, type OpenSandboxRunPromptOptions, Outcome, OutputAdapter, type PairwiseOptions, type PairwiseVerdict, PanelSpec, Persona, PipelineStage, type ProfileRichness, type ProfileRichnessThresholds, type PromotionGateOptions, type PromotionVerdict, type RegistryAnalyzeProjection, RenderCorpusToInstructionsOptions, type ReservationTicket, type ResolveSandboxClientOptions, ResultBlobStore, RouterConfig, type RunAgenticOptions, RunPersonifiedOptions, type SandboxCapabilities, SandboxClient, type SandboxLineage, type SandboxLineageHandle, type SandboxRun, SandboxRunAbortError, type SandboxToolPartState, Scope, ScopeAnalyst, ScopeAnalyzeInput, ScopeWidenGate, type SessionCapableBox, type SessionTraceBox, Settled, ShapeRegistry, type Shell, type ShotPersona, type ShotSpec, Spend, SteerContext, type SteeringDecision, type Strategy, type StrategyCtx, type StrategyEvolutionConfig, type StrategyResult, type StreamAgentTurnOptions, type SuperviseOptions, type SuperviseSurfaceOptions, type SuperviseSurfaceResult, SupervisedResult, Supervisor, type SupervisorAgentDeps, type SupervisorProfile, type SurfaceScore, type SurfaceWorkerConfig, type SurfaceWorkerOut, ToolLoopChat, ToolLoopCompactionOptions, type TraceSource, type TrajectoryAnalysis, TrajectoryReport, TrajectoryReportOptions, type TurnResult, UsageEvent, type UsageSink, Validator, type VerifierEnvironmentOptions, VerifySpec, type WatchTraceOptions, type WaterfallCollector, type WaterfallReport, type WaterfallSpan, WidenSpec, WinnerStrategy, type Workspace, type WorkspaceCommit, type WorkspaceRun, acquireSandbox, adaptiveRefine, analyzeTrace, anytimeReport, asAuthoredProfile, assertModelAllowed, assertStrategyContract, assertTraceDerivedFindings, assessAuthoredProfile, auditIntent, authorStrategy, authoredWorker, breadthStrategy, buildSteerContext, builtinShapes, collectAgentTurn, completionAuthorizes, contentAddress, createBudgetPool, createInMemoryRunContext, createInbox, createMcpEnvironment, createPushTraceSource, createSandboxLineage, createSandboxToolPartState, createScope, createScopeAnalyst, createShapeRegistry, createSupervisor, createVerifierEnvironment, createWaterfallCollector, decodeToolPart, defaultAnalystInstruction, defaultAuditorInstruction, defaultDelegateBudget, defaultProfileRichnessThresholds, defaultSelectWinner, defaultToolDetectors, defineLeaderboard, definePersona, defineStrategy, delegate, depthStrategy, deterministicCompletion, discriminatingMeans, driverAgent, dumbDriver, equalKOnCost, extractLlmCallEvent, failuresAnalyst, fanout, finalizeBestDelivered, flatWidenGate, gitWorkspace, harvestCorpus, inProcessSandboxClient, inlineSandboxClient, jjWorkspace, leaderboard, localShell, loopCampaignDispatch, loopDispatch, loopUntil, mapSandboxEvent, mapSandboxToolEvent, naiveDriver, observe, openSandboxRun, pairwiseSignificance, panel, pickChampion, pipeline, printBenchmarkReport, probeSandboxCapabilities, profileRichnessFinding, promotionGate, refine, registerShape, registryScopeAnalyst, renderAnytimeTable, renderCorpusToInstructions, renderLeaderboardHtml, renderLeaderboardMarkdown, renderLeaderboardSvg, renderPairwiseMarkdown, renderReport, reportLoopUsage, resolveSandboxClient, runAgentic, runBenchmark, runInWorkspace, runLoop, runPersonified, runStrategyEvolution, sample, sampleThenRefine, sandboxSessionTraceSource, selectChampion, selectValidWinner, sentinelCompletion, serveCoordinationMcp, settledToIteration, spendFromUsageEvents, stopSentinel, strategyAuthorContract, streamAgentTurn, sumSandboxUsage, supervise, superviseSurface, supervisorAgent, supervisorInstructions, trajectoryReport, verify, watchTrace, widen, workerFromBackend };
+export { Agent, AgentRunSpec, type AgentTurnBackend, type AgentTurnUsage, AgenticOptions, AgenticSurface, AgenticTask, AgenticTool, AnalystRegistry, type AnytimeReport, type AnytimeStrategySummary, type AnytimeTaskCurve, type ApplyContinuation, ArtifactHandle, AssertTraceDerivedFindings, type AuditIntentInput, type AuditIntentOptions, type AuthorStrategyOptions, type AuthoredProfile, type AuthoredStrategy, type BenchmarkCell, type BenchmarkConfig, type BenchmarkLift, type BenchmarkReport, type BenchmarkStrategySummary, type BenchmarkTaskRow, Budget, type BudgetPool, type BudgetReadout, type ChampionPick, type ChampionPolicy, type CheckpointCapableBox, type CollectedAgentTurn, CombinatorShape, type CompletionAnalyst, type CompletionEvidence, type CompletionPolicy, type CompletionVerdict, CoordinationEvent, type CoordinationMcpHandle, Corpus, CorpusFilter, CorpusRecord, type CreateScopeAnalystOptions, type CriuCapableClient, DefinePersonaInput, type DefinedLeaderboard, type DelegateOptions, type Deliverable, DeliverableSpec, Driver, type DriverAgentOptions, type DumbDriverOptions, type Environment, EqualKArm, EqualKOnCostOptions, EqualKVerdict, type EvolutionArchiveNode, type EvolutionAuthor, type EvolutionBandInfo, type EvolutionCandidate, type EvolutionGeneration, type EvolutionReport, ExecCtx, ExecutorConfig, ExecutorFactory, ExecutorRegistry, FanoutOptions, FanoutWinnerSelector, FileCorpus, type ForkCapableBox, type GitWorkspaceOptions, type HarvestCorpusOptions, type HarvestFailure, type HarvestReport, InMemoryCorpus, InMemoryResultBlobStore, type InMemoryRunContext, type InMemoryRunContextOptions, InMemorySpawnJournal, type InProcessOnPrompt, type InProcessPromptCtx, type InProcessSandboxClientOptions, type Inbox, type InboxMessage, type IntentAudit, type Interval, Iteration, type Leaderboard, type LeaderboardBenchScore, type LeaderboardBenchTask, type LeaderboardBenchmarkAdapter, type LeaderboardFlagSpec, type LeaderboardIterationInfo, type LeaderboardOptions, type LeaderboardRow, type LeaderboardRunContext, type LeaderboardScenario, type LeaderboardScore, type LeaderboardSpec, type LoopCampaignDispatchOptions, type LoopDispatchOptions, LoopLineageOptions, type LoopOptionsForDispatch, LoopResult, LoopShape, LoopTokenUsage, LoopUntilSpec, LoopWinner, MakeWorkerAgent, type McpEndpoint, type McpEnvironmentOptions, MountRecorder, type NaiveDriverOptions, type Observation, type ObserveInput, type ObserveOptions, type OpenSandboxRunBeforeStartContext, type OpenSandboxRunOptions, type OpenSandboxRunPromptOptions, Outcome, OutputAdapter, type PairwiseOptions, type PairwiseVerdict, PanelSpec, Persona, PipelineStage, type ProfileRichness, type ProfileRichnessThresholds, type PromotionGateOptions, type PromotionVerdict, type RegistryAnalyzeProjection, RenderCorpusToInstructionsOptions, type ReservationTicket, type ResolveSandboxClientOptions, ResultBlobStore, RouterConfig, RunPersonifiedOptions, type SandboxCapabilities, SandboxClient, type SandboxLineage, type SandboxLineageHandle, type SandboxRun, SandboxRunAbortError, type SandboxToolPartState, Scope, ScopeAnalyst, ScopeAnalyzeInput, ScopeWidenGate, type SessionCapableBox, type SessionTraceBox, Settled, ShapeRegistry, type Shell, Spend, SteerContext, type SteeringDecision, Strategy, type StrategyEvolutionConfig, type StreamAgentTurnOptions, type SuperviseSurfaceOptions, type SuperviseSurfaceResult, SupervisedResult, Supervisor, SupervisorProfile, SurfaceScore, type SurfaceWorkerConfig, type SurfaceWorkerOut, ToolLoopChat, ToolLoopCompactionOptions, type TraceSource, type TrajectoryAnalysis, TrajectoryReport, TrajectoryReportOptions, type TurnResult, UsageEvent, type UsageSink, Validator, type VerifierEnvironmentOptions, VerifySpec, type WatchTraceOptions, type WaterfallCollector, type WaterfallReport, type WaterfallSpan, WidenSpec, WinnerStrategy, type Workspace, type WorkspaceCommit, type WorkspaceRun, acquireSandbox, analyzeTrace, anytimeReport, asAuthoredProfile, assertModelAllowed, assertStrategyContract, assertTraceDerivedFindings, assessAuthoredProfile, auditIntent, authorStrategy, authoredWorker, buildSteerContext, builtinShapes, collectAgentTurn, completionAuthorizes, contentAddress, createBudgetPool, createInMemoryRunContext, createInbox, createMcpEnvironment, createPushTraceSource, createSandboxLineage, createSandboxToolPartState, createScope, createScopeAnalyst, createShapeRegistry, createSupervisor, createVerifierEnvironment, createWaterfallCollector, decodeToolPart, defaultAnalystInstruction, defaultAuditorInstruction, defaultDelegateBudget, defaultProfileRichnessThresholds, defaultSelectWinner, defaultToolDetectors, defineLeaderboard, definePersona, delegate, deterministicCompletion, discriminatingMeans, driverAgent, dumbDriver, equalKOnCost, extractLlmCallEvent, failuresAnalyst, fanout, finalizeBestDelivered, flatWidenGate, gitWorkspace, harvestCorpus, inProcessSandboxClient, inlineSandboxClient, jjWorkspace, leaderboard, localShell, loopCampaignDispatch, loopDispatch, loopUntil, mapSandboxEvent, mapSandboxToolEvent, naiveDriver, observe, openSandboxRun, pairwiseSignificance, panel, pickChampion, pipeline, printBenchmarkReport, probeSandboxCapabilities, profileRichnessFinding, promotionGate, registerShape, registryScopeAnalyst, renderAnytimeTable, renderCorpusToInstructions, renderLeaderboardHtml, renderLeaderboardMarkdown, renderLeaderboardSvg, renderPairwiseMarkdown, renderReport, reportLoopUsage, resolveSandboxClient, runBenchmark, runInWorkspace, runLoop, runPersonified, runStrategyEvolution, sandboxSessionTraceSource, selectChampion, selectValidWinner, sentinelCompletion, serveCoordinationMcp, settledToIteration, spendFromUsageEvents, stopSentinel, strategyAuthorContract, streamAgentTurn, sumSandboxUsage, superviseSurface, supervisorInstructions, trajectoryReport, verify, watchTrace, widen };