@colbymchenry/codegraph-darwin-x64 0.9.7 → 0.9.8

package/lib/dist/mcp/tools.js

@@ -43,7 +43,14 @@ exports.getExploreBudget = getExploreBudget;
 exports.getExploreOutputBudget = getExploreOutputBudget;
 exports.formatStaleBanner = formatStaleBanner;
 exports.formatStaleFooter = formatStaleFooter;
-const index_1 = __importStar(require("../index"));
+exports.getStaticTools = getStaticTools;
+const directory_1 = require("../directory");
+// Lazy-load the heavy CodeGraph 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;
+// CodeGraph is pulled in only when a tool actually opens a project. require() is
+// sync + cached (CommonJS build).
+const loadCodeGraph = () => require('../index').default;
 const worktree_1 = require("../sync/worktree");
 const crypto_1 = require("crypto");
 const fs_1 = require("fs");
@@ -206,6 +213,21 @@ function getExploreOutputBudget(fileCount) {
 function exploreLineNumbersEnabled() {
     return process.env.CODEGRAPH_EXPLORE_LINENUMS !== '0';
 }
+/**
+ * Adaptive explore sizing (default ON). `codegraph_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 `CODEGRAPH_ADAPTIVE_EXPLORE=0` to disable.
+ */
+function adaptiveExploreEnabled() {
+    return process.env.CODEGRAPH_ADAPTIVE_EXPLORE !== '0' && process.env.CODEGRAPH_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
@@ -536,6 +558,19 @@ exports.tools = [
         },
     },
 ];
+/**
+ * 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-CodeGraph 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.CODEGRAPH_MCP_TOOLS;
+    if (!raw || !raw.trim())
+        return exports.tools;
+    const allow = new Set(raw.split(',').map(s => s.trim().replace(/^codegraph_/, '')).filter(Boolean));
+    return allow.size ? exports.tools.filter(t => allow.has(t.name.replace(/^codegraph_/, ''))) : exports.tools;
+}
 /**
  * Tool handler that executes tools against a CodeGraph instance
  *
@@ -718,7 +753,7 @@ class ToolHandler {
             }
         }
         // Walk up parent directories to find nearest .codegraph/
-        const resolvedRoot = (0, index_1.findNearestCodeGraphRoot)(projectPath);
+        const resolvedRoot = (0, directory_1.findNearestCodeGraphRoot)(projectPath);
         if (!resolvedRoot) {
             throw new Error(`CodeGraph not initialized in ${projectPath}. Run 'codegraph init' in that project first.`);
         }
@@ -740,7 +775,7 @@ class ToolHandler {
             return cg;
         }
         // Open and cache under both paths
-        const cg = index_1.default.openSync(resolvedRoot);
+        const cg = loadCodeGraph().openSync(resolvedRoot);
         this.projectCache.set(resolvedRoot, cg);
         if (projectPath !== resolvedRoot) {
             this.projectCache.set(projectPath, cg);
@@ -1463,10 +1498,30 @@ class ToolHandler {
             - (isLessCanonicalPath(b) ? LESS_CANONICAL_PENALTY : 0);
         const fromCands = fromMatches.nodes;
         const toCands = toMatches.nodes;
+        // Candidate relevance: an overloaded name (Alamofire has 44 `request`s, most
+        // of them EMPTY EventMonitor protocol-conformance stubs `func request(…){}`)
+        // floods the pool with no-op decls. Shared-dir-prefix alone then MISLEADS —
+        // two unrelated `Source/Features/` delegate stubs outscore the real
+        // `Source/Core/Session.request` × `Source/Core/…task` pair the agent meant,
+        // so trace resolves to stubs, finds no path, and the agent reads by line.
+        // Penalize empty stubs and test-file symbols so a substantive entry point
+        // wins; among real methods this is ~flat, so path-proximity still decides
+        // (cosmos EndBlocker disambiguation is unaffected — none of its candidates
+        // are stubs/tests).
+        const isTestPath = (p) => /(^|\/)(tests?|specs?|__tests__|testdata|mocks?|fixtures?)\//i.test(p) || /\.(test|spec)\.[a-z]+$/i.test(p);
+        const nodeRelevance = (n) => {
+            const bodyLines = Math.max(0, (n.endLine ?? n.startLine) - n.startLine);
+            let s = Math.min(bodyLines, 20); // a substantive body is more likely the meant symbol
+            if (bodyLines <= 1)
+                s -= 40; // empty/one-line stub (protocol no-op, decl-only) — almost never the trace endpoint
+            if (isTestPath(n.filePath))
+                s -= 150; // a Source/ symbol is meant over a Tests/ same-named one
+            return s;
+        };
         const pairs = [];
         for (const f of fromCands) {
             for (const t of toCands) {
-                pairs.push({ f, t, score: scorePair(f.filePath, t.filePath) });
+                pairs.push({ f, t, score: scorePair(f.filePath, t.filePath) + nodeRelevance(f) + nodeRelevance(t) });
             }
         }
         // Sort by shared prefix desc, then by FTS order (already encoded in the
@@ -1718,6 +1773,14 @@ class ToolHandler {
                 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;
     }
     /**
@@ -1811,6 +1874,7 @@ class ToolHandler {
      * 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
@@ -1822,7 +1886,7 @@ class ToolHandler {
                     .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 '';
+                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.
@@ -1832,24 +1896,34 @@ class ToolHandler {
                     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 pick = cands.length <= 3
+                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))
+                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 '';
+                return EMPTY;
             const MAX_HOPS = 7;
             let best = null;
             // BFS the full call graph (incl. synth edges) from each named seed, but
@@ -1895,22 +1969,66 @@ class ToolHandler {
                 if (!best || chain.length > best.length)
                     best = chain;
             }
-            if (!best || best.length < 3)
-                return '';
-            const out = ['## Flow (call path among the symbols you queried)', ''];
-            for (let i = 0; i < best.length; i++) {
-                const step = best[i];
-                if (step.edge) {
-                    const sy = this.synthEdgeNote(step.edge);
-                    out.push(`   ↓ ${sy ? sy.compact : step.edge.kind}`);
+            const hasMain = !!best && best.length >= 3;
+            const pathIds = new Set((best ?? []).map((s) => s.node.id));
+            // Supplementary: dynamic-dispatch (synthesized) edges incident to a NAMED
+            // symbol — the indirect hops an agent would otherwise grep/Read to
+            // reconstruct ("where do the appended `validators` actually run?"). The
+            // synth edge IS that answer, so surface it even when the OTHER end wasn't
+            // named (e.g. the agent names `validate` but not the `didCompleteTask`
+            // that drains the collection). On-topic by construction: only heuristic
+            // edges touching a symbol the agent named; skipped when the hop already
+            // shows in the main chain.
+            const synthLines = [];
+            const synthSeen = new Set();
+            for (const n of named.values()) {
+                if (synthLines.length >= 6)
+                    break;
+                for (const { node: other, edge } of [...cg.getCallers(n.id), ...cg.getCallees(n.id)]) {
+                    if (synthLines.length >= 6)
+                        break;
+                    if (edge.provenance !== 'heuristic' || other.id === n.id)
+                        continue;
+                    if (pathIds.has(edge.source) && pathIds.has(edge.target))
+                        continue; // already in the main chain
+                    const src = edge.source === n.id ? n : other;
+                    const tgt = edge.source === n.id ? other : n;
+                    const key = `${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}:${step.node.startLine})`);
                 }
-                out.push(`${i + 1}. ${step.node.name} (${step.node.filePath}:${step.node.startLine})`);
+                out.push('');
             }
-            out.push('', '> Full source for these symbols is below; codegraph_trace(from,to) for the exact path between two endpoints.', '');
-            return out.join('\n');
+            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; codegraph_trace(from,to) for the exact path between two endpoints.', '');
+            // 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 '';
+            return EMPTY;
         }
     }
     /**
@@ -1991,9 +2109,42 @@ class ToolHandler {
                 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);
+            for (const t of tokens) {
+                const cands = this.findAllSymbols(cg, t).nodes
+                    .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
+                // (`request` = 44, mostly stubs) injects only the single most substantive
+                // one, so the build-overload flood doesn't crowd the subgraph.
+                for (const n of cands.slice(0, cands.length <= 3 ? cands.length : 1)) {
+                    if (!subgraph.nodes.has(n.id)) {
+                        subgraph.nodes.set(n.id, n);
+                        namedSeedIds.add(n.id);
+                    }
+                }
+            }
+        }
         // Step 2: Group nodes by file, score by relevance
         const fileGroups = new Map();
-        const entryNodeIds = new Set(subgraph.roots);
+        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) {
@@ -2008,8 +2159,16 @@ class ToolHandler {
                 continue;
             const group = fileGroups.get(node.filePath) || { nodes: [], score: 0 };
             group.nodes.push(node);
-            // Score: entry point nodes worth 10, directly connected worth 3, others worth 1
-            if (entryNodeIds.has(node.id)) {
+            // 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)) {
@@ -2042,20 +2201,18 @@ class ToolHandler {
                 /\bicons?\b/.test(lp) ||
                 /\bi18n\b/.test(lp));
         };
-        // Tiny-tier hard-exclude: on small projects (`excludeLowValueFiles`
-        // budget flag), one slipped test/spec file dominates the per-file budget
-        // (cobra's `command_test.go` displaced `args.go` and contributed ~10KB of
-        // pure noise to "How does cobra parse commands?"). The sort-step
-        // deprioritization isn't enough at small N. Skip the hard-exclude when
-        // the query itself is about tests — that's the legitimate "explore the
-        // tests" case where the agent does want them.
-        if (budget.excludeLowValueFiles) {
+        // 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));
-                // Only apply the hard-filter if we still have at least 2 non-test
-                // candidates after the cut — otherwise the agent is asking about an
-                // area where tests are the only signal, and we should not strip them.
                 if (nonLow.length >= 2) {
                     relevantFiles = nonLow;
                 }
@@ -2132,6 +2289,64 @@ class ToolHandler {
             }
         }
         // 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.');
@@ -2142,8 +2357,15 @@ class ToolHandler {
         for (const [filePath, group] of sortedFiles) {
             if (filesIncluded >= maxFiles)
                 break;
-            if (totalChars > budget.maxOutputChars * 0.9)
-                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;
@@ -2156,6 +2378,131 @@ class ToolHandler {
             }
             const fileLines = fileContent.split('\n');
             const lang = group.nodes[0]?.language || '';
+            // Adaptive sizing (CODEGRAPH_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, Math.min(n.endLine, n.startLine + 220)).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;
+                const bodyCap = budget.maxCharsPerFile * 2;
+                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, Math.min(n.endLine, n.startLine + 220)).join('\n').length;
+                    // Spine methods (prio 0) ALWAYS get a full body — the cap governs the
+                    // off-path extras (unique-named, family base), never the flow path itself.
+                    if (prio(n) > 0 && 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 = Math.min(n.endLine, n.startLine + 220);
+                        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 codegraph_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 — codegraph_explore a signature by name for its body; do NOT Read)'
+                        : 'skeleton (signatures only — codegraph_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-small-file rule: if a relevant file is small enough to afford,
             // return it ENTIRELY instead of clustering. Clustering exists to tame
             // god-files (App.tsx ~13k lines); on a ~134-line component a cluster is a
@@ -2205,14 +2552,33 @@ class ToolHandler {
             // Alamofire is the canonical case: the `Session` class spans ~1,400
             // lines). We want the granular symbols inside, not the envelope.
             const ENVELOPE_KINDS = new Set(['file', 'module', 'class', 'struct', 'interface', 'enum', 'namespace', 'protocol', 'trait', 'component']);
-            const ranges = group.nodes
-                .filter(n => n.startLine > 0 && n.endLine > 0)
+            // 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))
@@ -2313,6 +2679,13 @@ class ToolHandler {
                     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) {
@@ -2325,7 +2698,7 @@ class ToolHandler {
                     projectedChars += sectionLen;
                     continue;
                 }
-                if (projectedChars + sectionLen > budget.maxCharsPerFile)
+                if (projectedChars + sectionLen > fileBudget)
                     continue;
                 chosenIndices.add(rc.idx);
                 projectedChars += sectionLen;
@@ -2371,22 +2744,22 @@ class ToolHandler {
                 ? `${headerSymbols.join(', ')}, +${omittedCount} more`
                 : headerSymbols.join(', ');
             const fileHeader = `#### ${filePath} — ${headerSuffix}`;
-            // Respect the total output cap on a file-by-file basis.
-            if (totalChars + fileSection.length + 200 > budget.maxOutputChars) {
+            // 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) {
                 const remaining = budget.maxOutputChars - totalChars - 200;
                 if (remaining < 500)
-                    break;
-                const trimmed = fileSection.slice(0, remaining) + '\n... (trimmed) ...';
-                lines.push(fileHeader);
-                lines.push('');
-                lines.push('```' + lang);
-                lines.push(trimmed);
-                lines.push('```');
-                lines.push('');
-                totalChars += trimmed.length + 200;
-                filesIncluded++;
+                    continue; // incidental file, no room — skip it, keep scanning for necessary ones
+                fileSection = fileSection.slice(0, remaining) + '\n... (trimmed) ...';
                 anyFileTrimmed = true;
-                break;
             }
             lines.push(fileHeader);
             lines.push('');
@@ -2449,11 +2822,20 @@ class ToolHandler {
         // maxOutputChars (observed 30k against a 28k tier cap). A fat explore
         // payload persists in the agent's context and is re-read as cache-input
         // on every subsequent turn, so the overrun is paid many times over.
-        const output = this.buildFlowFromNamedSymbols(cg, query) + lines.join('\n');
-        if (output.length > budget.maxOutputChars) {
-            const cut = output.slice(0, budget.maxOutputChars);
+        // Final ceiling. The render loop is now the authority on WHAT to emit — it
+        // renders necessary files (named/spine) even past maxOutputChars and caps
+        // only incidental ones, all bounded by maxFiles + per-file true-spine — so
+        // this is a SAFETY ceiling above that necessary content, not a hard cut
+        // through it. Cutting at a flat maxOutputChars here undid the whole point:
+        // Alamofire's loop assembles build+validators-exec+validate (~15K) and a 13K
+        // slice dropped the validate phase the agent then Read. Allow necessary
+        // overflow up to 1.5× (still bounds a pathological monolith).
+        const output = flow.text + lines.join('\n');
+        const hardCeiling = Math.round(budget.maxOutputChars * 1.5);
+        if (output.length > hardCeiling) {
+            const cut = output.slice(0, hardCeiling);
             const lastNewline = cut.lastIndexOf('\n');
-            const safe = lastNewline > budget.maxOutputChars * 0.8 ? cut.slice(0, lastNewline) : cut;
+            const safe = lastNewline > hardCeiling * 0.8 ? cut.slice(0, lastNewline) : cut;
             return this.textResult(safe + '\n\n... (output truncated to budget; the source above is complete and verbatim — treat it as already Read. For any area not covered, run another codegraph_explore with the specific names — do NOT Read these files.)');
         }
         return this.textResult(output);