fix(scout): a first reply answers what was asked in two or three sentences, then one next step

backend/__tests__/unit/config/nativeAgents.scoutReplyLength.test.js

@@ -0,0 +1,33 @@
+/**
+ * Scout's first replies stay short (F60, staging walk 2026-11-02).
+ *
+ * Asked what it could do, Scout named five things in one sentence, and the reply
+ * folded behind "Show more" on a phone. The prompt told it to answer the literal
+ * question, and the literal answer to "what can you do?" is the whole catalogue.
+ * Measured offline on the same model before this rule (9 first posts, 3 samples
+ * each): median 265 characters, 6 of 27 replies past the desktop fold, 19 of 27
+ * past it at phone width. With it: 181, 0 of 27, 4 of 27.
+ */
+const { scoutApp } = require('../../../config/native-agents/scout');
+
+describe('Scout reply length', () => {
+  const prompt = scoutApp.systemPrompt;
+
+  it('budgets every reply and names the case that ran long', () => {
+    expect(prompt).toContain('two or three short sentences, about 200 characters');
+    expect(prompt).toContain('Never volunteer a list of what you or Commonly can do');
+    expect(prompt).toContain('name the one thing most useful to them now');
+  });
+
+  it('exempts what the person asked for and every caveat the rules require', () => {
+    expect(prompt).toContain('a caveat these rules tell you to state, may run longer');
+    expect(prompt).toContain('Asked which tools, models or paths work, name them');
+  });
+
+  it('repeats the budget in the hard rules, the last thing the model reads, with its exemption', () => {
+    const hardRules = prompt.slice(prompt.lastIndexOf('ROLE-SPECIFIC HARD RULES:'));
+    expect(hardRules).toContain('A first impression is also short');
+    expect(hardRules).toContain('unless a command or a caveat these rules require needs more');
+    expect(hardRules).toContain('never a tour of what you can do');
+  });
+});

backend/config/native-agents/scout.ts

@@ -106,8 +106,8 @@ export const scoutApp = {
     + 'the person\'s own message: '
     + `"${startLabel('ask-scout')}", "${startLabel('connect-agent')}" and "${startLabel('start-pod')}". `
     + `For "${startLabel('ask-scout')}", answer in two or three short sentences: `
-    + 'name what you can do for them here, then offer one thing you can do right '
-    + 'now in this pod, such as putting their first task on the board. '
+    + 'name the one thing most useful to them here, not a list, then offer to do '
+    + 'it right now in this pod, such as putting their first task on the board. '
     + `For "${startLabel('connect-agent')}", find out which tool they run `
     + '(Claude Code or Codex) and their operating system, one question per turn, '
     + 'then propose the connect card when the card rule below allows it; in that first reply, also say in one short '
@@ -316,6 +316,14 @@ export const scoutApp = {
     + 'HOW YOU ANSWER:\n'
     + '- Answer the literal question first, in your first sentence, with the '
     + 'concrete step or command. Then stop.\n'
+    // F60 (staging walk 2026-11-02): asked what it can do, Scout named five
+    // things in one sentence, and the reply folded behind "Show more" on a phone.
+    + '- Keep a reply to two or three short sentences, about 200 characters, and '
+    + 'make the last one the next step. Only a command or step they asked for, or '
+    + 'a caveat these rules tell you to state, may run longer. Never volunteer a '
+    + 'list of what you or Commonly can do: asked what you can do, what this is, '
+    + 'or just greeted, name the one thing most useful to them now and offer to '
+    + 'do it. Asked which tools, models or paths work, name them.\n'
     + '- Ask at most one question per turn, and only when the answer truly '
     + 'depends on it (their OS, or which tool they run).\n'
     + '- Before ANY terminal command: if you do not know their operating '
@@ -365,7 +373,10 @@ export const scoutApp = {
     'Never claim the user\'s own agent is connected or working, or that it '
     + 'is not, unless commonly_agent_status said so in this turn. And '
     + 'remember you are the product\'s first impression: an invented product '
-    + 'fact from you is worse than one from anyone else.',
+    + 'fact from you is worse than one from anyone else. A first impression is '
+    + 'also short: two or three short sentences, the last one the next step, '
+    + 'unless a command or a caveat these rules require needs more, and never '
+    + 'a tour of what you can do.',
   ),
   model: 'deepseek-v4-flash',
   triggers: ['mention', 'chat.message'],