codegraph/src/mcp/tools.ts

/**
 * MCP Tool Definitions
 *
 * Defines the tools exposed by the CodeGraph MCP server.
 */

import CodeGraph, { findNearestCodeGraphRoot } from '../index';
import type { Node, Edge, SearchResult, Subgraph, TaskContext, NodeKind } from '../types';
import { createHash } from 'crypto';
import {
  constants as fsConstants,
  closeSync,
  existsSync,
  lstatSync,
  openSync,
  readFileSync,
  writeSync,
} from 'fs';
import { clamp, validatePathWithinRoot, validateProjectPath } from '../utils';
import { tmpdir } from 'os';
import { join } from 'path';

/** Maximum output length to prevent context bloat (characters) */
const MAX_OUTPUT_LENGTH = 15000;

/**
 * Maximum length for free-form string inputs (query, task, symbol).
 * Bounds memory and CPU when a buggy or hostile MCP client sends a
 * huge payload — without this an attacker could ship a 100MB string
 * and force a full FTS5 scan / OOM the server. 10 000 characters is
 * far beyond any realistic legitimate query.
 */
const MAX_INPUT_LENGTH = 10_000;

/**
 * Maximum length for path-like string inputs (projectPath, path
 * filter, glob pattern). Paths beyond a few thousand chars are
 * never legitimate and signal abuse or a bug upstream.
 */
const MAX_PATH_LENGTH = 4_096;

/**
 * Rust path roots that have no file-system equivalent — `crate` is the
 * current crate, `super` is the parent module, `self` is the current
 * module. Used by `matchesSymbol` to strip these before file-path
 * matching so `crate::configurator::stage_apply::run` resolves the
 * same as `configurator::stage_apply::run`.
 */
const RUST_PATH_PREFIXES = new Set(['crate', 'super', 'self']);

/**
 * Node kinds that contain other symbols. For these, `codegraph_node` with
 * `includeCode=true` returns a structural outline (member names + signatures
 * + line numbers) instead of the full body, which for a large class is a
 * multi-thousand-character wall of source that bloats the agent's context.
 */
const CONTAINER_NODE_KINDS = new Set<NodeKind>([
  'class', 'struct', 'interface', 'trait', 'protocol', 'enum', 'namespace', 'module',
]);

/** Last `::` / `.` / `/`-separated segment of a qualified symbol. */
function lastQualifierPart(symbol: string): string {
  const parts = symbol.split(/::|[./]/).filter((p) => p.length > 0);
  return parts[parts.length - 1] ?? symbol;
}

/**
 * Calculate the recommended number of codegraph_explore calls based on project size.
 * Larger codebases need more exploration calls to cover their surface area,
 * but smaller ones should use fewer to avoid unnecessary overhead.
 */
export function getExploreBudget(fileCount: number): number {
  if (fileCount < 500) return 1;
  if (fileCount < 5000) return 2;
  if (fileCount < 15000) return 3;
  if (fileCount < 25000) return 4;
  return 5;
}

/**
 * Adaptive output budget for `codegraph_explore`, scaled to project size.
 *
 * Smaller codebases get a tighter total cap, fewer default files, smaller
 * per-file cap, and tighter clustering — so a focused query on a 100-file
 * project doesn't dump a whole file's worth of source into the agent's
 * context. Larger codebases keep the generous defaults because the
 * agent's native discovery cost (grep + find + many Reads) genuinely
 * dwarfs a fat explore call at that scale.
 *
 * Meta-text (relationships map, "additional relevant files" list,
 * completeness signal, budget note) is gated off for tiny projects
 * where one rich call is the whole story and the extra prose is just
 * overhead.
 *
 * Tier breakpoints mirror `getExploreBudget` so a project sits in the
 * same tier across both knobs.
 */
export interface ExploreOutputBudget {
  /** Hard cap on total output characters. */
  maxOutputChars: number;
  /** Default `maxFiles` when the caller didn't specify one. */
  defaultMaxFiles: number;
  /** Cap on contiguous source returned per file (across all its clusters). */
  maxCharsPerFile: number;
  /** Cluster gap threshold in lines — tighter clustering on small projects. */
  gapThreshold: number;
  /** Max symbols listed in the per-file header (`#### path — sym(kind), ...`). */
  maxSymbolsInFileHeader: number;
  /** Max edges shown per relationship kind in the Relationships section. */
  maxEdgesPerRelationshipKind: number;
  /** Include the "Relationships" section. */
  includeRelationships: boolean;
  /** Include the "Additional relevant files (not shown)" trailing list. */
  includeAdditionalFiles: boolean;
  /** Include the "Complete source code is included above…" reminder. */
  includeCompletenessSignal: boolean;
  /** Include the explore-budget reminder at the end. */
  includeBudgetNote: boolean;
}

export function getExploreOutputBudget(fileCount: number): ExploreOutputBudget {
  if (fileCount < 500) {
    return {
      maxOutputChars: 18000,
      defaultMaxFiles: 5,
      maxCharsPerFile: 3800,
      gapThreshold: 8,
      maxSymbolsInFileHeader: 6,
      maxEdgesPerRelationshipKind: 6,
      includeRelationships: true,
      includeAdditionalFiles: false,
      includeCompletenessSignal: false,
      includeBudgetNote: false,
    };
  }
  if (fileCount < 5000) {
    return {
      // Sized so ONE explore can cover a flow that centers on a god-file (e.g.
      // excalidraw's 415 KB App.tsx): the previous 2500/file returned <1% of such
      // a file, forcing the agent to Read it anyway. Per-file must also stay ≥ the
      // smaller <500 tier (3800) — the old 2500 was non-monotonic. Tokens are
      // cheap relative to a 5–10 Read round-trip spiral; favor sufficiency.
      maxOutputChars: 28000,
      defaultMaxFiles: 10,
      maxCharsPerFile: 6500,
      gapThreshold: 12,
      maxSymbolsInFileHeader: 10,
      maxEdgesPerRelationshipKind: 10,
      includeRelationships: true,
      includeAdditionalFiles: true,
      includeCompletenessSignal: true,
      includeBudgetNote: true,
    };
  }
  if (fileCount < 15000) {
    return {
      maxOutputChars: 35000,
      defaultMaxFiles: 12,
      maxCharsPerFile: 7000,
      gapThreshold: 15,
      maxSymbolsInFileHeader: 15,
      maxEdgesPerRelationshipKind: 15,
      includeRelationships: true,
      includeAdditionalFiles: true,
      includeCompletenessSignal: true,
      includeBudgetNote: true,
    };
  }
  return {
    maxOutputChars: 38000,
    defaultMaxFiles: 14,
    maxCharsPerFile: 7000,
    gapThreshold: 15,
    maxSymbolsInFileHeader: 15,
    maxEdgesPerRelationshipKind: 15,
    includeRelationships: true,
    includeAdditionalFiles: true,
    includeCompletenessSignal: true,
    includeBudgetNote: true,
  };
}

/**
 * Whether `codegraph_explore` should prefix source lines with their line
 * numbers (cat -n style: `<num>\t<code>`).
 *
 * Line numbers let the agent cite `file:line` straight from the explore
 * payload instead of re-Reading the file just to find a line number — the
 * dominant residual cost on precise-tracing questions (#185 follow-up).
 *
 * Defaults ON. Set `CODEGRAPH_EXPLORE_LINENUMS=0` to disable (used by the
 * A/B harness to measure the payload-cost vs. read-savings tradeoff).
 */
function exploreLineNumbersEnabled(): boolean {
  return process.env.CODEGRAPH_EXPLORE_LINENUMS !== '0';
}

/**
 * Prefix each line of a source slice with its 1-based line number, matching
 * the Read tool's `cat -n` convention (number + tab) so the agent treats it
 * the same way it treats Read output.
 *
 * @param slice  contiguous source text (already extracted from the file)
 * @param firstLineNumber  the 1-based line number of the slice's first line
 */
function numberSourceLines(slice: string, firstLineNumber: number): string {
  const out: string[] = [];
  const split = slice.split('\n');
  for (let i = 0; i < split.length; i++) {
    out.push(`${firstLineNumber + i}\t${split[i]}`);
  }
  return out.join('\n');
}

/**
 * Mark a Claude session as having consulted MCP tools.
 * This enables Grep/Glob/Bash commands that would otherwise be blocked.
 *
 * Why the explicit openSync + O_NOFOLLOW dance instead of plain writeFileSync:
 * tmpdir() is world-writable on Linux (mode 1777), so on a shared multi-user
 * machine any other local user can pre-create `codegraph-consulted-<hash>` as
 * a symlink pointing at a file the victim owns. The old `writeFileSync` would
 * happily follow that link and overwrite the target's contents with the ISO
 * timestamp string (CWE-59). The session-id hash provides the predictability
 * gate, but it's defense-in-depth: if a session id ever surfaces in logs,
 * argv, or telemetry the attack becomes trivial, and the right fix is to not
 * follow links from /tmp paths in the first place.
 */
function markSessionConsulted(sessionId: string): void {
  try {
    const hash = createHash('md5').update(sessionId).digest('hex').slice(0, 16);
    const markerPath = join(tmpdir(), `codegraph-consulted-${hash}`);
    // Refuse to follow a pre-planted symlink at the marker path (CWE-59).
    // O_NOFOLLOW (below) is the atomic, TOCTOU-free guard on POSIX, but it is
    // `undefined` on Windows (libuv ignores it), so the bitwise-OR silently
    // drops it and openSync would follow the link. This lstat check closes that
    // gap cross-platform; ENOENT (path is free) falls through to create it.
    try {
      if (lstatSync(markerPath).isSymbolicLink()) return;
    } catch {
      // No existing entry (or stat failed) — nothing to refuse; proceed.
    }
    // O_NOFOLLOW makes openSync throw ELOOP if markerPath is already a symlink.
    // O_CREAT + O_TRUNC keep the original "create-or-overwrite" semantics, and
    // mode 0o600 prevents readback by other local users (the marker payload is
    // benign, but narrowing the exposure costs nothing).
    const flags = fsConstants.O_WRONLY | fsConstants.O_CREAT | fsConstants.O_TRUNC | fsConstants.O_NOFOLLOW;
    const fd = openSync(markerPath, flags, 0o600);
    try {
      writeSync(fd, new Date().toISOString());
    } finally {
      closeSync(fd);
    }
  } catch {
    // Silently fail - don't break MCP on marker write failure. ELOOP from a
    // planted symlink lands here too, which is the intended behavior: refuse
    // to write rather than overwrite an attacker-chosen target.
  }
}

/**
 * MCP Tool definition
 */
export interface ToolDefinition {
  name: string;
  description: string;
  inputSchema: {
    type: 'object';
    properties: Record<string, PropertySchema>;
    required?: string[];
  };
}

interface PropertySchema {
  type: string;
  description: string;
  enum?: string[];
  default?: unknown;
}

/**
 * Tool execution result
 */
export interface ToolResult {
  content: Array<{
    type: 'text';
    text: string;
  }>;
  isError?: boolean;
}

/**
 * Common projectPath property for cross-project queries
 */
const projectPathProperty: PropertySchema = {
  type: 'string',
  description: 'Path to a different project with .codegraph/ initialized. If omitted, uses current project. Use this to query other codebases.',
};

/**
 * All CodeGraph MCP tools
 *
 * Designed for minimal context usage - use codegraph_context as the primary tool,
 * and only use other tools for targeted follow-up queries.
 *
 * All tools support cross-project queries via the optional `projectPath` parameter.
 */
export const tools: ToolDefinition[] = [
  {
    name: 'codegraph_search',
    description: 'Quick symbol search by name. Returns locations only (no code). Use codegraph_context instead for comprehensive task context.',
    inputSchema: {
      type: 'object',
      properties: {
        query: {
          type: 'string',
          description: 'Symbol name or partial name (e.g., "auth", "signIn", "UserService")',
        },
        kind: {
          type: 'string',
          description: 'Filter by node kind',
          enum: ['function', 'method', 'class', 'interface', 'type', 'variable', 'route', 'component'],
        },
        limit: {
          type: 'number',
          description: 'Maximum results (default: 10)',
          default: 10,
        },
        projectPath: projectPathProperty,
      },
      required: ['query'],
    },
  },
  {
    name: 'codegraph_context',
    description: 'PRIMARY TOOL — call this FIRST for any "how does X work", architecture, feature, or bug-context question. Composes search + node + callers + callees and returns entry points, related symbols, and key code in ONE call — usually enough to answer with no further search/Read/Grep. Prefer this over chaining codegraph_search + codegraph_node, and over codegraph_explore. NOTE: provides CODE context, not product requirements; for new features still clarify UX/edge cases with the user.',
    inputSchema: {
      type: 'object',
      properties: {
        task: {
          type: 'string',
          description: 'Description of the task, bug, or feature to build context for',
        },
        maxNodes: {
          type: 'number',
          description: 'Maximum symbols to include (default: 20)',
          default: 20,
        },
        includeCode: {
          type: 'boolean',
          description: 'Include code snippets for key symbols (default: true)',
          default: true,
        },
        projectPath: projectPathProperty,
      },
      required: ['task'],
    },
  },
  {
    name: 'codegraph_callers',
    description: 'Find all functions/methods that call a specific symbol. Useful for understanding usage patterns and impact of changes.',
    inputSchema: {
      type: 'object',
      properties: {
        symbol: {
          type: 'string',
          description: 'Name of the function, method, or class to find callers for',
        },
        limit: {
          type: 'number',
          description: 'Maximum number of callers to return (default: 20)',
          default: 20,
        },
        projectPath: projectPathProperty,
      },
      required: ['symbol'],
    },
  },
  {
    name: 'codegraph_callees',
    description: 'Find all functions/methods that a specific symbol calls. Useful for understanding dependencies and code flow.',
    inputSchema: {
      type: 'object',
      properties: {
        symbol: {
          type: 'string',
          description: 'Name of the function, method, or class to find callees for',
        },
        limit: {
          type: 'number',
          description: 'Maximum number of callees to return (default: 20)',
          default: 20,
        },
        projectPath: projectPathProperty,
      },
      required: ['symbol'],
    },
  },
  {
    name: 'codegraph_impact',
    description: 'Analyze the impact radius of changing a symbol. Shows what code could be affected by modifications.',
    inputSchema: {
      type: 'object',
      properties: {
        symbol: {
          type: 'string',
          description: 'Name of the symbol to analyze impact for',
        },
        depth: {
          type: 'number',
          description: 'How many levels of dependencies to traverse (default: 2)',
          default: 2,
        },
        projectPath: projectPathProperty,
      },
      required: ['symbol'],
    },
  },
  {
    name: 'codegraph_node',
    description: 'Get ONE symbol\'s details (location, signature, docstring) PLUS its TRAIL — what it calls and what calls it, each with file:line. Pass includeCode=true for source (functions return their body; containers return a member outline). Use this to WALK the call graph hop-by-hop — node a symbol, then node one of its trail entries — the structural, no-Read way to follow "what calls/triggers/handles X" across files. For a broad first overview of many symbols at once use codegraph_explore; use node to drill along a specific path from there. (If a trail is empty on a non-leaf, that hop is likely dynamic dispatch — read just that line.) Source returned with includeCode is the verbatim live file content — identical to Read.',
    inputSchema: {
      type: 'object',
      properties: {
        symbol: {
          type: 'string',
          description: 'Name of the symbol to get details for',
        },
        includeCode: {
          type: 'boolean',
          description: 'Include full source code (default: false to minimize context)',
          default: false,
        },
        projectPath: projectPathProperty,
      },
      required: ['symbol'],
    },
  },
  {
    name: 'codegraph_explore',
    description: 'Returns source for SEVERAL related symbols grouped by file, plus a relationship map, in ONE capped call. This is the efficient way to inspect many related symbols at once — strongly prefer it over a series of codegraph_node or Read calls (each separate call re-reads the whole context, so 8 node calls cost far more than 1 explore). Use it after codegraph_context when you need to see the actual source of several symbols. Query with specific symbol/file/code terms, NOT natural-language sentences — run codegraph_search first to find names. Bad: "how are agent prompts loaded and passed to the CLI". Good: "renderStaticScene drawElementOnCanvas ShapeCache renderElement.ts". The code it returns is the VERBATIM live file source (byte-for-byte identical to Read), line-numbered — not a summary; treat files it shows as already Read, no need to re-open them.',
    inputSchema: {
      type: 'object',
      properties: {
        query: {
          type: 'string',
          description: 'Symbol names, file names, or short code terms to explore (e.g., "AuthService loginUser session-manager", "GraphTraverser BFS impact traversal.ts"). Use codegraph_search first to find relevant names.',
        },
        maxFiles: {
          type: 'number',
          description: 'Maximum number of files to include source code from (default: 12)',
          default: 12,
        },
        projectPath: projectPathProperty,
      },
      required: ['query'],
    },
  },
  {
    name: 'codegraph_status',
    description: 'Get the status of the CodeGraph index, including statistics about indexed files, nodes, and edges.',
    inputSchema: {
      type: 'object',
      properties: {
        projectPath: projectPathProperty,
      },
    },
  },
  {
    name: 'codegraph_files',
    description: 'REQUIRED for file/folder exploration. Get the project file structure from the CodeGraph index. Returns a tree view of all indexed files with metadata (language, symbol count). Much faster than Glob/filesystem scanning. Use this FIRST when exploring project structure, finding files, or understanding codebase organization.',
    inputSchema: {
      type: 'object',
      properties: {
        path: {
          type: 'string',
          description: 'Filter to files under this directory path (e.g., "src/components"). Returns all files if not specified.',
        },
        pattern: {
          type: 'string',
          description: 'Filter files matching this glob pattern (e.g., "*.tsx", "**/*.test.ts")',
        },
        format: {
          type: 'string',
          description: 'Output format: "tree" (hierarchical, default), "flat" (simple list), "grouped" (by language)',
          enum: ['tree', 'flat', 'grouped'],
          default: 'tree',
        },
        includeMetadata: {
          type: 'boolean',
          description: 'Include file metadata like language and symbol count (default: true)',
          default: true,
        },
        maxDepth: {
          type: 'number',
          description: 'Maximum directory depth to show (default: unlimited)',
        },
        projectPath: projectPathProperty,
      },
    },
  },
  {
    name: 'codegraph_trace',
    description: 'Trace the CALL PATH between two symbols — "how does <from> reach/become <to>?" Returns the chain of functions from one to the other (each hop with file:line and its body inlined, plus the outgoing calls of the destination itself) in ONE call. This is something grep/Read structurally cannot do: there is no text pattern for "the path from A to B". Ideal for flow questions — how an update triggers a render, how a request reaches a handler, how a QuerySet becomes SQL. If no static path exists the chain likely breaks at dynamic dispatch (callbacks/descriptors/metaclasses); the tool says where and points you to codegraph_node to bridge it.',
    inputSchema: {
      type: 'object',
      properties: {
        from: {
          type: 'string',
          description: 'Symbol the flow starts at (e.g., "QuerySet", "handleRequest", "mutateElement")',
        },
        to: {
          type: 'string',
          description: 'Symbol the flow should reach (e.g., "execute_sql", "render", "setState")',
        },
        projectPath: projectPathProperty,
      },
      required: ['from', 'to'],
    },
  },
];

/**
 * Tool handler that executes tools against a CodeGraph instance
 *
 * Supports cross-project queries via the projectPath parameter.
 * Other projects are opened on-demand and cached for performance.
 */
export class ToolHandler {
  // Cache of opened CodeGraph instances for cross-project queries
  private projectCache: Map<string, CodeGraph> = new Map();
  // The directory the server last searched for a default project. Surfaced in
  // the "not initialized" error so users can see why detection missed.
  private defaultProjectHint: string | null = null;

  constructor(private cg: CodeGraph | null) {}

  /**
   * Update the default CodeGraph instance (e.g. after lazy initialization)
   */
  setDefaultCodeGraph(cg: CodeGraph): void {
    this.cg = cg;
  }

  /**
   * Record the directory the server tried to resolve the default project from.
   * Used only to make the "no default project" error actionable.
   */
  setDefaultProjectHint(searchedPath: string): void {
    this.defaultProjectHint = searchedPath;
  }

  /**
   * Whether a default CodeGraph instance is available
   */
  hasDefaultCodeGraph(): boolean {
    return this.cg !== null;
  }

  /**
   * Optional allowlist of exposed tools, parsed from the CODEGRAPH_MCP_TOOLS
   * env var (comma-separated short names, e.g. "trace,search,node,context").
   * Unset/empty → every tool is exposed. Lets an operator (or an A/B harness)
   * trim the tool surface without rebuilding the client config; the ablated
   * tool is then truly absent from ListTools rather than merely denied on call.
   * Matching is on the short form, so "trace" and "codegraph_trace" both work.
   */
  private toolAllowlist(): Set<string> | null {
    const raw = process.env.CODEGRAPH_MCP_TOOLS;
    if (!raw || !raw.trim()) return null;
    const short = (s: string) => s.trim().replace(/^codegraph_/, '');
    const set = new Set(raw.split(',').map(short).filter(Boolean));
    return set.size ? set : null;
  }

  /** Whether a tool name passes the CODEGRAPH_MCP_TOOLS allowlist (if any). */
  private isToolAllowed(name: string): boolean {
    const allow = this.toolAllowlist();
    return !allow || allow.has(name.replace(/^codegraph_/, ''));
  }

  /**
   * Get tool definitions with dynamic descriptions based on project size.
   * The codegraph_explore tool description includes a budget recommendation
   * scaled to the number of indexed files. Honors the CODEGRAPH_MCP_TOOLS
   * allowlist so a trimmed surface is reflected in ListTools.
   */
  getTools(): ToolDefinition[] {
    const allow = this.toolAllowlist();
    const visible = allow
      ? tools.filter(t => allow.has(t.name.replace(/^codegraph_/, '')))
      : tools;
    if (!this.cg) return visible;

    try {
      const stats = this.cg.getStats();
      const budget = getExploreBudget(stats.fileCount);

      return visible.map(tool => {
        if (tool.name === 'codegraph_explore') {
          return {
            ...tool,
            description: `${tool.description} Budget: make at most ${budget} calls for this project (${stats.fileCount.toLocaleString()} files indexed).`,
          };
        }
        return tool;
      });
    } catch {
      return visible;
    }
  }

  /**
   * Get CodeGraph instance for a project
   *
   * If projectPath is provided, opens that project's CodeGraph (cached).
   * Otherwise returns the default CodeGraph instance.
   *
   * Walks up parent directories to find the nearest .codegraph/ folder,
   * similar to how git finds .git/ directories.
   */
  private getCodeGraph(projectPath?: string): CodeGraph {
    if (!projectPath) {
      if (!this.cg) {
        const searched = this.defaultProjectHint ?? process.cwd();
        throw new Error(
          'No CodeGraph project is loaded for this session.\n' +
          `Searched for a .codegraph/ directory starting from: ${searched}\n` +
          'The index is likely fine — this is a working-directory detection issue: ' +
          "the MCP client launched the server outside your project and didn't report the " +
          'workspace root. Fix it either way:\n' +
          '  • Pass projectPath to the tool call, e.g. projectPath: "/absolute/path/to/your/project"\n' +
          '  • Or add --path to the server\'s MCP config args: ["serve", "--mcp", "--path", "/absolute/path/to/your/project"]'
        );
      }
      return this.cg;
    }

    // Check cache first (using original path as key)
    if (this.projectCache.has(projectPath)) {
      return this.projectCache.get(projectPath)!;
    }

    // Reject sensitive system directories before opening. Only validate a
    // path that actually exists — a nested or not-yet-created sub-path of a
    // real project must still be allowed to resolve UP to its .codegraph/
    // root below (issue #238), so we don't run the existence-checking
    // validator on paths that are meant to walk up.
    if (existsSync(projectPath)) {
      const pathError = validateProjectPath(projectPath);
      if (pathError) {
        throw new Error(pathError);
      }
    }

    // Walk up parent directories to find nearest .codegraph/
    const resolvedRoot = findNearestCodeGraphRoot(projectPath);

    if (!resolvedRoot) {
      throw new Error(`CodeGraph not initialized in ${projectPath}. Run 'codegraph init' in that project first.`);
    }

    // If the path resolves to the default project, reuse the already-open
    // default instance rather than opening a SECOND connection to the same DB.
    // A duplicate connection serializes reads against the watcher's auto-sync
    // writes; on the wasm backend (no WAL) that surfaces as intermittent
    // "database is locked" on concurrent tool calls. See issue #238. Deliberately
    // not cached under projectPath — the server owns and closes the default
    // instance, so routing it through projectCache.closeAll() would double-close it.
    if (this.cg && this.cg.getProjectRoot() === resolvedRoot) {
      return this.cg;
    }

    // Check if we already have this resolved root cached (different path, same project)
    if (this.projectCache.has(resolvedRoot)) {
      const cg = this.projectCache.get(resolvedRoot)!;
      // Cache under original path too for faster future lookups
      this.projectCache.set(projectPath, cg);
      return cg;
    }

    // Open and cache under both paths
    const cg = CodeGraph.openSync(resolvedRoot);
    this.projectCache.set(resolvedRoot, cg);
    if (projectPath !== resolvedRoot) {
      this.projectCache.set(projectPath, cg);
    }
    return cg;
  }

  /**
   * Close all cached project connections
   */
  closeAll(): void {
    for (const cg of this.projectCache.values()) {
      cg.close();
    }
    this.projectCache.clear();
  }

  /**
   * Validate that a value is a non-empty string within length bounds.
   *
   * The `maxLength` cap protects against MCP clients that ship huge
   * payloads (10MB+ query strings either by accident or maliciously).
   * Without this, a single oversized input can pin the FTS5 index or
   * exhaust memory before any real work runs.
   */
  private validateString(
    value: unknown,
    name: string,
    maxLength: number = MAX_INPUT_LENGTH
  ): string | ToolResult {
    if (typeof value !== 'string' || value.length === 0) {
      return this.errorResult(`${name} must be a non-empty string`);
    }
    if (value.length > maxLength) {
      return this.errorResult(
        `${name} exceeds maximum length of ${maxLength} characters (got ${value.length})`
      );
    }
    return value;
  }

  /**
   * Validate an optional path-like string input. Returns the value if
   * valid (or undefined), or a ToolResult with the error.
   */
  private validateOptionalPath(
    value: unknown,
    name: string
  ): string | undefined | ToolResult {
    if (value === undefined || value === null) return undefined;
    if (typeof value !== 'string') {
      return this.errorResult(`${name} must be a string`);
    }
    if (value.length > MAX_PATH_LENGTH) {
      return this.errorResult(
        `${name} exceeds maximum length of ${MAX_PATH_LENGTH} characters (got ${value.length})`
      );
    }
    return value;
  }

  /**
   * Execute a tool by name
   */
  async execute(toolName: string, args: Record<string, unknown>): Promise<ToolResult> {
    try {
      // Honor the optional tool allowlist (CODEGRAPH_MCP_TOOLS): a trimmed
      // surface rejects ablated tools defensively even if a client cached them.
      if (!this.isToolAllowed(toolName)) {
        return this.errorResult(`Tool ${toolName} is disabled via CODEGRAPH_MCP_TOOLS`);
      }
      // Cross-cutting input validation. All tools accept an optional
      // `projectPath` and most accept either `query`, `task`, or
      // `symbol` — bound their lengths centrally so individual handlers
      // can stay focused on tool-specific logic.
      const pathCheck = this.validateOptionalPath(args.projectPath, 'projectPath');
      if (typeof pathCheck === 'object' && pathCheck !== undefined) {
        return pathCheck;
      }
      // The `path` and `pattern` properties used by codegraph_files are
      // also path-shaped — apply the same cap.
      if (args.path !== undefined) {
        const check = this.validateOptionalPath(args.path, 'path');
        if (typeof check === 'object' && check !== undefined) return check;
      }
      if (args.pattern !== undefined) {
        const check = this.validateOptionalPath(args.pattern, 'pattern');
        if (typeof check === 'object' && check !== undefined) return check;
      }

      switch (toolName) {
        case 'codegraph_search':
          return await this.handleSearch(args);
        case 'codegraph_context':
          return await this.handleContext(args);
        case 'codegraph_callers':
          return await this.handleCallers(args);
        case 'codegraph_callees':
          return await this.handleCallees(args);
        case 'codegraph_impact':
          return await this.handleImpact(args);
        case 'codegraph_explore':
          return await this.handleExplore(args);
        case 'codegraph_node':
          return await this.handleNode(args);
        case 'codegraph_status':
          return await this.handleStatus(args);
        case 'codegraph_files':
          return await this.handleFiles(args);
        case 'codegraph_trace':
          return await this.handleTrace(args);
        default:
          return this.errorResult(`Unknown tool: ${toolName}`);
      }
    } catch (err) {
      return this.errorResult(`Tool execution failed: ${err instanceof Error ? err.message : String(err)}`);
    }
  }

  /**
   * Handle codegraph_search
   */
  private async handleSearch(args: Record<string, unknown>): Promise<ToolResult> {
    const query = this.validateString(args.query, 'query');
    if (typeof query !== 'string') return query;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const kind = args.kind as string | undefined;
    const rawLimit = Number(args.limit) || 10;
    const limit = clamp(rawLimit, 1, 100);

    const results = cg.searchNodes(query, {
      limit,
      kinds: kind ? [kind as NodeKind] : undefined,
    });

    if (results.length === 0) {
      return this.textResult(`No results found for "${query}"`);
    }

    const formatted = this.formatSearchResults(results);
    return this.textResult(this.truncateOutput(formatted));
  }

  /**
   * Handle codegraph_context
   */
  private async handleContext(args: Record<string, unknown>): Promise<ToolResult> {
    const task = this.validateString(args.task, 'task');
    if (typeof task !== 'string') return task;

    // Mark session as consulted (enables Grep/Glob/Bash)
    const sessionId = process.env.CLAUDE_SESSION_ID;
    if (sessionId) {
      markSessionConsulted(sessionId);
    }

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const maxNodes = (args.maxNodes as number) || 20;
    const includeCode = args.includeCode !== false;

    const context = await cg.buildContext(task, {
      maxNodes,
      includeCode,
      format: 'markdown',
    });

    // Detect if this looks like a feature request (vs bug fix or exploration)
    const isFeatureQuery = this.looksLikeFeatureRequest(task);
    const reminder = isFeatureQuery
      ? '\n\n⚠️ **Ask user:** UX preferences, edge cases, acceptance criteria'
      : '';

    // buildContext returns string when format is 'markdown'
    if (typeof context === 'string') {
      return this.textResult(this.truncateOutput(context + reminder));
    }

    // If it returns TaskContext, format it
    return this.textResult(this.truncateOutput(this.formatTaskContext(context) + reminder));
  }

  /**
   * Heuristic to detect if a query looks like a feature request
   */
  private looksLikeFeatureRequest(task: string): boolean {
    const featureKeywords = [
      'add', 'create', 'implement', 'build', 'enable', 'allow',
      'new feature', 'support for', 'ability to', 'want to',
      'should be able', 'need to add', 'swap', 'edit', 'modify'
    ];
    const bugKeywords = [
      'fix', 'bug', 'error', 'broken', 'crash', 'issue', 'problem',
      'not working', 'fails', 'undefined', 'null'
    ];
    const explorationKeywords = [
      'how does', 'where is', 'what is', 'find', 'show me',
      'explain', 'understand', 'explore'
    ];

    const lowerTask = task.toLowerCase();

    // If it's clearly a bug or exploration, not a feature
    if (bugKeywords.some(k => lowerTask.includes(k))) return false;
    if (explorationKeywords.some(k => lowerTask.includes(k))) return false;

    // If it matches feature keywords, it's likely a feature request
    return featureKeywords.some(k => lowerTask.includes(k));
  }

  /**
   * Handle codegraph_callers
   */
  private async handleCallers(args: Record<string, unknown>): Promise<ToolResult> {
    const symbol = this.validateString(args.symbol, 'symbol');
    if (typeof symbol !== 'string') return symbol;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const limit = clamp((args.limit as number) || 20, 1, 100);

    const allMatches = this.findAllSymbols(cg, symbol);
    if (allMatches.nodes.length === 0) {
      return this.textResult(`Symbol "${symbol}" not found in the codebase`);
    }

    // Aggregate callers across all matching symbols
    const seen = new Set<string>();
    const allCallers: Node[] = [];
    for (const node of allMatches.nodes) {
      for (const c of cg.getCallers(node.id)) {
        if (!seen.has(c.node.id)) {
          seen.add(c.node.id);
          allCallers.push(c.node);
        }
      }
    }

    if (allCallers.length === 0) {
      return this.textResult(`No callers found for "${symbol}"${allMatches.note}`);
    }

    const formatted = this.formatNodeList(allCallers.slice(0, limit), `Callers of ${symbol}`) + allMatches.note;
    return this.textResult(this.truncateOutput(formatted));
  }

  /**
   * Handle codegraph_callees
   */
  private async handleCallees(args: Record<string, unknown>): Promise<ToolResult> {
    const symbol = this.validateString(args.symbol, 'symbol');
    if (typeof symbol !== 'string') return symbol;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const limit = clamp((args.limit as number) || 20, 1, 100);

    const allMatches = this.findAllSymbols(cg, symbol);
    if (allMatches.nodes.length === 0) {
      return this.textResult(`Symbol "${symbol}" not found in the codebase`);
    }

    // Aggregate callees across all matching symbols
    const seen = new Set<string>();
    const allCallees: Node[] = [];
    for (const node of allMatches.nodes) {
      for (const c of cg.getCallees(node.id)) {
        if (!seen.has(c.node.id)) {
          seen.add(c.node.id);
          allCallees.push(c.node);
        }
      }
    }

    if (allCallees.length === 0) {
      return this.textResult(`No callees found for "${symbol}"${allMatches.note}`);
    }

    const formatted = this.formatNodeList(allCallees.slice(0, limit), `Callees of ${symbol}`) + allMatches.note;
    return this.textResult(this.truncateOutput(formatted));
  }

  /**
   * Handle codegraph_impact
   */
  private async handleImpact(args: Record<string, unknown>): Promise<ToolResult> {
    const symbol = this.validateString(args.symbol, 'symbol');
    if (typeof symbol !== 'string') return symbol;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const depth = clamp((args.depth as number) || 2, 1, 10);

    const allMatches = this.findAllSymbols(cg, symbol);
    if (allMatches.nodes.length === 0) {
      return this.textResult(`Symbol "${symbol}" not found in the codebase`);
    }

    // Aggregate impact across all matching symbols
    const mergedNodes = new Map<string, Node>();
    const mergedEdges: Edge[] = [];
    const seenEdges = new Set<string>();

    for (const node of allMatches.nodes) {
      const impact = cg.getImpactRadius(node.id, depth);
      for (const [id, n] of impact.nodes) {
        mergedNodes.set(id, n);
      }
      for (const e of impact.edges) {
        const key = `${e.source}->${e.target}:${e.kind}`;
        if (!seenEdges.has(key)) {
          seenEdges.add(key);
          mergedEdges.push(e);
        }
      }
    }

    const mergedImpact = {
      nodes: mergedNodes,
      edges: mergedEdges,
      roots: allMatches.nodes.map(n => n.id),
    };

    const formatted = this.formatImpact(symbol, mergedImpact) + allMatches.note;
    return this.textResult(this.truncateOutput(formatted));
  }

  /**
   * Handle codegraph_trace — shortest CALL PATH between two symbols.
   *
   * Exposes GraphTraverser.findPath: the chain of functions from `from` to `to`,
   * each hop annotated with file:line and the call-site line. This is the
   * capability grep/Read structurally cannot provide. When no static path
   * exists, the chain has almost certainly broken at dynamic dispatch
   * (callbacks, descriptors, metaclasses) — we say so and surface the start
   * symbol's outgoing calls so the agent bridges the one missing hop with
   * codegraph_node rather than blindly reading.
   */
  private async handleTrace(args: Record<string, unknown>): Promise<ToolResult> {
    const from = this.validateString(args.from, 'from');
    if (typeof from !== 'string') return from;
    const to = this.validateString(args.to, 'to');
    if (typeof to !== 'string') return to;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const fromMatches = this.findAllSymbols(cg, from);
    if (fromMatches.nodes.length === 0) return this.textResult(`Symbol "${from}" not found in the codebase`);
    const toMatches = this.findAllSymbols(cg, to);
    if (toMatches.nodes.length === 0) return this.textResult(`Symbol "${to}" not found in the codebase`);

    // Trace along call edges only — a true call path. Names can map to several
    // nodes, so try a few from×to candidate pairs until a usable path turns up.
    //
    // MAX_HOPS guard: a BFS shortest path longer than this on a dense call graph
    // is almost always a spurious wander through unrelated code (django's
    // `_fetch_all → … → execute_sql` BFS detours through prefetch/filter), not
    // the real execution flow — and a confident-but-wrong 15-hop trace is worse
    // than none. Over-cap paths are rejected and reported as "no direct path"
    // (which, on real code, means the flow breaks at dynamic dispatch).
    const edgeKinds: Edge['kind'][] = ['calls'];
    const MAX_HOPS = 7;
    const fromTry = fromMatches.nodes.slice(0, 3);
    const toTry = toMatches.nodes.slice(0, 3);
    let path: Array<{ node: Node; edge: Edge | null }> | null = null;
    let overCap: Array<{ node: Node; edge: Edge | null }> | null = null;
    for (const f of fromTry) {
      for (const t of toTry) {
        const p = cg.findPath(f.id, t.id, edgeKinds);
        if (!p || p.length <= 1) continue;
        if (p.length <= MAX_HOPS) { path = p; break; }
        if (!overCap || p.length < overCap.length) overCap = p;
      }
      if (path) break;
    }

    if (!path) {
      // No static path — almost always a dynamic-dispatch break. Surface the
      // start symbol's outgoing calls so the agent can bridge the gap.
      const start = fromTry[0]!;
      const callees = cg.getCallees(start.id).slice(0, 10)
        .map(c => `${c.node.name} (${c.node.filePath}:${c.node.startLine})`);
      const lines = [
        `No direct call path from "${from}" to "${to}".`,
        '',
        (overCap
          ? `(Only a ${overCap.length}-hop indirect chain connects them — almost certainly a BFS wander through unrelated code, not the real flow.) `
          : '') +
        'The direct chain most likely breaks at **dynamic dispatch** (a callback, descriptor, ' +
        'metaclass, or attribute-as-callable) that static parsing cannot resolve into an edge. ' +
        `Inspect \`${start.name}\` (${start.filePath}:${start.startLine}) with codegraph_node ` +
        '(includeCode=true) — its body usually shows the dynamic call to follow next.',
      ];
      if (callees.length > 0) {
        lines.push('', `**${start.name} statically calls:** ${callees.join(', ')}`);
      }
      return this.textResult(lines.join('\n') + fromMatches.note + toMatches.note);
    }

    const lines: string[] = [
      `## Trace: ${from} → ${to}`,
      '',
      `Full execution path below — ${path.length} hops, each with its body, plus what the destination calls. This is the complete flow; answer from it.`,
      '',
      `${path.length} hops:`,
      '',
    ];
    // Inline what each hop needs so the agent doesn't Read/Grep to get it: the
    // call-site source line, the registration site for dynamic-dispatch hops, AND
    // the hop's own body (capped per hop so the trace stays path-scoped). Earlier
    // versions inlined only the call-site line, which left agents calling explore
    // or Read for the bodies — the exact follow-up the ablation experiment measured.
    const fileCache = new Map<string, string[]>();
    for (let i = 0; i < path.length; i++) {
      const step = path[i]!;
      if (step.edge) {
        const synth = this.synthEdgeNote(step.edge);
        if (synth) {
          lines.push(`   ↓ ${synth.label}`);
          if (synth.registeredAt) {
            const regSrc = this.sourceLineAt(cg, synth.registeredAt, fileCache);
            lines.push(`     ↳ registered at ${synth.registeredAt}${regSrc ? `   ${regSrc}` : ''}`);
          }
        } else {
          // The call happens in the PREVIOUS hop's file at edge.line.
          const prev = path[i - 1];
          const ref = prev && step.edge.line ? `${prev.node.filePath}:${step.edge.line}` : undefined;
          const callSrc = this.sourceLineAt(cg, ref, fileCache);
          lines.push(`   ↓ ${step.edge.kind}${step.edge.line ? `@${step.edge.line}` : ''}${callSrc ? `   ${callSrc}` : ''}`);
        }
      }
      lines.push(`${i + 1}. ${step.node.name} (${step.node.filePath}:${step.node.startLine}-${step.node.endLine})`);
      const body = this.sourceRangeAt(cg, step.node.filePath, step.node.startLine, step.node.endLine, fileCache, 60, 1800);
      if (body) lines.push(body);
    }
    // The "last mile": what the destination does next. Agents otherwise explore/Read
    // for exactly this (e.g. renderStaticScene → _renderStaticScene → the canvas draw),
    // so inlining the destination's callees is what actually stops the investigation —
    // sufficiency, not a "don't explore" instruction.
    const dest = path[path.length - 1]!.node;
    const destCallees = cg.getCallees(dest.id)
      .filter(c => !path.some(p => p.node.id === c.node.id))
      .slice(0, 6);
    if (destCallees.length > 0) {
      lines.push('', `### \`${dest.name}\` then calls (the destination's immediate work):`);
      for (const c of destCallees) {
        lines.push('', `- ${c.node.name} (${c.node.filePath}:${c.node.startLine}-${c.node.endLine})`);
        const body = this.sourceRangeAt(cg, c.node.filePath, c.node.startLine, c.node.endLine, fileCache, 16, 600);
        if (body) lines.push(body);
      }
    }
    lines.push('', '> Full path + every hop body + the destination\'s calls are inlined above — the complete flow. Answer from it; a Read is only needed to chase a specific local variable\'s data-flow.');
    return this.textResult(this.truncateOutput(lines.join('\n')));
  }

  /**
   * Describe a synthesized (dynamic-dispatch) edge for human output: how the
   * callback was wired up — the bridge static parsing can't see. Returns null
   * for ordinary static edges. Used by trace + the node trail so a synthesized
   * hop reads as "registered via onUpdate at App.tsx:3148", not a bare arrow.
   */
  private synthEdgeNote(edge: Edge | null): { label: string; compact: string; registeredAt?: string } | null {
    if (!edge || edge.provenance !== 'heuristic') return null;
    const m = edge.metadata as Record<string, unknown> | undefined;
    const registeredAt = typeof m?.registeredAt === 'string' ? m.registeredAt : undefined;
    const at = registeredAt ? ` @${registeredAt}` : '';
    if (m?.synthesizedBy === 'callback') {
      const via = m.via ? `\`${String(m.via)}\`` : 'a registrar';
      const field = m.field ? ` on .${String(m.field)}` : '';
      return {
        label: `callback — registered via ${via}${field} (dynamic dispatch)`,
        compact: `dynamic: callback via ${via}${at}`,
        registeredAt,
      };
    }
    if (m?.synthesizedBy === 'event-emitter') {
      const ev = m.event ? `\`${String(m.event)}\`` : 'an event';
      return {
        label: `event ${ev} — emit → handler (dynamic dispatch)`,
        compact: `dynamic: event ${ev}${at}`,
        registeredAt,
      };
    }
    if (m?.synthesizedBy === 'react-render') {
      return {
        label: `React re-render — \`setState\` re-runs render() (dynamic dispatch)`,
        compact: `dynamic: React re-render via setState${at}`,
        registeredAt,
      };
    }
    if (m?.synthesizedBy === 'jsx-render') {
      const child = m.via ? `<${String(m.via)}>` : 'a child component';
      return {
        label: `renders ${child} (JSX child — dynamic dispatch)`,
        compact: `dynamic: renders ${child}`,
        registeredAt,
      };
    }
    if (m?.synthesizedBy === 'vue-handler') {
      const ev = m.event ? `@${String(m.event)}` : 'a template event';
      return {
        label: `Vue template handler — bound to ${ev} (dynamic dispatch)`,
        compact: `dynamic: Vue ${ev} handler`,
        registeredAt,
      };
    }
    if (m?.synthesizedBy === 'interface-impl') {
      return {
        label: `interface/abstract dispatch — runs the implementation override (dynamic dispatch)`,
        compact: `dynamic: interface → impl${at}`,
        registeredAt,
      };
    }
    return null;
  }

  /**
   * Read one trimmed source line at "relpath:line" (relative to the project
   * root). `cache` holds split file contents so a multi-hop trace reads each
   * file at most once. Returns null if the file/line can't be resolved.
   */
  private sourceLineAt(cg: CodeGraph, ref: string | undefined, cache: Map<string, string[]>): string | null {
    if (!ref) return null;
    const i = ref.lastIndexOf(':');
    if (i < 0) return null;
    const filePath = ref.slice(0, i);
    const line = parseInt(ref.slice(i + 1), 10);
    if (!Number.isFinite(line) || line < 1) return null;
    let fileLines = cache.get(filePath);
    if (!fileLines) {
      const abs = validatePathWithinRoot(cg.getProjectRoot(), filePath);
      if (!abs || !existsSync(abs)) return null;
      try { fileLines = readFileSync(abs, 'utf-8').split('\n'); } catch { return null; }
      cache.set(filePath, fileLines);
    }
    const raw = fileLines[line - 1];
    if (raw == null) return null;
    const t = raw.trim();
    return t.length > 160 ? t.slice(0, 157) + '…' : t;
  }

  /**
   * Read a hop's body — filePath lines [startLine..endLine] — for inlining into
   * a trace, capped (lines + chars) so the whole path stays path-scoped even on
   * a 7-hop chain. Dedents to the body's own indentation and marks truncation.
   * Shares `cache` with sourceLineAt so each file is read at most once per trace.
   */
  private sourceRangeAt(
    cg: CodeGraph,
    filePath: string,
    startLine: number,
    endLine: number,
    cache: Map<string, string[]>,
    maxLines = 28,
    maxChars = 1200
  ): string | null {
    if (!Number.isFinite(startLine) || startLine < 1) return null;
    let fileLines = cache.get(filePath);
    if (!fileLines) {
      const abs = validatePathWithinRoot(cg.getProjectRoot(), filePath);
      if (!abs || !existsSync(abs)) return null;
      try { fileLines = readFileSync(abs, 'utf-8').split('\n'); } catch { return null; }
      cache.set(filePath, fileLines);
    }
    const end = Number.isFinite(endLine) && endLine >= startLine ? endLine : startLine;
    let slice = fileLines.slice(startLine - 1, end);
    if (slice.length === 0) return null;
    let omitted = 0;
    if (slice.length > maxLines) { omitted = slice.length - maxLines; slice = slice.slice(0, maxLines); }
    const nonBlank = slice.filter(l => l.trim().length > 0);
    const dedent = nonBlank.length ? Math.min(...nonBlank.map(l => l.length - l.trimStart().length)) : 0;
    let text = slice.map((l, i) => `      ${startLine + i}\t${l.slice(dedent)}`).join('\n');
    if (text.length > maxChars) {
      text = text.slice(0, maxChars).replace(/\n[^\n]*$/, '');
      omitted = Math.max(omitted, 1);
    }
    if (omitted > 0) text += `\n      … (+${omitted} more line${omitted === 1 ? '' : 's'})`;
    return text;
  }

  /**
   * Flow-from-named-symbols: an agent's codegraph_explore query is a bag of
   * symbol names that usually spans the flow it's investigating (e.g.
   * "PmsProductController getList PmsProductService list PmsProductServiceImpl").
   * Surface the longest call chain AMONG those named symbols — scoped to what the
   * agent explicitly named, so (unlike a fuzzy relevance set) there's no
   * wrong-feature wandering. Rides synthesized edges, so controller→service-
   * interface→impl shows up. Returns '' if no chain of >=3 nodes exists.
   *
   * Ambiguous tokens (Java `list` → dozens of nodes) are disambiguated by
   * CO-NAMING: the agent names the class too, so we keep only `list` candidates
   * whose qualifiedName contains another named token (`PmsProductServiceImpl::list`),
   * dropping unrelated `OmsOrderService::list`.
   */
  private buildFlowFromNamedSymbols(cg: CodeGraph, query: string): string {
    try {
      const CALLABLE = new Set(['method', 'function', 'component', 'constructor']);
      const tokens = [...new Set(
        query.split(/[\s,]+/)
          .map((t) => t.replace(/\.[A-Za-z0-9]+$/, '').trim()) // strip file ext: Create.cs → Create
          .filter((t) => /^[A-Za-z_$][\w$]{2,}$/.test(t))
      )].slice(0, 16);
      if (tokens.length < 2) return '';
      const lower = tokens.map((t) => t.toLowerCase());
      const named = new Map<string, Node>();
      for (const t of tokens) {
        const cands = this.findAllSymbols(cg, t).nodes.filter((n) => CALLABLE.has(n.kind));
        // Disambiguate by co-naming: for an ambiguous name keep only candidates
        // qualified by another named token; a specific name (<=3 hits) keeps all.
        const pick = cands.length <= 3
          ? cands
          : cands.filter((n) => {
              // Match qualifiedName SEGMENTS (Class::method), not substrings —
              // "list" is a substring of "getList" but not a segment of it.
              const segs = (n.qualifiedName || '').toLowerCase().split(/::|\./).filter(Boolean);
              return lower.some((o) => o !== t.toLowerCase() && segs.includes(o));
            });
        for (const n of pick.slice(0, 6)) named.set(n.id, n);
        if (named.size > 40) break;
      }
      if (named.size < 2) return '';
      const MAX_HOPS = 7;
      let best: Array<{ node: Node; edge: Edge | null }> | null = null;
      // BFS the full call graph (incl. synth edges) from each named seed, but
      // only ACCEPT a sink that is also named — both ends anchored to symbols the
      // agent named, so the chain stays on-topic while bridging intermediates
      // (e.g. the exact interface overload) that the token resolution missed.
      for (const seed of [...named.values()].slice(0, 8)) {
        const parent = new Map<string, { prev: string | null; edge: Edge | null; node: Node }>();
        parent.set(seed.id, { prev: null, edge: null, node: seed });
        const q: Array<{ id: string; depth: number; streak: number }> = [{ id: seed.id, depth: 0, streak: 0 }];
        let deep: string | null = null, deepDepth = 0;
        const MAX_BRIDGE = 1; // ≤1 consecutive UNNAMED hop: bridge one missing intermediate, never wander a god-function's fan-out
        for (let h = 0; h < q.length && parent.size < 1500; h++) {
          const { id, depth, streak } = q[h]!;
          if (id !== seed.id && named.has(id) && depth > deepDepth) { deep = id; deepDepth = depth; }
          if (depth >= MAX_HOPS - 1) continue;
          for (const c of cg.getCallees(id)) {
            if (c.edge.kind !== 'calls' || parent.has(c.node.id)) continue;
            const newStreak = named.has(c.node.id) ? 0 : streak + 1;
            if (newStreak > MAX_BRIDGE) continue;
            parent.set(c.node.id, { prev: id, edge: c.edge, node: c.node });
            q.push({ id: c.node.id, depth: depth + 1, streak: newStreak });
          }
        }
        if (!deep) continue;
        const chain: Array<{ node: Node; edge: Edge | null }> = [];
        let cur: string | null = deep;
        while (cur) { const p = parent.get(cur); if (!p) break; chain.push({ node: p.node, edge: p.edge }); cur = p.prev; }
        chain.reverse();
        if (!best || chain.length > best.length) best = chain;
      }
      if (!best || best.length < 3) return '';
      const out = ['## Flow (call path among the symbols you queried)', ''];
      for (let i = 0; i < best.length; i++) {
        const step = best[i]!;
        if (step.edge) { const sy = this.synthEdgeNote(step.edge); out.push(`   ↓ ${sy ? sy.compact : step.edge.kind}`); }
        out.push(`${i + 1}. ${step.node.name} (${step.node.filePath}:${step.node.startLine})`);
      }
      out.push('', '> Full source for these symbols is below; codegraph_trace(from,to) for the exact path between two endpoints.', '');
      return out.join('\n');
    } catch {
      return '';
    }
  }

  /**
   * Handle codegraph_explore — deep exploration in a single call
   *
   * Strategy: find relevant symbols via graph traversal, group by file,
   * then read contiguous file sections covering all symbols per file.
   * This replaces multiple codegraph_node + Read calls.
   *
   * Output size is adaptive to project file count via
   * `getExploreOutputBudget` — see #185 for why a fixed 35k cap was a
   * tax on small projects while earning its keep on large ones.
   */
  private async handleExplore(args: Record<string, unknown>): Promise<ToolResult> {
    const query = this.validateString(args.query, 'query');
    if (typeof query !== 'string') return query;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const projectRoot = cg.getProjectRoot();

    // Resolve adaptive output budget from project size. Falls back to the
    // largest-tier defaults if stats aren't available, which preserves
    // pre-#185 behavior for callers that hit the rare stats failure.
    let budget: ExploreOutputBudget;
    try {
      budget = getExploreOutputBudget(cg.getStats().fileCount);
    } catch {
      budget = getExploreOutputBudget(Infinity);
    }
    const maxFiles = clamp((args.maxFiles as number) || budget.defaultMaxFiles, 1, 20);

    // Step 1: Find relevant context with generous parameters.
    // Use a large maxNodes budget — explore has its own 35k char output limit
    // that prevents context bloat, so more nodes just means better coverage
    // across entry points (especially for large files like Svelte components).
    const subgraph = await cg.findRelevantContext(query, {
      searchLimit: 8,
      traversalDepth: 3,
      maxNodes: 200,
      minScore: 0.2,
    });

    if (subgraph.nodes.size === 0) {
      return this.textResult(`No relevant code found for "${query}"`);
    }

    // Graph-aware glue: findRelevantContext builds the subgraph from name/text
    // search, so a method that BRIDGES named symbols — e.g. App.tsx's
    // triggerRender, which calls the named triggerUpdate — is never a search hit
    // and gets missed, forcing the agent to Read the file to trace it. Pull in
    // the callers/callees of the entry (root) nodes, but ONLY those that live in
    // files the subgraph already surfaces (where the agent reads to fill gaps),
    // so we add wiring without dragging in unrelated files. These get an
    // importance boost below so they survive the per-file cluster budget.
    const glueNodeIds = new Set<string>();
    const subgraphFiles = new Set<string>();
    for (const n of subgraph.nodes.values()) subgraphFiles.add(n.filePath);
    const GLUE_NODE_CAP = 60;
    for (const rootId of subgraph.roots) {
      if (glueNodeIds.size >= GLUE_NODE_CAP) break;
      let neighbors: Node[] = [];
      try {
        neighbors = [
          ...cg.getCallers(rootId).map(c => c.node),
          ...cg.getCallees(rootId).map(c => c.node),
        ];
      } catch {
        continue;
      }
      for (const nb of neighbors) {
        if (glueNodeIds.size >= GLUE_NODE_CAP) break;
        if (subgraph.nodes.has(nb.id)) continue;
        if (!subgraphFiles.has(nb.filePath)) continue;
        subgraph.nodes.set(nb.id, nb);
        glueNodeIds.add(nb.id);
      }
    }

    // Step 2: Group nodes by file, score by relevance
    const fileGroups = new Map<string, { nodes: Node[]; score: number }>();
    const entryNodeIds = new Set(subgraph.roots);

    // Build a set of nodes directly connected to entry points (depth 1)
    const connectedToEntry = new Set<string>();
    for (const edge of subgraph.edges) {
      if (entryNodeIds.has(edge.source)) connectedToEntry.add(edge.target);
      if (entryNodeIds.has(edge.target)) connectedToEntry.add(edge.source);
    }

    for (const node of subgraph.nodes.values()) {
      // Skip import/export nodes — they add noise without information
      if (node.kind === 'import' || node.kind === 'export') continue;

      const group = fileGroups.get(node.filePath) || { nodes: [], score: 0 };
      group.nodes.push(node);
      // Score: entry point nodes worth 10, directly connected worth 3, others worth 1
      if (entryNodeIds.has(node.id)) {
        group.score += 10;
      } else if (connectedToEntry.has(node.id)) {
        group.score += 3;
      } else {
        group.score += 1;
      }
      fileGroups.set(node.filePath, group);
    }

    // Only include files that have entry points or nodes directly connected to entry points
    const relevantFiles = [...fileGroups.entries()].filter(([, group]) => group.score >= 3);

    // Extract query terms for relevance checking
    const queryTerms = query.toLowerCase().split(/\s+/).filter(t => t.length >= 3);

    // Sort files: highest relevance first, deprioritize low-value files
    const sortedFiles = relevantFiles.sort((a, b) => {
      const aPath = a[0].toLowerCase();
      const bPath = b[0].toLowerCase();

      // Check if any node name or file path relates to query terms
      const hasQueryRelevance = (filePath: string, nodes: Node[]) => {
        const fp = filePath.toLowerCase();
        if (queryTerms.some(t => fp.includes(t))) return true;
        return nodes.some(n => queryTerms.some(t => n.name.toLowerCase().includes(t)));
      };

      const aRelevant = hasQueryRelevance(aPath, a[1].nodes);
      const bRelevant = hasQueryRelevance(bPath, b[1].nodes);
      if (aRelevant !== bRelevant) return aRelevant ? -1 : 1;

      // Deprioritize test files, icon files, and i18n files
      const isLowValue = (p: string) =>
        /\/(tests?|__tests?__|spec)\//i.test(p) ||
        /\bicons?\b/i.test(p) ||
        /\bi18n\b/i.test(p);
      const aLow = isLowValue(aPath);
      const bLow = isLowValue(bPath);
      if (aLow !== bLow) return aLow ? 1 : -1;

      if (a[1].score !== b[1].score) return b[1].score - a[1].score;
      return b[1].nodes.length - a[1].nodes.length;
    });

    // Step 3: Build relationship map
    const lines: string[] = [
      `## Exploration: ${query}`,
      '',
      `Found ${subgraph.nodes.size} symbols across ${fileGroups.size} files.`,
      '',
    ];

    // Relationship map — show how symbols connect
    const significantEdges = subgraph.edges.filter(e =>
      e.kind !== 'contains' // skip contains — it's implied by file grouping
    );

    if (budget.includeRelationships && significantEdges.length > 0) {
      lines.push('### Relationships');
      lines.push('');

      // Group edges by kind for readability
      const byKind = new Map<string, Array<{ source: string; target: string }>>();
      for (const edge of significantEdges) {
        const sourceNode = subgraph.nodes.get(edge.source);
        const targetNode = subgraph.nodes.get(edge.target);
        if (!sourceNode || !targetNode) continue;

        const group = byKind.get(edge.kind) || [];
        group.push({ source: sourceNode.name, target: targetNode.name });
        byKind.set(edge.kind, group);
      }

      for (const [kind, edges] of byKind) {
        const cap = budget.maxEdgesPerRelationshipKind;
        const shown = edges.slice(0, cap);
        lines.push(`**${kind}:**`);
        for (const e of shown) {
          lines.push(`- ${e.source} → ${e.target}`);
        }
        if (edges.length > cap) {
          lines.push(`- ... and ${edges.length - cap} more`);
        }
        lines.push('');
      }
    }

    // Step 4: Read contiguous file sections
    lines.push('### Source Code');
    lines.push('');
    lines.push('> The code below is the **verbatim, current on-disk source** of these files — re-read from disk on this call and line-numbered, byte-for-byte identical to what the Read tool returns. It is NOT a summary, outline, or stale cache. Treat each block as a Read you have already performed: do not Read a file shown here.');
    lines.push('');

    let totalChars = lines.join('\n').length;
    let filesIncluded = 0;
    let anyFileTrimmed = false;

    for (const [filePath, group] of sortedFiles) {
      if (filesIncluded >= maxFiles) break;
      if (totalChars > budget.maxOutputChars * 0.9) break;

      const absPath = validatePathWithinRoot(projectRoot, filePath);
      if (!absPath || !existsSync(absPath)) continue;

      let fileContent: string;
      try {
        fileContent = readFileSync(absPath, 'utf-8');
      } catch {
        continue;
      }

      const fileLines = fileContent.split('\n');
      const lang = group.nodes[0]?.language || '';

      // Whole-small-file rule: if a relevant file is small enough to afford,
      // return it ENTIRELY instead of clustering. Clustering exists to tame
      // god-files (App.tsx ~13k lines); on a ~134-line component a cluster is a
      // lossy subset of a file the agent will just Read in full anyway — costing
      // a round-trip and a re-read every later turn. Reserve clustering for files
      // too big to ship whole. Still bounded by the total maxOutputChars check.
      const WHOLE_FILE_MAX_LINES = 220;
      const WHOLE_FILE_MAX_CHARS = budget.maxCharsPerFile * 3;
      if (fileLines.length <= WHOLE_FILE_MAX_LINES && fileContent.length <= WHOLE_FILE_MAX_CHARS) {
        const body = fileContent.replace(/\n+$/, '');
        let wholeSection = exploreLineNumbersEnabled() ? numberSourceLines(body, 1) : body;
        const uniqSymbols = [...new Set(
          group.nodes
            .filter(n => n.kind !== 'import' && n.kind !== 'export')
            .map(n => `${n.name}(${n.kind})`)
        )];
        const headerNames = uniqSymbols.slice(0, budget.maxSymbolsInFileHeader);
        const omitted = uniqSymbols.length - headerNames.length;
        const wholeHeader = `#### ${filePath} — ${omitted > 0 ? `${headerNames.join(', ')}, +${omitted} more` : headerNames.join(', ')}`;

        if (totalChars + wholeSection.length + 200 > budget.maxOutputChars) {
          const remaining = budget.maxOutputChars - totalChars - 200;
          if (remaining < 500) break;
          wholeSection = wholeSection.slice(0, remaining) + '\n... (trimmed) ...';
          anyFileTrimmed = true;
        }
        lines.push(wholeHeader, '', '```' + lang, wholeSection, '```', '');
        totalChars += wholeSection.length + 200;
        filesIncluded++;
        continue;
      }

      // Cluster nearby symbols to avoid reading huge gaps between distant symbols.
      // Sort by start line, then merge overlapping/adjacent ranges (within the
      // adaptive gap threshold). Include both node ranges AND edge source
      // locations so template sections with component usages/calls are
      // covered (not just script block symbols).
      //
      // Each range carries an `importance` score so we can rank clusters
      // when the per-file budget forces us to drop some: entry-point nodes
      // are worth 10, directly-connected nodes 3, peripheral nodes 1, and
      // bare edge-source lines 2 (less than a connected node but more than
      // a peripheral one — they hint at a reference but aren't a definition).
      // Container kinds whose body can span most/all of a file. When such a
      // node covers most of the file we drop it from the ranges: keeping it
      // would merge every method inside it into one giant cluster spanning
      // the whole file, which then tail-trims down to just the container's
      // opening lines (its header/declarations) and buries the methods the
      // query actually asked about (#185 follow-up — Session.swift in
      // Alamofire is the canonical case: the `Session` class spans ~1,400
      // lines). We want the granular symbols inside, not the envelope.
      const ENVELOPE_KINDS = new Set(['file', 'module', 'class', 'struct', 'interface', 'enum', 'namespace', 'protocol', 'trait', 'component']);
      const ranges: Array<{ start: number; end: number; name: string; kind: string; importance: number }> = group.nodes
        .filter(n => n.startLine > 0 && n.endLine > 0)
        // Drop whole-file envelope nodes (containers covering >50% of the file).
        .filter(n => !(ENVELOPE_KINDS.has(n.kind) && (n.endLine - n.startLine + 1) > fileLines.length * 0.5))
        .map(n => {
          let importance = 1;
          if (entryNodeIds.has(n.id)) importance = 10;
          else if (glueNodeIds.has(n.id)) importance = 6; // bridging caller/callee of an entry
          else if (connectedToEntry.has(n.id)) importance = 3;
          return { start: n.startLine, end: n.endLine, name: n.name, kind: n.kind, importance };
        });

      // Add edge source locations in this file — captures template references
      // (component usages, event handlers) that aren't nodes themselves.
      // Query edges directly from the DB (not just the subgraph) because BFS
      // traversal may have pruned template reference targets due to node budget.
      const edgeLines = new Set<string>(); // dedup by "line:name"
      for (const node of group.nodes) {
        const outgoing = cg.getOutgoingEdges(node.id);
        for (const edge of outgoing) {
          if (!edge.line || edge.line <= 0 || edge.kind === 'contains') continue;
          const key = `${edge.line}:${edge.target}`;
          if (edgeLines.has(key)) continue;
          edgeLines.add(key);
          // Look up target name from subgraph first, fall back to edge kind
          const targetNode = subgraph.nodes.get(edge.target);
          const targetName = targetNode?.name ?? edge.kind;
          ranges.push({ start: edge.line, end: edge.line, name: targetName, kind: edge.kind, importance: 2 });
        }
      }

      ranges.sort((a, b) => a.start - b.start);

      if (ranges.length === 0) continue;

      const gapThreshold = budget.gapThreshold;
      const clusters: Array<{ start: number; end: number; symbols: string[]; score: number; maxImportance: number }> = [];
      let current = {
        start: ranges[0]!.start,
        end: ranges[0]!.end,
        symbols: [`${ranges[0]!.name}(${ranges[0]!.kind})`],
        score: ranges[0]!.importance,
        maxImportance: ranges[0]!.importance,
      };

      for (let i = 1; i < ranges.length; i++) {
        const r = ranges[i]!;
        if (r.start <= current.end + gapThreshold) {
          current.end = Math.max(current.end, r.end);
          current.symbols.push(`${r.name}(${r.kind})`);
          current.score += r.importance;
          current.maxImportance = Math.max(current.maxImportance, r.importance);
        } else {
          clusters.push(current);
          current = {
            start: r.start,
            end: r.end,
            symbols: [`${r.name}(${r.kind})`],
            score: r.importance,
            maxImportance: r.importance,
          };
        }
      }
      clusters.push(current);

      // Build file section output from clusters, capped by per-file budget.
      // The pathological case (#185): a file like Session.swift where every
      // method is adjacent collapses into one cluster spanning the whole
      // file, and dumping that into the agent's context is most of the
      // token cost on small projects. We pick clusters in priority order
      // until the per-file char cap is hit. Truly enormous single clusters
      // get tail-trimmed with a marker.
      const contextPadding = 3;
      const withLineNumbers = exploreLineNumbersEnabled();
      const buildSection = (c: { start: number; end: number }): string => {
        const startIdx = Math.max(0, c.start - 1 - contextPadding);
        const endIdx = Math.min(fileLines.length, c.end + contextPadding);
        const slice = fileLines.slice(startIdx, endIdx).join('\n');
        // startIdx is 0-based, so the slice's first line is line startIdx + 1.
        return withLineNumbers ? numberSourceLines(slice, startIdx + 1) : slice;
      };
      // Language-neutral separator (no `//` — not a comment in Python, Ruby,
      // etc.). With line numbers on, the line-number jump also signals the gap.
      const GAP_MARKER = '\n\n... (gap) ...\n\n';

      // Rank clusters for inclusion under the per-file cap. Entry-point
      // clusters come first: a cluster containing a query entry point
      // (importance 10) must outrank a dense block of mere declarations,
      // otherwise on a large file like Session.swift the top-of-file class
      // header + property list (many adjacent low-importance nodes, high
      // density) wins the budget and buries the actual methods the query
      // asked about (perform/didCreateURLRequest/task live deep in the
      // file). Within the same importance tier, prefer density (score per
      // line) so we still favor focused clusters over sprawling ones, then
      // smaller span as a cheap-to-include tiebreak.
      const rankedClusters = clusters
        .map((c, i) => ({ idx: i, span: c.end - c.start + 1, c }))
        .sort((a, b) => {
          if (b.c.maxImportance !== a.c.maxImportance) return b.c.maxImportance - a.c.maxImportance;
          const densityA = a.c.score / a.span;
          const densityB = b.c.score / b.span;
          if (densityB !== densityA) return densityB - densityA;
          if (b.c.score !== a.c.score) return b.c.score - a.c.score;
          return a.span - b.span;
        });

      const chosenIndices = new Set<number>();
      let projectedChars = 0;
      for (const rc of rankedClusters) {
        const sectionLen = buildSection(rc.c).length + (chosenIndices.size > 0 ? GAP_MARKER.length : 0);
        // Always take the top-ranked cluster, even if oversize, so we don't
        // return an empty file section (agent would then re-Read the file,
        // negating the savings).
        if (chosenIndices.size === 0) {
          chosenIndices.add(rc.idx);
          projectedChars += sectionLen;
          continue;
        }
        if (projectedChars + sectionLen > budget.maxCharsPerFile) continue;
        chosenIndices.add(rc.idx);
        projectedChars += sectionLen;
      }

      // Emit chosen clusters in source order so the file reads top-to-bottom.
      let fileSection = '';
      const allSymbols: string[] = [];
      let fileTrimmed = false;
      for (let i = 0; i < clusters.length; i++) {
        if (!chosenIndices.has(i)) continue;
        const cluster = clusters[i]!;
        const section = buildSection(cluster);
        if (fileSection.length > 0) fileSection += GAP_MARKER;
        fileSection += section;
        allSymbols.push(...cluster.symbols);
      }

      // If a single chosen cluster is still oversize (long monolithic
      // function), tail-trim it. Better one trimmed view than nothing.
      if (fileSection.length > budget.maxCharsPerFile) {
        fileSection = fileSection.slice(0, budget.maxCharsPerFile) + '\n... (trimmed) ...';
        fileTrimmed = true;
      }
      if (chosenIndices.size < clusters.length || fileTrimmed) {
        anyFileTrimmed = true;
      }

      // Dedupe + cap the symbols list shown in the per-file header. Some
      // files (Session.swift in Alamofire) produced 3.4KB symbol lists
      // from cluster scoring + edge-source lines, dwarfing the per-file
      // body cap. Show top names by frequency, with a "+N more" tail.
      const symbolCounts = new Map<string, number>();
      for (const s of allSymbols) {
        symbolCounts.set(s, (symbolCounts.get(s) ?? 0) + 1);
      }
      const sortedSymbols = [...symbolCounts.entries()]
        .sort((a, b) => b[1] - a[1])
        .map(([name]) => name);
      const headerCap = budget.maxSymbolsInFileHeader;
      const headerSymbols = sortedSymbols.slice(0, headerCap);
      const omittedCount = sortedSymbols.length - headerSymbols.length;
      const headerSuffix = omittedCount > 0
        ? `${headerSymbols.join(', ')}, +${omittedCount} more`
        : headerSymbols.join(', ');
      const fileHeader = `#### ${filePath} — ${headerSuffix}`;

      // Respect the total output cap on a file-by-file basis.
      if (totalChars + fileSection.length + 200 > budget.maxOutputChars) {
        const remaining = budget.maxOutputChars - totalChars - 200;
        if (remaining < 500) break;
        const trimmed = fileSection.slice(0, remaining) + '\n... (trimmed) ...';

        lines.push(fileHeader);
        lines.push('');
        lines.push('```' + lang);
        lines.push(trimmed);
        lines.push('```');
        lines.push('');
        totalChars += trimmed.length + 200;
        filesIncluded++;
        anyFileTrimmed = true;
        break;
      }

      lines.push(fileHeader);
      lines.push('');
      lines.push('```' + lang);
      lines.push(fileSection);
      lines.push('```');
      lines.push('');

      totalChars += fileSection.length + 200;
      filesIncluded++;
    }

    // Add remaining files as references (from both relevant and peripheral files).
    // Small projects (per budget) skip this — the relevant story already fits
    // in the source section, and a trailing pointer list is pure overhead.
    if (budget.includeAdditionalFiles) {
      const remainingRelevant = sortedFiles.slice(filesIncluded);
      const peripheralFiles = [...fileGroups.entries()]
        .filter(([, group]) => group.score < 3)
        .sort((a, b) => b[1].score - a[1].score);
      const remainingFiles = [...remainingRelevant, ...peripheralFiles];
      if (remainingFiles.length > 0) {
        lines.push('### Not shown above — explore these names for their source');
        lines.push('');
        for (const [filePath, group] of remainingFiles.slice(0, 10)) {
          const symbols = group.nodes.map(n => `${n.name}:${n.startLine}`).join(', ');
          lines.push(`- ${filePath}: ${symbols}`);
        }
        if (remainingFiles.length > 10) {
          lines.push(`- ... and ${remainingFiles.length - 10} more files`);
        }
      }
    }

    // Add completeness signal so agents know they don't need to re-read these files.
    // On small projects the budget gates this off — but if we actually had to
    // trim or drop clusters, surface a brief note so the agent knows it can
    // still Read for more detail.
    if (budget.includeCompletenessSignal) {
      lines.push('');
      lines.push('---');
      lines.push(`> **Complete source for ${filesIncluded} files is included above — do NOT re-read them.** If your question also needs files/symbols listed under "Not shown above" (or any area this call didn't cover), make ANOTHER codegraph_explore targeting those names — it returns the same source with line numbers and is cheaper and more complete than reading. Reserve Read for a single specific line range explore can't surface.`);
    } else if (anyFileTrimmed) {
      lines.push('');
      lines.push(`> Some file sections were trimmed for size. For a specific symbol you still need, run another \`codegraph_explore\` (or \`codegraph_node\`) with its exact name — line-numbered source, cheaper and more complete than Read.`);
    }

    // Add explore budget note based on project size
    if (budget.includeBudgetNote) {
      try {
        const stats = cg.getStats();
        const callBudget = getExploreBudget(stats.fileCount);
        lines.push('');
        lines.push(`> **Explore budget: ${callBudget} calls for this project (${stats.fileCount.toLocaleString()} files indexed).** Each call covers ~6 files; if your question spans more, spend your remaining calls on the uncovered area BEFORE falling back to Read — another explore is cheaper and more complete than reading those files. Synthesize once you've used ${callBudget}.`);
      } catch {
        // Stats unavailable — skip budget note
      }
    }

    // Hard-cap to the adaptive budget. The per-file loop bounds the source
    // sections, but the relationship map, additional-files list, and
    // completeness/budget notes can still push the assembled output past
    // maxOutputChars (observed 30k against a 28k tier cap). A fat explore
    // payload persists in the agent's context and is re-read as cache-input
    // on every subsequent turn, so the overrun is paid many times over.
    const output = this.buildFlowFromNamedSymbols(cg, query) + lines.join('\n');
    if (output.length > budget.maxOutputChars) {
      const cut = output.slice(0, budget.maxOutputChars);
      const lastNewline = cut.lastIndexOf('\n');
      const safe = lastNewline > budget.maxOutputChars * 0.8 ? cut.slice(0, lastNewline) : cut;
      return this.textResult(safe + '\n\n... (output truncated to budget; the source above is complete and verbatim — treat it as already Read. For any area not covered, run another codegraph_explore with the specific names — do NOT Read these files.)');
    }
    return this.textResult(output);
  }

  /**
   * Handle codegraph_node
   */
  private async handleNode(args: Record<string, unknown>): Promise<ToolResult> {
    const symbol = this.validateString(args.symbol, 'symbol');
    if (typeof symbol !== 'string') return symbol;

    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    // Default to false to minimize context usage
    const includeCode = args.includeCode === true;

    const match = this.findSymbol(cg, symbol);
    if (!match) {
      return this.textResult(`Symbol "${symbol}" not found in the codebase`);
    }

    let code: string | null = null;
    let outline: string | null = null;

    if (includeCode) {
      // For container symbols (class/interface/struct/…), the full body is the
      // sum of every method body — a wall of source (e.g. a 10k-char class)
      // that bloats context and is rarely needed in full. Return a structural
      // outline (members + signatures + line numbers) instead; the agent can
      // Read or codegraph_node a specific method for its body. Leaf symbols
      // (function/method/etc.) return their full body as before.
      if (CONTAINER_NODE_KINDS.has(match.node.kind)) {
        outline = this.buildContainerOutline(cg, match.node);
      }
      if (!outline) {
        code = await cg.getCode(match.node.id);
      }
    }

    const trail = this.formatTrail(cg, match.node);
    const formatted = this.formatNodeDetails(match.node, code, outline) + trail + match.note;
    return this.textResult(this.truncateOutput(formatted));
  }

  /**
   * Build the "trail" for a symbol: its direct callees (what it calls) and
   * callers (what calls it), each with file:line — so codegraph_node doubles as
   * the structural Grep→Read→expand primitive: a spot PLUS where to go next.
   * Capped to stay cheap. Walk the graph by calling codegraph_node on a trail
   * entry; no Read needed for covered hops. Empty edges on a non-leaf often mean
   * dynamic dispatch the static graph couldn't resolve — that absence is itself
   * a signal (read that one hop) rather than a dead end.
   */
  private formatTrail(cg: CodeGraph, node: Node): string {
    const TRAIL_CAP = 12;
    const fmt = (e: { node: Node; edge: Edge }) => {
      const base = `${e.node.name} (${e.node.filePath}:${e.node.startLine})`;
      const synth = this.synthEdgeNote(e.edge);
      return synth ? `${base} [${synth.compact}]` : base;
    };
    const collect = (edges: Array<{ node: Node; edge: Edge }>): Array<{ node: Node; edge: Edge }> => {
      const seen = new Set<string>([node.id]);
      const out: Array<{ node: Node; edge: Edge }> = [];
      for (const e of edges) {
        if (seen.has(e.node.id)) continue;
        seen.add(e.node.id);
        out.push(e);
      }
      return out;
    };
    const callees = collect(cg.getCallees(node.id));
    const callers = collect(cg.getCallers(node.id));
    if (callees.length === 0 && callers.length === 0) return '';
    const lines: string[] = ['', '### Trail — codegraph_node any of these to follow it (no Read needed)'];
    if (callees.length > 0) {
      lines.push(`**Calls →** ${callees.slice(0, TRAIL_CAP).map(fmt).join(', ')}${callees.length > TRAIL_CAP ? `, +${callees.length - TRAIL_CAP} more` : ''}`);
    }
    if (callers.length > 0) {
      lines.push(`**Called by ←** ${callers.slice(0, TRAIL_CAP).map(fmt).join(', ')}${callers.length > TRAIL_CAP ? `, +${callers.length - TRAIL_CAP} more` : ''}`);
    }
    return lines.join('\n');
  }

  /**
   * Handle codegraph_status
   */
  private async handleStatus(args: Record<string, unknown>): Promise<ToolResult> {
    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const stats = cg.getStats();

    const lines: string[] = [
      '## CodeGraph Status',
      '',
      `**Files indexed:** ${stats.fileCount}`,
      `**Total nodes:** ${stats.nodeCount}`,
      `**Total edges:** ${stats.edgeCount}`,
      `**Database size:** ${(stats.dbSizeBytes / 1024 / 1024).toFixed(2)} MB`,
    ];

    // Surface the active SQLite backend (node:sqlite, Node's built-in real
    // SQLite — full WAL + FTS5, no native build).
    lines.push(`**Backend:** node:sqlite (Node built-in) — full WAL + FTS5`);

    // Effective journal mode. 'wal' ⇒ concurrent reads never block on a writer;
    // anything else ⇒ they can ("database is locked"). node:sqlite supports WAL
    // everywhere, so a non-wal mode means the filesystem can't (network/
    // virtualized mounts, WSL2 /mnt). See issue #238.
    const journalMode = cg.getJournalMode();
    if (journalMode === 'wal') {
      lines.push(`**Journal mode:** wal (concurrent reads safe)`);
    } else {
      lines.push(
        `**Journal mode:** ⚠ ${journalMode || 'unknown'} — WAL not active, so reads ` +
        `can block on a concurrent write (WAL appears unsupported on this filesystem)`
      );
    }

    lines.push('', '### Nodes by Kind:');

    for (const [kind, count] of Object.entries(stats.nodesByKind)) {
      if ((count as number) > 0) {
        lines.push(`- ${kind}: ${count}`);
      }
    }

    lines.push('', '### Languages:');
    for (const [lang, count] of Object.entries(stats.filesByLanguage)) {
      if ((count as number) > 0) {
        lines.push(`- ${lang}: ${count}`);
      }
    }

    return this.textResult(lines.join('\n'));
  }

  /**
   * Handle codegraph_files - get project file structure from the index
   */
  private async handleFiles(args: Record<string, unknown>): Promise<ToolResult> {
    const cg = this.getCodeGraph(args.projectPath as string | undefined);
    const pathFilter = args.path as string | undefined;
    const pattern = args.pattern as string | undefined;
    const format = (args.format as 'tree' | 'flat' | 'grouped') || 'tree';
    const includeMetadata = args.includeMetadata !== false;
    const maxDepth = args.maxDepth != null ? clamp(args.maxDepth as number, 1, 20) : undefined;

    // Get all files from the index
    const allFiles = cg.getFiles();

    if (allFiles.length === 0) {
      return this.textResult('No files indexed. Run `codegraph index` first.');
    }

    // Filter by path prefix
    let files = pathFilter
      ? allFiles.filter(f => f.path.startsWith(pathFilter) || f.path.startsWith('./' + pathFilter))
      : allFiles;

    // Filter by glob pattern
    if (pattern) {
      const regex = this.globToRegex(pattern);
      files = files.filter(f => regex.test(f.path));
    }

    if (files.length === 0) {
      return this.textResult(`No files found matching the criteria.`);
    }

    // Format output
    let output: string;
    switch (format) {
      case 'flat':
        output = this.formatFilesFlat(files, includeMetadata);
        break;
      case 'grouped':
        output = this.formatFilesGrouped(files, includeMetadata);
        break;
      case 'tree':
      default:
        output = this.formatFilesTree(files, includeMetadata, maxDepth);
        break;
    }

    return this.textResult(this.truncateOutput(output));
  }

  /**
   * Convert glob pattern to regex
   */
  private globToRegex(pattern: string): RegExp {
    const escaped = pattern
      .replace(/[.+^${}()|[\]\\]/g, '\\$&')  // Escape special regex chars except * and ?
      .replace(/\*\*/g, '{{GLOBSTAR}}')       // Temp placeholder for **
      .replace(/\*/g, '[^/]*')                // * matches anything except /
      .replace(/\?/g, '[^/]')                 // ? matches single char except /
      .replace(/\{\{GLOBSTAR\}\}/g, '.*');    // ** matches anything including /
    return new RegExp(escaped);
  }

  /**
   * Format files as a flat list
   */
  private formatFilesFlat(files: { path: string; language: string; nodeCount: number }[], includeMetadata: boolean): string {
    const lines: string[] = [`## Files (${files.length})`, ''];

    for (const file of files.sort((a, b) => a.path.localeCompare(b.path))) {
      if (includeMetadata) {
        lines.push(`- ${file.path} (${file.language}, ${file.nodeCount} symbols)`);
      } else {
        lines.push(`- ${file.path}`);
      }
    }

    return lines.join('\n');
  }

  /**
   * Format files grouped by language
   */
  private formatFilesGrouped(files: { path: string; language: string; nodeCount: number }[], includeMetadata: boolean): string {
    const byLang = new Map<string, typeof files>();

    for (const file of files) {
      const existing = byLang.get(file.language) || [];
      existing.push(file);
      byLang.set(file.language, existing);
    }

    const lines: string[] = [`## Files by Language (${files.length} total)`, ''];

    // Sort languages by file count (descending)
    const sortedLangs = [...byLang.entries()].sort((a, b) => b[1].length - a[1].length);

    for (const [lang, langFiles] of sortedLangs) {
      lines.push(`### ${lang} (${langFiles.length})`);
      for (const file of langFiles.sort((a, b) => a.path.localeCompare(b.path))) {
        if (includeMetadata) {
          lines.push(`- ${file.path} (${file.nodeCount} symbols)`);
        } else {
          lines.push(`- ${file.path}`);
        }
      }
      lines.push('');
    }

    return lines.join('\n');
  }

  /**
   * Format files as a tree structure
   */
  private formatFilesTree(
    files: { path: string; language: string; nodeCount: number }[],
    includeMetadata: boolean,
    maxDepth?: number
  ): string {
    // Build tree structure
    interface TreeNode {
      name: string;
      children: Map<string, TreeNode>;
      file?: { language: string; nodeCount: number };
    }

    const root: TreeNode = { name: '', children: new Map() };

    for (const file of files) {
      const parts = file.path.split('/');
      let current = root;

      for (let i = 0; i < parts.length; i++) {
        const part = parts[i];
        if (!part) continue;

        if (!current.children.has(part)) {
          current.children.set(part, { name: part, children: new Map() });
        }
        current = current.children.get(part)!;

        // If this is the last part, it's a file
        if (i === parts.length - 1) {
          current.file = { language: file.language, nodeCount: file.nodeCount };
        }
      }
    }

    // Render tree
    const lines: string[] = [`## Project Structure (${files.length} files)`, ''];

    const renderNode = (node: TreeNode, prefix: string, isLast: boolean, depth: number): void => {
      if (maxDepth !== undefined && depth > maxDepth) return;

      const connector = isLast ? '└── ' : '├── ';
      const childPrefix = isLast ? '    ' : '│   ';

      if (node.name) {
        let line = prefix + connector + node.name;
        if (node.file && includeMetadata) {
          line += ` (${node.file.language}, ${node.file.nodeCount} symbols)`;
        }
        lines.push(line);
      }

      const children = [...node.children.values()];
      // Sort: directories first, then files, both alphabetically
      children.sort((a, b) => {
        const aIsDir = a.children.size > 0 && !a.file;
        const bIsDir = b.children.size > 0 && !b.file;
        if (aIsDir !== bIsDir) return aIsDir ? -1 : 1;
        return a.name.localeCompare(b.name);
      });

      for (let i = 0; i < children.length; i++) {
        const child = children[i]!;
        const nextPrefix = node.name ? prefix + childPrefix : prefix;
        renderNode(child, nextPrefix, i === children.length - 1, depth + 1);
      }
    };

    renderNode(root, '', true, 0);

    return lines.join('\n');
  }

  // =========================================================================
  // Symbol resolution helpers
  // =========================================================================

  /**
   * Find a symbol by name, handling disambiguation when multiple matches exist.
   * Returns the best match and a note about alternatives if any.
   */
  /**
   * Check if a node matches a symbol query.
   *
   * Accepts simple names (`run`) and three flavors of qualifier:
   *   - dotted     `Session.request`         (TS/JS/Python)
   *   - colon-pair `stage_apply::run`        (Rust, C++, Ruby)
   *   - slash      `configurator/stage_apply` (path-ish)
   *
   * Multi-level qualifiers compose: `crate::configurator::stage_apply::run`
   * works. Rust path prefixes (`crate`, `super`, `self`) are stripped so
   * the canonical `crate::module::symbol` form resolves.
   *
   * Resolution order, last part must always equal `node.name`:
   *   1. Suffix-match against `qualifiedName` (handles class-scoped methods
   *      where the extractor builds the qualified name from the AST stack)
   *   2. File-path containment (handles file-derived modules in Rust/
   *      Python — `stage_apply::run` matches a `run` in `stage_apply.rs`)
   */
  private matchesSymbol(node: Node, symbol: string): boolean {
    // Simple name match
    if (node.name === symbol) return true;
    // File basename match (e.g., "product-card" matches "product-card.liquid")
    if (node.kind === 'file' && node.name.replace(/\.[^.]+$/, '') === symbol) return true;

    // Qualified-name lookups: split on any supported separator. `\w` keeps
    // identifier chars (incl. `_`) intact; everything else is treated as
    // a separator we tolerate.
    if (!/[.\/]|::/.test(symbol)) return false;
    const parts = symbol.split(/::|[./]/).filter((p) => p.length > 0);
    if (parts.length < 2) return false;

    const lastPart = parts[parts.length - 1]!;
    if (node.name !== lastPart) return false;

    // Stage 1: qualified-name suffix match. The extractor joins the
    // semantic hierarchy with `::`, so `Session.request` and
    // `Session::request` both become `Session::request` here.
    const colonSuffix = parts.join('::');
    if (node.qualifiedName.includes(colonSuffix)) return true;

    // Stage 2: file-path containment. Rust modules and Python packages
    // are not in `qualifiedName` — they're encoded in the file path. So
    // `stage_apply::run` matches a `run` in any file whose path
    // contains a `stage_apply` segment (with or without an extension).
    //
    // Filter out Rust path prefixes that have no file-system equivalent.
    const containerHints = parts.slice(0, -1).filter((p) => !RUST_PATH_PREFIXES.has(p));
    if (containerHints.length === 0) return false;

    const segments = node.filePath.split('/').filter((s) => s.length > 0);
    return containerHints.every((hint) =>
      segments.some((seg) => seg === hint || seg.replace(/\.[^.]+$/, '') === hint)
    );
  }

  private findSymbol(cg: CodeGraph, symbol: string): { node: Node; note: string } | null {
    // Use higher limit for qualified lookups (e.g., "Session.request",
    // "stage_apply::run") since the target may rank lower in FTS when
    // there are many partial matches across the qualifier parts.
    const isQualified = /[.\/]|::/.test(symbol);
    const limit = isQualified ? 50 : 10;
    let results = cg.searchNodes(symbol, { limit });

    // FTS strips colons as a special char, so `stage_apply::run` searches
    // for the literal `stage_applyrun` and finds nothing. Re-search by
    // the bare last part and let `matchesSymbol` filter by qualifier.
    if (isQualified && results.length === 0) {
      const tail = lastQualifierPart(symbol);
      if (tail && tail !== symbol) results = cg.searchNodes(tail, { limit });
    }

    if (results.length === 0 || !results[0]) {
      return null;
    }

    const exactMatches = results.filter(r => this.matchesSymbol(r.node, symbol));

    if (exactMatches.length === 1) {
      return { node: exactMatches[0]!.node, note: '' };
    }

    if (exactMatches.length > 1) {
      // Multiple exact matches - pick first, note the others
      const picked = exactMatches[0]!.node;
      const others = exactMatches.slice(1).map(r =>
        `${r.node.name} (${r.node.kind}) at ${r.node.filePath}:${r.node.startLine}`
      );
      const note = `\n\n> **Note:** ${exactMatches.length} symbols named "${symbol}". Showing results for \`${picked.filePath}:${picked.startLine}\`. Others: ${others.join(', ')}`;
      return { node: picked, note };
    }

    // No exact match. For qualified lookups, don't silently fall back
    // to a fuzzy result — the user typed a specific qualifier, and
    // resolving `stage_apply::nonexistent_fn` to the unrelated
    // `stage_apply.rs` file would be actively misleading (#173).
    if (isQualified) return null;
    return { node: results[0]!.node, note: '' };
  }

  /**
   * Find ALL symbols matching a name. Used by callers/callees/impact to aggregate
   * results across all matching symbols (e.g., multiple classes with an `execute` method).
   */
  private findAllSymbols(cg: CodeGraph, symbol: string): { nodes: Node[]; note: string } {
    let results = cg.searchNodes(symbol, { limit: 50 });

    // Mirror the fallback in `findSymbol` for qualified queries — FTS
    // strips colons, so a module-qualified lookup needs a second pass
    // by the bare last part.
    if (results.length === 0 && /[.\/]|::/.test(symbol)) {
      const tail = lastQualifierPart(symbol);
      if (tail && tail !== symbol) results = cg.searchNodes(tail, { limit: 50 });
    }

    if (results.length === 0) {
      return { nodes: [], note: '' };
    }

    const exactMatches = results.filter(r => this.matchesSymbol(r.node, symbol));

    if (exactMatches.length <= 1) {
      const node = exactMatches[0]?.node ?? results[0]!.node;
      return { nodes: [node], note: '' };
    }

    const locations = exactMatches.map(r =>
      `${r.node.kind} at ${r.node.filePath}:${r.node.startLine}`
    );
    const note = `\n\n> **Note:** Aggregated results across ${exactMatches.length} symbols named "${symbol}": ${locations.join(', ')}`;
    return { nodes: exactMatches.map(r => r.node), note };
  }

  /**
   * Truncate output if it exceeds the maximum length
   */
  private truncateOutput(text: string): string {
    if (text.length <= MAX_OUTPUT_LENGTH) return text;
    const truncated = text.slice(0, MAX_OUTPUT_LENGTH);
    const lastNewline = truncated.lastIndexOf('\n');
    const cutPoint = lastNewline > MAX_OUTPUT_LENGTH * 0.8 ? lastNewline : MAX_OUTPUT_LENGTH;
    return truncated.slice(0, cutPoint) + '\n\n... (output truncated)';
  }

  // =========================================================================
  // Formatting helpers (compact by default to reduce context usage)
  // =========================================================================

  private formatSearchResults(results: SearchResult[]): string {
    const lines: string[] = [`## Search Results (${results.length} found)`, ''];

    for (const result of results) {
      const { node } = result;
      const location = node.startLine ? `:${node.startLine}` : '';
      // Compact format: one line per result with key info
      lines.push(`### ${node.name} (${node.kind})`);
      lines.push(`${node.filePath}${location}`);
      if (node.signature) lines.push(`\`${node.signature}\``);
      lines.push('');
    }

    return lines.join('\n');
  }

  private formatNodeList(nodes: Node[], title: string): string {
    const lines: string[] = [`## ${title} (${nodes.length} found)`, ''];

    for (const node of nodes) {
      const location = node.startLine ? `:${node.startLine}` : '';
      // Compact: just name, kind, location
      lines.push(`- ${node.name} (${node.kind}) - ${node.filePath}${location}`);
    }

    return lines.join('\n');
  }

  private formatImpact(symbol: string, impact: Subgraph): string {
    const nodeCount = impact.nodes.size;

    // Compact format: just list affected symbols grouped by file
    const lines: string[] = [
      `## Impact: "${symbol}" affects ${nodeCount} symbols`,
      '',
    ];

    // Group by file
    const byFile = new Map<string, Node[]>();
    for (const node of impact.nodes.values()) {
      const existing = byFile.get(node.filePath) || [];
      existing.push(node);
      byFile.set(node.filePath, existing);
    }

    for (const [file, nodes] of byFile) {
      lines.push(`**${file}:**`);
      // Compact: inline list
      const nodeList = nodes.map(n => `${n.name}:${n.startLine}`).join(', ');
      lines.push(nodeList);
      lines.push('');
    }

    return lines.join('\n');
  }

  /**
   * Build a compact structural outline of a container symbol from its
   * indexed children (methods, fields, properties, …) — name, kind,
   * line number, and signature — so the agent gets the shape of a class
   * without the full source of every method. Returns '' when the container
   * has no indexed children, so the caller can fall back to full source.
   */
  private buildContainerOutline(cg: CodeGraph, node: Node): string {
    const children = cg.getChildren(node.id)
      .filter(c => c.kind !== 'import' && c.kind !== 'export')
      .sort((a, b) => (a.startLine ?? 0) - (b.startLine ?? 0));
    if (children.length === 0) return '';

    const lines = [`**Members (${children.length}):**`, ''];
    for (const c of children) {
      const loc = c.startLine ? `:${c.startLine}` : '';
      const sig = c.signature ? ` — \`${c.signature}\`` : '';
      lines.push(`- ${c.name} (${c.kind})${loc}${sig}`);
    }
    return lines.join('\n');
  }

  private formatNodeDetails(node: Node, code: string | null, outline?: string | null): string {
    const location = node.startLine ? `:${node.startLine}` : '';
    const lines: string[] = [
      `## ${node.name} (${node.kind})`,
      '',
      `**Location:** ${node.filePath}${location}`,
    ];

    if (node.signature) {
      lines.push(`**Signature:** \`${node.signature}\``);
    }

    // Only include docstring if it's short and useful
    if (node.docstring && node.docstring.length < 200) {
      lines.push('', node.docstring);
    }

    if (outline) {
      lines.push('', outline, '',
        `> Structural outline only. Read \`${node.filePath}\` or call codegraph_node on a specific member for its body.`);
    } else if (code) {
      // Line-numbered (cat -n style, like codegraph_explore and Read) so the
      // agent can cite/edit exact lines without re-Reading the file for them.
      const numbered = node.startLine ? numberSourceLines(code, node.startLine) : code;
      lines.push('', '```' + node.language, numbered, '```');
    }

    return lines.join('\n');
  }

  private formatTaskContext(context: TaskContext): string {
    return context.summary || 'No context found';
  }

  private textResult(text: string): ToolResult {
    return {
      content: [{ type: 'text', text }],
    };
  }

  private errorResult(message: string): ToolResult {
    return {
      content: [{ type: 'text', text: `Error: ${message}` }],
      isError: true,
    };
  }
}