@askexenow/exe-os 0.9.112 → 0.9.113

package/README.md

@@ -1,6 +1,6 @@
 # Exe OS
 
-![npm version](https://img.shields.io/npm/v/@askexenow/exe-os.svg) ![License: CC-BY-NC-4.0](https://img.shields.io/badge/license-CC--BY--NC--4.0-blue.svg) ![Node.js](https://img.shields.io/badge/node-%3E%3D20.0.0-brightgreen.svg)
+![npm version](https://img.shields.io/npm/v/@askexenow/exe-os.svg) ![License: AGPL-3.0](https://img.shields.io/badge/license-AGPL--3.0-blue.svg) ![Node.js](https://img.shields.io/badge/node-%3E%3D20.0.0-brightgreen.svg)
 
 **Hire the team you couldn't afford.** AI employee operating system with persistent memory, identity, and multi-agent orchestration.
 
@@ -51,9 +51,10 @@ Best for: managing multiple projects and employees.
 | Feature | What it does |
 |---------|-------------|
 | **Persistent memory** | Every interaction stored in encrypted SQLCipher + vector search. Searchable across sessions. |
-| **Three-layer cognition** | Identity (who you are) + Expertise (what you've learned) + Experience (what you remember) |
+| **Four-layer cognition** | Company Procedures (org-wide rules) + Identity (who you are) + Expertise (what you've learned) + Experience (what you remember) |
 | **Multi-agent orchestration** | COO coordinates CTO, CMO, engineers. Parallel task execution via tmux. |
 | **Task system** | Create, assign, review, chain tasks. Auto-dispatch. Review pipeline with cascading approval. |
+| **Company procedures** | Org-wide rules that every employee follows. Set once, cascades everywhere. |
 | **Identity injection** | Each employee gets a permanent identity doc. Claude Code's `--agent` flag replaces the default system prompt. |
 | **Skill learning** | Agents learn procedures from repeated patterns. Corrections become permanent behaviors. |
 | **Cloud sync** | End-to-end encrypted memory sync across devices. Your key, your data. |
@@ -80,6 +81,7 @@ You talk to your COO. Your COO delegates. Each employee has:
 - **Identity** — permanent role doc that defines who they are
 - **Task queue** — auto-chains through assigned work
 - **Behavioral expertise** — corrections accumulate as permanent rules
+- **Company procedures** — org-wide rules that cascade to every employee
 
 ---
 
@@ -100,7 +102,7 @@ npm install -g @askexenow/exe-os
 # Run `wsl --install` in PowerShell first, then follow Linux steps inside WSL2
 ```
 
-Requires Node.js 22+ and tmux. See **[docs/install.md](docs/install.md)** for full platform-specific instructions and troubleshooting.
+Requires Node.js 22+ and tmux. See **[docs/install.md](docs/install.md)** for full platform-specific instructions and **[docs/quickstart.md](docs/quickstart.md)** for the 5-minute getting started guide.
 
 ### 2. Run Setup
 
@@ -136,8 +138,8 @@ exe-os                     # launch the TUI dashboard
 
 | Plan | Price | Employees | Memories | Devices |
 |------|-------|-----------|----------|---------|
-| Free | $0 | 1 (COO) | 5,000 | 1 |
-| Solopreneur | $97/mo | 5 | 100,000 | 2 |
+| Free | $0 | 1 (COO) | 50,000 | 1 |
+| Solopreneur | $97/mo | 5 | 250,000 | 2 |
 | Company | $297/mo | 20 | 1,000,000 | 10 |
 | Agency | $497/mo | 100 | 10,000,000 | 50 |
 | Enterprise | Custom | Unlimited | Unlimited | Unlimited |
@@ -202,7 +204,7 @@ See [docs/Agency-Playbook.md](docs/Agency-Playbook.md) for multi-client VPS depl
 exe-os/
   src/
     lib/          # Core: memory, tasks, identity, behaviors, search, sync, encryption
-    mcp/          # MCP server (20+ tools for Claude Code integration)
+    mcp/          # MCP server (100+ tools for Claude Code integration)
     adapters/     # Claude Code hooks (ingest, session-start, pre-tool-use, etc.)
     tui/          # Ink-based terminal dashboard (Mode 2)
     runtime/      # Agent loop, tool registry, permissions (Mode 2 standalone)
@@ -244,4 +246,4 @@ Full guide: **[docs/install.md](docs/install.md)**
 
 ## License
 
-CC-BY-NC-4.0 (non-commercial). Commercial licenses available at [askexe.com](https://askexe.com).
+AGPL-3.0. Commercial licenses available at [askexe.com](https://askexe.com).

package/dist/bin/agentic-ontology-backfill.js

@@ -3034,6 +3034,22 @@ async function ensureSchema() {
   } catch (e) {
     logCatchDebug("migration", e);
   }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN visibility TEXT DEFAULT 'private'`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN strength REAL DEFAULT 1.0`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
 }
 async function disposeDatabase() {
   if (_walCheckpointTimer) {
@@ -3600,11 +3616,17 @@ var init_platform_procedures = __esm({
         content: "Founder -> coordinator (the executive agent, internally routed as 'COO') -> CTO/CMO. CTO -> engineers. CMO -> content production. Never skip levels: the coordinator does not bypass managers for specialist work. Specialists report to their manager. If you need cross-team info, use ask_team_memory \u2014 don't read other agents' task folders. Each level owns dispatch downward and review upward."
       },
       {
-        title: "Customer orchestration maturity \u2014 recommend, never trap",
+        title: "Orchestration phase guidance \u2014 recommend, never trap",
         domain: "workflow",
         priority: "p1",
         content: "New customers start best in Phase 1: founder \u2194 coordinator/Chief of Staff, building company context. Suggest Phase 2 executives when domain work repeats; suggest Phase 3 parallel execution only when review/permission gates are ready. This is guidance, not a blocker: users may jump phases anytime. Never overwrite their phase, role titles, identities, or custom org design."
       },
+      {
+        title: "Routing slot vs display title \u2014 internal 'coo' is plumbing, not your name",
+        domain: "identity",
+        priority: "p0",
+        content: "These procedures reference 'COO' as a shorthand for the coordinator role. This is an INTERNAL routing slot used by exe-os code (chain-of-command checks, dispatch logic, session detection). It is NOT your display title. Your actual title comes from your identity file's `title:` field \u2014 that is what you use externally: introductions, sign-offs, team comms, and any user-facing text. If your identity says `title: AI Chief of Staff`, you are the AI Chief of Staff. The routing slot stays `role: coo` for code compatibility \u2014 never rename it, but also never introduce yourself as 'COO' unless your identity file explicitly says so. The founder chose your title; respect it."
+      },
       {
         title: "Single dispatch path \u2014 create_task only",
         domain: "workflow",
@@ -3638,6 +3660,12 @@ var init_platform_procedures = __esm({
         priority: "p0",
         content: "NEVER: (1) Access the database directly \u2014 it's SQLCipher encrypted, always fails. Use MCP tools only. (2) Manually spawn tmux sessions \u2014 create_task handles it. (3) Run git checkout main \u2014 agents work in worktrees. (4) Modify another agent's in-progress task. (5) Push to remote \u2014 the COO reviews and pushes. (6) Skip update_task(done) \u2014 it's the ONLY way your work gets reviewed. (7) Run git init."
       },
+      {
+        title: "Destructive operations \u2014 mandatory reviewer gate",
+        domain: "security",
+        priority: "p0",
+        content: "Before ANY destructive operation (delete, remove, overwrite, drop, reset, force-push, truncate), you MUST: (1) Have your full task spec accessible \u2014 if you cannot read it, STOP and report to your reviewer. Never improvise destructive actions. (2) Confirm with your reviewer (assigned_by or COO) before executing. (3) If the task spec explicitly authorizes the operation, proceed \u2014 but log it. Violation = immediate task failure. This applies to ALL agents regardless of role."
+      },
       {
         title: "Customer patch triage \u2014 upstream bug vs customization",
         domain: "support",
@@ -3923,10 +3951,24 @@ function stableId(memoryId, type, content) {
   return createHash2("sha256").update(`${memoryId}:${type}:${content}`).digest("hex").slice(0, 32);
 }
 function cleanText(text) {
-  return text.replace(/```[\s\S]*?```/g, " ").replace(/<[^>]+>/g, " ").replace(/\s+/g, " ").trim();
-}
-function splitSentences(text) {
-  return cleanText(text).split(/(?<=[.!?])\s+|\n+/).map((s) => s.trim()).filter((s) => s.length >= 24 && s.length <= MAX_SENTENCE_CHARS);
+  let cleaned = text.replace(
+    /```(\w*)\n(.*?)(?:\n[\s\S]*?)```/g,
+    (_m, lang, firstLine) => `[code${lang ? `:${lang}` : ""}] ${firstLine.trim()}`
+  );
+  cleaned = cleaned.replace(/<[^>]+>/g, " ").replace(/\s+/g, " ").trim();
+  return cleaned;
+}
+function splitSegments(text) {
+  const cleaned = cleanText(text);
+  const segments = cleaned.split(/(?<=[.!?:;])\s+|\n{2,}|(?<=\))\s+(?=[A-Z])|\s*[|│]\s*/).map((s) => s.trim()).filter((s) => s.length >= MIN_SEGMENT_CHARS && s.length <= MAX_SEGMENT_CHARS);
+  if (segments.length === 0 && cleaned.length >= MIN_SEGMENT_CHARS) {
+    const lines = cleaned.split(/\n+/).map((l) => l.trim()).filter((l) => l.length >= MIN_SEGMENT_CHARS && l.length <= MAX_SEGMENT_CHARS);
+    if (lines.length > 0) return lines;
+    if (cleaned.length >= MIN_SEGMENT_CHARS) {
+      return [cleaned.slice(0, MAX_SEGMENT_CHARS)];
+    }
+  }
+  return segments;
 }
 function inferCardType(sentence, toolName) {
   const lower = sentence.toLowerCase();
@@ -3958,12 +4000,12 @@ function predicateFor(type) {
   }
 }
 function extractMemoryCards(row) {
-  const sentences = splitSentences(row.raw_text);
+  const segments = splitSegments(row.raw_text);
   const cards = [];
-  for (const sentence of sentences) {
+  for (const sentence of segments) {
     const type = inferCardType(sentence, row.tool_name);
     const subject = extractSubject(sentence, row.agent_id);
-    const content = sentence.length > MAX_SENTENCE_CHARS ? `${sentence.slice(0, MAX_SENTENCE_CHARS - 1)}\u2026` : sentence;
+    const content = sentence.length > MAX_SEGMENT_CHARS ? `${sentence.slice(0, MAX_SEGMENT_CHARS - 1)}\u2026` : sentence;
     cards.push({
       id: stableId(row.id, type, content),
       memory_id: row.id,
@@ -4059,13 +4101,14 @@ Source memory: ${String(row.source_ref ?? row.memory_id)}`,
     last_accessed: String(row.timestamp)
   }));
 }
-var MAX_CARDS_PER_MEMORY, MAX_SENTENCE_CHARS;
+var MAX_CARDS_PER_MEMORY, MAX_SEGMENT_CHARS, MIN_SEGMENT_CHARS;
 var init_memory_cards = __esm({
   "src/lib/memory-cards.ts"() {
     "use strict";
     init_database();
-    MAX_CARDS_PER_MEMORY = 6;
-    MAX_SENTENCE_CHARS = 360;
+    MAX_CARDS_PER_MEMORY = 8;
+    MAX_SEGMENT_CHARS = 500;
+    MIN_SEGMENT_CHARS = 20;
   }
 });
 

package/dist/bin/agentic-reflection-backfill.js

@@ -3034,6 +3034,22 @@ async function ensureSchema() {
   } catch (e) {
     logCatchDebug("migration", e);
   }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN visibility TEXT DEFAULT 'private'`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN strength REAL DEFAULT 1.0`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
 }
 async function disposeDatabase() {
   if (_walCheckpointTimer) {
@@ -3600,11 +3616,17 @@ var init_platform_procedures = __esm({
         content: "Founder -> coordinator (the executive agent, internally routed as 'COO') -> CTO/CMO. CTO -> engineers. CMO -> content production. Never skip levels: the coordinator does not bypass managers for specialist work. Specialists report to their manager. If you need cross-team info, use ask_team_memory \u2014 don't read other agents' task folders. Each level owns dispatch downward and review upward."
       },
       {
-        title: "Customer orchestration maturity \u2014 recommend, never trap",
+        title: "Orchestration phase guidance \u2014 recommend, never trap",
         domain: "workflow",
         priority: "p1",
         content: "New customers start best in Phase 1: founder \u2194 coordinator/Chief of Staff, building company context. Suggest Phase 2 executives when domain work repeats; suggest Phase 3 parallel execution only when review/permission gates are ready. This is guidance, not a blocker: users may jump phases anytime. Never overwrite their phase, role titles, identities, or custom org design."
       },
+      {
+        title: "Routing slot vs display title \u2014 internal 'coo' is plumbing, not your name",
+        domain: "identity",
+        priority: "p0",
+        content: "These procedures reference 'COO' as a shorthand for the coordinator role. This is an INTERNAL routing slot used by exe-os code (chain-of-command checks, dispatch logic, session detection). It is NOT your display title. Your actual title comes from your identity file's `title:` field \u2014 that is what you use externally: introductions, sign-offs, team comms, and any user-facing text. If your identity says `title: AI Chief of Staff`, you are the AI Chief of Staff. The routing slot stays `role: coo` for code compatibility \u2014 never rename it, but also never introduce yourself as 'COO' unless your identity file explicitly says so. The founder chose your title; respect it."
+      },
       {
         title: "Single dispatch path \u2014 create_task only",
         domain: "workflow",
@@ -3638,6 +3660,12 @@ var init_platform_procedures = __esm({
         priority: "p0",
         content: "NEVER: (1) Access the database directly \u2014 it's SQLCipher encrypted, always fails. Use MCP tools only. (2) Manually spawn tmux sessions \u2014 create_task handles it. (3) Run git checkout main \u2014 agents work in worktrees. (4) Modify another agent's in-progress task. (5) Push to remote \u2014 the COO reviews and pushes. (6) Skip update_task(done) \u2014 it's the ONLY way your work gets reviewed. (7) Run git init."
       },
+      {
+        title: "Destructive operations \u2014 mandatory reviewer gate",
+        domain: "security",
+        priority: "p0",
+        content: "Before ANY destructive operation (delete, remove, overwrite, drop, reset, force-push, truncate), you MUST: (1) Have your full task spec accessible \u2014 if you cannot read it, STOP and report to your reviewer. Never improvise destructive actions. (2) Confirm with your reviewer (assigned_by or COO) before executing. (3) If the task spec explicitly authorizes the operation, proceed \u2014 but log it. Violation = immediate task failure. This applies to ALL agents regardless of role."
+      },
       {
         title: "Customer patch triage \u2014 upstream bug vs customization",
         domain: "support",

package/dist/bin/agentic-semantic-label.js

@@ -3034,6 +3034,22 @@ async function ensureSchema() {
   } catch (e) {
     logCatchDebug("migration", e);
   }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN visibility TEXT DEFAULT 'private'`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN strength REAL DEFAULT 1.0`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
 }
 async function disposeDatabase() {
   if (_walCheckpointTimer) {
@@ -3600,11 +3616,17 @@ var init_platform_procedures = __esm({
         content: "Founder -> coordinator (the executive agent, internally routed as 'COO') -> CTO/CMO. CTO -> engineers. CMO -> content production. Never skip levels: the coordinator does not bypass managers for specialist work. Specialists report to their manager. If you need cross-team info, use ask_team_memory \u2014 don't read other agents' task folders. Each level owns dispatch downward and review upward."
       },
       {
-        title: "Customer orchestration maturity \u2014 recommend, never trap",
+        title: "Orchestration phase guidance \u2014 recommend, never trap",
         domain: "workflow",
         priority: "p1",
         content: "New customers start best in Phase 1: founder \u2194 coordinator/Chief of Staff, building company context. Suggest Phase 2 executives when domain work repeats; suggest Phase 3 parallel execution only when review/permission gates are ready. This is guidance, not a blocker: users may jump phases anytime. Never overwrite their phase, role titles, identities, or custom org design."
       },
+      {
+        title: "Routing slot vs display title \u2014 internal 'coo' is plumbing, not your name",
+        domain: "identity",
+        priority: "p0",
+        content: "These procedures reference 'COO' as a shorthand for the coordinator role. This is an INTERNAL routing slot used by exe-os code (chain-of-command checks, dispatch logic, session detection). It is NOT your display title. Your actual title comes from your identity file's `title:` field \u2014 that is what you use externally: introductions, sign-offs, team comms, and any user-facing text. If your identity says `title: AI Chief of Staff`, you are the AI Chief of Staff. The routing slot stays `role: coo` for code compatibility \u2014 never rename it, but also never introduce yourself as 'COO' unless your identity file explicitly says so. The founder chose your title; respect it."
+      },
       {
         title: "Single dispatch path \u2014 create_task only",
         domain: "workflow",
@@ -3638,6 +3660,12 @@ var init_platform_procedures = __esm({
         priority: "p0",
         content: "NEVER: (1) Access the database directly \u2014 it's SQLCipher encrypted, always fails. Use MCP tools only. (2) Manually spawn tmux sessions \u2014 create_task handles it. (3) Run git checkout main \u2014 agents work in worktrees. (4) Modify another agent's in-progress task. (5) Push to remote \u2014 the COO reviews and pushes. (6) Skip update_task(done) \u2014 it's the ONLY way your work gets reviewed. (7) Run git init."
       },
+      {
+        title: "Destructive operations \u2014 mandatory reviewer gate",
+        domain: "security",
+        priority: "p0",
+        content: "Before ANY destructive operation (delete, remove, overwrite, drop, reset, force-push, truncate), you MUST: (1) Have your full task spec accessible \u2014 if you cannot read it, STOP and report to your reviewer. Never improvise destructive actions. (2) Confirm with your reviewer (assigned_by or COO) before executing. (3) If the task spec explicitly authorizes the operation, proceed \u2014 but log it. Violation = immediate task failure. This applies to ALL agents regardless of role."
+      },
       {
         title: "Customer patch triage \u2014 upstream bug vs customization",
         domain: "support",

package/dist/bin/backfill-conversations.js

@@ -3174,6 +3174,22 @@ async function ensureSchema() {
   } catch (e) {
     logCatchDebug("migration", e);
   }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN visibility TEXT DEFAULT 'private'`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN strength REAL DEFAULT 1.0`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
 }
 async function disposeDatabase() {
   if (_walCheckpointTimer) {
@@ -3740,11 +3756,17 @@ var init_platform_procedures = __esm({
         content: "Founder -> coordinator (the executive agent, internally routed as 'COO') -> CTO/CMO. CTO -> engineers. CMO -> content production. Never skip levels: the coordinator does not bypass managers for specialist work. Specialists report to their manager. If you need cross-team info, use ask_team_memory \u2014 don't read other agents' task folders. Each level owns dispatch downward and review upward."
       },
       {
-        title: "Customer orchestration maturity \u2014 recommend, never trap",
+        title: "Orchestration phase guidance \u2014 recommend, never trap",
         domain: "workflow",
         priority: "p1",
         content: "New customers start best in Phase 1: founder \u2194 coordinator/Chief of Staff, building company context. Suggest Phase 2 executives when domain work repeats; suggest Phase 3 parallel execution only when review/permission gates are ready. This is guidance, not a blocker: users may jump phases anytime. Never overwrite their phase, role titles, identities, or custom org design."
       },
+      {
+        title: "Routing slot vs display title \u2014 internal 'coo' is plumbing, not your name",
+        domain: "identity",
+        priority: "p0",
+        content: "These procedures reference 'COO' as a shorthand for the coordinator role. This is an INTERNAL routing slot used by exe-os code (chain-of-command checks, dispatch logic, session detection). It is NOT your display title. Your actual title comes from your identity file's `title:` field \u2014 that is what you use externally: introductions, sign-offs, team comms, and any user-facing text. If your identity says `title: AI Chief of Staff`, you are the AI Chief of Staff. The routing slot stays `role: coo` for code compatibility \u2014 never rename it, but also never introduce yourself as 'COO' unless your identity file explicitly says so. The founder chose your title; respect it."
+      },
       {
         title: "Single dispatch path \u2014 create_task only",
         domain: "workflow",
@@ -3778,6 +3800,12 @@ var init_platform_procedures = __esm({
         priority: "p0",
         content: "NEVER: (1) Access the database directly \u2014 it's SQLCipher encrypted, always fails. Use MCP tools only. (2) Manually spawn tmux sessions \u2014 create_task handles it. (3) Run git checkout main \u2014 agents work in worktrees. (4) Modify another agent's in-progress task. (5) Push to remote \u2014 the COO reviews and pushes. (6) Skip update_task(done) \u2014 it's the ONLY way your work gets reviewed. (7) Run git init."
       },
+      {
+        title: "Destructive operations \u2014 mandatory reviewer gate",
+        domain: "security",
+        priority: "p0",
+        content: "Before ANY destructive operation (delete, remove, overwrite, drop, reset, force-push, truncate), you MUST: (1) Have your full task spec accessible \u2014 if you cannot read it, STOP and report to your reviewer. Never improvise destructive actions. (2) Confirm with your reviewer (assigned_by or COO) before executing. (3) If the task spec explicitly authorizes the operation, proceed \u2014 but log it. Violation = immediate task failure. This applies to ALL agents regardless of role."
+      },
       {
         title: "Customer patch triage \u2014 upstream bug vs customization",
         domain: "support",
@@ -4063,10 +4091,24 @@ function stableId(memoryId, type, content) {
   return createHash2("sha256").update(`${memoryId}:${type}:${content}`).digest("hex").slice(0, 32);
 }
 function cleanText(text) {
-  return text.replace(/```[\s\S]*?```/g, " ").replace(/<[^>]+>/g, " ").replace(/\s+/g, " ").trim();
+  let cleaned = text.replace(
+    /```(\w*)\n(.*?)(?:\n[\s\S]*?)```/g,
+    (_m, lang, firstLine) => `[code${lang ? `:${lang}` : ""}] ${firstLine.trim()}`
+  );
+  cleaned = cleaned.replace(/<[^>]+>/g, " ").replace(/\s+/g, " ").trim();
+  return cleaned;
 }
-function splitSentences(text) {
-  return cleanText(text).split(/(?<=[.!?])\s+|\n+/).map((s) => s.trim()).filter((s) => s.length >= 24 && s.length <= MAX_SENTENCE_CHARS);
+function splitSegments(text) {
+  const cleaned = cleanText(text);
+  const segments = cleaned.split(/(?<=[.!?:;])\s+|\n{2,}|(?<=\))\s+(?=[A-Z])|\s*[|│]\s*/).map((s) => s.trim()).filter((s) => s.length >= MIN_SEGMENT_CHARS && s.length <= MAX_SEGMENT_CHARS);
+  if (segments.length === 0 && cleaned.length >= MIN_SEGMENT_CHARS) {
+    const lines = cleaned.split(/\n+/).map((l) => l.trim()).filter((l) => l.length >= MIN_SEGMENT_CHARS && l.length <= MAX_SEGMENT_CHARS);
+    if (lines.length > 0) return lines;
+    if (cleaned.length >= MIN_SEGMENT_CHARS) {
+      return [cleaned.slice(0, MAX_SEGMENT_CHARS)];
+    }
+  }
+  return segments;
 }
 function inferCardType(sentence, toolName) {
   const lower = sentence.toLowerCase();
@@ -4098,12 +4140,12 @@ function predicateFor(type) {
   }
 }
 function extractMemoryCards(row) {
-  const sentences = splitSentences(row.raw_text);
+  const segments = splitSegments(row.raw_text);
   const cards = [];
-  for (const sentence of sentences) {
+  for (const sentence of segments) {
     const type = inferCardType(sentence, row.tool_name);
     const subject = extractSubject(sentence, row.agent_id);
-    const content = sentence.length > MAX_SENTENCE_CHARS ? `${sentence.slice(0, MAX_SENTENCE_CHARS - 1)}\u2026` : sentence;
+    const content = sentence.length > MAX_SEGMENT_CHARS ? `${sentence.slice(0, MAX_SEGMENT_CHARS - 1)}\u2026` : sentence;
     cards.push({
       id: stableId(row.id, type, content),
       memory_id: row.id,
@@ -4199,13 +4241,14 @@ Source memory: ${String(row.source_ref ?? row.memory_id)}`,
     last_accessed: String(row.timestamp)
   }));
 }
-var MAX_CARDS_PER_MEMORY, MAX_SENTENCE_CHARS;
+var MAX_CARDS_PER_MEMORY, MAX_SEGMENT_CHARS, MIN_SEGMENT_CHARS;
 var init_memory_cards = __esm({
   "src/lib/memory-cards.ts"() {
     "use strict";
     init_database();
-    MAX_CARDS_PER_MEMORY = 6;
-    MAX_SENTENCE_CHARS = 360;
+    MAX_CARDS_PER_MEMORY = 8;
+    MAX_SEGMENT_CHARS = 500;
+    MIN_SEGMENT_CHARS = 20;
   }
 });
 

package/dist/bin/backfill-responses.js

@@ -3174,6 +3174,22 @@ async function ensureSchema() {
   } catch (e) {
     logCatchDebug("migration", e);
   }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN visibility TEXT DEFAULT 'private'`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN strength REAL DEFAULT 1.0`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
 }
 async function disposeDatabase() {
   if (_walCheckpointTimer) {
@@ -3740,11 +3756,17 @@ var init_platform_procedures = __esm({
         content: "Founder -> coordinator (the executive agent, internally routed as 'COO') -> CTO/CMO. CTO -> engineers. CMO -> content production. Never skip levels: the coordinator does not bypass managers for specialist work. Specialists report to their manager. If you need cross-team info, use ask_team_memory \u2014 don't read other agents' task folders. Each level owns dispatch downward and review upward."
       },
       {
-        title: "Customer orchestration maturity \u2014 recommend, never trap",
+        title: "Orchestration phase guidance \u2014 recommend, never trap",
         domain: "workflow",
         priority: "p1",
         content: "New customers start best in Phase 1: founder \u2194 coordinator/Chief of Staff, building company context. Suggest Phase 2 executives when domain work repeats; suggest Phase 3 parallel execution only when review/permission gates are ready. This is guidance, not a blocker: users may jump phases anytime. Never overwrite their phase, role titles, identities, or custom org design."
       },
+      {
+        title: "Routing slot vs display title \u2014 internal 'coo' is plumbing, not your name",
+        domain: "identity",
+        priority: "p0",
+        content: "These procedures reference 'COO' as a shorthand for the coordinator role. This is an INTERNAL routing slot used by exe-os code (chain-of-command checks, dispatch logic, session detection). It is NOT your display title. Your actual title comes from your identity file's `title:` field \u2014 that is what you use externally: introductions, sign-offs, team comms, and any user-facing text. If your identity says `title: AI Chief of Staff`, you are the AI Chief of Staff. The routing slot stays `role: coo` for code compatibility \u2014 never rename it, but also never introduce yourself as 'COO' unless your identity file explicitly says so. The founder chose your title; respect it."
+      },
       {
         title: "Single dispatch path \u2014 create_task only",
         domain: "workflow",
@@ -3778,6 +3800,12 @@ var init_platform_procedures = __esm({
         priority: "p0",
         content: "NEVER: (1) Access the database directly \u2014 it's SQLCipher encrypted, always fails. Use MCP tools only. (2) Manually spawn tmux sessions \u2014 create_task handles it. (3) Run git checkout main \u2014 agents work in worktrees. (4) Modify another agent's in-progress task. (5) Push to remote \u2014 the COO reviews and pushes. (6) Skip update_task(done) \u2014 it's the ONLY way your work gets reviewed. (7) Run git init."
       },
+      {
+        title: "Destructive operations \u2014 mandatory reviewer gate",
+        domain: "security",
+        priority: "p0",
+        content: "Before ANY destructive operation (delete, remove, overwrite, drop, reset, force-push, truncate), you MUST: (1) Have your full task spec accessible \u2014 if you cannot read it, STOP and report to your reviewer. Never improvise destructive actions. (2) Confirm with your reviewer (assigned_by or COO) before executing. (3) If the task spec explicitly authorizes the operation, proceed \u2014 but log it. Violation = immediate task failure. This applies to ALL agents regardless of role."
+      },
       {
         title: "Customer patch triage \u2014 upstream bug vs customization",
         domain: "support",
@@ -4063,10 +4091,24 @@ function stableId(memoryId, type, content) {
   return createHash2("sha256").update(`${memoryId}:${type}:${content}`).digest("hex").slice(0, 32);
 }
 function cleanText(text) {
-  return text.replace(/```[\s\S]*?```/g, " ").replace(/<[^>]+>/g, " ").replace(/\s+/g, " ").trim();
-}
-function splitSentences(text) {
-  return cleanText(text).split(/(?<=[.!?])\s+|\n+/).map((s) => s.trim()).filter((s) => s.length >= 24 && s.length <= MAX_SENTENCE_CHARS);
+  let cleaned = text.replace(
+    /```(\w*)\n(.*?)(?:\n[\s\S]*?)```/g,
+    (_m, lang, firstLine) => `[code${lang ? `:${lang}` : ""}] ${firstLine.trim()}`
+  );
+  cleaned = cleaned.replace(/<[^>]+>/g, " ").replace(/\s+/g, " ").trim();
+  return cleaned;
+}
+function splitSegments(text) {
+  const cleaned = cleanText(text);
+  const segments = cleaned.split(/(?<=[.!?:;])\s+|\n{2,}|(?<=\))\s+(?=[A-Z])|\s*[|│]\s*/).map((s) => s.trim()).filter((s) => s.length >= MIN_SEGMENT_CHARS && s.length <= MAX_SEGMENT_CHARS);
+  if (segments.length === 0 && cleaned.length >= MIN_SEGMENT_CHARS) {
+    const lines = cleaned.split(/\n+/).map((l) => l.trim()).filter((l) => l.length >= MIN_SEGMENT_CHARS && l.length <= MAX_SEGMENT_CHARS);
+    if (lines.length > 0) return lines;
+    if (cleaned.length >= MIN_SEGMENT_CHARS) {
+      return [cleaned.slice(0, MAX_SEGMENT_CHARS)];
+    }
+  }
+  return segments;
 }
 function inferCardType(sentence, toolName) {
   const lower = sentence.toLowerCase();
@@ -4098,12 +4140,12 @@ function predicateFor(type) {
   }
 }
 function extractMemoryCards(row) {
-  const sentences = splitSentences(row.raw_text);
+  const segments = splitSegments(row.raw_text);
   const cards = [];
-  for (const sentence of sentences) {
+  for (const sentence of segments) {
     const type = inferCardType(sentence, row.tool_name);
     const subject = extractSubject(sentence, row.agent_id);
-    const content = sentence.length > MAX_SENTENCE_CHARS ? `${sentence.slice(0, MAX_SENTENCE_CHARS - 1)}\u2026` : sentence;
+    const content = sentence.length > MAX_SEGMENT_CHARS ? `${sentence.slice(0, MAX_SEGMENT_CHARS - 1)}\u2026` : sentence;
     cards.push({
       id: stableId(row.id, type, content),
       memory_id: row.id,
@@ -4199,13 +4241,14 @@ Source memory: ${String(row.source_ref ?? row.memory_id)}`,
     last_accessed: String(row.timestamp)
   }));
 }
-var MAX_CARDS_PER_MEMORY, MAX_SENTENCE_CHARS;
+var MAX_CARDS_PER_MEMORY, MAX_SEGMENT_CHARS, MIN_SEGMENT_CHARS;
 var init_memory_cards = __esm({
   "src/lib/memory-cards.ts"() {
     "use strict";
     init_database();
-    MAX_CARDS_PER_MEMORY = 6;
-    MAX_SENTENCE_CHARS = 360;
+    MAX_CARDS_PER_MEMORY = 8;
+    MAX_SEGMENT_CHARS = 500;
+    MIN_SEGMENT_CHARS = 20;
   }
 });
 

package/dist/bin/backfill-vectors.js

@@ -3170,6 +3170,22 @@ async function ensureSchema() {
   } catch (e) {
     logCatchDebug("migration", e);
   }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN visibility TEXT DEFAULT 'private'`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
+  try {
+    await client.execute({
+      sql: `ALTER TABLE memories ADD COLUMN strength REAL DEFAULT 1.0`,
+      args: []
+    });
+  } catch (e) {
+    logCatchDebug("migration", e);
+  }
 }
 async function disposeDatabase() {
   if (_walCheckpointTimer) {
@@ -3736,11 +3752,17 @@ var init_platform_procedures = __esm({
         content: "Founder -> coordinator (the executive agent, internally routed as 'COO') -> CTO/CMO. CTO -> engineers. CMO -> content production. Never skip levels: the coordinator does not bypass managers for specialist work. Specialists report to their manager. If you need cross-team info, use ask_team_memory \u2014 don't read other agents' task folders. Each level owns dispatch downward and review upward."
       },
       {
-        title: "Customer orchestration maturity \u2014 recommend, never trap",
+        title: "Orchestration phase guidance \u2014 recommend, never trap",
         domain: "workflow",
         priority: "p1",
         content: "New customers start best in Phase 1: founder \u2194 coordinator/Chief of Staff, building company context. Suggest Phase 2 executives when domain work repeats; suggest Phase 3 parallel execution only when review/permission gates are ready. This is guidance, not a blocker: users may jump phases anytime. Never overwrite their phase, role titles, identities, or custom org design."
       },
+      {
+        title: "Routing slot vs display title \u2014 internal 'coo' is plumbing, not your name",
+        domain: "identity",
+        priority: "p0",
+        content: "These procedures reference 'COO' as a shorthand for the coordinator role. This is an INTERNAL routing slot used by exe-os code (chain-of-command checks, dispatch logic, session detection). It is NOT your display title. Your actual title comes from your identity file's `title:` field \u2014 that is what you use externally: introductions, sign-offs, team comms, and any user-facing text. If your identity says `title: AI Chief of Staff`, you are the AI Chief of Staff. The routing slot stays `role: coo` for code compatibility \u2014 never rename it, but also never introduce yourself as 'COO' unless your identity file explicitly says so. The founder chose your title; respect it."
+      },
       {
         title: "Single dispatch path \u2014 create_task only",
         domain: "workflow",
@@ -3774,6 +3796,12 @@ var init_platform_procedures = __esm({
         priority: "p0",
         content: "NEVER: (1) Access the database directly \u2014 it's SQLCipher encrypted, always fails. Use MCP tools only. (2) Manually spawn tmux sessions \u2014 create_task handles it. (3) Run git checkout main \u2014 agents work in worktrees. (4) Modify another agent's in-progress task. (5) Push to remote \u2014 the COO reviews and pushes. (6) Skip update_task(done) \u2014 it's the ONLY way your work gets reviewed. (7) Run git init."
       },
+      {
+        title: "Destructive operations \u2014 mandatory reviewer gate",
+        domain: "security",
+        priority: "p0",
+        content: "Before ANY destructive operation (delete, remove, overwrite, drop, reset, force-push, truncate), you MUST: (1) Have your full task spec accessible \u2014 if you cannot read it, STOP and report to your reviewer. Never improvise destructive actions. (2) Confirm with your reviewer (assigned_by or COO) before executing. (3) If the task spec explicitly authorizes the operation, proceed \u2014 but log it. Violation = immediate task failure. This applies to ALL agents regardless of role."
+      },
       {
         title: "Customer patch triage \u2014 upstream bug vs customization",
         domain: "support",