@tangle-network/agent-runtime 0.75.1 → 0.77.0

package/dist/intelligence.d.ts

@@ -1,6 +1,6 @@
 import { c as LoopTraceEvent } from './types-B-jWSfcu.js';
 import { AgentProfileMcpServer } from '@tangle-network/agent-interface';
-import { T as ToolSpec } from './router-client-C7kp_ECN.js';
+import { T as ToolSpec } from './router-client-CMAWGv1h.js';
 import '@tangle-network/agent-eval';
 import '@tangle-network/sandbox';
 

package/dist/loop-runner-bin.js

@@ -2,10 +2,10 @@
 import {
   parseLoopRunnerArgv,
   runLoopRunnerCli
-} from "./chunk-QKNBYHMK.js";
+} from "./chunk-KKHMDE5I.js";
 import "./chunk-FNMGYYSS.js";
-import "./chunk-SA5GCF2X.js";
-import "./chunk-ZKMOIEOB.js";
+import "./chunk-CEW5BMGN.js";
+import "./chunk-RQYPH2QE.js";
 import "./chunk-VMNEQHJR.js";
 import "./chunk-IODKUOBA.js";
 import "./chunk-T2HVQVB4.js";

package/dist/loops.d.ts

@@ -1,11 +1,11 @@
 import { ChatClient, AnalystFinding, AgentProfile, AnalystRunInputs, ToolSpan, StreamingDetector, DetectorSignal, buildTrajectory } from '@tangle-network/agent-eval';
 export { AnalystFinding, DefaultVerdict, computeFindingId, makeFinding } from '@tangle-network/agent-eval';
-import { SandboxInstance, CreateSandboxOptions, SandboxEvent } from '@tangle-network/sandbox';
+import { SandboxEvent, SandboxInstance, CreateSandboxOptions } from '@tangle-network/sandbox';
 export { AgentProfile, CreateSandboxOptions, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
 import { e as ResultBlobStore, S as SpawnJournal, N as NodeId, n as SpawnEvent, m as ExecutorFactory, d as Agent, B as Budget, j as Scope, h as Settled, g as SupervisedResult, l as DeliverableSpec, i as Spend, U as UsageEvent, E as ExecutorRegistry, o as Supervisor } from './worktree-CDxqwxGo.js';
 export { A as AgentSpec, k as Executor, p as ExecutorContext, q as ExecutorResult, s as SupervisorOpts, T as TreeView, t as WidenGate, u as gateOnDeliverable } from './worktree-CDxqwxGo.js';
-import { ak as MakeWorkerAgent, o as CoordinationTools, A as AnalystRegistry, n as CoordinationEvent, ar as QuestionPolicy, a$ as ExecutorConfig } from './coordination-c_7Olmtq.js';
-export { b0 as BusEvent, b1 as BusRecord, b2 as BusStats, b3 as EventBus, b4 as PublishOptions, b5 as cliWorktreeExecutor, b6 as createEventBus, b7 as createExecutor, b8 as createExecutorRegistry } from './coordination-c_7Olmtq.js';
+import { ak as MakeWorkerAgent, o as CoordinationTools, A as AnalystRegistry, n as CoordinationEvent, ar as QuestionPolicy, a$ as ExecutorConfig } from './coordination-CZe4lTxy.js';
+export { b0 as BusEvent, b1 as BusRecord, b2 as BusStats, b3 as EventBus, b4 as PublishOptions, b5 as cliWorktreeExecutor, b6 as createEventBus, b7 as createExecutor, b8 as createExecutorRegistry } from './coordination-CZe4lTxy.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-B-jWSfcu.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-B-jWSfcu.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-CljF1L2v.js';
@@ -13,8 +13,8 @@ export { A as AuthoredHarness, t as DefinePersona, u as EqualKOnCost, v as Fanou
 import { Scenario, 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 } from './router-client-C7kp_ECN.js';
-export { b as RouterChatResult, c as RouterChatToolsResult, d as RouterToolCall, e as RouterToolLoopResult, T as ToolSpec, r as routerBrain, f as routerChatWithTools, g as routerChatWithUsage, h as routerToolLoop } from './router-client-C7kp_ECN.js';
+import { R as RouterConfig, a as ToolLoopChat } from './router-client-CMAWGv1h.js';
+export { b as RouterChatResult, c as RouterChatToolsResult, d as RouterToolCall, e as RouterToolLoopResult, T as ToolSpec, r as routerBrain, f as routerChatWithTools, g as routerChatWithUsage, h as routerToolLoop } from './router-client-CMAWGv1h.js';
 import { stuckLoopView, toolWasteView } from '@tangle-network/agent-eval/pipelines';
 import './substrate-rNj6TDc3.js';
 import './otel-export-BKmNwiCb.js';
@@ -470,6 +470,55 @@ interface HarvestReport {
 }
 declare function harvestCorpus(opts: HarvestCorpusOptions): Promise<HarvestReport>;
 
+/** Context handed to each `onPrompt` call. */
+interface InProcessPromptCtx {
+    /** 0-based round index — increments per `streamPrompt` on the SAME box (so a
+     *  refine driver's round N can differ from round N-1). Fresh boxes start at 0. */
+    round: number;
+    /** Absolute path of this box's workspace, when a `workdir` was configured.
+     *  Write the deliverable / fixtures here; `fs.read`/`fs.write`/`exec` operate
+     *  over it. `undefined` for pure event-only boxes. */
+    workdir?: string;
+    /** Cooperative cancellation channel for this turn. */
+    signal: AbortSignal;
+}
+/**
+ * The user callback: given a prompt and its round, produce the box's event
+ * stream for that turn. Return a plain `SandboxEvent[]` (the common case) or an
+ * async iterable for streaming. The callback may also write files into
+ * `ctx.workdir` (read back via `fs.read` or graded by `exec`).
+ */
+type InProcessOnPrompt = (prompt: string, ctx: InProcessPromptCtx) => SandboxEvent[] | AsyncIterable<SandboxEvent> | Promise<SandboxEvent[]>;
+/** @experimental */
+interface InProcessSandboxClientOptions {
+    /** The per-turn behavior — see {@link InProcessOnPrompt}. */
+    onPrompt: InProcessOnPrompt;
+    /**
+     * Opt in to a REAL filesystem-backed box. When set, each `create()` mints a
+     * fresh temp directory (prefixed `<workdir>-`) and the box exposes
+     * `fs.read`/`fs.write` and `exec` over it; `delete()` removes the dir. Omit
+     * for a pure event-only box (no `fs`/`exec` members), which is all a driver
+     * or fanout loop needs.
+     */
+    workdir?: string;
+    /**
+     * Override the box `id`. A string is used verbatim; a function receives the
+     * 0-based create-sequence and returns the id (e.g. machine-keyed placement
+     * demos). Default `in-process-<seq>`. The id is the value `describePlacement`
+     * tags, so set it when a demo's output reads on a meaningful sandbox id.
+     */
+    id?: string | ((seq: number) => string);
+}
+/**
+ * Adapt a single `onPrompt(prompt, ctx)` callback into a `SandboxClient` for
+ * `runLoop` / `openSandboxRun`. Returns a PROPERLY-TYPED `SandboxClient`: the
+ * lone `SandboxInstance` cast (object literal → `declare class`) lives inside
+ * this function, so call sites stay cast-free.
+ *
+ * @experimental
+ */
+declare function inProcessSandboxClient(options: InProcessSandboxClientOptions): SandboxClient;
+
 /**
  * Adapt an `ExecutorFactory` into a `SandboxClient` for `runLoop`. The factory is
  * instantiated fresh per `streamPrompt` (mirrors the per-spawn executor lifecycle):
@@ -696,6 +745,12 @@ 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. */
@@ -1877,6 +1932,128 @@ interface OpenSandboxRunOptions {
  */
 declare function openSandboxRun<Out>(client: SandboxClient, options: OpenSandboxRunOptions, deliverable: Deliverable<Out>): Promise<SandboxRun<Out>>;
 
+/**
+ * @experimental
+ *
+ * Leak-free steering drivers — the non-LLM controls for the driven loop.
+ *
+ * These are the two sibling `Driver`s of the `refine` reference driver
+ * (examples/driver-loop/driver-loop.ts), differing ONLY in how much of the
+ * prior `verdict` their `plan()` is allowed to read. The amount of the verdict
+ * a driver reads IS the experimental axis:
+ *
+ *   • refine  — reads `verdict.notes` (and/or `verdict.scores`): the grader's
+ *               findings. A reviewer/judge LLM turns those findings into the
+ *               next prompt. This is the "coached" condition.
+ *   • dumb    — reads ONLY `verdict.valid` (the pass/fail boolean). No findings,
+ *               no scores, no LLM. The next prompt is one of two fixed strings
+ *               keyed on pass/fail. This is the leak-free pass/fail control.
+ *   • naive   — reads NOTHING from the verdict. The next prompt is one fixed
+ *               continuation string every round. This is the leak-free
+ *               no-signal control.
+ *
+ * The `dumb → refine` gap measured on the same task therefore isolates how much
+ * the grader's findings (its notes/scores) inflate the result beyond a bare
+ * pass/fail signal: it is exactly the delta between reading `verdict.valid` and
+ * reading `verdict.notes`. The `naive → dumb` gap isolates the value of the
+ * pass/fail bit alone. Run all three against one task to attribute a loop's
+ * lift to its actual source instead of crediting it to the grader for free.
+ *
+ * GENERALITY CONTRACT
+ * -------------------
+ * These builders carry ZERO domain coupling:
+ *   • The continuation strings are PARAMETERS. The caller passes its own text;
+ *     the substrate hardcodes none.
+ *   • The Task shape is OPAQUE. The builder never assumes a `prompt` field (or
+ *     any field). The caller supplies `applyContinuation(task, text) → Task`,
+ *     which folds a steering string into the caller's own Task shape — exactly
+ *     symmetric with the caller-supplied `AgentRunSpec.taskToPrompt`. For a Task
+ *     that is `{ prompt: string; … }`, the fold is the one-liner
+ *     `(task, text) => ({ ...task, prompt: text })`.
+ *
+ * They consume nothing the loop kernel does not already give `plan()`/`decide()`
+ * — `runLoop` requires no change. A benchmark picks a driver by name and the
+ * kernel does the rest.
+ */
+
+/**
+ * Terminal-or-continue decision shared by all three steering drivers. The
+ * non-terminal `'refine'` keeps the loop running another shot; the terminal
+ * `'pick-winner'`/`'fail'` stop it (`isTerminalDecision` in run-loop.ts treats
+ * `'pick-winner'` and `'fail'` as terminal and any other string as a request
+ * for another round). Identical to the reference refine driver's decision set.
+ */
+type SteeringDecision = 'refine' | 'pick-winner' | 'fail';
+/**
+ * Fold a steering string into the caller's Task shape, producing the Task for
+ * the next shot. The substrate never assumes how a Task carries its prompt, so
+ * the caller supplies this — the same way it supplies `taskToPrompt`. The
+ * original `task` is passed so the fold can preserve task-level fields (ids,
+ * fixtures, feature names) and replace only the instruction.
+ */
+type ApplyContinuation<Task> = (task: Task, continuation: string) => Task;
+/** Options for {@link naiveDriver}. */
+interface NaiveDriverOptions<Task> {
+    /**
+     * The fixed continuation issued every round after shot 0. The same string is
+     * sent whether the prior shot passed inspection or not — the naive driver
+     * reads no part of the verdict. Domain text is the caller's; the substrate
+     * supplies none.
+     */
+    continuation: string;
+    /** Folds `continuation` into the caller's Task shape for the next shot. */
+    applyContinuation: ApplyContinuation<Task>;
+    /** Hard shot cap. The loop stops refining once history reaches this length. */
+    maxIterations: number;
+    /** Trace-event identifier. Default `'naive'`. */
+    name?: string;
+}
+/**
+ * `naiveDriver` — the no-signal steering control.
+ *
+ * `plan()` runs the initial `task` at shot 0, then issues the SAME fixed
+ * `continuation` every subsequent round until a shot is valid or the cap is
+ * hit. It reads NOTHING from `history[last].verdict` — not `.valid`, not
+ * `.notes`, not `.scores`. It is the floor a coached loop must beat to earn its
+ * coaching: any lift over naive that is not also present in `dumb` is
+ * attributable to the pass/fail bit, and any lift of `refine` over `dumb` is
+ * attributable to the grader's findings.
+ */
+declare function naiveDriver<Task, Output>(options: NaiveDriverOptions<Task>): Driver<Task, Output, SteeringDecision>;
+/** Options for {@link dumbDriver}. */
+interface DumbDriverOptions<Task> {
+    /**
+     * Continuation issued when the prior shot's verdict is valid. In a
+     * stop-on-pass loop this is rarely reached (a valid shot ends the loop), but
+     * it is required so the driver is total over the pass/fail bit; pass a
+     * confirmation/keep-going string.
+     */
+    onPass: string;
+    /** Continuation issued when the prior shot's verdict is NOT valid. */
+    onFail: string;
+    /** Folds the chosen continuation into the caller's Task shape. */
+    applyContinuation: ApplyContinuation<Task>;
+    /** Hard shot cap. The loop stops refining once history reaches this length. */
+    maxIterations: number;
+    /** Trace-event identifier. Default `'dumb'`. */
+    name?: string;
+}
+/**
+ * `dumbDriver` — the pass/fail-only steering control.
+ *
+ * `plan()` runs the initial `task` at shot 0, then reads ONLY
+ * `history[last].verdict.valid` (the boolean) and issues `onPass` or `onFail`
+ * accordingly. It MUST NOT read `.notes` or `.scores` — that boundary is the
+ * leak-free firewall. A `verdict` with no `valid` set (or no verdict) is
+ * treated as not-valid, so the driver is total and never throws on a
+ * grader/transport gap.
+ *
+ * The `dumb → refine` gap is the headline measurement: refine reads the
+ * grader's `notes`, dumb reads only the pass/fail bit, so the difference is
+ * exactly the value the findings add over a bare boolean.
+ */
+declare function dumbDriver<Task, Output>(options: DumbDriverOptions<Task>): Driver<Task, Output, SteeringDecision>;
+
 /**
  * authorStrategy — the agent-authored layer as a package primitive (software-3.0): an
  * LLM reads a benchmark's per-task LOSSES + the defineStrategy contract and writes a NEW
@@ -3065,4 +3242,4 @@ 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, type AnytimeReport, type AnytimeStrategySummary, type AnytimeTaskCurve, 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 Environment, EqualKArm, EqualKOnCostOptions, EqualKVerdict, type EvolutionArchiveNode, type EvolutionAuthor, type EvolutionBandInfo, type EvolutionCandidate, type EvolutionGeneration, type EvolutionReport, ExecCtx, ExecutorConfig, FanoutOptions, FanoutWinnerSelector, FileCorpus, type ForkCapableBox, type GitWorkspaceOptions, type HarvestCorpusOptions, type HarvestFailure, type HarvestReport, InMemoryCorpus, InMemoryResultBlobStore, type InMemoryRunContext, type InMemoryRunContextOptions, InMemorySpawnJournal, type Inbox, type InboxMessage, type IntentAudit, Iteration, type LoopDispatchOptions, LoopLineageOptions, type LoopOptionsForDispatch, LoopResult, LoopShape, LoopTokenUsage, LoopUntilSpec, LoopWinner, type McpEndpoint, type McpEnvironmentOptions, MountRecorder, type Observation, type ObserveInput, type ObserveOptions, type OpenSandboxRunOptions, Outcome, OutputAdapter, PanelSpec, Persona, PipelineStage, type ProfileRichness, type ProfileRichnessThresholds, type PromotionGateOptions, type PromotionVerdict, type RegistryAnalyzeProjection, RenderCorpusToInstructionsOptions, type ReservationTicket, 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 Strategy, type StrategyCtx, type StrategyEvolutionConfig, type StrategyResult, type SuperviseOptions, SupervisedResult, Supervisor, type SupervisorAgentDeps, type SupervisorProfile, type SurfaceScore, ToolLoopChat, 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, equalKOnCost, extractLlmCallEvent, fanout, finalizeBestDelivered, flatWidenGate, gitWorkspace, harvestCorpus, inlineSandboxClient, jjWorkspace, localShell, loopDispatch, loopUntil, mapSandboxEvent, observe, openSandboxRun, panel, pickChampion, pipeline, printBenchmarkReport, probeSandboxCapabilities, profileRichnessFinding, promotionGate, refine, registerShape, registryScopeAnalyst, renderAnytimeTable, renderCorpusToInstructions, renderReport, reportLoopUsage, runAgentic, runBenchmark, runInWorkspace, runLoop, runPersonified, runStrategyEvolution, sample, sampleThenRefine, sandboxSessionTraceSource, selectChampion, selectValidWinner, sentinelCompletion, serveCoordinationMcp, settledToIteration, spendFromUsageEvents, stopSentinel, strategyAuthorContract, supervise, supervisorAgent, supervisorInstructions, trajectoryReport, verify, watchTrace, widen, workerFromBackend };
+export { Agent, AgentRunSpec, type AgenticOptions, type AgenticRunResult, type AgenticSurface, type AgenticTask, type AgenticTool, 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, 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, Iteration, type LoopDispatchOptions, LoopLineageOptions, type LoopOptionsForDispatch, LoopResult, LoopShape, LoopTokenUsage, LoopUntilSpec, LoopWinner, type McpEndpoint, type McpEnvironmentOptions, MountRecorder, type NaiveDriverOptions, type Observation, type ObserveInput, type ObserveOptions, type OpenSandboxRunOptions, Outcome, OutputAdapter, PanelSpec, Persona, PipelineStage, type ProfileRichness, type ProfileRichnessThresholds, type PromotionGateOptions, type PromotionVerdict, type RegistryAnalyzeProjection, RenderCorpusToInstructionsOptions, type ReservationTicket, 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, SupervisedResult, Supervisor, type SupervisorAgentDeps, type SupervisorProfile, type SurfaceScore, ToolLoopChat, 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, fanout, finalizeBestDelivered, flatWidenGate, gitWorkspace, harvestCorpus, inProcessSandboxClient, inlineSandboxClient, jjWorkspace, localShell, loopDispatch, loopUntil, mapSandboxEvent, naiveDriver, observe, openSandboxRun, panel, pickChampion, pipeline, printBenchmarkReport, probeSandboxCapabilities, profileRichnessFinding, promotionGate, refine, registerShape, registryScopeAnalyst, renderAnytimeTable, renderCorpusToInstructions, renderReport, reportLoopUsage, runAgentic, runBenchmark, runInWorkspace, runLoop, runPersonified, runStrategyEvolution, sample, sampleThenRefine, sandboxSessionTraceSource, selectChampion, selectValidWinner, sentinelCompletion, serveCoordinationMcp, settledToIteration, spendFromUsageEvents, stopSentinel, strategyAuthorContract, supervise, supervisorAgent, supervisorInstructions, trajectoryReport, verify, watchTrace, widen, workerFromBackend };

package/dist/loops.js

@@ -29,17 +29,20 @@ import {
   depthStrategy,
   deterministicCompletion,
   discriminatingMeans,
+  dumbDriver,
   equalKOnCost,
   fanout,
   flatWidenGate,
   gitWorkspace,
   harvestCorpus,
+  inProcessSandboxClient,
   inlineSandboxClient,
   jjWorkspace,
   localShell,
   loopDispatch,
   loopUntil,
   makeFinding,
+  naiveDriver,
   observe,
   openSandboxRun,
   panel,
@@ -73,7 +76,7 @@ import {
   watchTrace,
   widen,
   worktreeFanout
-} from "./chunk-SA5GCF2X.js";
+} from "./chunk-CEW5BMGN.js";
 import {
   InMemoryResultBlobStore,
   InMemorySpawnJournal,
@@ -115,7 +118,7 @@ import {
   supervisorAgent,
   supervisorInstructions,
   workerFromBackend
-} from "./chunk-ZKMOIEOB.js";
+} from "./chunk-RQYPH2QE.js";
 import "./chunk-VMNEQHJR.js";
 import {
   extractLlmCallEvent,
@@ -180,6 +183,7 @@ export {
   deterministicCompletion,
   discriminatingMeans,
   driverAgent,
+  dumbDriver,
   equalKOnCost,
   extractLlmCallEvent,
   fanout,
@@ -188,6 +192,7 @@ export {
   gateOnDeliverable,
   gitWorkspace,
   harvestCorpus,
+  inProcessSandboxClient,
   inlineSandboxClient,
   jjWorkspace,
   localShell,
@@ -195,6 +200,7 @@ export {
   loopUntil,
   makeFinding,
   mapSandboxEvent,
+  naiveDriver,
   observe,
   openSandboxRun,
   panel,

package/dist/mcp/bin.js

@@ -9,7 +9,7 @@ import {
   DelegationTaskQueue,
   FileDelegationStore,
   createMcpServer
-} from "../chunk-ZKMOIEOB.js";
+} from "../chunk-RQYPH2QE.js";
 import "../chunk-VMNEQHJR.js";
 import "../chunk-T2HVQVB4.js";
 import "../chunk-O2UPHN7X.js";

package/dist/mcp/index.d.ts

@@ -1,6 +1,6 @@
 import { S as SandboxClient, d as LoopSandboxPlacement } from '../types-B-jWSfcu.js';
-import { F as FleetHandle, D as DelegationExecutor, a as DelegationTaskQueue, b as FeedbackStore, c as DelegateFeedbackResult, d as DelegateFeedbackArgs, U as UiAuditorDelegate, e as DelegateUiAuditArgs, f as DelegateUiAuditResult, g as DelegationHistoryResult, h as DelegationHistoryArgs, i as DelegationStatusResult, j as DelegationStatusArgs } from '../coordination-c_7Olmtq.js';
-export { A as AnalystRegistry, C as CappedDelegationTrace, k as CoderDelegate, l as CoderReview, m as CoderReviewer, n as CoordinationEvent, o as CoordinationTools, p as CoordinationToolsOptions, q as DELEGATE_DESCRIPTION, r as DELEGATE_INPUT_SCHEMA, s as DELEGATE_TOOL_NAME, t as DELEGATION_TRACE_MAX_BYTES, u as DELEGATION_TRACE_MAX_SPANS, v as DelegateArgs, w as DelegateCodeArgs, x as DelegateCodeConfig, y as DelegateCodeResult, z as DelegateHandlerOptions, B as DelegateResearchArgs, E as DelegateResearchConfig, G as DelegateResearchResult, H as DelegateResult, I as DelegateRunCtx, J as DelegateUiAuditConfig, K as DelegateUiAuditRoute, L as DelegationError, M as DelegationFeedbackSnapshot, N as DelegationHistoryEntry, O as DelegationPersistenceError, P as DelegationProfile, Q as DelegationProgress, R as DelegationRecord, S as DelegationResultPayload, T as DelegationResumeContext, V as DelegationResumeDriver, W as DelegationResumeTick, X as DelegationRunContext, Y as DelegationStateCorruptError, Z as DelegationStatus, _ as DelegationStore, $ as DelegationTaskQueueOptions, a0 as DelegationTraceCaps, a1 as DelegationTraceCollector, a2 as DelegationTraceSpan, a3 as DetachedSessionDelegateOptions, a4 as DetachedSessionRefParts, a5 as DetachedTurn, a6 as DetachedTurnResumeDriverOptions, a7 as DetachedWinnerSelection, a8 as DriveTurnCapableBox, a9 as DriveTurnTick, aa as FeedbackEvent, ab as FeedbackRating, ac as FeedbackRefersTo, ad as FileDelegationStore, ae as FileDelegationStoreOptions, af as FleetWorkspaceExecutorOptions, ag as InMemoryDelegationStore, ah as InMemoryFeedbackStore, ai as JsonRpcMessage, aj as JsonRpcResponse, ak as MakeWorkerAgent, al as McpServer, am as McpServerOptions, an as McpToolDescriptor, ao as McpTransport, ap as Question, aq as QuestionDecision, ar as QuestionPolicy, as as QuestionRecord, at as ResearchOutputShape, au as ResearchSource, av as RunDetachedTurnOptions, aw as SettleDetachedCoderTurnOptions, ax as SettledWorker, ay as SiblingSandboxExecutorOptions, az as SubmitInput, aA as SubmitOutput, aB as TraceContext, aC as UiAuditorDelegationOutput, aD as buildDelegationTraceSpans, aE as capDelegationTrace, aF as coderTaskFromArgs, aG as composeLoopTraceEmitters, aH as createCoordinationTools, aI as createDelegateHandler, aJ as createDelegationTraceCollector, aK as createDetachedTurnResumeDriver, aL as createFleetWorkspaceExecutor, aM as createInProcessTransport, aN as createMcpServer, aO as createPropagatingTraceEmitter, aP as createSiblingSandboxExecutor, aQ as detachedSessionDelegate, aR as detachedTurnEvents, aS as eventToSnapshot, aT as formatDetachedSessionRef, aU as hashIdempotencyInput, aV as parseDetachedSessionRef, aW as readTraceContextFromEnv, aX as runDetachedTurn, aY as settleDetachedCoderTurn, aZ as traceContextToEnv, a_ as validateDelegateArgs } from '../coordination-c_7Olmtq.js';
+import { F as FleetHandle, D as DelegationExecutor, a as DelegationTaskQueue, b as FeedbackStore, c as DelegateFeedbackResult, d as DelegateFeedbackArgs, U as UiAuditorDelegate, e as DelegateUiAuditArgs, f as DelegateUiAuditResult, g as DelegationHistoryResult, h as DelegationHistoryArgs, i as DelegationStatusResult, j as DelegationStatusArgs } from '../coordination-CZe4lTxy.js';
+export { A as AnalystRegistry, C as CappedDelegationTrace, k as CoderDelegate, l as CoderReview, m as CoderReviewer, n as CoordinationEvent, o as CoordinationTools, p as CoordinationToolsOptions, q as DELEGATE_DESCRIPTION, r as DELEGATE_INPUT_SCHEMA, s as DELEGATE_TOOL_NAME, t as DELEGATION_TRACE_MAX_BYTES, u as DELEGATION_TRACE_MAX_SPANS, v as DelegateArgs, w as DelegateCodeArgs, x as DelegateCodeConfig, y as DelegateCodeResult, z as DelegateHandlerOptions, B as DelegateResearchArgs, E as DelegateResearchConfig, G as DelegateResearchResult, H as DelegateResult, I as DelegateRunCtx, J as DelegateUiAuditConfig, K as DelegateUiAuditRoute, L as DelegationError, M as DelegationFeedbackSnapshot, N as DelegationHistoryEntry, O as DelegationPersistenceError, P as DelegationProfile, Q as DelegationProgress, R as DelegationRecord, S as DelegationResultPayload, T as DelegationResumeContext, V as DelegationResumeDriver, W as DelegationResumeTick, X as DelegationRunContext, Y as DelegationStateCorruptError, Z as DelegationStatus, _ as DelegationStore, $ as DelegationTaskQueueOptions, a0 as DelegationTraceCaps, a1 as DelegationTraceCollector, a2 as DelegationTraceSpan, a3 as DetachedSessionDelegateOptions, a4 as DetachedSessionRefParts, a5 as DetachedTurn, a6 as DetachedTurnResumeDriverOptions, a7 as DetachedWinnerSelection, a8 as DriveTurnCapableBox, a9 as DriveTurnTick, aa as FeedbackEvent, ab as FeedbackRating, ac as FeedbackRefersTo, ad as FileDelegationStore, ae as FileDelegationStoreOptions, af as FleetWorkspaceExecutorOptions, ag as InMemoryDelegationStore, ah as InMemoryFeedbackStore, ai as JsonRpcMessage, aj as JsonRpcResponse, ak as MakeWorkerAgent, al as McpServer, am as McpServerOptions, an as McpToolDescriptor, ao as McpTransport, ap as Question, aq as QuestionDecision, ar as QuestionPolicy, as as QuestionRecord, at as ResearchOutputShape, au as ResearchSource, av as RunDetachedTurnOptions, aw as SettleDetachedCoderTurnOptions, ax as SettledWorker, ay as SiblingSandboxExecutorOptions, az as SubmitInput, aA as SubmitOutput, aB as TraceContext, aC as UiAuditorDelegationOutput, aD as buildDelegationTraceSpans, aE as capDelegationTrace, aF as coderTaskFromArgs, aG as composeLoopTraceEmitters, aH as createCoordinationTools, aI as createDelegateHandler, aJ as createDelegationTraceCollector, aK as createDetachedTurnResumeDriver, aL as createFleetWorkspaceExecutor, aM as createInProcessTransport, aN as createMcpServer, aO as createPropagatingTraceEmitter, aP as createSiblingSandboxExecutor, aQ as detachedSessionDelegate, aR as detachedTurnEvents, aS as eventToSnapshot, aT as formatDetachedSessionRef, aU as hashIdempotencyInput, aV as parseDetachedSessionRef, aW as readTraceContextFromEnv, aX as runDetachedTurn, aY as settleDetachedCoderTurn, aZ as traceContextToEnv, a_ as validateDelegateArgs } from '../coordination-CZe4lTxy.js';
 import { L as LocalHarness, r as runLocalHarness } from '../local-harness-BE_h8szs.js';
 export { a as LocalHarnessResult, R as RunLocalHarnessOptions } from '../local-harness-BE_h8szs.js';
 import { G as GitRunner } from '../worktree-CDxqwxGo.js';
@@ -12,7 +12,7 @@ import { U as UiLens } from '../substrate-rNj6TDc3.js';
 import '@tangle-network/sandbox';
 import '@tangle-network/agent-interface';
 import '../otel-export-BKmNwiCb.js';
-import '../router-client-C7kp_ECN.js';
+import '../router-client-CMAWGv1h.js';
 import 'node:child_process';
 
 /**

package/dist/mcp/index.js

@@ -6,7 +6,7 @@ import {
 import {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
-} from "../chunk-MRWXCFV5.js";
+} from "../chunk-LFL7K5FM.js";
 import {
   createKbGate
 } from "../chunk-FNMGYYSS.js";
@@ -14,7 +14,7 @@ import {
   assertTraceDerivedFindings,
   runCoderChecks,
   selectValidWinner
-} from "../chunk-SA5GCF2X.js";
+} from "../chunk-CEW5BMGN.js";
 import {
   DELEGATE_DESCRIPTION,
   DELEGATE_FEEDBACK_DESCRIPTION,
@@ -68,7 +68,7 @@ import {
   validateDelegateUiAuditArgs,
   validateDelegationHistoryArgs,
   validateDelegationStatusArgs
-} from "../chunk-ZKMOIEOB.js";
+} from "../chunk-RQYPH2QE.js";
 import "../chunk-VMNEQHJR.js";
 import "../chunk-T2HVQVB4.js";
 import {

package/dist/{router-client-C7kp_ECN.d.ts → router-client-CMAWGv1h.d.ts}

@@ -40,6 +40,14 @@ interface RouterConfig {
     routerBaseUrl: string;
     routerKey: string;
     model: string;
+    /**
+     * Optional completion transport. When set, `routerChatWithUsage` / `routerChatWithTools` call it
+     * with the OpenAI-shape request body and use the parsed `/chat/completions` JSON it returns,
+     * INSTEAD of `fetch(routerBaseUrl + '/chat/completions')`. When absent the fetch path runs
+     * unchanged — the live router stays the default. The injection seam an offline benchmark uses to
+     * drive the worker with no network: a deterministic in-process responder satisfies it, no server.
+     */
+    complete?: (body: Record<string, unknown>) => Promise<unknown>;
 }
 interface RouterChatResult {
     content: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.75.1",
+  "version": "0.77.0",
   "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": {
@@ -75,7 +75,7 @@
   "scripts": {
     "build": "tsup",
     "dev": "tsup --watch",
-    "prepare": "tsup",
+    "prepare": "(git rev-parse --git-dir > /dev/null 2>&1 && git config core.hooksPath .githooks) || true; tsup",
     "test": "vitest run",
     "test:watch": "vitest",
     "lint": "biome check src tests examples",
@@ -83,18 +83,19 @@
     "typecheck": "tsc --noEmit && pnpm run typecheck:examples",
     "typecheck:examples": "tsc --noEmit -p tsconfig.examples.json",
     "verify:package": "node scripts/verify-package-exports.mjs",
-    "docs:api": "typedoc",
+    "docs:api": "typedoc && node scripts/gen-primitive-catalog.mjs",
     "docs:freshness": "node scripts/check-docs-freshness.mjs",
     "docs:check": "pnpm run docs:api && git diff --exit-code -- docs/api && pnpm run docs:freshness"
   },
   "devDependencies": {
     "@biomejs/biome": "^2.4.15",
-    "@tangle-network/agent-eval": ">=0.97.0 <1.0.0",
+    "@tangle-network/agent-eval": ">=0.100.0 <1.0.0",
     "@tangle-network/agent-interface": ">=0.10.0 <1.0.0",
     "@tangle-network/sandbox": ">=0.8.0 <1.0.0",
     "@types/node": "^25.9.3",
     "playwright": "^1.61.0",
     "tsup": "^8.0.0",
+    "tsx": "^4.22.4",
     "typedoc": "0.28.19",
     "typedoc-plugin-markdown": "4.12.0",
     "typescript": "^5.7.0",