@tangle-network/agent-runtime 0.46.0 → 0.47.0

package/dist/agent.d.ts

@@ -1,6 +1,6 @@
 import * as _tangle_network_agent_eval from '@tangle-network/agent-eval';
 import { TraceAnalystKindSpec, AnalystFinding } from '@tangle-network/agent-eval';
-import { R as RuntimeStreamEvent, S as SandboxClient, O as OutputAdapter, A as AgentRunSpec } from './types-DnYoHvvZ.js';
+import { R as RuntimeStreamEvent, S as SandboxClient, O as OutputAdapter, A as AgentRunSpec } from './types-nBMuollC.js';
 import { A as AgentSurfaces } from './improvement-adapter-BC4HhuAR.js';
 export { C as CreateSurfaceImprovementAdapterOpts, D as DraftPatchInput, a as DraftPatchOutput, R as ResolvedSurface, S as SurfaceImprovementEdit, b as SurfaceValidationIssue, c as createSurfaceImprovementAdapter, r as renderSurfaceIssues, d as resolveSubjectPath, v as validateSurfaces } from './improvement-adapter-BC4HhuAR.js';
 import { K as KnowledgeAdapter, a as RunAnalystLoopResult } from './types-p8dWBIXL.js';

package/dist/agent.js

@@ -3,7 +3,7 @@ import {
 } from "./chunk-7JITYN6T.js";
 import {
   createSandboxForSpec
-} from "./chunk-65FQLI4V.js";
+} from "./chunk-72JQCHOZ.js";
 import {
   mapSandboxEvent
 } from "./chunk-GSUO5QS6.js";

package/dist/analyst-loop.d.ts

@@ -1,5 +1,5 @@
 import { TraceAnalystByteBudgets, TraceAnalysisStore } from '@tangle-network/agent-eval';
-import { I as Iteration } from './types-DnYoHvvZ.js';
+import { I as Iteration } from './types-nBMuollC.js';
 import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-p8dWBIXL.js';
 export { A as AnalystLoopEvent, b as AnalystRegistryLike, c as AnalystRegistryStreamingLike, d as AutoApplyPolicy, F as FindingsStoreLike, I as ImprovementAdapter, e as ImprovementEditBatch, f as ImprovementReport, K as KnowledgeAdapter, g as KnowledgeProposalBatch, h as KnowledgeReport } from './types-p8dWBIXL.js';
 import '@tangle-network/sandbox';

package/dist/{chunk-I42NHLKX.js → chunk-5YDS7BLC.js}

@@ -4,7 +4,7 @@ import {
 } from "./chunk-KADIJAD4.js";
 import {
   runLoop
-} from "./chunk-65FQLI4V.js";
+} from "./chunk-72JQCHOZ.js";
 
 // src/mcp/executor.ts
 function createSiblingSandboxExecutor(options) {
@@ -91,7 +91,11 @@ function createDefaultCoderDelegate(options) {
     const variants = Math.max(1, Math.trunc(args.variants ?? 1));
     ctx.report({ iteration: 0, phase: "starting" });
     if (variants <= 1) {
-      const { agentRunSpec, output, validator } = coderProfile({ task });
+      const { agentRunSpec, output, validator } = coderProfile({
+        task,
+        ...options.harness ? { harness: options.harness } : {},
+        ...options.model ? { model: options.model } : {}
+      });
       const result2 = await runLoop({
         driver: singleShotDriver,
         agentRun: agentRunSpec,
@@ -113,9 +117,10 @@ function createDefaultCoderDelegate(options) {
       ctx.report({ iteration: 1, phase: "completed" });
       return chosen2;
     }
-    const fanout = multiHarnessCoderFanout(
-      fanoutHarnesses && fanoutHarnesses.length > 0 ? { harnesses: fanoutHarnesses.slice(0, variants) } : { harnesses: void 0 }
-    );
+    const fanout = multiHarnessCoderFanout({
+      ...fanoutHarnesses && fanoutHarnesses.length > 0 ? { harnesses: fanoutHarnesses.slice(0, variants) } : {},
+      ...options.fanoutModels ? { models: options.fanoutModels.slice(0, variants) } : {}
+    });
     const agentRuns = fanout.agentRuns.slice(0, variants);
     const result = await runLoop({
       driver: fanout.driver,
@@ -210,4 +215,4 @@ export {
   createFleetWorkspaceExecutor,
   createDefaultCoderDelegate
 };
-//# sourceMappingURL=chunk-I42NHLKX.js.map
+//# sourceMappingURL=chunk-5YDS7BLC.js.map

package/dist/chunk-5YDS7BLC.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/mcp/executor.ts","../src/mcp/delegates.ts"],"sourcesContent":["/**\n * @experimental\n *\n * Delegation executors — the layer between MCP delegates and the sandbox\n * substrate. Each executor exposes a {@link SandboxClient} the kernel\n * consumes plus a placement tag so the trace pipeline can correlate workers\n * with their physical placement.\n *\n * Two implementations ship in-box:\n *\n * - {@link createSiblingSandboxExecutor} — every delegation spawns a fresh\n *   sandbox sibling to the caller. Default when the MCP server runs as a\n *   standalone CLI mounted outside a fleet.\n *\n * - {@link createFleetWorkspaceExecutor} — delegations dispatch onto machines\n *   in the caller's existing fleet so worker diffs land directly on the\n *   caller's filesystem (the fleet's shared workspace). Selected when the\n *   parent sandbox passes `TANGLE_FLEET_ID` into the MCP server's env.\n */\n\nimport type { CreateSandboxOptions, SandboxInstance } from '@tangle-network/sandbox'\nimport type { LoopSandboxPlacement, SandboxClient } from '../runtime'\n\n/** @experimental */\nexport interface DelegationExecutor {\n  /** Sandbox client the kernel calls. Returned with `describePlacement` set. */\n  readonly client: SandboxClient\n  /** Best-effort one-liner used in stderr boot logs and diagnostics. */\n  describe(): string\n}\n\n/** @experimental */\nexport interface SiblingSandboxExecutorOptions {\n  client: SandboxClient\n}\n\n/**\n * Wrap a raw sandbox SDK client so the kernel emits\n * `loop.iteration.dispatch` events with `{ placement: 'sibling', sandboxId }`.\n *\n * The returned client `.create()` delegates to the underlying client; the\n * only added behavior is a `describePlacement` tag the kernel reads.\n *\n * @experimental\n */\nexport function createSiblingSandboxExecutor(\n  options: SiblingSandboxExecutorOptions,\n): DelegationExecutor {\n  const underlying = options.client\n  const client: SandboxClient = {\n    create(opts?: CreateSandboxOptions): Promise<SandboxInstance> {\n      return underlying.create(opts)\n    },\n    describePlacement(box: SandboxInstance): LoopSandboxPlacement {\n      return { kind: 'sibling', sandboxId: readId(box) }\n    },\n  }\n  return {\n    client,\n    describe(): string {\n      return 'sibling-sandbox (each delegation = fresh sandbox via client.create)'\n    },\n  }\n}\n\n/**\n * Minimal `SandboxFleet` surface the fleet executor calls. Declared\n * structurally so tests can pass an in-memory stub without instantiating the\n * sandbox SDK.\n *\n * @experimental\n */\nexport interface FleetHandle {\n  readonly fleetId: string\n  /** Machine ids in dispatch-eligible order. The executor round-robins. */\n  readonly ids: ReadonlyArray<string>\n  /** Resolve a machine id to its `SandboxInstance` — that machine is mounted\n   * on the fleet's shared workspace, so any diff the worker writes lands on\n   * every other fleet machine's filesystem too. */\n  sandbox(machineId: string): Promise<SandboxInstance>\n}\n\n/** @experimental */\nexport interface FleetWorkspaceExecutorOptions {\n  fleet: FleetHandle\n  /**\n   * Override the machine-selection policy. Default = round-robin across\n   * `fleet.ids`, skipping the optional `excludeMachineIds` set (typically the\n   * coordinator machine the MCP server is running on).\n   */\n  selectMachine?: (call: { callIndex: number; ids: ReadonlyArray<string> }) => string\n  /**\n   * Machine ids to skip during default round-robin. Set to the caller's own\n   * machineId so workers don't compete with the orchestrator on the same VM.\n   */\n  excludeMachineIds?: ReadonlyArray<string>\n}\n\n/**\n * Build an executor that resolves each delegated iteration to an existing\n * machine in `fleet`. The fleet's shared-workspace policy means the worker\n * machine sees the caller's filesystem — diffs land in-place with no\n * cross-sandbox copy step.\n *\n * @experimental\n */\nexport function createFleetWorkspaceExecutor(\n  options: FleetWorkspaceExecutorOptions,\n): DelegationExecutor {\n  const fleet = options.fleet\n  const exclude = new Set(options.excludeMachineIds ?? [])\n  let callIndex = 0\n  // machineId-by-sandboxId, populated as we resolve machines so\n  // `describePlacement` can recover the assignment from the SandboxInstance\n  // the kernel hands back.\n  const placementBySandboxId = new Map<string, { machineId: string }>()\n\n  const client: SandboxClient = {\n    async create(): Promise<SandboxInstance> {\n      const ids = fleet.ids.filter((id) => !exclude.has(id))\n      if (ids.length === 0) {\n        throw new Error(\n          `agent-runtime: fleet ${fleet.fleetId} has no eligible worker machines (ids=[${fleet.ids.join(',')}], excluded=[${[...exclude].join(',')}])`,\n        )\n      }\n      const selector = options.selectMachine\n      const machineId = selector ? selector({ callIndex, ids }) : ids[callIndex % ids.length]\n      callIndex += 1\n      if (typeof machineId !== 'string' || machineId.length === 0) {\n        throw new Error('agent-runtime: fleet executor selectMachine returned an empty machine id')\n      }\n      const box = await fleet.sandbox(machineId)\n      const sandboxId = readId(box)\n      if (sandboxId) placementBySandboxId.set(sandboxId, { machineId })\n      return box\n    },\n    describePlacement(box: SandboxInstance): LoopSandboxPlacement {\n      const sandboxId = readId(box)\n      const recorded = sandboxId ? placementBySandboxId.get(sandboxId) : undefined\n      return {\n        kind: 'fleet',\n        sandboxId,\n        fleetId: fleet.fleetId,\n        machineId: recorded?.machineId,\n      }\n    },\n  }\n\n  return {\n    client,\n    describe(): string {\n      const excluded = exclude.size > 0 ? ` (excluded=[${[...exclude].join(',')}])` : ''\n      return `fleet-workspace (fleetId=${fleet.fleetId}, machines=[${fleet.ids.join(',')}]${excluded})`\n    },\n  }\n}\n\nfunction readId(box: SandboxInstance): string | undefined {\n  const raw = (box as unknown as { id?: unknown }).id\n  return typeof raw === 'string' && raw.length > 0 ? raw : undefined\n}\n","/**\n * @experimental\n *\n * Delegate factories — the layer between MCP tool handlers and the\n * underlying `runLoop` runners.\n *\n * The MCP server is profile-agnostic: it owns the task queue + feedback\n * store + transport. Each `*Delegate` is the closure that the queue\n * invokes when a task runs. Consumers can override either delegate to\n * inject custom drivers, mocks, fleet-aware dispatchers, etc.\n *\n * The default coder delegate is wired here because we own\n * `coderProfile` / `multiHarnessCoderFanout`. The default researcher\n * delegate is **not** wired in this file — `agent-knowledge` cannot be\n * imported from `agent-runtime` without inducing a cycle. Consumers\n * pass `researcherDelegate` explicitly when constructing the server.\n */\n\nimport { type CoderOutput, coderProfile, multiHarnessCoderFanout } from '../profiles/coder'\nimport type { Iteration, LoopTraceEmitter, SandboxClient } from '../runtime'\nimport { runLoop } from '../runtime'\nimport { createSiblingSandboxExecutor, type DelegationExecutor } from './executor'\nimport type {\n  CoderTask,\n  DelegateCodeArgs,\n  DelegateResearchArgs,\n  DelegateUiAuditArgs,\n  DelegationProgress,\n  ResearchOutputShape,\n  UiAuditorDelegationOutput,\n} from './types'\n\n/** @experimental */\nexport interface DelegateRunCtx {\n  signal: AbortSignal\n  report(progress: DelegationProgress): void\n}\n\n/** @experimental */\nexport type CoderDelegate = (\n  args: DelegateCodeArgs,\n  ctx: DelegateRunCtx,\n) => Promise<import('../profiles/coder').CoderOutput>\n\n/** @experimental */\nexport type ResearcherDelegate = (\n  args: DelegateResearchArgs,\n  ctx: DelegateRunCtx,\n) => Promise<ResearchOutputShape>\n\n/**\n * UI-auditor delegate — fully consumer-injected. agent-runtime ships no\n * default factory because the inputs are workspace path + judge function\n * + (optionally) a `SandboxClient`, and the judge is the consumer's\n * model seam. See `createInProcessUiAuditClient` + `uiAuditorProfile` in\n * `@tangle-network/agent-runtime/profiles` for the canonical wiring.\n *\n * @experimental\n */\nexport type UiAuditorDelegate = (\n  args: DelegateUiAuditArgs,\n  ctx: DelegateRunCtx,\n) => Promise<UiAuditorDelegationOutput>\n\n/** @experimental Structured review verdict over a coder candidate. */\nexport interface CoderReview {\n  /** Gate: only approved candidates are eligible to win. */\n  approved: boolean\n  /** Reviewer's recommendation — surfaced in traces. */\n  recommendation: 'ship' | 'approve-with-nits' | 'changes-requested' | 'reject'\n  /** Readiness 0..1, used by the `highest-readiness` winner-selection strategy. */\n  readiness: number\n  notes?: string\n}\n\n/**\n * @experimental\n *\n * Optional adversarial reviewer over a coder candidate that already passed\n * mechanical validation (tests/typecheck/forbidden/diff/no-op/secrets). Folded\n * from the ai-trading-blueprint delegation MCP: a candidate is only eligible to\n * win if the reviewer approves it. The reviewer is the consumer's seam — an LLM\n * judge, a `pnpm review` command, anything returning a `CoderReview`.\n */\nexport type CoderReviewer = (\n  output: import('../profiles/coder').CoderOutput,\n  task: CoderTask,\n  ctx: { signal: AbortSignal },\n) => Promise<CoderReview> | CoderReview\n\n/**\n * @experimental Winner-selection strategy among validated (+ reviewed)\n * candidates. `highest-readiness` requires a `reviewer`. Default `highest-score`\n * (the kernel's behavior — preserves backward compatibility).\n */\nexport type CoderWinnerSelection =\n  | 'highest-score'\n  | 'smallest-diff'\n  | 'highest-readiness'\n  | 'first-approved'\n\n/** @experimental */\nexport interface CreateDefaultCoderDelegateOptions {\n  /**\n   * Execution placement. Pass a {@link DelegationExecutor} (sibling or fleet)\n   * to control where worker iterations land. `sandboxClient` is a\n   * convenience shorthand that wraps the client in a sibling executor — pass\n   * one or the other, not both.\n   */\n  executor?: DelegationExecutor\n  /**\n   * Convenience shorthand for sibling placement. Equivalent to\n   * `executor: createSiblingSandboxExecutor({ client: sandboxClient })`.\n   */\n  sandboxClient?: SandboxClient\n  /** Backend harness for the single-coder path. Default comes from `coderProfile`. */\n  harness?: string\n  /** Model override for the single-coder path. */\n  model?: string\n  /** Default `['claude-code', 'codex', 'opencode/zai-coding-plan/glm-5.1']` when variants > 1. */\n  fanoutHarnesses?: string[]\n  /** Optional per-harness model override for `variants > 1`. */\n  fanoutModels?: (string | undefined)[]\n  /** Hard cap on the kernel's per-batch concurrency. Default 4. */\n  maxConcurrency?: number\n  /**\n   * Optional adversarial reviewer. When set, a candidate must pass mechanical\n   * validation AND `reviewer.approved` to be eligible to win — empty/secret/\n   * test-failing patches are already gone; this catches the \"compiles + passes\n   * but wrong/unsafe\" class the deterministic validator can't see.\n   */\n  reviewer?: CoderReviewer\n  /** Winner-selection strategy among eligible candidates. Default `highest-score`. */\n  winnerSelection?: CoderWinnerSelection\n  /**\n   * Loop trace emitter forwarded into every delegated `runLoop`. Wire\n   * `createPropagatingTraceEmitter(readTraceContextFromEnv())` here (the bin\n   * does) so delegated build-loops export their topology spans to the OTLP /\n   * Tangle Intelligence sink when `OTEL_EXPORTER_OTLP_ENDPOINT` is set — and\n   * are a cheap no-op when it isn't. Configurable by construction.\n   */\n  traceEmitter?: LoopTraceEmitter\n}\n\n/**\n * Build a coder delegate that drives `runLoop` against the project's\n * sandbox client + coder profile. When `args.variants > 1` it switches\n * to the multi-harness fanout topology.\n *\n * @experimental\n */\nexport function createDefaultCoderDelegate(\n  options: CreateDefaultCoderDelegateOptions,\n): CoderDelegate {\n  const executor = resolveExecutor(options)\n  const sandboxClient = executor.client\n  const fanoutHarnesses = options.fanoutHarnesses\n  const maxConcurrency = options.maxConcurrency ?? 4\n  const traceEmitter = options.traceEmitter\n  return async (args, ctx) => {\n    const task: CoderTask = {\n      goal: buildCoderGoal(args),\n      repoRoot: args.repoRoot,\n      testCmd: args.config?.testCmd,\n      typecheckCmd: args.config?.typecheckCmd,\n      forbiddenPaths: args.config?.forbiddenPaths,\n      maxDiffLines: args.config?.maxDiffLines,\n    }\n    const variants = Math.max(1, Math.trunc(args.variants ?? 1))\n    ctx.report({ iteration: 0, phase: 'starting' })\n    if (variants <= 1) {\n      const { agentRunSpec, output, validator } = coderProfile({\n        task,\n        ...(options.harness ? { harness: options.harness } : {}),\n        ...(options.model ? { model: options.model } : {}),\n      })\n      const result = await runLoop({\n        driver: singleShotDriver,\n        agentRun: agentRunSpec,\n        output,\n        validator,\n        task,\n        ctx: { sandboxClient, signal: ctx.signal, ...(traceEmitter ? { traceEmitter } : {}) },\n        maxIterations: 1,\n        maxConcurrency,\n      })\n      const chosen = await pickCoderWinner({\n        iterations: result.iterations,\n        reviewer: options.reviewer,\n        selection: options.winnerSelection ?? 'highest-score',\n        task,\n        signal: ctx.signal,\n      })\n      if (!chosen) throw new Error(noWinnerMessage(options.reviewer))\n      ctx.report({ iteration: 1, phase: 'completed' })\n      return chosen\n    }\n    const fanout = multiHarnessCoderFanout({\n      ...(fanoutHarnesses && fanoutHarnesses.length > 0\n        ? { harnesses: fanoutHarnesses.slice(0, variants) }\n        : {}),\n      ...(options.fanoutModels ? { models: options.fanoutModels.slice(0, variants) } : {}),\n    })\n    const agentRuns = fanout.agentRuns.slice(0, variants)\n    const result = await runLoop({\n      driver: fanout.driver,\n      agentRuns,\n      output: fanout.output,\n      validator: fanout.validator,\n      task,\n      ctx: { sandboxClient, signal: ctx.signal, ...(traceEmitter ? { traceEmitter } : {}) },\n      maxIterations: variants,\n      maxConcurrency: Math.min(maxConcurrency, variants),\n    })\n    const chosen = await pickCoderWinner({\n      iterations: result.iterations,\n      reviewer: options.reviewer,\n      selection: options.winnerSelection ?? 'highest-score',\n      task,\n      signal: ctx.signal,\n    })\n    if (!chosen) throw new Error(noWinnerMessage(options.reviewer))\n    ctx.report({ iteration: agentRuns.length, phase: 'completed' })\n    return chosen\n  }\n}\n\ninterface PickCoderWinnerArgs {\n  iterations: ReadonlyArray<Iteration<CoderTask, CoderOutput>>\n  reviewer: CoderReviewer | undefined\n  selection: CoderWinnerSelection\n  task: CoderTask\n  signal: AbortSignal\n}\n\ninterface CoderCandidate {\n  index: number\n  output: CoderOutput\n  score: number\n  readiness: number\n}\n\n/**\n * Pick the winning coder candidate from a finished loop's iterations:\n *   1. keep only mechanically-VALID candidates (the validator already gated\n *      tests/typecheck/forbidden/diff/no-op/secrets),\n *   2. if a `reviewer` is wired, keep only those it APPROVES,\n *   3. select among survivors by the chosen strategy.\n * Returns `undefined` when nothing survives — the delegate fails loud.\n */\nasync function pickCoderWinner(args: PickCoderWinnerArgs): Promise<CoderOutput | undefined> {\n  const valid: CoderCandidate[] = []\n  for (const iter of args.iterations) {\n    if (iter.output === undefined || iter.error || iter.verdict?.valid !== true) continue\n    valid.push({\n      index: iter.index,\n      output: iter.output,\n      score: iter.verdict.score ?? 0,\n      readiness: iter.verdict.score ?? 0,\n    })\n  }\n  if (valid.length === 0) return undefined\n\n  let eligible = valid\n  if (args.reviewer) {\n    eligible = []\n    for (const c of valid) {\n      const review = await args.reviewer(c.output, args.task, { signal: args.signal })\n      if (review.approved) eligible.push({ ...c, readiness: review.readiness })\n    }\n    if (eligible.length === 0) return undefined\n  }\n\n  return selectCoderCandidate(eligible, args.selection).output\n}\n\n/** Apply the winner-selection strategy; ties broken by earliest iteration. */\nfunction selectCoderCandidate(\n  candidates: CoderCandidate[],\n  selection: CoderWinnerSelection,\n): CoderCandidate {\n  const diffLines = (c: CoderCandidate) =>\n    c.output.diffStats.insertions + c.output.diffStats.deletions\n  const sorted = [...candidates].sort((a, b) => {\n    switch (selection) {\n      case 'smallest-diff':\n        return diffLines(a) - diffLines(b) || a.index - b.index\n      case 'highest-readiness':\n        return b.readiness - a.readiness || a.index - b.index\n      case 'first-approved':\n        return a.index - b.index\n      default:\n        return b.score - a.score || a.index - b.index\n    }\n  })\n  return sorted[0]!\n}\n\nfunction noWinnerMessage(reviewer: CoderReviewer | undefined): string {\n  return reviewer\n    ? 'coder delegate: no candidate passed validation + review'\n    : 'coder delegate: no candidate passed validation'\n}\n\nfunction buildCoderGoal(args: DelegateCodeArgs): string {\n  if (!args.contextHint) return args.goal\n  return [args.goal, '', '## Context', args.contextHint].join('\\n')\n}\n\nfunction resolveExecutor(options: CreateDefaultCoderDelegateOptions): DelegationExecutor {\n  if (options.executor && options.sandboxClient) {\n    throw new Error('createDefaultCoderDelegate: pass exactly one of `executor` or `sandboxClient`')\n  }\n  if (options.executor) return options.executor\n  if (options.sandboxClient) {\n    return createSiblingSandboxExecutor({ client: options.sandboxClient })\n  }\n  throw new Error('createDefaultCoderDelegate: `executor` or `sandboxClient` is required')\n}\n\n/**\n * Single-shot driver — plan one task on iteration 0, stop after one\n * iteration. Used by the coder delegate when `variants <= 1`. Keeps the\n * runLoop kernel-level accounting (timing, cost, trace emission) while\n * skipping fanout/refine topology overhead.\n */\nconst singleShotDriver = {\n  name: 'mcp-single-shot',\n  async plan<Task>(task: Task, history: ReadonlyArray<unknown>): Promise<Task[]> {\n    return history.length === 0 ? [task] : []\n  },\n  decide(history: ReadonlyArray<unknown>): 'pick-winner' | 'fail' {\n    return history.length > 0 ? 'pick-winner' : 'fail'\n  },\n}\n"],"mappings":";;;;;;;;;AA6CO,SAAS,6BACd,SACoB;AACpB,QAAM,aAAa,QAAQ;AAC3B,QAAM,SAAwB;AAAA,IAC5B,OAAO,MAAuD;AAC5D,aAAO,WAAW,OAAO,IAAI;AAAA,IAC/B;AAAA,IACA,kBAAkB,KAA4C;AAC5D,aAAO,EAAE,MAAM,WAAW,WAAW,OAAO,GAAG,EAAE;AAAA,IACnD;AAAA,EACF;AACA,SAAO;AAAA,IACL;AAAA,IACA,WAAmB;AACjB,aAAO;AAAA,IACT;AAAA,EACF;AACF;AA2CO,SAAS,6BACd,SACoB;AACpB,QAAM,QAAQ,QAAQ;AACtB,QAAM,UAAU,IAAI,IAAI,QAAQ,qBAAqB,CAAC,CAAC;AACvD,MAAI,YAAY;AAIhB,QAAM,uBAAuB,oBAAI,IAAmC;AAEpE,QAAM,SAAwB;AAAA,IAC5B,MAAM,SAAmC;AACvC,YAAM,MAAM,MAAM,IAAI,OAAO,CAAC,OAAO,CAAC,QAAQ,IAAI,EAAE,CAAC;AACrD,UAAI,IAAI,WAAW,GAAG;AACpB,cAAM,IAAI;AAAA,UACR,wBAAwB,MAAM,OAAO,0CAA0C,MAAM,IAAI,KAAK,GAAG,CAAC,gBAAgB,CAAC,GAAG,OAAO,EAAE,KAAK,GAAG,CAAC;AAAA,QAC1I;AAAA,MACF;AACA,YAAM,WAAW,QAAQ;AACzB,YAAM,YAAY,WAAW,SAAS,EAAE,WAAW,IAAI,CAAC,IAAI,IAAI,YAAY,IAAI,MAAM;AACtF,mBAAa;AACb,UAAI,OAAO,cAAc,YAAY,UAAU,WAAW,GAAG;AAC3D,cAAM,IAAI,MAAM,0EAA0E;AAAA,MAC5F;AACA,YAAM,MAAM,MAAM,MAAM,QAAQ,SAAS;AACzC,YAAM,YAAY,OAAO,GAAG;AAC5B,UAAI,UAAW,sBAAqB,IAAI,WAAW,EAAE,UAAU,CAAC;AAChE,aAAO;AAAA,IACT;AAAA,IACA,kBAAkB,KAA4C;AAC5D,YAAM,YAAY,OAAO,GAAG;AAC5B,YAAM,WAAW,YAAY,qBAAqB,IAAI,SAAS,IAAI;AACnE,aAAO;AAAA,QACL,MAAM;AAAA,QACN;AAAA,QACA,SAAS,MAAM;AAAA,QACf,WAAW,UAAU;AAAA,MACvB;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AAAA,IACL;AAAA,IACA,WAAmB;AACjB,YAAM,WAAW,QAAQ,OAAO,IAAI,eAAe,CAAC,GAAG,OAAO,EAAE,KAAK,GAAG,CAAC,OAAO;AAChF,aAAO,4BAA4B,MAAM,OAAO,eAAe,MAAM,IAAI,KAAK,GAAG,CAAC,IAAI,QAAQ;AAAA,IAChG;AAAA,EACF;AACF;AAEA,SAAS,OAAO,KAA0C;AACxD,QAAM,MAAO,IAAoC;AACjD,SAAO,OAAO,QAAQ,YAAY,IAAI,SAAS,IAAI,MAAM;AAC3D;;;ACTO,SAAS,2BACd,SACe;AACf,QAAM,WAAW,gBAAgB,OAAO;AACxC,QAAM,gBAAgB,SAAS;AAC/B,QAAM,kBAAkB,QAAQ;AAChC,QAAM,iBAAiB,QAAQ,kBAAkB;AACjD,QAAM,eAAe,QAAQ;AAC7B,SAAO,OAAO,MAAM,QAAQ;AAC1B,UAAM,OAAkB;AAAA,MACtB,MAAM,eAAe,IAAI;AAAA,MACzB,UAAU,KAAK;AAAA,MACf,SAAS,KAAK,QAAQ;AAAA,MACtB,cAAc,KAAK,QAAQ;AAAA,MAC3B,gBAAgB,KAAK,QAAQ;AAAA,MAC7B,cAAc,KAAK,QAAQ;AAAA,IAC7B;AACA,UAAM,WAAW,KAAK,IAAI,GAAG,KAAK,MAAM,KAAK,YAAY,CAAC,CAAC;AAC3D,QAAI,OAAO,EAAE,WAAW,GAAG,OAAO,WAAW,CAAC;AAC9C,QAAI,YAAY,GAAG;AACjB,YAAM,EAAE,cAAc,QAAQ,UAAU,IAAI,aAAa;AAAA,QACvD;AAAA,QACA,GAAI,QAAQ,UAAU,EAAE,SAAS,QAAQ,QAAQ,IAAI,CAAC;AAAA,QACtD,GAAI,QAAQ,QAAQ,EAAE,OAAO,QAAQ,MAAM,IAAI,CAAC;AAAA,MAClD,CAAC;AACD,YAAMA,UAAS,MAAM,QAAQ;AAAA,QAC3B,QAAQ;AAAA,QACR,UAAU;AAAA,QACV;AAAA,QACA;AAAA,QACA;AAAA,QACA,KAAK,EAAE,eAAe,QAAQ,IAAI,QAAQ,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC,EAAG;AAAA,QACpF,eAAe;AAAA,QACf;AAAA,MACF,CAAC;AACD,YAAMC,UAAS,MAAM,gBAAgB;AAAA,QACnC,YAAYD,QAAO;AAAA,QACnB,UAAU,QAAQ;AAAA,QAClB,WAAW,QAAQ,mBAAmB;AAAA,QACtC;AAAA,QACA,QAAQ,IAAI;AAAA,MACd,CAAC;AACD,UAAI,CAACC,QAAQ,OAAM,IAAI,MAAM,gBAAgB,QAAQ,QAAQ,CAAC;AAC9D,UAAI,OAAO,EAAE,WAAW,GAAG,OAAO,YAAY,CAAC;AAC/C,aAAOA;AAAA,IACT;AACA,UAAM,SAAS,wBAAwB;AAAA,MACrC,GAAI,mBAAmB,gBAAgB,SAAS,IAC5C,EAAE,WAAW,gBAAgB,MAAM,GAAG,QAAQ,EAAE,IAChD,CAAC;AAAA,MACL,GAAI,QAAQ,eAAe,EAAE,QAAQ,QAAQ,aAAa,MAAM,GAAG,QAAQ,EAAE,IAAI,CAAC;AAAA,IACpF,CAAC;AACD,UAAM,YAAY,OAAO,UAAU,MAAM,GAAG,QAAQ;AACpD,UAAM,SAAS,MAAM,QAAQ;AAAA,MAC3B,QAAQ,OAAO;AAAA,MACf;AAAA,MACA,QAAQ,OAAO;AAAA,MACf,WAAW,OAAO;AAAA,MAClB;AAAA,MACA,KAAK,EAAE,eAAe,QAAQ,IAAI,QAAQ,GAAI,eAAe,EAAE,aAAa,IAAI,CAAC,EAAG;AAAA,MACpF,eAAe;AAAA,MACf,gBAAgB,KAAK,IAAI,gBAAgB,QAAQ;AAAA,IACnD,CAAC;AACD,UAAM,SAAS,MAAM,gBAAgB;AAAA,MACnC,YAAY,OAAO;AAAA,MACnB,UAAU,QAAQ;AAAA,MAClB,WAAW,QAAQ,mBAAmB;AAAA,MACtC;AAAA,MACA,QAAQ,IAAI;AAAA,IACd,CAAC;AACD,QAAI,CAAC,OAAQ,OAAM,IAAI,MAAM,gBAAgB,QAAQ,QAAQ,CAAC;AAC9D,QAAI,OAAO,EAAE,WAAW,UAAU,QAAQ,OAAO,YAAY,CAAC;AAC9D,WAAO;AAAA,EACT;AACF;AAyBA,eAAe,gBAAgB,MAA6D;AAC1F,QAAM,QAA0B,CAAC;AACjC,aAAW,QAAQ,KAAK,YAAY;AAClC,QAAI,KAAK,WAAW,UAAa,KAAK,SAAS,KAAK,SAAS,UAAU,KAAM;AAC7E,UAAM,KAAK;AAAA,MACT,OAAO,KAAK;AAAA,MACZ,QAAQ,KAAK;AAAA,MACb,OAAO,KAAK,QAAQ,SAAS;AAAA,MAC7B,WAAW,KAAK,QAAQ,SAAS;AAAA,IACnC,CAAC;AAAA,EACH;AACA,MAAI,MAAM,WAAW,EAAG,QAAO;AAE/B,MAAI,WAAW;AACf,MAAI,KAAK,UAAU;AACjB,eAAW,CAAC;AACZ,eAAW,KAAK,OAAO;AACrB,YAAM,SAAS,MAAM,KAAK,SAAS,EAAE,QAAQ,KAAK,MAAM,EAAE,QAAQ,KAAK,OAAO,CAAC;AAC/E,UAAI,OAAO,SAAU,UAAS,KAAK,EAAE,GAAG,GAAG,WAAW,OAAO,UAAU,CAAC;AAAA,IAC1E;AACA,QAAI,SAAS,WAAW,EAAG,QAAO;AAAA,EACpC;AAEA,SAAO,qBAAqB,UAAU,KAAK,SAAS,EAAE;AACxD;AAGA,SAAS,qBACP,YACA,WACgB;AAChB,QAAM,YAAY,CAAC,MACjB,EAAE,OAAO,UAAU,aAAa,EAAE,OAAO,UAAU;AACrD,QAAM,SAAS,CAAC,GAAG,UAAU,EAAE,KAAK,CAAC,GAAG,MAAM;AAC5C,YAAQ,WAAW;AAAA,MACjB,KAAK;AACH,eAAO,UAAU,CAAC,IAAI,UAAU,CAAC,KAAK,EAAE,QAAQ,EAAE;AAAA,MACpD,KAAK;AACH,eAAO,EAAE,YAAY,EAAE,aAAa,EAAE,QAAQ,EAAE;AAAA,MAClD,KAAK;AACH,eAAO,EAAE,QAAQ,EAAE;AAAA,MACrB;AACE,eAAO,EAAE,QAAQ,EAAE,SAAS,EAAE,QAAQ,EAAE;AAAA,IAC5C;AAAA,EACF,CAAC;AACD,SAAO,OAAO,CAAC;AACjB;AAEA,SAAS,gBAAgB,UAA6C;AACpE,SAAO,WACH,4DACA;AACN;AAEA,SAAS,eAAe,MAAgC;AACtD,MAAI,CAAC,KAAK,YAAa,QAAO,KAAK;AACnC,SAAO,CAAC,KAAK,MAAM,IAAI,cAAc,KAAK,WAAW,EAAE,KAAK,IAAI;AAClE;AAEA,SAAS,gBAAgB,SAAgE;AACvF,MAAI,QAAQ,YAAY,QAAQ,eAAe;AAC7C,UAAM,IAAI,MAAM,+EAA+E;AAAA,EACjG;AACA,MAAI,QAAQ,SAAU,QAAO,QAAQ;AACrC,MAAI,QAAQ,eAAe;AACzB,WAAO,6BAA6B,EAAE,QAAQ,QAAQ,cAAc,CAAC;AAAA,EACvE;AACA,QAAM,IAAI,MAAM,uEAAuE;AACzF;AAQA,IAAM,mBAAmB;AAAA,EACvB,MAAM;AAAA,EACN,MAAM,KAAW,MAAY,SAAkD;AAC7E,WAAO,QAAQ,WAAW,IAAI,CAAC,IAAI,IAAI,CAAC;AAAA,EAC1C;AAAA,EACA,OAAO,SAAyD;AAC9D,WAAO,QAAQ,SAAS,IAAI,gBAAgB;AAAA,EAC9C;AACF;","names":["result","chosen"]}

package/dist/{chunk-65FQLI4V.js → chunk-72JQCHOZ.js}

@@ -1024,6 +1024,7 @@ function createSandboxLineage(client, capabilities, options = {}) {
     if (signal.aborted) throwAbort();
     const opts = buildBackendOptions(spec.profile, spec.sandboxOverrides);
     const box = await acquireSandbox(client, opts, { signal });
+    await spec.prepareBox?.(box, { signal });
     owned.push(box);
     return box;
   };
@@ -1052,6 +1053,7 @@ function createSandboxLineage(client, capabilities, options = {}) {
         if (checkpointId !== void 0) {
           const box2 = await forkFromCheckpoint(parent.box, checkpointId, signal);
           owned.push(box2);
+          await spec.prepareBox?.(box2, { signal });
           const sessionId2 = mintSessionId();
           return {
             handle: { box: box2, sessionId: sessionId2 },
@@ -1475,6 +1477,7 @@ async function executeIteration(args) {
     if (args.validator) {
       slot.verdict = await args.validator.validate(slot.output, {
         iteration: args.item.index,
+        ...box ? { box } : {},
         signal: args.signal,
         traceEmitter: args.ctx.traceEmitter
       });
@@ -1571,7 +1574,9 @@ function readSandboxId(box) {
 async function createSandboxForSpec(client, spec, signal) {
   const opts = buildBackendOptions(spec.profile, spec.sandboxOverrides);
   if (signal.aborted) throwAbort();
-  return acquireSandbox(client, opts, { signal });
+  const box = await acquireSandbox(client, opts, { signal });
+  await spec.prepareBox?.(box, { signal });
+  return box;
 }
 function finalize(args) {
   const winner = args.options.selectWinner ? args.options.selectWinner(args.iterations) : args.options.driver.selectWinner?.(args.iterations) ?? defaultSelectWinner(args.iterations);
@@ -1732,6 +1737,162 @@ function loopDispatch(opts) {
   return (profile, scenario, ctx) => runLoopForCell(opts, scenario, profile, ctx);
 }
 
+// src/runtime/observe.ts
+import { makeFinding } from "@tangle-network/agent-eval";
+var observerId = "observe/trace";
+function summarizeTrace(trace, maxLines) {
+  const lines = [];
+  for (const ev of trace) {
+    const e = ev;
+    const t = (e.type ?? "").toLowerCase();
+    const d = e.data ?? {};
+    const part = d.part ?? {};
+    if (part.type === "tool")
+      lines.push(`tool:${part.tool}${part.state?.status ? `(${part.state.status})` : ""}`);
+    else if (t.includes("error"))
+      lines.push(`ERROR: ${String(d.message ?? d.detail ?? "").slice(0, 200)}`);
+    else if (t === "status" && typeof d.status === "string") lines.push(`status:${d.status}`);
+    else if (t.includes("tool")) lines.push(`tool-event:${t}`);
+  }
+  const out = [];
+  for (const ln of lines) {
+    const prev = out[out.length - 1];
+    const m = prev?.match(/^(.*?)(?: x(\d+))?$/);
+    if (m && m[1] === ln) out[out.length - 1] = `${ln} x${(Number(m[2]) || 1) + 1}`;
+    else out.push(ln);
+  }
+  return out.slice(0, maxLines).join("\n") || "(no tool/error events in trace)";
+}
+var findingsSchema = {
+  name: "observer_findings",
+  schema: {
+    type: "object",
+    additionalProperties: false,
+    properties: {
+      findings: {
+        type: "array",
+        items: {
+          type: "object",
+          additionalProperties: false,
+          properties: {
+            area: {
+              type: "string",
+              description: "tool-use | cost | verification | process | failure | latency"
+            },
+            severity: { type: "string", enum: ["critical", "high", "medium", "low", "info"] },
+            claim: {
+              type: "string",
+              description: "what you OBSERVED in the trace (a fact, with the evidence)"
+            },
+            recommended_action: {
+              type: "string",
+              description: "the concrete change for the agent or operator"
+            },
+            audience: {
+              type: "string",
+              enum: ["agent", "operator"],
+              description: "who should act on this"
+            },
+            confidence: { type: "number" }
+          },
+          required: ["area", "severity", "claim", "recommended_action", "audience", "confidence"]
+        }
+      }
+    },
+    required: ["findings"]
+  }
+};
+async function observe(input, opts) {
+  const traceSummary = summarizeTrace(input.trace, opts.maxTraceLines ?? 80);
+  const res = await opts.chat.chat(
+    {
+      ...opts.model ? { model: opts.model } : {},
+      jsonSchema: findingsSchema,
+      messages: [
+        {
+          role: "system",
+          content: "You are a third-person OBSERVER watching an AI agent work. You see its TRACE (what it did), not its grader. From the trace, name SPECIFIC, behavior-grounded findings: wasted/duplicated tool calls, thrash/retries, token/cost waste, missing verification, failure patterns. For each, a concrete recommended_action, and whether the AGENT (fix its skills/prompt/tools) or the OPERATOR (fix framing/decomposition/config) should act. Only claim what the trace shows. No findings if the run was clean."
+        },
+        {
+          role: "user",
+          content: `TASK: ${input.task}
+
+OUTCOME: ${input.outcome ?? "unknown"}
+
+FINAL OUTPUT (truncated):
+${input.output.slice(0, 1200)}
+
+TRACE (in order; "xN" = repeated):
+${traceSummary}`
+        }
+      ]
+    },
+    { ...opts.signal ? { signal: opts.signal } : {} }
+  );
+  const parsed = parseFindings(res.content);
+  const producedAt = input.runId ? `${input.runId}` : observerId;
+  const findings = parsed.map(
+    (f) => makeFinding({
+      analyst_id: observerId,
+      area: `${f.area}`,
+      severity: f.severity,
+      claim: f.claim,
+      recommended_action: f.recommended_action,
+      confidence: typeof f.confidence === "number" ? f.confidence : 0.5,
+      evidence_refs: [],
+      // The observer reads BEHAVIOR, never the judge verdict — firewall provenance.
+      derived_from_judge: false,
+      metadata: { audience: f.audience },
+      ...input.runId ? { subject: input.runId } : {}
+    })
+  );
+  const learned = [];
+  if (opts.corpus) {
+    for (const f of findings) {
+      const record = {
+        schemaVersion: "1.0.0",
+        id: f.finding_id,
+        runId: input.runId ?? observerId,
+        producedAt: f.produced_at ?? producedAt,
+        area: f.area,
+        claim: f.recommended_action ?? f.claim,
+        ...f.claim ? { rationale: f.claim } : {},
+        tags: [...opts.tags ?? [], `audience:${f.metadata?.audience ?? "agent"}`],
+        confidence: f.confidence,
+        evidence: [{ kind: "finding", uri: f.finding_id }]
+      };
+      const r = await opts.corpus.append(record);
+      if (r.succeeded) learned.push(record);
+    }
+  }
+  return { findings, learned, report: renderReport(findings) };
+}
+function parseFindings(content) {
+  let obj;
+  try {
+    obj = JSON.parse(content);
+  } catch {
+    const m = content.match(/\{[\s\S]*\}/);
+    obj = m ? JSON.parse(m[0]) : { findings: [] };
+  }
+  const arr = obj.findings;
+  return Array.isArray(arr) ? arr : [];
+}
+function renderReport(findings) {
+  if (findings.length === 0) return "\u2713 clean run \u2014 the observer found nothing to change.";
+  const audience = (f) => f.metadata?.audience ?? "agent";
+  const forAgent = findings.filter((f) => audience(f) === "agent");
+  const forOperator = findings.filter((f) => audience(f) === "operator");
+  const block = (title, fs) => fs.length === 0 ? "" : `**${title}**
+${fs.map((f) => `- [${f.severity}] ${f.claim}
+  \u2192 ${f.recommended_action ?? ""}`).join("\n")}
+`;
+  return [
+    block("For the agent (fix skills / prompt / tools)", forAgent),
+    block("For you (the operator)", forOperator)
+  ].filter(Boolean).join("\n");
+}
+
 // src/runtime/supervise/scope.ts
 function createScope(args) {
   const children = /* @__PURE__ */ new Map();
@@ -4028,6 +4189,70 @@ function errorMessage(error) {
   return error instanceof Error ? error.message : String(error);
 }
 
+// src/runtime/workspace.ts
+function localShell() {
+  return async (args, cwd) => {
+    const { execFile } = await import("child_process");
+    const [bin, ...rest] = args;
+    return new Promise((resolve) => {
+      execFile(
+        bin ?? "",
+        rest,
+        { cwd, encoding: "utf-8", maxBuffer: 64 * 1024 * 1024 },
+        (err, stdout, stderr) => {
+          resolve({
+            stdout: stdout ?? "",
+            stderr: stderr ?? "",
+            code: err ? err.code ?? 1 : 0
+          });
+        }
+      );
+    });
+  };
+}
+function gitWorkspace(opts) {
+  const shell = opts.shell ?? localShell();
+  const branch = opts.branch ?? "main";
+  const cfg = opts.noHooks === false ? [] : ["-c", "core.hooksPath=/dev/null"];
+  const ident = ["-c", "user.email=workspace@tangle.local", "-c", "user.name=workspace"];
+  const run = async (args, cwd) => {
+    const res = await shell(["git", ...cfg, ...ident, ...args], cwd);
+    if (res.code !== 0) {
+      throw new Error(
+        `git ${args.join(" ")} failed (${res.code}): ${tail(res.stderr || res.stdout)}`
+      );
+    }
+    return res.stdout;
+  };
+  return {
+    ref: opts.ref,
+    materialize: (dir) => run(["clone", "--branch", branch, opts.ref, dir]).then(() => {
+    }),
+    async commit(dir, message) {
+      await run(["add", "-A"], dir);
+      const status = await run(["status", "--porcelain"], dir);
+      if (!status.trim()) return { ok: true, rev: (await run(["rev-parse", "HEAD"], dir)).trim() };
+      await run(["commit", "-m", message], dir);
+      const pull = await shell(["git", ...cfg, ...ident, "pull", "--rebase", "origin", branch], dir);
+      if (pull.code !== 0) {
+        await shell(["git", ...cfg, "rebase", "--abort"], dir).catch(() => {
+        });
+        return { ok: false, conflict: tail(pull.stderr || pull.stdout) };
+      }
+      const push = await shell(["git", ...cfg, ...ident, "push", "origin", branch], dir);
+      if (push.code !== 0) return { ok: false, conflict: tail(push.stderr || push.stdout) };
+      return { ok: true, rev: (await run(["rev-parse", "HEAD"], dir)).trim() };
+    },
+    async head() {
+      const out = await run(["ls-remote", opts.ref, `refs/heads/${branch}`]);
+      return out.split(/\s+/)[0] ?? "";
+    }
+  };
+}
+function tail(s) {
+  return s.slice(-400);
+}
+
 export {
   contentAddress,
   InMemoryResultBlobStore,
@@ -4059,6 +4284,8 @@ export {
   createSandboxForSpec,
   defaultSelectWinner,
   loopDispatch,
+  observe,
+  renderReport,
   createScope,
   settledToIteration,
   pipeline,
@@ -4084,6 +4311,8 @@ export {
   runPersonified,
   trajectoryReport,
   equalKOnCost,
-  openSandboxRun
+  openSandboxRun,
+  localShell,
+  gitWorkspace
 };
-//# sourceMappingURL=chunk-65FQLI4V.js.map
+//# sourceMappingURL=chunk-72JQCHOZ.js.map