feat(parser): raise default token budget from 2048 to 4096 (#845)

.bench/baseline.json

@@ -1,19 +1,19 @@
 {
-  "capturedAt": "2026-05-06T00:45:56.229Z",
+  "capturedAt": "2026-05-06T00:58:34.994Z",
   "node": "v22.13.0",
   "platform": "darwin-arm64",
   "options": {
     "baseLatencyMs": 1500,
     "perTokenMs": 2,
     "maxConcurrent": 6,
-    "maxTokens": 2048
+    "maxTokens": 4096
   },
   "results": [
     {
       "fixture": "tiny",
       "fileCount": 5,
       "approxTokens": 790,
-      "durationMs": 1,
+      "durationMs": 2,
       "llmCalls": 0,
       "llmTotalMs": 0,
       "llmTotalPromptTokens": 0
@@ -22,55 +22,55 @@
       "fixture": "medium",
       "fileCount": 25,
       "approxTokens": 36150,
-      "durationMs": 31137,
-      "llmCalls": 20,
-      "llmTotalMs": 106348,
-      "llmTotalPromptTokens": 34237
+      "durationMs": 29267,
+      "llmCalls": 19,
+      "llmTotalMs": 109679,
+      "llmTotalPromptTokens": 36895
     },
     {
       "fixture": "large",
       "fileCount": 50,
       "approxTokens": 83410,
-      "durationMs": 72093,
-      "llmCalls": 41,
-      "llmTotalMs": 244101,
-      "llmTotalPromptTokens": 74197
+      "durationMs": 59992,
+      "llmCalls": 30,
+      "llmTotalMs": 228089,
+      "llmTotalPromptTokens": 74609
     },
     {
       "fixture": "feature-add",
       "fileCount": 14,
       "approxTokens": 17600,
-      "durationMs": 15967,
+      "durationMs": 19591,
       "llmCalls": 11,
-      "llmTotalMs": 54727,
-      "llmTotalPromptTokens": 18937
+      "llmTotalMs": 59354,
+      "llmTotalPromptTokens": 20707
     },
     {
       "fixture": "refactor",
       "fileCount": 30,
       "approxTokens": 32650,
-      "durationMs": 33999,
-      "llmCalls": 28,
-      "llmTotalMs": 153888,
-      "llmTotalPromptTokens": 52430
+      "durationMs": 41340,
+      "llmCalls": 20,
+      "llmTotalMs": 143983,
+      "llmTotalPromptTokens": 53548
     },
     {
       "fixture": "initial-commit",
       "fileCount": 50,
       "approxTokens": 83410,
-      "durationMs": 72285,
-      "llmCalls": 41,
-      "llmTotalMs": 245148,
-      "llmTotalPromptTokens": 74546
+      "durationMs": 60034,
+      "llmCalls": 30,
+      "llmTotalMs": 229291,
+      "llmTotalPromptTokens": 74948
     },
     {
       "fixture": "docs-update",
       "fileCount": 9,
       "approxTokens": 15050,
-      "durationMs": 18570,
-      "llmCalls": 8,
-      "llmTotalMs": 56293,
-      "llmTotalPromptTokens": 13908
+      "durationMs": 18563,
+      "llmCalls": 7,
+      "llmTotalMs": 52225,
+      "llmTotalPromptTokens": 13139
     },
     {
       "fixture": "dep-bump",

bin/benchmark.ts

@@ -71,7 +71,11 @@ const DEFAULT_OPTIONS: BenchOptions = {
   baseLatencyMs: 1500,
   perTokenMs: 2,
   maxConcurrent: 6,
-  maxTokens: 2048,
+  // Match the canonical service tokenLimit from `langchain/utils.ts`
+  // (raised from 2048 to 4096 in PR 1 of #845). The bench mirrors
+  // the most-common production budget so per-PR diffs reflect what
+  // real users will see.
+  maxTokens: 4096,
 }
 
 type BenchResult = {

src/lib/parsers/default/index.ts

@@ -49,10 +49,17 @@ export async function fileChangeParser({
   // 1. Pre-process large files to prevent bias
   // 2. Group by directory and assess token count
   // 3. Wave-based parallel summarization until under budget
+  //
+  // The 4096 fallback (#845) matches the default service configs
+  // for openai / anthropic / ollama (`langchain/utils.ts`). It's a
+  // safety net for users with custom service definitions that omit
+  // `tokenLimit` — without it those users hit a degenerate 2048
+  // budget that triggers needless pre-summarization on diffs the
+  // model could absorb whole.
   logger.startTimer()
   const summary = await summarizeDiffs(diffs, {
     tokenizer,
-    maxTokens: maxTokens || 2048,
+    maxTokens: maxTokens || 4096,
     minTokensForSummary,
     maxFileTokens,
     maxConcurrent,

src/lib/parsers/default/utils/summarizeDiffs.ts

@@ -247,7 +247,15 @@ export async function summarizeDiffs(
   {
     tokenizer,
     logger,
-    maxTokens = 2048,
+    // Default raised to 4096 (#845) so the budget matches the
+    // canonical service configs in `langchain/utils.ts`. The
+    // previous 2048 default came from an earlier era when 4k
+    // context was a stretch for fast models; today every shipped
+    // service overrides it to 4096 anyway. Keeping this in sync
+    // with the service defaults means a caller that omits
+    // `maxTokens` doesn't accidentally fall into a tighter budget
+    // than the rest of the system assumes.
+    maxTokens = 4096,
     minTokensForSummary = 400,
     maxFileTokens,
     maxConcurrent = 6,