@tangle-network/agent-runtime 0.83.0 → 0.84.0

package/dist/index.js

@@ -1,7 +1,7 @@
 import {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
-} from "./chunk-J3N6GODQ.js";
+} from "./chunk-G7HQI6DB.js";
 import {
   DEFAULT_ROUTER_BASE_URL,
   cleanModelId,
@@ -20,16 +20,16 @@ import {
   runLoopRunnerCli,
   selfImproveLoopRunner,
   worktreeLoopRunner
-} from "./chunk-UJ3ONGHR.js";
+} from "./chunk-XBSWWF6A.js";
 import "./chunk-SGKPNBXE.js";
-import "./chunk-QDVLTRCP.js";
+import "./chunk-UJW6EVNY.js";
 import {
   assertModelAllowed,
   composeRuntimeHooks,
   defineRuntimeHooks,
   notifyRuntimeDecisionPoint,
   notifyRuntimeHookEvent
-} from "./chunk-SSA6OERV.js";
+} from "./chunk-XN5TIJGP.js";
 import {
   INTELLIGENCE_WIRE_VERSION,
   buildLoopOtelSpans,

package/dist/loop-runner-bin.js

@@ -2,10 +2,10 @@
 import {
   parseLoopRunnerArgv,
   runLoopRunnerCli
-} from "./chunk-UJ3ONGHR.js";
+} from "./chunk-XBSWWF6A.js";
 import "./chunk-SGKPNBXE.js";
-import "./chunk-QDVLTRCP.js";
-import "./chunk-SSA6OERV.js";
+import "./chunk-UJW6EVNY.js";
+import "./chunk-XN5TIJGP.js";
 import "./chunk-UD4BHQMI.js";
 import "./chunk-BZF3KQ6G.js";
 import "./chunk-YPA5MLVE.js";

package/dist/loops.d.ts

@@ -1,4 +1,4 @@
-import { ChatClient, RunRecord, AnalystFinding, AgentProfile, AnalystRunInputs, ToolSpan, StreamingDetector, DetectorSignal, buildTrajectory } from '@tangle-network/agent-eval';
+import { ChatClient, RunRecord, HarnessType, AgentProfile, AnalystFinding, AnalystRunInputs, ToolSpan, StreamingDetector, DetectorSignal, buildTrajectory } from '@tangle-network/agent-eval';
 export { AnalystFinding, DefaultVerdict, computeFindingId, makeFinding } from '@tangle-network/agent-eval';
 import { SandboxEvent, SandboxInstance, CreateSandboxOptions } from '@tangle-network/sandbox';
 export { AgentProfile, CreateSandboxOptions, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
@@ -8,10 +8,10 @@ import { ak as MakeWorkerAgent, A as AnalystRegistry, o as CoordinationTools, n
 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 } 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 { 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 { Scenario, DispatchFn, ProfileDispatchFn } from '@tangle-network/agent-eval/campaign';
 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';
@@ -494,6 +494,158 @@ declare function deterministicCompletion<Task, Output>(check: (output: Output, h
     reasons?: string;
 }): CompletionAnalyst<Task, Output>;
 
+/** Structured per-case verdict a `score` function may return (a bare number is
+ *  shorthand for `{ composite }`). `composite` is the [0,1] leaderboard score;
+ *  `dimensions` are recorded as extra judge dimensions. */
+interface LeaderboardScore {
+    composite: number;
+    dimensions?: Record<string, number>;
+    notes?: string;
+}
+/** The campaign scenario a case is wrapped into: the case rides along so
+ *  judges and hooks can reach the full domain payload, not just its id. */
+interface LeaderboardScenario<TCase> extends Scenario {
+    case: TCase;
+}
+/** One extra CLI flag a spec declares. Parsed by `run()` as `--<name> <value>`
+ *  and surfaced to every hook via `ctx.args`. */
+interface LeaderboardFlagSpec {
+    default?: string;
+    description: string;
+}
+/** Resolved run configuration handed to `setup` / `teardown` / `export`. */
+interface LeaderboardRunContext {
+    name: string;
+    /** Execution backend name (`--backend`), a key of `backends`. */
+    backend: string;
+    runDir: string;
+    exportDir: string;
+    /** Every parsed flag (standard + `spec.flags`), by name without `--`. */
+    args: Record<string, string | undefined>;
+    harnesses: readonly HarnessType[];
+    /** Snapshot-stamped model ids (`name@snapshot`) — the eval identity models. */
+    models: readonly string[];
+    caseIds: readonly string[];
+    shots: number;
+    reps: number;
+}
+/** Structurally `BenchTask` (bench registry shape) — declared locally so this
+ *  module adds no dependency on a benchmark package. */
+interface LeaderboardBenchTask {
+    id: string;
+    prompt: string;
+    split?: string;
+    metadata?: Record<string, unknown>;
+}
+/** Structurally `BenchScore` (bench registry shape). */
+interface LeaderboardBenchScore {
+    resolved: boolean;
+    score: number;
+    detail?: string;
+}
+/** Structurally `BenchmarkAdapter` (bench registry shape): `name`,
+ *  `preflight()`, `loadTasks()`, deterministic `judge()`, `goldArtifact()`. */
+interface LeaderboardBenchmarkAdapter {
+    readonly name: string;
+    preflight(): Promise<void>;
+    loadTasks(opts?: {
+        limit?: number;
+        split?: string;
+        ids?: string[];
+    }): Promise<LeaderboardBenchTask[]>;
+    judge(task: LeaderboardBenchTask, artifact: string): Promise<LeaderboardBenchScore>;
+    goldArtifact(task: LeaderboardBenchTask): Promise<string | undefined>;
+}
+interface LeaderboardSpec<TCase> {
+    /** Leaderboard name — the scenario `kind`, default profile name, and report title. */
+    name: string;
+    /** The case corpus. Every case needs a stable string id (see `caseId`). */
+    cases: TCase[];
+    /** Stable id extractor. Default: the case's own `id` property (fail-loud
+     *  when absent or not a string). */
+    caseId?: (c: TCase) => string;
+    /** The per-case task prompt. May be async (e.g. built by shelling out to a
+     *  reference implementation); resolved ONCE per case before dispatch. */
+    prompt: (c: TCase) => string | Promise<string>;
+    /** The domain grader: agent output text → score. Used BOTH as the per-shot
+     *  validator (a shot with `composite > 0` stops the naive retry loop) and,
+     *  wrapped as a campaign judge, as the recorded leaderboard score. */
+    score: (output: string, c: TCase) => number | LeaderboardScore;
+    /** Harness × model axes for `expandProfileAxes`. Defaults: the canonical
+     *  `CODING_HARNESSES` × the base profile's `model.default`. `--harnesses` /
+     *  `--models` override per run. */
+    axis?: {
+        harnesses?: readonly HarnessType[];
+        models?: readonly string[];
+    };
+    /** Base profile the axes expand over (prompt/tools/skills held fixed).
+     *  Default: a minimal `{ name, model: { default: <first model> } }`. */
+    baseProfile?: AgentProfile;
+    /**
+     * Execution-backend registry: `--backend <name>` picks the factory that
+     * yields the `SandboxClient` every cell runs on. Merged over the defaults:
+     *   - `sandbox` — throws with guidance (a product must supply its real
+     *     Sandbox-backed client; the facade has no credentials).
+     *   - `cli-bridge` — `resolveSandboxClient({ backend: 'bridge' })` reading
+     *     `CLI_BRIDGE_URL` + `BRIDGE_BEARER`/`CLI_BRIDGE_BEARER`; the per-cell
+     *     harness/model ride in via `sandboxOverrides.backend`.
+     */
+    backends?: Record<string, (() => SandboxClient) | undefined>;
+    /** Extra `--flag value` CLI args `run()` parses and surfaces via `ctx.args`. */
+    flags?: Record<string, LeaderboardFlagSpec>;
+    /** Extra fields merged into each cell's `backend.model` create override —
+     *  e.g. `{ provider: 'openai-compat', apiKey, baseUrl }` for a router-backed
+     *  sandbox. The cell's bare model id is set by the facade from the axis. */
+    modelBackend?: Record<string, unknown>;
+    /** Runs once before the matrix (fetch fixtures, warm caches). */
+    setup?: (ctx: LeaderboardRunContext) => Promise<void> | void;
+    /** Runs once after the matrix, even on failure (reap boxes, close handles). */
+    teardown?: (ctx: LeaderboardRunContext) => Promise<void> | void;
+    /** Per-cell event tap: the raw sandbox events of each parsed iteration,
+     *  with the case — the seam for domain metric capture (search counts,
+     *  citations) without a substrate change. */
+    onCellEvents?: (events: readonly SandboxEvent[], c: TCase) => void;
+    /** Output decode override: raw events → the scored output text. Default:
+     *  the sandbox SDK's `collectAgentResponseText` (final answer text; empty
+     *  string when the stream carried none — which then scores 0). */
+    parseOutput?: (events: readonly SandboxEvent[], c: TCase) => string;
+    /** Result export. Default: write `matrix-result.json` under the run dir and
+     *  print (+ write) the ranked leaderboard markdown under the export dir. */
+    export?: (result: RunProfileMatrixResult<string, LeaderboardScenario<TCase>>, ctx: LeaderboardRunContext) => Promise<void> | void;
+    /** LEVEL 2 — full dispatch replacement (in-process products bring their own).
+     *  The default is `loopDispatch` + `naiveDriver` over the resolved backend. */
+    dispatch?: ProfileDispatchFn<LeaderboardScenario<TCase>, string>;
+    /** LEVEL 2 — full judge replacement. Default: `score` wrapped as one judge. */
+    judges?: JudgeConfig<string, LeaderboardScenario<TCase>>[];
+    /** Naive-retry shot cap per cell (`--shots`). Default 1. */
+    shots?: number;
+    /** Replicates per cell (`--reps`). Default 1. */
+    reps?: number;
+    /** Passthrough overrides spread onto the final `runProfileMatrix` call
+     *  (e.g. `maxConcurrency`, `costCeiling`, `integrity`, `storage`) — spread
+     *  LAST, so anything the facade wired can be overridden. */
+    matrix?: Partial<RunProfileMatrixOptions<LeaderboardScenario<TCase>, string>>;
+}
+interface DefinedLeaderboard<TCase> {
+    /**
+     * Parse flags, run the matrix, export, and return the raw result.
+     *
+     * Standard flags: `--backend <name>` (default `sandbox`), `--harnesses a,b`,
+     * `--models m1,m2`, `--cases id1,id2`, `--shots N`, `--reps N`,
+     * `--model-snapshot <tag>`, `--run-dir <path>`, `--export-dir <path>`,
+     * plus every `spec.flags` entry. `argv` defaults to `process.argv.slice(2)`.
+     *
+     * The default run dir is FRESH per invocation (timestamp+pid under the OS
+     * tmpdir). `runProfileMatrix` caches cells by run dir, and a stable default
+     * would silently reuse a prior FAILED zero-token cell and skip dispatch —
+     * only an explicit `--run-dir` opts into that resume behavior.
+     */
+    run(argv?: string[]): Promise<RunProfileMatrixResult<string, LeaderboardScenario<TCase>>>;
+    /** The same domain surface in the structural `BenchmarkAdapter` shape. */
+    toBenchmarkAdapter(): LeaderboardBenchmarkAdapter;
+}
+declare function defineLeaderboard<TCase>(spec: LeaderboardSpec<TCase>): DefinedLeaderboard<TCase>;
+
 /**
  * The third-person observer — the connective tissue that closes the loop.
  *
@@ -3626,4 +3778,4 @@ declare function runInWorkspace<T>(ws: Workspace, body: (cwd: string) => Promise
     commitOnInvalid?: boolean;
 }): Promise<WorkspaceRun<T>>;
 
-export { Agent, AgentRunSpec, 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, CombinatorShape, type CompletionAnalyst, type CompletionEvidence, type CompletionPolicy, type CompletionVerdict, CoordinationEvent, type CoordinationMcpHandle, Corpus, CorpusFilter, CorpusRecord, type CreateScopeAnalystOptions, type CriuCapableClient, DefinePersonaInput, 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 LeaderboardOptions, type LeaderboardRow, 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, 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, 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 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, completionAuthorizes, contentAddress, createBudgetPool, createInMemoryRunContext, createInbox, createMcpEnvironment, createPushTraceSource, createSandboxLineage, createScope, createScopeAnalyst, createShapeRegistry, createSupervisor, createVerifierEnvironment, createWaterfallCollector, decodeToolPart, defaultAnalystInstruction, defaultAuditorInstruction, defaultDelegateBudget, defaultProfileRichnessThresholds, defaultSelectWinner, defaultToolDetectors, 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, 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, sumSandboxUsage, supervise, superviseSurface, supervisorAgent, supervisorInstructions, trajectoryReport, verify, watchTrace, widen, workerFromBackend };
+export { Agent, AgentRunSpec, 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, 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 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, 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, 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 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, completionAuthorizes, contentAddress, createBudgetPool, createInMemoryRunContext, createInbox, createMcpEnvironment, createPushTraceSource, createSandboxLineage, 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, 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, sumSandboxUsage, supervise, superviseSurface, supervisorAgent, supervisorInstructions, trajectoryReport, verify, watchTrace, widen, workerFromBackend };

package/dist/loops.js

@@ -24,6 +24,7 @@ import {
   defaultAnalystInstruction,
   defaultAuditorInstruction,
   defaultToolDetectors,
+  defineLeaderboard,
   definePersona,
   defineStrategy,
   depthStrategy,
@@ -86,7 +87,7 @@ import {
   watchTrace,
   widen,
   worktreeFanout
-} from "./chunk-QDVLTRCP.js";
+} from "./chunk-UJW6EVNY.js";
 import {
   InMemoryResultBlobStore,
   InMemorySpawnJournal,
@@ -128,7 +129,7 @@ import {
   supervisorAgent,
   supervisorInstructions,
   workerFromBackend
-} from "./chunk-SSA6OERV.js";
+} from "./chunk-XN5TIJGP.js";
 import "./chunk-UD4BHQMI.js";
 import {
   createAgentEnvironmentProviderRegistry,
@@ -195,6 +196,7 @@ export {
   defaultProfileRichnessThresholds,
   defaultSelectWinner,
   defaultToolDetectors,
+  defineLeaderboard,
   definePersona,
   defineStrategy,
   delegate,

package/dist/mcp/bin.js

@@ -9,7 +9,7 @@ import {
   DelegationTaskQueue,
   FileDelegationStore,
   createMcpServer
-} from "../chunk-SSA6OERV.js";
+} from "../chunk-XN5TIJGP.js";
 import "../chunk-UD4BHQMI.js";
 import "../chunk-BZF3KQ6G.js";
 import "../chunk-4J6RBI3K.js";

package/dist/mcp/index.js

@@ -6,7 +6,7 @@ import {
 import {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
-} from "../chunk-J3N6GODQ.js";
+} from "../chunk-G7HQI6DB.js";
 import {
   createKbGate
 } from "../chunk-SGKPNBXE.js";
@@ -14,7 +14,7 @@ import {
   assertTraceDerivedFindings,
   runCoderChecks,
   selectValidWinner
-} from "../chunk-QDVLTRCP.js";
+} from "../chunk-UJW6EVNY.js";
 import {
   DELEGATE_DESCRIPTION,
   DELEGATE_FEEDBACK_DESCRIPTION,
@@ -64,7 +64,7 @@ import {
   validateDelegateUiAuditArgs,
   validateDelegationHistoryArgs,
   validateDelegationStatusArgs
-} from "../chunk-SSA6OERV.js";
+} from "../chunk-XN5TIJGP.js";
 import "../chunk-UD4BHQMI.js";
 import {
   deleteBoxSafe,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.83.0",
+  "version": "0.84.0",
   "description": "Shared task-lifecycle skeleton for agents: a recursive loop kernel for chat turns, one-shot tasks, and multi-attempt loops, with trace capture and eval-gated self-improvement. Domain behavior lives in adapters; scoring and ship-gates in @tangle-network/agent-eval.",
   "homepage": "https://github.com/tangle-network/agent-runtime#readme",
   "repository": {
@@ -94,7 +94,7 @@
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
-    "@tangle-network/agent-eval": ">=0.100.0 <1.0.0",
+    "@tangle-network/agent-eval": "^0.103.1",
     "@tangle-network/agent-interface": ">=0.14.0 <1.0.0",
     "@tangle-network/sandbox": ">=0.8.0 <1.0.0",
     "@types/node": "^25.9.3",
@@ -123,7 +123,7 @@
   "license": "MIT",
   "packageManager": "pnpm@10.28.0",
   "peerDependencies": {
-    "@tangle-network/agent-eval": ">=0.97.0 <1.0.0",
+    "@tangle-network/agent-eval": ">=0.101.0 <1.0.0",
     "@tangle-network/agent-interface": ">=0.14.0 <1.0.0",
     "@tangle-network/sandbox": ">=0.8.0 <1.0.0",
     "playwright": "^1.40.0"