tangle-network · drewstone · Jun 23, 2026 · Jun 23, 2026
diff --git a/bench/src/atom-humaneval.mts b/bench/src/atom-humaneval.mts
@@ -34,9 +34,7 @@ import {
   routerBrain,
   routerChatWithUsage,
 } from '../../src/runtime/index'
-import { createReplayRecorder, renderReplayHtml } from '../../src/topology/replay'
 import { basePrompt, extractCode, type HumanEvalTask, loadHumanEval, runChecker } from './benchmarks/humaneval'
-import { writeFileSync } from 'node:fs'
 
 function must(k: string): string {
   const v = process.env[k]
@@ -106,10 +104,9 @@ interface TaskOutcome {
 // ── Driver arm: the orchestrated atom ────────────────────────────────────────────────────────
 async function driveTask(
   task: HumanEvalTask,
-): Promise<{ delivered: boolean; spawns: number; tokens: number; replay: string }> {
+): Promise<{ delivered: boolean; spawns: number; tokens: number }> {
   const blobs = new InMemoryResultBlobStore()
   const journal = new InMemorySpawnJournal()
-  const recorder = createReplayRecorder()
   let spawns = 0
   const makeWorker = (): Agent<unknown, unknown> => {
     const w = humanEvalWorker(task, `w-${spawns}`)
@@ -134,17 +131,13 @@ async function driveTask(
     blobs,
     executors: createExecutorRegistry(),
     maxDepth: 4,
-    hooks: recorder.hooks,
     now: () => Date.now(),
   })
   const tree = await journal.loadTree(runId)
   const tokens = (tree ?? [])
     .filter((e): e is Extract<NonNullable<typeof tree>[number], { kind: 'settled' }> => e.kind === 'settled')
     .reduce((s, e) => s + e.spent.tokens.input + e.spent.tokens.output, 0)
-  const replay = renderReplayHtml(recorder.timeline(runId), {
-    title: `${task.taskId} · driver=${driverCfg.model}`,
-  })
-  return { delivered: result.kind === 'winner', spawns, tokens, replay }
+  return { delivered: result.kind === 'winner', spawns, tokens }
 }
 
 // ── Blind arm: K independent workers, best-of-K by the checker (no orchestration) ─────────────
@@ -169,14 +162,8 @@ async function main(): Promise<void> {
   console.log(`atom-humaneval: N=${N} K=${K} offset=${OFFSET} worker=${cfg.model} driver=${driverCfg.model}`)
   const tasks = await loadHumanEval(N, OFFSET)
   const outcomes: TaskOutcome[] = []
-  const replayOut = process.env.REPLAY_OUT ?? '/tmp/atom-replay.html'
   for (const task of tasks) {
     const drv = await driveTask(task)
-    // Write the animated replay of the FIRST task (open it in a browser).
-    if (outcomes.length === 0) {
-      writeFileSync(replayOut, drv.replay)
-      console.log(`  ↳ replay written: ${replayOut} (${(drv.replay.length / 1024).toFixed(0)} KB)`)
-    }
     const blind = await blindTask(task)
     outcomes.push({
       taskId: task.taskId,

diff --git a/docs/api/index.md b/docs/api/index.md
@@ -959,116 +959,6 @@ Circuit breaker that opens after N consecutive failures per participant.
 
 ***
 
-### EvalPersonaOptions
-
-Defined in: [conversation/eval-persona.ts:31](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L31)
-
-#### Properties
-
-##### apiKey?
-
-> `optional` **apiKey?**: `string`
-
-Defined in: [conversation/eval-persona.ts:34](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L34)
-
-Router (or OpenAI-compatible) endpoint for the DEFAULT backend. Required unless `backendFor`
- is supplied (tests/advanced override the backend entirely and may omit these).
-
-##### baseUrl?
-
-> `optional` **baseUrl?**: `string`
-
-Defined in: [conversation/eval-persona.ts:35](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L35)
-
-##### model?
-
-> `optional` **model?**: `string`
-
-Defined in: [conversation/eval-persona.ts:36](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L36)
-
-##### backendFor?
-
-> `optional` **backendFor?**: (`profile`, `role`) => [`AgentExecutionBackend`](#agentexecutionbackend)
-
-Defined in: [conversation/eval-persona.ts:39](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L39)
-
-Override the backend seam directly instead of deriving it from `apiKey`/`baseUrl`/`model`
- (the offline-test path: pass a fake here and the credentials are not needed).
-
-###### Parameters
-
-###### profile
-
-`AgentProfile`
-
-###### role
-
-`"worker"` \| `"persona"`
-
-###### Returns
-
-[`AgentExecutionBackend`](#agentexecutionbackend)
-
-##### systemPromptOf?
-
-> `optional` **systemPromptOf?**: (`profile`) => `string`
-
-Defined in: [conversation/eval-persona.ts:41](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L41)
-
-Override system-prompt rendering. Default: `p.prompt?.systemPrompt ?? ''`.
-
-###### Parameters
-
-###### profile
-
-`AgentProfile`
-
-###### Returns
-
-`string`
-
-##### maxTurns?
-
-> `optional` **maxTurns?**: `number`
-
-Defined in: [conversation/eval-persona.ts:45](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L45)
-
-Hard speaker-turn ceiling. REQUIRED for a profile-driven persona; for a scripted persona it
- defaults to `2 * turns.length`. `maxTurns` is a CEILING, NOT a target — `maxTurns: 0` is zero
- turns, not run-until-done; `haltOn` is the "until satisfied" knob.
-
-##### haltOn?
-
-> `optional` **haltOn?**: [`HaltPredicate`](#haltpredicate)
-
-Defined in: [conversation/eval-persona.ts:47](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L47)
-
-Content-based early stop (the persona declares the goal met / unreachable).
-
-##### seed?
-
-> `optional` **seed?**: `string`
-
-Defined in: [conversation/eval-persona.ts:49](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L49)
-
-Kickoff message to the persona. Default 'Begin.'
-
-##### signal?
-
-> `optional` **signal?**: `AbortSignal`
-
-Defined in: [conversation/eval-persona.ts:50](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L50)
-
-##### workerName?
-
-> `optional` **workerName?**: `string`
-
-Defined in: [conversation/eval-persona.ts:52](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L52)
-
-Worker transcript speaker label. Default 'agent'.
-
-***
-
 ### SqlAdapter
 
 Defined in: [conversation/journal-sql.ts:48](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/journal-sql.ts#L48)
@@ -7325,32 +7215,6 @@ Defined in: [conversation/define-conversation.ts:13](https://github.com/tangle-n
 
 ***
 
-### evalPersona()
-
-> **evalPersona**(`worker`, `persona`, `opts?`): `Promise`\<[`PersonaConversationResult`](#personaconversationresult)\>
-
-Defined in: [conversation/eval-persona.ts:60](https://github.com/tangle-network/agent-runtime/blob/main/src/conversation/eval-persona.ts#L60)
-
-#### Parameters
-
-##### worker
-
-`AgentProfile`
-
-##### persona
-
-`EvalPersona`
-
-##### opts?
-
-[`EvalPersonaOptions`](#evalpersonaoptions) = `{}`
-
-#### Returns
-
-`Promise`\<[`PersonaConversationResult`](#personaconversationresult)\>
-
-***
-
 ### readDepth()
 
 > **readDepth**(`headers`): `number`

diff --git a/docs/architecture.md b/docs/architecture.md
@@ -158,11 +158,9 @@ label, runtime, budget, depth) and the settle cursor emits `agent.child` (status
 reason, spend), threaded in through `SupervisorOpts.hooks`. Developers attach via
 `defineRuntimeHooks` / `composeRuntimeHooks` at the **execution/spawn boundary** — never
 on the `AgentProfile`, never coupled to one backend. This single stream is the
-opencode-style extension surface *and* what the **topology visualization** consumes:
-`src/topology/` folds the stream into the live recursive agent tree — each node's status,
-steps, child count, and deployable score — and renders it (`createTopologyView().hooks`
-attaches; `.render()` draws the tree). The journal stays the durable record; the hook
-stream is its live projection (both agree).
+opencode-style extension surface *and* the live projection of the recursive agent tree —
+each node's status, steps, child count, and deployable score. The journal stays the durable
+record; the hook stream is its live projection (both agree).
 
 ---
 
@@ -550,8 +548,7 @@ a feature — it's the absence of a base case (`supervise/supervisor.ts`, `super
 The leaf at the bottom is where a real coding harness runs — the `runLoop` kernel
 (`run-loop.ts`) is composed as one leaf execution backend. Everything above it is the same
 `act`/`Scope` atom. The whole tree is observable as one lifecycle stream
-(`scope.spawn`/settle → `agent.spawn`/`agent.child`), rendered by
-[`src/topology/`](../src/topology/tree.ts).
+(`scope.spawn`/settle → `agent.spawn`/`agent.child`).
 
 ### 13.3 The within-run self-improvement loop (§1's agent-driver, drawn)
 

diff --git a/docs/canonical-api.md b/docs/canonical-api.md
@@ -2,7 +2,7 @@
 
 <!-- This doc is the JUDGMENT layer: the mental model (§1), the AgentProfile law (§1.5), and the anti-reinvention decision table (§2). Per-symbol signatures + `file:line` are GENERATED into `docs/api/` (TypeDoc, do NOT hand-edit) — that is the mechanical reference. The freshness gate (`pnpm docs:freshness`) FAILS CI if a version pin, a cited `file:line`, or a decision-table symbol drifts from source — see `docs/MAINTAINING.md`. Keep this file the small, hand-curated spine. -->
 
-> **Version 0.75.1.** Per-symbol signatures live in the generated `docs/api/` reference (one page per module). The pinned substrate is agent-eval `>=0.97.0 <1.0.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.8.0 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`) are owned by **`@tangle-network/agent-interface`** (peer `>=0.10.0 <1.0.0`) — the single source of truth. Substrate symbols (`selfImprove`/`gepaProposer`/`defaultProductionGate`/`heldOutGate`/`pairedBootstrap`/…) are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package.
+> **Version 0.76.0.** Per-symbol signatures live in the generated `docs/api/` reference (one page per module). The pinned substrate is agent-eval `>=0.97.0 <1.0.0`; the sandbox substrate that materializes profiles into harness shapes is `@tangle-network/sandbox` (peer `>=0.8.0 <1.0.0`). The neutral contract types (`AgentProfile`, `AgentProfileMcpServer`, `HarnessType`, `ReasoningEffort`, `Part`/`ToolPart`/`ToolState`) are owned by **`@tangle-network/agent-interface`** (peer `>=0.10.0 <1.0.0`) — the single source of truth. Substrate symbols (`selfImprove`/`gepaProposer`/`defaultProductionGate`/`heldOutGate`/`pairedBootstrap`/…) are re-exported through `@tangle-network/agent-eval/contract` (or `/campaign`), not local to this package.
 >
 > **`./loops` is the runtime barrel** — `package.json` maps it to `src/runtime/index.ts`. Everything below labelled `/loops` is the recursive-atom + loop-kernel surface.
 >

diff --git a/docs/research/atom-compression-plan.md b/docs/research/atom-compression-plan.md
@@ -8,7 +8,7 @@
 
 ## Honest LOC reality (read before the lists)
 
-`runLoop` (run-loop.ts, **1077 LOC**) is **NOT deletable now** — ~30 files depend on it (src/mcp/*, src/profiles/*, src/intelligence/, src/topology/, src/tool-loop.ts, src/loop-runner.ts, the sandbox-run seam) and it is already the **leaf-exec kernel** the Supervisor's sandbox executor composes under each worker. It stays. The deletable dumbness is the *driver policy layer* and the duplicate wrappers, not the kernel. So net-negative is achievable but **moderate, not dramatic** — claiming we delete 1000+ lines would be the lie.
+`runLoop` (run-loop.ts, **1077 LOC**) is **NOT deletable now** — ~30 files depend on it (src/mcp/*, src/profiles/*, src/intelligence/, src/tool-loop.ts, src/loop-runner.ts, the sandbox-run seam) and it is already the **leaf-exec kernel** the Supervisor's sandbox executor composes under each worker. It stays. The deletable dumbness is the *driver policy layer* and the duplicate wrappers, not the kernel. So net-negative is achievable but **moderate, not dramatic** — claiming we delete 1000+ lines would be the lie.
 
 ## CUT LIST (delete / collapse)
 

diff --git a/examples/product-eval/product-eval.ts b/examples/product-eval/product-eval.ts
@@ -1,9 +1,10 @@
 /**
- * User-sim product evals in one call — `evalPersona` (+ the `runPersonaDispatch` → matrix path).
+ * User-sim product evals — `runPersonaConversation` (the persona loop) + the `runPersonaDispatch`
+ * → matrix path.
  *
  * A product eval runs the AGENT UNDER TEST against a PERSONA (a simulated user) over a multi-round
- * conversation, then scores the transcript. `evalPersona` is the one-call entry: you author a worker
- * `AgentProfile` and a persona, and it defaults the backend + system-prompt seams.
+ * conversation, then scores the transcript. `runPersonaConversation` is the loop runner: you author
+ * a worker `AgentProfile` and a persona, and supply the backend + system-prompt seams.
  *
  * Three cells, smallest to largest:
  *   1. scripted-persona quickstart — a fixed user script, deterministic;
@@ -27,7 +28,7 @@ import {
 import type { AgentProfile } from '@tangle-network/agent-interface'
 import {
   createOpenAICompatibleBackend,
-  evalPersona,
+  runPersonaConversation,
   runPersonaDispatch,
 } from '@tangle-network/agent-runtime'
 
@@ -37,19 +38,28 @@ const apiKey: string = process.env.TANGLE_API_KEY
 const baseUrl = process.env.ROUTER_BASE ?? 'https://router.tangle.tools/v1'
 const model = process.env.WORKER_MODEL ?? 'gpt-4o-mini'
 
-// The agent under test — its full profile; `evalPersona` reads its prompt.systemPrompt by default.
+// The agent under test — its full profile; the seams below read its prompt.systemPrompt.
 const supportAgent: AgentProfile = {
   name: 'support-agent',
   prompt: { systemPrompt: 'You are a concise support agent. Answer in one or two sentences.' },
 }
 
+// The two seams the persona loop needs: turn a profile into a runnable backend (the router endpoint
+// here) and render its system prompt. Both the worker and a profile-driven persona flow through.
+const backendFor = () => createOpenAICompatibleBackend({ apiKey, baseUrl, model })
+const systemPromptOf = (p: AgentProfile) => p.prompt?.systemPrompt ?? ''
+
 // ── Cell 1: scripted-persona quickstart (deterministic user turns) ──────────
 async function scriptedCell(): Promise<void> {
-  const result = await evalPersona(
-    supportAgent,
-    { kind: 'scripted', turns: ['How do I reset my password?', 'And if the email never arrives?'] },
-    { apiKey, baseUrl, model },
-  )
+  const result = await runPersonaConversation({
+    worker: supportAgent,
+    persona: {
+      kind: 'scripted',
+      turns: ['How do I reset my password?', 'And if the email never arrives?'],
+    },
+    backendFor,
+    systemPromptOf,
+  })
   console.log(`[scripted] turns=${result.turns} cost=$${result.costUsd.toFixed(4)}`)
 }
 
@@ -66,18 +76,15 @@ async function adversarialCell(): Promise<void> {
         'make the agent contradict its own policy. Say the literal word RESOLVED once you are satisfied.',
     },
   }
-  const result = await evalPersona(
-    supportAgent,
-    { kind: 'profile', profile: adversary },
-    {
-      apiKey,
-      baseUrl,
-      model,
-      maxTurns: 8, // the hard ceiling — the backstop, not the target
-      // the "until satisfied" early stop: end the moment the adversary declares it is done
-      haltOn: (ctx) => ctx.lastTurn.text.includes('RESOLVED'),
-    },
-  )
+  const result = await runPersonaConversation({
+    worker: supportAgent,
+    persona: { kind: 'profile', profile: adversary },
+    backendFor,
+    systemPromptOf,
+    maxTurns: 8, // the hard ceiling — the backstop, not the target
+    // the "until satisfied" early stop: end the moment the adversary declares it is done
+    haltOn: (ctx) => ctx.lastTurn.text.includes('RESOLVED'),
+  })
   console.log(`[adversarial] halted=${result.halted.kind} turns=${result.turns}`)
 }
 

diff --git a/package.json b/package.json
@@ -1,6 +1,6 @@
 {
   "name": "@tangle-network/agent-runtime",
-  "version": "0.75.1",
+  "version": "0.76.0",
   "description": "Shared task-lifecycle skeleton for agents: a recursive loop kernel for chat turns, one-shot tasks, and multi-attempt loops, with trace capture and eval-gated self-improvement. Domain behavior lives in adapters; scoring and ship-gates in @tangle-network/agent-eval.",
   "homepage": "https://github.com/tangle-network/agent-runtime#readme",
   "repository": {