@sonnechasser/ntrp 1.4.6 → 1.4.7

package/dist/index.js

@@ -232,6 +232,30 @@ function severityLabel(severity) {
       return severity.toUpperCase();
   }
 }
+function pickGoverningFinding(findings) {
+  if (findings.length === 0) return null;
+  return findings.find((f) => f.severity === "critical") ?? findings.find((f) => f.severity === "warning") ?? findings[0] ?? null;
+}
+function findingHeadline(findingText) {
+  const trimmed = findingText.trim();
+  if (!trimmed) return "";
+  const match = trimmed.match(/^(.+?[.!?])(?:\s|$)/);
+  return (match?.[1] ?? trimmed.split(/\n/)[0] ?? trimmed).trim();
+}
+function formatTheCallSection(findings) {
+  const pick3 = pickGoverningFinding(findings);
+  if (!pick3) return "";
+  const headline = findingHeadline(pick3.finding);
+  if (!headline) return "";
+  const alreadyHasDollar = /\$[\d.,]+/.test(headline);
+  const dollar = !alreadyHasDollar && pick3.dollar_value != null && pick3.dollar_value > 0 ? `**${formatDollarValue(pick3.dollar_value)}** \u2014 ` : "";
+  return [
+    "## The Call",
+    "",
+    `${dollar}${headline}`,
+    ""
+  ].join("\n");
+}
 var VITAL_SIGN_LABELS, DOLLAR_LABELS;
 var init_formatters = __esm({
   "src/output/formatters.ts"() {
@@ -20719,7 +20743,7 @@ function buildFindingsFindingHint(prefs) {
   const voice2 = prefs ?? loadVoicePrefs();
   const register = voice2.personality === "robotic" ? "STE-100 sentences. No contractions, slang, or filler." : voice2.personality === "composed" ? "Composed consultant sentences. Light contractions OK." : voice2.personality === "casual" ? "Casual consultant sentences. Warm, still short." : "Loose consultant sentences. Personality in the wrap, not the math.";
   const roast = voice2.roast === "light" ? "Light roast: what is working, then what to improve." : voice2.roast === "medium" ? "Medium roast: name the miss. No padding." : voice2.roast === "dark" ? "Dark roast: surgical. Verdict + dollar in one breath." : "Heavy roast: maximum roast after the number, never instead of it.";
-  return `Pyramid-shaped, 2-3 sentences: (1) HEADLINE \u2014 verdict + dollar figure in one short sentence (\u226420 words); (2) EVIDENCE \u2014 the one or two numbers that prove it; (3) SO-WHAT \u2014 the consequence or the action. ${register} ${roast} Keep dollar figures and play names. Do not invent numbers. An executive should be able to repeat sentence 1 from memory.`;
+  return `Pyramid-shaped, 2-3 sentences: (1) HEADLINE \u2014 verdict + dollar figure in one short sentence (\u226420 words); (2) DRIVERS \u2014 the one or two distinct, non-overlapping numbers that prove it; (3) SO-WHAT \u2014 the consequence or the action. ${register} ${roast} Keep dollar figures and play names. Do not invent numbers. An executive should be able to repeat sentence 1 from memory.`;
 }
 function buildFindingsSchemaBlock(prefs) {
   const finding = buildFindingsFindingHint(prefs);
@@ -21915,10 +21939,10 @@ function wrapForTarget(target, analysisBlock, conversationBlock, openQuestions,
   const contextLabel = handoffInstructionPrefix(ctx.analysis.primary);
   const forWhom = audiencePhrase(ctx.scope?.audience);
   const instructions = {
-    deck: `produce an executive review deck outline for ${company}, framed for ${forWhom}. Structure: (1) Headline number; (2) Pipeline health; (3) Top 3 risks with dollar impact; (4) Recommended plays; (5) Asks / decisions. Use the Metric definitions appendix so every slide can briefly remind the room what the number means for this audience.`,
+    deck: `produce an executive review deck outline for ${company}, framed for ${forWhom}. Pyramid structure: (1) HEADLINE \u2014 the governing number and verdict in one line; (2) DRIVERS \u2014 2\u20133 distinct, non-overlapping risks or causes with dollar impact (not a full scorecard); (3) SO-WHAT \u2014 recommended plays; (4) ASKS / decisions the room must make. Optionally one pipeline-health slide after the headline if it supports the call. Use the Metric definitions appendix so every slide can briefly remind the room what the number means for this audience.`,
     asana: `produce an Asana project plan with sections and tasks tied to findings for ${forWhom}. Prioritize by dollar impact. Reference metric definitions when a task owner needs to know what "good" looks like.`,
     clay: `produce a Clay table specification to operationalize the highest-impact finding for ${forWhom}.`,
-    plan: `produce a prioritized action plan for ${forWhom} with problem, play, first 3 steps, owner, and leading indicator per item. Ground indicators in the Metric definitions appendix.`
+    plan: `produce a prioritized action plan for ${forWhom}, answer-first: open with the governing call (highest-dollar problem), then items each with problem, play, first 3 steps, owner, and leading indicator. Ground indicators in the Metric definitions appendix. End with the sharpest asks / decisions.`
   };
   const parts = [
     `# NTRP handoff \u2192 ${target}`,
@@ -22825,6 +22849,7 @@ ${opts.conversationBlock}
   const socraticCraft = `HOW YOU THINK WITH THE USER (socratic partner \u2014 not a search box, not a strategist):
 - Pre-explore funnel gates are handled by the CLI; you speak in the think channel after analysis.
 - You are a thinking partner. Prefer sharp questions that unlock the user's idea over dumping answers.
+- Probing turns: lead with the sharpest question that unlocks the next insight. Concluding turns: use pyramid (headline \u2192 drivers \u2192 so-what) \u2014 answer first, then structure.
 - Protect imagination: give "what if" space before the smell-test \u2014 do not kill creativity in the first sentence.
 - Steelman the user's position before you attack it; then deliver one hard pushback grounded in evidence when possible.
 - Ground claims about THIS pipeline in tool results. Label speculation explicitly ("hypothesis:", "speculation:").
@@ -22835,7 +22860,7 @@ ${opts.conversationBlock}
 - You have continuity via prior think-channel messages. Never repeat an angle already covered unless asked.`;
   const jobSection = `YOUR JOB (THINK CHANNEL \u2014 always deep):
 - Decide whether you need tools, a direct answer, or both. Do not re-call a tool whose result you already have.
-- Lead with the answer or the question that matters most, then structure.
+- Probing: lead with the question that matters most. Concluding: lead with the answer (headline), then drivers and so-what.
 - When you use numbers, include dollar values where available and lead with financial impact.
 - Descriptive exploration stays in this channel; plan-of-attack questions hand off via draft_strategy.
 - After a successful draft_strategy tool result: reply with at most 1\u20132 sentences introducing the handoff. Do not write the plan.`;
@@ -22873,6 +22898,9 @@ ${EXECUTION_BIAS_BLOCK}
 GTM ENGINEERING (how recommendations become systems):
 ${GTM_ENGINEERING_BLOCK}
 
+IDEA GATE (when recommending a play or before draft_strategy \u2014 silent; probing/imagination turns skip this until you conclude):
+${HEILMEIER_GATE}
+
 OUTPUT DOCTRINE (how answers are structured for recall):
 ${PYRAMID_OUTPUT_BLOCK}
 
@@ -23273,12 +23301,18 @@ ${commandStyleRule}`;
 - Respond in plain text (not JSON). You do NOT need to emit the findings schema \u2014 only emit structured findings if the user explicitly asks for them.
 - Include specific numbers from tool results, never guess.
 - When you have enough information, answer directly. Don't call tools you don't need.`;
-  const executionBiasSection = responseMode === "brief" ? "" : `EXECUTION BIAS (how you work):
+  const executionBiasSection = responseMode === "brief" ? `OUTPUT DOCTRINE (brief \u2014 one headline + \u22642 drivers + one so-what):
+${PYRAMID_OUTPUT_BLOCK}
+
+` : `EXECUTION BIAS (how you work):
 ${EXECUTION_BIAS_BLOCK}
 
 GTM ENGINEERING (how your recommendations become systems):
 ${GTM_ENGINEERING_BLOCK}
 
+IDEA GATE (when recommending action \u2014 silent):
+${HEILMEIER_GATE}
+
 OUTPUT DOCTRINE (how answers are structured for recall):
 ${PYRAMID_OUTPUT_BLOCK}
 
@@ -25940,6 +25974,11 @@ function generateMarkdownReport(data) {
   lines.push("");
   lines.push(`*Generated: ${data.generatedAt}*`);
   lines.push("");
+  const theCall = formatTheCallSection(data.findings);
+  if (theCall) {
+    lines.push(theCall.trimEnd());
+    lines.push("");
+  }
   lines.push("## Overall Health");
   lines.push("");
   lines.push(`${statusEmoji(data.health.overall_status)} **Score: ${formatScore(data.health.overall_score)} / 100** (${data.health.overall_status})`);
@@ -26373,6 +26412,11 @@ function exportToNotes(data) {
   const body = [];
   body.push("# GTM Health Diagnosis");
   body.push("");
+  const theCall = formatTheCallSection(findings);
+  if (theCall) {
+    body.push(theCall.trimEnd());
+    body.push("");
+  }
   body.push(`**Overall Score:** ${aggregate.overall_score}/100 (${aggregate.overall_status})`);
   body.push(`**Gating Vital Sign:** ${VITAL_SIGN_LABELS[aggregate.gating_vital_sign]}`);
   body.push(`**Date:** ${now2.toISOString().split("T")[0]} ${timeStr.replace(/(\d{2})(\d{2})/, "$1:$2")}`);
@@ -28436,10 +28480,13 @@ ${companyContextSection4()}${operatorSection4()}HOW YOU THINK (the strategist me
 8. RESPECT CAPACITY. Total effort must fit the team that actually exists. A brilliant plan the team cannot staff is a bad plan.
 
 ALTITUDE CONTRACT (the plan must work at every altitude a client reads it at):
-- 30,000 FT: summary_30k is a situation-complication-resolution narrative \u2014 where the business stands, what gates what, and what leadership should expect by when. A board member reads only this; it must survive being forwarded unedited.
+- 30,000 FT: summary_30k is answer-first \u2014 open with the governing thought and what leadership should expect by when, then the situation and complication that make it true. A board member reads only this; it must survive being forwarded unedited.
 - 10,000 FT: each workstream's title + problem line is a delegation unit \u2014 a one-liner an owner could receive in Slack and know what they own, why it's theirs, and what number they move.
 - GROUND LEVEL: actions are Monday-morning prescriptive. You have built this before \u2014 read the linked play's full detail with get_play_detail and prescribe its known-good sequence adapted to THIS company's numbers and constraints. The first action of every workstream must be startable within 48 hours with no new tooling.
 
+OUTPUT DOCTRINE (narrative fields \u2014 summary_30k and workstream problem lines \u2014 follow pyramid recall):
+${PYRAMID_OUTPUT_BLOCK}
+
 ${buildUserVisibleProseBlock()}
 
 GTM ENGINEERING (plans install systems, not heroics):
@@ -28522,6 +28569,8 @@ function buildStressTestMessage() {
 6. EXPECTATION HONESTY: are target ranges defensible from the baseline and effort, or heroic? Widen ranges or lower confidence rather than promising what the data does not support. Where the local play track record shows a play has hit or missed here before, weight confidence accordingly.
 7. CONTINGENCY QUALITY: is each trigger observable on a specific date, and is each fallback a real pre-decision (alternate play, descope, escalate) rather than "monitor closely"?
 8. ALTITUDE CHECK: does summary_30k survive being forwarded to a board member unedited? Is each workstream title + problem a self-contained delegation one-liner? Is every first action startable within 48 hours?
+9. HEILMEIER GATE (silent \u2014 never name this to the user). Run every workstream through this filter. Failures: cut the workstream, demote it to assumptions, or rewrite until it passes \u2014 then emit final JSON.
+${HEILMEIER_GATE}
 
 Then respond with the FINAL revised plan as strict JSON in the same schema \u2014 no markdown fences, no prose. Fold what you learned into constraints, assumptions, risks, and confidence. This version is the one that gets saved and reviewed against, so make every number one you are willing to be checked on.`;
 }
@@ -33182,11 +33231,14 @@ function buildRecapSystemPrompt(companyBlock) {
     "",
     companyBlock ? `## Company Context
 ${companyBlock}` : "",
+    "",
+    "OUTPUT DOCTRINE (how the recap is structured for recall):",
+    PYRAMID_OUTPUT_BLOCK,
     "",
     "Produce a concise session recap in markdown, pyramid-shaped:",
-    "1. **The Story** \u2014 one situation-complication-resolution sentence: where the business stood, what this session surfaced, and what that means. This is the line the client repeats tomorrow.",
-    "2. **Key Findings** \u2014 3-5 bullets, each headline-first with its number and dollar value; no bullet without a figure unless none exists.",
-    "3. **Next Steps** \u2014 2-3 actions, each with an owner-shaped verb and, where a play applies, the play named.",
+    "1. **The Story** \u2014 answer-first: open with the resolution (what leadership should take away), then the situation and complication that make it true. One or two sentences. This is the line the client repeats tomorrow.",
+    "2. **Key Findings** \u2014 3-5 bullets as drivers: each headline-first with its number and dollar value; distinct and non-overlapping; no bullet without a figure unless none exists.",
+    "3. **Next Steps** \u2014 2-3 so-what actions, each with an owner-shaped verb and, where a play applies, the play named.",
     "4. **Open Loops** \u2014 anything left unresolved or promised for next time (omit the section if none).",
     "",
     "Keep the total output under 300 words.",
@@ -37048,7 +37100,7 @@ function formatCatalogLine(meta) {
   const note = DESTRUCTIVE_COMMAND_NOTES[meta.name] ? ` (${DESTRUCTIVE_COMMAND_NOTES[meta.name]})` : "";
   return `- /${meta.name}${args} \u2014 ${meta.description}${note}`;
 }
-var ANALYST_FILE_NAME, ANALYST_FILE_MAX_CHARS, EXECUTION_BIAS_BLOCK, ANALYST_INSTINCT_BLOCK, SAFETY_BLOCK, VITAL_SIGNS_BLOCK, PLAYBOOK_BLOCK, DESTRUCTIVE_COMMAND_NOTES, METRICS_BLOCK, GTM_ENGINEERING_BLOCK, PYRAMID_OUTPUT_BLOCK;
+var ANALYST_FILE_NAME, ANALYST_FILE_MAX_CHARS, EXECUTION_BIAS_BLOCK, ANALYST_INSTINCT_BLOCK, SAFETY_BLOCK, VITAL_SIGNS_BLOCK, PLAYBOOK_BLOCK, DESTRUCTIVE_COMMAND_NOTES, METRICS_BLOCK, GTM_ENGINEERING_BLOCK, HEILMEIER_GATE, PYRAMID_OUTPUT_BLOCK;
 var init_prompt_parts = __esm({
   "src/ai/prompt-parts.ts"() {
     "use strict";
@@ -37127,12 +37179,23 @@ Instrument trust: every metric here carries confidence and reliability_gate fiel
 - EVERY FIX GETS A MECHANISM. One-time cleanups decay in a quarter. Pair each cleanup with the mechanism that keeps it fixed: a routing rule, an SLA with a report behind it, an enrichment waterfall, a signal-triggered task, an alert in the channel reps already work in.
 - INSTRUMENT WHAT YOU CHANGE. A system you can't measure is a system you can't defend at the next QBR. Name the metric each mechanism should move and where it will be read.
 Restraint: you diagnose and prescribe the system; you do not build it here. Name the mechanism class, not a vendor shopping list.`;
+    HEILMEIER_GATE = `Before you recommend an action, play, or workstream \u2014 run this gate. Failures are silent: kill, demote to an assumption, or ask one sharper question. Never say Heilmeier, DARPA, or catechism to the user.
+- END STATE. Name the precise outcome in one line. Vague goals die here.
+- STATUS QUO. How is it done today, and what is the hard limit of that approach?
+- NEWNESS. What is new in this approach, and why might it work when the status quo fails? If nothing is new, kill it as a restatement.
+- STAKE. Who cares? If it works, what difference does it make \u2014 dollars recovered or a decision unlocked? Name the cost of doing nothing. No stake \u2192 do not recommend.
+- RISKS AND PAYOFFS. State the main downside and the upside in the same breath.
+- EFFORT. Rough cost in team hours or capacity. If it cannot fit the team that exists, cut or re-sequence \u2014 do not soft-pedal.
+- TIME. Bound by sales-cycle and metric-update physics. Dates faster than evidence can appear are invalid.
+- EXAMS. Name a mid-term check and a final exam: dated, measurable, with an instrument that exists. No exams \u2192 demote to assumption or drop.
+- KILL RULE. If newness, stake, or exams fail: do not recommend. Demote, drop, or ask one sharper question. Surviving ideas then get pyramid packaging.`;
     PYRAMID_OUTPUT_BLOCK = `Structure everything the way a client remembers it \u2014 pyramid, answer first:
-- HEADLINE FIRST. Open with the verdict and the number in one sentence (\u226415 words where possible): what is true and what it costs. Never open with methodology or context.
-- THEN THE DRIVERS. Support the headline with 2-3 distinct, non-overlapping drivers, each with its own number. If two points share a cause, merge them.
+- HEADLINE FIRST. Open with the verdict and the number in one sentence (\u226420 words): what is true and what it costs. Never open with methodology or context.
+- THEN THE DRIVERS. Support the headline with 2-3 drivers that are distinct, non-overlapping, and together explain the headline \u2014 each with its own number. If two points share a cause, merge them.
 - THEN THE SO-WHAT. Close with what it means for the decision at hand: the action, the owner-shaped next step, or the sharpest next cut.
 - THE RECALL TEST. A busy executive should be able to repeat your headline and one number to their CEO an hour later. If they couldn't, tighten it.
-- ALTITUDE CONTROL. Answer at the altitude asked: a high-level question gets the 30,000-ft story (one narrative sentence, three numbers max) with an offer to descend; a "how do we fix it" / plan-of-attack question prefers draft_strategy (or, if answering one play inline: one play + mechanism + what to verify \u2014 observe and recommend, never invent a multi-week Phase 1/2/3 program).`;
+- ALTITUDE CONTROL. Answer at the altitude asked: a high-level question gets the 30,000-ft story (one narrative sentence, three numbers max) with an offer to descend; a "how do we fix it" / plan-of-attack question prefers draft_strategy (or, if answering one play inline: one play + mechanism + what to verify \u2014 observe and recommend, never invent a multi-week Phase 1/2/3 program).
+- BRIEF MODE. When brevity is required: one headline + \u22642 drivers + one so-what. Still answer-first; never open with a scorecard.`;
   }
 });
 
@@ -37141,7 +37204,7 @@ function buildMetricsPlaybookBlock() {
   const plays = getMetricsPlays();
   return plays.map((p) => `- "${p.name}" (id: ${p.id}) \u2014 when ${p.trigger_condition}: ${p.why}`).join("\n");
 }
-function buildSystemPrompt2() {
+function buildMetricsFindingsSystemPrompt() {
   const profile = loadProfile();
   const profileBlock = buildCompanyProfileBlock();
   const motionLabel = motionBenchmarkLabel(profile?.sales_motion);
@@ -37220,7 +37283,7 @@ async function generateMetricsFindings(input, ctx) {
   const { response, meta } = await completeWithFailover(
     {
       surface: "metrics_findings",
-      system: buildSystemPrompt2(),
+      system: buildMetricsFindingsSystemPrompt(),
       messages: [{ role: "user", content: userMessage }],
       max_tokens: 4096
     },

package/dist/mcp/server.js

@@ -10730,7 +10730,7 @@ function formatCatalogLine(meta) {
   const note = DESTRUCTIVE_COMMAND_NOTES[meta.name] ? ` (${DESTRUCTIVE_COMMAND_NOTES[meta.name]})` : "";
   return `- /${meta.name}${args} \u2014 ${meta.description}${note}`;
 }
-var ANALYST_FILE_NAME, ANALYST_FILE_MAX_CHARS, EXECUTION_BIAS_BLOCK, ANALYST_INSTINCT_BLOCK, SAFETY_BLOCK, VITAL_SIGNS_BLOCK, PLAYBOOK_BLOCK, DESTRUCTIVE_COMMAND_NOTES, METRICS_BLOCK, GTM_ENGINEERING_BLOCK, PYRAMID_OUTPUT_BLOCK;
+var ANALYST_FILE_NAME, ANALYST_FILE_MAX_CHARS, EXECUTION_BIAS_BLOCK, ANALYST_INSTINCT_BLOCK, SAFETY_BLOCK, VITAL_SIGNS_BLOCK, PLAYBOOK_BLOCK, DESTRUCTIVE_COMMAND_NOTES, METRICS_BLOCK, GTM_ENGINEERING_BLOCK, HEILMEIER_GATE, PYRAMID_OUTPUT_BLOCK;
 var init_prompt_parts = __esm({
   "src/ai/prompt-parts.ts"() {
     "use strict";
@@ -10809,12 +10809,23 @@ Instrument trust: every metric here carries confidence and reliability_gate fiel
 - EVERY FIX GETS A MECHANISM. One-time cleanups decay in a quarter. Pair each cleanup with the mechanism that keeps it fixed: a routing rule, an SLA with a report behind it, an enrichment waterfall, a signal-triggered task, an alert in the channel reps already work in.
 - INSTRUMENT WHAT YOU CHANGE. A system you can't measure is a system you can't defend at the next QBR. Name the metric each mechanism should move and where it will be read.
 Restraint: you diagnose and prescribe the system; you do not build it here. Name the mechanism class, not a vendor shopping list.`;
+    HEILMEIER_GATE = `Before you recommend an action, play, or workstream \u2014 run this gate. Failures are silent: kill, demote to an assumption, or ask one sharper question. Never say Heilmeier, DARPA, or catechism to the user.
+- END STATE. Name the precise outcome in one line. Vague goals die here.
+- STATUS QUO. How is it done today, and what is the hard limit of that approach?
+- NEWNESS. What is new in this approach, and why might it work when the status quo fails? If nothing is new, kill it as a restatement.
+- STAKE. Who cares? If it works, what difference does it make \u2014 dollars recovered or a decision unlocked? Name the cost of doing nothing. No stake \u2192 do not recommend.
+- RISKS AND PAYOFFS. State the main downside and the upside in the same breath.
+- EFFORT. Rough cost in team hours or capacity. If it cannot fit the team that exists, cut or re-sequence \u2014 do not soft-pedal.
+- TIME. Bound by sales-cycle and metric-update physics. Dates faster than evidence can appear are invalid.
+- EXAMS. Name a mid-term check and a final exam: dated, measurable, with an instrument that exists. No exams \u2192 demote to assumption or drop.
+- KILL RULE. If newness, stake, or exams fail: do not recommend. Demote, drop, or ask one sharper question. Surviving ideas then get pyramid packaging.`;
     PYRAMID_OUTPUT_BLOCK = `Structure everything the way a client remembers it \u2014 pyramid, answer first:
-- HEADLINE FIRST. Open with the verdict and the number in one sentence (\u226415 words where possible): what is true and what it costs. Never open with methodology or context.
-- THEN THE DRIVERS. Support the headline with 2-3 distinct, non-overlapping drivers, each with its own number. If two points share a cause, merge them.
+- HEADLINE FIRST. Open with the verdict and the number in one sentence (\u226420 words): what is true and what it costs. Never open with methodology or context.
+- THEN THE DRIVERS. Support the headline with 2-3 drivers that are distinct, non-overlapping, and together explain the headline \u2014 each with its own number. If two points share a cause, merge them.
 - THEN THE SO-WHAT. Close with what it means for the decision at hand: the action, the owner-shaped next step, or the sharpest next cut.
 - THE RECALL TEST. A busy executive should be able to repeat your headline and one number to their CEO an hour later. If they couldn't, tighten it.
-- ALTITUDE CONTROL. Answer at the altitude asked: a high-level question gets the 30,000-ft story (one narrative sentence, three numbers max) with an offer to descend; a "how do we fix it" / plan-of-attack question prefers draft_strategy (or, if answering one play inline: one play + mechanism + what to verify \u2014 observe and recommend, never invent a multi-week Phase 1/2/3 program).`;
+- ALTITUDE CONTROL. Answer at the altitude asked: a high-level question gets the 30,000-ft story (one narrative sentence, three numbers max) with an offer to descend; a "how do we fix it" / plan-of-attack question prefers draft_strategy (or, if answering one play inline: one play + mechanism + what to verify \u2014 observe and recommend, never invent a multi-week Phase 1/2/3 program).
+- BRIEF MODE. When brevity is required: one headline + \u22642 drivers + one so-what. Still answer-first; never open with a scorecard.`;
   }
 });
 
@@ -10902,7 +10913,7 @@ function buildFindingsFindingHint(prefs) {
   const voice2 = prefs ?? loadVoicePrefs();
   const register = voice2.personality === "robotic" ? "STE-100 sentences. No contractions, slang, or filler." : voice2.personality === "composed" ? "Composed consultant sentences. Light contractions OK." : voice2.personality === "casual" ? "Casual consultant sentences. Warm, still short." : "Loose consultant sentences. Personality in the wrap, not the math.";
   const roast = voice2.roast === "light" ? "Light roast: what is working, then what to improve." : voice2.roast === "medium" ? "Medium roast: name the miss. No padding." : voice2.roast === "dark" ? "Dark roast: surgical. Verdict + dollar in one breath." : "Heavy roast: maximum roast after the number, never instead of it.";
-  return `Pyramid-shaped, 2-3 sentences: (1) HEADLINE \u2014 verdict + dollar figure in one short sentence (\u226420 words); (2) EVIDENCE \u2014 the one or two numbers that prove it; (3) SO-WHAT \u2014 the consequence or the action. ${register} ${roast} Keep dollar figures and play names. Do not invent numbers. An executive should be able to repeat sentence 1 from memory.`;
+  return `Pyramid-shaped, 2-3 sentences: (1) HEADLINE \u2014 verdict + dollar figure in one short sentence (\u226420 words); (2) DRIVERS \u2014 the one or two distinct, non-overlapping numbers that prove it; (3) SO-WHAT \u2014 the consequence or the action. ${register} ${roast} Keep dollar figures and play names. Do not invent numbers. An executive should be able to repeat sentence 1 from memory.`;
 }
 function buildFindingsSchemaBlock(prefs) {
   const finding = buildFindingsFindingHint(prefs);
@@ -14302,7 +14313,7 @@ function buildMetricsPlaybookBlock() {
   const plays = getMetricsPlays();
   return plays.map((p) => `- "${p.name}" (id: ${p.id}) \u2014 when ${p.trigger_condition}: ${p.why}`).join("\n");
 }
-function buildSystemPrompt() {
+function buildMetricsFindingsSystemPrompt() {
   const profile = loadProfile();
   const profileBlock = buildCompanyProfileBlock();
   const motionLabel = motionBenchmarkLabel(profile?.sales_motion);
@@ -14381,7 +14392,7 @@ async function generateMetricsFindings(input, ctx) {
   const { response, meta } = await completeWithFailover(
     {
       surface: "metrics_findings",
-      system: buildSystemPrompt(),
+      system: buildMetricsFindingsSystemPrompt(),
       messages: [{ role: "user", content: userMessage }],
       max_tokens: 4096
     },
@@ -17001,10 +17012,10 @@ function wrapForTarget(target, analysisBlock, conversationBlock, openQuestions,
   const contextLabel = handoffInstructionPrefix(ctx.analysis.primary);
   const forWhom = audiencePhrase(ctx.scope?.audience);
   const instructions = {
-    deck: `produce an executive review deck outline for ${company}, framed for ${forWhom}. Structure: (1) Headline number; (2) Pipeline health; (3) Top 3 risks with dollar impact; (4) Recommended plays; (5) Asks / decisions. Use the Metric definitions appendix so every slide can briefly remind the room what the number means for this audience.`,
+    deck: `produce an executive review deck outline for ${company}, framed for ${forWhom}. Pyramid structure: (1) HEADLINE \u2014 the governing number and verdict in one line; (2) DRIVERS \u2014 2\u20133 distinct, non-overlapping risks or causes with dollar impact (not a full scorecard); (3) SO-WHAT \u2014 recommended plays; (4) ASKS / decisions the room must make. Optionally one pipeline-health slide after the headline if it supports the call. Use the Metric definitions appendix so every slide can briefly remind the room what the number means for this audience.`,
     asana: `produce an Asana project plan with sections and tasks tied to findings for ${forWhom}. Prioritize by dollar impact. Reference metric definitions when a task owner needs to know what "good" looks like.`,
     clay: `produce a Clay table specification to operationalize the highest-impact finding for ${forWhom}.`,
-    plan: `produce a prioritized action plan for ${forWhom} with problem, play, first 3 steps, owner, and leading indicator per item. Ground indicators in the Metric definitions appendix.`
+    plan: `produce a prioritized action plan for ${forWhom}, answer-first: open with the governing call (highest-dollar problem), then items each with problem, play, first 3 steps, owner, and leading indicator. Ground indicators in the Metric definitions appendix. End with the sharpest asks / decisions.`
   };
   const parts = [
     `# NTRP handoff \u2192 ${target}`,
@@ -17399,10 +17410,13 @@ ${companyContextSection2()}${operatorSection2()}HOW YOU THINK (the strategist me
 8. RESPECT CAPACITY. Total effort must fit the team that actually exists. A brilliant plan the team cannot staff is a bad plan.
 
 ALTITUDE CONTRACT (the plan must work at every altitude a client reads it at):
-- 30,000 FT: summary_30k is a situation-complication-resolution narrative \u2014 where the business stands, what gates what, and what leadership should expect by when. A board member reads only this; it must survive being forwarded unedited.
+- 30,000 FT: summary_30k is answer-first \u2014 open with the governing thought and what leadership should expect by when, then the situation and complication that make it true. A board member reads only this; it must survive being forwarded unedited.
 - 10,000 FT: each workstream's title + problem line is a delegation unit \u2014 a one-liner an owner could receive in Slack and know what they own, why it's theirs, and what number they move.
 - GROUND LEVEL: actions are Monday-morning prescriptive. You have built this before \u2014 read the linked play's full detail with get_play_detail and prescribe its known-good sequence adapted to THIS company's numbers and constraints. The first action of every workstream must be startable within 48 hours with no new tooling.
 
+OUTPUT DOCTRINE (narrative fields \u2014 summary_30k and workstream problem lines \u2014 follow pyramid recall):
+${PYRAMID_OUTPUT_BLOCK}
+
 ${buildUserVisibleProseBlock()}
 
 GTM ENGINEERING (plans install systems, not heroics):
@@ -17485,6 +17499,8 @@ function buildStressTestMessage() {
 6. EXPECTATION HONESTY: are target ranges defensible from the baseline and effort, or heroic? Widen ranges or lower confidence rather than promising what the data does not support. Where the local play track record shows a play has hit or missed here before, weight confidence accordingly.
 7. CONTINGENCY QUALITY: is each trigger observable on a specific date, and is each fallback a real pre-decision (alternate play, descope, escalate) rather than "monitor closely"?
 8. ALTITUDE CHECK: does summary_30k survive being forwarded to a board member unedited? Is each workstream title + problem a self-contained delegation one-liner? Is every first action startable within 48 hours?
+9. HEILMEIER GATE (silent \u2014 never name this to the user). Run every workstream through this filter. Failures: cut the workstream, demote it to assumptions, or rewrite until it passes \u2014 then emit final JSON.
+${HEILMEIER_GATE}
 
 Then respond with the FINAL revised plan as strict JSON in the same schema \u2014 no markdown fences, no prose. Fold what you learned into constraints, assumptions, risks, and confidence. This version is the one that gets saved and reviewed against, so make every number one you are willing to be checked on.`;
 }
@@ -25665,6 +25681,7 @@ ${opts.conversationBlock}
   const socraticCraft = `HOW YOU THINK WITH THE USER (socratic partner \u2014 not a search box, not a strategist):
 - Pre-explore funnel gates are handled by the CLI; you speak in the think channel after analysis.
 - You are a thinking partner. Prefer sharp questions that unlock the user's idea over dumping answers.
+- Probing turns: lead with the sharpest question that unlocks the next insight. Concluding turns: use pyramid (headline \u2192 drivers \u2192 so-what) \u2014 answer first, then structure.
 - Protect imagination: give "what if" space before the smell-test \u2014 do not kill creativity in the first sentence.
 - Steelman the user's position before you attack it; then deliver one hard pushback grounded in evidence when possible.
 - Ground claims about THIS pipeline in tool results. Label speculation explicitly ("hypothesis:", "speculation:").
@@ -25675,7 +25692,7 @@ ${opts.conversationBlock}
 - You have continuity via prior think-channel messages. Never repeat an angle already covered unless asked.`;
   const jobSection = `YOUR JOB (THINK CHANNEL \u2014 always deep):
 - Decide whether you need tools, a direct answer, or both. Do not re-call a tool whose result you already have.
-- Lead with the answer or the question that matters most, then structure.
+- Probing: lead with the question that matters most. Concluding: lead with the answer (headline), then drivers and so-what.
 - When you use numbers, include dollar values where available and lead with financial impact.
 - Descriptive exploration stays in this channel; plan-of-attack questions hand off via draft_strategy.
 - After a successful draft_strategy tool result: reply with at most 1\u20132 sentences introducing the handoff. Do not write the plan.`;
@@ -25713,6 +25730,9 @@ ${EXECUTION_BIAS_BLOCK}
 GTM ENGINEERING (how recommendations become systems):
 ${GTM_ENGINEERING_BLOCK}
 
+IDEA GATE (when recommending a play or before draft_strategy \u2014 silent; probing/imagination turns skip this until you conclude):
+${HEILMEIER_GATE}
+
 OUTPUT DOCTRINE (how answers are structured for recall):
 ${PYRAMID_OUTPUT_BLOCK}
 
@@ -25779,7 +25799,7 @@ function operatorSection4() {
 
 `;
 }
-function buildSystemPrompt2() {
+function buildSystemPrompt() {
   const stable = `You are an expert GTM health investigator. You have tools to query a local database of CRM and pipeline data. Your job is to investigate the health scores you've been given and discover the specific root causes behind any problems.
 
 ${companyContextSection4()}${operatorSection4()}INVESTIGATION APPROACH:
@@ -25949,12 +25969,18 @@ ${commandStyleRule}`;
 - Respond in plain text (not JSON). You do NOT need to emit the findings schema \u2014 only emit structured findings if the user explicitly asks for them.
 - Include specific numbers from tool results, never guess.
 - When you have enough information, answer directly. Don't call tools you don't need.`;
-  const executionBiasSection = responseMode === "brief" ? "" : `EXECUTION BIAS (how you work):
+  const executionBiasSection = responseMode === "brief" ? `OUTPUT DOCTRINE (brief \u2014 one headline + \u22642 drivers + one so-what):
+${PYRAMID_OUTPUT_BLOCK}
+
+` : `EXECUTION BIAS (how you work):
 ${EXECUTION_BIAS_BLOCK}
 
 GTM ENGINEERING (how your recommendations become systems):
 ${GTM_ENGINEERING_BLOCK}
 
+IDEA GATE (when recommending action \u2014 silent):
+${HEILMEIER_GATE}
+
 OUTPUT DOCTRINE (how answers are structured for recall):
 ${PYRAMID_OUTPUT_BLOCK}
 
@@ -26049,7 +26075,7 @@ async function* agenticFindings(computeResult, divergences, options) {
         responseMode
       })
     }
-  ) : buildSystemPrompt2();
+  ) : buildSystemPrompt();
   const analysisReady = isAnalysisReady(options.ctx);
   const tools2 = useTools ? mode === "think" ? buildThinkTools() : mode === "fresh" ? buildFreshNlTools({ analysisReady }) : buildInvestigationTools() : [];
   const toolCtx = { computeResult, divergences };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@sonnechasser/ntrp",
-  "version": "1.4.6",
+  "version": "1.4.7",
   "description": "GTM Health Diagnostic CLI — local pipeline analysis tool",
   "homepage": "https://ntrp.sonnechasser.com",
   "repository": {