@forwardimpact/libeval 0.1.47 → 0.1.48

package/README.md

@@ -69,6 +69,20 @@ Inbox lines on resume:
 Async means the lead can issue Asks, end its turn, and plan in the gap
 while participants work in parallel — nothing blocks the LLM thread.
 
+### Discuss-mode replies
+
+In discussion mode, Answer calls routed to the lead are captured as
+thread replies delivered via the bridge callback. The lead delegates work
+via Ask; each agent's Answer becomes a separate reply posted to the
+discussion thread. No explicit reply tool is needed on the lead surface —
+the message bus intercepts answers and appends them to `ctx.replies[]`.
+
+`RequestForComment` is a separate coordination tool available on agent
+roles (facilitated agents and discuss agents). It queues an intent to
+open a new Discussion thread for long-horizon coordination on open
+questions; these are accumulated in `ctx.rfcs[]`, separate from the
+thread replies in `ctx.replies[]`.
+
 ## Orchestration loop
 
 Each participant drains the bus (or waits), runs/resumes the LLM with
@@ -84,15 +98,15 @@ only feeds the summary's `success`/`verdict`.
 
 ## Tool surface, by role
 
-| Role         | Ask | Answer | Announce | RollCall | Conclude | Other                                    |
-| ------------ | --- | ------ | -------- | -------- | -------- | ---------------------------------------- |
-| Facilitator  | ✓   | ✓      | ✓        | ✓        | ✓        |                                          |
-| Fac. agent   | ✓   | ✓      | ✓        | ✓        |          |                                          |
-| Supervisor   | ✓   | ✓      | ✓        | ✓        | ✓        |                                          |
-| Sup. agent   | ✓   | ✓      | ✓        | ✓        |          |                                          |
-| Discuss lead | ✓   | ✓      | ✓        | ✓        |          | `RequestForComment`, `Recess`, `Adjourn` |
-| Discuss agt  | ✓   | ✓      | ✓        | ✓        |          |                                          |
-| Judge        |     |        |          |          | ✓        |                                          |
+| Role         | Ask | Answer | Announce | RollCall | Conclude | Other                          |
+| ------------ | --- | ------ | -------- | -------- | -------- | ------------------------------ |
+| Facilitator  | ✓   | ✓      | ✓        | ✓        | ✓        |                                |
+| Fac. agent   | ✓   | ✓      | ✓        | ✓        |          | `RequestForComment`            |
+| Supervisor   | ✓   | ✓      | ✓        | ✓        | ✓        |                                |
+| Sup. agent   | ✓   | ✓      | ✓        | ✓        |          |                                |
+| Discuss lead | ✓   | ✓      | ✓        | ✓        |          | `Recess`, `Adjourn`            |
+| Discuss agt  | ✓   | ✓      | ✓        | ✓        |          | `RequestForComment`            |
+| Judge        |     |        |          |          | ✓        |                                |
 
 Ask's `to` accepts a participant name on multi-participant roles
 (facilitator, discuss lead, all participants). The supervise pair has
@@ -152,10 +166,10 @@ downloadable through retention.
 | ----------------------------------------------------------- | -------------------------------------------------------------------- |
 | `agent-runner.js`                                           | One Claude Agent SDK session; emits NDJSON via the redactor.         |
 | `message-bus.js`                                            | Per-participant queues + `waitForMessages` Promise wakeup.           |
-| `orchestration-toolkit.js`                                  | Shared Ask/Answer/Announce/Conclude/RollCall handlers + builders.    |
+| `orchestration-toolkit.js`                                  | Shared Ask/Answer/Announce/Conclude/RollCall/RequestForComment handlers + builders. |
 | `orchestration-loop.js`                                     | Unified lead+participant loop; reminder/violation handling.          |
 | `facilitator.js` / `supervisor.js` / `discusser.js` / `judge.js` | Per-mode class + factory + system prompt.                       |
-| `discuss-tools.js`                                          | Discuss-only `RequestForComment`/`Recess`/`Adjourn`.                 |
+| `discuss-tools.js`                                          | Discuss-only `Recess`/`Adjourn`.                                     |
 | `trace-collector.js` / `trace-query.js` / `trace-github.js` | Trace ingestion / querying / GitHub-attachment helpers.              |
 | `redaction.js`                                              | Env-var allowlist + credential-shape pattern redaction.              |
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@forwardimpact/libeval",
-  "version": "0.1.47",
+  "version": "0.1.48",
   "description": "Agent evaluation framework — prove whether agent changes improved outcomes with reproducible evidence.",
   "keywords": [
     "eval",

package/src/commands/discuss.js

@@ -52,7 +52,7 @@ export function parseDiscussOptions(values) {
     taskContent,
     taskAmend,
     agentConfigs,
-    leadProfile: values["lead-profile"] ?? "release-engineer",
+    leadProfile: values["lead-profile"] ?? undefined,
     leadModel: values["lead-model"] ?? "claude-opus-4-7[1m]",
     agentModel: values["agent-model"] ?? "claude-opus-4-7[1m]",
     maxTurns,

package/src/discuss-tools.js

@@ -1,15 +1,19 @@
 /**
  * DiscussTools — discuss-mode tool servers. The lead's surface extends the
- * base set with three discuss-only terminal tools:
+ * base set with two discuss-only terminal tools:
  *
- * - `RequestForComment` posts a fire-and-forget message to a human channel
- *   via the bridge; the reply arrives on a later workflow run.
  * - `Recess` suspends the session with a resumption trigger.
  * - `Adjourn` ends the discussion with a verdict.
  *
- * `Conclude` is absent — discuss mode ends via Adjourn or Recess. The
- * agent surface is identical to the facilitated agent's: Ask / Answer /
- * Announce / RollCall, with Ask defaulting to the lead.
+ * `Conclude` is absent — discuss mode ends via Adjourn or Recess.
+ *
+ * `RequestForComment` is an agent-level coordination tool — available on
+ * discuss agents and facilitated agents (not leads). It opens a new
+ * Discussion thread for long-horizon coordination on open questions.
+ *
+ * In discuss mode, each agent Answer routed to the lead is captured as a
+ * thread reply delivered via the bridge callback — no explicit reply tool
+ * is needed on the lead surface.
  */
 
 import { tool } from "@anthropic-ai/claude-agent-sdk";
@@ -19,16 +23,17 @@ import {
   baseTools,
   concludeSession,
   orchestrationServer,
+  requestForCommentTool,
 } from "./orchestration-toolkit.js";
 
-/** System prompt appended for discuss-mode agent runners. */
+/** System prompt for discuss-mode agent participants. L0 mechanics only per COALIGNED. */
 export const DISCUSS_AGENT_SYSTEM_PROMPT =
-  "You participate in an asynchronous discussion. " +
-  "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 lead or another participant and returns immediately with {askIds:[N]}; the reply arrives on a later turn as `[answer#N] <participant>: <text>` in your inbox. " +
-  "Announce broadcasts a message to every other participant — use this for unsolicited remarks or to reply to an Announce. " +
-  "RollCall lists participants.";
+  "You are a participant in a discussion.\n" +
+  "Each question arrives as `[ask#N] <name>: <text>`.\n" +
+  "Quote N as askId on your `Answer` to route the reply correctly.\n" +
+  "Your `Answer` is posted to the discussion thread as a separate reply.\n" +
+  "If the task already contains a completed response with no new human input after it, `Answer` that no further action is needed.\n" +
+  "Do not redo completed work.";
 
 const RESUME_TRIGGER_SCHEMA = z
   .object({
@@ -42,16 +47,6 @@ const RESUME_TRIGGER_SCHEMA = z
 export function createDiscussLeadToolServer(ctx) {
   return orchestrationServer([
     ...baseTools(ctx, { from: "lead", defaultTo: undefined, broadcast: true }),
-    tool(
-      "RequestForComment",
-      "Post a fire-and-forget message to a channel via the bridge. Returns a correlation id; the reply arrives on a later workflow run.",
-      {
-        channel: z.string(),
-        body: z.string(),
-        addressees: z.array(z.string()).optional(),
-      },
-      createRequestForCommentHandler(ctx),
-    ),
     tool(
       "Recess",
       "Suspend the run. The bridge re-dispatches the workflow when the trigger fires.",
@@ -73,33 +68,10 @@ export function createDiscussLeadToolServer(ctx) {
 
 /** Discuss-mode agent tool server. */
 export function createDiscussAgentToolServer(ctx, { from }) {
-  return orchestrationServer(
-    baseTools(ctx, { from, defaultTo: "lead", broadcast: true }),
-  );
-}
-
-/** RequestForComment handler — queues structured replies on `ctx.replies[]`. */
-export function createRequestForCommentHandler(ctx) {
-  return async ({ channel, body, addressees }) => {
-    const correlationId = `rfc_${++ctx.rfcCounter}`;
-    const addresseeList = addressees?.length ? addressees : [null];
-    for (const addressee of addresseeList) {
-      ctx.replies.push({
-        ...(addressee && { addressee }),
-        body,
-        ...(ctx.discussionId && { thread_id: ctx.discussionId }),
-        correlation_id: correlationId,
-      });
-    }
-    return {
-      content: [
-        {
-          type: "text",
-          text: JSON.stringify({ correlation_id: correlationId, channel }),
-        },
-      ],
-    };
-  };
+  return orchestrationServer([
+    ...baseTools(ctx, { from, defaultTo: "lead", broadcast: true }),
+    requestForCommentTool(ctx),
+  ]);
 }
 
 /**

package/src/discusser.js

@@ -1,19 +1,23 @@
 /**
  * Discusser — async, suspendable orchestration on top of a within-run
- * `OrchestrationLoop`. The lead role uses `DiscussTools` (Adjourn / Recess
- * / RequestForComment) instead of the facilitator's Conclude.
+ * `OrchestrationLoop`. The lead role uses `DiscussTools` (Adjourn / Recess)
+ * instead of the facilitator's Conclude.
  *
  * Discuss mode is a sibling of facilitate mode, not a subset of it. The
  * within-run turn loop is shared via `OrchestrationLoop`, but the lead
  * role, tool set, system prompts, and participant naming all stay
  * mode-local.
+ *
+ * Each agent Answer routed to the lead is captured as a thread reply
+ * delivered via the bridge callback — no explicit reply tool is needed
+ * on the lead surface.
  */
 
 import { Writable } from "node:stream";
 import { resolve } from "node:path";
 
 import { createAgentRunner } from "./agent-runner.js";
-import { composeProfilePrompt } from "./profile-prompt.js";
+import { composeSystemPrompt } from "./profile-prompt.js";
 import { SequenceCounter } from "./sequence-counter.js";
 import { createMessageBus } from "./message-bus.js";
 import { createOrchestrationContext } from "./orchestration-toolkit.js";
@@ -24,18 +28,18 @@ import {
 } from "./discuss-tools.js";
 import { OrchestrationLoop } from "./orchestration-loop.js";
 
-/** System prompt appended for the lead (Chair) runner in discuss mode. */
+/** System prompt for the discuss-mode lead. L0 mechanics only per COALIGNED. */
 export const DISCUSS_SYSTEM_PROMPT =
-  "You lead an asynchronous discussion across multiple participants and a human channel. " +
-  "Ask sends a question and returns immediately with {askIds:[N,…]}. The reply arrives on a later turn as `[answer#N] <participant>: <text>` in your inbox — between turns you can plan, reflect, or send more Asks while participants work in parallel. End your turn with text after you've asked everything you intend to; the orchestrator wakes you when the next message lands. " +
-  "Answer replies to an ask a participant addressed to you (you'll see it tagged `[ask#N] <participant>: …` 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. " +
-  "RollCall returns the participant roster. " +
-  "RequestForComment posts a message to the human thread via the bridge. Every reply you want the human to see MUST go through RequestForComment — the bridge delivers only queued replies, not your text output. " +
-  "Recess suspends the run with a resumption trigger (responses / elapsed / either); any open Asks get a synthetic '[no answer: session concluded]' on the asker's queue so nothing dangles. " +
-  "Adjourn ends the discussion with a verdict ('adjourned' / 'failed') and a summary. " +
-  "Multiple Ask / Announce calls in one assistant turn dispatch in parallel — issue them as parallel tool_use blocks rather than sending the same question both broadcast and individually. " +
-  "You MUST call RequestForComment with your response before calling Adjourn. You MUST end every run by calling Adjourn or Recess — never end a turn with only text *after* every Ask round has resolved.";
+  "You lead a discussion.\n" +
+  "You have no tools to perform work yourself.\n" +
+  "Use `RollCall` to list participants.\n" +
+  "Use `Ask` to delegate work to the best-suited participant.\n" +
+  "Participants are domain experts; state the task, not how to do it.\n" +
+  "Each participant's `Answer` is posted to the discussion thread as a separate reply.\n" +
+  "`Ask` returns {askIds:[N,…]} immediately.\n" +
+  "Answers arrive on your next turn as `[answer#N] <participant>: <text>`.\n" +
+  "Multiple `Ask` calls in one turn run participants concurrently.\n" +
+  "Wait for all participants to `Answer` before calling `Adjourn` or `Recess`.";
 
 /**
  * Augment a base orchestration context with discuss-mode fields.
@@ -47,6 +51,7 @@ export function augmentContextForDiscuss(ctx, discussionId) {
   ctx.discussionId = discussionId;
   ctx.recessTrigger = null;
   ctx.replies = [];
+  ctx.rfcs = [];
   ctx.rfcCounter = 0;
   ctx.outcome = null;
   return ctx;
@@ -141,6 +146,7 @@ export class Discusser {
       ...(this.ctx.summary && { summary: this.ctx.summary }),
       ...(this.ctx.outcome && { outcome: this.ctx.outcome }),
       replies: this.ctx.replies,
+      ...(this.ctx.rfcs?.length && { rfcs: this.ctx.rfcs }),
       ...(this.ctx.recessTrigger && { trigger: this.ctx.recessTrigger }),
       ...(this.discussionId && { discussion_id: this.discussionId }),
     };
@@ -228,6 +234,20 @@ export function createDiscusser({
   const messageBus = createMessageBus({
     participants: ["lead", ...resolvedConfigs.map((a) => a.name)],
   });
+
+  // Intercept answers routed to the lead — each becomes a discussion reply.
+  const originalAnswer = messageBus.answer.bind(messageBus);
+  messageBus.answer = (from, to, text, askId) => {
+    if (to === "lead" && from !== "@orchestrator") {
+      ctx.replies.push({
+        body: text,
+        agent: from,
+        ...(ctx.discussionId && { thread_id: ctx.discussionId }),
+      });
+    }
+    originalAnswer(from, to, text, askId);
+  };
+
   ctx.messageBus = messageBus;
   if (ctx.participants.length === 0) {
     ctx.participants = [
@@ -236,16 +256,6 @@ export function createDiscusser({
     ];
   }
 
-  const systemPromptFor = (profile, trailer) => {
-    if (!trailer) throw new Error("trailer is required");
-    return profile
-      ? composeProfilePrompt(profile, {
-          profilesDir: resolvedProfilesDir,
-          trailer,
-        })
-      : { type: "preset", preset: "claude_code", append: trailer };
-  };
-
   let discusser;
   const leadServer = createDiscussLeadToolServer(ctx);
 
@@ -268,26 +278,44 @@ export function createDiscusser({
       onLine: (line) => discusser.loop.emitLine(config.name, line),
       mcpServers: { orchestration: agentServer },
       settingSources: ["project"],
-      systemPrompt: systemPromptFor(config.agentProfile, agentTrailer),
+      systemPrompt: composeSystemPrompt({
+        role: "agent",
+        profile: config.agentProfile,
+        profilesDir: resolvedProfilesDir,
+        trailer: agentTrailer,
+      }),
       redactor,
     });
 
     return { name: config.name, role: config.role, runner };
   });
 
-  const defaultDisallowed = ["Agent", "Task", "TaskOutput", "TaskStop"];
+  const defaultDisallowed = [
+    "Agent",
+    "Task",
+    "TaskOutput",
+    "TaskStop",
+    "Bash",
+    "Write",
+    "Edit",
+  ];
   const leadRunner = createAgentRunner({
     cwd: resolvedLeadCwd,
     query,
     output: devNull,
     model: leadModel ?? "claude-opus-4-7[1m]",
     maxTurns: maxTurns ?? 80,
-    allowedTools: ["Bash", "Read", "Glob", "Grep", "Write", "Edit"],
+    allowedTools: ["Read", "Glob", "Grep"],
     disallowedTools: defaultDisallowed,
     onLine: (line) => discusser.loop.emitLine("lead", line),
     mcpServers: { orchestration: leadServer },
     settingSources: ["project"],
-    systemPrompt: systemPromptFor(leadProfile, DISCUSS_SYSTEM_PROMPT),
+    systemPrompt: composeSystemPrompt({
+      role: "lead",
+      profile: leadProfile,
+      profilesDir: resolvedProfilesDir,
+      trailer: DISCUSS_SYSTEM_PROMPT,
+    }),
     redactor,
   });
 

package/src/facilitator.js

@@ -9,7 +9,7 @@
 import { Writable } from "node:stream";
 import { resolve } from "node:path";
 import { createAgentRunner } from "./agent-runner.js";
-import { composeProfilePrompt } from "./profile-prompt.js";
+import { composeSystemPrompt } from "./profile-prompt.js";
 import { createMessageBus } from "./message-bus.js";
 import {
   createOrchestrationContext,
@@ -18,26 +18,25 @@ import {
 } from "./orchestration-toolkit.js";
 import { OrchestrationLoop } from "./orchestration-loop.js";
 
-/** System prompt appended for the facilitator runner. */
+/** System prompt for the facilitator lead. L0 mechanics only per COALIGNED. */
 export const FACILITATOR_SYSTEM_PROMPT =
-  "You coordinate multiple participants via these tools: " +
-  "Ask sends a question and returns immediately with {askIds:[N,…]}. The reply arrives on a later turn as `[answer#N] <participant>: <text>` in your inbox — between turns you can plan, reflect, or send more Asks while participants work in parallel. End your turn with text after you've asked everything you intend to; the orchestrator wakes you again as soon as a reply (or any message) lands. " +
-  "Answer replies to an ask a participant addressed to you (you'll see it tagged `[ask#N] <participant>: …` 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. " +
-  "RollCall returns the participant roster. " +
-  "Conclude ends the session with a verdict ('success' or 'failure') and a summary. " +
-  "Multiple Ask / Announce calls in one assistant turn dispatch in parallel — issue them as parallel tool_use blocks rather than sending the same question both broadcast and individually. " +
-  "You MUST end every session with Conclude — never end a turn with only text *after* every Ask round has resolved. " +
-  "If you can answer the task yourself, still call Conclude with verdict='success' and the answer as the summary.";
-
-/** System prompt appended for facilitated agent runners. */
+  "You are the facilitator.\n" +
+  "You have no tools to perform work yourself.\n" +
+  "Use `RollCall` to list participants.\n" +
+  "Use `Ask` to delegate work to the best-suited participant.\n" +
+  "Participants are domain experts; state the task, not how to do it.\n" +
+  "`Ask` returns {askIds:[N,…]} immediately.\n" +
+  "Answers arrive on your next turn as `[answer#N] <participant>: <text>`.\n" +
+  "Multiple `Ask` calls in one turn run participants concurrently.\n" +
+  "Wait for all participants to `Answer` before calling `Conclude`.";
+
+/** System prompt for facilitated agent participants. L0 mechanics only per COALIGNED. */
 export const FACILITATED_AGENT_SYSTEM_PROMPT =
-  "You participate in a coordinated session. " +
-  "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 another participant and returns immediately with {askIds:[N]}; the reply arrives on a later turn as `[answer#N] <participant>: <text>` in your inbox. " +
-  "Announce broadcasts a message to every other participant — use this for unsolicited remarks or to reply to an Announce. " +
-  "RollCall lists participants.";
+  "You are a participant in a facilitated session.\n" +
+  "Each question arrives as `[ask#N] <name>: <text>`.\n" +
+  "Quote N as askId on your `Answer` to route the reply correctly.\n" +
+  "If the task already contains a completed response with no new human input after it, `Answer` that no further action is needed.\n" +
+  "Do not redo completed work.";
 
 /**
  * Facilitate-mode wrapper around `OrchestrationLoop`. The lead is named
@@ -113,15 +112,6 @@ export function createFacilitator({
   if (!redactor) throw new Error("redactor is required");
   const resolvedProfilesDir =
     profilesDir ?? resolve(facilitatorCwd, ".claude/agents");
-  const systemPromptFor = (profile, trailer) => {
-    if (!trailer) throw new Error("trailer is required");
-    return profile
-      ? composeProfilePrompt(profile, {
-          profilesDir: resolvedProfilesDir,
-          trailer,
-        })
-      : { type: "preset", preset: "claude_code", append: trailer };
-  };
   const ctx = createOrchestrationContext();
   const messageBus = createMessageBus({
     participants: ["facilitator", ...agentConfigs.map((a) => a.name)],
@@ -155,17 +145,27 @@ export function createFacilitator({
       onLine: (line) => facilitator.emitLine(config.name, line),
       mcpServers: { orchestration: agentServer },
       settingSources: ["project"],
-      systemPrompt: systemPromptFor(config.agentProfile, agentTrailer),
+      systemPrompt: composeSystemPrompt({
+        role: "agent",
+        profile: config.agentProfile,
+        profilesDir: resolvedProfilesDir,
+        trailer: agentTrailer,
+      }),
       redactor,
     });
 
     return { name: config.name, role: config.role, runner };
   });
 
-  // Block the SDK's sub-agent spawn tools on the facilitator: its job is to
-  // coordinate participants through the libeval orchestration harness, not
-  // to fan work out to ad-hoc Claude Code sub-agents. Mirrors the supervisor.
-  const defaultDisallowed = ["Agent", "Task", "TaskOutput", "TaskStop"];
+  const defaultDisallowed = [
+    "Agent",
+    "Task",
+    "TaskOutput",
+    "TaskStop",
+    "Bash",
+    "Write",
+    "Edit",
+  ];
   const disallowedTools = facilitatorDisallowedTools
     ? [...new Set([...defaultDisallowed, ...facilitatorDisallowedTools])]
     : defaultDisallowed;
@@ -176,22 +176,17 @@ export function createFacilitator({
     output: devNull,
     model: facilitatorModel ?? model,
     maxTurns: maxTurns ?? 80,
-    allowedTools: facilitatorAllowedTools ?? [
-      "Bash",
-      "Read",
-      "Glob",
-      "Grep",
-      "Write",
-      "Edit",
-    ],
+    allowedTools: facilitatorAllowedTools ?? ["Read", "Glob", "Grep"],
     disallowedTools,
     onLine: (line) => facilitator.emitLine("facilitator", line),
     mcpServers: { orchestration: facilitatorServer },
     settingSources: ["project"],
-    systemPrompt: systemPromptFor(
-      facilitatorProfile,
-      FACILITATOR_SYSTEM_PROMPT,
-    ),
+    systemPrompt: composeSystemPrompt({
+      role: "lead",
+      profile: facilitatorProfile,
+      profilesDir: resolvedProfilesDir,
+      trailer: FACILITATOR_SYSTEM_PROMPT,
+    }),
     redactor,
   });
 

package/src/index.js

@@ -8,7 +8,11 @@ export {
   parseGitRemote,
 } from "./trace-github.js";
 export { AgentRunner, createAgentRunner } from "./agent-runner.js";
-export { composeProfilePrompt } from "./profile-prompt.js";
+export {
+  composeProfilePrompt,
+  composeLeadPrompt,
+  composeSystemPrompt,
+} from "./profile-prompt.js";
 export {
   Supervisor,
   createSupervisor,
@@ -19,6 +23,7 @@ export { TeeWriter, createTeeWriter } from "./tee-writer.js";
 export { SequenceCounter, createSequenceCounter } from "./sequence-counter.js";
 export {
   createOrchestrationContext,
+  createRequestForCommentHandler,
   createSupervisorToolServer,
   createSupervisedAgentToolServer,
   createFacilitatorToolServer,

package/src/orchestration-toolkit.js

@@ -5,15 +5,15 @@
  *
  * **Tool surface, by role:**
  *
- *   |             | Ask | Answer | Announce | RollCall | Conclude | …extras |
- *   |-------------|-----|--------|----------|----------|----------|---------|
- *   | Facilitator |  ✓  |   ✓    |    ✓     |    ✓     |    ✓     |         |
- *   | Fac. agent  |  ✓  |   ✓    |    ✓     |    ✓     |          |         |
- *   | Supervisor  |  ✓  |   ✓    |    ✓     |    ✓     |    ✓     |         |
- *   | Sup. agent  |  ✓  |   ✓    |    ✓     |    ✓     |          |         |
- *   | Discuss lead|  ✓  |   ✓    |    ✓     |    ✓     |          | RFC / Recess / Adjourn |
- *   | Discuss agt |  ✓  |   ✓    |    ✓     |    ✓     |          |         |
- *   | Judge       |     |        |          |          |    ✓     |         |
+ *   |             | Ask | Answer | Announce | RollCall | Conclude | …extras              |
+ *   |-------------|-----|--------|----------|----------|----------|-----------------------|
+ *   | Facilitator |  ✓  |   ✓    |    ✓     |    ✓     |    ✓     |                       |
+ *   | Fac. agent  |  ✓  |   ✓    |    ✓     |    ✓     |          | RFC                   |
+ *   | Supervisor  |  ✓  |   ✓    |    ✓     |    ✓     |    ✓     |                       |
+ *   | Sup. agent  |  ✓  |   ✓    |    ✓     |    ✓     |          |                       |
+ *   | Discuss lead|  ✓  |   ✓    |    ✓     |    ✓     |          | Recess / Adjourn      |
+ *   | Discuss agt |  ✓  |   ✓    |    ✓     |    ✓     |          | RFC                   |
+ *   | Judge       |     |        |          |          |    ✓     |                       |
  *
  * **Ask is async.** Ask returns `{askIds:[…]}` immediately and posts the
  * question to the addressee's bus queue. The reply arrives on the asker's
@@ -337,11 +337,12 @@ export function createFacilitatorToolServer(ctx) {
   ]);
 }
 
-/** Facilitated agent tools: Ask + Answer + Announce + RollCall. */
+/** Facilitated agent tools: Ask + Answer + Announce + RollCall + RequestForComment. */
 export function createFacilitatedAgentToolServer(ctx, { from }) {
-  return orchestrationServer(
-    baseTools(ctx, { from, defaultTo: "facilitator", broadcast: true }),
-  );
+  return orchestrationServer([
+    ...baseTools(ctx, { from, defaultTo: "facilitator", broadcast: true }),
+    requestForCommentTool(ctx),
+  ]);
 }
 
 /**
@@ -352,6 +353,42 @@ export function createJudgeToolServer(ctx) {
   return orchestrationServer([concludeTool(ctx)]);
 }
 
+// --- RequestForComment (agent-level coordination tool) ---
+
+/** RequestForComment handler — queues RFC intent on `ctx.rfcs[]`. */
+export function createRequestForCommentHandler(ctx) {
+  return async ({ channel, body, addressees }) => {
+    if (!ctx.rfcs) ctx.rfcs = [];
+    if (typeof ctx.rfcCounter !== "number") ctx.rfcCounter = 0;
+    const correlationId = `rfc_${++ctx.rfcCounter}`;
+    const addresseeList = addressees?.length ? addressees : [null];
+    for (const addressee of addresseeList) {
+      ctx.rfcs.push({
+        ...(addressee && { addressee }),
+        body,
+        channel,
+        ...(ctx.discussionId && { thread_id: ctx.discussionId }),
+        correlation_id: correlationId,
+      });
+    }
+    return jsonResult({ correlation_id: correlationId, channel });
+  };
+}
+
+/** Build the RequestForComment tool definition. */
+function requestForCommentTool(ctx) {
+  return tool(
+    "RequestForComment",
+    "Open a new Discussion thread for long-horizon coordination on an open question. The bridge creates the thread; replies arrive asynchronously on future runs.",
+    {
+      channel: z.string(),
+      body: z.string(),
+      addressees: z.array(z.string()).optional(),
+    },
+    createRequestForCommentHandler(ctx),
+  );
+}
+
 // Re-export the building blocks discuss-tools.js needs to assemble its
-// own lead tool surface (it has three extra terminal tools).
-export { baseTools, orchestrationServer };
+// own lead tool surface (it has two extra terminal tools).
+export { baseTools, orchestrationServer, requestForCommentTool };

package/src/profile-prompt.js

@@ -1,22 +1,28 @@
 /**
- * Compose an SDK `systemPrompt` value from a `.claude/agents/<name>.md` file.
+ * System prompt composition for agent runners.
  *
- * Pure function. Reads the profile file, strips YAML frontmatter, and returns
- * the SDK-shaped `{ type: "preset", preset: "claude_code", append }` object
- * with the profile body — plus an optional mode-specific trailer — in the
- * `append` slot. Callers in libeval pass the result straight into an
- * `AgentRunner`'s `systemPrompt` input so the profile reaches the main-thread
- * system prompt without going through the SDK's top-level `agent` option.
+ * Two helpers:
+ *
+ * - `composeProfilePrompt(name, opts)` — profile + `claude_code` preset.
+ *   Used by agent participants that need the full Claude Code tool surface.
+ *
+ * - `composeLeadPrompt(opts)` — plain string, no preset. Used by lead
+ *   roles (supervisor, facilitator, discuss lead) that should only see
+ *   the orchestration instructions and optionally a profile body.
+ *
+ * - `composeSystemPrompt(opts)` — unified entry point. Delegates to one
+ *   of the above based on `opts.role`.
  */
 
 import { readFileSync } from "node:fs";
 import { join } from "node:path";
 
 /**
+ * Compose a `claude_code`-preset system prompt from a profile file.
  * @param {string} name - Profile basename (no `.md` suffix)
  * @param {object} opts
  * @param {string} opts.profilesDir - Directory containing `<name>.md`
- * @param {string} [opts.trailer] - Optional mode-specific trailer appended after a blank line
+ * @param {string} [opts.trailer] - Mode-specific trailer appended after a blank line
  * @returns {{type: "preset", preset: "claude_code", append: string}}
  */
 export function composeProfilePrompt(name, { profilesDir, trailer }) {
@@ -27,6 +33,45 @@ export function composeProfilePrompt(name, { profilesDir, trailer }) {
   return { type: "preset", preset: "claude_code", append };
 }
 
+/**
+ * Compose a plain-string system prompt for a lead role (no Claude Code preset).
+ * @param {object} opts
+ * @param {string} [opts.profile] - Profile basename (no `.md` suffix)
+ * @param {string} [opts.profilesDir] - Directory containing profile files
+ * @param {string} opts.trailer - Mode-specific orchestration instructions
+ * @returns {string}
+ */
+export function composeLeadPrompt({ profile, profilesDir, trailer }) {
+  if (!trailer) throw new Error("trailer is required");
+  if (!profile) return trailer;
+  const path = join(profilesDir, `${profile}.md`);
+  const raw = readFileSync(path, "utf8");
+  const body = stripFrontmatter(raw).trim();
+  return `${body}\n\n${trailer}`;
+}
+
+/**
+ * Unified entry point for composing system prompts.
+ *
+ * @param {object} opts
+ * @param {"lead"|"agent"} opts.role - `"lead"` produces a plain string;
+ *   `"agent"` produces a `claude_code` preset object.
+ * @param {string} [opts.profile] - Profile basename
+ * @param {string} [opts.profilesDir]
+ * @param {string} opts.trailer - Mode-specific instructions
+ * @returns {string | {type: "preset", preset: "claude_code", append: string}}
+ */
+export function composeSystemPrompt({ role, profile, profilesDir, trailer }) {
+  if (!trailer) throw new Error("trailer is required");
+  if (role === "lead") {
+    return composeLeadPrompt({ profile, profilesDir, trailer });
+  }
+  if (profile) {
+    return composeProfilePrompt(profile, { profilesDir, trailer });
+  }
+  return { type: "preset", preset: "claude_code", append: trailer };
+}
+
 /**
  * Strip a leading YAML frontmatter fence (`---\n…\n---\n`) from a markdown
  * string. Returns the input unchanged when no frontmatter is present.

package/src/supervisor.js

@@ -18,7 +18,7 @@
 import { Writable } from "node:stream";
 import { resolve } from "node:path";
 import { createAgentRunner } from "./agent-runner.js";
-import { composeProfilePrompt } from "./profile-prompt.js";
+import { composeSystemPrompt } from "./profile-prompt.js";
 import { createMessageBus } from "./message-bus.js";
 import {
   createOrchestrationContext,
@@ -27,23 +27,23 @@ import {
 } from "./orchestration-toolkit.js";
 import { OrchestrationLoop } from "./orchestration-loop.js";
 
-/** System prompt appended for the supervisor runner in supervise mode. */
+/** System prompt for the supervisor lead. L0 mechanics only per COALIGNED. */
 export const SUPERVISOR_SYSTEM_PROMPT =
-  "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.";
+  "You supervise one agent.\n" +
+  "You have no tools to perform work yourself.\n" +
+  "Use `Ask` to delegate work to the agent.\n" +
+  "`Ask` returns {askIds:[N]} immediately.\n" +
+  "The reply arrives on your next turn as `[answer#N] agent: <text>`.\n" +
+  "If the agent goes off-track, send a corrective `Ask`.\n" +
+  "End every session by calling `Conclude`.";
 
-/** System prompt appended for the agent runner in supervise mode. */
+/** System prompt for the supervised agent. L0 mechanics only per COALIGNED. */
 export const AGENT_SYSTEM_PROMPT =
-  "A supervisor watches your work. " +
-  "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.";
+  "A supervisor directs your work.\n" +
+  "Each question arrives as `[ask#N] supervisor: <text>`.\n" +
+  "Quote N as askId on your `Answer` to route the reply correctly.\n" +
+  "If the task already contains a completed response with no new human input after it, `Answer` that no further action is needed.\n" +
+  "Do not redo completed work.";
 
 /**
  * Supervise-mode wrapper around `OrchestrationLoop`. The lead is
@@ -148,15 +148,6 @@ export function createSupervisor({
   if (!redactor) throw new Error("redactor is required");
   const resolvedProfilesDir =
     profilesDir ?? resolve(supervisorCwd, ".claude/agents");
-  const systemPromptFor = (profile, trailer) => {
-    if (!trailer) throw new Error("trailer is required");
-    return profile
-      ? composeProfilePrompt(profile, {
-          profilesDir: resolvedProfilesDir,
-          trailer,
-        })
-      : { type: "preset", preset: "claude_code", append: trailer };
-  };
 
   const ctx = createOrchestrationContext();
   const messageBus = createMessageBus({
@@ -183,15 +174,25 @@ export function createSupervisor({
     allowedTools,
     onLine: (line) => supervisor.emitLine("agent", line),
     settingSources: ["project"],
-    systemPrompt: systemPromptFor(agentProfile, AGENT_SYSTEM_PROMPT),
+    systemPrompt: composeSystemPrompt({
+      role: "agent",
+      profile: agentProfile,
+      profilesDir: resolvedProfilesDir,
+      trailer: AGENT_SYSTEM_PROMPT,
+    }),
     mcpServers: { orchestration: agentServer, ...agentMcpServers },
     redactor,
   });
 
-  // 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 defaultDisallowed = [
+    "Agent",
+    "Task",
+    "TaskOutput",
+    "TaskStop",
+    "Bash",
+    "Write",
+    "Edit",
+  ];
   const disallowedTools = supervisorDisallowedTools
     ? [...new Set([...defaultDisallowed, ...supervisorDisallowedTools])]
     : defaultDisallowed;
@@ -202,18 +203,16 @@ export function createSupervisor({
     output: devNull,
     model: supervisorModel ?? model,
     maxTurns: perRunBudget,
-    allowedTools: supervisorAllowedTools ?? [
-      "Bash",
-      "Read",
-      "Glob",
-      "Grep",
-      "Write",
-      "Edit",
-    ],
+    allowedTools: supervisorAllowedTools ?? ["Read", "Glob", "Grep"],
     disallowedTools,
     onLine: (line) => supervisor.emitLine("supervisor", line),
     settingSources: ["project"],
-    systemPrompt: systemPromptFor(supervisorProfile, SUPERVISOR_SYSTEM_PROMPT),
+    systemPrompt: composeSystemPrompt({
+      role: "lead",
+      profile: supervisorProfile,
+      profilesDir: resolvedProfilesDir,
+      trailer: SUPERVISOR_SYSTEM_PROMPT,
+    }),
     mcpServers: { orchestration: supervisorServer },
     redactor,
   });