@tangle-network/agent-runtime 0.44.0 → 0.46.0

package/dist/{kb-gate-D0ZIhFOU.d.ts → kb-gate-2Gwpz_27.d.ts}

@@ -1,12 +1,13 @@
-import { CoderOutput, CoderTask } from './profiles.js';
-import { a as LoopSandboxClient, d as LoopTraceEmitter } from './types-DbJzz2uf.js';
+import { C as CoderOutput, a as CoderTask } from './coder-DCWFQpmJ.js';
+import { S as SandboxClient, e as LoopTraceEmitter } from './types-DnYoHvvZ.js';
 import { SandboxInstance } from '@tangle-network/sandbox';
+import { a as UiLens, U as UiFinding } from './substrate-CUgk7F7s.js';
 
 /**
  * @experimental
  *
  * Delegation executors — the layer between MCP delegates and the sandbox
- * substrate. Each executor exposes a {@link LoopSandboxClient} the kernel
+ * substrate. Each executor exposes a {@link SandboxClient} the kernel
  * consumes plus a placement tag so the trace pipeline can correlate workers
  * with their physical placement.
  *
@@ -25,13 +26,13 @@ import { SandboxInstance } from '@tangle-network/sandbox';
 /** @experimental */
 interface DelegationExecutor {
     /** Sandbox client the kernel calls. Returned with `describePlacement` set. */
-    readonly client: LoopSandboxClient;
+    readonly client: SandboxClient;
     /** Best-effort one-liner used in stderr boot logs and diagnostics. */
     describe(): string;
 }
 /** @experimental */
 interface SiblingSandboxExecutorOptions {
-    client: LoopSandboxClient;
+    client: SandboxClient;
 }
 /**
  * Wrap a raw sandbox SDK client so the kernel emits
@@ -102,7 +103,7 @@ declare function createFleetWorkspaceExecutor(options: FleetWorkspaceExecutorOpt
  */
 
 /** @experimental */
-type DelegationProfile = 'coder' | 'researcher';
+type DelegationProfile = 'coder' | 'researcher' | 'ui-auditor';
 /** @experimental */
 type DelegationStatus = 'pending' | 'running' | 'completed' | 'failed' | 'cancelled';
 /**
@@ -223,7 +224,73 @@ type DelegationResultPayload = {
 } | {
     profile: 'researcher';
     output: ResearchOutputShape;
+} | {
+    profile: 'ui-auditor';
+    output: UiAuditorDelegationOutput;
 };
+/**
+ * Wire-shape of a completed UI-audit delegation. The `findings` array
+ * contains every finding persisted to the workspace during the run,
+ * already enriched with `id` and `createdAt` by the writer. `workspaceDir`
+ * is the absolute path to the workspace; `indexFile` is the workspace-
+ * relative path to the regenerated index.md.
+ *
+ * @experimental
+ */
+interface UiAuditorDelegationOutput {
+    workspaceDir: string;
+    indexFile: string;
+    findings: UiFinding[];
+    /** Total iterations the loop ran for this delegation. */
+    iterations: number;
+}
+/** @experimental */
+type UiAuditLensFilter = readonly UiLens[];
+/** Optional per-route capture spec the agent surfaces over the wire. */
+interface DelegateUiAuditRoute {
+    /** Stable route name (used in screenshot filenames + finding metadata). */
+    name: string;
+    /** Fully-qualified URL. */
+    url: string;
+    /** Viewports to capture at. Defaults to `[{ width: 1280, height: 800 }]`. */
+    viewports?: readonly {
+        width: number;
+        height: number;
+    }[];
+    /** Default false. Full-page captures for the broad lenses. */
+    fullPage?: boolean;
+    /** Selector to wait for before capture. */
+    waitFor?: string;
+}
+/** @experimental */
+interface DelegateUiAuditConfig {
+    /**
+     * Lenses to iterate. Default: every lens except `'other'`. Order is
+     * preserved — the driver iterates lens-by-lens.
+     */
+    lenses?: UiAuditLensFilter;
+    /** Maximum total iterations across all (lens × route) pairs. Default 33 (11 lenses × 3 routes). */
+    maxIterations?: number;
+    /** Maximum concurrent iterations within a single plan() round. Default 2. */
+    maxConcurrency?: number;
+    /** Free-form product context surfaced to the judge. */
+    productContext?: string;
+}
+/** @experimental */
+interface DelegateUiAuditArgs {
+    /** Workspace root for the audit (absolute path). */
+    workspaceDir: string;
+    /** Routes to audit. Must be non-empty. */
+    routes: readonly DelegateUiAuditRoute[];
+    /** Multi-tenant scope. */
+    namespace?: string;
+    config?: DelegateUiAuditConfig;
+}
+/** @experimental */
+interface DelegateUiAuditResult {
+    taskId: string;
+    estimatedDurationMs?: number;
+}
 /**
  * Loose shape of a research output over the wire — the substrate cannot
  * import the `ResearchOutput` type from agent-knowledge without inducing
@@ -274,7 +341,7 @@ interface DelegationHistoryEntry {
     taskId: string;
     profile: DelegationProfile;
     namespace?: string;
-    args: DelegateCodeArgs | DelegateResearchArgs;
+    args: DelegateCodeArgs | DelegateResearchArgs | DelegateUiAuditArgs;
     status: DelegationStatus;
     feedback?: DelegationFeedbackSnapshot[];
     costUsd?: number;
@@ -295,6 +362,16 @@ interface DelegateRunCtx {
 type CoderDelegate = (args: DelegateCodeArgs, ctx: DelegateRunCtx) => Promise<CoderOutput>;
 /** @experimental */
 type ResearcherDelegate = (args: DelegateResearchArgs, ctx: DelegateRunCtx) => Promise<ResearchOutputShape>;
+/**
+ * UI-auditor delegate — fully consumer-injected. agent-runtime ships no
+ * default factory because the inputs are workspace path + judge function
+ * + (optionally) a `SandboxClient`, and the judge is the consumer's
+ * model seam. See `createInProcessUiAuditClient` + `uiAuditorProfile` in
+ * `@tangle-network/agent-runtime/profiles` for the canonical wiring.
+ *
+ * @experimental
+ */
+type UiAuditorDelegate = (args: DelegateUiAuditArgs, ctx: DelegateRunCtx) => Promise<UiAuditorDelegationOutput>;
 /** @experimental Structured review verdict over a coder candidate. */
 interface CoderReview {
     /** Gate: only approved candidates are eligible to win. */
@@ -336,7 +413,7 @@ interface CreateDefaultCoderDelegateOptions {
      * Convenience shorthand for sibling placement. Equivalent to
      * `executor: createSiblingSandboxExecutor({ client: sandboxClient })`.
      */
-    sandboxClient?: LoopSandboxClient;
+    sandboxClient?: SandboxClient;
     /** Default `['claude-code', 'codex', 'opencode/zai-coding-plan/glm-5.1']` when variants > 1. */
     fanoutHarnesses?: string[];
     /** Hard cap on the kernel's per-batch concurrency. Default 4. */
@@ -443,4 +520,4 @@ interface CreateKbGateOptions {
  */
 declare function createKbGate(options?: CreateKbGateOptions): (candidate: FactCandidate) => Promise<KbGateResult>;
 
-export { type DelegateRunCtx as A, type FactJudge as B, type CoderReviewer as C, type DelegateCodeArgs as D, type FactJudgeVerdict as E, type FactCandidate as F, type FeedbackRating as G, type FeedbackRefersTo as H, type FleetWorkspaceExecutorOptions as I, type ResearchOutputShape as J, type KbGateResult as K, createDefaultCoderDelegate as L, createFleetWorkspaceExecutor as M, createKbGate as N, createSiblingSandboxExecutor as O, type ResearcherDelegate as R, type SiblingSandboxExecutorOptions as S, type CoderWinnerSelection as a, type CreateKbGateOptions as b, type FleetHandle as c, type DelegationExecutor as d, type DelegateFeedbackArgs as e, type DelegationFeedbackSnapshot as f, type DelegationProfile as g, type DelegateResearchArgs as h, type DelegationStatus as i, type DelegationProgress as j, type DelegationResultPayload as k, type DelegationError as l, type DelegationStatusResult as m, type DelegationHistoryArgs as n, type DelegationHistoryEntry as o, type CoderDelegate as p, type DelegateCodeResult as q, type DelegateFeedbackResult as r, type ResearchSource as s, type DelegateResearchResult as t, type DelegationHistoryResult as u, type DelegationStatusArgs as v, type CoderReview as w, type CreateDefaultCoderDelegateOptions as x, type DelegateCodeConfig as y, type DelegateResearchConfig as z };
+export { type DelegateCodeConfig as A, type DelegateResearchConfig as B, type CoderReviewer as C, type DelegateCodeArgs as D, type DelegateRunCtx as E, type FactCandidate as F, type DelegateUiAuditConfig as G, type DelegateUiAuditRoute as H, type FactJudge as I, type FactJudgeVerdict as J, type FeedbackRating as K, type FeedbackRefersTo as L, type FleetWorkspaceExecutorOptions as M, type KbGateResult as N, type ResearchOutputShape as O, type UiAuditorDelegationOutput as P, createDefaultCoderDelegate as Q, type ResearcherDelegate as R, type SiblingSandboxExecutorOptions as S, createFleetWorkspaceExecutor as T, type UiAuditorDelegate as U, createKbGate as V, createSiblingSandboxExecutor as W, type CoderWinnerSelection as a, type CreateKbGateOptions as b, type FleetHandle as c, type DelegationExecutor as d, type DelegateFeedbackArgs as e, type DelegationFeedbackSnapshot as f, type DelegationProfile as g, type DelegateResearchArgs as h, type DelegateUiAuditArgs as i, type DelegationStatus as j, type DelegationProgress as k, type DelegationResultPayload as l, type DelegationError as m, type DelegationStatusResult as n, type DelegationHistoryArgs as o, type DelegationHistoryEntry as p, type CoderDelegate as q, type DelegateCodeResult as r, type DelegateFeedbackResult as s, type ResearchSource as t, type DelegateResearchResult as u, type DelegateUiAuditResult as v, type DelegationHistoryResult as w, type DelegationStatusArgs as x, type CoderReview as y, type CreateDefaultCoderDelegateOptions as z };

package/dist/{loop-runner-bin-BLMa8He3.d.ts → loop-runner-bin-D-K6bRp3.d.ts}

@@ -1,10 +1,10 @@
 import { Scenario } from '@tangle-network/agent-eval/campaign';
+import { SelfImproveOptions, SelfImproveResult } from '@tangle-network/agent-eval/contract';
 import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-p8dWBIXL.js';
-import { O as OptimizePromptOptions, a as OptimizePromptResult } from './optimize-prompt-D-urF2wW.js';
-import { T as TopologyPlanner, D as DynamicDecision } from './dynamic-wUgp6UKs.js';
-import { a as LoopSandboxClient, O as OutputAdapter, V as Validator, A as AgentRunSpec, b as LoopResult } from './types-DbJzz2uf.js';
-import { D as DelegateCodeArgs, C as CoderReviewer, a as CoderWinnerSelection, F as FactCandidate, b as CreateKbGateOptions } from './kb-gate-D0ZIhFOU.js';
-import { CoderOutput } from './profiles.js';
+import { D as DelegateCodeArgs, C as CoderReviewer, a as CoderWinnerSelection, F as FactCandidate, b as CreateKbGateOptions } from './kb-gate-2Gwpz_27.js';
+import { C as CoderOutput } from './coder-DCWFQpmJ.js';
+import { S as SandboxClient, O as OutputAdapter, V as Validator, A as AgentRunSpec, b as LoopResult } from './types-DnYoHvvZ.js';
+import { T as TopologyPlanner, C as CreateDriverOptions, D as DriverDecision } from './driver-C-mtBo7h.js';
 
 /**
  * @experimental
@@ -19,8 +19,8 @@ import { CoderOutput } from './profiles.js';
  *   review       → code mode with a REQUIRED reviewer (the gate is the point)
  *   research     → research-in-a-loop with valid-only KB growth (createKbGate)
  *   audit        → analyze trace/run data → findings (runAnalystLoop, caller-wired)
- *   self-improve → identity-gated prompt optimization (optimizePrompt, caller-wired)
- *   dynamic      → agent-authored topology (runLoop + createDynamicDriver)
+ *   self-improve → closed-loop text/config optimization (selfImprove, held-out gated)
+ *   dynamic      → agent-authored topology (runLoop + createDriver)
  *
  * It is intentionally a thin façade: the value is that EVERY product reuses the
  * one hardened engine instead of forking delegation logic. The dispatcher owns
@@ -67,7 +67,7 @@ interface RunDelegatedLoopOptions {
 declare function runDelegatedLoop<T = unknown>(mode: DelegatedLoopMode, registry: DelegatedLoopRegistry, options?: RunDelegatedLoopOptions): Promise<DelegatedLoopResult<T>>;
 /** @experimental Options for the default `code`/`review` runner. */
 interface CoderLoopRunnerOptions {
-    sandboxClient: LoopSandboxClient;
+    sandboxClient: SandboxClient;
     /** What to build — the delegate args (goal, repoRoot, variants, config, …). */
     args: DelegateCodeArgs;
     /** Adversarial reviewer. REQUIRED for `review` mode (see `reviewLoopRunner`). */
@@ -90,8 +90,8 @@ declare function reviewLoopRunner(options: CoderLoopRunnerOptions & {
 }): DelegatedLoopRunner<CoderOutput>;
 /** @experimental Options for the default `dynamic` runner. */
 interface DynamicLoopRunnerOptions<Task, Output> {
-    sandboxClient: LoopSandboxClient;
-    /** The agent-authored topology planner (e.g. `createSandboxPlanner(...)`). */
+    sandboxClient: SandboxClient;
+    /** The agent-authored topology planner (sync or async; an async planner is where an LLM call goes). */
     planner: TopologyPlanner<Task, Output>;
     task: Task;
     output: OutputAdapter<Output>;
@@ -101,9 +101,13 @@ interface DynamicLoopRunnerOptions<Task, Output> {
     agentRuns?: AgentRunSpec<Task>[];
     maxIterations?: number;
     maxFanout?: number;
+    /** Optional trace-analyst hook forwarded to the dynamic driver so the loop runs
+     *  `f(trace, findings)` — see `CreateDriverOptions.analyze`. Caller-side
+     *  seam to `runAnalystLoop`; keeps this runner analyst-free. */
+    analyze?: CreateDriverOptions<Task, Output>['analyze'];
 }
 /** @experimental `dynamic` mode — agent-authored topology over `runLoop`. */
-declare function dynamicLoopRunner<Task, Output>(o: DynamicLoopRunnerOptions<Task, Output>): DelegatedLoopRunner<LoopResult<Task, Output, DynamicDecision>>;
+declare function dynamicLoopRunner<Task, Output>(o: DynamicLoopRunnerOptions<Task, Output>): DelegatedLoopRunner<LoopResult<Task, Output, DriverDecision>>;
 /** @experimental A fact rejected at the KB gate — surfaced, never dropped. */
 interface VetoedFact {
     candidate: FactCandidate;
@@ -142,8 +146,8 @@ interface ResearchLoopRunnerOptions {
  * never silently dropped) so the caller audits vs retries.
  */
 declare function researchLoopRunner(o: ResearchLoopRunnerOptions): DelegatedLoopRunner<ResearchLoopResult>;
-/** @experimental `self-improve` mode — identity-gated prompt optimization. */
-declare function selfImproveLoopRunner<TScenario extends Scenario, TArtifact>(options: OptimizePromptOptions<TScenario, TArtifact>): DelegatedLoopRunner<OptimizePromptResult<TArtifact, TScenario>>;
+/** @experimental `self-improve` mode — agent-eval's one-call closed loop (held-out gated). */
+declare function selfImproveLoopRunner<TScenario extends Scenario, TArtifact>(options: SelfImproveOptions<TScenario, TArtifact>): DelegatedLoopRunner<SelfImproveResult<TScenario, TArtifact>>;
 /** @experimental `audit` mode — analyst loop over captured trace/run data. */
 declare function auditLoopRunner<TProposal = unknown, TEdit = unknown>(options: RunAnalystLoopOpts): DelegatedLoopRunner<RunAnalystLoopResult<TProposal, TEdit>>;
 

package/dist/loop-runner-bin.d.ts

@@ -1,11 +1,13 @@
 #!/usr/bin/env node
-export { L as LoopRunnerCliArgs, f as LoopRunnerCliResult, p as parseLoopRunnerArgv, o as runLoopRunnerCli } from './loop-runner-bin-BLMa8He3.js';
+export { L as LoopRunnerCliArgs, f as LoopRunnerCliResult, p as parseLoopRunnerArgv, o as runLoopRunnerCli } from './loop-runner-bin-D-K6bRp3.js';
 import '@tangle-network/agent-eval/campaign';
+import '@tangle-network/agent-eval/contract';
 import './types-p8dWBIXL.js';
 import '@tangle-network/agent-eval';
-import './optimize-prompt-D-urF2wW.js';
-import './dynamic-wUgp6UKs.js';
-import './types-DbJzz2uf.js';
+import './kb-gate-2Gwpz_27.js';
+import './coder-DCWFQpmJ.js';
 import '@tangle-network/sandbox';
-import './kb-gate-D0ZIhFOU.js';
-import './profiles.js';
+import './types-DnYoHvvZ.js';
+import './runtime-hooks-C7JwKb9E.js';
+import './substrate-CUgk7F7s.js';
+import './driver-C-mtBo7h.js';

package/dist/loop-runner-bin.js

@@ -2,15 +2,13 @@
 import {
   parseLoopRunnerArgv,
   runLoopRunnerCli
-} from "./chunk-GFKVVRQ7.js";
+} from "./chunk-GN75RGM6.js";
 import "./chunk-FNMGYYSS.js";
-import "./chunk-SKUZZCHE.js";
-import "./chunk-XBUG326M.js";
-import "./chunk-VOX6Z3II.js";
-import "./chunk-S7JXV32P.js";
-import "./chunk-3HMHSN22.js";
-import "./chunk-PY6NMZYX.js";
-import "./chunk-SQSCRJ7U.js";
+import "./chunk-I42NHLKX.js";
+import "./chunk-KADIJAD4.js";
+import "./chunk-HNUXAZIJ.js";
+import "./chunk-65FQLI4V.js";
+import "./chunk-GSUO5QS6.js";
 import "./chunk-DGUM43GV.js";
 export {
   parseLoopRunnerArgv,