@tangle-network/agent-runtime 0.90.0 → 0.90.1
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/dist/agent.d.ts +1 -1
- package/dist/agent.js +4 -4
- package/dist/analyst-loop.d.ts +1 -1
- package/dist/{chunk-4WXGK6GV.js → chunk-7ON74BQO.js} +2 -2
- package/dist/{chunk-7LO5GMAO.js → chunk-CMYMTRGA.js} +14 -5
- package/dist/chunk-CMYMTRGA.js.map +1 -0
- package/dist/{chunk-UD4BHQMI.js → chunk-IVGYLCFH.js} +2 -2
- package/dist/{chunk-UD4BHQMI.js.map → chunk-IVGYLCFH.js.map} +1 -1
- package/dist/{chunk-XMOU4WEY.js → chunk-OOL3675H.js} +12 -2
- package/dist/chunk-OOL3675H.js.map +1 -0
- package/dist/{chunk-XQEISTK2.js → chunk-QK4DV5PR.js} +2 -2
- package/dist/{chunk-74SBMDTO.js → chunk-R2VAJGR3.js} +2 -2
- package/dist/{chunk-QSNSMJSZ.js → chunk-RYBVU4M3.js} +963 -957
- package/dist/chunk-RYBVU4M3.js.map +1 -0
- package/dist/{chunk-MHK62APK.js → chunk-WRUSWK4F.js} +3 -3
- package/dist/{chunk-4IBAMGBE.js → chunk-ZV4LXYCJ.js} +420 -23
- package/dist/chunk-ZV4LXYCJ.js.map +1 -0
- package/dist/{completion-gate-BDaBP8dd.d.ts → completion-gate-DkAnUmpb.d.ts} +2 -2
- package/dist/{coordination-DxHduZg7.d.ts → coordination-rRj5hjJK.d.ts} +5 -5
- package/dist/environment-provider.d.ts +2 -2
- package/dist/index.d.ts +112 -16
- package/dist/index.js +155 -10
- package/dist/index.js.map +1 -1
- package/dist/intelligence.d.ts +227 -83
- package/dist/intelligence.js +259 -166
- package/dist/intelligence.js.map +1 -1
- package/dist/knowledge.d.ts +6 -6
- package/dist/knowledge.js +4 -4
- package/dist/lifecycle.d.ts +2 -2
- package/dist/lifecycle.js +2 -2
- package/dist/{local-harness-sI0S_XNA.d.ts → local-harness-dcD5WTTr.d.ts} +3 -0
- package/dist/{loop-runner-bin-D3GmB2Ru.d.ts → loop-runner-bin-DTbZVGfM.d.ts} +2 -2
- package/dist/loop-runner-bin.d.ts +6 -6
- package/dist/loop-runner-bin.js +5 -5
- package/dist/loops.d.ts +16 -256
- package/dist/loops.js +30 -4
- package/dist/mcp/bin.js +4 -4
- package/dist/mcp/index.d.ts +9 -9
- package/dist/mcp/index.js +6 -6
- package/dist/{mcp-serve-verifier-FL7-ZEb_.d.ts → mcp-serve-verifier-XsX8rkB9.d.ts} +12 -1
- package/dist/{openai-tools-D0xmzo0y.d.ts → openai-tools-C4ZfUD4L.d.ts} +1 -1
- package/dist/profiles.d.ts +1 -1
- package/dist/structural-rollout-MwlpgQ-6.d.ts +446 -0
- package/dist/{supervise-BpCdssu0.d.ts → supervise-DPmYPk0j.d.ts} +3 -3
- package/dist/{types-Dnk189QA.d.ts → types-SyuwunY_.d.ts} +1 -1
- package/dist/{types-DAJQRIUD.d.ts → types-eMNgWgFi.d.ts} +2 -2
- package/dist/{worktree-fanout-CfRXYmgV.d.ts → worktree-fanout-BDFQIO-Y.d.ts} +234 -234
- package/package.json +1 -1
- package/skills/build-with-agent-runtime/SKILL.md +20 -17
- package/dist/chunk-4IBAMGBE.js.map +0 -1
- package/dist/chunk-7LO5GMAO.js.map +0 -1
- package/dist/chunk-QSNSMJSZ.js.map +0 -1
- package/dist/chunk-XMOU4WEY.js.map +0 -1
- /package/dist/{chunk-4WXGK6GV.js.map → chunk-7ON74BQO.js.map} +0 -0
- /package/dist/{chunk-XQEISTK2.js.map → chunk-QK4DV5PR.js.map} +0 -0
- /package/dist/{chunk-74SBMDTO.js.map → chunk-R2VAJGR3.js.map} +0 -0
- /package/dist/{chunk-MHK62APK.js.map → chunk-WRUSWK4F.js.map} +0 -0
package/dist/knowledge.d.ts
CHANGED
|
@@ -1,15 +1,15 @@
|
|
|
1
1
|
import { KnowledgeReadinessSpec, BuildEvalKnowledgeBundleOptions, KnowledgeBaseQualityOptions, KnowledgeImprovementResult, KnowledgeImprovementOptions } from '@tangle-network/agent-knowledge';
|
|
2
|
-
import { E as ExecutorConfig } from './coordination-
|
|
3
|
-
import { a as SuperviseOptions, S as SupervisorProfile } from './supervise-
|
|
4
|
-
import { B as Budget, b as SupervisedResult } from './types-
|
|
5
|
-
import { D as DeliverableSpec } from './completion-gate-
|
|
2
|
+
import { E as ExecutorConfig } from './coordination-rRj5hjJK.js';
|
|
3
|
+
import { a as SuperviseOptions, S as SupervisorProfile } from './supervise-DPmYPk0j.js';
|
|
4
|
+
import { B as Budget, b as SupervisedResult } from './types-eMNgWgFi.js';
|
|
5
|
+
import { D as DeliverableSpec } from './completion-gate-DkAnUmpb.js';
|
|
6
6
|
import '@tangle-network/agent-interface';
|
|
7
7
|
import './substrate-DO2GHNg2.js';
|
|
8
|
-
import './types-
|
|
8
|
+
import './types-SyuwunY_.js';
|
|
9
9
|
import '@tangle-network/agent-eval';
|
|
10
10
|
import '@tangle-network/sandbox';
|
|
11
11
|
import './router-client-DJImUDlm.js';
|
|
12
|
-
import './local-harness-
|
|
12
|
+
import './local-harness-dcD5WTTr.js';
|
|
13
13
|
import 'node:child_process';
|
|
14
14
|
import './environment-provider.js';
|
|
15
15
|
import '@tangle-network/agent-interface/environment-provider';
|
package/dist/knowledge.js
CHANGED
|
@@ -6,14 +6,14 @@ import {
|
|
|
6
6
|
knowledgeReadinessDeliverable,
|
|
7
7
|
runKnowledgeImprovementJob,
|
|
8
8
|
runSupervisedKnowledgeUpdate
|
|
9
|
-
} from "./chunk-
|
|
9
|
+
} from "./chunk-R2VAJGR3.js";
|
|
10
10
|
import "./chunk-5QOB7H74.js";
|
|
11
|
-
import "./chunk-
|
|
11
|
+
import "./chunk-RYBVU4M3.js";
|
|
12
12
|
import "./chunk-DPEUKJRO.js";
|
|
13
|
-
import "./chunk-
|
|
13
|
+
import "./chunk-IVGYLCFH.js";
|
|
14
14
|
import "./chunk-BZF3KQ6G.js";
|
|
15
15
|
import "./chunk-FVJ7M3DA.js";
|
|
16
|
-
import "./chunk-
|
|
16
|
+
import "./chunk-CMYMTRGA.js";
|
|
17
17
|
import "./chunk-YEJR7IXO.js";
|
|
18
18
|
import "./chunk-DGUM43GV.js";
|
|
19
19
|
export {
|
package/dist/lifecycle.d.ts
CHANGED
|
@@ -2,8 +2,8 @@ import { AgentProfile } from '@tangle-network/agent-interface';
|
|
|
2
2
|
import { a as ProfileArtifact, A as ArtifactKind, b as ArtifactStatus, c as ArtifactInput, E as EvalRunner, d as EvalResult, G as GenerateContext, C as CandidateGenerator, e as PromotionVerdict, P as PromotionGate } from './generator-YkAQrOoD.js';
|
|
3
3
|
export { f as ArtifactPayloads, H as HeldOutPromotionGateOptions, M as MarginalLift, g as MeasureMarginalLiftOptions, h as heldOutPromotionGate, m as measureMarginalLift, t as thresholdPromotionGate } from './generator-YkAQrOoD.js';
|
|
4
4
|
import { LlmClientOptions, AnalystFinding } from '@tangle-network/agent-eval';
|
|
5
|
-
import { M as McpServeSpec } from './mcp-serve-verifier-
|
|
6
|
-
import { L as LocalHarness } from './local-harness-
|
|
5
|
+
import { M as McpServeSpec } from './mcp-serve-verifier-XsX8rkB9.js';
|
|
6
|
+
import { L as LocalHarness } from './local-harness-dcD5WTTr.js';
|
|
7
7
|
import '@tangle-network/agent-eval/campaign';
|
|
8
8
|
import 'node:child_process';
|
|
9
9
|
|
package/dist/lifecycle.js
CHANGED
|
@@ -39,6 +39,9 @@ interface RunLocalHarnessOptions {
|
|
|
39
39
|
command?: string;
|
|
40
40
|
args: ReadonlyArray<string>;
|
|
41
41
|
};
|
|
42
|
+
/** Allow autonomous Claude edits without an interactive permission prompt.
|
|
43
|
+
* Use only when `cwd` is an isolated candidate worktree. */
|
|
44
|
+
dangerouslySkipPermissions?: boolean;
|
|
42
45
|
/** Wall-clock kill deadline (ms). Default 5 min. Subprocess SIGTERMed on expiry. */
|
|
43
46
|
timeoutMs?: number;
|
|
44
47
|
/** Caller cancellation. SIGTERM is sent on abort. */
|
|
@@ -2,8 +2,8 @@ import { Scenario } from '@tangle-network/agent-eval/campaign';
|
|
|
2
2
|
import { SelfImproveOptions, SelfImproveResult } from '@tangle-network/agent-eval/contract';
|
|
3
3
|
import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-BC3bZpH0.js';
|
|
4
4
|
import { F as FactCandidate, C as CreateKbGateOptions } from './kb-gate-CwHO0vz6.js';
|
|
5
|
-
import { B as Budget } from './types-
|
|
6
|
-
import { A as AuthoredHarness, W as WinnerStrategy, a as WorktreeFanoutOptions, b as WorktreePatchArtifact } from './worktree-fanout-
|
|
5
|
+
import { B as Budget } from './types-eMNgWgFi.js';
|
|
6
|
+
import { A as AuthoredHarness, W as WinnerStrategy, a as WorktreeFanoutOptions, b as WorktreePatchArtifact } from './worktree-fanout-BDFQIO-Y.js';
|
|
7
7
|
|
|
8
8
|
/**
|
|
9
9
|
*
|
|
@@ -1,15 +1,15 @@
|
|
|
1
1
|
#!/usr/bin/env node
|
|
2
|
-
export { L as LoopRunnerCliArgs, e as LoopRunnerCliResult, p as parseLoopRunnerArgv, k as runLoopRunnerCli } from './loop-runner-bin-
|
|
2
|
+
export { L as LoopRunnerCliArgs, e as LoopRunnerCliResult, p as parseLoopRunnerArgv, k as runLoopRunnerCli } from './loop-runner-bin-DTbZVGfM.js';
|
|
3
3
|
import '@tangle-network/agent-eval/campaign';
|
|
4
4
|
import '@tangle-network/agent-eval/contract';
|
|
5
5
|
import './types-BC3bZpH0.js';
|
|
6
6
|
import '@tangle-network/agent-eval';
|
|
7
7
|
import './kb-gate-CwHO0vz6.js';
|
|
8
|
-
import './types-
|
|
8
|
+
import './types-eMNgWgFi.js';
|
|
9
9
|
import '@tangle-network/agent-interface';
|
|
10
10
|
import '@tangle-network/sandbox';
|
|
11
|
-
import './types-
|
|
12
|
-
import './worktree-fanout-
|
|
13
|
-
import './completion-gate-
|
|
14
|
-
import './local-harness-
|
|
11
|
+
import './types-SyuwunY_.js';
|
|
12
|
+
import './worktree-fanout-BDFQIO-Y.js';
|
|
13
|
+
import './completion-gate-DkAnUmpb.js';
|
|
14
|
+
import './local-harness-dcD5WTTr.js';
|
|
15
15
|
import 'node:child_process';
|
package/dist/loop-runner-bin.js
CHANGED
|
@@ -2,16 +2,16 @@
|
|
|
2
2
|
import {
|
|
3
3
|
parseLoopRunnerArgv,
|
|
4
4
|
runLoopRunnerCli
|
|
5
|
-
} from "./chunk-
|
|
5
|
+
} from "./chunk-WRUSWK4F.js";
|
|
6
6
|
import "./chunk-SGKPNBXE.js";
|
|
7
|
-
import "./chunk-
|
|
7
|
+
import "./chunk-ZV4LXYCJ.js";
|
|
8
8
|
import "./chunk-ZQZX77MM.js";
|
|
9
|
-
import "./chunk-
|
|
9
|
+
import "./chunk-RYBVU4M3.js";
|
|
10
10
|
import "./chunk-DPEUKJRO.js";
|
|
11
|
-
import "./chunk-
|
|
11
|
+
import "./chunk-IVGYLCFH.js";
|
|
12
12
|
import "./chunk-BZF3KQ6G.js";
|
|
13
13
|
import "./chunk-FVJ7M3DA.js";
|
|
14
|
-
import "./chunk-
|
|
14
|
+
import "./chunk-CMYMTRGA.js";
|
|
15
15
|
import "./chunk-YEJR7IXO.js";
|
|
16
16
|
import "./chunk-DGUM43GV.js";
|
|
17
17
|
export {
|
package/dist/loops.d.ts
CHANGED
|
@@ -2,28 +2,30 @@ import { ChatClient, RunRecord, HarnessType, AgentProfile, AnalystFinding, Analy
|
|
|
2
2
|
export { AnalystFinding, DefaultVerdict, computeFindingId, makeFinding } from '@tangle-network/agent-eval';
|
|
3
3
|
import { SandboxEvent, SandboxInstance, CreateSandboxOptions, PromptOptions, TaskOptions } from '@tangle-network/sandbox';
|
|
4
4
|
export { AgentProfile, CreateSandboxOptions, SandboxEvent, SandboxInstance } from '@tangle-network/sandbox';
|
|
5
|
-
import { a as ResultBlobStore,
|
|
6
|
-
export {
|
|
7
|
-
import { M as MakeWorkerAgent, A as AnalystRegistry, o as CoordinationTools, n as CoordinationEvent, as as QuestionPolicy, E as ExecutorConfig } from './coordination-
|
|
8
|
-
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-
|
|
9
|
-
import { R as RuntimeHooks, I as Iteration, S as SandboxClient, D as Driver, A as AgentRunSpec,
|
|
10
|
-
export {
|
|
5
|
+
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';
|
|
6
|
+
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';
|
|
7
|
+
import { M as MakeWorkerAgent, A as AnalystRegistry, o as CoordinationTools, n as CoordinationEvent, as as QuestionPolicy, E as ExecutorConfig } from './coordination-rRj5hjJK.js';
|
|
8
|
+
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';
|
|
9
|
+
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';
|
|
10
|
+
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';
|
|
11
11
|
import { RunProfileMatrixResult, Scenario, ProfileDispatchFn, JudgeConfig, RunProfileMatrixOptions, DispatchFn } from '@tangle-network/agent-eval/campaign';
|
|
12
12
|
export { AgentEnvironmentProviderRef, AgentEnvironmentProviderRegistry, ProviderAsSandboxClientOptions, ProviderExecutorOptions, SandboxClientProviderOptions, createAgentEnvironmentProviderRegistry, providerAsExecutor, providerAsSandboxClient, resolveAgentEnvironmentProvider, sandboxClientAsProvider } from './environment-provider.js';
|
|
13
|
-
import {
|
|
14
|
-
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-
|
|
13
|
+
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';
|
|
14
|
+
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';
|
|
15
|
+
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';
|
|
16
|
+
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';
|
|
15
17
|
import { b as AnalystRegistryLike } from './types-BC3bZpH0.js';
|
|
16
18
|
import { AgentProfile as AgentProfile$1 } from '@tangle-network/agent-interface';
|
|
17
19
|
import { R as RouterConfig, T as ToolLoopChat, a as ToolLoopCompactionOptions } from './router-client-DJImUDlm.js';
|
|
18
20
|
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';
|
|
19
|
-
import { D as DeliverableSpec } from './completion-gate-
|
|
20
|
-
export { g as WorktreeCommandResult, h as gateOnDeliverable } from './completion-gate-
|
|
21
|
-
import { S as SupervisorProfile } from './supervise-
|
|
22
|
-
export { D as DriveHarness, a as SuperviseOptions, b as SupervisorAgentDeps, s as supervise, c as supervisorAgent, w as workerFromBackend } from './supervise-
|
|
21
|
+
import { D as DeliverableSpec } from './completion-gate-DkAnUmpb.js';
|
|
22
|
+
export { g as WorktreeCommandResult, h as gateOnDeliverable } from './completion-gate-DkAnUmpb.js';
|
|
23
|
+
import { S as SupervisorProfile } from './supervise-DPmYPk0j.js';
|
|
24
|
+
export { D as DriveHarness, a as SuperviseOptions, b as SupervisorAgentDeps, s as supervise, c as supervisorAgent, w as workerFromBackend } from './supervise-DPmYPk0j.js';
|
|
23
25
|
import { stuckLoopView, toolWasteView } from '@tangle-network/agent-eval/pipelines';
|
|
24
26
|
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';
|
|
25
27
|
import './substrate-DO2GHNg2.js';
|
|
26
|
-
import './local-harness-
|
|
28
|
+
import './local-harness-dcD5WTTr.js';
|
|
27
29
|
import 'node:child_process';
|
|
28
30
|
|
|
29
31
|
/**
|
|
@@ -1060,248 +1062,6 @@ declare function loopCampaignDispatch<Task, Output, Decision, TScenario extends
|
|
|
1060
1062
|
*/
|
|
1061
1063
|
declare function loopDispatch<Task, Output, Decision, TScenario extends Scenario, TArtifact>(opts: LoopDispatchOptions<Task, Output, Decision, TScenario, TArtifact>): ProfileDispatchFn<TScenario, TArtifact>;
|
|
1062
1064
|
|
|
1063
|
-
/**
|
|
1064
|
-
* The general agentic primitive — sequential (depth) and parallel (breadth) over a shared,
|
|
1065
|
-
* checkable artifact, driven through the keystone Supervisor as one recursive `Agent.act`.
|
|
1066
|
-
*
|
|
1067
|
-
* The domain lives behind ONE seam — `AgenticSurface` (open an artifact, list tools, call a tool,
|
|
1068
|
-
* score the artifact, close it). EnterpriseOps implements it (seed a gym DB, MCP tools, SQL
|
|
1069
|
-
* verifier); Commit0/AppWorld/terminal-bench implement it the same way (a repo workspace, shell
|
|
1070
|
-
* tools, the test suite). The drivers below are domain-blind: they run over any surface.
|
|
1071
|
-
*
|
|
1072
|
-
* Two shapes, the agent's POMDP rollout as the unit:
|
|
1073
|
-
* - DEPTH one persistent artifact carried across shots. Each shot the agent works the tool loop;
|
|
1074
|
-
* between shots a trace-analyst (selector≠judge: reads the trajectory, never the score)
|
|
1075
|
-
* steers the resumed session toward what's unfinished. shot n stands on shot n-1's
|
|
1076
|
-
* artifact state + history. This is continuation — long-horizon, same artifact.
|
|
1077
|
-
* - BREADTH K independent artifacts, each a fresh rollout, the deployable verifier picks the best.
|
|
1078
|
-
*
|
|
1079
|
-
* Both are an `Agent` whose `act` spawns leaf shots through `scope.spawn` and reacts via
|
|
1080
|
-
* `scope.next()` — so the conserved budget pool meters them (equal-k by construction), the journal
|
|
1081
|
-
* records the tree, and the same primitive nests. `runAgentic` runs the chosen driver through
|
|
1082
|
-
* `createSupervisor().run`. The leaf (one shot over a handle) is resolved per-spawn from a
|
|
1083
|
-
* surface-closed registry — the open `Executor` seam, not bespoke per-benchmark glue.
|
|
1084
|
-
*/
|
|
1085
|
-
|
|
1086
|
-
interface AgenticTask {
|
|
1087
|
-
readonly id: string;
|
|
1088
|
-
readonly systemPrompt: string;
|
|
1089
|
-
readonly userPrompt: string;
|
|
1090
|
-
/** Opaque domain payload the surface reads (EOPS: servers/verifiers/tools). Drivers never read it. */
|
|
1091
|
-
readonly meta?: Record<string, unknown>;
|
|
1092
|
-
}
|
|
1093
|
-
interface ArtifactHandle {
|
|
1094
|
-
readonly id: string;
|
|
1095
|
-
readonly surface: string;
|
|
1096
|
-
/** Opaque per-artifact context the surface stashes (EOPS: the seeded gym server + db id). */
|
|
1097
|
-
readonly ctx?: unknown;
|
|
1098
|
-
}
|
|
1099
|
-
interface AgenticTool {
|
|
1100
|
-
readonly type: 'function';
|
|
1101
|
-
readonly function: {
|
|
1102
|
-
name: string;
|
|
1103
|
-
description?: string;
|
|
1104
|
-
parameters: Record<string, unknown>;
|
|
1105
|
-
};
|
|
1106
|
-
}
|
|
1107
|
-
interface SurfaceScore {
|
|
1108
|
-
passes: number;
|
|
1109
|
-
total: number;
|
|
1110
|
-
/** Checks excluded as malformed (data defect, not the agent). `total === 0` ⇒ unscoreable. */
|
|
1111
|
-
errored: number;
|
|
1112
|
-
}
|
|
1113
|
-
/** A stateful, checkable environment an agent operates over with tools. Open behind one interface. */
|
|
1114
|
-
interface AgenticSurface {
|
|
1115
|
-
readonly name: string;
|
|
1116
|
-
open(task: AgenticTask): Promise<ArtifactHandle>;
|
|
1117
|
-
tools(task: AgenticTask, handle: ArtifactHandle): Promise<AgenticTool[]>;
|
|
1118
|
-
call(handle: ArtifactHandle, name: string, args: Record<string, unknown>): Promise<string>;
|
|
1119
|
-
score(task: AgenticTask, handle: ArtifactHandle): Promise<SurfaceScore>;
|
|
1120
|
-
close(handle: ArtifactHandle): Promise<void>;
|
|
1121
|
-
}
|
|
1122
|
-
interface AgenticOptions {
|
|
1123
|
-
routerBaseUrl: string;
|
|
1124
|
-
routerKey: string;
|
|
1125
|
-
model: string;
|
|
1126
|
-
/** Optional completion transport (see `RouterConfig.complete`): when set, BOTH legs of an
|
|
1127
|
-
* offline run use it instead of `fetch`-ing the router — the worker's tool loop (threaded into
|
|
1128
|
-
* its `routerToolLoop` cfg) AND the analyst's critic (its `ChatClient` is bound to this same
|
|
1129
|
-
* transport). One injected responder serves both, as a localhost mock endpoint would. Absent ⇒
|
|
1130
|
-
* the live router fetch path (the default). */
|
|
1131
|
-
complete?: (body: Record<string, unknown>) => Promise<unknown>;
|
|
1132
|
-
temperature?: number;
|
|
1133
|
-
/** Completion cap per worker turn — REQUIRED for thinking models (they burn unbounded
|
|
1134
|
-
* budgets on reasoning and return empty content without it). Omitted ⇒ provider default. */
|
|
1135
|
-
maxTokens?: number;
|
|
1136
|
-
/** Turns the agent may take within ONE shot before the driver intervenes. */
|
|
1137
|
-
innerTurns?: number;
|
|
1138
|
-
/** The depth STEERER's analyst instruction (observe()'s system prompt). The knob a
|
|
1139
|
-
* prompt optimizer (GEPA) tunes — the analyst IS the steerer. Omitted ⇒ the default. */
|
|
1140
|
-
analystInstruction?: string;
|
|
1141
|
-
/** The critic's model — lets the analyst be a stronger (or cheaper) model than the
|
|
1142
|
-
* worker. Omitted ⇒ the worker's `model`. */
|
|
1143
|
-
analystModel?: string;
|
|
1144
|
-
/** Across-run learning: when set, the analyst's observe() pass appends trace-derived
|
|
1145
|
-
* facts here (the flywheel write side). Read-back is opt-in via `corpusReadback`
|
|
1146
|
-
* because unconditional priming can pollute context on some domains. */
|
|
1147
|
-
corpus?: Corpus;
|
|
1148
|
-
/** Tags written onto learned facts (and used by the caller's priming query). */
|
|
1149
|
-
corpusTags?: string[];
|
|
1150
|
-
/** In-context learning: when set, query `corpus` before each depth shot and inject
|
|
1151
|
-
* the top trace-derived facts as guidance for the active run. No corpus means no read-back. */
|
|
1152
|
-
corpusReadback?: CorpusReadbackOptions;
|
|
1153
|
-
}
|
|
1154
|
-
interface CorpusReadbackOptions {
|
|
1155
|
-
/** Minimum confidence for a fact to be injected. Default 0.7. */
|
|
1156
|
-
minConfidence?: number;
|
|
1157
|
-
/** Extra tags a fact must carry, in addition to `corpusTags`. */
|
|
1158
|
-
tags?: ReadonlyArray<string>;
|
|
1159
|
-
/** Max facts injected per shot. Default 3. */
|
|
1160
|
-
maxFacts?: number;
|
|
1161
|
-
/** Default false: only facts tagged `audience:agent` are injected into the worker. */
|
|
1162
|
-
includeOperatorFacts?: boolean;
|
|
1163
|
-
}
|
|
1164
|
-
type Msg = Record<string, unknown>;
|
|
1165
|
-
interface ShotResult {
|
|
1166
|
-
messages: Msg[];
|
|
1167
|
-
score: number;
|
|
1168
|
-
passes: number;
|
|
1169
|
-
total: number;
|
|
1170
|
-
completions: number;
|
|
1171
|
-
toolErrors: number;
|
|
1172
|
-
}
|
|
1173
|
-
interface AgenticRunResult {
|
|
1174
|
-
/** The strategy name (built-in 'depth'/'breadth' or a custom strategy's name). */
|
|
1175
|
-
mode: string;
|
|
1176
|
-
score: number;
|
|
1177
|
-
resolved: boolean;
|
|
1178
|
-
completions: number;
|
|
1179
|
-
/** DEPTH: score after each shot — the progress-over-rounds curve. BREADTH: best-so-far per rollout. */
|
|
1180
|
-
progression: number[];
|
|
1181
|
-
shots: number;
|
|
1182
|
-
/** The cost vector, stamped by `runAgentic` from the Supervisor's conserved pool: real
|
|
1183
|
-
* router tokens, priced usd (0 when the model is unpriced — never fabricated), wall ms. */
|
|
1184
|
-
usd: number;
|
|
1185
|
-
ms: number;
|
|
1186
|
-
tokens: {
|
|
1187
|
-
input: number;
|
|
1188
|
-
output: number;
|
|
1189
|
-
};
|
|
1190
|
-
}
|
|
1191
|
-
/** DEPTH: one persistent artifact, carried across analyst-steered shots. */
|
|
1192
|
-
declare function depthStrategy(surface: AgenticSurface, task: AgenticTask, opts: AgenticOptions, cfg: {
|
|
1193
|
-
maxShots: number;
|
|
1194
|
-
}): Agent<unknown, Outcome<unknown>>;
|
|
1195
|
-
/** BREADTH: K independent rollouts (each own artifact), verifier picks the best. */
|
|
1196
|
-
declare function breadthStrategy(_surface: AgenticSurface, task: AgenticTask, opts: AgenticOptions, cfg: {
|
|
1197
|
-
width: number;
|
|
1198
|
-
}): Agent<unknown, Outcome<unknown>>;
|
|
1199
|
-
/**
|
|
1200
|
-
* A Strategy is HOW you spend the compute budget to beat the Environment's check — it
|
|
1201
|
-
* builds the driver `Agent` the Supervisor runs. This is the OPEN extension point: a dev
|
|
1202
|
-
* authors their own by implementing `driver()` to return an Agent whose `act()` spawns
|
|
1203
|
-
* shots/analysts via `scope.spawn` / `scope.next` / `scope.send`. The two built-ins are
|
|
1204
|
-
* the reference implementations to copy:
|
|
1205
|
-
* sample — K INDEPENDENT attempts, keep the best-verifying (best-of-N / resample).
|
|
1206
|
-
* refine — attempt → observe() reads the trace → steer the next → repeat (iterate).
|
|
1207
|
-
* (A multi-agent "team" is just a Strategy whose driver spawns several different agents.)
|
|
1208
|
-
*/
|
|
1209
|
-
interface Strategy {
|
|
1210
|
-
readonly name: string;
|
|
1211
|
-
driver(surface: AgenticSurface, task: AgenticTask, opts: AgenticOptions, budget: number): Agent<unknown, Outcome<unknown>>;
|
|
1212
|
-
}
|
|
1213
|
-
/** Built-in `Strategy`: K independent attempts, keep the best-verifying (best-of-N / resample). */
|
|
1214
|
-
declare const sample: Strategy;
|
|
1215
|
-
/** Built-in `Strategy`: attempt → `observe()` reads the trace → steer the next attempt → repeat (deepen one lineage). */
|
|
1216
|
-
declare const refine: Strategy;
|
|
1217
|
-
/** A role for one shot — multi-agent loops (researcher + engineer, a panel of k
|
|
1218
|
-
* researchers) give each shot its own system prompt and optionally its own model. */
|
|
1219
|
-
interface ShotPersona {
|
|
1220
|
-
/** Replaces the task's systemPrompt for a FRESH shot; on a carried conversation it is
|
|
1221
|
-
* injected as a hand-off message (the transcript's earlier roles stay intact). */
|
|
1222
|
-
systemPrompt?: string;
|
|
1223
|
-
/** Per-shot model override (e.g. a stronger model for the engineer shot). */
|
|
1224
|
-
model?: string;
|
|
1225
|
-
}
|
|
1226
|
-
interface ShotSpec {
|
|
1227
|
-
/** present ⇒ continue this artifact (depth); absent ⇒ the shot opens a fresh one (sample/restart). */
|
|
1228
|
-
handle?: ArtifactHandle;
|
|
1229
|
-
messages?: Msg[];
|
|
1230
|
-
steer?: string;
|
|
1231
|
-
persona?: ShotPersona;
|
|
1232
|
-
/** Restrict THIS shot to a subset of the domain's tools (by name) — focus a shot on
|
|
1233
|
-
* the relevant capabilities. Restriction-only; unknown names throw. Omitted ⇒ all. */
|
|
1234
|
-
tools?: string[];
|
|
1235
|
-
}
|
|
1236
|
-
interface StrategyResult {
|
|
1237
|
-
score: number;
|
|
1238
|
-
resolved: boolean;
|
|
1239
|
-
completions: number;
|
|
1240
|
-
progression: number[];
|
|
1241
|
-
shots: number;
|
|
1242
|
-
}
|
|
1243
|
-
/** Artifact lifecycle a strategy may manage itself — open/close ONLY. Raw `call`/`score`
|
|
1244
|
-
* are withheld: scores reach the body solely through `shot()`'s ShotResult (the
|
|
1245
|
-
* harness-verified channel), so a body cannot peek the check or fabricate around it. */
|
|
1246
|
-
interface StrategyArtifacts {
|
|
1247
|
-
readonly name: string;
|
|
1248
|
-
open(task: AgenticTask): Promise<ArtifactHandle>;
|
|
1249
|
-
close(handle: ArtifactHandle): Promise<void>;
|
|
1250
|
-
}
|
|
1251
|
-
/** What a strategy body composes with: the artifact lifecycle, the budget, and the two steps. */
|
|
1252
|
-
interface StrategyCtx {
|
|
1253
|
-
/** Open/close artifacts the body manages itself (e.g. one persistent handle for depth). */
|
|
1254
|
-
readonly surface: StrategyArtifacts;
|
|
1255
|
-
readonly task: AgenticTask;
|
|
1256
|
-
readonly opts: AgenticOptions;
|
|
1257
|
-
readonly budget: number;
|
|
1258
|
-
readonly scope: Scope<Outcome<unknown>>;
|
|
1259
|
-
/** Run ONE worker shot; its harness-scored result, or null if it went down. */
|
|
1260
|
-
shot(spec?: ShotSpec): Promise<ShotResult | null>;
|
|
1261
|
-
/** The firewalled critic reads the trajectory → a steer string, or null on COMPLETE/down. */
|
|
1262
|
-
critique(messages: Msg[]): Promise<string | null>;
|
|
1263
|
-
/** The RAW analyst channel: the firewalled critic answers `instruction` over the
|
|
1264
|
-
* trajectory verbatim — no findings extraction, so verdict-shaped formats
|
|
1265
|
-
* (CONTINUE/STOP decisions, calibrated predictions) survive. Same firewall:
|
|
1266
|
-
* trajectory in, never scores. Null when the analyst went down. */
|
|
1267
|
-
consult(messages: Msg[], instruction: string): Promise<string | null>;
|
|
1268
|
-
/** The tools THIS artifact's task actually offers (names + descriptions only — never
|
|
1269
|
-
* the implementations). Tool sets vary per task on heterogeneous domains; a strategy
|
|
1270
|
-
* that restricts shots MUST select from this list, never from hardcoded names. */
|
|
1271
|
-
listTools(handle: ArtifactHandle): Promise<Array<{
|
|
1272
|
-
name: string;
|
|
1273
|
-
description?: string;
|
|
1274
|
-
}>>;
|
|
1275
|
-
}
|
|
1276
|
-
/** Author a Strategy from the composable steps — the open, compact way. */
|
|
1277
|
-
declare function defineStrategy(name: string, run: (ctx: StrategyCtx) => Promise<StrategyResult>): Strategy;
|
|
1278
|
-
/** A NEW strategy, authored from the steps (~20 lines): refine, but when a steered shot
|
|
1279
|
-
* fails to improve the score it ABANDONS that line and restarts fresh (branch-when-stuck)
|
|
1280
|
-
* — the widen/MCTS idea the depth-stuck failure motivated. Scored keep-best (the best
|
|
1281
|
-
* checkpoint across all lines), the deployable metric. This is the "experts build BETTER
|
|
1282
|
-
* optimizations" path: a new technique, compact, with zero Supervisor ceremony. */
|
|
1283
|
-
declare const adaptiveRefine: Strategy;
|
|
1284
|
-
/** The explore-then-exploit MIX: spend ⌈budget/2⌉ on independent samples (kept open),
|
|
1285
|
-
* then refine the best-verifying line with the remaining budget. Sample's basin escape +
|
|
1286
|
-
* refine's accumulation — the third built-in, authored from the public steps. */
|
|
1287
|
-
declare const sampleThenRefine: Strategy;
|
|
1288
|
-
interface RunAgenticOptions extends AgenticOptions {
|
|
1289
|
-
surface: AgenticSurface;
|
|
1290
|
-
task: AgenticTask;
|
|
1291
|
-
/** Lifecycle observability — every spawn/settle (shots, analysts) streams here live.
|
|
1292
|
-
* The seam online watchdogs/route-auditors subscribe to. */
|
|
1293
|
-
hooks?: RuntimeHooks;
|
|
1294
|
-
/** A Strategy (the open way) — author/pass your own. Overrides `mode` when present. */
|
|
1295
|
-
strategy?: Strategy;
|
|
1296
|
-
/** Built-in shorthand: 'depth'→refine, 'breadth'→sample. Default 'depth'. */
|
|
1297
|
-
mode?: 'depth' | 'breadth';
|
|
1298
|
-
/** budget: refine→max shots; sample→rollout width. */
|
|
1299
|
-
budget: number;
|
|
1300
|
-
rootBudget?: Budget;
|
|
1301
|
-
}
|
|
1302
|
-
/** Run a Strategy through the keystone Supervisor — `Agent.act` over a conserved-budget Scope. */
|
|
1303
|
-
declare function runAgentic(opts: RunAgenticOptions): Promise<AgenticRunResult>;
|
|
1304
|
-
|
|
1305
1065
|
/**
|
|
1306
1066
|
* runBenchmark — the packaged optimization suite. Define a domain by implementing an
|
|
1307
1067
|
* `Environment` (open / tools / call / score / close); get the optimization strategies
|
|
@@ -3994,4 +3754,4 @@ declare function runInWorkspace<T>(ws: Workspace, body: (cwd: string) => Promise
|
|
|
3994
3754
|
commitOnInvalid?: boolean;
|
|
3995
3755
|
}): Promise<WorkspaceRun<T>>;
|
|
3996
3756
|
|
|
3997
|
-
export { Agent, AgentRunSpec, type AgentTurnBackend, type AgentTurnUsage,
|
|
3757
|
+
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 };
|
package/dist/loops.js
CHANGED
|
@@ -12,8 +12,11 @@ import {
|
|
|
12
12
|
breadthStrategy,
|
|
13
13
|
buildSteerContext,
|
|
14
14
|
builtinShapes,
|
|
15
|
+
canDisplace,
|
|
15
16
|
collectAgentTurn,
|
|
17
|
+
compareCheckOutcomes,
|
|
16
18
|
completionAuthorizes,
|
|
19
|
+
composeCheckSources,
|
|
17
20
|
computeFindingId,
|
|
18
21
|
createMcpEnvironment,
|
|
19
22
|
createPushTraceSource,
|
|
@@ -24,6 +27,8 @@ import {
|
|
|
24
27
|
decodeToolPart,
|
|
25
28
|
defaultAnalystInstruction,
|
|
26
29
|
defaultAuditorInstruction,
|
|
30
|
+
defaultExtractCandidate,
|
|
31
|
+
defaultStructuralRolloutPolicy,
|
|
27
32
|
defaultToolDetectors,
|
|
28
33
|
defineLeaderboard,
|
|
29
34
|
definePersona,
|
|
@@ -35,6 +40,7 @@ import {
|
|
|
35
40
|
equalKOnCost,
|
|
36
41
|
failuresAnalyst,
|
|
37
42
|
fanout,
|
|
43
|
+
filterAuthoredAsserts,
|
|
38
44
|
flatWidenGate,
|
|
39
45
|
gitWorkspace,
|
|
40
46
|
harvestCorpus,
|
|
@@ -47,8 +53,10 @@ import {
|
|
|
47
53
|
loopDispatch,
|
|
48
54
|
loopUntil,
|
|
49
55
|
makeFinding,
|
|
56
|
+
modelAuthoredChecks,
|
|
50
57
|
naiveDriver,
|
|
51
58
|
observe,
|
|
59
|
+
officialChecksFromMeta,
|
|
52
60
|
openSandboxRun,
|
|
53
61
|
pairwiseSignificance,
|
|
54
62
|
panel,
|
|
@@ -68,6 +76,7 @@ import {
|
|
|
68
76
|
renderPairwiseMarkdown,
|
|
69
77
|
renderReport,
|
|
70
78
|
reportLoopUsage,
|
|
79
|
+
resolveEntrySymbol,
|
|
71
80
|
resolveSandboxClient,
|
|
72
81
|
runAgentic,
|
|
73
82
|
runBenchmark,
|
|
@@ -76,20 +85,24 @@ import {
|
|
|
76
85
|
runStrategyEvolution,
|
|
77
86
|
sample,
|
|
78
87
|
sampleThenRefine,
|
|
88
|
+
sandboxCheckRunner,
|
|
79
89
|
sandboxSessionTraceSource,
|
|
90
|
+
selectBestIndex,
|
|
80
91
|
selectChampion,
|
|
81
92
|
selectValidWinner,
|
|
82
93
|
sentinelCompletion,
|
|
83
94
|
stopSentinel,
|
|
84
95
|
strategyAuthorContract,
|
|
85
96
|
streamAgentTurn,
|
|
97
|
+
structuralRollout,
|
|
86
98
|
superviseSurface,
|
|
87
99
|
trajectoryReport,
|
|
88
100
|
verify,
|
|
101
|
+
visibleCheckScore,
|
|
89
102
|
watchTrace,
|
|
90
103
|
widen,
|
|
91
104
|
worktreeFanout
|
|
92
|
-
} from "./chunk-
|
|
105
|
+
} from "./chunk-ZV4LXYCJ.js";
|
|
93
106
|
import {
|
|
94
107
|
InMemoryResultBlobStore,
|
|
95
108
|
InMemorySpawnJournal,
|
|
@@ -131,9 +144,9 @@ import {
|
|
|
131
144
|
supervisorAgent,
|
|
132
145
|
supervisorInstructions,
|
|
133
146
|
workerFromBackend
|
|
134
|
-
} from "./chunk-
|
|
147
|
+
} from "./chunk-RYBVU4M3.js";
|
|
135
148
|
import "./chunk-DPEUKJRO.js";
|
|
136
|
-
import "./chunk-
|
|
149
|
+
import "./chunk-IVGYLCFH.js";
|
|
137
150
|
import {
|
|
138
151
|
createAgentEnvironmentProviderRegistry,
|
|
139
152
|
providerAsExecutor,
|
|
@@ -148,7 +161,7 @@ import {
|
|
|
148
161
|
mapSandboxToolEvent,
|
|
149
162
|
sumSandboxUsage
|
|
150
163
|
} from "./chunk-FVJ7M3DA.js";
|
|
151
|
-
import "./chunk-
|
|
164
|
+
import "./chunk-CMYMTRGA.js";
|
|
152
165
|
import "./chunk-YEJR7IXO.js";
|
|
153
166
|
import "./chunk-DGUM43GV.js";
|
|
154
167
|
export {
|
|
@@ -172,9 +185,12 @@ export {
|
|
|
172
185
|
breadthStrategy,
|
|
173
186
|
buildSteerContext,
|
|
174
187
|
builtinShapes,
|
|
188
|
+
canDisplace,
|
|
175
189
|
cliWorktreeExecutor,
|
|
176
190
|
collectAgentTurn,
|
|
191
|
+
compareCheckOutcomes,
|
|
177
192
|
completionAuthorizes,
|
|
193
|
+
composeCheckSources,
|
|
178
194
|
computeFindingId,
|
|
179
195
|
contentAddress,
|
|
180
196
|
createAgentEnvironmentProviderRegistry,
|
|
@@ -199,8 +215,10 @@ export {
|
|
|
199
215
|
defaultAnalystInstruction,
|
|
200
216
|
defaultAuditorInstruction,
|
|
201
217
|
defaultDelegateBudget,
|
|
218
|
+
defaultExtractCandidate,
|
|
202
219
|
defaultProfileRichnessThresholds,
|
|
203
220
|
defaultSelectWinner,
|
|
221
|
+
defaultStructuralRolloutPolicy,
|
|
204
222
|
defaultToolDetectors,
|
|
205
223
|
defineLeaderboard,
|
|
206
224
|
definePersona,
|
|
@@ -215,6 +233,7 @@ export {
|
|
|
215
233
|
extractLlmCallEvent,
|
|
216
234
|
failuresAnalyst,
|
|
217
235
|
fanout,
|
|
236
|
+
filterAuthoredAsserts,
|
|
218
237
|
finalizeBestDelivered,
|
|
219
238
|
flatWidenGate,
|
|
220
239
|
gateOnDeliverable,
|
|
@@ -231,8 +250,10 @@ export {
|
|
|
231
250
|
makeFinding,
|
|
232
251
|
mapSandboxEvent,
|
|
233
252
|
mapSandboxToolEvent,
|
|
253
|
+
modelAuthoredChecks,
|
|
234
254
|
naiveDriver,
|
|
235
255
|
observe,
|
|
256
|
+
officialChecksFromMeta,
|
|
236
257
|
openSandboxRun,
|
|
237
258
|
pairwiseSignificance,
|
|
238
259
|
panel,
|
|
@@ -257,6 +278,7 @@ export {
|
|
|
257
278
|
renderReport,
|
|
258
279
|
reportLoopUsage,
|
|
259
280
|
resolveAgentEnvironmentProvider,
|
|
281
|
+
resolveEntrySymbol,
|
|
260
282
|
resolveSandboxClient,
|
|
261
283
|
routerBrain,
|
|
262
284
|
routerChatWithTools,
|
|
@@ -270,8 +292,10 @@ export {
|
|
|
270
292
|
runStrategyEvolution,
|
|
271
293
|
sample,
|
|
272
294
|
sampleThenRefine,
|
|
295
|
+
sandboxCheckRunner,
|
|
273
296
|
sandboxClientAsProvider,
|
|
274
297
|
sandboxSessionTraceSource,
|
|
298
|
+
selectBestIndex,
|
|
275
299
|
selectChampion,
|
|
276
300
|
selectValidWinner,
|
|
277
301
|
sentinelCompletion,
|
|
@@ -281,6 +305,7 @@ export {
|
|
|
281
305
|
stopSentinel,
|
|
282
306
|
strategyAuthorContract,
|
|
283
307
|
streamAgentTurn,
|
|
308
|
+
structuralRollout,
|
|
284
309
|
sumSandboxUsage,
|
|
285
310
|
supervise,
|
|
286
311
|
superviseSurface,
|
|
@@ -288,6 +313,7 @@ export {
|
|
|
288
313
|
supervisorInstructions,
|
|
289
314
|
trajectoryReport,
|
|
290
315
|
verify,
|
|
316
|
+
visibleCheckScore,
|
|
291
317
|
watchTrace,
|
|
292
318
|
widen,
|
|
293
319
|
workerFromBackend,
|