memorix 1.2.2 → 1.2.3

package/src/orchestrate/output-budget.ts

@@ -1,80 +1,80 @@
-/**
- * Output Budget — Phase 7, Step 4: Trim large outputs to a token-friendly size.
- *
- * When gate outputs, agent tail outputs, or other large strings need to be
- * embedded in prompts (e.g., fix prompts, ledger entries), this module trims
- * them to a configurable budget while preserving the most useful parts
- * (beginning + end).
- *
- * Full output is optionally persisted to disk for debugging.
- */
-
-import { writeFileSync, mkdirSync } from 'node:fs';
-import { join, dirname } from 'node:path';
-
-// ── Constants ──────────────────────────────────────────────────────
-
-/** Default output budget in bytes */
-export const DEFAULT_OUTPUT_BUDGET = 4096;
-
-// ── Core ───────────────────────────────────────────────────────────
-
-/**
- * Trim text to a byte budget. If the text exceeds the budget, keep
- * the first half and last half with an omission marker in between.
- *
- * Returns the trimmed text. Never throws.
- */
-export function trimToBudget(text: string, budget: number = DEFAULT_OUTPUT_BUDGET): string {
-  if (text.length <= budget) return text;
-
-  const half = Math.floor(budget / 2) - 40; // Reserve space for marker
-  if (half <= 0) return text.slice(0, budget);
-
-  const omitted = text.length - half * 2;
-  return (
-    text.slice(0, half) +
-    `\n\n... (${omitted} bytes omitted — full output saved to disk) ...\n\n` +
-    text.slice(-half)
-  );
-}
-
-/**
- * Persist full output to a file and return the trimmed version.
- *
- * If disk write fails, returns the trimmed text without file reference.
- * Never throws.
- */
-export function trimAndPersist(
-  text: string,
-  filePath: string,
-  budget: number = DEFAULT_OUTPUT_BUDGET,
-): { trimmed: string; persisted: boolean; fullPath?: string } {
-  if (text.length <= budget) {
-    return { trimmed: text, persisted: false };
-  }
-
-  // Persist full output to disk (best-effort)
-  let persisted = false;
-  try {
-    mkdirSync(dirname(filePath), { recursive: true });
-    writeFileSync(filePath, text, 'utf-8');
-    persisted = true;
-  } catch {
-    // Disk write failed — degrade gracefully
-  }
-
-  const half = Math.floor(budget / 2) - 60; // Reserve space for marker + path
-  if (half <= 0) {
-    return { trimmed: text.slice(0, budget), persisted, fullPath: persisted ? filePath : undefined };
-  }
-
-  const omitted = text.length - half * 2;
-  const pathNote = persisted ? ` Full output at: ${filePath}` : '';
-  const trimmed =
-    text.slice(0, half) +
-    `\n\n... (${omitted} bytes omitted.${pathNote}) ...\n\n` +
-    text.slice(-half);
-
-  return { trimmed, persisted, fullPath: persisted ? filePath : undefined };
-}
+/**
+ * Output Budget — Phase 7, Step 4: Trim large outputs to a token-friendly size.
+ *
+ * When gate outputs, agent tail outputs, or other large strings need to be
+ * embedded in prompts (e.g., fix prompts, ledger entries), this module trims
+ * them to a configurable budget while preserving the most useful parts
+ * (beginning + end).
+ *
+ * Full output is optionally persisted to disk for debugging.
+ */
+
+import { writeFileSync, mkdirSync } from 'node:fs';
+import { join, dirname } from 'node:path';
+
+// ── Constants ──────────────────────────────────────────────────────
+
+/** Default output budget in bytes */
+export const DEFAULT_OUTPUT_BUDGET = 4096;
+
+// ── Core ───────────────────────────────────────────────────────────
+
+/**
+ * Trim text to a byte budget. If the text exceeds the budget, keep
+ * the first half and last half with an omission marker in between.
+ *
+ * Returns the trimmed text. Never throws.
+ */
+export function trimToBudget(text: string, budget: number = DEFAULT_OUTPUT_BUDGET): string {
+  if (text.length <= budget) return text;
+
+  const half = Math.floor(budget / 2) - 40; // Reserve space for marker
+  if (half <= 0) return text.slice(0, budget);
+
+  const omitted = text.length - half * 2;
+  return (
+    text.slice(0, half) +
+    `\n\n... (${omitted} bytes omitted — full output saved to disk) ...\n\n` +
+    text.slice(-half)
+  );
+}
+
+/**
+ * Persist full output to a file and return the trimmed version.
+ *
+ * If disk write fails, returns the trimmed text without file reference.
+ * Never throws.
+ */
+export function trimAndPersist(
+  text: string,
+  filePath: string,
+  budget: number = DEFAULT_OUTPUT_BUDGET,
+): { trimmed: string; persisted: boolean; fullPath?: string } {
+  if (text.length <= budget) {
+    return { trimmed: text, persisted: false };
+  }
+
+  // Persist full output to disk (best-effort)
+  let persisted = false;
+  try {
+    mkdirSync(dirname(filePath), { recursive: true });
+    writeFileSync(filePath, text, 'utf-8');
+    persisted = true;
+  } catch {
+    // Disk write failed — degrade gracefully
+  }
+
+  const half = Math.floor(budget / 2) - 60; // Reserve space for marker + path
+  if (half <= 0) {
+    return { trimmed: text.slice(0, budget), persisted, fullPath: persisted ? filePath : undefined };
+  }
+
+  const omitted = text.length - half * 2;
+  const pathNote = persisted ? ` Full output at: ${filePath}` : '';
+  const trimmed =
+    text.slice(0, half) +
+    `\n\n... (${omitted} bytes omitted.${pathNote}) ...\n\n` +
+    text.slice(-half);
+
+  return { trimmed, persisted, fullPath: persisted ? filePath : undefined };
+}

package/src/orchestrate/permission.ts

@@ -1,152 +1,152 @@
-/**
- * Permission — Phase 7, Step 10: Risk classification + monitoring for agent tools.
- *
- * Classifies tool calls by risk tier and records them for post-run audit.
- * Does NOT block or restrict agent tools (Non-Invasiveness principle).
- * The primary use case is monitoring and reporting:
- *   - Which high-risk tools were used and how often
- *   - Whether agents used filesystem-destructive or network operations
- *   - Per-task risk summary in the evidence directory
- *
- * Three tiers:
- *   - safe:      read-only operations (Read, Grep, List, etc.)
- *   - moderate:  write operations (Edit, Write, Execute safe commands)
- *   - dangerous: destructive/network/exec operations (Delete, Bash, Deploy, etc.)
- */
-
-// ── Types ──────────────────────────────────────────────────────────
-
-export type RiskTier = 'safe' | 'moderate' | 'dangerous';
-
-export interface ToolUsageRecord {
-  tool: string;
-  tier: RiskTier;
-  count: number;
-  firstSeen: number;
-  lastSeen: number;
-}
-
-export interface TaskRiskProfile {
-  taskId: string;
-  /** Highest risk tier used by this task */
-  maxTier: RiskTier;
-  /** Total tool calls */
-  totalCalls: number;
-  /** Tool usage breakdown */
-  tools: ToolUsageRecord[];
-}
-
-// ── Classification Registry ────────────────────────────────────────
-
-const SAFE_TOOLS = new Set([
-  'read_file', 'grep_search', 'find_by_name', 'list_dir', 'code_search',
-  'read_notebook', 'read_url_content', 'view_content_chunk',
-  'memorix_search', 'memorix_detail', 'memorix_search_reasoning',
-  'mcp_query', 'search_web',
-]);
-
-const MODERATE_TOOLS = new Set([
-  'edit', 'multi_edit', 'write_to_file', 'edit_notebook',
-  'memorix_store', 'memorix_store_reasoning', 'memorix_resolve',
-  'memorix_session_start', 'memorix_handoff', 'memorix_poll',
-  'git_commit', 'git_add', 'git_checkout',
-]);
-
-const DANGEROUS_TOOLS = new Set([
-  'run_command', 'bash', 'execute_command', 'terminal',
-  'delete_file', 'remove_directory',
-  'deploy_web_app', 'browser_navigate', 'browser_click',
-  'http_request', 'fetch_url',
-]);
-
-// ── Core ───────────────────────────────────────────────────────────
-
-/**
- * Classify a tool name into a risk tier.
- * Unknown tools default to 'moderate' (conservative but not alarming).
- */
-export function classifyTool(toolName: string): RiskTier {
-  const name = toolName.toLowerCase().replace(/[^a-z0-9_]/g, '_');
-  if (SAFE_TOOLS.has(name)) return 'safe';
-  if (DANGEROUS_TOOLS.has(name)) return 'dangerous';
-  if (MODERATE_TOOLS.has(name)) return 'moderate';
-  // Heuristic fallback: tools with 'write', 'delete', 'exec', 'run' in name
-  if (/delete|remove|exec|run|deploy|bash|terminal/.test(name)) return 'dangerous';
-  if (/write|edit|create|update|store|commit/.test(name)) return 'moderate';
-  return 'moderate'; // Unknown → moderate (safe default)
-}
-
-/**
- * Task-level tool usage tracker. Create one per dispatched task.
- */
-export class TaskToolTracker {
-  private tools = new Map<string, ToolUsageRecord>();
-  readonly taskId: string;
-
-  constructor(taskId: string) {
-    this.taskId = taskId;
-  }
-
-  /** Record a tool call. Call this when an agent uses a tool. */
-  record(toolName: string): void {
-    const tier = classifyTool(toolName);
-    const now = Date.now();
-    const existing = this.tools.get(toolName);
-
-    if (existing) {
-      existing.count++;
-      existing.lastSeen = now;
-    } else {
-      this.tools.set(toolName, {
-        tool: toolName,
-        tier,
-        count: 1,
-        firstSeen: now,
-        lastSeen: now,
-      });
-    }
-  }
-
-  /** Get the risk profile for this task. */
-  getProfile(): TaskRiskProfile {
-    const tools = Array.from(this.tools.values());
-    const totalCalls = tools.reduce((sum, t) => sum + t.count, 0);
-
-    let maxTier: RiskTier = 'safe';
-    for (const t of tools) {
-      if (t.tier === 'dangerous') { maxTier = 'dangerous'; break; }
-      if (t.tier === 'moderate') maxTier = 'moderate';
-    }
-
-    return {
-      taskId: this.taskId,
-      maxTier,
-      totalCalls,
-      tools,
-    };
-  }
-
-  /** Format a human-readable risk summary. */
-  formatSummary(): string {
-    const profile = this.getProfile();
-    const lines = [`Risk: ${profile.maxTier} (${profile.totalCalls} tool calls)`];
-
-    const dangerous = profile.tools.filter(t => t.tier === 'dangerous');
-    if (dangerous.length > 0) {
-      lines.push('  Dangerous tools:');
-      for (const t of dangerous) {
-        lines.push(`    - ${t.tool}: ${t.count}x`);
-      }
-    }
-
-    return lines.join('\n');
-  }
-}
-
-/**
- * Compare two risk tiers. Returns positive if a > b, negative if a < b, 0 if equal.
- */
-export function compareTiers(a: RiskTier, b: RiskTier): number {
-  const order: Record<RiskTier, number> = { safe: 0, moderate: 1, dangerous: 2 };
-  return order[a] - order[b];
-}
+/**
+ * Permission — Phase 7, Step 10: Risk classification + monitoring for agent tools.
+ *
+ * Classifies tool calls by risk tier and records them for post-run audit.
+ * Does NOT block or restrict agent tools (Non-Invasiveness principle).
+ * The primary use case is monitoring and reporting:
+ *   - Which high-risk tools were used and how often
+ *   - Whether agents used filesystem-destructive or network operations
+ *   - Per-task risk summary in the evidence directory
+ *
+ * Three tiers:
+ *   - safe:      read-only operations (Read, Grep, List, etc.)
+ *   - moderate:  write operations (Edit, Write, Execute safe commands)
+ *   - dangerous: destructive/network/exec operations (Delete, Bash, Deploy, etc.)
+ */
+
+// ── Types ──────────────────────────────────────────────────────────
+
+export type RiskTier = 'safe' | 'moderate' | 'dangerous';
+
+export interface ToolUsageRecord {
+  tool: string;
+  tier: RiskTier;
+  count: number;
+  firstSeen: number;
+  lastSeen: number;
+}
+
+export interface TaskRiskProfile {
+  taskId: string;
+  /** Highest risk tier used by this task */
+  maxTier: RiskTier;
+  /** Total tool calls */
+  totalCalls: number;
+  /** Tool usage breakdown */
+  tools: ToolUsageRecord[];
+}
+
+// ── Classification Registry ────────────────────────────────────────
+
+const SAFE_TOOLS = new Set([
+  'read_file', 'grep_search', 'find_by_name', 'list_dir', 'code_search',
+  'read_notebook', 'read_url_content', 'view_content_chunk',
+  'memorix_search', 'memorix_detail', 'memorix_search_reasoning',
+  'mcp_query', 'search_web',
+]);
+
+const MODERATE_TOOLS = new Set([
+  'edit', 'multi_edit', 'write_to_file', 'edit_notebook',
+  'memorix_store', 'memorix_store_reasoning', 'memorix_resolve',
+  'memorix_session_start', 'memorix_handoff', 'memorix_poll',
+  'git_commit', 'git_add', 'git_checkout',
+]);
+
+const DANGEROUS_TOOLS = new Set([
+  'run_command', 'bash', 'execute_command', 'terminal',
+  'delete_file', 'remove_directory',
+  'deploy_web_app', 'browser_navigate', 'browser_click',
+  'http_request', 'fetch_url',
+]);
+
+// ── Core ───────────────────────────────────────────────────────────
+
+/**
+ * Classify a tool name into a risk tier.
+ * Unknown tools default to 'moderate' (conservative but not alarming).
+ */
+export function classifyTool(toolName: string): RiskTier {
+  const name = toolName.toLowerCase().replace(/[^a-z0-9_]/g, '_');
+  if (SAFE_TOOLS.has(name)) return 'safe';
+  if (DANGEROUS_TOOLS.has(name)) return 'dangerous';
+  if (MODERATE_TOOLS.has(name)) return 'moderate';
+  // Heuristic fallback: tools with 'write', 'delete', 'exec', 'run' in name
+  if (/delete|remove|exec|run|deploy|bash|terminal/.test(name)) return 'dangerous';
+  if (/write|edit|create|update|store|commit/.test(name)) return 'moderate';
+  return 'moderate'; // Unknown → moderate (safe default)
+}
+
+/**
+ * Task-level tool usage tracker. Create one per dispatched task.
+ */
+export class TaskToolTracker {
+  private tools = new Map<string, ToolUsageRecord>();
+  readonly taskId: string;
+
+  constructor(taskId: string) {
+    this.taskId = taskId;
+  }
+
+  /** Record a tool call. Call this when an agent uses a tool. */
+  record(toolName: string): void {
+    const tier = classifyTool(toolName);
+    const now = Date.now();
+    const existing = this.tools.get(toolName);
+
+    if (existing) {
+      existing.count++;
+      existing.lastSeen = now;
+    } else {
+      this.tools.set(toolName, {
+        tool: toolName,
+        tier,
+        count: 1,
+        firstSeen: now,
+        lastSeen: now,
+      });
+    }
+  }
+
+  /** Get the risk profile for this task. */
+  getProfile(): TaskRiskProfile {
+    const tools = Array.from(this.tools.values());
+    const totalCalls = tools.reduce((sum, t) => sum + t.count, 0);
+
+    let maxTier: RiskTier = 'safe';
+    for (const t of tools) {
+      if (t.tier === 'dangerous') { maxTier = 'dangerous'; break; }
+      if (t.tier === 'moderate') maxTier = 'moderate';
+    }
+
+    return {
+      taskId: this.taskId,
+      maxTier,
+      totalCalls,
+      tools,
+    };
+  }
+
+  /** Format a human-readable risk summary. */
+  formatSummary(): string {
+    const profile = this.getProfile();
+    const lines = [`Risk: ${profile.maxTier} (${profile.totalCalls} tool calls)`];
+
+    const dangerous = profile.tools.filter(t => t.tier === 'dangerous');
+    if (dangerous.length > 0) {
+      lines.push('  Dangerous tools:');
+      for (const t of dangerous) {
+        lines.push(`    - ${t.tool}: ${t.count}x`);
+      }
+    }
+
+    return lines.join('\n');
+  }
+}
+
+/**
+ * Compare two risk tiers. Returns positive if a > b, negative if a < b, 0 if equal.
+ */
+export function compareTiers(a: RiskTier, b: RiskTier): number {
+  const order: Record<RiskTier, number> = { safe: 0, moderate: 1, dangerous: 2 };
+  return order[a] - order[b];
+}