@forwardimpact/libeval 0.1.44 → 0.1.45

package/src/supervisor.js

@@ -1,478 +1,99 @@
 /**
- * Supervisor — orchestrates a relay loop between an agent and a supervisor,
- * both running as AgentRunner instances. The supervisor receives the task first,
- * introduces itself, and delegates work to the agent. The loop then alternates:
- * agent → supervisor → agent.
+ * Supervisor — supervise-mode wrapper around `OrchestrationLoop`. One
+ * named participant (`"agent"`) coordinated by a lead participant
+ * (`"supervisor"`). Structurally the same as `Facilitator` with a
+ * single agent; differs only in role names, prompts, and pass-through
+ * accessors.
  *
- * Signaling uses orchestration tools (Ask / Announce / Redirect / Conclude)
- * via in-process MCP servers; the supervisor has no Answer tool — agent replies
- * are routed back through the relay loop. The Ask/Answer contract is enforced
- * at turn boundaries: an unanswered Ask triggers one synthetic reminder and
- * then a `protocol_violation` trace event plus a null-answer injection so the
- * session advances without silent deadlock.
+ * Ask is async (same contract as facilitate / discuss): returns
+ * `{askIds:[N]}` immediately; the agent's reply arrives on the
+ * supervisor's next turn as `[answer#N] agent: <text>`. The supervisor
+ * sees the agent at each Ask boundary, plans the next step, and
+ * eventually calls Conclude.
  *
- * Follows OO+DI: constructor injection, factory function, tests bypass factory.
+ * For tighter feedback loops, size the agent's per-turn budget down
+ * (smaller `maxTurns` on the agent runner) so each Ask returns sooner.
  */
 
 import { Writable } from "node:stream";
 import { resolve } from "node:path";
 import { createAgentRunner } from "./agent-runner.js";
 import { composeProfilePrompt } from "./profile-prompt.js";
-import { TraceCollector } from "./trace-collector.js";
-import { SequenceCounter } from "./sequence-counter.js";
 import { createMessageBus } from "./message-bus.js";
 import {
   createOrchestrationContext,
-  createSupervisorToolServer,
   createSupervisedAgentToolServer,
-  checkPendingAsk,
+  createSupervisorToolServer,
 } from "./orchestration-toolkit.js";
-import { formatMessages } from "./orchestrator-helpers.js";
+import { OrchestrationLoop } from "./orchestration-loop.js";
 
 /** System prompt appended for the supervisor runner in supervise mode. */
 export const SUPERVISOR_SYSTEM_PROMPT =
-  "You supervise one agent. " +
-  "Ask sends a question to the agent; the reply arrives via Answer. " +
-  "Answer replies to an ask the agent addressed to you. " +
-  "Announce sends a message with no reply obligation. " +
-  "Redirect interrupts the agent with replacement instructions. " +
-  "Conclude ends the session with a verdict ('success' or 'failure') and a summary; " +
-  "the verdict reflects whether the agent's work meets the criteria stated in the task.";
+  "You supervise one agent named `agent`. " +
+  "Ask sends a question and returns immediately with {askIds:[N]}. The reply arrives on a later turn as `[answer#N] agent: <text>` in your inbox — between turns you can plan and reflect while the agent works. End your turn with text after asking; the orchestrator wakes you when the agent replies. " +
+  "Answer replies to an ask the agent addressed to you (you'll see it tagged `[ask#N] agent: …` in your inbox). Quote askId from the [ask#N] tag; omit it and the handler auto-picks the only pending ask or routes your message as an Announce. " +
+  "Announce delivers a message with no reply obligation. " +
+  "Conclude ends the session with a verdict ('success' or 'failure') and a summary; the verdict reflects whether the agent's work meets the criteria stated in the task. " +
+  "You MUST end every session with Conclude — never end a turn with only text *after* every Ask round has resolved. " +
+  "If the agent goes off-track, course-correct by issuing a new Ask with corrected instructions; each Ask carries a fresh askId, so a follow-up never collides with an earlier one.";
 
 /** System prompt appended for the agent runner in supervise mode. */
 export const AGENT_SYSTEM_PROMPT =
   "A supervisor watches your work. " +
-  "Answer replies to an ask addressed to you. " +
-  "Ask sends a question to the supervisor; the reply arrives via Answer. " +
-  "Announce sends a message with no reply expected.";
+  "Each question you receive carries an [ask#N] header — quote that N back as the askId field on Answer so the reply pairs with the right question. " +
+  "Answer replies to an ask addressed to you. askId is optional: omit it and the handler auto-picks if exactly one ask is owed to you, otherwise it routes your message as an Announce. " +
+  "Ask sends a question to the supervisor and returns immediately with {askIds:[N]}; the reply arrives on a later turn as `[answer#N] supervisor: <text>` in your inbox. " +
+  "Announce sends a message with no reply expected — use this for unsolicited remarks or to reply to an Announce.";
 
 /**
- * Maximum number of mid-turn interventions allowed within a single agent turn.
- * Bounded so a looping supervisor exhausts its quota fast (observability) but
- * leaves headroom for legitimate "intervene, observe, intervene again" patterns.
- * The outer exchange budget still bounds overall runtime.
+ * Supervise-mode wrapper around `OrchestrationLoop`. The lead is
+ * `"supervisor"`, one participant is `"agent"`, mode tag is `"supervised"`.
  */
-const MAX_INTERVENTIONS_PER_TURN = 5;
-
-/**
- * Default cap on supervisor↔agent exchanges in a single run. Not exposed via
- * CLI — `--max-turns` governs the per-runner invocation budget instead. When
- * a `--max-exchanges` flag is added this becomes the default for that flag.
- */
-const DEFAULT_MAX_EXCHANGES = 100;
-
-/** Orchestrate a relay loop between a supervisor LLM and an agent LLM with mid-turn review. */
-export class Supervisor {
+export class Supervisor extends OrchestrationLoop {
   /**
    * @param {object} deps
-   * @param {import("./agent-runner.js").AgentRunner} deps.agentRunner - Runs the agent sessions
-   * @param {import("./agent-runner.js").AgentRunner} deps.supervisorRunner - Runs the supervisor sessions
-   * @param {import("stream").Writable} deps.output - Stream to emit tagged NDJSON to
-   * @param {number} [deps.maxTurns] - Maximum supervisor ↔ agent exchanges
-   * @param {object} [deps.ctx] - Orchestration context (injected by factory)
-   * @param {import("./message-bus.js").MessageBus} [deps.messageBus] - Two-participant message bus ("supervisor" / "agent")
-   * @param {string} [deps.taskAmend] - Opaque addendum appended to the task before delivery.
+   * @param {import("./agent-runner.js").AgentRunner} deps.supervisorRunner
+   * @param {import("./agent-runner.js").AgentRunner} deps.agentRunner
+   * @param {import("./message-bus.js").MessageBus} deps.messageBus
+   * @param {import("stream").Writable} deps.output
+   * @param {object} deps.ctx
+   * @param {object} deps.redactor
+   * @param {string} [deps.taskAmend]
    */
   constructor({
-    agentRunner,
     supervisorRunner,
+    agentRunner,
+    messageBus,
     output,
-    maxTurns,
     ctx,
-    messageBus,
     taskAmend,
     redactor,
   }) {
     if (!agentRunner) throw new Error("agentRunner is required");
     if (!supervisorRunner) throw new Error("supervisorRunner is required");
     if (!output) throw new Error("output is required");
-    if (!redactor) throw new Error("redactor is required");
-    this.redactor = redactor;
-    this.agentRunner = agentRunner;
-    this.supervisorRunner = supervisorRunner;
-    this.output = output;
-    this.maxTurns = maxTurns ?? 100;
-    this.ctx = ctx ?? createOrchestrationContext();
-    this.messageBus =
-      messageBus ?? createMessageBus({ participants: ["supervisor", "agent"] });
-    if (!this.ctx.messageBus) this.ctx.messageBus = this.messageBus;
-    this.counter = new SequenceCounter();
-    this.taskAmend = taskAmend ?? null;
-    /** @type {"agent"|"supervisor"} */
-    this.currentSource = "agent";
-    /** @type {number} */
-    this.currentTurn = 0;
-  }
-
-  /**
-   * Run the supervisor ↔ agent relay loop.
-   * @param {string} task - The initial task for the supervisor
-   * @returns {Promise<{success: boolean, turns: number, concluded: boolean}>}
-   */
-  async run(task) {
-    const initialTask = this.taskAmend ? `${task}\n\n${this.taskAmend}` : task;
-    this.currentSource = "supervisor";
-    this.currentTurn = 0;
-    let supervisorResult = await this.supervisorRunner.run(initialTask);
-
-    if (supervisorResult.error) {
-      this.emitSummary({ success: false, turns: 0 });
-      return { success: false, turns: 0, concluded: false };
-    }
-
-    if (this.ctx.concluded) {
-      const success = this.ctx.verdict === "success";
-      this.emitSummary({
-        success,
-        verdict: this.ctx.verdict,
-        turns: 0,
-        summary: this.ctx.summary,
-      });
-      return { success, turns: 0, concluded: true };
-    }
-
-    let pendingRelay = null;
-    const turnLimit = this.maxTurns === 0 ? Infinity : this.maxTurns;
-    for (let turn = 1; turn <= turnLimit; turn++) {
-      const relay =
-        pendingRelay ?? this.#buildInitialRelay(supervisorResult.text);
-
-      const turnOutcome = await this.#runAgentTurn(turn, relay);
-      if (turnOutcome.exit) {
-        return { ...turnOutcome.exit, concluded: this.ctx.concluded };
-      }
-
-      const reviewOutcome = await this.#endOfTurnReview(turn);
-      if (reviewOutcome.exit) {
-        return { ...reviewOutcome.exit, concluded: this.ctx.concluded };
-      }
-      supervisorResult = reviewOutcome.supervisorResult;
-      pendingRelay = reviewOutcome.relay ?? null;
-    }
-
-    this.emitSummary({ success: false, turns: this.maxTurns });
-    return { success: false, turns: this.maxTurns, concluded: false };
-  }
-
-  #buildInitialRelay(fallbackText) {
-    const queued = this.messageBus.drain("agent");
-    if (queued.length > 0) return formatMessages(queued);
-    return this.extractLastText(this.supervisorRunner, fallbackText);
-  }
-
-  #checkAsk(name) {
-    return checkPendingAsk({
-      ctx: this.ctx,
-      messageBus: this.messageBus,
-      addresseeName: name,
+    super({
+      leadRunner: supervisorRunner,
+      agents: [{ name: "agent", role: "agent", runner: agentRunner }],
+      messageBus,
+      output,
+      leadName: "supervisor",
       mode: "supervised",
-      emitViolation: (e) => this.emitOrchestratorEvent(e),
+      ctx,
+      taskAmend,
+      redactor,
     });
   }
 
-  /**
-   * Drive the agent through one turn, allowing the supervisor to interrupt
-   * via the Redirect tool. Returns either an `exit` outcome (the loop should
-   * return immediately) or `{exit: null}` (proceed to end-of-turn review).
-   * @param {number} turn
-   * @param {string} initialRelay
-   * @returns {Promise<{exit: {success: boolean, turns: number}|null}>}
-   */
-  async #runAgentTurn(turn, initialRelay) {
-    let relay = initialRelay;
-    let interventions = 0;
-    let agentCalled = this.agentRunner.sessionId !== null;
-
-    this.agentRunner.onBatch = (batchLines, ctx) =>
-      this.#midTurnReview(turn, batchLines, ctx);
-
-    try {
-      while (true) {
-        this.currentSource = "agent";
-        this.currentTurn = turn;
-        const agentResult = agentCalled
-          ? await this.agentRunner.resume(relay)
-          : await this.agentRunner.run(relay);
-        agentCalled = true;
-
-        const outcome = this.#classifyAgentOutcome(
-          agentResult,
-          turn,
-          interventions,
-        );
-
-        if (outcome.type === "exit") return { exit: outcome.exit };
-        if (outcome.type === "intervention_limit") return { exit: null };
-
-        if (outcome.type === "redirect") {
-          interventions++;
-          relay = outcome.relay;
-          this.emitOrchestratorEvent({ type: "intervention_relayed", turn });
-          continue;
-        }
-
-        const askRelay = this.#drainAgentAskRelay();
-        if (askRelay) {
-          relay = askRelay;
-          continue;
-        }
-
-        return { exit: null };
-      }
-    } finally {
-      this.agentRunner.onBatch = null;
-    }
-  }
-
-  /**
-   * Classify the outcome of a single agent execution within #runAgentTurn.
-   * @returns {{type: string, exit?: object|null, relay?: string}}
-   */
-  #classifyAgentOutcome(agentResult, turn, interventions) {
-    if (agentResult.error && !agentResult.aborted) {
-      this.emitSummary({ success: false, turns: turn });
-      return { type: "exit", exit: { success: false, turns: turn } };
-    }
-
-    if (this.ctx.concluded) {
-      const success = this.ctx.verdict === "success";
-      this.emitSummary({
-        success,
-        verdict: this.ctx.verdict,
-        turns: turn,
-        summary: this.ctx.summary,
-      });
-      return { type: "exit", exit: { success, turns: turn } };
-    }
-
-    if (agentResult.aborted && this.ctx.redirect) {
-      const redirect = this.ctx.redirect;
-      this.ctx.redirect = null;
-      if (interventions + 1 >= MAX_INTERVENTIONS_PER_TURN) {
-        this.emitOrchestratorEvent({ type: "intervention_limit", turn });
-        return { type: "intervention_limit" };
-      }
-      return { type: "redirect", relay: redirect.message };
-    }
-
-    return { type: "continue" };
-  }
-
-  /**
-   * If the agent has an unanswered ask, drain reminders and return a
-   * formatted relay string. Returns null when no relay is needed.
-   * @returns {string|null}
-   */
-  #drainAgentAskRelay() {
-    if (this.#checkAsk("agent") !== "recheck" || this.ctx.concluded)
-      return null;
-    const reminders = this.messageBus.drain("agent");
-    return reminders.length > 0 ? formatMessages(reminders) : null;
-  }
-
-  /**
-   * Mid-turn supervisor review fired from inside the agent's onBatch hook.
-   * Runs the supervisor's LLM against the batch and aborts the agent if
-   * the supervisor calls Redirect or Conclude.
-   * @param {number} turn
-   * @param {string[]} batchLines
-   * @param {{abort: () => void}} ctx
-   */
-  async #midTurnReview(turn, batchLines, { abort }) {
-    const batchTranscript = this.renderBatch(batchLines);
-    this.emitOrchestratorEvent({ type: "mid_turn_review", turn });
-
-    this.currentSource = "supervisor";
-    this.ctx.redirect = null;
-
-    await this.supervisorRunner.resume(
-      `The agent is mid-turn. Latest batch:\n\n${batchTranscript}\n\n` +
-        `Review and use your tools if action is needed.`,
-    );
-    this.currentSource = "agent";
-
-    if (this.ctx.redirect) {
-      this.emitOrchestratorEvent({ type: "intervention_requested", turn });
-      abort();
-      return;
-    }
-    if (this.ctx.concluded) {
-      this.emitOrchestratorEvent({ type: "complete_requested", turn });
-      abort();
-    }
-  }
-
-  /**
-   * End-of-turn supervisor review. Returns either an exit outcome (error or
-   * completion) or the supervisor result so the outer loop can build the
-   * next turn's relay.
-   * @param {number} turn
-   * @returns {Promise<{exit: {success: boolean, turns: number}|null, supervisorResult?: object, relay?: string}>}
-   */
-  async #endOfTurnReview(turn) {
-    const queuedForSupervisor = this.messageBus.drain("supervisor");
-    const agentTranscript = this.extractTranscript(this.agentRunner);
-    this.currentSource = "supervisor";
-    this.currentTurn = turn;
-    this.ctx.redirect = null;
-
-    const reviewPrompt =
-      queuedForSupervisor.length > 0
-        ? `The agent reported:\n\n${agentTranscript}\n\n` +
-          `Agent messages:\n${formatMessages(queuedForSupervisor)}\n\n` +
-          `Review and decide how to proceed.`
-        : `The agent reported:\n\n${agentTranscript}\n\nReview the agent's work and decide how to proceed.`;
-
-    let supervisorResult = await this.supervisorRunner.resume(reviewPrompt);
-
-    if (supervisorResult.error) {
-      this.emitSummary({ success: false, turns: turn });
-      return { exit: { success: false, turns: turn } };
-    }
-
-    if (this.ctx.concluded) {
-      const success = this.ctx.verdict === "success";
-      this.emitSummary({
-        success,
-        verdict: this.ctx.verdict,
-        turns: turn,
-        summary: this.ctx.summary,
-      });
-      return { exit: { success, turns: turn } };
-    }
-
-    if (this.#checkAsk("supervisor") === "recheck" && !this.ctx.concluded) {
-      const reminders = this.messageBus.drain("supervisor");
-      if (reminders.length > 0) {
-        supervisorResult = await this.supervisorRunner.resume(
-          formatMessages(reminders),
-        );
-        if (this.ctx.concluded) {
-          const success = this.ctx.verdict === "success";
-          this.emitSummary({
-            success,
-            verdict: this.ctx.verdict,
-            turns: turn,
-            summary: this.ctx.summary,
-          });
-          return { exit: { success, turns: turn } };
-        }
-        this.#checkAsk("supervisor");
-      }
-    }
-
-    if (this.ctx.redirect) {
-      const redirect = this.ctx.redirect;
-      this.ctx.redirect = null;
-      return { exit: null, supervisorResult, relay: redirect.message };
-    }
-
-    const queuedForAgent = this.messageBus.drain("agent");
-    const relay =
-      queuedForAgent.length > 0 ? formatMessages(queuedForAgent) : undefined;
-    return { exit: null, supervisorResult, relay };
+  /** Readability shims for tests that read the runners by their domain names. */
+  /** Readability shim — exposes the lead runner under its mode-specific name. */
+  get supervisorRunner() {
+    return this.leadRunner;
   }
 
-  /**
-   * Extract a human-readable transcript from an AgentRunner's buffered output.
-   * @param {import("./agent-runner.js").AgentRunner} runner
-   * @returns {string}
-   */
-  extractTranscript(runner) {
-    const lines = runner.drainOutput();
-    const collector = new TraceCollector();
-    for (const line of lines) {
-      collector.addLine(line);
-    }
-    return collector.toText() || "[The agent produced no output.]";
-  }
-
-  /**
-   * Extract only the last assistant text block from an AgentRunner's buffer.
-   * @param {import("./agent-runner.js").AgentRunner} runner
-   * @param {string} fallback
-   * @returns {string}
-   */
-  extractLastText(runner, fallback) {
-    const lines = runner.buffer;
-    for (let i = lines.length - 1; i >= 0; i--) {
-      const event = JSON.parse(lines[i]);
-      if (event.type !== "assistant") continue;
-      const content = event.message?.content ?? event.content;
-      if (!Array.isArray(content)) continue;
-      for (let j = content.length - 1; j >= 0; j--) {
-        if (content[j].type === "text" && content[j].text) {
-          return content[j].text;
-        }
-      }
-    }
-    return fallback;
-  }
-
-  /**
-   * Emit a single NDJSON line tagged with the current source and seq.
-   * @param {string} line - Raw NDJSON line from the runner
-   */
-  emitLine(line) {
-    const event = JSON.parse(line);
-    const tagged = {
-      source: this.currentSource,
-      seq: this.counter.next(),
-      event,
-    };
-    this.output.write(JSON.stringify(this.redactor.redactValue(tagged)) + "\n");
-  }
-
-  /**
-   * Render a batch of buffered NDJSON lines as human-readable text.
-   * @param {string[]} batchLines
-   * @returns {string}
-   */
-  renderBatch(batchLines) {
-    if (batchLines.length === 0) return "[empty]";
-    const collector = new TraceCollector();
-    for (const line of batchLines) {
-      collector.addLine(line);
-    }
-    return collector.toText() || "[empty]";
-  }
-
-  /**
-   * Emit an orchestrator-source NDJSON line.
-   * @param {{type: string, turn?: number}} event
-   */
-  emitOrchestratorEvent(event) {
-    this.output.write(
-      JSON.stringify(
-        this.redactor.redactValue({
-          source: "orchestrator",
-          seq: this.counter.next(),
-          event,
-        }),
-      ) + "\n",
-    );
-  }
-
-  /**
-   * Emit a final orchestrator summary line, wrapped in the universal envelope.
-   * @param {{success: boolean, verdict?: string|null, turns: number, summary?: string}} result
-   */
-  emitSummary(result) {
-    this.output.write(
-      JSON.stringify(
-        this.redactor.redactValue({
-          source: "orchestrator",
-          seq: this.counter.next(),
-          event: {
-            type: "summary",
-            success: result.success,
-            ...(result.verdict && { verdict: result.verdict }),
-            turns: result.turns,
-            ...(result.summary && { summary: result.summary }),
-          },
-        }),
-      ) + "\n",
-    );
+  /** Readability shim — exposes the single agent runner directly. */
+  get agentRunner() {
+    return this.agents[0].runner;
   }
 }
 
@@ -483,24 +104,26 @@ const devNull = new Writable({
 });
 
 /**
- * Factory function — wires both AgentRunners with their respective configs.
+ * Factory — wires the supervisor + agent runners and the orchestration
+ * context. Mirrors the facilitator factory in shape.
+ *
  * @param {object} deps
  * @param {string} deps.supervisorCwd
  * @param {string} deps.agentCwd
  * @param {function} deps.query
  * @param {import("stream").Writable} deps.output
- * @param {string} [deps.model] - Default model for both runners.
- * @param {string} [deps.agentModel] - Agent model override (falls back to `model`).
- * @param {string} [deps.supervisorModel] - Supervisor model override (falls back to `model`).
- * @param {number} [deps.maxTurns] - Per-runner invocation budget for both the supervisor and the agent (default 200; 0 = unlimited). Outer supervisor↔agent exchanges are bounded separately by `DEFAULT_MAX_EXCHANGES` (passes through to unlimited when `maxTurns === 0`).
+ * @param {string} [deps.model]
+ * @param {string} [deps.agentModel]
+ * @param {string} [deps.supervisorModel]
+ * @param {number} [deps.maxTurns] - Per-runner SDK turn budget (default 200).
  * @param {string[]} [deps.allowedTools]
  * @param {string[]} [deps.supervisorAllowedTools]
  * @param {string[]} [deps.supervisorDisallowedTools]
- * @param {string} [deps.supervisorProfile] - Supervisor profile name; resolved into the main-thread system prompt via `composeProfilePrompt`.
- * @param {string} [deps.agentProfile] - Agent profile name; resolved into the main-thread system prompt via `composeProfilePrompt`.
- * @param {string} [deps.profilesDir] - Directory containing `<name>.md` profile files. Defaults to `<supervisorCwd>/.claude/agents`. Resolved once from the orchestrator's cwd so profiles travel with the project, not with a per-agent sandbox.
- * @param {string} [deps.taskAmend] - Opaque addendum appended to the task before delivery.
- * @param {Record<string, object>} [deps.agentMcpServers] - Additional MCP servers exposed to the agent (merged alongside the orchestration server).
+ * @param {string} [deps.supervisorProfile]
+ * @param {string} [deps.agentProfile]
+ * @param {string} [deps.profilesDir]
+ * @param {string} [deps.taskAmend]
+ * @param {Record<string, object>} [deps.agentMcpServers]
  * @returns {Supervisor}
  */
 export function createSupervisor({
@@ -513,8 +136,8 @@ export function createSupervisor({
   supervisorModel,
   maxTurns,
   allowedTools,
-  supervisorDisallowedTools,
   supervisorAllowedTools,
+  supervisorDisallowedTools,
   supervisorProfile,
   agentProfile,
   profilesDir,
@@ -534,7 +157,6 @@ export function createSupervisor({
         })
       : { type: "preset", preset: "claude_code", append: trailer };
   };
-  let supervisor;
 
   const ctx = createOrchestrationContext();
   const messageBus = createMessageBus({
@@ -546,36 +168,29 @@ export function createSupervisor({
     { name: "agent", role: "agent" },
   ];
 
-  const supervisorServer = createSupervisorToolServer(ctx);
-  const agentServer = createSupervisedAgentToolServer(ctx);
-
-  const onLine = (line) => supervisor.emitLine(line);
+  let supervisor;
+  const perRunBudget = maxTurns ?? 200;
 
-  // `maxTurns` is the per-runner invocation budget — matches `run` and
-  // `facilitate` semantics. The outer supervisor↔agent exchange loop is
-  // bounded separately by `DEFAULT_MAX_EXCHANGES`; when --max-exchanges is
-  // added it will become a parameter. `maxTurns === 0` propagates through
-  // to mean unlimited on both axes.
-  const perInvocationTurns = maxTurns ?? 200;
-  const exchangeBudget = maxTurns === 0 ? 0 : DEFAULT_MAX_EXCHANGES;
+  const agentServer = createSupervisedAgentToolServer(ctx);
+  const supervisorServer = createSupervisorToolServer(ctx);
 
   const agentRunner = createAgentRunner({
     cwd: agentCwd,
     query,
     output: devNull,
     model: agentModel ?? model,
-    maxTurns: perInvocationTurns,
+    maxTurns: perRunBudget,
     allowedTools,
-    onLine,
+    onLine: (line) => supervisor.emitLine("agent", line),
     settingSources: ["project"],
     systemPrompt: systemPromptFor(agentProfile, AGENT_SYSTEM_PROMPT),
     mcpServers: { orchestration: agentServer, ...agentMcpServers },
     redactor,
   });
 
-  // Block the SDK's sub-agent spawn tools on the supervisor: its job is to
-  // coordinate the agent through the libeval orchestration harness, not to
-  // fan work out to ad-hoc Claude Code sub-agents. Mirrors the facilitator.
+  // Block the SDK's sub-agent spawn tools on the supervisor: it should
+  // coordinate the agent through orchestration tools, not fan work out
+  // to ad-hoc Claude Code sub-agents.
   const defaultDisallowed = ["Agent", "Task", "TaskOutput", "TaskStop"];
   const disallowedTools = supervisorDisallowedTools
     ? [...new Set([...defaultDisallowed, ...supervisorDisallowedTools])]
@@ -586,7 +201,7 @@ export function createSupervisor({
     query,
     output: devNull,
     model: supervisorModel ?? model,
-    maxTurns: perInvocationTurns,
+    maxTurns: perRunBudget,
     allowedTools: supervisorAllowedTools ?? [
       "Bash",
       "Read",
@@ -596,7 +211,7 @@ export function createSupervisor({
       "Edit",
     ],
     disallowedTools,
-    onLine,
+    onLine: (line) => supervisor.emitLine("supervisor", line),
     settingSources: ["project"],
     systemPrompt: systemPromptFor(supervisorProfile, SUPERVISOR_SYSTEM_PROMPT),
     mcpServers: { orchestration: supervisorServer },
@@ -604,12 +219,11 @@ export function createSupervisor({
   });
 
   supervisor = new Supervisor({
-    agentRunner,
     supervisorRunner,
+    agentRunner,
+    messageBus,
     output,
-    maxTurns: exchangeBudget,
     ctx,
-    messageBus,
     taskAmend,
     redactor,
   });