@tangle-network/agent-runtime 0.71.1 → 0.73.0

package/README.md

@@ -240,10 +240,10 @@ on a never-touched holdout slice.
 `runDelegatedLoop` is one entrypoint a worker agent or a scheduled routine calls to run a disciplined loop in a chosen mode, over the hardened engines below. It fails loud on an unwired mode; a thrown engine is captured as `{ ok: false }`, so unattended runs record rather than crash.
 
 ```ts
-import { runDelegatedLoop, coderLoopRunner, researchLoopRunner, type DelegatedLoopRegistry } from '@tangle-network/agent-runtime'
+import { runDelegatedLoop, worktreeLoopRunner, researchLoopRunner, type DelegatedLoopRegistry } from '@tangle-network/agent-runtime'
 
 const registry: DelegatedLoopRegistry = {
-  code: coderLoopRunner({ sandboxClient, args: { goal: 'fix the flaky retry test', repoRoot: '/repo' }, reviewer, winnerSelection: 'smallest-diff' }),
+  code: worktreeLoopRunner({ repoRoot: '/repo', taskPrompt: 'fix the flaky retry test', harnesses, budget }),
   research: researchLoopRunner({ research, gate: { selfArtifactKinds: ['spec'] }, maxRounds: 3 }),
 }
 const result = await runDelegatedLoop('code', registry)
@@ -251,7 +251,7 @@ const result = await runDelegatedLoop('code', registry)
 
 Modes: `code`, `review`, `research`, `audit`, `self-improve`, `dynamic`. The `agent-runtime-loop` bin runs the registry from a cron or routine and exits 0 (ok), 1 (recorded failure), or 2 (usage or config error).
 
-The coder delegate (`createDefaultCoderDelegate`, `/mcp`) has default-on safety gates: no-op rejection (an empty patch cannot pass trivially), an always-on secret-path floor (`.env`, keys, wallets), an optional `reviewer` gate, and a `winnerSelection` policy (`highest-score`, `smallest-diff`, `highest-readiness`, `first-approved`).
+`worktreeLoopRunner` (`code` mode, the generic recursive path) authors one `AgentProfile` per harness and runs them as a `worktreeFanout` (each leaf `gateOnDeliverable`), winner by the shared valid-only selector. The sandbox-session counterpart is `detachedSessionDelegate` (`/mcp`): it drives the in-box harness over a `SandboxClient` to a mechanically-validated patch, with default-on safety gates — no-op rejection, an always-on secret-path floor (`.env`, keys, wallets), an optional `reviewer` gate, and a `winnerSelection` policy. Its worker profile is a parameter the caller authors (`workerProfile`); omit it for a minimal model-only default.
 
 The knowledge-base gate (`createKbGate`, `/mcp`) is fail-closed: a fact's `verbatimPassage` must appear in its `sourceText`, the asserted value must be in the passage, and citations cannot point at self-generated artifacts. `researchLoopRunner` wraps it with a correct-on-veto loop that re-researches the vetoed gaps up to `maxRounds`, then returns the unverified ones rather than dropping them.
 
@@ -271,15 +271,15 @@ The shape: `loop` to `loop.round` (move plus rationale) to `loop.iteration` (age
 
 ## MCP delegation server
 
-Expose the delegation tools (`delegate_code`, `delegate_research`, `delegate_feedback`, `delegation_status`, `delegation_history`) to a sandbox coding agent. Mount the canonical server instead of forking delegation logic.
+Expose the delegation tools to a sandbox coding agent: the generic `delegate` verb (one intent → a supervisor that authors + drives its own worker, returns the delivered output with its real spend) plus the queue-bound `delegate_feedback`, `delegation_status`, `delegation_history` (and `delegate_ui_audit` when a UI-audit runner is wired). Mount the canonical server instead of forking delegation logic.
 
 ```ts
-import { createMcpServer, createDefaultCoderDelegate } from '@tangle-network/agent-runtime/mcp'
+import { createMcpServer } from '@tangle-network/agent-runtime/mcp'
 
-const server = createMcpServer({ coderDelegate: createDefaultCoderDelegate({ sandboxClient }), researcherDelegate })
+const server = createMcpServer({ delegateSupervisor: { router, backend, deliverable } })
 ```
 
-Or mount the `agent-runtime-mcp` stdio bin on a production `AgentProfile.mcp`.
+Or mount the `agent-runtime-mcp` stdio bin on a production `AgentProfile.mcp` with `MCP_ENABLE_DELEGATE=1`.
 
 Delegation state is in-memory by default — a server restart drops pending delegations and history. Set `AGENT_RUNTIME_DELEGATION_STATE_FILE=/path/state.json` on the bin (or construct via `DelegationTaskQueue.restore({ store: new FileDelegationStore({ filePath }) })`) to persist records across restarts: `delegation_status`/`delegation_history` keep answering for prior runs, idempotency keys dedupe resubmissions, and in-flight records either resume through the `resumeDelegate` seam (when submitted with a `detachedSessionRef`) or settle as failed with an explicit driver-restart error. A corrupt state file refuses to load (`DelegationStateCorruptError`); `AGENT_RUNTIME_DELEGATION_STATE_RECOVER=1` archives it and starts empty. `AGENT_RUNTIME_DELEGATION_RETAIN_TERMINAL=<n>` caps retained terminal records.
 
@@ -342,7 +342,7 @@ Six subpaths — the public surface:
 | `@tangle-network/agent-runtime` | chat turns, delegated loop-runner, OTEL export, errors, model resolution |
 | `.../agent` | `defineAgent` plus surface and outcome adapters |
 | `.../loops` | **the optimization suite** (`Environment`, `defineStrategy`, `runBenchmark`, `runStrategyEvolution`, `authorStrategy`, `promotionGate`) + the recursive atom (`Supervisor`/`Scope`, `createExecutor`), the `runLoop` kernel, the `Driver` type, `loopDispatch` |
-| `.../profiles` | `coderProfile`, `researcherProfile`, the `uiAuditorProfile` presets + the UI-audit workspace I/O helpers |
+| `.../profiles` | `coderTaskToPrompt` (the coder task formatter), the `uiAuditorProfile` presets + the UI-audit workspace I/O helpers |
 | `.../intelligence` | `withTangleIntelligence`, `createIntelligenceClient` — Observe + the provable-OFF billing boundary |
 | `.../mcp` | `createMcpServer`, `createDefaultCoderDelegate`, `createKbGate`, the `agent-runtime-mcp` bin |
 

package/dist/agent.d.ts

@@ -1,12 +1,11 @@
 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, a as OutputAdapter, A as AgentRunSpec } from './types-CdnEAE3U.js';
+import { R as RuntimeStreamEvent, S as SandboxClient, a as OutputAdapter, A as AgentRunSpec } from './types-JufmXF2a.js';
 import { A as AgentSurfaces } from './improvement-adapter-CioiEE2z.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-CioiEE2z.js';
 import { K as KnowledgeAdapter, a as RunAnalystLoopResult } from './types-BC3bZpH0.js';
-import { AgentProfile } from '@tangle-network/agent-interface';
+import { AgentProfile, AgentProfileFileMount, AgentProfileMcpServer } from '@tangle-network/agent-interface';
 import { SandboxEvent } from '@tangle-network/sandbox';
-import { C as ComposeProductionAgentProfileOptions } from './delegation-profile-Bvfro2m1.js';
 
 /**
  * The full agent manifest. Each agent ships ONE of these.
@@ -387,12 +386,13 @@ declare function measureOutcome<TProposal, TEdit>(result: RunAnalystLoopResult<T
  *
  * The point of this adapter is parity: the eval substrate must run the agent
  * through the SAME profile the production chat turn uses, or scorecard numbers
- * grade a profile that never ships. `createSandboxAct` composes the production
- * profile via {@link composeProductionAgentProfile}, boots a sandbox with it
- * through the loop kernel's own {@link createSandboxForSpec}, streams the
- * `streamPrompt` events mapped to the `RuntimeStreamEvent` vocabulary, and
- * resolves the `OutputAdapter`-parsed output for rubric scoring — satisfying
- * the `act` streaming contract with one code path shared by chat and eval.
+ * grade a profile that never ships. `createSandboxAct` boots a sandbox with the
+ * agent's profile (the caller's `baseProfile`, with optional per-persona
+ * overrides) through the loop kernel's own {@link createSandboxForSpec},
+ * streams the `streamPrompt` events mapped to the `RuntimeStreamEvent`
+ * vocabulary, and resolves the `OutputAdapter`-parsed output for rubric scoring
+ * — satisfying the `act` streaming contract with one code path shared by chat
+ * and eval.
  *
  * Agents with a bespoke streaming chat turn should wire THAT into `act`
  * directly (the contract is designed for it); this adapter is the default for
@@ -400,8 +400,22 @@ declare function measureOutcome<TProposal, TEdit>(result: RunAnalystLoopResult<T
  * agents agent-builder generates.
  */
 
+/** Per-persona profile-merge slots applied over the base profile (§1.5: the caller authors the
+ *  per-persona profile). Each slot overlays the base; an absent slot leaves the base untouched. */
+interface SandboxActComposeOverrides {
+    /** Replace the base profile's system prompt (e.g. a workspace-augmented prompt). */
+    systemPrompt?: string;
+    /** Extra file mounts layered after the base profile's `resources.files`. */
+    extraFiles?: AgentProfileFileMount[];
+    /** Override the profile `name`. Defaults to the base profile's name. */
+    name?: string;
+    /** Box built-in tool ON/OFF flags merged over the base profile's `tools` (overlay wins per key). */
+    tools?: Record<string, boolean>;
+    /** MCP connections merged over the base profile's `mcp` (overlay wins per key). */
+    mcpConnections?: Record<string, AgentProfileMcpServer>;
+}
 interface CreateSandboxActOptions<TPersona, TRunOutput> {
-    /** Canonical agent profile — the same one the prod chat turn composes from. */
+    /** Canonical agent profile — the same one the prod chat turn uses. */
     baseProfile: AgentProfile;
     /** Sandbox client used to boot the per-run sandbox. */
     sandboxClient: SandboxClient;
@@ -410,12 +424,10 @@ interface CreateSandboxActOptions<TPersona, TRunOutput> {
     /** Sandbox event stream → typed output the rubric scores. */
     output: OutputAdapter<TRunOutput>;
     /**
-     * Per-persona composition overrides (workspace-augmented system prompt,
-     * extra file mounts, sandbox key). Merged into
-     * {@link composeProductionAgentProfile}; `env` here is overridden by the
-     * top-level `env` option when both are set.
+     * Per-persona profile overrides (workspace-augmented system prompt, extra
+     * file mounts, tool flags, MCP connections). Overlaid onto `baseProfile`.
      */
-    compose?: (persona: TPersona) => ComposeProductionAgentProfileOptions;
+    compose?: (persona: TPersona) => SandboxActComposeOverrides;
     /** Sandbox-SDK overrides forwarded to `createSandboxForSpec`. */
     sandboxOverrides?: AgentRunSpec<unknown>['sandboxOverrides'];
     /** Stable run name surfaced in mapped `llm_call` events. */
@@ -424,8 +436,6 @@ interface CreateSandboxActOptions<TPersona, TRunOutput> {
     mapEvent?: (event: SandboxEvent, opts: {
         agentRunName?: string;
     }) => RuntimeStreamEvent | undefined;
-    /** Environment source for delegation-MCP composition. Defaults to `process.env`. */
-    env?: Record<string, string | undefined>;
 }
 /**
  * Build an `AgentRuntime.act` implementation backed by a single prod-profile

package/dist/agent.js

@@ -1,13 +1,12 @@
-import {
-  composeProductionAgentProfile
-} from "./chunk-XRYEZPR6.js";
+import "./chunk-HPYWEFVY.js";
 import {
   createSandboxForSpec
-} from "./chunk-5ISW5JUF.js";
+} from "./chunk-NCH4XUZ7.js";
 import "./chunk-VMNEQHJR.js";
 import {
   mapSandboxEvent
-} from "./chunk-VLF5RHEQ.js";
+} from "./chunk-T2HVQVB4.js";
+import "./chunk-NBV35BR6.js";
 import "./chunk-WIR4HOOJ.js";
 import {
   __require
@@ -507,10 +506,7 @@ function meanComposite(rows) {
 function createSandboxAct(options) {
   const mapEvent = options.mapEvent ?? mapSandboxEvent;
   return (persona, ctx) => {
-    const profile = composeProductionAgentProfile(options.baseProfile, {
-      ...options.compose?.(persona) ?? {},
-      ...options.env ? { env: options.env } : {}
-    });
+    const profile = applyComposeOverrides(options.baseProfile, options.compose?.(persona));
     const agentRunName = options.name ?? profile.name ?? "agent";
     const message = options.buildPrompt(persona);
     const signal = ctx.signal ?? new AbortController().signal;
@@ -546,6 +542,22 @@ function createSandboxAct(options) {
     return { events: events(), output };
   };
 }
+function applyComposeOverrides(base, overrides) {
+  if (!overrides) return base;
+  const prompt = overrides.systemPrompt ? { ...base.prompt, systemPrompt: overrides.systemPrompt } : base.prompt;
+  const mergedTools = overrides.tools ? { ...base.tools ?? {}, ...overrides.tools } : base.tools;
+  const mergedMcp = overrides.mcpConnections ? { ...base.mcp ?? {}, ...overrides.mcpConnections } : base.mcp;
+  const baseFiles = base.resources?.files ?? [];
+  const mergedFiles = overrides.extraFiles?.length ? [...baseFiles, ...overrides.extraFiles] : [...baseFiles];
+  return {
+    ...base,
+    name: overrides.name ?? base.name,
+    prompt,
+    ...mergedTools ? { tools: mergedTools } : {},
+    ...mergedMcp ? { mcp: mergedMcp } : {},
+    resources: { ...base.resources, files: mergedFiles }
+  };
+}
 export {
   AgentManifestError,
   collectAgentRun,

package/dist/agent.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/agent/surfaces.ts","../src/agent/define-agent.ts","../src/agent/improvement-adapter.ts","../src/agent/knowledge-adapter.ts","../src/agent/outcome.ts","../src/agent/sandbox-act.ts"],"sourcesContent":["/**\n * `AgentSurfaces` — declarative map of the mutable file/directory paths\n * the self-improvement loop can edit on behalf of an agent.\n *\n * The substrate uses this map to resolve every parsed `FindingSubject`\n * (from agent-eval) to a real on-disk path. No per-vertical glue;\n * no fabricated paths; no silent `existsSync(...)` skips that hide\n * misconfiguration from the operator.\n *\n * Surfaces are validated at `defineAgent` time — missing paths fail\n * loud with a list of every offender. A surface that's not needed\n * (e.g. an agent with no RAG corpora) is simply omitted; the loop\n * refuses to route those subjects rather than fabricating a target.\n */\n\nimport { existsSync } from 'node:fs'\nimport { isAbsolute, join } from 'node:path'\nimport type { FindingSubject } from '@tangle-network/agent-eval'\n\n/**\n * Surface declarations. Every path is repo-relative (or absolute) at\n * `defineAgent` time. At resolution time, paths are joined against the\n * agent's `repoRoot`.\n *\n * `systemPrompt`, `tools`, `personas` are DIRECTORIES; the loop appends\n * `<section>.md`, `<tool>/README.md`, `<persona-id>.yaml` etc.\n * `rubric`, `outputSchema` are SINGLE FILES; the loop edits them in\n * place.\n *\n * `knowledge` is the agent-knowledge root (typically `.agent-knowledge`);\n * `applyKnowledgeWriteBlocks` writes pages relative to it.\n *\n * Optional surfaces (`scaffolding`, `memory`, `rag`, `outputSchema`)\n * can be omitted — the loop will reject findings targeting them with a\n * clear log message instead of fabricating a path.\n */\nexport interface AgentSurfaces {\n  /** Directory containing one markdown file per system-prompt section. */\n  systemPrompt: string\n  /** Directory containing one subdir per tool (`<tool>/README.md`). */\n  tools: string\n  /** Single file (TypeScript module) defining the rubric weights + dimensions. */\n  rubric: string\n  /** Knowledge-base root; typically `.agent-knowledge`. */\n  knowledge: string\n  /** Directory containing one YAML/JSON file per persona. */\n  personas: string\n  /** Optional: directory containing scaffolding rules (precondition checks, retry policies). */\n  scaffolding?: string\n  /** Optional: memory store path (JSONL / SQLite / DB). */\n  memory?: string\n  /** Optional: directory containing RAG corpora (`<corpus>/<doc-id>.md`). */\n  rag?: string\n  /** Optional: single file defining the output schema (Zod / JSON Schema). */\n  outputSchema?: string\n}\n\nexport interface ResolvedSurface {\n  /** Absolute filesystem path the operator can `cat` / `vim`. */\n  absolutePath: string\n  /** Repo-relative path for PR descriptions, diffs, audit logs. */\n  repoRelativePath: string\n  /** Whether the path currently exists on disk. */\n  exists: boolean\n  /** The substrate's intent: edit an existing file or create a new one. */\n  intent: 'edit-existing' | 'create-new'\n}\n\n/**\n * Resolve a parsed `FindingSubject` to the file path the substrate\n * should edit (or create) on disk.\n *\n * Returns `null` when:\n *   - the subject targets a surface the agent didn't declare\n *     (e.g. `rag:*` when `surfaces.rag` is undefined), OR\n *   - the subject is a `cluster` (failure-mode emits these as evidence,\n *     not actionable mutations — they don't route to a file).\n *\n * Returns a `ResolvedSurface` with `intent: 'create-new'` when the\n * subject names a path that doesn't yet exist (e.g. a new wiki page).\n * The caller chooses whether to honour the create — for tightly-managed\n * surfaces like `systemPrompt` it's usually a contract violation\n * (the analyst named a section that doesn't exist); for `knowledge`\n * it's the whole point.\n */\nexport function resolveSubjectPath(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ResolvedSurface | null {\n  const candidates = candidatePathsForSubject(subject, surfaces)\n  if (candidates.length === 0) return null\n\n  // Probe candidates in order, preferring the first one that exists on disk.\n  // Lets the substrate accept both the flat `<section>.md` convention and\n  // the skill-dir `<section>/SKILL.md` convention without forcing one layout.\n  // When none exists, fall back to the first candidate (canonical create-new).\n  for (const rel of candidates) {\n    const abs = isAbsolute(rel) ? rel : join(repoRoot, rel)\n    if (existsSync(abs)) {\n      return { absolutePath: abs, repoRelativePath: rel, exists: true, intent: 'edit-existing' }\n    }\n  }\n  const fallback = candidates[0]!\n  const fallbackAbs = isAbsolute(fallback) ? fallback : join(repoRoot, fallback)\n  return {\n    absolutePath: fallbackAbs,\n    repoRelativePath: fallback,\n    exists: false,\n    intent: 'create-new',\n  }\n}\n\nfunction candidatePathsForSubject(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n): ReadonlyArray<string> {\n  switch (subject.kind) {\n    case 'knowledge.wiki':\n    case 'knowledge.stale':\n      return [join(surfaces.knowledge, `${subject.slug}.md`)]\n    case 'knowledge.claim':\n      // Claims land in a per-topic claims directory under the knowledge root.\n      return [join(surfaces.knowledge, 'claims', `${slugify(subject.topic)}.md`)]\n    case 'knowledge.raw':\n      return [join(surfaces.knowledge, 'raw', `${subject.sourceId}.md`)]\n    case 'system-prompt': {\n      const slug = slugify(subject.section)\n      // Prefer flat layout for create-new (canonical); probe skill-dir layout\n      // in case the existing repo (tax/legal/gtm/creative) uses\n      // `<section>/SKILL.md` already.\n      return [\n        join(surfaces.systemPrompt, `${slug}.md`),\n        join(surfaces.systemPrompt, slug, 'SKILL.md'),\n        join(surfaces.systemPrompt, slug, 'index.md'),\n      ]\n    }\n    case 'tool-doc':\n      if (subject.aspect) {\n        return [join(surfaces.tools, subject.tool, `${slugify(subject.aspect)}.md`)]\n      }\n      // tool-doc default: `<tool>/README.md`; also probe `<tool>.md` for flat\n      // tool-list repos.\n      return [\n        join(surfaces.tools, subject.tool, 'README.md'),\n        join(surfaces.tools, `${subject.tool}.md`),\n      ]\n    case 'new-tool':\n      return [join(surfaces.tools, subject.name, 'README.md')]\n    case 'rag':\n      if (!surfaces.rag) return []\n      return [join(surfaces.rag, subject.corpus, `${subject.docId}.md`)]\n    case 'memory':\n      if (!surfaces.memory) return []\n      return [join(surfaces.memory, `${slugify(subject.key)}.json`)]\n    case 'scaffolding':\n      if (!surfaces.scaffolding) return []\n      return [join(surfaces.scaffolding, `${slugify(subject.concern)}.md`)]\n    case 'output-schema':\n      if (!surfaces.outputSchema) return []\n      return [surfaces.outputSchema]\n    case 'websearch.outdated':\n    case 'prior-run-summary':\n      // Stale signals don't map to a single file — handled by the knowledge\n      // adapter as `agent-knowledge:stale:*` after operator review.\n      return []\n    case 'cluster':\n      // failure-mode cluster labels are evidence, not mutations.\n      return []\n  }\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9-]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 200) || 'untitled'\n  )\n}\n\n/**\n * Validate that every declared surface exists on disk under `repoRoot`.\n *\n * Returns an array of `SurfaceValidationIssue` — empty when all required\n * surfaces resolve. `defineAgent` throws with the issues rendered, so\n * a misconfigured manifest fails at startup (not at the first finding\n * the loop produces 20 minutes later).\n */\nexport interface SurfaceValidationIssue {\n  surface: keyof AgentSurfaces\n  path: string\n  reason: 'missing' | 'not-directory' | 'not-file'\n}\n\nexport function validateSurfaces(\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ReadonlyArray<SurfaceValidationIssue> {\n  const issues: SurfaceValidationIssue[] = []\n  const dirSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'systemPrompt',\n    'tools',\n    'personas',\n    'knowledge',\n  ]\n  const fileSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['rubric']\n  const optionalDirSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['scaffolding', 'memory', 'rag']\n  const optionalFileSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['outputSchema']\n\n  for (const key of dirSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    }\n  }\n  for (const key of fileSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    }\n  }\n  for (const key of [...optionalDirSurfaces, ...optionalFileSurfaces]) {\n    const p = surfaces[key] as string | undefined\n    if (p === undefined) continue\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    }\n  }\n  return issues\n}\n\nexport function renderSurfaceIssues(\n  issues: ReadonlyArray<SurfaceValidationIssue>,\n  repoRoot: string,\n): string {\n  if (issues.length === 0) return ''\n  const lines = issues.map(\n    (i) => `  - ${i.surface}: ${i.path ? `\"${i.path}\"` : '<not set>'} (${i.reason})`,\n  )\n  return [\n    `Agent surface validation failed against repoRoot=${repoRoot}:`,\n    ...lines,\n    '',\n    'Fix the manifest: every required surface must point at an existing',\n    'directory (systemPrompt / tools / personas / knowledge) or file',\n    '(rubric). Optional surfaces (scaffolding / memory / rag / outputSchema)',\n    'may be omitted; the loop will reject findings targeting omitted',\n    'surfaces rather than fabricating a path.',\n  ].join('\\n')\n}\n","/**\n * `defineAgent` — typed, validated manifest API for declarative agent\n * configuration. The substrate consumes this manifest to wire the\n * canonical eval pattern + analyst self-improvement loop without any\n * per-vertical glue.\n *\n * Design goal: scale to 1000s of vertical agents. Every agent declares\n * its surfaces, rubric, runtime, and analyst configuration in ~50 lines.\n * No per-vertical `ImprovementAdapter`. No per-vertical CLI. No\n * fabricated paths.\n *\n * Validation: `defineAgent` runs `validateSurfaces` synchronously and\n * throws a structured error if any required surface is missing on\n * disk. The cost is one filesystem stat per surface (cheap); the\n * benefit is a manifest that can't ship broken.\n */\n\nimport type { TraceAnalystKindSpec } from '@tangle-network/agent-eval'\nimport type { RuntimeStreamEvent } from '../types'\nimport { type AgentSurfaces, renderSurfaceIssues, validateSurfaces } from './surfaces'\n\n// ── manifest ─────────────────────────────────────────────────────────\n\n/**\n * The full agent manifest. Each agent ships ONE of these.\n *\n * Generics:\n *   `TPersona` — the agent's persona shape (loaded from\n *     `surfaces.personas`). Defaults to `unknown` so the substrate's\n *     persona discovery (`loadPersonas`) can accept anything; per-agent\n *     code re-narrows when it matters.\n *   `TRunOutput` — the shape `runtime.act` returns. Used by the rubric\n *     scorers and emitted into the trace.\n */\nexport interface AgentManifest<TPersona = unknown, TRunOutput = unknown> {\n  /**\n   * Stable identifier — used as `projectId` in traces, as the analyst\n   * loop's `runId` prefix, and as the namespace under which findings\n   * are persisted. MUST match the agent's repo name to keep\n   * cross-repo telemetry joinable.\n   */\n  id: string\n\n  /**\n   * Filesystem root the substrate resolves surface paths against.\n   * Typically `process.cwd()` or a fixed absolute path. Use an\n   * absolute path when the agent's tests may run from subdirectories\n   * (vitest sometimes shifts cwd).\n   */\n  repoRoot: string\n\n  /**\n   * Map of mutable surfaces the self-improvement loop can edit. See\n   * `AgentSurfaces` — required: `systemPrompt`, `tools`, `rubric`,\n   * `knowledge`, `personas`. Optional: `scaffolding`, `memory`, `rag`,\n   * `outputSchema`.\n   *\n   * Every required path is validated at `defineAgent` time. Missing\n   * paths throw with the full list of offenders.\n   */\n  surfaces: AgentSurfaces\n\n  /**\n   * Rubric the substrate uses to score each run. Dimensions × weights\n   * × judges. The substrate computes the weighted composite and\n   * stamps it into the RunRecord.\n   */\n  rubric: AgentRubric<TRunOutput>\n\n  /**\n   * Runtime adapter — how the substrate INVOKES the agent against a\n   * persona. The `act` function takes a persona + a context (with the\n   * tracer the substrate threads through for span emission) and\n   * returns the run output the rubric will score.\n   *\n   * The agent's existing production runtime goes in here; the\n   * substrate is intentionally thin around it.\n   */\n  runtime: AgentRuntime<TPersona, TRunOutput>\n\n  /**\n   * Persona discovery — the substrate loads personas via this function\n   * at eval start. Can read from `surfaces.personas`, an API, or be\n   * hardcoded. The substrate calls it once per `runAgentEval` call;\n   * persona ordering is preserved.\n   */\n  personas: () => Promise<ReadonlyArray<TPersona>>\n\n  /**\n   * Analyst kinds the substrate runs against each persona's trace.\n   * Defaults to `DEFAULT_TRACE_ANALYST_KINDS` from agent-eval. Per-agent\n   * authors can prune (e.g. skip `knowledge-poisoning` when there's no\n   * knowledge base) or extend (custom domain kinds).\n   *\n   * Empty array disables the loop — useful for `pnpm eval --no-analyst`.\n   */\n  analystKinds: ReadonlyArray<TraceAnalystKindSpec>\n\n  /**\n   * Analyst LLM configuration. The substrate uses these for all four\n   * kinds (override per-kind via `analystKinds` if needed).\n   */\n  analyst: AnalystConfig\n\n  /**\n   * Auto-apply policy. Knowledge / improvement edits land only when\n   * `enabled === true` AND the source finding's confidence meets the\n   * threshold. `mode` controls how applies happen: `'write'` mutates\n   * files in-place; `'open-pr'` writes to a branch and opens a PR.\n   *\n   * Default: knowledge auto-applies at confidence ≥0.85 in `'write'`\n   * mode (wiki edits are git-reversible); improvement stays at\n   * `enabled: false` until the agent author has measured precision.\n   */\n  autoApply?: AutoApplyPolicy\n}\n\nexport interface AgentRubric<TRunOutput> {\n  /** Dimensions composing the weighted score. Weights sum to 1.0 by convention. */\n  dimensions: ReadonlyArray<RubricDimension<TRunOutput>>\n  /**\n   * Optional judges layered on top of deterministic dimensions. Each\n   * judge returns a score per dimension; the substrate averages judges\n   * (mean by default) for the LLM contribution.\n   */\n  judges?: ReadonlyArray<JudgeConfig<TRunOutput>>\n}\n\nexport interface RubricDimension<TRunOutput> {\n  /** Unique identifier — appears in finding subjects (`rubric:<id>`). */\n  id: string\n  /** 0..1 — weight in the composite. */\n  weight: number\n  /**\n   * Deterministic scorer: given the persona + run output, returns a\n   * 0..1 score. The substrate sums weight × score across dimensions\n   * for the deterministic composite; judges supplement subjective dims.\n   */\n  score: (input: { persona: unknown; output: TRunOutput }) => number\n  /** Optional human-readable label for reports. */\n  label?: string\n}\n\nexport interface JudgeConfig<TRunOutput> {\n  /** Judge identifier — appears in trace spans + manifest. */\n  id: string\n  /** Model snapshot to invoke. Pin the snapshot (`claude-sonnet-4-6@2025-04-15`); the validator rejects bare aliases. */\n  model: string\n  /** Dimensions this judge scores. */\n  dimensions: ReadonlyArray<string>\n  /**\n   * Optional rubric anchors — text examples the judge sees as a\n   * few-shot prompt to calibrate. STRONGLY recommended for subjective\n   * dimensions; required by the calibration gate (Pearson ≥0.7).\n   */\n  anchors?: ReadonlyArray<{ input: string; output: TRunOutput; expected: Record<string, number> }>\n}\n\nexport interface AgentRuntime<TPersona, TRunOutput> {\n  /**\n   * Invoke the agent against one persona. Returns BOTH:\n   *   - `events`: an `AsyncIterable<RuntimeStreamEvent>` the chat-centric\n   *     product consumes verbatim (SSE / WebSocket / inline render).\n   *     **Streaming is mandatory — never collapse this to a single Promise.**\n   *     The agent's existing `runChatTurn` (or equivalent async generator)\n   *     plugs in here directly.\n   *   - `output`: a `Promise<TRunOutput>` resolved AFTER the event stream\n   *     drains. The eval substrate awaits this for rubric scoring; chat\n   *     products usually ignore it (they already rendered incrementally).\n   *\n   * Implementation contract:\n   *   1. `act` MUST return immediately (synchronous construction of the\n   *      `events` iterator + the `output` promise).\n   *   2. Iterating `events` drives the underlying LLM/tool calls — the\n   *      caller chooses when to consume.\n   *   3. `output` resolves only after the iterator yields its terminal\n   *      event (typically `task_end`); see `collectAgentRun` helper.\n   *\n   * `ctx.emitter` is the substrate-threaded `TraceEmitter` — runtimes\n   * SHOULD record LLM/tool spans through it for capture integrity.\n   * `ctx.deadlineMs` is wall-clock; the runtime SHOULD honour for graceful\n   * cancel. `ctx.signal` is the standard abort signal.\n   */\n  act: (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput>\n}\n\nexport interface AgentRunInvocation<TRunOutput> {\n  /** Live stream of typed runtime events. Consumed by chat UX directly. */\n  events: AsyncIterable<RuntimeStreamEvent>\n  /** Final structured output the rubric scores. Resolves after `events` drains. */\n  output: Promise<TRunOutput>\n}\n\n/**\n * Stub for agents whose `runtime.act` is not yet wired to the substrate's\n * eval path. Preserves the streaming contract (empty event stream + a\n * rejected `output` promise that tells the caller exactly what to fix).\n *\n * Per-vertical manifests usually start with this stub and replace it with\n * the agent's real streaming runtime (`runChatTurn` or equivalent) once\n * the eval path consumes the manifest end-to-end.\n */\nexport function unimplementedAgentRun<TRunOutput = unknown>(\n  reason = 'AgentRuntime.act is not yet wired for this manifest',\n): AgentRunInvocation<TRunOutput> {\n  return {\n    events: (async function* empty(): AsyncIterable<RuntimeStreamEvent> {})(),\n    output: Promise.reject(new Error(reason)),\n  }\n}\n\n/**\n * Drain `act`'s `events` into an array AND await its `output`. Useful for\n * eval / outcome-measurement code paths that don't care about live\n * rendering. The events array is preserved so the substrate can inspect\n * tool calls / readiness / questions retrospectively.\n *\n * IMPORTANT: chat-centric UX MUST NOT call this — it defeats streaming\n * (no incremental render). Use `for await (const ev of invocation.events)`\n * directly in the chat surface.\n */\nexport async function collectAgentRun<TRunOutput>(\n  invocation: AgentRunInvocation<TRunOutput>,\n): Promise<{ events: ReadonlyArray<RuntimeStreamEvent>; output: TRunOutput }> {\n  const events: RuntimeStreamEvent[] = []\n  for await (const ev of invocation.events) events.push(ev)\n  const output = await invocation.output\n  return { events, output }\n}\n\nexport interface AgentRunContext {\n  /** Substrate-managed trace emitter. */\n  emitter: import('@tangle-network/agent-eval').TraceEmitter\n  /** Stable run id for this persona × variant cell. */\n  runId: string\n  /** Variant the runtime is exercising (e.g. `'baseline'`, `'source-grounded'`). */\n  variantId?: string\n  /** Wall-clock deadline (epoch ms). The runtime SHOULD honour for graceful cancel. */\n  deadlineMs?: number\n  /** Optional abort signal. */\n  signal?: AbortSignal\n}\n\nexport interface AnalystConfig {\n  /** Model the analyst kinds use. Override per-kind via `analystKinds[i].cost.models`. */\n  model: string\n  /** Optional total budget across all kinds for one run. Substrate enforces via `BudgetGuard`. */\n  budgetUsd?: number\n  /** Backend hint for the AxAIService factory — same shape every kind uses. */\n  backend?: {\n    name?: 'openai' | 'router'\n    apiKey?: string\n    baseUrl?: string\n  }\n}\n\nexport interface AutoApplyPolicy {\n  knowledge?: {\n    enabled: boolean\n    confidenceThreshold?: number\n    mode?: 'write' | 'open-pr'\n  }\n  improvement?: {\n    enabled: boolean\n    confidenceThreshold?: number\n    mode?: 'write' | 'open-pr'\n  }\n}\n\n// ── factory + validation ─────────────────────────────────────────────\n\nexport class AgentManifestError extends Error {\n  constructor(\n    message: string,\n    public readonly agentId: string,\n    public readonly issues: ReadonlyArray<unknown> = [],\n  ) {\n    super(message)\n    this.name = 'AgentManifestError'\n  }\n}\n\n/**\n * Construct a validated agent manifest. Throws `AgentManifestError`\n * if any required surface is missing on disk.\n *\n * Generics: pass your persona / output types if you want narrowed\n * `runtime.act` signatures:\n *   `defineAgent<TaxPersona, TaxRunOutput>({ ... })`\n *\n * Most callers don't need the generics — the substrate operates on\n * `unknown` payloads internally and the manifest's `score` /\n * `runtime.act` see the typed shapes via TypeScript inference at\n * the call site.\n */\nexport function defineAgent<TPersona = unknown, TRunOutput = unknown>(\n  manifest: AgentManifest<TPersona, TRunOutput>,\n): AgentManifest<TPersona, TRunOutput> {\n  if (!manifest.id || manifest.id.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `id` is required', manifest.id ?? '')\n  }\n  if (!manifest.repoRoot || manifest.repoRoot.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `repoRoot` is required', manifest.id)\n  }\n  const issues = validateSurfaces(manifest.surfaces, manifest.repoRoot)\n  if (issues.length > 0) {\n    throw new AgentManifestError(\n      renderSurfaceIssues(issues, manifest.repoRoot),\n      manifest.id,\n      issues,\n    )\n  }\n  // Lightweight rubric sanity: weights sum to ~1.0 (no hard requirement —\n  // the substrate normalizes — but flag wildly miscalibrated weights).\n  const total = manifest.rubric.dimensions.reduce((acc, d) => acc + d.weight, 0)\n  if (manifest.rubric.dimensions.length > 0 && (total < 0.5 || total > 1.5)) {\n    throw new AgentManifestError(\n      `defineAgent(${manifest.id}): rubric dimension weights sum to ${total.toFixed(3)} — should be ~1.0`,\n      manifest.id,\n    )\n  }\n  return manifest\n}\n","/**\n * Substrate-default `ImprovementAdapter` — surfaces-driven, LLM-drafted\n * patches, optional auto-apply or PR-open.\n *\n * This is the one ImprovementAdapter every vertical agent uses. The\n * substrate parses each finding's `subject` via\n * `parseFindingSubject` (agent-eval), resolves it to a real file path\n * via the agent's `AgentSurfaces`, reads the current content, and asks\n * an LLM to draft a unified-diff patch given the finding + current\n * content + per-kind editing-discipline rules.\n *\n * Auto-apply gates on the source-finding's confidence and the\n * autoApply.improvement policy. Two modes:\n *   `write` — apply the patch in-place via `git apply -p0`. Operator\n *     reviews via `git diff`.\n *   `open-pr` — write to a branch, commit, push, open a PR via `gh`.\n *     Operator reviews via the PR UI.\n *\n * Fail-loud rules:\n *   - Findings whose subject doesn't parse → counted in `errors`.\n *   - Findings whose subject targets an undeclared surface → counted in\n *     `errors` with the offending kind in the message.\n *   - Findings whose target path doesn't exist AND the kind isn't a\n *     create-new variant (`new-tool`, `knowledge.wiki`) → counted in\n *     `errors` with the resolved path in the message.\n *   - LLM drafts that fail JSON-schema validation → counted in\n *     `errors` with the schema issue.\n *\n * No silent skips. Every dropped finding has a recorded reason the\n * loop's report surfaces.\n */\n\nimport { spawnSync } from 'node:child_process'\nimport { readFileSync } from 'node:fs'\nimport type { AnalystFinding, FindingSubject } from '@tangle-network/agent-eval'\nimport { parseFindingSubject } from '@tangle-network/agent-eval/analyst'\nimport type { ImprovementAdapter } from '../analyst-loop/types'\nimport type { AgentSurfaces, ResolvedSurface } from './surfaces'\nimport { resolveSubjectPath } from './surfaces'\n\n// ── proposal shape ───────────────────────────────────────────────────\n\nexport interface SurfaceImprovementEdit {\n  /** Stable id derived from the source finding so re-proposals are idempotent. */\n  id: string\n  /** The finding that produced this edit — for revert + audit trail. */\n  sourceFindingId: string\n  /** Parsed subject; included so the apply step doesn't re-parse. */\n  subject: FindingSubject\n  /** Resolved on-disk target. */\n  target: ResolvedSurface\n  /** SHA-256 of the current file content the patch was drafted against. */\n  baseSha256: string\n  /** Unified-diff patch the LLM drafted (relative to `target.absolutePath`). */\n  patch: string\n  /** One-line summary the operator sees in the report / PR title. */\n  summary: string\n  /** Multi-line rationale for the PR body — finding context + LLM reasoning. */\n  rationale: string\n  /** Carry-forward from the finding so the apply gate can check the threshold. */\n  confidence: number\n  /** Carry-forward severity for prioritization. */\n  severity: AnalystFinding['severity']\n}\n\nexport interface CreateSurfaceImprovementAdapterOpts {\n  surfaces: AgentSurfaces\n  repoRoot: string\n  /**\n   * LLM-draft callback. Given a finding + current file content + the\n   * resolved target, returns a unified-diff patch + summary + rationale.\n   *\n   * Required — the substrate doesn't ship a hardcoded prompt; the agent\n   * author picks the model (Haiku for cheap routine drafts, Sonnet for\n   * substantive prompt rewrites, etc.) via this callback.\n   */\n  draftPatch: (input: DraftPatchInput) => Promise<DraftPatchOutput>\n  /**\n   * Apply mode:\n   *   `write` — `git apply` in-place; operator reviews via `git diff`\n   *   `open-pr` — branch + commit + push + `gh pr create`\n   *   `none` — never apply; collect proposals for the report only\n   *\n   * The `apply` method honours this even when the loop calls it; the\n   * effective behaviour is also gated on the per-finding confidence\n   * threshold via `runAnalystLoop`'s `autoApply` policy.\n   */\n  mode?: 'write' | 'open-pr' | 'none'\n  /** When `mode === 'open-pr'`, the base branch new PRs target. Default: `main`. */\n  baseBranch?: string\n  /** Required for `mode === 'open-pr'` — the GH owner/repo (`tangle-network/tax-agent`). */\n  ghRepo?: string\n  /**\n   * When the resolved target doesn't exist, allow the substrate to\n   * CREATE the file (for `knowledge.wiki`, `new-tool` subjects). Default\n   * true for those kinds, false for `system-prompt` / `rubric` / etc.\n   * (named sections that don't exist are a contract violation, not a\n   * scaffolding opportunity).\n   */\n  allowCreateForKinds?: ReadonlyArray<FindingSubject['kind']>\n}\n\nexport interface DraftPatchInput {\n  finding: AnalystFinding\n  subject: FindingSubject\n  target: ResolvedSurface\n  /** Current file content (empty string when `intent === 'create-new'`). */\n  currentContent: string\n}\n\nexport interface DraftPatchOutput {\n  /** Unified diff against the current file content. Empty string skips this finding. */\n  patch: string\n  /** One-line summary for the operator. */\n  summary: string\n  /** Multi-line rationale for the PR body. */\n  rationale: string\n}\n\n// ── factory ──────────────────────────────────────────────────────────\n\nconst DEFAULT_CREATE_KINDS: ReadonlyArray<FindingSubject['kind']> = [\n  'knowledge.wiki',\n  'knowledge.claim',\n  'knowledge.raw',\n  'new-tool',\n]\n\nexport function createSurfaceImprovementAdapter(\n  opts: CreateSurfaceImprovementAdapterOpts,\n): ImprovementAdapter<SurfaceImprovementEdit> {\n  const mode = opts.mode ?? 'none'\n  const allowCreate = opts.allowCreateForKinds ?? DEFAULT_CREATE_KINDS\n\n  return {\n    async proposeFromFindings(findings) {\n      const edits: SurfaceImprovementEdit[] = []\n      const errors: Array<{ findingId: string; subject: string; message: string }> = []\n      let skipped = 0\n\n      for (const f of findings) {\n        const subject = parseFindingSubject(f.subject)\n        if (subject === null) {\n          if (f.subject !== undefined) {\n            errors.push({\n              findingId: f.finding_id,\n              subject: f.subject,\n              message: 'subject does not parse against the finding-subject grammar',\n            })\n          } else {\n            // Subject-less findings are descriptive, not actionable —\n            // legitimate; count in `skipped` not `errors`.\n            skipped += 1\n          }\n          continue\n        }\n\n        // `cluster` findings (failure-mode) are evidence, not mutations.\n        if (subject.kind === 'cluster') {\n          skipped += 1\n          continue\n        }\n\n        // `agent-knowledge:*` findings flow to the KnowledgeAdapter;\n        // the ImprovementAdapter skips them so subjects don't double-route.\n        if (subject.kind.startsWith('knowledge.')) {\n          skipped += 1\n          continue\n        }\n\n        const target = resolveSubjectPath(subject, opts.surfaces, opts.repoRoot)\n        if (target === null) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `subject kind \"${subject.kind}\" targets an undeclared surface; declare it in AgentSurfaces or stop emitting this subject`,\n          })\n          continue\n        }\n\n        if (target.intent === 'create-new' && !allowCreate.includes(subject.kind)) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `target ${target.repoRelativePath} does not exist; the kind \"${subject.kind}\" requires an existing target (analyst named a section that isn't in the codebase)`,\n          })\n          continue\n        }\n\n        const currentContent = target.exists ? readFileSync(target.absolutePath, 'utf-8') : ''\n\n        let draft: DraftPatchOutput\n        try {\n          draft = await opts.draftPatch({ finding: f, subject, target, currentContent })\n        } catch (err) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `draftPatch threw: ${err instanceof Error ? err.message : String(err)}`,\n          })\n          continue\n        }\n\n        if (draft.patch.trim().length === 0) {\n          skipped += 1\n          continue\n        }\n\n        edits.push({\n          id: `imp-${f.finding_id}`,\n          sourceFindingId: f.finding_id,\n          subject,\n          target,\n          baseSha256: sha256(currentContent),\n          patch: draft.patch,\n          summary: draft.summary,\n          rationale: draft.rationale,\n          confidence: f.confidence,\n          severity: f.severity,\n        })\n      }\n\n      return { edits, skipped, errors }\n    },\n\n    async apply(edits) {\n      const applied: string[] = []\n      const warnings: string[] = []\n\n      if (mode === 'none') {\n        warnings.push(\n          'createSurfaceImprovementAdapter: mode=none; no edits applied — adjust manifest.autoApply.improvement.mode',\n        )\n        return { applied, warnings }\n      }\n\n      if (mode === 'open-pr' && !opts.ghRepo) {\n        warnings.push(\n          'createSurfaceImprovementAdapter: mode=open-pr requires `ghRepo`; falling back to no-op',\n        )\n        return { applied, warnings }\n      }\n\n      for (const edit of edits) {\n        // Race-detection: confirm the file content hasn't moved since the\n        // patch was drafted. A diff applied against drifted content is a\n        // recipe for silent corruption.\n        const current = edit.target.exists ? readFileSync(edit.target.absolutePath, 'utf-8') : ''\n        if (sha256(current) !== edit.baseSha256) {\n          warnings.push(\n            `${edit.target.repoRelativePath}: base SHA mismatch; file changed after draft. Skipping.`,\n          )\n          continue\n        }\n\n        const ok = applyPatchInPlace(edit, opts.repoRoot)\n        if (!ok) {\n          warnings.push(`${edit.target.repoRelativePath}: git apply failed`)\n          continue\n        }\n        applied.push(edit.target.repoRelativePath)\n      }\n\n      if (mode === 'open-pr' && applied.length > 0 && opts.ghRepo) {\n        const prUrl = openPullRequest(\n          applied,\n          edits.filter((e) => applied.includes(e.target.repoRelativePath)),\n          opts.repoRoot,\n          opts.ghRepo,\n          opts.baseBranch ?? 'main',\n        )\n        if (prUrl) warnings.push(`opened PR: ${prUrl}`)\n        else warnings.push('PR creation failed; edits are committed to a local branch only')\n      }\n\n      return { applied, warnings }\n    },\n  }\n}\n\n// ── apply helpers ────────────────────────────────────────────────────\n\nfunction applyPatchInPlace(edit: SurfaceImprovementEdit, repoRoot: string): boolean {\n  const result = spawnSync('git', ['apply', '--whitespace=fix', '-p0', '-'], {\n    cwd: repoRoot,\n    input: edit.patch,\n    encoding: 'utf-8',\n  })\n  return result.status === 0\n}\n\nfunction openPullRequest(\n  paths: ReadonlyArray<string>,\n  edits: ReadonlyArray<SurfaceImprovementEdit>,\n  repoRoot: string,\n  ghRepo: string,\n  baseBranch: string,\n): string | null {\n  const branch = `analyst-loop/${Date.now()}-${edits[0]?.sourceFindingId.slice(0, 12) ?? 'edits'}`\n  // Create branch, stage, commit\n  const checkout = spawnSync('git', ['checkout', '-b', branch], { cwd: repoRoot })\n  if (checkout.status !== 0) return null\n  const add = spawnSync('git', ['add', ...paths], { cwd: repoRoot })\n  if (add.status !== 0) return null\n  const title = `analyst-loop: ${edits[0]?.summary ?? `${edits.length} improvement edits`}`\n  const body = [\n    `Automated analyst-loop edits — review carefully before merge.`,\n    '',\n    `Source findings:`,\n    ...edits.map(\n      (e) =>\n        `  - ${e.sourceFindingId} (confidence ${e.confidence.toFixed(2)}, severity ${e.severity})`,\n    ),\n    '',\n    'Rationales:',\n    ...edits.map((e) => `\\n## ${e.target.repoRelativePath}\\n\\n${e.rationale}`),\n  ].join('\\n')\n  const commit = spawnSync('git', ['commit', '-m', title, '-m', body], { cwd: repoRoot })\n  if (commit.status !== 0) return null\n  const push = spawnSync('git', ['push', '-u', 'origin', branch], { cwd: repoRoot })\n  if (push.status !== 0) return null\n  const pr = spawnSync(\n    'gh',\n    [\n      'pr',\n      'create',\n      '--repo',\n      ghRepo,\n      '--title',\n      title,\n      '--body',\n      body,\n      '--base',\n      baseBranch,\n      '--head',\n      branch,\n    ],\n    { cwd: repoRoot, encoding: 'utf-8' },\n  )\n  if (pr.status !== 0) return null\n  return pr.stdout.trim()\n}\n\nfunction sha256(s: string): string {\n  // node:crypto is dynamic-imported lazily so the adapter can be tested in\n  // environments without crypto (browser tests, mocked envs).\n  const crypto = require('node:crypto') as typeof import('node:crypto')\n  return crypto.createHash('sha256').update(s, 'utf-8').digest('hex')\n}\n","/**\n * Substrate-default `KnowledgeAdapter` — wraps agent-knowledge's\n * `proposeFromFindings` + `applyKnowledgeWriteBlocks` with substrate\n * defaults (auto-lint after apply, source linkage via finding id).\n *\n * Every agent that ships a `.agent-knowledge/` tree uses this adapter\n * unmodified. Per-agent customization happens at the manifest level\n * (`autoApply.knowledge.confidenceThreshold`, etc.), not by writing a\n * new adapter.\n *\n * Lint discipline: after each apply we run agent-knowledge's\n * `lintKnowledgeIndex` to catch broken links / circular claims /\n * duplicate pages introduced by the new writes. Findings that fail the\n * post-apply lint are recorded in `warnings`; the apply itself is not\n * rolled back (lint failures are soft — humans review the wiki state).\n */\n\nimport type { AnalystFinding } from '@tangle-network/agent-eval'\nimport type { KnowledgeAdapter } from '../analyst-loop/types'\n\nexport interface CreateSurfaceKnowledgeAdapterOpts {\n  /** `.agent-knowledge/` root (absolute path the substrate writes blocks against). */\n  knowledgeRoot: string\n}\n\n/**\n * Build the adapter. We accept the agent-knowledge functions as DI so\n * the substrate stays decoupled from a specific agent-knowledge\n * version — the agent author imports them in their manifest module\n * and hands them to the factory.\n *\n * `proposeFromFindings(findings)` returns\n *   `{ proposals: KnowledgeProposal[]; skipped: number; errors: ... }`.\n *\n * `applyKnowledgeWriteBlocks(root, content)` returns\n *   `{ written: string[]; warnings: string[] }`.\n *\n * `lintKnowledgeIndex(index)` (optional) returns `KnowledgeLintFinding[]`.\n */\nexport interface KnowledgeAdapterDeps<TProposal> {\n  proposeFromFindings: (findings: ReadonlyArray<AnalystFinding>) => {\n    proposals: TProposal[]\n    skipped: number\n    errors: Array<{ findingId: string; subject: string; message: string }>\n  }\n  applyKnowledgeWriteBlocks: (\n    root: string,\n    proposalText: string,\n  ) => Promise<{ written: string[]; warnings: string[] }>\n  /**\n   * Optional post-apply lint hook. The substrate runs it after each\n   * batch of writes; failures land in `warnings` (the apply is not\n   * rolled back — lint signals drift to review, not block).\n   */\n  lintAfterApply?: (root: string) => Promise<ReadonlyArray<string>>\n}\n\nexport function createSurfaceKnowledgeAdapter<TProposal>(\n  opts: CreateSurfaceKnowledgeAdapterOpts,\n  deps: KnowledgeAdapterDeps<TProposal>,\n): KnowledgeAdapter<TProposal> {\n  return {\n    proposeFromFindings(findings) {\n      const batch = deps.proposeFromFindings(findings)\n      return {\n        proposals: batch.proposals,\n        skipped: batch.skipped,\n        errors: batch.errors,\n      }\n    },\n    async apply(proposals) {\n      const written: string[] = []\n      const warnings: string[] = []\n      for (const p of proposals) {\n        const proposalText = renderProposalAsWriteBlock(p)\n        if (proposalText === null) {\n          warnings.push(\n            `proposal has no writeBlocks/content; skipping (sourceFindingId=${getSourceFindingId(p)})`,\n          )\n          continue\n        }\n        try {\n          const r = await deps.applyKnowledgeWriteBlocks(opts.knowledgeRoot, proposalText)\n          written.push(...r.written)\n          warnings.push(...r.warnings)\n        } catch (err) {\n          warnings.push(\n            `applyKnowledgeWriteBlocks failed: ${err instanceof Error ? err.message : String(err)}`,\n          )\n        }\n      }\n      if (deps.lintAfterApply && written.length > 0) {\n        try {\n          const lintIssues = await deps.lintAfterApply(opts.knowledgeRoot)\n          for (const issue of lintIssues) warnings.push(`lint: ${issue}`)\n        } catch (err) {\n          warnings.push(\n            `lintAfterApply failed: ${err instanceof Error ? err.message : String(err)}`,\n          )\n        }\n      }\n      return { written, warnings }\n    },\n  }\n}\n\n/**\n * Pluck the canonical write-block text from a proposal regardless of\n * which exact agent-knowledge version produced it. We accept either:\n *   - `{ writeBlocks: Array<{ path, content }> }` — the typed shape\n *     1.3.0+ emits\n *   - `{ proposalText: string }` — legacy single-block shape\n *   - `{ content: string }` — minimal raw form\n *\n * Returns `null` when nothing parseable is present (warned upstream).\n */\nfunction renderProposalAsWriteBlock(p: unknown): string | null {\n  if (!p || typeof p !== 'object') return null\n  const obj = p as Record<string, unknown>\n  if (Array.isArray(obj.writeBlocks)) {\n    const blocks = obj.writeBlocks as Array<{ path?: string; content?: string }>\n    if (blocks.length === 0) return null\n    return blocks\n      .map((b) => (typeof b.content === 'string' ? b.content : ''))\n      .filter((s) => s.length > 0)\n      .join('\\n\\n')\n  }\n  if (typeof obj.proposalText === 'string') return obj.proposalText\n  if (typeof obj.content === 'string') return obj.content\n  return null\n}\n\nfunction getSourceFindingId(p: unknown): string {\n  if (!p || typeof p !== 'object') return '<unknown>'\n  const obj = p as Record<string, unknown>\n  if (typeof obj.sourceFindingId === 'string') return obj.sourceFindingId\n  if (typeof obj.id === 'string') return obj.id\n  return '<unknown>'\n}\n","/**\n * `OutcomeMeasurement` — the missing metric that turns the analyst\n * loop from \"observability\" into \"self-improvement\".\n *\n * Without this hook, the loop reports process counts (`findings: 42`,\n * `applied: 7`) and never proves the applied edits actually improved\n * anything. With this hook, the substrate re-runs the cohort against\n * the same personas after each apply pass and reports a composite\n * score delta. A negative delta is the substrate's strongest signal\n * to either roll back or surface for review.\n *\n * Wiring is intentionally simple: pass the manifest + the `runAgentEval`\n * function and a list of `personaIds` to re-run. The wrapper:\n *   1. Captures the baseline composite from the just-finished run.\n *   2. After `runAnalystLoop` returns, re-invokes `runAgentEval` against\n *      the same persona slice.\n *   3. Computes the delta and appends to `loop-report.json`.\n *   4. If `rollbackOnRegression` and delta < 0, reverts applied edits.\n */\n\nimport type { RunAnalystLoopResult } from '../analyst-loop/types'\n\nexport interface OutcomeMeasurement {\n  /** Baseline composite before applies — captured from the most-recent eval run. */\n  baselineComposite: number\n  /** Composite after re-running the cohort with applied edits. */\n  afterComposite: number\n  /** `afterComposite - baselineComposite`. Positive = the loop improved the agent. */\n  delta: number\n  /** Per-persona deltas for finer-grained review. */\n  perPersona: ReadonlyArray<{ personaId: string; before: number; after: number; delta: number }>\n  /** When the substrate rolled back applies due to regression, the paths reverted. */\n  rolledBackPaths: ReadonlyArray<string>\n}\n\nexport interface OutcomeMeasurementOpts {\n  /** Composite scores from the run that produced the findings. */\n  baseline: ReadonlyArray<{ personaId: string; composite: number }>\n  /**\n   * Re-run callback — the substrate invokes this after applies. The\n   * agent author provides their `runAgentEval`-equivalent so the\n   * substrate can ask \"score this persona slice now.\"\n   *\n   * The callback SHOULD reuse the same cohort + judges + variant as\n   * the baseline run; only the agent's mutable surfaces have changed.\n   */\n  reRunCohort: (\n    personaIds: ReadonlyArray<string>,\n  ) => Promise<ReadonlyArray<{ personaId: string; composite: number }>>\n  /** When `true`, applied edits are reverted on negative delta. Default `false`. */\n  rollbackOnRegression?: boolean\n  /** Callback to revert a list of paths (typically `git checkout HEAD --`). */\n  revert?: (paths: ReadonlyArray<string>) => Promise<void>\n}\n\n/**\n * Run `runAnalystLoop` and stamp an `OutcomeMeasurement` onto the\n * result. The substrate calls this after each canonical eval; the\n * delta lands in `loop-report.json` for cross-run trend analysis.\n *\n * The function returns the original `RunAnalystLoopResult` enriched\n * with `outcome` so callers stay backwards-compatible (the field is\n * optional on the type; missing means no measurement was wired).\n */\nexport async function measureOutcome<TProposal, TEdit>(\n  result: RunAnalystLoopResult<TProposal, TEdit>,\n  opts: OutcomeMeasurementOpts,\n): Promise<RunAnalystLoopResult<TProposal, TEdit> & { outcome: OutcomeMeasurement }> {\n  const applied = result.knowledge?.applied ?? []\n  const improvementsApplied = result.improvement?.applied ?? []\n  const allApplied = [...applied, ...improvementsApplied]\n\n  // No applies → no outcome to measure. Return a zero-delta to keep the\n  // shape stable for consumers; baseline / after equal.\n  if (allApplied.length === 0) {\n    return {\n      ...result,\n      outcome: {\n        baselineComposite: meanComposite(opts.baseline),\n        afterComposite: meanComposite(opts.baseline),\n        delta: 0,\n        perPersona: opts.baseline.map((b) => ({\n          personaId: b.personaId,\n          before: b.composite,\n          after: b.composite,\n          delta: 0,\n        })),\n        rolledBackPaths: [],\n      },\n    }\n  }\n\n  const personaIds = opts.baseline.map((b) => b.personaId)\n  const after = await opts.reRunCohort(personaIds)\n  const afterByPersona = new Map(after.map((r) => [r.personaId, r.composite]))\n\n  const perPersona = opts.baseline.map((b) => {\n    const a = afterByPersona.get(b.personaId) ?? b.composite\n    return { personaId: b.personaId, before: b.composite, after: a, delta: a - b.composite }\n  })\n  const baselineComposite = meanComposite(opts.baseline)\n  const afterComposite = meanComposite(after)\n  const delta = afterComposite - baselineComposite\n\n  let rolledBackPaths: string[] = []\n  if (delta < 0 && opts.rollbackOnRegression && opts.revert) {\n    await opts.revert(allApplied)\n    rolledBackPaths = [...allApplied]\n  }\n\n  return {\n    ...result,\n    outcome: {\n      baselineComposite,\n      afterComposite,\n      delta,\n      perPersona,\n      rolledBackPaths,\n    },\n  }\n}\n\nfunction meanComposite(rows: ReadonlyArray<{ composite: number }>): number {\n  if (rows.length === 0) return 0\n  return rows.reduce((acc, r) => acc + r.composite, 0) / rows.length\n}\n","/**\n * Sandbox bridge for `AgentRuntime.act` — prod-faithful eval execution.\n *\n * The point of this adapter is parity: the eval substrate must run the agent\n * through the SAME profile the production chat turn uses, or scorecard numbers\n * grade a profile that never ships. `createSandboxAct` composes the production\n * profile via {@link composeProductionAgentProfile}, boots a sandbox with it\n * through the loop kernel's own {@link createSandboxForSpec}, streams the\n * `streamPrompt` events mapped to the `RuntimeStreamEvent` vocabulary, and\n * resolves the `OutputAdapter`-parsed output for rubric scoring — satisfying\n * the `act` streaming contract with one code path shared by chat and eval.\n *\n * Agents with a bespoke streaming chat turn should wire THAT into `act`\n * directly (the contract is designed for it); this adapter is the default for\n * agents whose turn is a plain prod-profile sandbox dispatch — notably the\n * agents agent-builder generates.\n */\n\nimport type { AgentProfile } from '@tangle-network/agent-interface'\nimport type { SandboxEvent } from '@tangle-network/sandbox'\nimport type { ComposeProductionAgentProfileOptions } from '../mcp/delegation-profile'\nimport { composeProductionAgentProfile } from '../mcp/delegation-profile'\nimport type { AgentRunSpec, OutputAdapter, SandboxClient } from '../runtime'\nimport { mapSandboxEvent } from '../runtime'\nimport { createSandboxForSpec } from '../runtime/run-loop'\nimport type { RuntimeStreamEvent } from '../types'\nimport type { AgentRunContext, AgentRunInvocation } from './define-agent'\n\nexport interface CreateSandboxActOptions<TPersona, TRunOutput> {\n  /** Canonical agent profile — the same one the prod chat turn composes from. */\n  baseProfile: AgentProfile\n  /** Sandbox client used to boot the per-run sandbox. */\n  sandboxClient: SandboxClient\n  /** Persona → prompt. Pure; the eval cell's input. */\n  buildPrompt: (persona: TPersona) => string\n  /** Sandbox event stream → typed output the rubric scores. */\n  output: OutputAdapter<TRunOutput>\n  /**\n   * Per-persona composition overrides (workspace-augmented system prompt,\n   * extra file mounts, sandbox key). Merged into\n   * {@link composeProductionAgentProfile}; `env` here is overridden by the\n   * top-level `env` option when both are set.\n   */\n  compose?: (persona: TPersona) => ComposeProductionAgentProfileOptions\n  /** Sandbox-SDK overrides forwarded to `createSandboxForSpec`. */\n  sandboxOverrides?: AgentRunSpec<unknown>['sandboxOverrides']\n  /** Stable run name surfaced in mapped `llm_call` events. */\n  name?: string\n  /** Override the `SandboxEvent → RuntimeStreamEvent` mapper. */\n  mapEvent?: (\n    event: SandboxEvent,\n    opts: { agentRunName?: string },\n  ) => RuntimeStreamEvent | undefined\n  /** Environment source for delegation-MCP composition. Defaults to `process.env`. */\n  env?: Record<string, string | undefined>\n}\n\n/**\n * Build an `AgentRuntime.act` implementation backed by a single prod-profile\n * sandbox run. The returned function honours the `act` contract: it returns\n * synchronously with a live `events` iterator and an `output` promise that\n * resolves only after the iterator drains.\n */\nexport function createSandboxAct<TPersona, TRunOutput>(\n  options: CreateSandboxActOptions<TPersona, TRunOutput>,\n): (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput> {\n  const mapEvent = options.mapEvent ?? mapSandboxEvent\n\n  return (persona: TPersona, ctx: AgentRunContext): AgentRunInvocation<TRunOutput> => {\n    const profile = composeProductionAgentProfile(options.baseProfile, {\n      ...(options.compose?.(persona) ?? {}),\n      ...(options.env ? { env: options.env } : {}),\n    })\n    const agentRunName = options.name ?? profile.name ?? 'agent'\n    const message = options.buildPrompt(persona)\n    const signal = ctx.signal ?? new AbortController().signal\n\n    const raw: SandboxEvent[] = []\n    let settle!: (value: TRunOutput) => void\n    let fail!: (err: unknown) => void\n    const output = new Promise<TRunOutput>((resolve, reject) => {\n      settle = resolve\n      fail = reject\n    })\n    // The output promise rejects when the stream errors; if the caller ignores\n    // `output` (chat UX) the rejection is still observed by the events iterator\n    // throwing. Attach a no-op catch so an ignored rejection is never \"unhandled\".\n    output.catch(() => {})\n\n    const spec: AgentRunSpec<unknown> = {\n      profile,\n      taskToPrompt: () => message,\n      name: agentRunName,\n      ...(options.sandboxOverrides ? { sandboxOverrides: options.sandboxOverrides } : {}),\n    }\n\n    async function* events(): AsyncIterable<RuntimeStreamEvent> {\n      try {\n        const box = await createSandboxForSpec(options.sandboxClient, spec, signal)\n        for await (const event of box.streamPrompt(message, { signal })) {\n          raw.push(event)\n          const mapped = mapEvent(event, { agentRunName })\n          if (mapped) yield mapped\n        }\n        settle(options.output.parse(raw))\n      } catch (err) {\n        fail(err)\n        throw err\n      }\n    }\n\n    return { events: events(), output }\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;;AAeA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,YAAY;AAqE1B,SAAS,mBACd,SACA,UACA,UACwB;AACxB,QAAM,aAAa,yBAAyB,SAAS,QAAQ;AAC7D,MAAI,WAAW,WAAW,EAAG,QAAO;AAMpC,aAAW,OAAO,YAAY;AAC5B,UAAM,MAAM,WAAW,GAAG,IAAI,MAAM,KAAK,UAAU,GAAG;AACtD,QAAI,WAAW,GAAG,GAAG;AACnB,aAAO,EAAE,cAAc,KAAK,kBAAkB,KAAK,QAAQ,MAAM,QAAQ,gBAAgB;AAAA,IAC3F;AAAA,EACF;AACA,QAAM,WAAW,WAAW,CAAC;AAC7B,QAAM,cAAc,WAAW,QAAQ,IAAI,WAAW,KAAK,UAAU,QAAQ;AAC7E,SAAO;AAAA,IACL,cAAc;AAAA,IACd,kBAAkB;AAAA,IAClB,QAAQ;AAAA,IACR,QAAQ;AAAA,EACV;AACF;AAEA,SAAS,yBACP,SACA,UACuB;AACvB,UAAQ,QAAQ,MAAM;AAAA,IACpB,KAAK;AAAA,IACL,KAAK;AACH,aAAO,CAAC,KAAK,SAAS,WAAW,GAAG,QAAQ,IAAI,KAAK,CAAC;AAAA,IACxD,KAAK;AAEH,aAAO,CAAC,KAAK,SAAS,WAAW,UAAU,GAAG,QAAQ,QAAQ,KAAK,CAAC,KAAK,CAAC;AAAA,IAC5E,KAAK;AACH,aAAO,CAAC,KAAK,SAAS,WAAW,OAAO,GAAG,QAAQ,QAAQ,KAAK,CAAC;AAAA,IACnE,KAAK,iBAAiB;AACpB,YAAM,OAAO,QAAQ,QAAQ,OAAO;AAIpC,aAAO;AAAA,QACL,KAAK,SAAS,cAAc,GAAG,IAAI,KAAK;AAAA,QACxC,KAAK,SAAS,cAAc,MAAM,UAAU;AAAA,QAC5C,KAAK,SAAS,cAAc,MAAM,UAAU;AAAA,MAC9C;AAAA,IACF;AAAA,IACA,KAAK;AACH,UAAI,QAAQ,QAAQ;AAClB,eAAO,CAAC,KAAK,SAAS,OAAO,QAAQ,MAAM,GAAG,QAAQ,QAAQ,MAAM,CAAC,KAAK,CAAC;AAAA,MAC7E;AAGA,aAAO;AAAA,QACL,KAAK,SAAS,OAAO,QAAQ,MAAM,WAAW;AAAA,QAC9C,KAAK,SAAS,OAAO,GAAG,QAAQ,IAAI,KAAK;AAAA,MAC3C;AAAA,IACF,KAAK;AACH,aAAO,CAAC,KAAK,SAAS,OAAO,QAAQ,MAAM,WAAW,CAAC;AAAA,IACzD,KAAK;AACH,UAAI,CAAC,SAAS,IAAK,QAAO,CAAC;AAC3B,aAAO,CAAC,KAAK,SAAS,KAAK,QAAQ,QAAQ,GAAG,QAAQ,KAAK,KAAK,CAAC;AAAA,IACnE,KAAK;AACH,UAAI,CAAC,SAAS,OAAQ,QAAO,CAAC;AAC9B,aAAO,CAAC,KAAK,SAAS,QAAQ,GAAG,QAAQ,QAAQ,GAAG,CAAC,OAAO,CAAC;AAAA,IAC/D,KAAK;AACH,UAAI,CAAC,SAAS,YAAa,QAAO,CAAC;AACnC,aAAO,CAAC,KAAK,SAAS,aAAa,GAAG,QAAQ,QAAQ,OAAO,CAAC,KAAK,CAAC;AAAA,IACtE,KAAK;AACH,UAAI,CAAC,SAAS,aAAc,QAAO,CAAC;AACpC,aAAO,CAAC,SAAS,YAAY;AAAA,IAC/B,KAAK;AAAA,IACL,KAAK;AAGH,aAAO,CAAC;AAAA,IACV,KAAK;AAEH,aAAO,CAAC;AAAA,EACZ;AACF;AAEA,SAAS,QAAQ,GAAmB;AAClC,SACE,EACG,YAAY,EACZ,QAAQ,gBAAgB,GAAG,EAC3B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,GAAG,KAAK;AAExB;AAgBO,SAAS,iBACd,UACA,UACuC;AACvC,QAAM,SAAmC,CAAC;AAC1C,QAAM,cAAkD;AAAA,IACtD;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,QAAM,eAAmD,CAAC,QAAQ;AAClE,QAAM,sBAA0D,CAAC,eAAe,UAAU,KAAK;AAC/F,QAAM,uBAA2D,CAAC,cAAc;AAEhF,aAAW,OAAO,aAAa;AAC7B,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,CAAC,GAAG;AACN,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,IAAI,QAAQ,UAAU,CAAC;AACzD;AAAA,IACF;AACA,UAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;AAChD,QAAI,CAAC,WAAW,GAAG,GAAG;AACpB,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,GAAG,QAAQ,UAAU,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,aAAW,OAAO,cAAc;AAC9B,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,CAAC,GAAG;AACN,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,IAAI,QAAQ,UAAU,CAAC;AACzD;AAAA,IACF;AACA,UAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;AAChD,QAAI,CAAC,WAAW,GAAG,GAAG;AACpB,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,GAAG,QAAQ,UAAU,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,aAAW,OAAO,CAAC,GAAG,qBAAqB,GAAG,oBAAoB,GAAG;AACnE,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,MAAM,OAAW;AACrB,UAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;AAChD,QAAI,CAAC,WAAW,GAAG,GAAG;AACpB,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,GAAG,QAAQ,UAAU,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,oBACd,QACA,UACQ;AACR,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,QAAM,QAAQ,OAAO;AAAA,IACnB,CAAC,MAAM,OAAO,EAAE,OAAO,KAAK,EAAE,OAAO,IAAI,EAAE,IAAI,MAAM,WAAW,KAAK,EAAE,MAAM;AAAA,EAC/E;AACA,SAAO;AAAA,IACL,oDAAoD,QAAQ;AAAA,IAC5D,GAAG;AAAA,IACH;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACb;;;AC5DO,SAAS,sBACd,SAAS,uDACuB;AAChC,SAAO;AAAA,IACL,SAAS,gBAAgB,QAA2C;AAAA,IAAC,GAAG;AAAA,IACxE,QAAQ,QAAQ,OAAO,IAAI,MAAM,MAAM,CAAC;AAAA,EAC1C;AACF;AAYA,eAAsB,gBACpB,YAC4E;AAC5E,QAAM,SAA+B,CAAC;AACtC,mBAAiB,MAAM,WAAW,OAAQ,QAAO,KAAK,EAAE;AACxD,QAAM,SAAS,MAAM,WAAW;AAChC,SAAO,EAAE,QAAQ,OAAO;AAC1B;AA2CO,IAAM,qBAAN,cAAiC,MAAM;AAAA,EAC5C,YACE,SACgB,SACA,SAAiC,CAAC,GAClD;AACA,UAAM,OAAO;AAHG;AACA;AAGhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAAA,EACA;AAKpB;AAeO,SAAS,YACd,UACqC;AACrC,MAAI,CAAC,SAAS,MAAM,SAAS,GAAG,KAAK,EAAE,WAAW,GAAG;AACnD,UAAM,IAAI,mBAAmB,iCAAiC,SAAS,MAAM,EAAE;AAAA,EACjF;AACA,MAAI,CAAC,SAAS,YAAY,SAAS,SAAS,KAAK,EAAE,WAAW,GAAG;AAC/D,UAAM,IAAI,mBAAmB,uCAAuC,SAAS,EAAE;AAAA,EACjF;AACA,QAAM,SAAS,iBAAiB,SAAS,UAAU,SAAS,QAAQ;AACpE,MAAI,OAAO,SAAS,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,oBAAoB,QAAQ,SAAS,QAAQ;AAAA,MAC7C,SAAS;AAAA,MACT;AAAA,IACF;AAAA,EACF;AAGA,QAAM,QAAQ,SAAS,OAAO,WAAW,OAAO,CAAC,KAAK,MAAM,MAAM,EAAE,QAAQ,CAAC;AAC7E,MAAI,SAAS,OAAO,WAAW,SAAS,MAAM,QAAQ,OAAO,QAAQ,MAAM;AACzE,UAAM,IAAI;AAAA,MACR,eAAe,SAAS,EAAE,sCAAsC,MAAM,QAAQ,CAAC,CAAC;AAAA,MAChF,SAAS;AAAA,IACX;AAAA,EACF;AACA,SAAO;AACT;;;AClSA,SAAS,iBAAiB;AAC1B,SAAS,oBAAoB;AAE7B,SAAS,2BAA2B;AAsFpC,IAAM,uBAA8D;AAAA,EAClE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,SAAS,gCACd,MAC4C;AAC5C,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,cAAc,KAAK,uBAAuB;AAEhD,SAAO;AAAA,IACL,MAAM,oBAAoB,UAAU;AAClC,YAAM,QAAkC,CAAC;AACzC,YAAM,SAAyE,CAAC;AAChF,UAAI,UAAU;AAEd,iBAAW,KAAK,UAAU;AACxB,cAAM,UAAU,oBAAoB,EAAE,OAAO;AAC7C,YAAI,YAAY,MAAM;AACpB,cAAI,EAAE,YAAY,QAAW;AAC3B,mBAAO,KAAK;AAAA,cACV,WAAW,EAAE;AAAA,cACb,SAAS,EAAE;AAAA,cACX,SAAS;AAAA,YACX,CAAC;AAAA,UACH,OAAO;AAGL,uBAAW;AAAA,UACb;AACA;AAAA,QACF;AAGA,YAAI,QAAQ,SAAS,WAAW;AAC9B,qBAAW;AACX;AAAA,QACF;AAIA,YAAI,QAAQ,KAAK,WAAW,YAAY,GAAG;AACzC,qBAAW;AACX;AAAA,QACF;AAEA,cAAM,SAAS,mBAAmB,SAAS,KAAK,UAAU,KAAK,QAAQ;AACvE,YAAI,WAAW,MAAM;AACnB,iBAAO,KAAK;AAAA,YACV,WAAW,EAAE;AAAA,YACb,SAAS,EAAE,WAAW;AAAA,YACtB,SAAS,iBAAiB,QAAQ,IAAI;AAAA,UACxC,CAAC;AACD;AAAA,QACF;AAEA,YAAI,OAAO,WAAW,gBAAgB,CAAC,YAAY,SAAS,QAAQ,IAAI,GAAG;AACzE,iBAAO,KAAK;AAAA,YACV,WAAW,EAAE;AAAA,YACb,SAAS,EAAE,WAAW;AAAA,YACtB,SAAS,UAAU,OAAO,gBAAgB,8BAA8B,QAAQ,IAAI;AAAA,UACtF,CAAC;AACD;AAAA,QACF;AAEA,cAAM,iBAAiB,OAAO,SAAS,aAAa,OAAO,cAAc,OAAO,IAAI;AAEpF,YAAI;AACJ,YAAI;AACF,kBAAQ,MAAM,KAAK,WAAW,EAAE,SAAS,GAAG,SAAS,QAAQ,eAAe,CAAC;AAAA,QAC/E,SAAS,KAAK;AACZ,iBAAO,KAAK;AAAA,YACV,WAAW,EAAE;AAAA,YACb,SAAS,EAAE,WAAW;AAAA,YACtB,SAAS,qBAAqB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UAChF,CAAC;AACD;AAAA,QACF;AAEA,YAAI,MAAM,MAAM,KAAK,EAAE,WAAW,GAAG;AACnC,qBAAW;AACX;AAAA,QACF;AAEA,cAAM,KAAK;AAAA,UACT,IAAI,OAAO,EAAE,UAAU;AAAA,UACvB,iBAAiB,EAAE;AAAA,UACnB;AAAA,UACA;AAAA,UACA,YAAY,OAAO,cAAc;AAAA,UACjC,OAAO,MAAM;AAAA,UACb,SAAS,MAAM;AAAA,UACf,WAAW,MAAM;AAAA,UACjB,YAAY,EAAE;AAAA,UACd,UAAU,EAAE;AAAA,QACd,CAAC;AAAA,MACH;AAEA,aAAO,EAAE,OAAO,SAAS,OAAO;AAAA,IAClC;AAAA,IAEA,MAAM,MAAM,OAAO;AACjB,YAAM,UAAoB,CAAC;AAC3B,YAAM,WAAqB,CAAC;AAE5B,UAAI,SAAS,QAAQ;AACnB,iBAAS;AAAA,UACP;AAAA,QACF;AACA,eAAO,EAAE,SAAS,SAAS;AAAA,MAC7B;AAEA,UAAI,SAAS,aAAa,CAAC,KAAK,QAAQ;AACtC,iBAAS;AAAA,UACP;AAAA,QACF;AACA,eAAO,EAAE,SAAS,SAAS;AAAA,MAC7B;AAEA,iBAAW,QAAQ,OAAO;AAIxB,cAAM,UAAU,KAAK,OAAO,SAAS,aAAa,KAAK,OAAO,cAAc,OAAO,IAAI;AACvF,YAAI,OAAO,OAAO,MAAM,KAAK,YAAY;AACvC,mBAAS;AAAA,YACP,GAAG,KAAK,OAAO,gBAAgB;AAAA,UACjC;AACA;AAAA,QACF;AAEA,cAAM,KAAK,kBAAkB,MAAM,KAAK,QAAQ;AAChD,YAAI,CAAC,IAAI;AACP,mBAAS,KAAK,GAAG,KAAK,OAAO,gBAAgB,oBAAoB;AACjE;AAAA,QACF;AACA,gBAAQ,KAAK,KAAK,OAAO,gBAAgB;AAAA,MAC3C;AAEA,UAAI,SAAS,aAAa,QAAQ,SAAS,KAAK,KAAK,QAAQ;AAC3D,cAAM,QAAQ;AAAA,UACZ;AAAA,UACA,MAAM,OAAO,CAAC,MAAM,QAAQ,SAAS,EAAE,OAAO,gBAAgB,CAAC;AAAA,UAC/D,KAAK;AAAA,UACL,KAAK;AAAA,UACL,KAAK,cAAc;AAAA,QACrB;AACA,YAAI,MAAO,UAAS,KAAK,cAAc,KAAK,EAAE;AAAA,YACzC,UAAS,KAAK,gEAAgE;AAAA,MACrF;AAEA,aAAO,EAAE,SAAS,SAAS;AAAA,IAC7B;AAAA,EACF;AACF;AAIA,SAAS,kBAAkB,MAA8B,UAA2B;AAClF,QAAM,SAAS,UAAU,OAAO,CAAC,SAAS,oBAAoB,OAAO,GAAG,GAAG;AAAA,IACzE,KAAK;AAAA,IACL,OAAO,KAAK;AAAA,IACZ,UAAU;AAAA,EACZ,CAAC;AACD,SAAO,OAAO,WAAW;AAC3B;AAEA,SAAS,gBACP,OACA,OACA,UACA,QACA,YACe;AACf,QAAM,SAAS,gBAAgB,KAAK,IAAI,CAAC,IAAI,MAAM,CAAC,GAAG,gBAAgB,MAAM,GAAG,EAAE,KAAK,OAAO;AAE9F,QAAM,WAAW,UAAU,OAAO,CAAC,YAAY,MAAM,MAAM,GAAG,EAAE,KAAK,SAAS,CAAC;AAC/E,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,QAAM,MAAM,UAAU,OAAO,CAAC,OAAO,GAAG,KAAK,GAAG,EAAE,KAAK,SAAS,CAAC;AACjE,MAAI,IAAI,WAAW,EAAG,QAAO;AAC7B,QAAM,QAAQ,iBAAiB,MAAM,CAAC,GAAG,WAAW,GAAG,MAAM,MAAM,oBAAoB;AACvF,QAAM,OAAO;AAAA,IACX;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG,MAAM;AAAA,MACP,CAAC,MACC,OAAO,EAAE,eAAe,gBAAgB,EAAE,WAAW,QAAQ,CAAC,CAAC,cAAc,EAAE,QAAQ;AAAA,IAC3F;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG,MAAM,IAAI,CAAC,MAAM;AAAA,KAAQ,EAAE,OAAO,gBAAgB;AAAA;AAAA,EAAO,EAAE,SAAS,EAAE;AAAA,EAC3E,EAAE,KAAK,IAAI;AACX,QAAM,SAAS,UAAU,OAAO,CAAC,UAAU,MAAM,OAAO,MAAM,IAAI,GAAG,EAAE,KAAK,SAAS,CAAC;AACtF,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,QAAM,OAAO,UAAU,OAAO,CAAC,QAAQ,MAAM,UAAU,MAAM,GAAG,EAAE,KAAK,SAAS,CAAC;AACjF,MAAI,KAAK,WAAW,EAAG,QAAO;AAC9B,QAAM,KAAK;AAAA,IACT;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA,EAAE,KAAK,UAAU,UAAU,QAAQ;AAAA,EACrC;AACA,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,SAAO,GAAG,OAAO,KAAK;AACxB;AAEA,SAAS,OAAO,GAAmB;AAGjC,QAAM,SAAS,UAAQ,QAAa;AACpC,SAAO,OAAO,WAAW,QAAQ,EAAE,OAAO,GAAG,OAAO,EAAE,OAAO,KAAK;AACpE;;;ACnSO,SAAS,8BACd,MACA,MAC6B;AAC7B,SAAO;AAAA,IACL,oBAAoB,UAAU;AAC5B,YAAM,QAAQ,KAAK,oBAAoB,QAAQ;AAC/C,aAAO;AAAA,QACL,WAAW,MAAM;AAAA,QACjB,SAAS,MAAM;AAAA,QACf,QAAQ,MAAM;AAAA,MAChB;AAAA,IACF;AAAA,IACA,MAAM,MAAM,WAAW;AACrB,YAAM,UAAoB,CAAC;AAC3B,YAAM,WAAqB,CAAC;AAC5B,iBAAW,KAAK,WAAW;AACzB,cAAM,eAAe,2BAA2B,CAAC;AACjD,YAAI,iBAAiB,MAAM;AACzB,mBAAS;AAAA,YACP,kEAAkE,mBAAmB,CAAC,CAAC;AAAA,UACzF;AACA;AAAA,QACF;AACA,YAAI;AACF,gBAAM,IAAI,MAAM,KAAK,0BAA0B,KAAK,eAAe,YAAY;AAC/E,kBAAQ,KAAK,GAAG,EAAE,OAAO;AACzB,mBAAS,KAAK,GAAG,EAAE,QAAQ;AAAA,QAC7B,SAAS,KAAK;AACZ,mBAAS;AAAA,YACP,qCAAqC,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UACvF;AAAA,QACF;AAAA,MACF;AACA,UAAI,KAAK,kBAAkB,QAAQ,SAAS,GAAG;AAC7C,YAAI;AACF,gBAAM,aAAa,MAAM,KAAK,eAAe,KAAK,aAAa;AAC/D,qBAAW,SAAS,WAAY,UAAS,KAAK,SAAS,KAAK,EAAE;AAAA,QAChE,SAAS,KAAK;AACZ,mBAAS;AAAA,YACP,0BAA0B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UAC5E;AAAA,QACF;AAAA,MACF;AACA,aAAO,EAAE,SAAS,SAAS;AAAA,IAC7B;AAAA,EACF;AACF;AAYA,SAAS,2BAA2B,GAA2B;AAC7D,MAAI,CAAC,KAAK,OAAO,MAAM,SAAU,QAAO;AACxC,QAAM,MAAM;AACZ,MAAI,MAAM,QAAQ,IAAI,WAAW,GAAG;AAClC,UAAM,SAAS,IAAI;AACnB,QAAI,OAAO,WAAW,EAAG,QAAO;AAChC,WAAO,OACJ,IAAI,CAAC,MAAO,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAAG,EAC3D,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC,EAC1B,KAAK,MAAM;AAAA,EAChB;AACA,MAAI,OAAO,IAAI,iBAAiB,SAAU,QAAO,IAAI;AACrD,MAAI,OAAO,IAAI,YAAY,SAAU,QAAO,IAAI;AAChD,SAAO;AACT;AAEA,SAAS,mBAAmB,GAAoB;AAC9C,MAAI,CAAC,KAAK,OAAO,MAAM,SAAU,QAAO;AACxC,QAAM,MAAM;AACZ,MAAI,OAAO,IAAI,oBAAoB,SAAU,QAAO,IAAI;AACxD,MAAI,OAAO,IAAI,OAAO,SAAU,QAAO,IAAI;AAC3C,SAAO;AACT;;;AC1EA,eAAsB,eACpB,QACA,MACmF;AACnF,QAAM,UAAU,OAAO,WAAW,WAAW,CAAC;AAC9C,QAAM,sBAAsB,OAAO,aAAa,WAAW,CAAC;AAC5D,QAAM,aAAa,CAAC,GAAG,SAAS,GAAG,mBAAmB;AAItD,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO;AAAA,MACL,GAAG;AAAA,MACH,SAAS;AAAA,QACP,mBAAmB,cAAc,KAAK,QAAQ;AAAA,QAC9C,gBAAgB,cAAc,KAAK,QAAQ;AAAA,QAC3C,OAAO;AAAA,QACP,YAAY,KAAK,SAAS,IAAI,CAAC,OAAO;AAAA,UACpC,WAAW,EAAE;AAAA,UACb,QAAQ,EAAE;AAAA,UACV,OAAO,EAAE;AAAA,UACT,OAAO;AAAA,QACT,EAAE;AAAA,QACF,iBAAiB,CAAC;AAAA,MACpB;AAAA,IACF;AAAA,EACF;AAEA,QAAM,aAAa,KAAK,SAAS,IAAI,CAAC,MAAM,EAAE,SAAS;AACvD,QAAM,QAAQ,MAAM,KAAK,YAAY,UAAU;AAC/C,QAAM,iBAAiB,IAAI,IAAI,MAAM,IAAI,CAAC,MAAM,CAAC,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC;AAE3E,QAAM,aAAa,KAAK,SAAS,IAAI,CAAC,MAAM;AAC1C,UAAM,IAAI,eAAe,IAAI,EAAE,SAAS,KAAK,EAAE;AAC/C,WAAO,EAAE,WAAW,EAAE,WAAW,QAAQ,EAAE,WAAW,OAAO,GAAG,OAAO,IAAI,EAAE,UAAU;AAAA,EACzF,CAAC;AACD,QAAM,oBAAoB,cAAc,KAAK,QAAQ;AACrD,QAAM,iBAAiB,cAAc,KAAK;AAC1C,QAAM,QAAQ,iBAAiB;AAE/B,MAAI,kBAA4B,CAAC;AACjC,MAAI,QAAQ,KAAK,KAAK,wBAAwB,KAAK,QAAQ;AACzD,UAAM,KAAK,OAAO,UAAU;AAC5B,sBAAkB,CAAC,GAAG,UAAU;AAAA,EAClC;AAEA,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS;AAAA,MACP;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,cAAc,MAAoD;AACzE,MAAI,KAAK,WAAW,EAAG,QAAO;AAC9B,SAAO,KAAK,OAAO,CAAC,KAAK,MAAM,MAAM,EAAE,WAAW,CAAC,IAAI,KAAK;AAC9D;;;AC9DO,SAAS,iBACd,SAC6E;AAC7E,QAAM,WAAW,QAAQ,YAAY;AAErC,SAAO,CAAC,SAAmB,QAAyD;AAClF,UAAM,UAAU,8BAA8B,QAAQ,aAAa;AAAA,MACjE,GAAI,QAAQ,UAAU,OAAO,KAAK,CAAC;AAAA,MACnC,GAAI,QAAQ,MAAM,EAAE,KAAK,QAAQ,IAAI,IAAI,CAAC;AAAA,IAC5C,CAAC;AACD,UAAM,eAAe,QAAQ,QAAQ,QAAQ,QAAQ;AACrD,UAAM,UAAU,QAAQ,YAAY,OAAO;AAC3C,UAAM,SAAS,IAAI,UAAU,IAAI,gBAAgB,EAAE;AAEnD,UAAM,MAAsB,CAAC;AAC7B,QAAI;AACJ,QAAI;AACJ,UAAM,SAAS,IAAI,QAAoB,CAAC,SAAS,WAAW;AAC1D,eAAS;AACT,aAAO;AAAA,IACT,CAAC;AAID,WAAO,MAAM,MAAM;AAAA,IAAC,CAAC;AAErB,UAAM,OAA8B;AAAA,MAClC;AAAA,MACA,cAAc,MAAM;AAAA,MACpB,MAAM;AAAA,MACN,GAAI,QAAQ,mBAAmB,EAAE,kBAAkB,QAAQ,iBAAiB,IAAI,CAAC;AAAA,IACnF;AAEA,oBAAgB,SAA4C;AAC1D,UAAI;AACF,cAAM,MAAM,MAAM,qBAAqB,QAAQ,eAAe,MAAM,MAAM;AAC1E,yBAAiB,SAAS,IAAI,aAAa,SAAS,EAAE,OAAO,CAAC,GAAG;AAC/D,cAAI,KAAK,KAAK;AACd,gBAAM,SAAS,SAAS,OAAO,EAAE,aAAa,CAAC;AAC/C,cAAI,OAAQ,OAAM;AAAA,QACpB;AACA,eAAO,QAAQ,OAAO,MAAM,GAAG,CAAC;AAAA,MAClC,SAAS,KAAK;AACZ,aAAK,GAAG;AACR,cAAM;AAAA,MACR;AAAA,IACF;AAEA,WAAO,EAAE,QAAQ,OAAO,GAAG,OAAO;AAAA,EACpC;AACF;","names":[]}
+{"version":3,"sources":["../src/agent/surfaces.ts","../src/agent/define-agent.ts","../src/agent/improvement-adapter.ts","../src/agent/knowledge-adapter.ts","../src/agent/outcome.ts","../src/agent/sandbox-act.ts"],"sourcesContent":["/**\n * `AgentSurfaces` — declarative map of the mutable file/directory paths\n * the self-improvement loop can edit on behalf of an agent.\n *\n * The substrate uses this map to resolve every parsed `FindingSubject`\n * (from agent-eval) to a real on-disk path. No per-vertical glue;\n * no fabricated paths; no silent `existsSync(...)` skips that hide\n * misconfiguration from the operator.\n *\n * Surfaces are validated at `defineAgent` time — missing paths fail\n * loud with a list of every offender. A surface that's not needed\n * (e.g. an agent with no RAG corpora) is simply omitted; the loop\n * refuses to route those subjects rather than fabricating a target.\n */\n\nimport { existsSync } from 'node:fs'\nimport { isAbsolute, join } from 'node:path'\nimport type { FindingSubject } from '@tangle-network/agent-eval'\n\n/**\n * Surface declarations. Every path is repo-relative (or absolute) at\n * `defineAgent` time. At resolution time, paths are joined against the\n * agent's `repoRoot`.\n *\n * `systemPrompt`, `tools`, `personas` are DIRECTORIES; the loop appends\n * `<section>.md`, `<tool>/README.md`, `<persona-id>.yaml` etc.\n * `rubric`, `outputSchema` are SINGLE FILES; the loop edits them in\n * place.\n *\n * `knowledge` is the agent-knowledge root (typically `.agent-knowledge`);\n * `applyKnowledgeWriteBlocks` writes pages relative to it.\n *\n * Optional surfaces (`scaffolding`, `memory`, `rag`, `outputSchema`)\n * can be omitted — the loop will reject findings targeting them with a\n * clear log message instead of fabricating a path.\n */\nexport interface AgentSurfaces {\n  /** Directory containing one markdown file per system-prompt section. */\n  systemPrompt: string\n  /** Directory containing one subdir per tool (`<tool>/README.md`). */\n  tools: string\n  /** Single file (TypeScript module) defining the rubric weights + dimensions. */\n  rubric: string\n  /** Knowledge-base root; typically `.agent-knowledge`. */\n  knowledge: string\n  /** Directory containing one YAML/JSON file per persona. */\n  personas: string\n  /** Optional: directory containing scaffolding rules (precondition checks, retry policies). */\n  scaffolding?: string\n  /** Optional: memory store path (JSONL / SQLite / DB). */\n  memory?: string\n  /** Optional: directory containing RAG corpora (`<corpus>/<doc-id>.md`). */\n  rag?: string\n  /** Optional: single file defining the output schema (Zod / JSON Schema). */\n  outputSchema?: string\n}\n\nexport interface ResolvedSurface {\n  /** Absolute filesystem path the operator can `cat` / `vim`. */\n  absolutePath: string\n  /** Repo-relative path for PR descriptions, diffs, audit logs. */\n  repoRelativePath: string\n  /** Whether the path currently exists on disk. */\n  exists: boolean\n  /** The substrate's intent: edit an existing file or create a new one. */\n  intent: 'edit-existing' | 'create-new'\n}\n\n/**\n * Resolve a parsed `FindingSubject` to the file path the substrate\n * should edit (or create) on disk.\n *\n * Returns `null` when:\n *   - the subject targets a surface the agent didn't declare\n *     (e.g. `rag:*` when `surfaces.rag` is undefined), OR\n *   - the subject is a `cluster` (failure-mode emits these as evidence,\n *     not actionable mutations — they don't route to a file).\n *\n * Returns a `ResolvedSurface` with `intent: 'create-new'` when the\n * subject names a path that doesn't yet exist (e.g. a new wiki page).\n * The caller chooses whether to honour the create — for tightly-managed\n * surfaces like `systemPrompt` it's usually a contract violation\n * (the analyst named a section that doesn't exist); for `knowledge`\n * it's the whole point.\n */\nexport function resolveSubjectPath(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ResolvedSurface | null {\n  const candidates = candidatePathsForSubject(subject, surfaces)\n  if (candidates.length === 0) return null\n\n  // Probe candidates in order, preferring the first one that exists on disk.\n  // Lets the substrate accept both the flat `<section>.md` convention and\n  // the skill-dir `<section>/SKILL.md` convention without forcing one layout.\n  // When none exists, fall back to the first candidate (canonical create-new).\n  for (const rel of candidates) {\n    const abs = isAbsolute(rel) ? rel : join(repoRoot, rel)\n    if (existsSync(abs)) {\n      return { absolutePath: abs, repoRelativePath: rel, exists: true, intent: 'edit-existing' }\n    }\n  }\n  const fallback = candidates[0]!\n  const fallbackAbs = isAbsolute(fallback) ? fallback : join(repoRoot, fallback)\n  return {\n    absolutePath: fallbackAbs,\n    repoRelativePath: fallback,\n    exists: false,\n    intent: 'create-new',\n  }\n}\n\nfunction candidatePathsForSubject(\n  subject: FindingSubject,\n  surfaces: AgentSurfaces,\n): ReadonlyArray<string> {\n  switch (subject.kind) {\n    case 'knowledge.wiki':\n    case 'knowledge.stale':\n      return [join(surfaces.knowledge, `${subject.slug}.md`)]\n    case 'knowledge.claim':\n      // Claims land in a per-topic claims directory under the knowledge root.\n      return [join(surfaces.knowledge, 'claims', `${slugify(subject.topic)}.md`)]\n    case 'knowledge.raw':\n      return [join(surfaces.knowledge, 'raw', `${subject.sourceId}.md`)]\n    case 'system-prompt': {\n      const slug = slugify(subject.section)\n      // Prefer flat layout for create-new (canonical); probe skill-dir layout\n      // in case the existing repo (tax/legal/gtm/creative) uses\n      // `<section>/SKILL.md` already.\n      return [\n        join(surfaces.systemPrompt, `${slug}.md`),\n        join(surfaces.systemPrompt, slug, 'SKILL.md'),\n        join(surfaces.systemPrompt, slug, 'index.md'),\n      ]\n    }\n    case 'tool-doc':\n      if (subject.aspect) {\n        return [join(surfaces.tools, subject.tool, `${slugify(subject.aspect)}.md`)]\n      }\n      // tool-doc default: `<tool>/README.md`; also probe `<tool>.md` for flat\n      // tool-list repos.\n      return [\n        join(surfaces.tools, subject.tool, 'README.md'),\n        join(surfaces.tools, `${subject.tool}.md`),\n      ]\n    case 'new-tool':\n      return [join(surfaces.tools, subject.name, 'README.md')]\n    case 'rag':\n      if (!surfaces.rag) return []\n      return [join(surfaces.rag, subject.corpus, `${subject.docId}.md`)]\n    case 'memory':\n      if (!surfaces.memory) return []\n      return [join(surfaces.memory, `${slugify(subject.key)}.json`)]\n    case 'scaffolding':\n      if (!surfaces.scaffolding) return []\n      return [join(surfaces.scaffolding, `${slugify(subject.concern)}.md`)]\n    case 'output-schema':\n      if (!surfaces.outputSchema) return []\n      return [surfaces.outputSchema]\n    case 'websearch.outdated':\n    case 'prior-run-summary':\n      // Stale signals don't map to a single file — handled by the knowledge\n      // adapter as `agent-knowledge:stale:*` after operator review.\n      return []\n    case 'cluster':\n      // failure-mode cluster labels are evidence, not mutations.\n      return []\n  }\n}\n\nfunction slugify(s: string): string {\n  return (\n    s\n      .toLowerCase()\n      .replace(/[^a-z0-9-]+/g, '-')\n      .replace(/^-+|-+$/g, '')\n      .slice(0, 200) || 'untitled'\n  )\n}\n\n/**\n * Validate that every declared surface exists on disk under `repoRoot`.\n *\n * Returns an array of `SurfaceValidationIssue` — empty when all required\n * surfaces resolve. `defineAgent` throws with the issues rendered, so\n * a misconfigured manifest fails at startup (not at the first finding\n * the loop produces 20 minutes later).\n */\nexport interface SurfaceValidationIssue {\n  surface: keyof AgentSurfaces\n  path: string\n  reason: 'missing' | 'not-directory' | 'not-file'\n}\n\nexport function validateSurfaces(\n  surfaces: AgentSurfaces,\n  repoRoot: string,\n): ReadonlyArray<SurfaceValidationIssue> {\n  const issues: SurfaceValidationIssue[] = []\n  const dirSurfaces: ReadonlyArray<keyof AgentSurfaces> = [\n    'systemPrompt',\n    'tools',\n    'personas',\n    'knowledge',\n  ]\n  const fileSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['rubric']\n  const optionalDirSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['scaffolding', 'memory', 'rag']\n  const optionalFileSurfaces: ReadonlyArray<keyof AgentSurfaces> = ['outputSchema']\n\n  for (const key of dirSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    }\n  }\n  for (const key of fileSurfaces) {\n    const p = surfaces[key] as string | undefined\n    if (!p) {\n      issues.push({ surface: key, path: '', reason: 'missing' })\n      continue\n    }\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    }\n  }\n  for (const key of [...optionalDirSurfaces, ...optionalFileSurfaces]) {\n    const p = surfaces[key] as string | undefined\n    if (p === undefined) continue\n    const abs = isAbsolute(p) ? p : join(repoRoot, p)\n    if (!existsSync(abs)) {\n      issues.push({ surface: key, path: p, reason: 'missing' })\n    }\n  }\n  return issues\n}\n\nexport function renderSurfaceIssues(\n  issues: ReadonlyArray<SurfaceValidationIssue>,\n  repoRoot: string,\n): string {\n  if (issues.length === 0) return ''\n  const lines = issues.map(\n    (i) => `  - ${i.surface}: ${i.path ? `\"${i.path}\"` : '<not set>'} (${i.reason})`,\n  )\n  return [\n    `Agent surface validation failed against repoRoot=${repoRoot}:`,\n    ...lines,\n    '',\n    'Fix the manifest: every required surface must point at an existing',\n    'directory (systemPrompt / tools / personas / knowledge) or file',\n    '(rubric). Optional surfaces (scaffolding / memory / rag / outputSchema)',\n    'may be omitted; the loop will reject findings targeting omitted',\n    'surfaces rather than fabricating a path.',\n  ].join('\\n')\n}\n","/**\n * `defineAgent` — typed, validated manifest API for declarative agent\n * configuration. The substrate consumes this manifest to wire the\n * canonical eval pattern + analyst self-improvement loop without any\n * per-vertical glue.\n *\n * Design goal: scale to 1000s of vertical agents. Every agent declares\n * its surfaces, rubric, runtime, and analyst configuration in ~50 lines.\n * No per-vertical `ImprovementAdapter`. No per-vertical CLI. No\n * fabricated paths.\n *\n * Validation: `defineAgent` runs `validateSurfaces` synchronously and\n * throws a structured error if any required surface is missing on\n * disk. The cost is one filesystem stat per surface (cheap); the\n * benefit is a manifest that can't ship broken.\n */\n\nimport type { TraceAnalystKindSpec } from '@tangle-network/agent-eval'\nimport type { RuntimeStreamEvent } from '../types'\nimport { type AgentSurfaces, renderSurfaceIssues, validateSurfaces } from './surfaces'\n\n// ── manifest ─────────────────────────────────────────────────────────\n\n/**\n * The full agent manifest. Each agent ships ONE of these.\n *\n * Generics:\n *   `TPersona` — the agent's persona shape (loaded from\n *     `surfaces.personas`). Defaults to `unknown` so the substrate's\n *     persona discovery (`loadPersonas`) can accept anything; per-agent\n *     code re-narrows when it matters.\n *   `TRunOutput` — the shape `runtime.act` returns. Used by the rubric\n *     scorers and emitted into the trace.\n */\nexport interface AgentManifest<TPersona = unknown, TRunOutput = unknown> {\n  /**\n   * Stable identifier — used as `projectId` in traces, as the analyst\n   * loop's `runId` prefix, and as the namespace under which findings\n   * are persisted. MUST match the agent's repo name to keep\n   * cross-repo telemetry joinable.\n   */\n  id: string\n\n  /**\n   * Filesystem root the substrate resolves surface paths against.\n   * Typically `process.cwd()` or a fixed absolute path. Use an\n   * absolute path when the agent's tests may run from subdirectories\n   * (vitest sometimes shifts cwd).\n   */\n  repoRoot: string\n\n  /**\n   * Map of mutable surfaces the self-improvement loop can edit. See\n   * `AgentSurfaces` — required: `systemPrompt`, `tools`, `rubric`,\n   * `knowledge`, `personas`. Optional: `scaffolding`, `memory`, `rag`,\n   * `outputSchema`.\n   *\n   * Every required path is validated at `defineAgent` time. Missing\n   * paths throw with the full list of offenders.\n   */\n  surfaces: AgentSurfaces\n\n  /**\n   * Rubric the substrate uses to score each run. Dimensions × weights\n   * × judges. The substrate computes the weighted composite and\n   * stamps it into the RunRecord.\n   */\n  rubric: AgentRubric<TRunOutput>\n\n  /**\n   * Runtime adapter — how the substrate INVOKES the agent against a\n   * persona. The `act` function takes a persona + a context (with the\n   * tracer the substrate threads through for span emission) and\n   * returns the run output the rubric will score.\n   *\n   * The agent's existing production runtime goes in here; the\n   * substrate is intentionally thin around it.\n   */\n  runtime: AgentRuntime<TPersona, TRunOutput>\n\n  /**\n   * Persona discovery — the substrate loads personas via this function\n   * at eval start. Can read from `surfaces.personas`, an API, or be\n   * hardcoded. The substrate calls it once per `runAgentEval` call;\n   * persona ordering is preserved.\n   */\n  personas: () => Promise<ReadonlyArray<TPersona>>\n\n  /**\n   * Analyst kinds the substrate runs against each persona's trace.\n   * Defaults to `DEFAULT_TRACE_ANALYST_KINDS` from agent-eval. Per-agent\n   * authors can prune (e.g. skip `knowledge-poisoning` when there's no\n   * knowledge base) or extend (custom domain kinds).\n   *\n   * Empty array disables the loop — useful for `pnpm eval --no-analyst`.\n   */\n  analystKinds: ReadonlyArray<TraceAnalystKindSpec>\n\n  /**\n   * Analyst LLM configuration. The substrate uses these for all four\n   * kinds (override per-kind via `analystKinds` if needed).\n   */\n  analyst: AnalystConfig\n\n  /**\n   * Auto-apply policy. Knowledge / improvement edits land only when\n   * `enabled === true` AND the source finding's confidence meets the\n   * threshold. `mode` controls how applies happen: `'write'` mutates\n   * files in-place; `'open-pr'` writes to a branch and opens a PR.\n   *\n   * Default: knowledge auto-applies at confidence ≥0.85 in `'write'`\n   * mode (wiki edits are git-reversible); improvement stays at\n   * `enabled: false` until the agent author has measured precision.\n   */\n  autoApply?: AutoApplyPolicy\n}\n\nexport interface AgentRubric<TRunOutput> {\n  /** Dimensions composing the weighted score. Weights sum to 1.0 by convention. */\n  dimensions: ReadonlyArray<RubricDimension<TRunOutput>>\n  /**\n   * Optional judges layered on top of deterministic dimensions. Each\n   * judge returns a score per dimension; the substrate averages judges\n   * (mean by default) for the LLM contribution.\n   */\n  judges?: ReadonlyArray<JudgeConfig<TRunOutput>>\n}\n\nexport interface RubricDimension<TRunOutput> {\n  /** Unique identifier — appears in finding subjects (`rubric:<id>`). */\n  id: string\n  /** 0..1 — weight in the composite. */\n  weight: number\n  /**\n   * Deterministic scorer: given the persona + run output, returns a\n   * 0..1 score. The substrate sums weight × score across dimensions\n   * for the deterministic composite; judges supplement subjective dims.\n   */\n  score: (input: { persona: unknown; output: TRunOutput }) => number\n  /** Optional human-readable label for reports. */\n  label?: string\n}\n\nexport interface JudgeConfig<TRunOutput> {\n  /** Judge identifier — appears in trace spans + manifest. */\n  id: string\n  /** Model snapshot to invoke. Pin the snapshot (`claude-sonnet-4-6@2025-04-15`); the validator rejects bare aliases. */\n  model: string\n  /** Dimensions this judge scores. */\n  dimensions: ReadonlyArray<string>\n  /**\n   * Optional rubric anchors — text examples the judge sees as a\n   * few-shot prompt to calibrate. STRONGLY recommended for subjective\n   * dimensions; required by the calibration gate (Pearson ≥0.7).\n   */\n  anchors?: ReadonlyArray<{ input: string; output: TRunOutput; expected: Record<string, number> }>\n}\n\nexport interface AgentRuntime<TPersona, TRunOutput> {\n  /**\n   * Invoke the agent against one persona. Returns BOTH:\n   *   - `events`: an `AsyncIterable<RuntimeStreamEvent>` the chat-centric\n   *     product consumes verbatim (SSE / WebSocket / inline render).\n   *     **Streaming is mandatory — never collapse this to a single Promise.**\n   *     The agent's existing `runChatTurn` (or equivalent async generator)\n   *     plugs in here directly.\n   *   - `output`: a `Promise<TRunOutput>` resolved AFTER the event stream\n   *     drains. The eval substrate awaits this for rubric scoring; chat\n   *     products usually ignore it (they already rendered incrementally).\n   *\n   * Implementation contract:\n   *   1. `act` MUST return immediately (synchronous construction of the\n   *      `events` iterator + the `output` promise).\n   *   2. Iterating `events` drives the underlying LLM/tool calls — the\n   *      caller chooses when to consume.\n   *   3. `output` resolves only after the iterator yields its terminal\n   *      event (typically `task_end`); see `collectAgentRun` helper.\n   *\n   * `ctx.emitter` is the substrate-threaded `TraceEmitter` — runtimes\n   * SHOULD record LLM/tool spans through it for capture integrity.\n   * `ctx.deadlineMs` is wall-clock; the runtime SHOULD honour for graceful\n   * cancel. `ctx.signal` is the standard abort signal.\n   */\n  act: (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput>\n}\n\nexport interface AgentRunInvocation<TRunOutput> {\n  /** Live stream of typed runtime events. Consumed by chat UX directly. */\n  events: AsyncIterable<RuntimeStreamEvent>\n  /** Final structured output the rubric scores. Resolves after `events` drains. */\n  output: Promise<TRunOutput>\n}\n\n/**\n * Stub for agents whose `runtime.act` is not yet wired to the substrate's\n * eval path. Preserves the streaming contract (empty event stream + a\n * rejected `output` promise that tells the caller exactly what to fix).\n *\n * Per-vertical manifests usually start with this stub and replace it with\n * the agent's real streaming runtime (`runChatTurn` or equivalent) once\n * the eval path consumes the manifest end-to-end.\n */\nexport function unimplementedAgentRun<TRunOutput = unknown>(\n  reason = 'AgentRuntime.act is not yet wired for this manifest',\n): AgentRunInvocation<TRunOutput> {\n  return {\n    events: (async function* empty(): AsyncIterable<RuntimeStreamEvent> {})(),\n    output: Promise.reject(new Error(reason)),\n  }\n}\n\n/**\n * Drain `act`'s `events` into an array AND await its `output`. Useful for\n * eval / outcome-measurement code paths that don't care about live\n * rendering. The events array is preserved so the substrate can inspect\n * tool calls / readiness / questions retrospectively.\n *\n * IMPORTANT: chat-centric UX MUST NOT call this — it defeats streaming\n * (no incremental render). Use `for await (const ev of invocation.events)`\n * directly in the chat surface.\n */\nexport async function collectAgentRun<TRunOutput>(\n  invocation: AgentRunInvocation<TRunOutput>,\n): Promise<{ events: ReadonlyArray<RuntimeStreamEvent>; output: TRunOutput }> {\n  const events: RuntimeStreamEvent[] = []\n  for await (const ev of invocation.events) events.push(ev)\n  const output = await invocation.output\n  return { events, output }\n}\n\nexport interface AgentRunContext {\n  /** Substrate-managed trace emitter. */\n  emitter: import('@tangle-network/agent-eval').TraceEmitter\n  /** Stable run id for this persona × variant cell. */\n  runId: string\n  /** Variant the runtime is exercising (e.g. `'baseline'`, `'source-grounded'`). */\n  variantId?: string\n  /** Wall-clock deadline (epoch ms). The runtime SHOULD honour for graceful cancel. */\n  deadlineMs?: number\n  /** Optional abort signal. */\n  signal?: AbortSignal\n}\n\nexport interface AnalystConfig {\n  /** Model the analyst kinds use. Override per-kind via `analystKinds[i].cost.models`. */\n  model: string\n  /** Optional total budget across all kinds for one run. Substrate enforces via `BudgetGuard`. */\n  budgetUsd?: number\n  /** Backend hint for the AxAIService factory — same shape every kind uses. */\n  backend?: {\n    name?: 'openai' | 'router'\n    apiKey?: string\n    baseUrl?: string\n  }\n}\n\nexport interface AutoApplyPolicy {\n  knowledge?: {\n    enabled: boolean\n    confidenceThreshold?: number\n    mode?: 'write' | 'open-pr'\n  }\n  improvement?: {\n    enabled: boolean\n    confidenceThreshold?: number\n    mode?: 'write' | 'open-pr'\n  }\n}\n\n// ── factory + validation ─────────────────────────────────────────────\n\nexport class AgentManifestError extends Error {\n  constructor(\n    message: string,\n    public readonly agentId: string,\n    public readonly issues: ReadonlyArray<unknown> = [],\n  ) {\n    super(message)\n    this.name = 'AgentManifestError'\n  }\n}\n\n/**\n * Construct a validated agent manifest. Throws `AgentManifestError`\n * if any required surface is missing on disk.\n *\n * Generics: pass your persona / output types if you want narrowed\n * `runtime.act` signatures:\n *   `defineAgent<TaxPersona, TaxRunOutput>({ ... })`\n *\n * Most callers don't need the generics — the substrate operates on\n * `unknown` payloads internally and the manifest's `score` /\n * `runtime.act` see the typed shapes via TypeScript inference at\n * the call site.\n */\nexport function defineAgent<TPersona = unknown, TRunOutput = unknown>(\n  manifest: AgentManifest<TPersona, TRunOutput>,\n): AgentManifest<TPersona, TRunOutput> {\n  if (!manifest.id || manifest.id.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `id` is required', manifest.id ?? '')\n  }\n  if (!manifest.repoRoot || manifest.repoRoot.trim().length === 0) {\n    throw new AgentManifestError('defineAgent: `repoRoot` is required', manifest.id)\n  }\n  const issues = validateSurfaces(manifest.surfaces, manifest.repoRoot)\n  if (issues.length > 0) {\n    throw new AgentManifestError(\n      renderSurfaceIssues(issues, manifest.repoRoot),\n      manifest.id,\n      issues,\n    )\n  }\n  // Lightweight rubric sanity: weights sum to ~1.0 (no hard requirement —\n  // the substrate normalizes — but flag wildly miscalibrated weights).\n  const total = manifest.rubric.dimensions.reduce((acc, d) => acc + d.weight, 0)\n  if (manifest.rubric.dimensions.length > 0 && (total < 0.5 || total > 1.5)) {\n    throw new AgentManifestError(\n      `defineAgent(${manifest.id}): rubric dimension weights sum to ${total.toFixed(3)} — should be ~1.0`,\n      manifest.id,\n    )\n  }\n  return manifest\n}\n","/**\n * Substrate-default `ImprovementAdapter` — surfaces-driven, LLM-drafted\n * patches, optional auto-apply or PR-open.\n *\n * This is the one ImprovementAdapter every vertical agent uses. The\n * substrate parses each finding's `subject` via\n * `parseFindingSubject` (agent-eval), resolves it to a real file path\n * via the agent's `AgentSurfaces`, reads the current content, and asks\n * an LLM to draft a unified-diff patch given the finding + current\n * content + per-kind editing-discipline rules.\n *\n * Auto-apply gates on the source-finding's confidence and the\n * autoApply.improvement policy. Two modes:\n *   `write` — apply the patch in-place via `git apply -p0`. Operator\n *     reviews via `git diff`.\n *   `open-pr` — write to a branch, commit, push, open a PR via `gh`.\n *     Operator reviews via the PR UI.\n *\n * Fail-loud rules:\n *   - Findings whose subject doesn't parse → counted in `errors`.\n *   - Findings whose subject targets an undeclared surface → counted in\n *     `errors` with the offending kind in the message.\n *   - Findings whose target path doesn't exist AND the kind isn't a\n *     create-new variant (`new-tool`, `knowledge.wiki`) → counted in\n *     `errors` with the resolved path in the message.\n *   - LLM drafts that fail JSON-schema validation → counted in\n *     `errors` with the schema issue.\n *\n * No silent skips. Every dropped finding has a recorded reason the\n * loop's report surfaces.\n */\n\nimport { spawnSync } from 'node:child_process'\nimport { readFileSync } from 'node:fs'\nimport type { AnalystFinding, FindingSubject } from '@tangle-network/agent-eval'\nimport { parseFindingSubject } from '@tangle-network/agent-eval/analyst'\nimport type { ImprovementAdapter } from '../analyst-loop/types'\nimport type { AgentSurfaces, ResolvedSurface } from './surfaces'\nimport { resolveSubjectPath } from './surfaces'\n\n// ── proposal shape ───────────────────────────────────────────────────\n\nexport interface SurfaceImprovementEdit {\n  /** Stable id derived from the source finding so re-proposals are idempotent. */\n  id: string\n  /** The finding that produced this edit — for revert + audit trail. */\n  sourceFindingId: string\n  /** Parsed subject; included so the apply step doesn't re-parse. */\n  subject: FindingSubject\n  /** Resolved on-disk target. */\n  target: ResolvedSurface\n  /** SHA-256 of the current file content the patch was drafted against. */\n  baseSha256: string\n  /** Unified-diff patch the LLM drafted (relative to `target.absolutePath`). */\n  patch: string\n  /** One-line summary the operator sees in the report / PR title. */\n  summary: string\n  /** Multi-line rationale for the PR body — finding context + LLM reasoning. */\n  rationale: string\n  /** Carry-forward from the finding so the apply gate can check the threshold. */\n  confidence: number\n  /** Carry-forward severity for prioritization. */\n  severity: AnalystFinding['severity']\n}\n\nexport interface CreateSurfaceImprovementAdapterOpts {\n  surfaces: AgentSurfaces\n  repoRoot: string\n  /**\n   * LLM-draft callback. Given a finding + current file content + the\n   * resolved target, returns a unified-diff patch + summary + rationale.\n   *\n   * Required — the substrate doesn't ship a hardcoded prompt; the agent\n   * author picks the model (Haiku for cheap routine drafts, Sonnet for\n   * substantive prompt rewrites, etc.) via this callback.\n   */\n  draftPatch: (input: DraftPatchInput) => Promise<DraftPatchOutput>\n  /**\n   * Apply mode:\n   *   `write` — `git apply` in-place; operator reviews via `git diff`\n   *   `open-pr` — branch + commit + push + `gh pr create`\n   *   `none` — never apply; collect proposals for the report only\n   *\n   * The `apply` method honours this even when the loop calls it; the\n   * effective behaviour is also gated on the per-finding confidence\n   * threshold via `runAnalystLoop`'s `autoApply` policy.\n   */\n  mode?: 'write' | 'open-pr' | 'none'\n  /** When `mode === 'open-pr'`, the base branch new PRs target. Default: `main`. */\n  baseBranch?: string\n  /** Required for `mode === 'open-pr'` — the GH owner/repo (`tangle-network/tax-agent`). */\n  ghRepo?: string\n  /**\n   * When the resolved target doesn't exist, allow the substrate to\n   * CREATE the file (for `knowledge.wiki`, `new-tool` subjects). Default\n   * true for those kinds, false for `system-prompt` / `rubric` / etc.\n   * (named sections that don't exist are a contract violation, not a\n   * scaffolding opportunity).\n   */\n  allowCreateForKinds?: ReadonlyArray<FindingSubject['kind']>\n}\n\nexport interface DraftPatchInput {\n  finding: AnalystFinding\n  subject: FindingSubject\n  target: ResolvedSurface\n  /** Current file content (empty string when `intent === 'create-new'`). */\n  currentContent: string\n}\n\nexport interface DraftPatchOutput {\n  /** Unified diff against the current file content. Empty string skips this finding. */\n  patch: string\n  /** One-line summary for the operator. */\n  summary: string\n  /** Multi-line rationale for the PR body. */\n  rationale: string\n}\n\n// ── factory ──────────────────────────────────────────────────────────\n\nconst DEFAULT_CREATE_KINDS: ReadonlyArray<FindingSubject['kind']> = [\n  'knowledge.wiki',\n  'knowledge.claim',\n  'knowledge.raw',\n  'new-tool',\n]\n\nexport function createSurfaceImprovementAdapter(\n  opts: CreateSurfaceImprovementAdapterOpts,\n): ImprovementAdapter<SurfaceImprovementEdit> {\n  const mode = opts.mode ?? 'none'\n  const allowCreate = opts.allowCreateForKinds ?? DEFAULT_CREATE_KINDS\n\n  return {\n    async proposeFromFindings(findings) {\n      const edits: SurfaceImprovementEdit[] = []\n      const errors: Array<{ findingId: string; subject: string; message: string }> = []\n      let skipped = 0\n\n      for (const f of findings) {\n        const subject = parseFindingSubject(f.subject)\n        if (subject === null) {\n          if (f.subject !== undefined) {\n            errors.push({\n              findingId: f.finding_id,\n              subject: f.subject,\n              message: 'subject does not parse against the finding-subject grammar',\n            })\n          } else {\n            // Subject-less findings are descriptive, not actionable —\n            // legitimate; count in `skipped` not `errors`.\n            skipped += 1\n          }\n          continue\n        }\n\n        // `cluster` findings (failure-mode) are evidence, not mutations.\n        if (subject.kind === 'cluster') {\n          skipped += 1\n          continue\n        }\n\n        // `agent-knowledge:*` findings flow to the KnowledgeAdapter;\n        // the ImprovementAdapter skips them so subjects don't double-route.\n        if (subject.kind.startsWith('knowledge.')) {\n          skipped += 1\n          continue\n        }\n\n        const target = resolveSubjectPath(subject, opts.surfaces, opts.repoRoot)\n        if (target === null) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `subject kind \"${subject.kind}\" targets an undeclared surface; declare it in AgentSurfaces or stop emitting this subject`,\n          })\n          continue\n        }\n\n        if (target.intent === 'create-new' && !allowCreate.includes(subject.kind)) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `target ${target.repoRelativePath} does not exist; the kind \"${subject.kind}\" requires an existing target (analyst named a section that isn't in the codebase)`,\n          })\n          continue\n        }\n\n        const currentContent = target.exists ? readFileSync(target.absolutePath, 'utf-8') : ''\n\n        let draft: DraftPatchOutput\n        try {\n          draft = await opts.draftPatch({ finding: f, subject, target, currentContent })\n        } catch (err) {\n          errors.push({\n            findingId: f.finding_id,\n            subject: f.subject ?? '',\n            message: `draftPatch threw: ${err instanceof Error ? err.message : String(err)}`,\n          })\n          continue\n        }\n\n        if (draft.patch.trim().length === 0) {\n          skipped += 1\n          continue\n        }\n\n        edits.push({\n          id: `imp-${f.finding_id}`,\n          sourceFindingId: f.finding_id,\n          subject,\n          target,\n          baseSha256: sha256(currentContent),\n          patch: draft.patch,\n          summary: draft.summary,\n          rationale: draft.rationale,\n          confidence: f.confidence,\n          severity: f.severity,\n        })\n      }\n\n      return { edits, skipped, errors }\n    },\n\n    async apply(edits) {\n      const applied: string[] = []\n      const warnings: string[] = []\n\n      if (mode === 'none') {\n        warnings.push(\n          'createSurfaceImprovementAdapter: mode=none; no edits applied — adjust manifest.autoApply.improvement.mode',\n        )\n        return { applied, warnings }\n      }\n\n      if (mode === 'open-pr' && !opts.ghRepo) {\n        warnings.push(\n          'createSurfaceImprovementAdapter: mode=open-pr requires `ghRepo`; falling back to no-op',\n        )\n        return { applied, warnings }\n      }\n\n      for (const edit of edits) {\n        // Race-detection: confirm the file content hasn't moved since the\n        // patch was drafted. A diff applied against drifted content is a\n        // recipe for silent corruption.\n        const current = edit.target.exists ? readFileSync(edit.target.absolutePath, 'utf-8') : ''\n        if (sha256(current) !== edit.baseSha256) {\n          warnings.push(\n            `${edit.target.repoRelativePath}: base SHA mismatch; file changed after draft. Skipping.`,\n          )\n          continue\n        }\n\n        const ok = applyPatchInPlace(edit, opts.repoRoot)\n        if (!ok) {\n          warnings.push(`${edit.target.repoRelativePath}: git apply failed`)\n          continue\n        }\n        applied.push(edit.target.repoRelativePath)\n      }\n\n      if (mode === 'open-pr' && applied.length > 0 && opts.ghRepo) {\n        const prUrl = openPullRequest(\n          applied,\n          edits.filter((e) => applied.includes(e.target.repoRelativePath)),\n          opts.repoRoot,\n          opts.ghRepo,\n          opts.baseBranch ?? 'main',\n        )\n        if (prUrl) warnings.push(`opened PR: ${prUrl}`)\n        else warnings.push('PR creation failed; edits are committed to a local branch only')\n      }\n\n      return { applied, warnings }\n    },\n  }\n}\n\n// ── apply helpers ────────────────────────────────────────────────────\n\nfunction applyPatchInPlace(edit: SurfaceImprovementEdit, repoRoot: string): boolean {\n  const result = spawnSync('git', ['apply', '--whitespace=fix', '-p0', '-'], {\n    cwd: repoRoot,\n    input: edit.patch,\n    encoding: 'utf-8',\n  })\n  return result.status === 0\n}\n\nfunction openPullRequest(\n  paths: ReadonlyArray<string>,\n  edits: ReadonlyArray<SurfaceImprovementEdit>,\n  repoRoot: string,\n  ghRepo: string,\n  baseBranch: string,\n): string | null {\n  const branch = `analyst-loop/${Date.now()}-${edits[0]?.sourceFindingId.slice(0, 12) ?? 'edits'}`\n  // Create branch, stage, commit\n  const checkout = spawnSync('git', ['checkout', '-b', branch], { cwd: repoRoot })\n  if (checkout.status !== 0) return null\n  const add = spawnSync('git', ['add', ...paths], { cwd: repoRoot })\n  if (add.status !== 0) return null\n  const title = `analyst-loop: ${edits[0]?.summary ?? `${edits.length} improvement edits`}`\n  const body = [\n    `Automated analyst-loop edits — review carefully before merge.`,\n    '',\n    `Source findings:`,\n    ...edits.map(\n      (e) =>\n        `  - ${e.sourceFindingId} (confidence ${e.confidence.toFixed(2)}, severity ${e.severity})`,\n    ),\n    '',\n    'Rationales:',\n    ...edits.map((e) => `\\n## ${e.target.repoRelativePath}\\n\\n${e.rationale}`),\n  ].join('\\n')\n  const commit = spawnSync('git', ['commit', '-m', title, '-m', body], { cwd: repoRoot })\n  if (commit.status !== 0) return null\n  const push = spawnSync('git', ['push', '-u', 'origin', branch], { cwd: repoRoot })\n  if (push.status !== 0) return null\n  const pr = spawnSync(\n    'gh',\n    [\n      'pr',\n      'create',\n      '--repo',\n      ghRepo,\n      '--title',\n      title,\n      '--body',\n      body,\n      '--base',\n      baseBranch,\n      '--head',\n      branch,\n    ],\n    { cwd: repoRoot, encoding: 'utf-8' },\n  )\n  if (pr.status !== 0) return null\n  return pr.stdout.trim()\n}\n\nfunction sha256(s: string): string {\n  // node:crypto is dynamic-imported lazily so the adapter can be tested in\n  // environments without crypto (browser tests, mocked envs).\n  const crypto = require('node:crypto') as typeof import('node:crypto')\n  return crypto.createHash('sha256').update(s, 'utf-8').digest('hex')\n}\n","/**\n * Substrate-default `KnowledgeAdapter` — wraps agent-knowledge's\n * `proposeFromFindings` + `applyKnowledgeWriteBlocks` with substrate\n * defaults (auto-lint after apply, source linkage via finding id).\n *\n * Every agent that ships a `.agent-knowledge/` tree uses this adapter\n * unmodified. Per-agent customization happens at the manifest level\n * (`autoApply.knowledge.confidenceThreshold`, etc.), not by writing a\n * new adapter.\n *\n * Lint discipline: after each apply we run agent-knowledge's\n * `lintKnowledgeIndex` to catch broken links / circular claims /\n * duplicate pages introduced by the new writes. Findings that fail the\n * post-apply lint are recorded in `warnings`; the apply itself is not\n * rolled back (lint failures are soft — humans review the wiki state).\n */\n\nimport type { AnalystFinding } from '@tangle-network/agent-eval'\nimport type { KnowledgeAdapter } from '../analyst-loop/types'\n\nexport interface CreateSurfaceKnowledgeAdapterOpts {\n  /** `.agent-knowledge/` root (absolute path the substrate writes blocks against). */\n  knowledgeRoot: string\n}\n\n/**\n * Build the adapter. We accept the agent-knowledge functions as DI so\n * the substrate stays decoupled from a specific agent-knowledge\n * version — the agent author imports them in their manifest module\n * and hands them to the factory.\n *\n * `proposeFromFindings(findings)` returns\n *   `{ proposals: KnowledgeProposal[]; skipped: number; errors: ... }`.\n *\n * `applyKnowledgeWriteBlocks(root, content)` returns\n *   `{ written: string[]; warnings: string[] }`.\n *\n * `lintKnowledgeIndex(index)` (optional) returns `KnowledgeLintFinding[]`.\n */\nexport interface KnowledgeAdapterDeps<TProposal> {\n  proposeFromFindings: (findings: ReadonlyArray<AnalystFinding>) => {\n    proposals: TProposal[]\n    skipped: number\n    errors: Array<{ findingId: string; subject: string; message: string }>\n  }\n  applyKnowledgeWriteBlocks: (\n    root: string,\n    proposalText: string,\n  ) => Promise<{ written: string[]; warnings: string[] }>\n  /**\n   * Optional post-apply lint hook. The substrate runs it after each\n   * batch of writes; failures land in `warnings` (the apply is not\n   * rolled back — lint signals drift to review, not block).\n   */\n  lintAfterApply?: (root: string) => Promise<ReadonlyArray<string>>\n}\n\nexport function createSurfaceKnowledgeAdapter<TProposal>(\n  opts: CreateSurfaceKnowledgeAdapterOpts,\n  deps: KnowledgeAdapterDeps<TProposal>,\n): KnowledgeAdapter<TProposal> {\n  return {\n    proposeFromFindings(findings) {\n      const batch = deps.proposeFromFindings(findings)\n      return {\n        proposals: batch.proposals,\n        skipped: batch.skipped,\n        errors: batch.errors,\n      }\n    },\n    async apply(proposals) {\n      const written: string[] = []\n      const warnings: string[] = []\n      for (const p of proposals) {\n        const proposalText = renderProposalAsWriteBlock(p)\n        if (proposalText === null) {\n          warnings.push(\n            `proposal has no writeBlocks/content; skipping (sourceFindingId=${getSourceFindingId(p)})`,\n          )\n          continue\n        }\n        try {\n          const r = await deps.applyKnowledgeWriteBlocks(opts.knowledgeRoot, proposalText)\n          written.push(...r.written)\n          warnings.push(...r.warnings)\n        } catch (err) {\n          warnings.push(\n            `applyKnowledgeWriteBlocks failed: ${err instanceof Error ? err.message : String(err)}`,\n          )\n        }\n      }\n      if (deps.lintAfterApply && written.length > 0) {\n        try {\n          const lintIssues = await deps.lintAfterApply(opts.knowledgeRoot)\n          for (const issue of lintIssues) warnings.push(`lint: ${issue}`)\n        } catch (err) {\n          warnings.push(\n            `lintAfterApply failed: ${err instanceof Error ? err.message : String(err)}`,\n          )\n        }\n      }\n      return { written, warnings }\n    },\n  }\n}\n\n/**\n * Pluck the canonical write-block text from a proposal regardless of\n * which exact agent-knowledge version produced it. We accept either:\n *   - `{ writeBlocks: Array<{ path, content }> }` — the typed shape\n *     1.3.0+ emits\n *   - `{ proposalText: string }` — legacy single-block shape\n *   - `{ content: string }` — minimal raw form\n *\n * Returns `null` when nothing parseable is present (warned upstream).\n */\nfunction renderProposalAsWriteBlock(p: unknown): string | null {\n  if (!p || typeof p !== 'object') return null\n  const obj = p as Record<string, unknown>\n  if (Array.isArray(obj.writeBlocks)) {\n    const blocks = obj.writeBlocks as Array<{ path?: string; content?: string }>\n    if (blocks.length === 0) return null\n    return blocks\n      .map((b) => (typeof b.content === 'string' ? b.content : ''))\n      .filter((s) => s.length > 0)\n      .join('\\n\\n')\n  }\n  if (typeof obj.proposalText === 'string') return obj.proposalText\n  if (typeof obj.content === 'string') return obj.content\n  return null\n}\n\nfunction getSourceFindingId(p: unknown): string {\n  if (!p || typeof p !== 'object') return '<unknown>'\n  const obj = p as Record<string, unknown>\n  if (typeof obj.sourceFindingId === 'string') return obj.sourceFindingId\n  if (typeof obj.id === 'string') return obj.id\n  return '<unknown>'\n}\n","/**\n * `OutcomeMeasurement` — the missing metric that turns the analyst\n * loop from \"observability\" into \"self-improvement\".\n *\n * Without this hook, the loop reports process counts (`findings: 42`,\n * `applied: 7`) and never proves the applied edits actually improved\n * anything. With this hook, the substrate re-runs the cohort against\n * the same personas after each apply pass and reports a composite\n * score delta. A negative delta is the substrate's strongest signal\n * to either roll back or surface for review.\n *\n * Wiring is intentionally simple: pass the manifest + the `runAgentEval`\n * function and a list of `personaIds` to re-run. The wrapper:\n *   1. Captures the baseline composite from the just-finished run.\n *   2. After `runAnalystLoop` returns, re-invokes `runAgentEval` against\n *      the same persona slice.\n *   3. Computes the delta and appends to `loop-report.json`.\n *   4. If `rollbackOnRegression` and delta < 0, reverts applied edits.\n */\n\nimport type { RunAnalystLoopResult } from '../analyst-loop/types'\n\nexport interface OutcomeMeasurement {\n  /** Baseline composite before applies — captured from the most-recent eval run. */\n  baselineComposite: number\n  /** Composite after re-running the cohort with applied edits. */\n  afterComposite: number\n  /** `afterComposite - baselineComposite`. Positive = the loop improved the agent. */\n  delta: number\n  /** Per-persona deltas for finer-grained review. */\n  perPersona: ReadonlyArray<{ personaId: string; before: number; after: number; delta: number }>\n  /** When the substrate rolled back applies due to regression, the paths reverted. */\n  rolledBackPaths: ReadonlyArray<string>\n}\n\nexport interface OutcomeMeasurementOpts {\n  /** Composite scores from the run that produced the findings. */\n  baseline: ReadonlyArray<{ personaId: string; composite: number }>\n  /**\n   * Re-run callback — the substrate invokes this after applies. The\n   * agent author provides their `runAgentEval`-equivalent so the\n   * substrate can ask \"score this persona slice now.\"\n   *\n   * The callback SHOULD reuse the same cohort + judges + variant as\n   * the baseline run; only the agent's mutable surfaces have changed.\n   */\n  reRunCohort: (\n    personaIds: ReadonlyArray<string>,\n  ) => Promise<ReadonlyArray<{ personaId: string; composite: number }>>\n  /** When `true`, applied edits are reverted on negative delta. Default `false`. */\n  rollbackOnRegression?: boolean\n  /** Callback to revert a list of paths (typically `git checkout HEAD --`). */\n  revert?: (paths: ReadonlyArray<string>) => Promise<void>\n}\n\n/**\n * Run `runAnalystLoop` and stamp an `OutcomeMeasurement` onto the\n * result. The substrate calls this after each canonical eval; the\n * delta lands in `loop-report.json` for cross-run trend analysis.\n *\n * The function returns the original `RunAnalystLoopResult` enriched\n * with `outcome` so callers stay backwards-compatible (the field is\n * optional on the type; missing means no measurement was wired).\n */\nexport async function measureOutcome<TProposal, TEdit>(\n  result: RunAnalystLoopResult<TProposal, TEdit>,\n  opts: OutcomeMeasurementOpts,\n): Promise<RunAnalystLoopResult<TProposal, TEdit> & { outcome: OutcomeMeasurement }> {\n  const applied = result.knowledge?.applied ?? []\n  const improvementsApplied = result.improvement?.applied ?? []\n  const allApplied = [...applied, ...improvementsApplied]\n\n  // No applies → no outcome to measure. Return a zero-delta to keep the\n  // shape stable for consumers; baseline / after equal.\n  if (allApplied.length === 0) {\n    return {\n      ...result,\n      outcome: {\n        baselineComposite: meanComposite(opts.baseline),\n        afterComposite: meanComposite(opts.baseline),\n        delta: 0,\n        perPersona: opts.baseline.map((b) => ({\n          personaId: b.personaId,\n          before: b.composite,\n          after: b.composite,\n          delta: 0,\n        })),\n        rolledBackPaths: [],\n      },\n    }\n  }\n\n  const personaIds = opts.baseline.map((b) => b.personaId)\n  const after = await opts.reRunCohort(personaIds)\n  const afterByPersona = new Map(after.map((r) => [r.personaId, r.composite]))\n\n  const perPersona = opts.baseline.map((b) => {\n    const a = afterByPersona.get(b.personaId) ?? b.composite\n    return { personaId: b.personaId, before: b.composite, after: a, delta: a - b.composite }\n  })\n  const baselineComposite = meanComposite(opts.baseline)\n  const afterComposite = meanComposite(after)\n  const delta = afterComposite - baselineComposite\n\n  let rolledBackPaths: string[] = []\n  if (delta < 0 && opts.rollbackOnRegression && opts.revert) {\n    await opts.revert(allApplied)\n    rolledBackPaths = [...allApplied]\n  }\n\n  return {\n    ...result,\n    outcome: {\n      baselineComposite,\n      afterComposite,\n      delta,\n      perPersona,\n      rolledBackPaths,\n    },\n  }\n}\n\nfunction meanComposite(rows: ReadonlyArray<{ composite: number }>): number {\n  if (rows.length === 0) return 0\n  return rows.reduce((acc, r) => acc + r.composite, 0) / rows.length\n}\n","/**\n * Sandbox bridge for `AgentRuntime.act` — prod-faithful eval execution.\n *\n * The point of this adapter is parity: the eval substrate must run the agent\n * through the SAME profile the production chat turn uses, or scorecard numbers\n * grade a profile that never ships. `createSandboxAct` boots a sandbox with the\n * agent's profile (the caller's `baseProfile`, with optional per-persona\n * overrides) through the loop kernel's own {@link createSandboxForSpec},\n * streams the `streamPrompt` events mapped to the `RuntimeStreamEvent`\n * vocabulary, and resolves the `OutputAdapter`-parsed output for rubric scoring\n * — satisfying the `act` streaming contract with one code path shared by chat\n * and eval.\n *\n * Agents with a bespoke streaming chat turn should wire THAT into `act`\n * directly (the contract is designed for it); this adapter is the default for\n * agents whose turn is a plain prod-profile sandbox dispatch — notably the\n * agents agent-builder generates.\n */\n\nimport type {\n  AgentProfile,\n  AgentProfileFileMount,\n  AgentProfileMcpServer,\n} from '@tangle-network/agent-interface'\nimport type { SandboxEvent } from '@tangle-network/sandbox'\nimport type { AgentRunSpec, OutputAdapter, SandboxClient } from '../runtime'\nimport { mapSandboxEvent } from '../runtime'\nimport { createSandboxForSpec } from '../runtime/run-loop'\nimport type { RuntimeStreamEvent } from '../types'\nimport type { AgentRunContext, AgentRunInvocation } from './define-agent'\n\n/** Per-persona profile-merge slots applied over the base profile (§1.5: the caller authors the\n *  per-persona profile). Each slot overlays the base; an absent slot leaves the base untouched. */\nexport interface SandboxActComposeOverrides {\n  /** Replace the base profile's system prompt (e.g. a workspace-augmented prompt). */\n  systemPrompt?: string\n  /** Extra file mounts layered after the base profile's `resources.files`. */\n  extraFiles?: AgentProfileFileMount[]\n  /** Override the profile `name`. Defaults to the base profile's name. */\n  name?: string\n  /** Box built-in tool ON/OFF flags merged over the base profile's `tools` (overlay wins per key). */\n  tools?: Record<string, boolean>\n  /** MCP connections merged over the base profile's `mcp` (overlay wins per key). */\n  mcpConnections?: Record<string, AgentProfileMcpServer>\n}\n\nexport interface CreateSandboxActOptions<TPersona, TRunOutput> {\n  /** Canonical agent profile — the same one the prod chat turn uses. */\n  baseProfile: AgentProfile\n  /** Sandbox client used to boot the per-run sandbox. */\n  sandboxClient: SandboxClient\n  /** Persona → prompt. Pure; the eval cell's input. */\n  buildPrompt: (persona: TPersona) => string\n  /** Sandbox event stream → typed output the rubric scores. */\n  output: OutputAdapter<TRunOutput>\n  /**\n   * Per-persona profile overrides (workspace-augmented system prompt, extra\n   * file mounts, tool flags, MCP connections). Overlaid onto `baseProfile`.\n   */\n  compose?: (persona: TPersona) => SandboxActComposeOverrides\n  /** Sandbox-SDK overrides forwarded to `createSandboxForSpec`. */\n  sandboxOverrides?: AgentRunSpec<unknown>['sandboxOverrides']\n  /** Stable run name surfaced in mapped `llm_call` events. */\n  name?: string\n  /** Override the `SandboxEvent → RuntimeStreamEvent` mapper. */\n  mapEvent?: (\n    event: SandboxEvent,\n    opts: { agentRunName?: string },\n  ) => RuntimeStreamEvent | undefined\n}\n\n/**\n * Build an `AgentRuntime.act` implementation backed by a single prod-profile\n * sandbox run. The returned function honours the `act` contract: it returns\n * synchronously with a live `events` iterator and an `output` promise that\n * resolves only after the iterator drains.\n */\nexport function createSandboxAct<TPersona, TRunOutput>(\n  options: CreateSandboxActOptions<TPersona, TRunOutput>,\n): (persona: TPersona, ctx: AgentRunContext) => AgentRunInvocation<TRunOutput> {\n  const mapEvent = options.mapEvent ?? mapSandboxEvent\n\n  return (persona: TPersona, ctx: AgentRunContext): AgentRunInvocation<TRunOutput> => {\n    const profile = applyComposeOverrides(options.baseProfile, options.compose?.(persona))\n    const agentRunName = options.name ?? profile.name ?? 'agent'\n    const message = options.buildPrompt(persona)\n    const signal = ctx.signal ?? new AbortController().signal\n\n    const raw: SandboxEvent[] = []\n    let settle!: (value: TRunOutput) => void\n    let fail!: (err: unknown) => void\n    const output = new Promise<TRunOutput>((resolve, reject) => {\n      settle = resolve\n      fail = reject\n    })\n    // The output promise rejects when the stream errors; if the caller ignores\n    // `output` (chat UX) the rejection is still observed by the events iterator\n    // throwing. Attach a no-op catch so an ignored rejection is never \"unhandled\".\n    output.catch(() => {})\n\n    const spec: AgentRunSpec<unknown> = {\n      profile,\n      taskToPrompt: () => message,\n      name: agentRunName,\n      ...(options.sandboxOverrides ? { sandboxOverrides: options.sandboxOverrides } : {}),\n    }\n\n    async function* events(): AsyncIterable<RuntimeStreamEvent> {\n      try {\n        const box = await createSandboxForSpec(options.sandboxClient, spec, signal)\n        for await (const event of box.streamPrompt(message, { signal })) {\n          raw.push(event)\n          const mapped = mapEvent(event, { agentRunName })\n          if (mapped) yield mapped\n        }\n        settle(options.output.parse(raw))\n      } catch (err) {\n        fail(err)\n        throw err\n      }\n    }\n\n    return { events: events(), output }\n  }\n}\n\n/** Overlay the per-persona overrides onto the base profile. Each slot merges over the base; an\n *  absent override leaves the base profile untouched. */\nfunction applyComposeOverrides(\n  base: AgentProfile,\n  overrides: SandboxActComposeOverrides | undefined,\n): AgentProfile {\n  if (!overrides) return base\n  const prompt = overrides.systemPrompt\n    ? { ...base.prompt, systemPrompt: overrides.systemPrompt }\n    : base.prompt\n  const mergedTools = overrides.tools ? { ...(base.tools ?? {}), ...overrides.tools } : base.tools\n  const mergedMcp = overrides.mcpConnections\n    ? { ...(base.mcp ?? {}), ...overrides.mcpConnections }\n    : base.mcp\n  const baseFiles = base.resources?.files ?? []\n  const mergedFiles: AgentProfileFileMount[] = overrides.extraFiles?.length\n    ? [...baseFiles, ...overrides.extraFiles]\n    : [...baseFiles]\n  return {\n    ...base,\n    name: overrides.name ?? base.name,\n    prompt,\n    ...(mergedTools ? { tools: mergedTools } : {}),\n    ...(mergedMcp ? { mcp: mergedMcp } : {}),\n    resources: { ...base.resources, files: mergedFiles },\n  }\n}\n"],"mappings":";;;;;;;;;;;;;;;AAeA,SAAS,kBAAkB;AAC3B,SAAS,YAAY,YAAY;AAqE1B,SAAS,mBACd,SACA,UACA,UACwB;AACxB,QAAM,aAAa,yBAAyB,SAAS,QAAQ;AAC7D,MAAI,WAAW,WAAW,EAAG,QAAO;AAMpC,aAAW,OAAO,YAAY;AAC5B,UAAM,MAAM,WAAW,GAAG,IAAI,MAAM,KAAK,UAAU,GAAG;AACtD,QAAI,WAAW,GAAG,GAAG;AACnB,aAAO,EAAE,cAAc,KAAK,kBAAkB,KAAK,QAAQ,MAAM,QAAQ,gBAAgB;AAAA,IAC3F;AAAA,EACF;AACA,QAAM,WAAW,WAAW,CAAC;AAC7B,QAAM,cAAc,WAAW,QAAQ,IAAI,WAAW,KAAK,UAAU,QAAQ;AAC7E,SAAO;AAAA,IACL,cAAc;AAAA,IACd,kBAAkB;AAAA,IAClB,QAAQ;AAAA,IACR,QAAQ;AAAA,EACV;AACF;AAEA,SAAS,yBACP,SACA,UACuB;AACvB,UAAQ,QAAQ,MAAM;AAAA,IACpB,KAAK;AAAA,IACL,KAAK;AACH,aAAO,CAAC,KAAK,SAAS,WAAW,GAAG,QAAQ,IAAI,KAAK,CAAC;AAAA,IACxD,KAAK;AAEH,aAAO,CAAC,KAAK,SAAS,WAAW,UAAU,GAAG,QAAQ,QAAQ,KAAK,CAAC,KAAK,CAAC;AAAA,IAC5E,KAAK;AACH,aAAO,CAAC,KAAK,SAAS,WAAW,OAAO,GAAG,QAAQ,QAAQ,KAAK,CAAC;AAAA,IACnE,KAAK,iBAAiB;AACpB,YAAM,OAAO,QAAQ,QAAQ,OAAO;AAIpC,aAAO;AAAA,QACL,KAAK,SAAS,cAAc,GAAG,IAAI,KAAK;AAAA,QACxC,KAAK,SAAS,cAAc,MAAM,UAAU;AAAA,QAC5C,KAAK,SAAS,cAAc,MAAM,UAAU;AAAA,MAC9C;AAAA,IACF;AAAA,IACA,KAAK;AACH,UAAI,QAAQ,QAAQ;AAClB,eAAO,CAAC,KAAK,SAAS,OAAO,QAAQ,MAAM,GAAG,QAAQ,QAAQ,MAAM,CAAC,KAAK,CAAC;AAAA,MAC7E;AAGA,aAAO;AAAA,QACL,KAAK,SAAS,OAAO,QAAQ,MAAM,WAAW;AAAA,QAC9C,KAAK,SAAS,OAAO,GAAG,QAAQ,IAAI,KAAK;AAAA,MAC3C;AAAA,IACF,KAAK;AACH,aAAO,CAAC,KAAK,SAAS,OAAO,QAAQ,MAAM,WAAW,CAAC;AAAA,IACzD,KAAK;AACH,UAAI,CAAC,SAAS,IAAK,QAAO,CAAC;AAC3B,aAAO,CAAC,KAAK,SAAS,KAAK,QAAQ,QAAQ,GAAG,QAAQ,KAAK,KAAK,CAAC;AAAA,IACnE,KAAK;AACH,UAAI,CAAC,SAAS,OAAQ,QAAO,CAAC;AAC9B,aAAO,CAAC,KAAK,SAAS,QAAQ,GAAG,QAAQ,QAAQ,GAAG,CAAC,OAAO,CAAC;AAAA,IAC/D,KAAK;AACH,UAAI,CAAC,SAAS,YAAa,QAAO,CAAC;AACnC,aAAO,CAAC,KAAK,SAAS,aAAa,GAAG,QAAQ,QAAQ,OAAO,CAAC,KAAK,CAAC;AAAA,IACtE,KAAK;AACH,UAAI,CAAC,SAAS,aAAc,QAAO,CAAC;AACpC,aAAO,CAAC,SAAS,YAAY;AAAA,IAC/B,KAAK;AAAA,IACL,KAAK;AAGH,aAAO,CAAC;AAAA,IACV,KAAK;AAEH,aAAO,CAAC;AAAA,EACZ;AACF;AAEA,SAAS,QAAQ,GAAmB;AAClC,SACE,EACG,YAAY,EACZ,QAAQ,gBAAgB,GAAG,EAC3B,QAAQ,YAAY,EAAE,EACtB,MAAM,GAAG,GAAG,KAAK;AAExB;AAgBO,SAAS,iBACd,UACA,UACuC;AACvC,QAAM,SAAmC,CAAC;AAC1C,QAAM,cAAkD;AAAA,IACtD;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF;AACA,QAAM,eAAmD,CAAC,QAAQ;AAClE,QAAM,sBAA0D,CAAC,eAAe,UAAU,KAAK;AAC/F,QAAM,uBAA2D,CAAC,cAAc;AAEhF,aAAW,OAAO,aAAa;AAC7B,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,CAAC,GAAG;AACN,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,IAAI,QAAQ,UAAU,CAAC;AACzD;AAAA,IACF;AACA,UAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;AAChD,QAAI,CAAC,WAAW,GAAG,GAAG;AACpB,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,GAAG,QAAQ,UAAU,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,aAAW,OAAO,cAAc;AAC9B,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,CAAC,GAAG;AACN,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,IAAI,QAAQ,UAAU,CAAC;AACzD;AAAA,IACF;AACA,UAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;AAChD,QAAI,CAAC,WAAW,GAAG,GAAG;AACpB,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,GAAG,QAAQ,UAAU,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,aAAW,OAAO,CAAC,GAAG,qBAAqB,GAAG,oBAAoB,GAAG;AACnE,UAAM,IAAI,SAAS,GAAG;AACtB,QAAI,MAAM,OAAW;AACrB,UAAM,MAAM,WAAW,CAAC,IAAI,IAAI,KAAK,UAAU,CAAC;AAChD,QAAI,CAAC,WAAW,GAAG,GAAG;AACpB,aAAO,KAAK,EAAE,SAAS,KAAK,MAAM,GAAG,QAAQ,UAAU,CAAC;AAAA,IAC1D;AAAA,EACF;AACA,SAAO;AACT;AAEO,SAAS,oBACd,QACA,UACQ;AACR,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,QAAM,QAAQ,OAAO;AAAA,IACnB,CAAC,MAAM,OAAO,EAAE,OAAO,KAAK,EAAE,OAAO,IAAI,EAAE,IAAI,MAAM,WAAW,KAAK,EAAE,MAAM;AAAA,EAC/E;AACA,SAAO;AAAA,IACL,oDAAoD,QAAQ;AAAA,IAC5D,GAAG;AAAA,IACH;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACb;;;AC5DO,SAAS,sBACd,SAAS,uDACuB;AAChC,SAAO;AAAA,IACL,SAAS,gBAAgB,QAA2C;AAAA,IAAC,GAAG;AAAA,IACxE,QAAQ,QAAQ,OAAO,IAAI,MAAM,MAAM,CAAC;AAAA,EAC1C;AACF;AAYA,eAAsB,gBACpB,YAC4E;AAC5E,QAAM,SAA+B,CAAC;AACtC,mBAAiB,MAAM,WAAW,OAAQ,QAAO,KAAK,EAAE;AACxD,QAAM,SAAS,MAAM,WAAW;AAChC,SAAO,EAAE,QAAQ,OAAO;AAC1B;AA2CO,IAAM,qBAAN,cAAiC,MAAM;AAAA,EAC5C,YACE,SACgB,SACA,SAAiC,CAAC,GAClD;AACA,UAAM,OAAO;AAHG;AACA;AAGhB,SAAK,OAAO;AAAA,EACd;AAAA,EALkB;AAAA,EACA;AAKpB;AAeO,SAAS,YACd,UACqC;AACrC,MAAI,CAAC,SAAS,MAAM,SAAS,GAAG,KAAK,EAAE,WAAW,GAAG;AACnD,UAAM,IAAI,mBAAmB,iCAAiC,SAAS,MAAM,EAAE;AAAA,EACjF;AACA,MAAI,CAAC,SAAS,YAAY,SAAS,SAAS,KAAK,EAAE,WAAW,GAAG;AAC/D,UAAM,IAAI,mBAAmB,uCAAuC,SAAS,EAAE;AAAA,EACjF;AACA,QAAM,SAAS,iBAAiB,SAAS,UAAU,SAAS,QAAQ;AACpE,MAAI,OAAO,SAAS,GAAG;AACrB,UAAM,IAAI;AAAA,MACR,oBAAoB,QAAQ,SAAS,QAAQ;AAAA,MAC7C,SAAS;AAAA,MACT;AAAA,IACF;AAAA,EACF;AAGA,QAAM,QAAQ,SAAS,OAAO,WAAW,OAAO,CAAC,KAAK,MAAM,MAAM,EAAE,QAAQ,CAAC;AAC7E,MAAI,SAAS,OAAO,WAAW,SAAS,MAAM,QAAQ,OAAO,QAAQ,MAAM;AACzE,UAAM,IAAI;AAAA,MACR,eAAe,SAAS,EAAE,sCAAsC,MAAM,QAAQ,CAAC,CAAC;AAAA,MAChF,SAAS;AAAA,IACX;AAAA,EACF;AACA,SAAO;AACT;;;AClSA,SAAS,iBAAiB;AAC1B,SAAS,oBAAoB;AAE7B,SAAS,2BAA2B;AAsFpC,IAAM,uBAA8D;AAAA,EAClE;AAAA,EACA;AAAA,EACA;AAAA,EACA;AACF;AAEO,SAAS,gCACd,MAC4C;AAC5C,QAAM,OAAO,KAAK,QAAQ;AAC1B,QAAM,cAAc,KAAK,uBAAuB;AAEhD,SAAO;AAAA,IACL,MAAM,oBAAoB,UAAU;AAClC,YAAM,QAAkC,CAAC;AACzC,YAAM,SAAyE,CAAC;AAChF,UAAI,UAAU;AAEd,iBAAW,KAAK,UAAU;AACxB,cAAM,UAAU,oBAAoB,EAAE,OAAO;AAC7C,YAAI,YAAY,MAAM;AACpB,cAAI,EAAE,YAAY,QAAW;AAC3B,mBAAO,KAAK;AAAA,cACV,WAAW,EAAE;AAAA,cACb,SAAS,EAAE;AAAA,cACX,SAAS;AAAA,YACX,CAAC;AAAA,UACH,OAAO;AAGL,uBAAW;AAAA,UACb;AACA;AAAA,QACF;AAGA,YAAI,QAAQ,SAAS,WAAW;AAC9B,qBAAW;AACX;AAAA,QACF;AAIA,YAAI,QAAQ,KAAK,WAAW,YAAY,GAAG;AACzC,qBAAW;AACX;AAAA,QACF;AAEA,cAAM,SAAS,mBAAmB,SAAS,KAAK,UAAU,KAAK,QAAQ;AACvE,YAAI,WAAW,MAAM;AACnB,iBAAO,KAAK;AAAA,YACV,WAAW,EAAE;AAAA,YACb,SAAS,EAAE,WAAW;AAAA,YACtB,SAAS,iBAAiB,QAAQ,IAAI;AAAA,UACxC,CAAC;AACD;AAAA,QACF;AAEA,YAAI,OAAO,WAAW,gBAAgB,CAAC,YAAY,SAAS,QAAQ,IAAI,GAAG;AACzE,iBAAO,KAAK;AAAA,YACV,WAAW,EAAE;AAAA,YACb,SAAS,EAAE,WAAW;AAAA,YACtB,SAAS,UAAU,OAAO,gBAAgB,8BAA8B,QAAQ,IAAI;AAAA,UACtF,CAAC;AACD;AAAA,QACF;AAEA,cAAM,iBAAiB,OAAO,SAAS,aAAa,OAAO,cAAc,OAAO,IAAI;AAEpF,YAAI;AACJ,YAAI;AACF,kBAAQ,MAAM,KAAK,WAAW,EAAE,SAAS,GAAG,SAAS,QAAQ,eAAe,CAAC;AAAA,QAC/E,SAAS,KAAK;AACZ,iBAAO,KAAK;AAAA,YACV,WAAW,EAAE;AAAA,YACb,SAAS,EAAE,WAAW;AAAA,YACtB,SAAS,qBAAqB,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UAChF,CAAC;AACD;AAAA,QACF;AAEA,YAAI,MAAM,MAAM,KAAK,EAAE,WAAW,GAAG;AACnC,qBAAW;AACX;AAAA,QACF;AAEA,cAAM,KAAK;AAAA,UACT,IAAI,OAAO,EAAE,UAAU;AAAA,UACvB,iBAAiB,EAAE;AAAA,UACnB;AAAA,UACA;AAAA,UACA,YAAY,OAAO,cAAc;AAAA,UACjC,OAAO,MAAM;AAAA,UACb,SAAS,MAAM;AAAA,UACf,WAAW,MAAM;AAAA,UACjB,YAAY,EAAE;AAAA,UACd,UAAU,EAAE;AAAA,QACd,CAAC;AAAA,MACH;AAEA,aAAO,EAAE,OAAO,SAAS,OAAO;AAAA,IAClC;AAAA,IAEA,MAAM,MAAM,OAAO;AACjB,YAAM,UAAoB,CAAC;AAC3B,YAAM,WAAqB,CAAC;AAE5B,UAAI,SAAS,QAAQ;AACnB,iBAAS;AAAA,UACP;AAAA,QACF;AACA,eAAO,EAAE,SAAS,SAAS;AAAA,MAC7B;AAEA,UAAI,SAAS,aAAa,CAAC,KAAK,QAAQ;AACtC,iBAAS;AAAA,UACP;AAAA,QACF;AACA,eAAO,EAAE,SAAS,SAAS;AAAA,MAC7B;AAEA,iBAAW,QAAQ,OAAO;AAIxB,cAAM,UAAU,KAAK,OAAO,SAAS,aAAa,KAAK,OAAO,cAAc,OAAO,IAAI;AACvF,YAAI,OAAO,OAAO,MAAM,KAAK,YAAY;AACvC,mBAAS;AAAA,YACP,GAAG,KAAK,OAAO,gBAAgB;AAAA,UACjC;AACA;AAAA,QACF;AAEA,cAAM,KAAK,kBAAkB,MAAM,KAAK,QAAQ;AAChD,YAAI,CAAC,IAAI;AACP,mBAAS,KAAK,GAAG,KAAK,OAAO,gBAAgB,oBAAoB;AACjE;AAAA,QACF;AACA,gBAAQ,KAAK,KAAK,OAAO,gBAAgB;AAAA,MAC3C;AAEA,UAAI,SAAS,aAAa,QAAQ,SAAS,KAAK,KAAK,QAAQ;AAC3D,cAAM,QAAQ;AAAA,UACZ;AAAA,UACA,MAAM,OAAO,CAAC,MAAM,QAAQ,SAAS,EAAE,OAAO,gBAAgB,CAAC;AAAA,UAC/D,KAAK;AAAA,UACL,KAAK;AAAA,UACL,KAAK,cAAc;AAAA,QACrB;AACA,YAAI,MAAO,UAAS,KAAK,cAAc,KAAK,EAAE;AAAA,YACzC,UAAS,KAAK,gEAAgE;AAAA,MACrF;AAEA,aAAO,EAAE,SAAS,SAAS;AAAA,IAC7B;AAAA,EACF;AACF;AAIA,SAAS,kBAAkB,MAA8B,UAA2B;AAClF,QAAM,SAAS,UAAU,OAAO,CAAC,SAAS,oBAAoB,OAAO,GAAG,GAAG;AAAA,IACzE,KAAK;AAAA,IACL,OAAO,KAAK;AAAA,IACZ,UAAU;AAAA,EACZ,CAAC;AACD,SAAO,OAAO,WAAW;AAC3B;AAEA,SAAS,gBACP,OACA,OACA,UACA,QACA,YACe;AACf,QAAM,SAAS,gBAAgB,KAAK,IAAI,CAAC,IAAI,MAAM,CAAC,GAAG,gBAAgB,MAAM,GAAG,EAAE,KAAK,OAAO;AAE9F,QAAM,WAAW,UAAU,OAAO,CAAC,YAAY,MAAM,MAAM,GAAG,EAAE,KAAK,SAAS,CAAC;AAC/E,MAAI,SAAS,WAAW,EAAG,QAAO;AAClC,QAAM,MAAM,UAAU,OAAO,CAAC,OAAO,GAAG,KAAK,GAAG,EAAE,KAAK,SAAS,CAAC;AACjE,MAAI,IAAI,WAAW,EAAG,QAAO;AAC7B,QAAM,QAAQ,iBAAiB,MAAM,CAAC,GAAG,WAAW,GAAG,MAAM,MAAM,oBAAoB;AACvF,QAAM,OAAO;AAAA,IACX;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG,MAAM;AAAA,MACP,CAAC,MACC,OAAO,EAAE,eAAe,gBAAgB,EAAE,WAAW,QAAQ,CAAC,CAAC,cAAc,EAAE,QAAQ;AAAA,IAC3F;AAAA,IACA;AAAA,IACA;AAAA,IACA,GAAG,MAAM,IAAI,CAAC,MAAM;AAAA,KAAQ,EAAE,OAAO,gBAAgB;AAAA;AAAA,EAAO,EAAE,SAAS,EAAE;AAAA,EAC3E,EAAE,KAAK,IAAI;AACX,QAAM,SAAS,UAAU,OAAO,CAAC,UAAU,MAAM,OAAO,MAAM,IAAI,GAAG,EAAE,KAAK,SAAS,CAAC;AACtF,MAAI,OAAO,WAAW,EAAG,QAAO;AAChC,QAAM,OAAO,UAAU,OAAO,CAAC,QAAQ,MAAM,UAAU,MAAM,GAAG,EAAE,KAAK,SAAS,CAAC;AACjF,MAAI,KAAK,WAAW,EAAG,QAAO;AAC9B,QAAM,KAAK;AAAA,IACT;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA,EAAE,KAAK,UAAU,UAAU,QAAQ;AAAA,EACrC;AACA,MAAI,GAAG,WAAW,EAAG,QAAO;AAC5B,SAAO,GAAG,OAAO,KAAK;AACxB;AAEA,SAAS,OAAO,GAAmB;AAGjC,QAAM,SAAS,UAAQ,QAAa;AACpC,SAAO,OAAO,WAAW,QAAQ,EAAE,OAAO,GAAG,OAAO,EAAE,OAAO,KAAK;AACpE;;;ACnSO,SAAS,8BACd,MACA,MAC6B;AAC7B,SAAO;AAAA,IACL,oBAAoB,UAAU;AAC5B,YAAM,QAAQ,KAAK,oBAAoB,QAAQ;AAC/C,aAAO;AAAA,QACL,WAAW,MAAM;AAAA,QACjB,SAAS,MAAM;AAAA,QACf,QAAQ,MAAM;AAAA,MAChB;AAAA,IACF;AAAA,IACA,MAAM,MAAM,WAAW;AACrB,YAAM,UAAoB,CAAC;AAC3B,YAAM,WAAqB,CAAC;AAC5B,iBAAW,KAAK,WAAW;AACzB,cAAM,eAAe,2BAA2B,CAAC;AACjD,YAAI,iBAAiB,MAAM;AACzB,mBAAS;AAAA,YACP,kEAAkE,mBAAmB,CAAC,CAAC;AAAA,UACzF;AACA;AAAA,QACF;AACA,YAAI;AACF,gBAAM,IAAI,MAAM,KAAK,0BAA0B,KAAK,eAAe,YAAY;AAC/E,kBAAQ,KAAK,GAAG,EAAE,OAAO;AACzB,mBAAS,KAAK,GAAG,EAAE,QAAQ;AAAA,QAC7B,SAAS,KAAK;AACZ,mBAAS;AAAA,YACP,qCAAqC,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UACvF;AAAA,QACF;AAAA,MACF;AACA,UAAI,KAAK,kBAAkB,QAAQ,SAAS,GAAG;AAC7C,YAAI;AACF,gBAAM,aAAa,MAAM,KAAK,eAAe,KAAK,aAAa;AAC/D,qBAAW,SAAS,WAAY,UAAS,KAAK,SAAS,KAAK,EAAE;AAAA,QAChE,SAAS,KAAK;AACZ,mBAAS;AAAA,YACP,0BAA0B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC;AAAA,UAC5E;AAAA,QACF;AAAA,MACF;AACA,aAAO,EAAE,SAAS,SAAS;AAAA,IAC7B;AAAA,EACF;AACF;AAYA,SAAS,2BAA2B,GAA2B;AAC7D,MAAI,CAAC,KAAK,OAAO,MAAM,SAAU,QAAO;AACxC,QAAM,MAAM;AACZ,MAAI,MAAM,QAAQ,IAAI,WAAW,GAAG;AAClC,UAAM,SAAS,IAAI;AACnB,QAAI,OAAO,WAAW,EAAG,QAAO;AAChC,WAAO,OACJ,IAAI,CAAC,MAAO,OAAO,EAAE,YAAY,WAAW,EAAE,UAAU,EAAG,EAC3D,OAAO,CAAC,MAAM,EAAE,SAAS,CAAC,EAC1B,KAAK,MAAM;AAAA,EAChB;AACA,MAAI,OAAO,IAAI,iBAAiB,SAAU,QAAO,IAAI;AACrD,MAAI,OAAO,IAAI,YAAY,SAAU,QAAO,IAAI;AAChD,SAAO;AACT;AAEA,SAAS,mBAAmB,GAAoB;AAC9C,MAAI,CAAC,KAAK,OAAO,MAAM,SAAU,QAAO;AACxC,QAAM,MAAM;AACZ,MAAI,OAAO,IAAI,oBAAoB,SAAU,QAAO,IAAI;AACxD,MAAI,OAAO,IAAI,OAAO,SAAU,QAAO,IAAI;AAC3C,SAAO;AACT;;;AC1EA,eAAsB,eACpB,QACA,MACmF;AACnF,QAAM,UAAU,OAAO,WAAW,WAAW,CAAC;AAC9C,QAAM,sBAAsB,OAAO,aAAa,WAAW,CAAC;AAC5D,QAAM,aAAa,CAAC,GAAG,SAAS,GAAG,mBAAmB;AAItD,MAAI,WAAW,WAAW,GAAG;AAC3B,WAAO;AAAA,MACL,GAAG;AAAA,MACH,SAAS;AAAA,QACP,mBAAmB,cAAc,KAAK,QAAQ;AAAA,QAC9C,gBAAgB,cAAc,KAAK,QAAQ;AAAA,QAC3C,OAAO;AAAA,QACP,YAAY,KAAK,SAAS,IAAI,CAAC,OAAO;AAAA,UACpC,WAAW,EAAE;AAAA,UACb,QAAQ,EAAE;AAAA,UACV,OAAO,EAAE;AAAA,UACT,OAAO;AAAA,QACT,EAAE;AAAA,QACF,iBAAiB,CAAC;AAAA,MACpB;AAAA,IACF;AAAA,EACF;AAEA,QAAM,aAAa,KAAK,SAAS,IAAI,CAAC,MAAM,EAAE,SAAS;AACvD,QAAM,QAAQ,MAAM,KAAK,YAAY,UAAU;AAC/C,QAAM,iBAAiB,IAAI,IAAI,MAAM,IAAI,CAAC,MAAM,CAAC,EAAE,WAAW,EAAE,SAAS,CAAC,CAAC;AAE3E,QAAM,aAAa,KAAK,SAAS,IAAI,CAAC,MAAM;AAC1C,UAAM,IAAI,eAAe,IAAI,EAAE,SAAS,KAAK,EAAE;AAC/C,WAAO,EAAE,WAAW,EAAE,WAAW,QAAQ,EAAE,WAAW,OAAO,GAAG,OAAO,IAAI,EAAE,UAAU;AAAA,EACzF,CAAC;AACD,QAAM,oBAAoB,cAAc,KAAK,QAAQ;AACrD,QAAM,iBAAiB,cAAc,KAAK;AAC1C,QAAM,QAAQ,iBAAiB;AAE/B,MAAI,kBAA4B,CAAC;AACjC,MAAI,QAAQ,KAAK,KAAK,wBAAwB,KAAK,QAAQ;AACzD,UAAM,KAAK,OAAO,UAAU;AAC5B,sBAAkB,CAAC,GAAG,UAAU;AAAA,EAClC;AAEA,SAAO;AAAA,IACL,GAAG;AAAA,IACH,SAAS;AAAA,MACP;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAEA,SAAS,cAAc,MAAoD;AACzE,MAAI,KAAK,WAAW,EAAG,QAAO;AAC9B,SAAO,KAAK,OAAO,CAAC,KAAK,MAAM,MAAM,EAAE,WAAW,CAAC,IAAI,KAAK;AAC9D;;;AChDO,SAAS,iBACd,SAC6E;AAC7E,QAAM,WAAW,QAAQ,YAAY;AAErC,SAAO,CAAC,SAAmB,QAAyD;AAClF,UAAM,UAAU,sBAAsB,QAAQ,aAAa,QAAQ,UAAU,OAAO,CAAC;AACrF,UAAM,eAAe,QAAQ,QAAQ,QAAQ,QAAQ;AACrD,UAAM,UAAU,QAAQ,YAAY,OAAO;AAC3C,UAAM,SAAS,IAAI,UAAU,IAAI,gBAAgB,EAAE;AAEnD,UAAM,MAAsB,CAAC;AAC7B,QAAI;AACJ,QAAI;AACJ,UAAM,SAAS,IAAI,QAAoB,CAAC,SAAS,WAAW;AAC1D,eAAS;AACT,aAAO;AAAA,IACT,CAAC;AAID,WAAO,MAAM,MAAM;AAAA,IAAC,CAAC;AAErB,UAAM,OAA8B;AAAA,MAClC;AAAA,MACA,cAAc,MAAM;AAAA,MACpB,MAAM;AAAA,MACN,GAAI,QAAQ,mBAAmB,EAAE,kBAAkB,QAAQ,iBAAiB,IAAI,CAAC;AAAA,IACnF;AAEA,oBAAgB,SAA4C;AAC1D,UAAI;AACF,cAAM,MAAM,MAAM,qBAAqB,QAAQ,eAAe,MAAM,MAAM;AAC1E,yBAAiB,SAAS,IAAI,aAAa,SAAS,EAAE,OAAO,CAAC,GAAG;AAC/D,cAAI,KAAK,KAAK;AACd,gBAAM,SAAS,SAAS,OAAO,EAAE,aAAa,CAAC;AAC/C,cAAI,OAAQ,OAAM;AAAA,QACpB;AACA,eAAO,QAAQ,OAAO,MAAM,GAAG,CAAC;AAAA,MAClC,SAAS,KAAK;AACZ,aAAK,GAAG;AACR,cAAM;AAAA,MACR;AAAA,IACF;AAEA,WAAO,EAAE,QAAQ,OAAO,GAAG,OAAO;AAAA,EACpC;AACF;AAIA,SAAS,sBACP,MACA,WACc;AACd,MAAI,CAAC,UAAW,QAAO;AACvB,QAAM,SAAS,UAAU,eACrB,EAAE,GAAG,KAAK,QAAQ,cAAc,UAAU,aAAa,IACvD,KAAK;AACT,QAAM,cAAc,UAAU,QAAQ,EAAE,GAAI,KAAK,SAAS,CAAC,GAAI,GAAG,UAAU,MAAM,IAAI,KAAK;AAC3F,QAAM,YAAY,UAAU,iBACxB,EAAE,GAAI,KAAK,OAAO,CAAC,GAAI,GAAG,UAAU,eAAe,IACnD,KAAK;AACT,QAAM,YAAY,KAAK,WAAW,SAAS,CAAC;AAC5C,QAAM,cAAuC,UAAU,YAAY,SAC/D,CAAC,GAAG,WAAW,GAAG,UAAU,UAAU,IACtC,CAAC,GAAG,SAAS;AACjB,SAAO;AAAA,IACL,GAAG;AAAA,IACH,MAAM,UAAU,QAAQ,KAAK;AAAA,IAC7B;AAAA,IACA,GAAI,cAAc,EAAE,OAAO,YAAY,IAAI,CAAC;AAAA,IAC5C,GAAI,YAAY,EAAE,KAAK,UAAU,IAAI,CAAC;AAAA,IACtC,WAAW,EAAE,GAAG,KAAK,WAAW,OAAO,YAAY;AAAA,EACrD;AACF;","names":[]}

package/dist/analyst-loop.d.ts

@@ -1,5 +1,5 @@
 import { TraceAnalystByteBudgets, TraceAnalysisStore } from '@tangle-network/agent-eval';
-import { I as Iteration } from './types-CdnEAE3U.js';
+import { I as Iteration } from './types-JufmXF2a.js';
 import { R as RunAnalystLoopOpts, a as RunAnalystLoopResult } from './types-BC3bZpH0.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-BC3bZpH0.js';
 import '@tangle-network/sandbox';

package/dist/analyst-loop.js

@@ -1,8 +1,9 @@
 import {
   iterationsToTraceStore,
   runAnalystLoop
-} from "./chunk-P5OKDSLB.js";
-import "./chunk-VLF5RHEQ.js";
+} from "./chunk-IODKUOBA.js";
+import "./chunk-T2HVQVB4.js";
+import "./chunk-NBV35BR6.js";
 import "./chunk-DGUM43GV.js";
 export {
   iterationsToTraceStore,

package/dist/{chunk-4KGQHS7U.js → chunk-7ODB76J5.js}

@@ -1,20 +1,14 @@
 import {
-  DELEGATE_CODE_DESCRIPTION,
-  DELEGATE_CODE_INPUT_SCHEMA,
-  DELEGATE_CODE_TOOL_NAME,
   DELEGATE_FEEDBACK_DESCRIPTION,
   DELEGATE_FEEDBACK_INPUT_SCHEMA,
   DELEGATE_FEEDBACK_TOOL_NAME,
-  DELEGATE_RESEARCH_DESCRIPTION,
-  DELEGATE_RESEARCH_INPUT_SCHEMA,
-  DELEGATE_RESEARCH_TOOL_NAME,
   DELEGATION_HISTORY_DESCRIPTION,
   DELEGATION_HISTORY_INPUT_SCHEMA,
   DELEGATION_HISTORY_TOOL_NAME,
   DELEGATION_STATUS_DESCRIPTION,
   DELEGATION_STATUS_INPUT_SCHEMA,
   DELEGATION_STATUS_TOOL_NAME
-} from "./chunk-5ISW5JUF.js";
+} from "./chunk-NCH4XUZ7.js";
 
 // src/mcp/openai-tools.ts
 function buildTool(name, description, parameters) {
@@ -25,16 +19,6 @@ function buildTool(name, description, parameters) {
 }
 function mcpToolsForRuntimeMcp() {
   return [
-    buildTool(
-      DELEGATE_CODE_TOOL_NAME,
-      DELEGATE_CODE_DESCRIPTION,
-      DELEGATE_CODE_INPUT_SCHEMA
-    ),
-    buildTool(
-      DELEGATE_RESEARCH_TOOL_NAME,
-      DELEGATE_RESEARCH_DESCRIPTION,
-      DELEGATE_RESEARCH_INPUT_SCHEMA
-    ),
     buildTool(
       DELEGATE_FEEDBACK_TOOL_NAME,
       DELEGATE_FEEDBACK_DESCRIPTION,
@@ -61,4 +45,4 @@ export {
   mcpToolsForRuntimeMcp,
   mcpToolsForRuntimeMcpSubset
 };
-//# sourceMappingURL=chunk-4KGQHS7U.js.map
+//# sourceMappingURL=chunk-7ODB76J5.js.map

package/dist/chunk-7ODB76J5.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/mcp/openai-tools.ts"],"sourcesContent":["/**\n * @experimental\n *\n * OpenAI Chat Completions `tools[]` projection of the queue-bound agent-runtime\n * MCP delegation tools.\n *\n * Use when configuring `createOpenAICompatibleBackend({ tools: ... })` so the\n * model can call `delegate_feedback`, `delegation_status`, and\n * `delegation_history` through the OpenAI-compat transport (tcloud, OpenRouter,\n * OpenAI direct, cli-bridge). The runtime surfaces tool calls as `tool_call`\n * stream events — execution is the caller's responsibility (typically the\n * parent sandbox runtime's MCP mount).\n *\n * Sandbox-SDK callers do NOT need this helper: the sandbox runtime mounts\n * MCP servers natively and the in-sandbox harness discovers tools via the\n * runtime, not via an OpenAI tools array.\n *\n * Tool name + description + JSON-schema are pulled from the canonical\n * `DELEGATE_*` constants exported by `./tools/*` so the projection cannot\n * drift from the server's own validators.\n */\n\nimport type { OpenAIChatTool } from '../types'\nimport {\n  DELEGATE_FEEDBACK_DESCRIPTION,\n  DELEGATE_FEEDBACK_INPUT_SCHEMA,\n  DELEGATE_FEEDBACK_TOOL_NAME,\n} from './tools/delegate-feedback'\nimport {\n  DELEGATION_HISTORY_DESCRIPTION,\n  DELEGATION_HISTORY_INPUT_SCHEMA,\n  DELEGATION_HISTORY_TOOL_NAME,\n} from './tools/delegation-history'\nimport {\n  DELEGATION_STATUS_DESCRIPTION,\n  DELEGATION_STATUS_INPUT_SCHEMA,\n  DELEGATION_STATUS_TOOL_NAME,\n} from './tools/delegation-status'\n\nfunction buildTool(\n  name: string,\n  description: string,\n  parameters: Readonly<Record<string, unknown>>,\n): OpenAIChatTool {\n  // `parameters` arrives as a deeply-readonly `as const` literal. The\n  // OpenAI-compat backend JSON-serializes the body so a shallow copy\n  // into a plain object is sufficient — and shields callers that mutate\n  // the returned descriptor from corrupting the source constant.\n  return {\n    type: 'function',\n    function: { name, description, parameters: { ...parameters } },\n  }\n}\n\n/**\n * @experimental\n *\n * Returns the queue-bound delegation tools projected into OpenAI Chat\n * Completions `tools[]` shape. The order is stable: `delegate_feedback`,\n * `delegation_status`, `delegation_history`.\n */\nexport function mcpToolsForRuntimeMcp(): OpenAIChatTool[] {\n  return [\n    buildTool(\n      DELEGATE_FEEDBACK_TOOL_NAME,\n      DELEGATE_FEEDBACK_DESCRIPTION,\n      DELEGATE_FEEDBACK_INPUT_SCHEMA as Readonly<Record<string, unknown>>,\n    ),\n    buildTool(\n      DELEGATION_STATUS_TOOL_NAME,\n      DELEGATION_STATUS_DESCRIPTION,\n      DELEGATION_STATUS_INPUT_SCHEMA as Readonly<Record<string, unknown>>,\n    ),\n    buildTool(\n      DELEGATION_HISTORY_TOOL_NAME,\n      DELEGATION_HISTORY_DESCRIPTION,\n      DELEGATION_HISTORY_INPUT_SCHEMA as Readonly<Record<string, unknown>>,\n    ),\n  ]\n}\n\n/**\n * @experimental\n *\n * Subset filter — return only the projected tools whose `function.name`\n * appears in `names`. Useful for curated mounts (e.g. only the queue-bound\n * delegation tools, omitting `delegate_feedback`). Unknown names are\n * silently ignored; pass an empty array to get an empty result.\n */\nexport function mcpToolsForRuntimeMcpSubset(names: ReadonlyArray<string>): OpenAIChatTool[] {\n  const allowed = new Set(names)\n  return mcpToolsForRuntimeMcp().filter((tool) => allowed.has(tool.function.name))\n}\n"],"mappings":";;;;;;;;;;;;;AAuCA,SAAS,UACP,MACA,aACA,YACgB;AAKhB,SAAO;AAAA,IACL,MAAM;AAAA,IACN,UAAU,EAAE,MAAM,aAAa,YAAY,EAAE,GAAG,WAAW,EAAE;AAAA,EAC/D;AACF;AASO,SAAS,wBAA0C;AACxD,SAAO;AAAA,IACL;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,IACA;AAAA,MACE;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACF;AAUO,SAAS,4BAA4B,OAAgD;AAC1F,QAAM,UAAU,IAAI,IAAI,KAAK;AAC7B,SAAO,sBAAsB,EAAE,OAAO,CAAC,SAAS,QAAQ,IAAI,KAAK,SAAS,IAAI,CAAC;AACjF;","names":[]}