@specship/specship 0.11.0

package/dist/mcp/tools.js

@@ -0,0 +1,3179 @@
+"use strict";
+/**
+ * MCP Tool Definitions
+ *
+ * Defines the tools exposed by the SpecShip MCP server.
+ */
+Object.defineProperty(exports, "__esModule", { value: true });
+exports.ToolHandler = exports.tools = void 0;
+exports.getExploreBudget = getExploreBudget;
+exports.getExploreOutputBudget = getExploreOutputBudget;
+exports.formatStaleBanner = formatStaleBanner;
+exports.formatStaleFooter = formatStaleFooter;
+exports.getStaticTools = getStaticTools;
+const directory_1 = require("../directory");
+const statusline_1 = require("../statusline");
+// Lazy-load the heavy SpecShip chain off the MCP startup path — see the same
+// helper in engine.ts. ToolHandler must load to answer tools/list (static
+// schemas), but it must NOT drag in sqlite/query layers before the daemon binds;
+// SpecShip is pulled in only when a tool actually opens a project. require() is
+// sync + cached (CommonJS build).
+const loadSpecShip = () => require('../index').default;
+const worktree_1 = require("../sync/worktree");
+const query_utils_1 = require("../search/query-utils");
+const fs_1 = require("fs");
+const utils_1 = require("../utils");
+const generated_detection_1 = require("../extraction/generated-detection");
+const path_1 = require("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, `specship_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([
+    'class', 'struct', 'interface', 'trait', 'protocol', 'enum', 'namespace', 'module',
+]);
+/** Last `::` / `.` / `/`-separated segment of a qualified symbol. */
+function lastQualifierPart(symbol) {
+    const parts = symbol.split(/::|[./]/).filter((p) => p.length > 0);
+    return parts[parts.length - 1] ?? symbol;
+}
+/**
+ * Calculate the recommended number of specship_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.
+ */
+function getExploreBudget(fileCount) {
+    if (fileCount < 500)
+        return 1;
+    if (fileCount < 5000)
+        return 2;
+    if (fileCount < 15000)
+        return 3;
+    if (fileCount < 25000)
+        return 4;
+    return 5;
+}
+function getExploreOutputBudget(fileCount) {
+    // Tiered budget, scaled to project size. The budget is a CEILING (relevance
+    // still gates WHAT is included), and it MUST stay under the agent's INLINE
+    // tool-result cap (~25K chars). Above that, the host externalizes the result
+    // to a file the agent then Reads back — re-introducing a read AND the
+    // cache-write cost — which is exactly what a 35K vscode explore did in the
+    // n=4 README A/B. So even large repos cap at ~24K: the answer is the handful
+    // of ~100-line flow windows the agent would have grep-located and read (it
+    // natively reads ~6–9 files, median 100-line ranges), NOT a sprawl of 12
+    // files. Concentration onto the flow emerges from this cap + the named-file-
+    // first sort dropping peripheral files. Invariant: a larger tier must never
+    // get a smaller `maxCharsPerFile` than a smaller tier.
+    if (fileCount < 150) {
+        return {
+            // ITER3: revert iter2's aggressive body shrink (forced Read fallback —
+            // the per-file 2.5K cap pushed the agent to Read instead of node).
+            // Back to the iter1 shape (13K/4/3.8K) but keep the test-file
+            // hard-exclude. The cost lever for this tier lives in steering the
+            // agent to stop after 1-2 calls, not in this budget.
+            maxOutputChars: 13000,
+            defaultMaxFiles: 4,
+            maxCharsPerFile: 3800,
+            gapThreshold: 7,
+            maxSymbolsInFileHeader: 5,
+            maxEdgesPerRelationshipKind: 4,
+            includeRelationships: false,
+            includeAdditionalFiles: false,
+            includeCompletenessSignal: false,
+            includeBudgetNote: false,
+            excludeLowValueFiles: true,
+        };
+    }
+    if (fileCount < 500) {
+        return {
+            // ITER3: same revert/keep-filter pattern as <150.
+            maxOutputChars: 18000,
+            defaultMaxFiles: 5,
+            maxCharsPerFile: 3800,
+            gapThreshold: 8,
+            maxSymbolsInFileHeader: 6,
+            maxEdgesPerRelationshipKind: 6,
+            includeRelationships: false,
+            includeAdditionalFiles: false,
+            includeCompletenessSignal: false,
+            includeBudgetNote: false,
+            excludeLowValueFiles: true,
+        };
+    }
+    if (fileCount < 5000) {
+        return {
+            // ~150-line per-file window (the native read unit) × ~6 files, capped at
+            // the ~24K inline ceiling so the response is never externalized. Per-file
+            // stays ≥ the <500 tier (3800) — monotonic.
+            maxOutputChars: 24000,
+            defaultMaxFiles: 8,
+            maxCharsPerFile: 6500,
+            gapThreshold: 12,
+            maxSymbolsInFileHeader: 10,
+            maxEdgesPerRelationshipKind: 10,
+            includeRelationships: true,
+            includeAdditionalFiles: true,
+            includeCompletenessSignal: true,
+            includeBudgetNote: true,
+            excludeLowValueFiles: false,
+        };
+    }
+    // Large + very-large repos: SAME ~24K inline ceiling (a bigger response just
+    // externalizes — see vscode). More files indexed → more CALLS via
+    // getExploreBudget, not a bigger single response. Per-file 7000 (≥ smaller
+    // tiers) gives the central file a ~180-line orientation window.
+    if (fileCount < 15000) {
+        return {
+            maxOutputChars: 24000,
+            defaultMaxFiles: 8,
+            maxCharsPerFile: 7000,
+            gapThreshold: 15,
+            maxSymbolsInFileHeader: 15,
+            maxEdgesPerRelationshipKind: 15,
+            includeRelationships: true,
+            includeAdditionalFiles: true,
+            includeCompletenessSignal: true,
+            includeBudgetNote: true,
+            excludeLowValueFiles: false,
+        };
+    }
+    return {
+        maxOutputChars: 24000,
+        defaultMaxFiles: 8,
+        maxCharsPerFile: 7000,
+        gapThreshold: 15,
+        maxSymbolsInFileHeader: 15,
+        maxEdgesPerRelationshipKind: 15,
+        includeRelationships: true,
+        includeAdditionalFiles: true,
+        includeCompletenessSignal: true,
+        includeBudgetNote: true,
+        excludeLowValueFiles: false,
+    };
+}
+/**
+ * Whether `specship_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 `SPECSHIP_EXPLORE_LINENUMS=0` to disable (used by the
+ * A/B harness to measure the payload-cost vs. read-savings tradeoff).
+ */
+function exploreLineNumbersEnabled() {
+    return process.env.SPECSHIP_EXPLORE_LINENUMS !== '0';
+}
+/**
+ * Adaptive explore sizing (default ON). `specship_explore` skeletonizes OFF-SPINE
+ * polymorphic-sibling files — a file whose class is one of ≥3 interchangeable
+ * implementations of a shared interface (e.g. OkHttp's `: Interceptor` classes) —
+ * to class + member signatures (bodies elided), keeping the on-spine exemplar full.
+ * This sizes the response to the answer instead of the budget cap on sibling-heavy
+ * flows (OkHttp interceptor-chain explore 28.5k→16.6k, ~28% cheaper than native
+ * search, reads flat). It is PROVABLY INERT elsewhere: distinct pipeline steps (no
+ * ≥3-implementer supertype, e.g. Excalidraw's `renderStaticScene`) and on-spine
+ * files keep full source — output is byte-identical to shipped on excalidraw /
+ * tokio / django / vscode / gin. Set `SPECSHIP_ADAPTIVE_EXPLORE=0` to disable.
+ */
+function adaptiveExploreEnabled() {
+    return process.env.SPECSHIP_ADAPTIVE_EXPLORE !== '0' && process.env.SPECSHIP_ADAPTIVE_EXPLORE !== 'false';
+}
+/**
+ * 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, firstLineNumber) {
+    const out = [];
+    const split = slice.split('\n');
+    for (let i = 0; i < split.length; i++) {
+        out.push(`${firstLineNumber + i}\t${split[i]}`);
+    }
+    return out.join('\n');
+}
+/**
+ * Per-file staleness banner emitted at the top of a tool response when the
+ * file watcher has pending events for files referenced by the response.
+ * The agent uses this to fall back to Read for those specific files
+ * without waiting for the debounced sync (issue #403).
+ */
+function formatStaleBanner(stale) {
+    const now = Date.now();
+    const lines = stale.map((p) => {
+        const ageMs = Math.max(0, now - p.lastSeenMs);
+        const label = p.indexing ? 'indexing in progress' : 'pending sync';
+        return `  - ${p.path} (edited ${ageMs}ms ago, ${label})`;
+    });
+    return ('⚠️ Some files referenced below were edited since the last index sync — ' +
+        'their specship entries may be stale:\n' +
+        lines.join('\n') +
+        '\nFor accurate content of those specific files, Read them directly. ' +
+        'The rest of this response is fresh.');
+}
+/**
+ * Compact footer listing pending files that are NOT referenced in this
+ * response. Gives the agent a complete project-wide freshness picture
+ * without bloating the main banner.
+ */
+function formatStaleFooter(stale) {
+    const MAX = 5;
+    const now = Date.now();
+    const shown = stale.slice(0, MAX);
+    const lines = shown.map((p) => {
+        const ageMs = Math.max(0, now - p.lastSeenMs);
+        return `  - ${p.path} (edited ${ageMs}ms ago)`;
+    });
+    const more = stale.length > MAX ? `\n  - …and ${stale.length - MAX} more` : '';
+    return (`(Note: ${stale.length} file(s) elsewhere in this project are pending index ` +
+        `sync but were not referenced above:\n${lines.join('\n')}${more})`);
+}
+/**
+ * Common projectPath property for cross-project queries
+ */
+const projectPathProperty = {
+    type: 'string',
+    description: 'Path to a different project with .specship/ initialized. If omitted, uses current project. Use this to query other codebases.',
+};
+/**
+ * All SpecShip MCP tools
+ *
+ * Designed for minimal context usage - use specship_explore as the primary tool
+ * (one call usually answers the whole question), and only use other tools for
+ * targeted follow-up queries.
+ *
+ * All tools support cross-project queries via the optional `projectPath` parameter.
+ */
+const spec_tools_1 = require("./spec-tools");
+const maintainability_tool_1 = require("./maintainability-tool");
+const fitness_tool_1 = require("./fitness-tool");
+const designer_tools_1 = require("./designer-tools");
+exports.tools = [
+    {
+        name: 'specship_search',
+        description: 'Quick symbol search by name. Returns locations only (no code). Use specship_explore instead to get the actual source / understand an area in one call.',
+        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: 'specship_callers',
+        description: 'List functions that call <symbol>. For the full flow, use specship_explore.',
+        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: 'specship_callees',
+        description: 'List functions that <symbol> calls. For the full flow, use specship_explore.',
+        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: 'specship_impact',
+        description: 'List symbols affected by changing <symbol>. Use before a refactor.',
+        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: 'specship_node',
+        description: 'SECONDARY (after specship_explore): get ONE symbol in full — its location, signature, callers/callees trail, and verbatim body (includeCode=true). When the name is AMBIGUOUS (an overloaded method, or the same method name on different types), it returns EVERY matching definition\'s full body in a single call — so you never need to Read a file to find the specific overload you want. For a heavily-overloaded name, pass `file` (and/or `line`) to pin the exact definition — e.g. the `file:line` a trail or another tool already showed you. Reach for this when explore trimmed a body you need. Use specship_explore for several related symbols or the full flow.',
+        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,
+                },
+                file: {
+                    type: 'string',
+                    description: 'Optional: disambiguate an overloaded name to the definition in this file (path or basename, e.g. "harness.rs").',
+                },
+                line: {
+                    type: 'number',
+                    description: 'Optional: disambiguate to the definition at/around this line (use with the file:line a trail showed you).',
+                },
+                projectPath: projectPathProperty,
+            },
+            required: ['symbol'],
+        },
+    },
+    {
+        name: 'specship_explore',
+        description: 'PRIMARY TOOL — call FIRST for almost any question: how does X work, architecture, a bug, where/what is X, or surveying an area. Returns the verbatim source of the relevant symbols grouped by file in ONE capped call (Read-equivalent — do NOT re-open shown files). Query can be a natural-language question OR a bag of symbol/file names. Usually the ONLY call you need — answers without further search/node/Read/Grep.',
+        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 specship_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: 'specship_status',
+        description: 'Index health check (files / nodes / edges). Skip unless debugging.',
+        inputSchema: {
+            type: 'object',
+            properties: {
+                projectPath: projectPathProperty,
+            },
+        },
+    },
+    {
+        name: 'specship_files',
+        description: 'Indexed file tree with language + symbol counts. Faster than Glob for project layout.',
+        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,
+            },
+        },
+    },
+    // Spec-layer tools (v5): see ./spec-tools.ts for handlers.
+    ...spec_tools_1.specToolDefinitions,
+    ...maintainability_tool_1.maintainabilityToolDefinitions,
+    ...fitness_tool_1.fitnessToolDefinitions,
+    // Designer tools: claude.ai/design driving, vendored from @pro-vi/designer.
+    // See ./designer-tools.ts for handlers. Drives a debug Chrome over CDP.
+    ...designer_tools_1.designerToolDefinitions,
+];
+/**
+ * Allowlist-filtered tool definitions WITHOUT an engine — the static surface the
+ * proxy answers `tools/list` with before any project is open. Mirrors
+ * `ToolHandler.getTools()` in the no-SpecShip case (the dynamic per-repo budget
+ * note in a description only adds once `cg` is loaded; the schemas are static).
+ */
+function getStaticTools() {
+    const raw = process.env.SPECSHIP_MCP_TOOLS;
+    if (!raw || !raw.trim())
+        return exports.tools;
+    const allow = new Set(raw.split(',').map(s => s.trim().replace(/^specship_/, '')).filter(Boolean));
+    return allow.size ? exports.tools.filter(t => allow.has(t.name.replace(/^specship_/, ''))) : exports.tools;
+}
+/**
+ * Tool handler that executes tools against a SpecShip instance
+ *
+ * Supports cross-project queries via the projectPath parameter.
+ * Other projects are opened on-demand and cached for performance.
+ */
+class ToolHandler {
+    cg;
+    // Cache of opened SpecShip instances for cross-project queries
+    projectCache = 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.
+    defaultProjectHint = null;
+    // Per-start-path cache of the git worktree/index mismatch (issue #155). The
+    // mismatch is a fixed property of (where the request came from → which
+    // .specship/ it resolves to), so the up-to-two `git rev-parse` spawns run
+    // once and every later tool call reuses the result — never shelling out to
+    // git on the hot path. `undefined` = not computed yet; `null` = no mismatch.
+    worktreeMismatchCache = new Map();
+    // Gate that the MCP engine pokes after `cg.open()` so the first tool call
+    // blocks on the post-open filesystem reconcile (catch-up sync). Without
+    // this, a tool call that races past `catchUpSync()` serves rows for files
+    // that were deleted (or edited) while no MCP server was running — and the
+    // per-file staleness banner can't help, because `getPendingFiles()` is
+    // populated by the watcher, not by catch-up. Cleared on first await so
+    // subsequent calls don't pay any cost.
+    catchUpGate = null;
+    constructor(cg) {
+        this.cg = cg;
+    }
+    /**
+     * Update the default SpecShip instance (e.g. after lazy initialization)
+     */
+    setDefaultSpecShip(cg) {
+        this.cg = cg;
+        // Reset the per-session call marker once, when this server process first
+        // attaches its code graph (REQ-STATUSLINE-004). Idempotent per process.
+        (0, statusline_1.initSession)(cg.getProjectRoot());
+    }
+    /**
+     * Engine-only: register the catch-up sync promise so the next `execute()`
+     * call awaits it before serving. The handler swallows rejections (the
+     * engine logs them) so a sync failure never propagates as a tool error;
+     * we still want to serve a best-effort result over the same potentially-
+     * stale data, which is what would have happened without the gate.
+     */
+    setCatchUpGate(p) {
+        this.catchUpGate = p;
+    }
+    /**
+     * Record the directory the server tried to resolve the default project from.
+     * Used only to make the "no default project" error actionable.
+     */
+    setDefaultProjectHint(searchedPath) {
+        this.defaultProjectHint = searchedPath;
+    }
+    /**
+     * Whether a default SpecShip instance is available
+     */
+    hasDefaultSpecShip() {
+        return this.cg !== null;
+    }
+    /**
+     * Optional allowlist of exposed tools, parsed from the SPECSHIP_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 "node" and "specship_node" both work.
+     */
+    toolAllowlist() {
+        const raw = process.env.SPECSHIP_MCP_TOOLS;
+        if (!raw || !raw.trim())
+            return null;
+        const short = (s) => s.trim().replace(/^specship_/, '');
+        const set = new Set(raw.split(',').map(short).filter(Boolean));
+        return set.size ? set : null;
+    }
+    /** Whether a tool name passes the SPECSHIP_MCP_TOOLS allowlist (if any). */
+    isToolAllowed(name) {
+        const allow = this.toolAllowlist();
+        return !allow || allow.has(name.replace(/^specship_/, ''));
+    }
+    /**
+     * Get tool definitions with dynamic descriptions based on project size.
+     * The specship_explore tool description includes a budget recommendation
+     * scaled to the number of indexed files. Honors the SPECSHIP_MCP_TOOLS
+     * allowlist so a trimmed surface is reflected in ListTools.
+     */
+    getTools() {
+        const allow = this.toolAllowlist();
+        let visible = allow
+            ? exports.tools.filter(t => allow.has(t.name.replace(/^specship_/, '')))
+            : exports.tools;
+        if (!this.cg)
+            return visible;
+        try {
+            const stats = this.cg.getStats();
+            const budget = getExploreBudget(stats.fileCount);
+            // Tiny-repo tool gating: on projects under TINY_REPO_FILE_THRESHOLD
+            // files, only expose the 5 core tools (search, context, node,
+            // explore, trace). The 5 omitted tools (callers, callees, impact,
+            // status, files) reduce to one grep at this scale.
+            //
+            // n=2 audits ruled out cutting below 5 tools:
+            // - 3-tool gate (search + context + trace): cost regressed on
+            //   cobra/ky/sinatra. The agent fell back to raw Reads to cover
+            //   what specship_node + specship_explore would have answered.
+            // - 1-tool gate (search only): catastrophic regression — express
+            //   went from -43% WIN to +107% LOSS. With only search, the agent
+            //   can't navigate the call graph structurally and reads everything.
+            //
+            // 5 is the empirical lower bound. Tools beyond search/context/
+            // node/explore/trace pay overhead that the agent doesn't recoup
+            // on tiny-repo flow questions.
+            // ITER4: raise threshold 150 → 500 so single-file frameworks
+            // (sinatra at 159, slim_framework around 200) also get the
+            // 5-tool surface. The empirical 5-tool floor was set on <150
+            // probes; iter3 measurement showed sinatra is structurally the
+            // SAME problem as cobra (single-file WITHOUT-arm Read wins),
+            // so it deserves the same gating.
+            const TINY_REPO_FILE_THRESHOLD = 500;
+            const TINY_REPO_CORE_TOOLS = new Set([
+                'specship_explore',
+                'specship_search',
+                'specship_node',
+            ]);
+            if (stats.fileCount < TINY_REPO_FILE_THRESHOLD) {
+                // Designer tools are not code-graph tools — the tiny-repo flow-question
+                // economics that justify trimming don't apply, so they always survive.
+                visible = visible.filter(t => TINY_REPO_CORE_TOOLS.has(t.name) || t.name.startsWith('designer_'));
+            }
+            return visible.map(tool => {
+                if (tool.name === 'specship_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 SpecShip instance for a project
+     *
+     * If projectPath is provided, opens that project's SpecShip (cached).
+     * Otherwise returns the default SpecShip instance.
+     *
+     * Walks up parent directories to find the nearest .specship/ folder,
+     * similar to how git finds .git/ directories.
+     */
+    getSpecShip(projectPath) {
+        if (!projectPath) {
+            if (!this.cg) {
+                const searched = this.defaultProjectHint ?? process.cwd();
+                throw new Error('No SpecShip project is loaded for this session.\n' +
+                    `Searched for a .specship/ 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 .specship/
+        // root below (issue #238), so we don't run the existence-checking
+        // validator on paths that are meant to walk up.
+        if ((0, fs_1.existsSync)(projectPath)) {
+            const pathError = (0, utils_1.validateProjectPath)(projectPath);
+            if (pathError) {
+                throw new Error(pathError);
+            }
+        }
+        // Walk up parent directories to find nearest .specship/
+        const resolvedRoot = (0, directory_1.findNearestSpecShipRoot)(projectPath);
+        if (!resolvedRoot) {
+            throw new Error(`SpecShip not initialized in ${projectPath}. Run 'specship 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 = loadSpecShip().openSync(resolvedRoot);
+        this.projectCache.set(resolvedRoot, cg);
+        if (projectPath !== resolvedRoot) {
+            this.projectCache.set(projectPath, cg);
+        }
+        return cg;
+    }
+    /**
+     * Close all cached project connections
+     */
+    closeAll() {
+        for (const cg of this.projectCache.values()) {
+            cg.close();
+        }
+        this.projectCache.clear();
+        this.worktreeMismatchCache.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.
+     */
+    validateString(value, name, maxLength = MAX_INPUT_LENGTH) {
+        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 ${value.length})`);
+        }
+        return value;
+    }
+    /**
+     * Validate an optional path-like string input. Returns the value if
+     * valid (or undefined), or a ToolResult with the error.
+     */
+    validateOptionalPath(value, name) {
+        if (value === undefined || value === null)
+            return undefined;
+        if (typeof value !== 'string') {
+            return this.errorResult(`${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;
+    }
+    /**
+     * Cached git worktree/index mismatch for a tool call's effective project.
+     *
+     * The "effective project" is what the request targets: an explicit
+     * `projectPath` arg, else the directory the server resolved its default
+     * project from (`defaultProjectHint`), else cwd. Memoized per start path —
+     * see `worktreeMismatchCache`. Best-effort: if the project can't be resolved
+     * (e.g. nothing initialized yet), it reports "no mismatch" so a tool is never
+     * broken by this check.
+     */
+    worktreeMismatchFor(projectPath) {
+        const startPath = projectPath ?? this.defaultProjectHint ?? process.cwd();
+        const cached = this.worktreeMismatchCache.get(startPath);
+        if (cached !== undefined)
+            return cached;
+        let mismatch = null;
+        try {
+            mismatch = (0, worktree_1.detectWorktreeIndexMismatch)(startPath, this.getSpecShip(projectPath).getProjectRoot());
+        }
+        catch {
+            // No resolvable project (or any other resolution error) → nothing to warn.
+            mismatch = null;
+        }
+        this.worktreeMismatchCache.set(startPath, mismatch);
+        return mismatch;
+    }
+    /**
+     * Prefix a successful read-tool result with a compact worktree-mismatch
+     * notice when the resolved index belongs to a different git working tree than
+     * the caller's (issue #155). Without this, an agent in a nested worktree
+     * silently trusts main-branch results. No-op on error results and when there
+     * is no mismatch. `specship_status` is excluded — it embeds its own verbose
+     * warning — so it stays out of this path.
+     */
+    withWorktreeNotice(result, projectPath) {
+        if (result.isError)
+            return result;
+        const mismatch = this.worktreeMismatchFor(projectPath);
+        if (!mismatch)
+            return result;
+        const notice = (0, worktree_1.worktreeMismatchNotice)(mismatch);
+        const [first, ...rest] = result.content;
+        if (first && first.type === 'text') {
+            return { ...result, content: [{ type: 'text', text: `${notice}\n\n${first.text}` }, ...rest] };
+        }
+        return result;
+    }
+    /**
+     * Annotate a successful read-tool result with per-file staleness — the
+     * non-blocking answer to issue #403. The file watcher tracks every event
+     * it sees per path; here we intersect "files referenced in this response"
+     * against that pending set and prepend a compact banner so the agent can
+     * fall back to Read for those *specific* files without waiting for the
+     * debounced sync to fire. Other pending files in the project (not
+     * referenced by this response) get a small footer so the agent has a
+     * complete picture without bloating the banner.
+     *
+     * Cost when nothing is pending — the common case — is one boolean check.
+     * No I/O, no parsing of markdown beyond a per-pending-file substring scan.
+     */
+    withStalenessNotice(result, projectPath) {
+        if (result.isError)
+            return result;
+        let cg;
+        try {
+            cg = this.getSpecShip(projectPath);
+        }
+        catch {
+            return result; // no default project — leave as is
+        }
+        // Cross-project `projectPath` calls open a cached SpecShip WITHOUT a
+        // watcher (watchers are only attached to the default session project).
+        // When the cross-project path happens to be the same project as the
+        // default cg, the cached instance is the wrong one — its pendingFiles is
+        // permanently empty. Detect the equal-path case and prefer the default
+        // cg so the staleness signal still fires when an agent passes the
+        // explicit projectPath form of its own project.
+        if (this.cg && cg !== this.cg) {
+            try {
+                const sameProject = (0, path_1.resolve)(this.cg.getProjectRoot()) === (0, path_1.resolve)(cg.getProjectRoot());
+                if (sameProject)
+                    cg = this.cg;
+            }
+            catch {
+                /* getProjectRoot may throw on a closed instance — leave cg as is */
+            }
+        }
+        // Defensive: some test fakes inject a partial SpecShip stub without the
+        // newer pending-files API. Treat missing/throwing as "no pending files."
+        let pending = [];
+        try {
+            pending = cg.getPendingFiles?.() ?? [];
+        }
+        catch {
+            return result;
+        }
+        if (pending.length === 0)
+            return result;
+        const [first, ...rest] = result.content;
+        if (!first || first.type !== 'text')
+            return result;
+        const text = first.text;
+        const inResponse = [];
+        const elsewhere = [];
+        for (const p of pending) {
+            // Substring match against the project-relative POSIX path — that's
+            // exactly the format both the watcher and every specship response
+            // emit, so a plain includes() is sufficient and avoids regex pitfalls.
+            if (text.includes(p.path))
+                inResponse.push(p);
+            else
+                elsewhere.push(p);
+        }
+        let banner = '';
+        if (inResponse.length > 0) {
+            banner = formatStaleBanner(inResponse);
+        }
+        let footer = '';
+        if (elsewhere.length > 0) {
+            footer = formatStaleFooter(elsewhere);
+        }
+        if (!banner && !footer)
+            return result;
+        const composed = [banner, text, footer].filter(Boolean).join('\n\n');
+        return { ...result, content: [{ type: 'text', text: composed }, ...rest] };
+    }
+    /**
+     * Execute a tool by name
+     */
+    async execute(toolName, args) {
+        try {
+            // Block the first tool call on the engine's post-open reconcile so we
+            // never serve rows for files deleted/edited while no MCP server was
+            // running. The gate is cleared after first await — subsequent calls
+            // pay nothing. Catch-up failures are logged by the engine; we
+            // proceed regardless so a transient sync error never breaks tools.
+            // Record this lookup into the per-session status-line marker
+            // (REQ-STATUSLINE-004). Best-effort and only for specship_* tools. The
+            // ENTIRE block is guarded — including resolving the project root — so a
+            // marker step can never turn a real tool call into an error (A3); a mock
+            // or partially-initialized code graph must pass straight through.
+            if (toolName.startsWith('specship_') && this.cg) {
+                try {
+                    const root = this.cg.getProjectRoot();
+                    (0, statusline_1.initSession)(root); // idempotent per process — fixes the session start
+                    (0, statusline_1.recordCall)(root, toolName);
+                }
+                catch { /* never let status-line bookkeeping affect the tool call */ }
+            }
+            if (this.catchUpGate) {
+                const gate = this.catchUpGate;
+                this.catchUpGate = null;
+                try {
+                    await gate;
+                }
+                catch { /* engine already logged */ }
+            }
+            // Honor the optional tool allowlist (SPECSHIP_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 SPECSHIP_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 specship_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;
+            }
+            // Read tools resolve through a single result variable so cross-cutting
+            // notices — worktree-index mismatch (issue #155) and per-file
+            // staleness (issue #403) — can be applied in one place. status embeds
+            // its own verbose worktree warning but still flows through the
+            // staleness wrapper so its pending-files section stays consistent
+            // with what the read tools surface.
+            let result;
+            switch (toolName) {
+                case 'specship_search':
+                    result = await this.handleSearch(args);
+                    break;
+                case 'specship_callers':
+                    result = await this.handleCallers(args);
+                    break;
+                case 'specship_callees':
+                    result = await this.handleCallees(args);
+                    break;
+                case 'specship_impact':
+                    result = await this.handleImpact(args);
+                    break;
+                case 'specship_explore':
+                    result = await this.handleExplore(args);
+                    break;
+                case 'specship_node':
+                    result = await this.handleNode(args);
+                    break;
+                case 'specship_status':
+                    // status embeds the pending-files list as a first-class section
+                    // (see handleStatus), so we skip the auto-banner wrapper here to
+                    // avoid duplicating the same info at the top of the response.
+                    return await this.handleStatus(args);
+                case 'specship_files':
+                    result = await this.handleFiles(args);
+                    break;
+                case 'specship_spec':
+                    result = await (0, spec_tools_1.handleSpecshipSpec)(this.getSpecShip(args.projectPath), args);
+                    break;
+                case 'specship_link_assert':
+                    result = await (0, spec_tools_1.handleSpecshipLinkAssert)(this.getSpecShip(args.projectPath), args);
+                    break;
+                case 'specship_link_verify':
+                    result = await (0, spec_tools_1.handleSpecshipLinkVerify)(this.getSpecShip(args.projectPath), args);
+                    break;
+                case 'specship_drifted':
+                    result = await (0, spec_tools_1.handleSpecshipDrifted)(this.getSpecShip(args.projectPath), args);
+                    break;
+                case 'specship_maintainability':
+                    result = await (0, maintainability_tool_1.handleSpecshipMaintainability)(this.getSpecShip(args.projectPath), args);
+                    break;
+                case 'specship_fitness':
+                    result = await (0, fitness_tool_1.handleSpecshipFitness)(this.getSpecShip(args.projectPath), args);
+                    break;
+                // Designer tools are independent of the code graph — return directly,
+                // bypassing the worktree/staleness notice wrappers (which assume a
+                // code-graph query).
+                case 'designer_session':
+                    return await (0, designer_tools_1.handleDesignerSession)(args);
+                case 'designer_prompt':
+                    return await (0, designer_tools_1.handleDesignerPrompt)(args);
+                case 'designer_ask':
+                    return await (0, designer_tools_1.handleDesignerAsk)(args);
+                case 'designer_list':
+                    return await (0, designer_tools_1.handleDesignerList)(args);
+                case 'designer_snapshot':
+                    return await (0, designer_tools_1.handleDesignerSnapshot)(args);
+                case 'designer_handoff':
+                    return await (0, designer_tools_1.handleDesignerHandoff)(args);
+                default:
+                    return this.errorResult(`Unknown tool: ${toolName}`);
+            }
+            const withWorktree = this.withWorktreeNotice(result, args.projectPath);
+            return this.withStalenessNotice(withWorktree, args.projectPath);
+        }
+        catch (err) {
+            return this.errorResult(`Tool execution failed: ${err instanceof Error ? err.message : String(err)}`);
+        }
+    }
+    /**
+     * Handle specship_search
+     */
+    async handleSearch(args) {
+        const query = this.validateString(args.query, 'query');
+        if (typeof query !== 'string')
+            return query;
+        const cg = this.getSpecShip(args.projectPath);
+        const kind = args.kind;
+        const rawLimit = Number(args.limit) || 10;
+        const limit = (0, utils_1.clamp)(rawLimit, 1, 100);
+        const results = cg.searchNodes(query, {
+            limit,
+            kinds: kind ? [kind] : undefined,
+        });
+        if (results.length === 0) {
+            return this.textResult(`No results found for "${query}"`);
+        }
+        // Down-rank generated files within the FTS-returned set so a search
+        // for "Send" surfaces the hand-written keeper before .pb.go stubs
+        // that share the name. Stable: only reorders generated vs. not.
+        const ranked = [...results].sort((a, b) => {
+            const aGen = (0, generated_detection_1.isGeneratedFile)(a.node.filePath) ? 1 : 0;
+            const bGen = (0, generated_detection_1.isGeneratedFile)(b.node.filePath) ? 1 : 0;
+            return aGen - bGen;
+        });
+        const formatted = this.formatSearchResults(ranked);
+        return this.textResult(this.truncateOutput(formatted));
+    }
+    /**
+     * Handle specship_callers
+     */
+    async handleCallers(args) {
+        const symbol = this.validateString(args.symbol, 'symbol');
+        if (typeof symbol !== 'string')
+            return symbol;
+        const cg = this.getSpecShip(args.projectPath);
+        const limit = (0, utils_1.clamp)(args.limit || 20, 1, 100);
+        const allMatches = this.findAllSymbols(cg, symbol);
+        if (allMatches.nodes.length === 0) {
+            return this.textResult(`Symbol "${symbol}" not found in the codebase`);
+        }
+        // Aggregate callers across all matching symbols
+        const seen = new Set();
+        const allCallers = [];
+        for (const node of allMatches.nodes) {
+            for (const c of cg.getCallers(node.id)) {
+                if (!seen.has(c.node.id)) {
+                    seen.add(c.node.id);
+                    allCallers.push(c.node);
+                }
+            }
+        }
+        if (allCallers.length === 0) {
+            return this.textResult(`No callers found for "${symbol}"${allMatches.note}`);
+        }
+        const formatted = this.formatNodeList(allCallers.slice(0, limit), `Callers of ${symbol}`) + allMatches.note;
+        return this.textResult(this.truncateOutput(formatted));
+    }
+    /**
+     * Handle specship_callees
+     */
+    async handleCallees(args) {
+        const symbol = this.validateString(args.symbol, 'symbol');
+        if (typeof symbol !== 'string')
+            return symbol;
+        const cg = this.getSpecShip(args.projectPath);
+        const limit = (0, utils_1.clamp)(args.limit || 20, 1, 100);
+        const allMatches = this.findAllSymbols(cg, symbol);
+        if (allMatches.nodes.length === 0) {
+            return this.textResult(`Symbol "${symbol}" not found in the codebase`);
+        }
+        // Aggregate callees across all matching symbols
+        const seen = new Set();
+        const allCallees = [];
+        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 ${symbol}`) + allMatches.note;
+        return this.textResult(this.truncateOutput(formatted));
+    }
+    /**
+     * Handle specship_impact
+     */
+    async handleImpact(args) {
+        const symbol = this.validateString(args.symbol, 'symbol');
+        if (typeof symbol !== 'string')
+            return symbol;
+        const cg = this.getSpecShip(args.projectPath);
+        const depth = (0, utils_1.clamp)(args.depth || 2, 1, 10);
+        const allMatches = this.findAllSymbols(cg, symbol);
+        if (allMatches.nodes.length === 0) {
+            return this.textResult(`Symbol "${symbol}" not found in the codebase`);
+        }
+        // Aggregate impact across all matching symbols
+        const mergedNodes = new Map();
+        const mergedEdges = [];
+        const seenEdges = new Set();
+        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}:${e.kind}`;
+                if (!seenEdges.has(key)) {
+                    seenEdges.add(key);
+                    mergedEdges.push(e);
+                }
+            }
+        }
+        const mergedImpact = {
+            nodes: mergedNodes,
+            edges: mergedEdges,
+            roots: allMatches.nodes.map(n => n.id),
+        };
+        const formatted = this.formatImpact(symbol, mergedImpact) + allMatches.note;
+        return this.textResult(this.truncateOutput(formatted));
+    }
+    /**
+     * Describe a synthesized (dynamic-dispatch) edge for human output: how the
+     * callback was wired up — the bridge static parsing can't see. Returns null
+     * for ordinary static edges. Used by trace + the node trail so a synthesized
+     * hop reads as "registered via onUpdate at App.tsx:3148", not a bare arrow.
+     */
+    synthEdgeNote(edge) {
+        if (!edge || edge.provenance !== 'heuristic')
+            return null;
+        const m = edge.metadata;
+        const registeredAt = typeof m?.registeredAt === 'string' ? m.registeredAt : undefined;
+        const at = registeredAt ? ` @${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,
+            };
+        }
+        if (m?.synthesizedBy === 'closure-collection') {
+            const field = m.field ? `\`${String(m.field)}\`` : 'a collection';
+            return {
+                label: `closure collection — runs handlers appended to ${field} (dynamic dispatch)`,
+                compact: `dynamic: runs ${field} handlers${at}`,
+                registeredAt,
+            };
+        }
+        return null;
+    }
+    /**
+     * Flow-from-named-symbols: an agent's specship_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`.
+     */
+    buildFlowFromNamedSymbols(cg, query) {
+        const EMPTY = { text: '', pathNodeIds: new Set(), namedNodeIds: new Set(), uniqueNamedNodeIds: new Set() };
+        try {
+            const CALLABLE = new Set(['method', 'function', 'component', 'constructor']);
+            // Strip only a REAL file extension (Create.cs → Create); KEEP qualified
+            // names (Class.method / Class::method) — the agent's most precise input,
+            // resolved exactly by findAllSymbols. (The old strip mangled Class.method
+            // into Class, throwing the method away.)
+            const FILE_EXT = /\.(?:java|kt|kts|ts|tsx|js|jsx|mjs|cjs|cs|py|go|rb|php|swift|rs|cpp|cc|cxx|c|h|hpp|scala|lua|dart|vue|svelte)$/i;
+            const tokens = [...new Set(query.split(/[\s,()[\]]+/)
+                    .map((t) => t.replace(FILE_EXT, '').trim())
+                    .filter((t) => t.length >= 3 && /^[A-Za-z_$][\w$]*(?:(?:::|\.)[\w$]+)*$/.test(t)))].slice(0, 16);
+            if (tokens.length < 2)
+                return EMPTY;
+            // Pool of name SEGMENTS (Class + method from every token) used to
+            // disambiguate an ambiguous SIMPLE name: keep a candidate only if its
+            // CONTAINER class is itself named in the query.
+            const segPool = new Set();
+            for (const t of tokens)
+                for (const s of t.toLowerCase().split(/::|\./))
+                    if (s)
+                        segPool.add(s);
+            const named = new Map();
+            // Nodes whose token is SPECIFIC — a (near-)unique callable name (<=3 defs in
+            // the whole graph). These are safe to SPARE a file on: the agent named THIS
+            // method (`getResponseWithInterceptorChain`, 1 def). A hyper-polymorphic name
+            // (`as_sql`, 110 defs across every Expression/Compiler subclass) is NOT here,
+            // so naming it doesn't keep every backend variant full and flood the budget.
+            const uniqueNamedNodeIds = new Set();
+            for (const t of tokens) {
+                const cands = this.findAllSymbols(cg, t).nodes.filter((n) => CALLABLE.has(n.kind));
+                // A qualified or otherwise-specific name (<=3 hits) keeps all; an
+                // ambiguous simple name keeps only candidates whose container is named.
+                const specific = cands.length <= 3;
+                const pick = specific
+                    ? cands
+                    : cands.filter((n) => {
+                        const segs = (n.qualifiedName || '').toLowerCase().split(/::|\./).filter(Boolean);
+                        const container = segs.length >= 2 ? segs[segs.length - 2] : '';
+                        return !!container && segPool.has(container);
+                    });
+                for (const n of pick.slice(0, 6)) {
+                    named.set(n.id, n);
+                    if (specific)
+                        uniqueNamedNodeIds.add(n.id);
+                }
+                if (named.size > 40)
+                    break;
+            }
+            if (named.size < 2)
+                return EMPTY;
+            const MAX_HOPS = 7;
+            let best = 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();
+                parent.set(seed.id, { prev: null, edge: null, node: seed });
+                const q = [{ id: seed.id, depth: 0, streak: 0 }];
+                let deep = 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 = [];
+                let cur = 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;
+            }
+            const hasMain = !!best && best.length >= 3;
+            const pathIds = new Set((best ?? []).map((s) => s.node.id));
+            // Supplementary: dynamic-dispatch (synthesized) edges incident to a NAMED
+            // symbol — the indirect hops an agent would otherwise grep/Read to
+            // reconstruct ("where do the appended `validators` actually run?"). The
+            // synth edge IS that answer, so surface it even when the OTHER end wasn't
+            // named (e.g. the agent names `validate` but not the `didCompleteTask`
+            // that drains the collection). On-topic by construction: only heuristic
+            // edges touching a symbol the agent named; skipped when the hop already
+            // shows in the main chain.
+            const synthLines = [];
+            const synthSeen = new Set();
+            for (const n of named.values()) {
+                if (synthLines.length >= 6)
+                    break;
+                for (const { node: other, edge } of [...cg.getCallers(n.id), ...cg.getCallees(n.id)]) {
+                    if (synthLines.length >= 6)
+                        break;
+                    if (edge.provenance !== 'heuristic' || other.id === n.id)
+                        continue;
+                    if (pathIds.has(edge.source) && pathIds.has(edge.target))
+                        continue; // already in the main chain
+                    const src = edge.source === n.id ? n : other;
+                    const tgt = edge.source === n.id ? other : n;
+                    const key = `${src.name}>${tgt.name}`;
+                    if (synthSeen.has(key))
+                        continue;
+                    synthSeen.add(key);
+                    const note = this.synthEdgeNote(edge);
+                    synthLines.push(`- ${src.name} → ${tgt.name}   [${note ? note.compact : edge.kind}]`);
+                }
+            }
+            if (!hasMain && synthLines.length === 0)
+                return EMPTY;
+            const out = [];
+            if (hasMain) {
+                out.push('## 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('');
+            }
+            if (synthLines.length) {
+                out.push('## Dynamic-dispatch links among your symbols', '(synthesized — the indirect hops grep/Read would reconstruct; the `@file:line` is the wiring site)', '', ...synthLines, '');
+            }
+            out.push('> Full source for these symbols is below — the call flow among them, followed by their bodies.', '');
+            // namedNodeIds = every callable the agent explicitly named (a superset of
+            // the spine). A file holding one is something the agent asked to SEE, so it
+            // must keep full source even if it's an off-spine polymorphic sibling — the
+            // agent named `getResponseWithInterceptorChain` / `SQLCompiler.execute_sql`
+            // as the mechanism, not as an interchangeable leaf. See the skeleton gate.
+            return { text: out.join('\n'), pathNodeIds: pathIds, namedNodeIds: new Set(named.keys()), uniqueNamedNodeIds };
+        }
+        catch {
+            return EMPTY;
+        }
+    }
+    /**
+     * Compact "blast radius" for the entry symbols of an explore result: who
+     * depends on each (callers) and which test files cover it — LOCATIONS ONLY,
+     * no source, so the agent knows what to update / re-verify before editing
+     * without reaching for a separate impact call. Always-on, but skips symbols
+     * that have no dependents (nothing to warn about), and returns '' when none
+     * qualify so a leaf-only exploration stays clean.
+     */
+    /**
+     * Domain-facts section for `specship_explore` (REQ-DOMAIN-005.A2). Domain
+     * facts live in the spec layer, not the node graph, so explore's symbol
+     * search never reaches them. When the query names a documented domain term
+     * or entity, surface the matching human-confirmed fact bodies inline so the
+     * agent gets the authoritative project semantics without a second call.
+     *
+     * Additive + silent: returns '' unless a domain fact matches a query token,
+     * so it never bloats a normal code exploration (the hot-path guard in
+     * CLAUDE.md). Facts are few and human-authored, so a linear scan + token
+     * match is cheap and precise enough.
+     */
+    buildDomainFactsSection(cg, query) {
+        let facts;
+        try {
+            facts = cg.getSpecQueries().getSpecsByKind('domain');
+        }
+        catch {
+            return '';
+        }
+        if (!facts || facts.length === 0)
+            return '';
+        // Tokens of length >= 4 keep the match precise — short words ("the", "id")
+        // would match almost any fact body.
+        const tokens = [...new Set(query.toLowerCase().split(/[^a-z0-9_]+/i).filter((t) => t.length >= 4))];
+        if (tokens.length === 0)
+            return '';
+        const matched = facts
+            .filter((f) => {
+            const hay = `${f.title} ${f.body ?? ''}`.toLowerCase();
+            return tokens.some((t) => hay.includes(t));
+        })
+            .slice(0, 5);
+        if (matched.length === 0)
+            return '';
+        const out = [
+            '### Domain facts',
+            '',
+            '_Human-confirmed domain knowledge matching this query (from the spec layer, not code). Treat as authoritative project context — no need to look further for these terms._',
+            '',
+        ];
+        for (const f of matched) {
+            const factType = f.metadata?.type;
+            const typeSuffix = typeof factType === 'string' ? ` · ${factType}` : '';
+            out.push(`#### ${f.id} — ${f.title}${typeSuffix}`);
+            const body = (f.body ?? '').trim();
+            out.push(body.length > 800 ? `${body.slice(0, 800)}…` : body || '_(empty)_');
+            out.push('');
+        }
+        return out.join('\n');
+    }
+    buildBlastRadiusSection(cg, subgraph) {
+        const ROOT_CAP = 5; // only the symbols the query actually targeted
+        const FILE_CAP = 4; // caller files listed per symbol before "+N more"
+        const MEANINGFUL = new Set([
+            'function', 'method', 'class', 'interface', 'struct', 'trait', 'protocol',
+            'enum', 'type_alias', 'component', 'constant', 'variable', 'property', 'field',
+        ]);
+        const rel = (p) => p.replace(/\\/g, '/');
+        const roots = subgraph.roots
+            .map((id) => subgraph.nodes.get(id))
+            .filter((n) => !!n && MEANINGFUL.has(n.kind))
+            .slice(0, ROOT_CAP);
+        if (roots.length === 0)
+            return '';
+        const entries = [];
+        for (const root of roots) {
+            let callers = [];
+            try {
+                callers = cg.getCallers(root.id);
+            }
+            catch { /* skip this root */ }
+            const seen = new Set();
+            const uniq = [];
+            for (const c of callers) {
+                if (c?.node && !seen.has(c.node.id)) {
+                    seen.add(c.node.id);
+                    uniq.push(c.node);
+                }
+            }
+            if (uniq.length === 0)
+                continue; // no blast radius → nothing to flag
+            const callerFiles = [...new Set(uniq.map((n) => rel(n.filePath)))];
+            const testFiles = callerFiles.filter((f) => (0, query_utils_1.isTestFile)(f));
+            const nonTest = callerFiles.filter((f) => !(0, query_utils_1.isTestFile)(f));
+            const shown = nonTest.slice(0, FILE_CAP).map((f) => `\`${f}\``).join(', ');
+            const more = nonTest.length > FILE_CAP ? ` +${nonTest.length - FILE_CAP} more` : '';
+            const where = nonTest.length > 0 ? ` in ${shown}${more}` : '';
+            const tests = testFiles.length > 0
+                ? `; tests: ${testFiles.slice(0, FILE_CAP).map((f) => `\`${f}\``).join(', ')}${testFiles.length > FILE_CAP ? ` +${testFiles.length - FILE_CAP}` : ''}`
+                : '; ⚠️ no covering tests found';
+            entries.push(`- \`${root.name}\` (${rel(root.filePath)}:${root.startLine}) — ${uniq.length} caller${uniq.length === 1 ? '' : 's'}${where}${tests}`);
+        }
+        if (entries.length === 0)
+            return '';
+        return [
+            '### Blast radius — what depends on these (update/verify before editing)',
+            '',
+            ...entries,
+            '',
+        ].join('\n');
+    }
+    /**
+     * Graph-connectivity relevance via Random-Walk-with-Restart (personalized
+     * PageRank) from the query's matched SEED nodes over the call/reference graph.
+     *
+     * This is the ranking signal text search (FTS/bm25) CANNOT provide, and it's
+     * specship's home turf: relevance by STRUCTURE, not words. A file whose
+     * symbols are call-connected to the matched cluster accrues walk mass and
+     * ranks high; a lone TEXT match — e.g. `LensSwitcher.swift` matched the word
+     * "switch" from `switchOrganization`, but calls none of `setUser`/`fetchUser`
+     * — gets only its own restart probability and ranks ~0. Immune to the
+     * tokenization trap that fools term matching, deterministic, no embeddings.
+     *
+     * Undirected adjacency (reachability both ways), restart α=0.25 to the seeds,
+     * power iteration to convergence. Bounded to the already-relevant subgraph, so
+     * it's a few hundred nodes × ~25 iterations — negligible cost.
+     */
+    computeGraphRelevance(nodeIds, edges, seedIds) {
+        const out = new Map();
+        const n = nodeIds.length;
+        if (n === 0)
+            return out;
+        const idx = new Map();
+        for (let i = 0; i < n; i++)
+            idx.set(nodeIds[i], i);
+        const RANK_EDGES = new Set([
+            'calls', 'references', 'extends', 'implements', 'overrides',
+            'instantiates', 'returns', 'type_of', 'imports',
+        ]);
+        const adj = Array.from({ length: n }, () => []);
+        for (const e of edges) {
+            if (!RANK_EDGES.has(e.kind))
+                continue;
+            const i = idx.get(e.source);
+            const j = idx.get(e.target);
+            if (i === undefined || j === undefined || i === j)
+                continue;
+            adj[i].push(j);
+            adj[j].push(i); // undirected — reachable either direction
+        }
+        // Restart vector: uniform over seeds present in the candidate set. (Falls
+        // back to uniform-over-all if no seed landed in the set, so we never return
+        // all-zero.)
+        const r = new Array(n).fill(0);
+        let rsum = 0;
+        for (const id of seedIds) {
+            const i = idx.get(id);
+            if (i !== undefined) {
+                r[i] = 1;
+                rsum += 1;
+            }
+        }
+        if (rsum === 0) {
+            for (let i = 0; i < n; i++)
+                r[i] = 1;
+            rsum = n;
+        }
+        for (let i = 0; i < n; i++)
+            r[i] /= rsum;
+        const alpha = 0.25;
+        let s = r.slice();
+        for (let iter = 0; iter < 25; iter++) {
+            const next = new Array(n).fill(0);
+            for (let i = 0; i < n; i++) {
+                const si = s[i];
+                if (si === 0)
+                    continue;
+                const d = adj[i].length;
+                if (d === 0) {
+                    next[i] += si;
+                    continue;
+                } // dangling: keep its mass
+                const share = si / d;
+                for (const j of adj[i])
+                    next[j] += share;
+            }
+            for (let i = 0; i < n; i++)
+                s[i] = (1 - alpha) * next[i] + alpha * r[i];
+        }
+        for (let i = 0; i < n; i++)
+            out.set(nodeIds[i], s[i]);
+        return out;
+    }
+    /**
+     * Handle specship_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 specship_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.
+     */
+    async handleExplore(args) {
+        const query = this.validateString(args.query, 'query');
+        if (typeof query !== 'string')
+            return query;
+        const cg = this.getSpecShip(args.projectPath);
+        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;
+        try {
+            budget = getExploreOutputBudget(cg.getStats().fileCount);
+        }
+        catch {
+            budget = getExploreOutputBudget(Infinity);
+        }
+        const maxFiles = (0, utils_1.clamp)(args.maxFiles || 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) {
+            // No code matched — but a documented domain term may still have a fact
+            // (it lives in the spec layer, not the node graph). Surface it so naming
+            // a pure domain concept isn't a dead end (REQ-DOMAIN-005.A2).
+            const domainOnly = this.buildDomainFactsSection(cg, query);
+            if (domainOnly) {
+                return this.textResult(`## Exploration: ${query}\n\n${domainOnly}`);
+            }
+            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();
+        const subgraphFiles = new Set();
+        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 = [];
+            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);
+            }
+        }
+        // Named-symbol seeding: findRelevantContext is an FTS/text rank, so a query
+        // that's a BAG of symbol names skewed toward one phase (Alamofire: 5 build
+        // terms, each a high-frequency name, vs 3 validate terms) lets the
+        // lower-frequency names fall below the search cut — their definitions, and
+        // whole files (Validation.swift), never get gathered, so they can never
+        // render and the agent Reads them. Resolve EACH named token to its
+        // substantive definition (skip empty stubs + test files, same relevance the
+        // trace endpoint picker uses) and inject it as an entry, so every symbol the
+        // agent explicitly named is in the subgraph and its file is scored.
+        const namedSeedIds = new Set();
+        {
+            const FILE_EXT = /\.(?:java|kt|kts|ts|tsx|js|jsx|mjs|cjs|cs|py|go|rb|php|swift|rs|cpp|cc|cxx|c|h|hpp|scala|lua|dart|vue|svelte)$/i;
+            const CALLABLE = new Set(['method', 'function', 'component', 'constructor']);
+            const isTestPath = (p) => /(^|\/)(tests?|specs?|__tests__|testdata|mocks?|fixtures?)\//i.test(p) || /\.(test|spec)\.[a-z]+$/i.test(p);
+            const bodyLines = (n) => Math.max(0, (n.endLine ?? n.startLine) - n.startLine);
+            const tokens = [...new Set(query.split(/[\s,()[\]]+/)
+                    .map((t) => t.replace(FILE_EXT, '').trim())
+                    .filter((t) => t.length >= 3 && /^[A-Za-z_$][\w$]*(?:(?:::|\.)[\w$]+)*$/.test(t)))].slice(0, 16);
+            // PascalCase tokens in the query are type/file disambiguators — when the
+            // agent writes "DataRequest task validate", the `task`/`validate` it wants
+            // are DataRequest's, NOT the same-named overloads in Validation.swift /
+            // Concurrency.swift / the abstract base. Used below to bias overloaded
+            // names toward the file/class the query also names.
+            const typeTokens = tokens.filter((o) => /^[A-Z][A-Za-z0-9]{3,}/.test(o));
+            const inNamedContext = (n) => typeTokens.some((ct) => {
+                const lc = ct.toLowerCase();
+                return n.filePath.toLowerCase().includes(lc) || n.qualifiedName.toLowerCase().includes(lc);
+            });
+            for (const t of tokens) {
+                // Enumerate ALL defs of a bare token via the direct index, not FTS — a
+                // 50+-overload name (tokio `poll`) ranks the wanted def (`Harness::poll`)
+                // below the FTS cut, so findAllSymbols would never see it and the
+                // type-token bias below couldn't pick the harness.rs one. (Same fix as
+                // specship_node's findSymbolMatches.) Qualified tokens keep findAllSymbols.
+                const isQual = /[.\/]|::/.test(t);
+                const raw = isQual ? this.findAllSymbols(cg, t).nodes : cg.getNodesByName(t);
+                const cands = raw
+                    .filter((n) => CALLABLE.has(n.kind) && !isTestPath(n.filePath))
+                    .sort((a, b) => (bodyLines(b) > 1 ? 1 : 0) - (bodyLines(a) > 1 ? 1 : 0) || bodyLines(b) - bodyLines(a));
+                // A specific name (<=3 defs) injects all its defs. An overloaded name
+                // (`validate` = 10, `request` = 44) would flood the subgraph, so inject
+                // only: the overloads whose file/class the query ALSO names (the agent
+                // told us which one it wants — DataRequest's, not Validation.swift's),
+                // capped; else fall back to the single most-substantive def. This is the
+                // explore-side mirror of specship_node's overload disambiguation.
+                let picks;
+                if (cands.length <= 3) {
+                    picks = cands;
+                }
+                else {
+                    const ctx = cands.filter(inNamedContext);
+                    picks = ctx.length > 0 ? ctx.slice(0, 4) : cands.slice(0, 1);
+                }
+                for (const n of picks) {
+                    if (!subgraph.nodes.has(n.id))
+                        subgraph.nodes.set(n.id, n);
+                    // Mark as a named seed EVEN IF the FTS gather already had it — being
+                    // "named by the agent" is independent of whether search happened to
+                    // surface it, and it drives the +50 score, the gate, and the
+                    // named-file sort below. (Previously only NEW injections were marked,
+                    // so a named symbol FTS already gathered never sorted to the top.)
+                    namedSeedIds.add(n.id);
+                }
+            }
+        }
+        // Step 2: Group nodes by file, score by relevance
+        const fileGroups = new Map();
+        const entryNodeIds = new Set([...subgraph.roots, ...namedSeedIds]);
+        // Build a set of nodes directly connected to entry points (depth 1)
+        const connectedToEntry = new Set();
+        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: a NAMED-SEED node (a symbol the agent named that FTS missed, now
+            // injected) is worth far more than a mere reference — its file is where the
+            // answer lives. Without this, an incidental file that name-drops the flow
+            // (Combine.swift references request/task → score 23 from connected nodes)
+            // outranks the file that DEFINES a named symbol (Validation.swift's
+            // `validate` → 10) and steals its render slot. Definition ≫ reference.
+            if (namedSeedIds.has(node.id)) {
+                group.score += 50;
+            }
+            else 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
+        let 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);
+        // Test/spec/icon/i18n file detector — used both for the pre-sort hard
+        // filter (tiny tier) and the comparator deprioritization (all tiers).
+        const isLowValue = (p) => {
+            const lp = p.toLowerCase();
+            return (/\/(tests?|__tests?__|spec)\//.test(lp) ||
+                /_test\.go$/.test(lp) ||
+                /(?:^|\/)test_[^/]+\.py$/.test(lp) ||
+                /_test\.py$/.test(lp) ||
+                /_spec\.rb$/.test(lp) ||
+                /_test\.rb$/.test(lp) ||
+                /\.(test|spec)\.[jt]sx?$/.test(lp) ||
+                /(test|spec|tests)\.(java|kt|scala)$/.test(lp) ||
+                /(tests?|spec)\.cs$/.test(lp) ||
+                /tests?\.swift$/.test(lp) ||
+                /_test\.dart$/.test(lp) ||
+                /\bicons?\b/.test(lp) ||
+                /\bi18n\b/.test(lp));
+        };
+        // Hard-exclude test/spec files (ALL tiers, not just tiny). One slipped test
+        // file dominates the per-file budget on small repos (cobra's `command_test.go`
+        // displaced `args.go`) AND wastes budget on large ones (Django's
+        // `custom_lookups/tests.py` ate ~2.3 KB of the 28 KB cap, crowding out the
+        // SQLCompiler mechanism the agent then Read). A test file almost never answers
+        // an architecture question. Skip when the query itself is about tests — the
+        // legitimate "explore the tests" case — and only cut if ≥2 non-test candidates
+        // remain (else tests are the only signal for this area).
+        {
+            const queryMentionsTests = /\b(test|tests|testing|spec|verify|verifies)\b/i.test(query);
+            if (!queryMentionsTests) {
+                const nonLow = relevantFiles.filter(([p]) => !isLowValue(p));
+                if (nonLow.length >= 2) {
+                    relevantFiles = nonLow;
+                }
+            }
+        }
+        // Secondary signal: how many DISTINCT query terms each file matches (path +
+        // symbol names). Kept only as a tiebreak — the PRIMARY relevance is graph
+        // connectivity below. (Term counting alone tied the real central file with
+        // incidental same-word matches; it's a weak text signal, not the ranker.)
+        const uniqueQueryTerms = [...new Set(queryTerms)].filter(t => t.length >= 3);
+        const fileTermHits = new Map();
+        for (const [fp, group] of relevantFiles) {
+            const hay = fp.toLowerCase() + ' ' + group.nodes.map(n => n.name.toLowerCase()).join(' ');
+            let hits = 0;
+            for (const t of uniqueQueryTerms)
+                if (hay.includes(t))
+                    hits++;
+            fileTermHits.set(fp, hits);
+        }
+        // PRIMARY relevance: graph connectivity (Random-Walk-with-Restart from the
+        // matched seeds — see computeGraphRelevance). Aggregate each file's nodes'
+        // walk mass. This is the signal text search lacks: the real cluster
+        // (org-user.storage.ts, call-connected to the matches) accrues mass; a lone
+        // text match (LensSwitcher.swift, matched "switch" but calls nothing in the
+        // flow) gets only its restart probability → ~0, and is dropped by the gate.
+        const nodeRwr = this.computeGraphRelevance([...subgraph.nodes.keys()], subgraph.edges, entryNodeIds);
+        const fileGraphScore = new Map();
+        for (const node of subgraph.nodes.values()) {
+            fileGraphScore.set(node.filePath, (fileGraphScore.get(node.filePath) ?? 0) + (nodeRwr.get(node.id) ?? 0));
+        }
+        const maxGraph = Math.max(0, ...fileGraphScore.values());
+        // Central file(s): the 1-2 most graph-central files that also match the
+        // query textually (so a connected hub-utility with no term match isn't
+        // mistaken for the subject). The heart of the answer — they earn the larger
+        // WHOLE-FILE ceiling below (a god-file central file still exceeds it and
+        // falls to generous full-method sectioning — never a whole dump).
+        const centralFiles = new Set([...fileGraphScore.entries()]
+            .filter(([fp, g]) => g > 0 && (fileTermHits.get(fp) ?? 0) >= 1)
+            .sort((a, b) => b[1] - a[1] || (fileTermHits.get(b[0]) ?? 0) - (fileTermHits.get(a[0]) ?? 0))
+            .slice(0, 2)
+            .map(([f]) => f));
+        // Files that DEFINE a symbol the agent named (or a subgraph root). These are
+        // the highest-relevance files there are — the agent asked for them by name —
+        // so the connectivity gate below must never drop them, even when their RWR
+        // mass is low (a leaf family file like codec.ts is call-connected to little
+        // but is exactly what the agent queried). Without this protection the gate
+        // prunes a named file and the agent Reads it back.
+        const entryFiles = new Set();
+        for (const id of entryNodeIds) {
+            const n = subgraph.nodes.get(id);
+            if (n)
+                entryFiles.add(n.filePath);
+        }
+        // Relevance gate (so the generous budget is a CEILING, not a target): keep a
+        // file only if it is STRUCTURALLY relevant by ANY of:
+        //   - graph score within a fraction of the top (it's on/near the flow), OR
+        //   - central (a query entry-point lives here), OR
+        //   - it DEFINES a symbol the agent named (entryFiles), OR
+        //   - it matches >= 2 DISTINCT named query terms — a strong text signal that
+        //     the agent is asking about this file even when nothing calls it (codec.ts:
+        //     the agent named `encode`/`Codec`/`JsonCodec`, all leaf classes with zero
+        //     RWR mass — graph alone wrongly drops it).
+        // A lone text match on one shared word (LensSwitcher: term=1, g~0) is still
+        // dropped, so the budget never fills with incidental files. Guarded so it
+        // never prunes below 2.
+        if (maxGraph > 0) {
+            const gated = relevantFiles.filter(([fp]) => (fileGraphScore.get(fp) ?? 0) >= maxGraph * 0.06
+                || centralFiles.has(fp)
+                || entryFiles.has(fp)
+                || (fileTermHits.get(fp) ?? 0) >= 2);
+            if (gated.length >= 2)
+                relevantFiles = gated;
+        }
+        // Sort files: graph-central first, then distinct-term match, then the
+        // existing low-value/generated/score tiebreaks.
+        // Files that DEFINE a symbol the agent NAMED. These sort first — ahead of
+        // graph connectivity — because the agent asked for them by name. Without
+        // this, a named leaf override reached only by dynamic dispatch (Alamofire's
+        // `DataRequest.task`/`validate`, low RWR mass) sorts below the high-
+        // connectivity abstract base (`Request.swift`) and the same-named overloads
+        // in other files (`Validation.swift`), falls outside the budget, and the
+        // agent Reads it. The named file is the answer — rank it at the top.
+        const namedSeedFiles = new Set();
+        for (const id of namedSeedIds) {
+            const n = subgraph.nodes.get(id);
+            if (n)
+                namedSeedFiles.add(n.filePath);
+        }
+        const sortedFiles = relevantFiles.sort((a, b) => {
+            const aPath = a[0].toLowerCase();
+            const bPath = b[0].toLowerCase();
+            // Agent-named files first (it asked for a symbol defined here by name).
+            const aNamed = namedSeedFiles.has(a[0]) ? 1 : 0;
+            const bNamed = namedSeedFiles.has(b[0]) ? 1 : 0;
+            if (aNamed !== bNamed)
+                return bNamed - aNamed;
+            // Graph connectivity is the next key (small epsilon so near-ties fall
+            // through to the text signal rather than coin-flipping on float noise).
+            const aG = fileGraphScore.get(a[0]) ?? 0;
+            const bG = fileGraphScore.get(b[0]) ?? 0;
+            if (Math.abs(aG - bG) > maxGraph * 0.01)
+                return bG - aG;
+            const aHits = fileTermHits.get(a[0]) ?? 0;
+            const bHits = fileTermHits.get(b[0]) ?? 0;
+            if (aHits !== bHits)
+                return bHits - aHits;
+            const aLow = isLowValue(aPath);
+            const bLow = isLowValue(bPath);
+            if (aLow !== bLow)
+                return aLow ? 1 : -1;
+            // Deprioritize generated source (.pb.go / .pulsar.go / _mocks.go / …) —
+            // the agent rarely needs to see the protobuf scaffold or gomock output
+            // when asking about the actual flow, and dumping their bodies inflates
+            // the response (the cosmos Q3 explore otherwise leads with
+            // `expected_keepers_mocks.go`, displacing the real `tally.go` content
+            // and forcing the agent to Read tally.go anyway).
+            const aGen = (0, generated_detection_1.isGeneratedFile)(a[0]);
+            const bGen = (0, generated_detection_1.isGeneratedFile)(b[0]);
+            if (aGen !== bGen)
+                return aGen ? 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 = [
+            `## Exploration: ${query}`,
+            '',
+            `Found ${subgraph.nodes.size} symbols across ${fileGroups.size} files.`,
+            '',
+        ];
+        // Domain facts (REQ-DOMAIN-005.A2): if the query names a documented domain
+        // term/entity, lead with the human-confirmed fact body. Placed near the top
+        // (before the source sections) so it is never lost to the output-budget
+        // truncation that drops trailing file sections.
+        const domainFacts = this.buildDomainFactsSection(cg, query);
+        if (domainFacts)
+            lines.push(domainFacts);
+        // Blast radius (always-on, compact): for the entry symbols, who depends on
+        // them + which tests cover them — locations only, no source — so the agent
+        // knows what to update/verify before editing without a separate call.
+        const blastRadius = this.buildBlastRadiusSection(cg, subgraph);
+        if (blastRadius)
+            lines.push(blastRadius);
+        // 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();
+            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
+        // Compute the flow spine once — used both to prepend the Flow section (below)
+        // and to gate adaptive source sizing: files on the spine get full source,
+        // off-spine peers skeletonize.
+        const flow = this.buildFlowFromNamedSymbols(cg, query);
+        // Polymorphic-sibling detector for adaptive sizing. A class that implements/
+        // extends a supertype shared by >= MIN_SIBLINGS classes is one of many
+        // INTERCHANGEABLE implementations (OkHttp's 14 `: Interceptor` classes —
+        // showing one + the rest as signatures is enough), as opposed to a DISTINCT
+        // pipeline step (Excalidraw's `renderStaticScene`, which shares no supertype and
+        // must stay full or the agent loses real content). Only off-spine sibling files
+        // skeletonize; distinct steps and on-spine files keep full source. Cache
+        // supertype→(has ≥N implementers) so this stays a handful of edge queries.
+        const MIN_SIBLINGS = 3;
+        const siblingSuper = new Map();
+        const isPolymorphicSibling = (nodes) => {
+            for (const n of nodes) {
+                for (const e of cg.getOutgoingEdges(n.id)) {
+                    if (e.kind !== 'implements' && e.kind !== 'extends')
+                        continue;
+                    let many = siblingSuper.get(e.target);
+                    if (many === undefined) {
+                        many = cg.getIncomingEdges(e.target)
+                            .filter((x) => x.kind === 'implements' || x.kind === 'extends').length >= MIN_SIBLINGS;
+                        siblingSuper.set(e.target, many);
+                    }
+                    if (many)
+                        return true;
+                }
+            }
+            return false;
+        };
+        // A file that DEFINES a polymorphic supertype (a class/interface with ≥
+        // MIN_SIBLINGS implementers) AND co-locates its subclasses is a redundant
+        // "family" file — Django's compiler.py holds `SQLCompiler` + its 4 subclasses
+        // (SQLInsert/Update/Delete/AggregateCompiler) in 2,266 lines. Such files are
+        // huge and read-anyway, so they should STILL skeletonize even when the agent
+        // named a method in them: a full one eats ~6.5K of the explore budget (Django
+        // is pinned at the 28K cap, truncating), starving the sibling files the agent
+        // then Reads. This flag OVERRIDES the named-callable spare below — it does NOT
+        // by itself spare a file. (OkHttp's RealCall implements the `Lockable` mixin
+        // but defines no ≥3-impl supertype, so the named spare keeps it full.)
+        const superMany = new Map();
+        const definesPolymorphicSupertype = (nodes) => {
+            for (const n of nodes) {
+                if (n.kind !== 'class' && n.kind !== 'interface' && n.kind !== 'struct'
+                    && n.kind !== 'trait' && n.kind !== 'protocol' && n.kind !== 'type_alias')
+                    continue;
+                let many = superMany.get(n.id);
+                if (many === undefined) {
+                    many = cg.getIncomingEdges(n.id)
+                        .filter((x) => x.kind === 'implements' || x.kind === 'extends').length >= MIN_SIBLINGS;
+                    superMany.set(n.id, many);
+                }
+                if (many)
+                    return true;
+            }
+            return false;
+        };
+        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;
+            // A file DEFINES a named/spine symbol (the answer) vs merely references the
+            // flow. Past 90% budget, stop pulling INCIDENTAL files — but keep scanning
+            // for necessary ones, which render even past the cap (bounded by maxFiles).
+            // Without this `continue` (was an unconditional `break`), the loop stopped
+            // after the build + validators-exec files and never reached the ranked-in
+            // validate-logic file (Alamofire's Validation.swift).
+            const fileNecessary = group.nodes.some(n => entryNodeIds.has(n.id) || flow.pathNodeIds.has(n.id) || flow.uniqueNamedNodeIds.has(n.id));
+            if (!fileNecessary && totalChars > budget.maxOutputChars * 0.9)
+                continue;
+            const absPath = (0, utils_1.validatePathWithinRoot)(projectRoot, filePath);
+            if (!absPath || !(0, fs_1.existsSync)(absPath))
+                continue;
+            let fileContent;
+            try {
+                fileContent = (0, fs_1.readFileSync)(absPath, 'utf-8');
+            }
+            catch {
+                continue;
+            }
+            const fileLines = fileContent.split('\n');
+            const lang = group.nodes[0]?.language || '';
+            // Adaptive sizing (SPECSHIP_ADAPTIVE_EXPLORE, default on): collapse a file
+            // to a per-symbol view when it's a redundant member of a polymorphic family.
+            // Engages iff ALL hold:
+            //   1. a flow spine exists,
+            //   2. no symbol in the file is on that spine (it's not the mechanism path),
+            //   3. it IS a polymorphic sibling (≥ MIN_SIBLINGS impls of a shared supertype),
+            //   4. it is NOT SPARED, where a file is spared iff the agent named a
+            //      (near-)UNIQUE callable in it (`getResponseWithInterceptorChain`, 1 def →
+            //      keep RealCall.kt full) UNLESS the file DEFINES the family supertype (a
+            //      base+subclasses "family" file like Django's compiler.py — collapse it).
+            //      Uniqueness matters: `as_sql` has 110 defs across every Compiler/Expression
+            //      subclass; naming it must NOT keep every backend variant + test file full
+            //      and flood the budget. That's why the spare reads uniqueNamedNodeIds.
+            // Within a collapsed file the render is PER-SYMBOL (condition B): a method the
+            // agent NAMED or that's on the spine is shown with its FULL body (so the agent
+            // doesn't Read the file back for it — Django's SQLCompiler.execute_sql/as_sql);
+            // every other symbol is just its signature. So the base mechanism survives while
+            // the file's other ~80 symbols + the redundant subclasses collapse to one line each.
+            const spareNamed = group.nodes.some(n => flow.uniqueNamedNodeIds.has(n.id));
+            const fileDefinesSuper = definesPolymorphicSupertype(group.nodes);
+            const spared = spareNamed && !fileDefinesSuper;
+            const CALLABLE_BODY = new Set(['method', 'function', 'constructor', 'component']);
+            const hasSpineNode = group.nodes.some(n => flow.pathNodeIds.has(n.id));
+            // On-spine god-file: the flow path runs THROUGH this file, but it also holds
+            // many OTHER named methods, and rendering all of them in full blows the
+            // per-file budget and starves the other flow files (Alamofire: the agent
+            // names ~7 Session.swift methods — the build spine PLUS off-path
+            // task/didCompleteTask — far past the whole response budget). Engage the
+            // per-symbol view to keep the SPINE full and collapse the off-path named
+            // methods to signatures. Only when there IS off-path content to shed —
+            // otherwise the spine is irreducible (a sequential flow has no redundancy),
+            // so leave it to the normal full render.
+            const namedBodyChars = group.nodes
+                .filter(n => CALLABLE_BODY.has(n.kind) && (flow.pathNodeIds.has(n.id) || flow.uniqueNamedNodeIds.has(n.id)))
+                .reduce((s, n) => s + fileLines.slice(n.startLine - 1, n.endLine).join('\n').length, 0);
+            const onSpineGodFile = hasSpineNode
+                && namedBodyChars > budget.maxCharsPerFile
+                && group.nodes.some(n => CALLABLE_BODY.has(n.kind) && flow.uniqueNamedNodeIds.has(n.id) && !flow.pathNodeIds.has(n.id));
+            if (adaptiveExploreEnabled() && flow.pathNodeIds.size > 0
+                && (onSpineGodFile || (!hasSpineNode && isPolymorphicSibling(group.nodes) && !spared))) {
+                const syms = group.nodes
+                    .filter(n => n.kind !== 'import' && n.kind !== 'export' && n.startLine > 0)
+                    .sort((a, b) => a.startLine - b.startLine);
+                // Pass 1: choose which symbols get a FULL body, by priority, greedily within
+                // a per-file body cap — so one huge family file can't body every named method
+                // and crowd out the other flow files (Django's query.py). A symbol earns a
+                // body if it's on-spine, or UNIQUELY named (`SQLCompiler.execute_sql`), or a
+                // co-named method WHEN this file DEFINES the family supertype (so the base
+                // `SQLCompiler.as_sql` body shows, but the 110 leaf `as_sql` overrides — and
+                // OkHttp's 5 `intercept`s if the agent names `intercept` — stay signatures).
+                const prio = (n) => !CALLABLE_BODY.has(n.kind) ? 99
+                    : flow.pathNodeIds.has(n.id) ? 0
+                        : flow.uniqueNamedNodeIds.has(n.id) ? 1
+                            : (fileDefinesSuper && flow.namedNodeIds.has(n.id)) ? 2 : 99;
+                // One ~250-line WINDOW per file. syms are taken by priority (spine first,
+                // then uniquely-named, then family-base), and the cap applies to ALL of
+                // them — including the spine — so a big-spine god-file (tokio's worker.rs:
+                // run→run_task→next_task→steal_work) can't eat the whole response and
+                // starve the co-flow file (harness.rs's poll). The native agent windows
+                // such a file too (~190 lines at a time), so this mimics, not truncates.
+                // Always emit ≥1 (never an empty section).
+                const bodyCap = budget.maxCharsPerFile * 1.5;
+                const bodyIds = new Set();
+                let bodyChars = 0;
+                for (const n of syms.filter(n => prio(n) < 99 && n.endLine >= n.startLine).sort((a, b) => prio(a) - prio(b))) {
+                    const sz = fileLines.slice(n.startLine - 1, n.endLine).join('\n').length;
+                    if (bodyChars + sz > bodyCap && bodyIds.size > 0)
+                        continue;
+                    bodyIds.add(n.id);
+                    bodyChars += sz;
+                }
+                // Pass 2: render in line order — full body for chosen symbols, else the
+                // signature line (capped, with a "+N more" tail so the structure map of a
+                // god-file doesn't itself bloat the budget).
+                const skel = [];
+                let coveredUntil = 0; // skip symbols already inside an emitted body
+                let sigCount = 0, sigDropped = 0;
+                const SIG_MAX = Math.max(12, budget.maxSymbolsInFileHeader * 2);
+                for (const n of syms) {
+                    if (n.startLine <= coveredUntil)
+                        continue;
+                    if (bodyIds.has(n.id)) {
+                        const end = n.endLine;
+                        const body = fileLines.slice(n.startLine - 1, end).join('\n');
+                        skel.push(exploreLineNumbersEnabled() ? numberSourceLines(body, n.startLine) : body);
+                        coveredUntil = end;
+                    }
+                    else {
+                        // Elide the body, emit the signature. node.startLine can point at a
+                        // decorator/annotation, so scan forward for the line that names the symbol.
+                        let lineNo = n.startLine;
+                        for (let k = 0; k < 4; k++) {
+                            if ((fileLines[n.startLine - 1 + k] || '').includes(n.name)) {
+                                lineNo = n.startLine + k;
+                                break;
+                            }
+                        }
+                        if (lineNo <= coveredUntil)
+                            continue;
+                        if (sigCount >= SIG_MAX) {
+                            sigDropped++;
+                            continue;
+                        }
+                        const sig = (fileLines[lineNo - 1] || '').trim();
+                        if (sig) {
+                            skel.push(exploreLineNumbersEnabled() ? `${lineNo}\t${sig}` : sig);
+                            sigCount++;
+                        }
+                    }
+                }
+                if (sigDropped > 0)
+                    skel.push(`… +${sigDropped} more (signatures elided)`);
+                if (skel.length > 0) {
+                    const names = [...new Set(group.nodes.filter(n => n.kind !== 'import' && n.kind !== 'export').map(n => n.name))]
+                        .slice(0, budget.maxSymbolsInFileHeader).join(', ');
+                    // Steer the agent to specship_explore for an elided body — NEVER to
+                    // Read. The old "Read for more" / "Read for a full body" tags invited
+                    // a Read of the very file just skeletonized; on a central, wanted file
+                    // (Session.swift, DataRequest.swift) that fired an over-investigation
+                    // spiral (the agent Read the skeletonized file, then kept digging).
+                    // CLAUDE.md: explore output must never tell the agent to Read.
+                    const tag = bodyIds.size > 0
+                        ? 'focused (the methods you named in full, the rest as signatures — specship_explore a signature by name for its body; do NOT Read)'
+                        : 'skeleton (signatures only — specship_explore a name for its full body; do NOT Read)';
+                    lines.push(`#### ${filePath} — ${names} · ${tag}`, '', '```' + lang, skel.join('\n'), '```', '');
+                    totalChars += skel.join('\n').length + 120;
+                    filesIncluded++;
+                    continue;
+                }
+            }
+            // Whole-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.
+            //
+            // CENTRAL files (where the query's entry points live) get a larger — but
+            // bounded — ceiling: they're the heart of the answer, the file(s) the agent
+            // would Read whole, so a genuinely small one comes back whole rather than as
+            // thin clusters. A LARGE central file (the 791-line org-user store) exceeds
+            // the ceiling and falls through to sectioning/clustering below — full method
+            // bodies + signatures — so we never dump (or overflow on) a whole god-file.
+            const isCentralFile = centralFiles.has(filePath);
+            // Central files get a slightly larger whole-file window than peripheral ones,
+            // but a TIGHT one (~1.5× the per-file cap): the native read of a central file
+            // is a ~150–250 line orientation window, NOT the whole file. A flat "whole
+            // central file" both overflowed the inline cap AND starved the co-flow files
+            // (worker.rs ate the budget, dropping harness.rs's poll). A larger central
+            // file falls through to per-method windowing/clustering below.
+            const WHOLE_FILE_MAX_LINES = isCentralFile ? 280 : 220;
+            const WHOLE_FILE_MAX_CHARS = isCentralFile
+                ? Math.min(Math.max(0, budget.maxOutputChars - totalChars - 200), Math.round(budget.maxCharsPerFile * 1.5))
+                : 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 (!fileNecessary && totalChars + wholeSection.length + 200 > budget.maxOutputChars) {
+                    // Don't slice a whole file mid-method: an incidental file that doesn't
+                    // fit is skipped; a necessary one (below) renders in full. Half a file
+                    // forces the Read this is meant to prevent.
+                    anyFileTrimmed = true;
+                    continue;
+                }
+                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']);
+            // Cluster from this file's gathered nodes PLUS any callable the agent NAMED that
+            // lives here. Explore's relevance gather can miss a named method def in a huge
+            // non-sibling file — Django's query.py is 3,040 lines and `_fetch_all` (L2237)
+            // was gathered only as call-reference edges, never as a def, so it formed no
+            // cluster and the agent Read it back. Inject named defs directly and rank them
+            // ABOVE connected/glue nodes (importance 9) so their cluster wins the per-file
+            // budget — the agent explicitly asked for these symbols.
+            const rangeNodes = new Map();
+            for (const n of group.nodes)
+                if (n.startLine > 0 && n.endLine > 0)
+                    rangeNodes.set(n.id, n);
+            for (const id of flow.namedNodeIds) {
+                if (rangeNodes.has(id))
+                    continue;
+                const n = cg.getNode(id);
+                if (n && n.filePath === filePath && n.startLine > 0 && n.endLine > 0)
+                    rangeNodes.set(id, n);
+            }
+            const ranges = [...rangeNodes.values()]
+                // 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 (flow.namedNodeIds.has(n.id))
+                    importance = 9; // agent named it → keep its cluster
+                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(); // 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}:${edge.target}`;
+                    if (edgeLines.has(key))
+                        continue;
+                    edgeLines.add(key);
+                    // Look up target name from subgraph first, fall back to edge kind
+                    const targetNode = subgraph.nodes.get(edge.target);
+                    const targetName = targetNode?.name ?? edge.kind;
+                    ranges.push({ start: edge.line, end: edge.line, name: targetName, kind: edge.kind, importance: 2 });
+                }
+            }
+            ranges.sort((a, b) => a.start - b.start);
+            if (ranges.length === 0)
+                continue;
+            const gapThreshold = budget.gapThreshold;
+            const clusters = [];
+            let current = {
+                start: ranges[0].start,
+                end: ranges[0].end,
+                symbols: [`${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) => {
+                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;
+            });
+            // Per-file budget is the SMALLER of the per-file cap and what's left of the
+            // total output cap — so selection (which ranks by importance) keeps the
+            // high-importance clusters and drops peripheral ones, instead of the
+            // downstream source-order trim slicing off whatever comes last in the file.
+            // That source-order slice is what cut Django's `_fetch_all` (L2237, importance
+            // 9 — agent-named) when query.py was the last of four big files to be emitted.
+            const fileBudget = Math.min(budget.maxCharsPerFile, Math.max(0, budget.maxOutputChars - totalChars - 200));
+            const chosenIndices = new Set();
+            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 > fileBudget)
+                    continue;
+                chosenIndices.add(rc.idx);
+                projectedChars += sectionLen;
+            }
+            // Emit chosen clusters in source order so the file reads top-to-bottom.
+            let fileSection = '';
+            const allSymbols = [];
+            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);
+            }
+            // A chosen cluster is a COMPLETE method-range — we never cut through a body.
+            // An oversize single cluster (a long monolithic function) renders in FULL:
+            // half a method is useless (the agent just Reads the rest for the other half),
+            // which is the very fallback explore exists to prevent. A pathological file is
+            // bounded by the per-file cluster SELECTION above + the total hard ceiling.
+            if (chosenIndices.size < clusters.length) {
+                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();
+            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}`;
+            // The total cap bounds INCIDENTAL files only. A file that DEFINES a symbol
+            // the agent named (or that's on the flow spine) renders even when the
+            // nominal total is used up — it's the answer, and the set is bounded by
+            // maxFiles AND by true-spine/named-seeding having already trimmed each file
+            // to its necessary content. A file that merely REFERENCES the flow
+            // (Combine.swift name-drops request/task) is incidental → still capped, so
+            // freed budget never leaks into noise. This is the last god-file layer:
+            // build (Session, true-spined) + validators-exec (Request) + validate
+            // (DataRequest/Validation) all render, instead of the cap dropping whichever
+            // phase the file order happened to put last.
+            if (!fileNecessary && totalChars + fileSection.length + 200 > budget.maxOutputChars) {
+                // Incidental file that doesn't fit: SKIP it whole — never slice mid-method.
+                // Keep scanning for necessary files (which bypass this cap and render in
+                // full, bounded by the hard ceiling).
+                anyFileTrimmed = true;
+                continue;
+            }
+            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 specship_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 \`specship_explore\` (or \`specship_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
+            }
+        }
+        // Final ceiling — an ABSOLUTE inline cap, not a multiple of the budget. The
+        // render loop renders necessary (named/spine) files even a bit past
+        // maxOutputChars and caps only incidental ones, so this is the last safety.
+        // It MUST stay under the host's inline tool-result limit (~25K chars): above
+        // that the result is externalized to a file the agent Reads back (a 35K
+        // vscode explore did exactly this in the n=4 A/B). So allow a little
+        // necessary overflow above the 24K budget, but hard-stop at 25K — never into
+        // externalize territory.
+        const output = flow.text + lines.join('\n');
+        const hardCeiling = Math.min(Math.round(budget.maxOutputChars * 1.5), 25000);
+        if (output.length > hardCeiling) {
+            // Cut at a FILE-SECTION boundary (the last `#### ` header before the
+            // ceiling) so we drop whole trailing file-sections rather than slicing
+            // through a method body — a half-rendered method just forces the Read this
+            // tool exists to prevent. Fall back to a line boundary only if no section
+            // header sits in the back half (degenerate single-giant-section case).
+            const cut = output.slice(0, hardCeiling);
+            const lastSection = cut.lastIndexOf('\n#### ');
+            const boundary = lastSection > hardCeiling * 0.5 ? lastSection : cut.lastIndexOf('\n');
+            const safe = boundary > 0 ? cut.slice(0, boundary) : 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 specship_explore with the specific names — do NOT Read these files.)');
+        }
+        return this.textResult(output);
+    }
+    /**
+     * Handle specship_node
+     */
+    async handleNode(args) {
+        const symbol = this.validateString(args.symbol, 'symbol');
+        if (typeof symbol !== 'string')
+            return symbol;
+        const cg = this.getSpecShip(args.projectPath);
+        // Default to false to minimize context usage
+        const includeCode = args.includeCode === true;
+        const fileHint = typeof args.file === 'string' && args.file.trim() ? args.file.trim() : undefined;
+        const lineHint = typeof args.line === 'number' && args.line > 0 ? args.line : undefined;
+        let matches = this.findSymbolMatches(cg, symbol);
+        if (matches.length === 0) {
+            return this.textResult(`Symbol "${symbol}" not found in the codebase`);
+        }
+        // Disambiguate a heavily-overloaded name to a specific definition the caller
+        // pinned by file/line (the `file:line` a trail or another tool showed it) —
+        // so it can fetch e.g. `Harness::poll` at harness.rs:153 out of 50+ `poll`s
+        // instead of Reading. file matches by path suffix/substring; line prefers the
+        // def whose body contains it, else the nearest start. Only narrows (never
+        // empties — if a hint matches nothing it's ignored).
+        if (matches.length > 1 && (fileHint || lineHint !== undefined)) {
+            const norm = (p) => p.replace(/\\/g, '/').toLowerCase();
+            let narrowed = matches;
+            if (fileHint) {
+                const fh = norm(fileHint);
+                const byFile = narrowed.filter((n) => norm(n.filePath).endsWith(fh) || norm(n.filePath).includes(fh));
+                if (byFile.length > 0)
+                    narrowed = byFile;
+            }
+            if (lineHint !== undefined && narrowed.length > 1) {
+                const containing = narrowed.filter((n) => n.startLine <= lineHint && (n.endLine ?? n.startLine) >= lineHint);
+                narrowed = containing.length > 0
+                    ? containing
+                    : [...narrowed].sort((a, b) => Math.abs(a.startLine - lineHint) - Math.abs(b.startLine - lineHint)).slice(0, 1);
+            }
+            if (narrowed.length > 0)
+                matches = narrowed;
+        }
+        // Single definition — the common case.
+        if (matches.length === 1) {
+            return this.textResult(this.truncateOutput(await this.renderNodeSection(cg, matches[0], includeCode)));
+        }
+        // Multiple definitions share this name — overloads, or same-named methods on
+        // different types (Alamofire `didCompleteTask`/`task`/`validate`, gin
+        // `reset`). Returning ONE forces the agent to guess, and when it guesses
+        // wrong it READS the file to find the right overload — the dominant
+        // specship_node read cause on Swift/Go. So return them ALL: pack as many
+        // FULL bodies as fit a char budget (the agent gets the one it needs in this
+        // one call, no follow-up parameter to learn), and list any remainder by
+        // file:line so a large overload set can't overflow the per-tool cap.
+        const header = `**${matches.length} definitions named "${symbol}"**`;
+        if (!includeCode) {
+            const list = matches.map((n) => `- \`${n.name}\` (${n.kind}) — ${n.filePath}:${n.startLine}`);
+            return this.textResult(this.truncateOutput([header, '', 'Re-query with `includeCode: true` to get every body in one call — no need to pick one first.', '', ...list].join('\n')));
+        }
+        const BODY_BUDGET = 12000; // leaves room under MAX_OUTPUT_LENGTH for the header + list
+        // The CHAR budget is the real limiter — keep the count cap high so a set of
+        // SHORT overloads (Alamofire's 10 `validate` variants, each a few lines) all
+        // render in full rather than relegating the one the agent wanted to a
+        // bodiless list. Only a set of many LARGE bodies hits the char budget first.
+        const HARD_CAP = 16;
+        const rendered = [];
+        const listed = [];
+        let used = 0;
+        for (const n of matches) {
+            if (rendered.length >= HARD_CAP) {
+                listed.push(n);
+                continue;
+            }
+            const section = await this.renderNodeSection(cg, n, true);
+            // Always emit the first; emit the rest only while within the char budget.
+            if (rendered.length === 0 || used + section.length <= BODY_BUDGET) {
+                rendered.push(section);
+                used += section.length;
+            }
+            else {
+                listed.push(n);
+            }
+        }
+        const out = [
+            header,
+            `Returning ${rendered.length} in full${listed.length ? `; ${listed.length} more listed below` : ''} — pick the one you need (no Read required).`,
+            '',
+            rendered.join('\n\n---\n\n'),
+        ];
+        if (listed.length) {
+            const LIST_CAP = 20;
+            const shownList = listed.slice(0, LIST_CAP);
+            out.push('', '### Other definitions', ...shownList.map((n) => `- \`${n.name}\` (${n.kind}) — ${n.filePath}:${n.startLine}`));
+            if (listed.length > LIST_CAP)
+                out.push(`- … +${listed.length - LIST_CAP} more`);
+            out.push('', `> Need one of these in full? Call specship_node again with \`file\` (e.g. \`"${listed[0].filePath.split('/').pop()}"\`) or \`line\` — do NOT Read it.`);
+        }
+        return this.textResult(this.truncateOutput(out.join('\n')));
+    }
+    /** Render one symbol: details + (optional) body/outline + its caller/callee trail. */
+    async renderNodeSection(cg, node, includeCode) {
+        let code = null;
+        let outline = null;
+        if (includeCode) {
+            // For container symbols (class/interface/struct/…), the full body is the
+            // sum of every method body — a wall of source. Return a structural outline
+            // (members + signatures + line numbers) instead; leaf symbols return their
+            // full body.
+            if (CONTAINER_NODE_KINDS.has(node.kind)) {
+                outline = this.buildContainerOutline(cg, node);
+            }
+            if (!outline) {
+                code = await cg.getCode(node.id);
+            }
+        }
+        return (this.formatNodeDetails(node, code, outline) +
+            this.formatTrail(cg, node) +
+            (0, spec_tools_1.renderLinkedSpecsForNode)(cg, node.id));
+    }
+    /**
+     * Build the "trail" for a symbol: its direct callees (what it calls) and
+     * callers (what calls it), each with file:line — so specship_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 specship_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.
+     */
+    formatTrail(cg, node) {
+        const TRAIL_CAP = 12;
+        const fmt = (e) => {
+            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) => {
+            const seen = new Set([node.id]);
+            const out = [];
+            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 = ['', '### Trail — specship_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 specship_status
+     */
+    async handleStatus(args) {
+        let cg = this.getSpecShip(args.projectPath);
+        // Same trick as withStalenessNotice — when an explicit projectPath
+        // resolves to the same project as the default session cg, prefer the
+        // default so getPendingFiles() (only populated by the default's watcher)
+        // is non-empty when there are pending edits.
+        if (this.cg && cg !== this.cg) {
+            try {
+                if ((0, path_1.resolve)(this.cg.getProjectRoot()) === (0, path_1.resolve)(cg.getProjectRoot())) {
+                    cg = this.cg;
+                }
+            }
+            catch { /* closed instance — leave as is */ }
+        }
+        const stats = cg.getStats();
+        // Warn when this index actually belongs to a different git working tree
+        // (e.g. the server resolved up from a nested worktree to the main checkout).
+        // Queries then reflect that tree's branch, not the worktree being edited.
+        // status shows the verbose, multi-line form; the read tools get the compact
+        // one-liner via withWorktreeNotice. Both share the cached detection.
+        const mismatch = this.worktreeMismatchFor(args.projectPath);
+        const lines = [
+            '## SpecShip Status',
+            '',
+        ];
+        if (mismatch) {
+            lines.push(`> ⚠ ${(0, worktree_1.worktreeMismatchWarning)(mismatch).replace(/\n/g, '\n> ')}`, '');
+        }
+        lines.push(`**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 > 0) {
+                lines.push(`- ${kind}: ${count}`);
+            }
+        }
+        lines.push('', '### Languages:');
+        for (const [lang, count] of Object.entries(stats.filesByLanguage)) {
+            if (count > 0) {
+                lines.push(`- ${lang}: ${count}`);
+            }
+        }
+        // Per-file freshness — the inverse of the auto-prepended staleness banner
+        // (issue #403). Surfacing it inside `status` gives the agent a single
+        // place to ask "is the index caught up?" rather than inferring from
+        // banners on other tool calls.
+        const pending = cg.getPendingFiles();
+        if (pending.length > 0) {
+            lines.push('', '### Pending sync:');
+            const now = Date.now();
+            for (const p of pending) {
+                const ageMs = Math.max(0, now - p.lastSeenMs);
+                const label = p.indexing ? 'indexing in progress' : 'pending sync';
+                lines.push(`- ${p.path} (edited ${ageMs}ms ago, ${label})`);
+            }
+        }
+        return this.textResult(lines.join('\n'));
+    }
+    /**
+     * Handle specship_files - get project file structure from the index
+     */
+    async handleFiles(args) {
+        const cg = this.getSpecShip(args.projectPath);
+        const pathFilter = args.path;
+        const pattern = args.pattern;
+        const format = args.format || 'tree';
+        const includeMetadata = args.includeMetadata !== false;
+        const maxDepth = args.maxDepth != null ? (0, utils_1.clamp)(args.maxDepth, 1, 20) : undefined;
+        // Get all files from the index
+        const allFiles = cg.getFiles();
+        if (allFiles.length === 0) {
+            return this.textResult('No files indexed. Run `specship index` first.');
+        }
+        // Filter by path prefix. Stored paths are project-relative POSIX (e.g.
+        // "src/foo.ts"), but agents commonly pass project-root variants like "/",
+        // ".", "./", "" or Windows-style "src\foo" — and prefixes with leading
+        // "/", "./" or "\". Normalize all of those before matching so the agent
+        // gets results instead of falling back to Read/Glob (see #426).
+        const normalizedFilter = pathFilter
+            ? pathFilter
+                .replace(/\\/g, '/')
+                .replace(/^(?:\.?\/+)+/, '')
+                .replace(/^\.$/, '')
+                .replace(/\/+$/, '')
+            : '';
+        let files = normalizedFilter
+            ? allFiles.filter(f => f.path === normalizedFilter || f.path.startsWith(normalizedFilter + '/'))
+            : allFiles;
+        // Filter by glob pattern
+        if (pattern) {
+            const regex = this.globToRegex(pattern);
+            files = files.filter(f => regex.test(f.path));
+        }
+        if (files.length === 0) {
+            return this.textResult(`No files found matching the criteria.`);
+        }
+        // Format output
+        let output;
+        switch (format) {
+            case 'flat':
+                output = this.formatFilesFlat(files, includeMetadata);
+                break;
+            case 'grouped':
+                output = this.formatFilesGrouped(files, includeMetadata);
+                break;
+            case 'tree':
+            default:
+                output = this.formatFilesTree(files, includeMetadata, maxDepth);
+                break;
+        }
+        return this.textResult(this.truncateOutput(output));
+    }
+    /**
+     * Convert glob pattern to regex
+     */
+    globToRegex(pattern) {
+        const escaped = pattern
+            .replace(/[.+^${}()|[\]\\]/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
+     */
+    formatFilesFlat(files, includeMetadata) {
+        const lines = [`## 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(`- ${file.path}`);
+            }
+        }
+        return lines.join('\n');
+    }
+    /**
+     * Format files grouped by language
+     */
+    formatFilesGrouped(files, includeMetadata) {
+        const byLang = new Map();
+        for (const file of files) {
+            const existing = byLang.get(file.language) || [];
+            existing.push(file);
+            byLang.set(file.language, existing);
+        }
+        const lines = [`## Files by Language (${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(`- ${file.path}`);
+                }
+            }
+            lines.push('');
+        }
+        return lines.join('\n');
+    }
+    /**
+     * Format files as a tree structure
+     */
+    formatFilesTree(files, includeMetadata, maxDepth) {
+        const root = { name: '', children: new Map() };
+        for (const file of files) {
+            const parts = file.path.split('/');
+            let current = root;
+            for (let i = 0; i < parts.length; i++) {
+                const part = parts[i];
+                if (!part)
+                    continue;
+                if (!current.children.has(part)) {
+                    current.children.set(part, { name: part, children: new Map() });
+                }
+                current = current.children.get(part);
+                // If this is the last part, it's a file
+                if (i === parts.length - 1) {
+                    current.file = { language: file.language, nodeCount: file.nodeCount };
+                }
+            }
+        }
+        // Render tree
+        const lines = [`## Project Structure (${files.length} files)`, ''];
+        const renderNode = (node, prefix, isLast, depth) => {
+            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`)
+     */
+    matchesSymbol(node, symbol) {
+        // 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));
+    }
+    /**
+     * Find ALL definitions matching a name, ranked, so specship_node can return
+     * every overload instead of guessing one (the wrong guess → a Read). Keepers
+     * rank before generated stubs (.pb.go etc.); stable within a group preserves
+     * FTS order. Returns [] when nothing matches; a qualified lookup that finds no
+     * exact match returns [] rather than a misleading fuzzy file hit (#173); a
+     * bare name with no exact match falls back to the single top fuzzy result.
+     */
+    findSymbolMatches(cg, symbol) {
+        const isQualified = /[.\/]|::/.test(symbol);
+        // For a bare name, enumerate EVERY exact-name definition via the direct index
+        // (not FTS, which caps + ranks): tokio's `poll` has 50+ defs and the one the
+        // caller wants (`Harness::poll` at harness.rs:153) ranks below any search cut,
+        // so it could be neither rendered nor pinned by the file/line disambiguator —
+        // and the agent Read it. With the full set, the multi-overload render + the
+        // file/line filter can both reach it.
+        if (!isQualified) {
+            const exact = cg.getNodesByName(symbol);
+            if (exact.length > 0) {
+                return [...exact].sort((a, b) => ((0, generated_detection_1.isGeneratedFile)(a.filePath) ? 1 : 0) - ((0, generated_detection_1.isGeneratedFile)(b.filePath) ? 1 : 0));
+            }
+            // No exact match — use the single top fuzzy result (e.g. a file basename).
+            const fuzzy = cg.searchNodes(symbol, { limit: 10 });
+            return fuzzy[0] ? [fuzzy[0].node] : [];
+        }
+        // Qualified lookup (`Session.request`, `stage_apply::run`): FTS + matchesSymbol.
+        const limit = 50;
+        let results = cg.searchNodes(symbol, { limit });
+        // FTS strips colons, so `stage_apply::run` searches 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)
+            return [];
+        const exactMatches = results.filter((r) => this.matchesSymbol(r.node, symbol));
+        if (exactMatches.length === 0) {
+            // No exact match — a qualified lookup must not fall back to a fuzzy file
+            // hit (#173); a bare name may use the single top fuzzy result.
+            return isQualified ? [] : results[0] ? [results[0].node] : [];
+        }
+        // Down-rank generated files (.pb.go, .pulsar.go, _grpc.pb.go, …) so a flow
+        // query prefers the keeper implementation over the protobuf-generated stub.
+        return [...exactMatches]
+            .sort((a, b) => ((0, generated_detection_1.isGeneratedFile)(a.node.filePath) ? 1 : 0) - ((0, generated_detection_1.isGeneratedFile)(b.node.filePath) ? 1 : 0))
+            .map((r) => r.node);
+    }
+    /**
+     * 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).
+     */
+    findAllSymbols(cg, symbol) {
+        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: '' };
+        }
+        // Same generated-file down-rank as findSymbol — keeps callers/callees
+        // /impact aggregation aligned (a query against "Send" returns the
+        // hand-written implementations before the protobuf scaffold).
+        const ranked = [...exactMatches].sort((a, b) => {
+            const aGen = (0, generated_detection_1.isGeneratedFile)(a.node.filePath) ? 1 : 0;
+            const bGen = (0, generated_detection_1.isGeneratedFile)(b.node.filePath) ? 1 : 0;
+            return aGen - bGen;
+        });
+        const locations = ranked.map(r => `${r.node.kind} at ${r.node.filePath}:${r.node.startLine}`);
+        const note = `\n\n> **Note:** Aggregated results across ${ranked.length} symbols named "${symbol}": ${locations.join(', ')}`;
+        return { nodes: ranked.map(r => r.node), note };
+    }
+    /**
+     * Truncate output if it exceeds the maximum length
+     */
+    truncateOutput(text) {
+        if (text.length <= MAX_OUTPUT_LENGTH)
+            return text;
+        const truncated = text.slice(0, MAX_OUTPUT_LENGTH);
+        const lastNewline = truncated.lastIndexOf('\n');
+        const cutPoint = lastNewline > MAX_OUTPUT_LENGTH * 0.8 ? lastNewline : MAX_OUTPUT_LENGTH;
+        return truncated.slice(0, cutPoint) + '\n\n... (output truncated)';
+    }
+    // =========================================================================
+    // Formatting helpers (compact by default to reduce context usage)
+    // =========================================================================
+    formatSearchResults(results) {
+        const lines = [`## Search Results (${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');
+    }
+    formatNodeList(nodes, title) {
+        const lines = [`## ${title} (${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');
+    }
+    formatImpact(symbol, impact) {
+        const nodeCount = impact.nodes.size;
+        // Compact format: just list affected symbols grouped by file
+        const lines = [
+            `## Impact: "${symbol}" affects ${nodeCount} symbols`,
+            '',
+        ];
+        // Group by file
+        const byFile = new Map();
+        for (const node of impact.nodes.values()) {
+            const existing = byFile.get(node.filePath) || [];
+            existing.push(node);
+            byFile.set(node.filePath, existing);
+        }
+        for (const [file, nodes] of byFile) {
+            lines.push(`**${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.
+     */
+    buildContainerOutline(cg, node) {
+        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');
+    }
+    formatNodeDetails(node, code, outline) {
+        const location = node.startLine ? `:${node.startLine}` : '';
+        const lines = [
+            `## ${node.name} (${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 specship_node on a specific member for its body.`);
+        }
+        else if (code) {
+            // Line-numbered (cat -n style, like specship_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');
+    }
+    textResult(text) {
+        return {
+            content: [{ type: 'text', text }],
+        };
+    }
+    errorResult(message) {
+        return {
+            content: [{ type: 'text', text: `Error: ${message}` }],
+            isError: true,
+        };
+    }
+}
+exports.ToolHandler = ToolHandler;
+//# sourceMappingURL=tools.js.map