@vellumai/assistant 0.11.7-dev.202609010322.2bc955e → 0.11.7-dev.202609011017.3dd6343

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@vellumai/assistant",
-  "version": "0.11.7-dev.202609010322.2bc955e",
+  "version": "0.11.7-dev.202609011017.3dd6343",
   "license": "MIT",
   "type": "module",
   "exports": {

package/src/__tests__/file-ops-service.test.ts

@@ -10,7 +10,8 @@ import { tmpdir } from "node:os";
 import { join } from "node:path";
 import { afterEach, describe, expect, test } from "bun:test";
 
-import { THRESHOLD_CHARS } from "../context/post-turn-tool-result-truncation.js";
+import { FILE_READ_TOOL_NAMES } from "../context/post-turn-tool-result-truncation.js";
+import { RESULT_TIME_SPOOL_EXEMPT_TOOLS } from "../context/tool-result-spool.js";
 import {
   FileSystemOps,
   type PathPolicy,
@@ -330,8 +331,12 @@ describe("FileSystemOps.readFileSafe", () => {
     expect(body).toBe("abcd");
   });
 
-  test("the read budget stays under the tool-result spool threshold", () => {
-    expect(READ_CHAR_BUDGET).toBeLessThan(THRESHOLD_CHARS);
+  test("file reads are result-time spool-exempt so a full window survives its turn", () => {
+    // The budget exceeds the spool threshold, so without the exemption a
+    // default read of a large file would be stubbed before the model saw it.
+    for (const name of FILE_READ_TOOL_NAMES) {
+      expect(RESULT_TIME_SPOOL_EXEMPT_TOOLS.has(name)).toBe(true);
+    }
   });
 });
 

package/src/__tests__/tool-result-spool.test.ts

@@ -140,11 +140,12 @@ describe("spoolAndStubOversizedToolResults", () => {
     expect(existsSync(join(convDir, TOOL_RESULT_DIR))).toBe(false);
   });
 
-  test("spools an oversized file read of an ordinary path", () => {
+  test("leaves an oversized file read of an ordinary path inline (explicit self-sized read)", () => {
     const blocks = [
       makeToolResult(LONG, "tu_read"),
       makeToolResult(LONG, "tu_host_read"),
     ];
+    const originals = blocks.map((b) => b);
 
     const count = spoolAndStubOversizedToolResults(blocks, {
       conversationDir: convDir,
@@ -154,13 +155,9 @@ describe("spoolAndStubOversizedToolResults", () => {
           : { name: "host_file_read", input: { path: "/Users/me/notes.md" } },
     });
 
-    expect(count).toBe(2);
-    expect((blocks[0] as { content: string }).content).toContain(
-      TRUNCATION_MARKER,
-    );
-    expect((blocks[1] as { content: string }).content).toContain(
-      TRUNCATION_MARKER,
-    );
+    expect(count).toBe(0);
+    expect(blocks).toEqual(originals);
+    expect(existsSync(join(convDir, TOOL_RESULT_DIR))).toBe(false);
   });
 
   test("leaves a file read of a spooled .tool-results path inline", () => {

package/src/context/tool-result-spool.ts

@@ -20,15 +20,22 @@ const AX_TREE_TAG = "<ax-tree>";
 
 /**
  * Tools whose results are explicit, self-sized reads the model paginates
- * itself: `web_fetch` takes `max_chars`/`start_index` and reports a
- * "Character Window: X-Y of Z", so the model pages by character index against
- * the window it requested. Stubbing such a result at result time hands the
- * model ~a tenth of the window it just sized — and it keeps paging blind
- * against content it never saw. Like the file-read tools, these explicit
- * reads are honored in full for the turn that made them; the post-turn pass
- * still truncates them at turn end, after the model has consumed the content.
+ * itself: each takes `max_chars`/`start_index` and reports where its window
+ * ended (`web_fetch`'s "Character Window: X-Y of Z", the file-read tools'
+ * "[Truncated: characters X-Y of Z]" notice), so the model pages by character
+ * index against the window it requested. Stubbing such a result at result
+ * time would hand the model a fraction of the window it just sized — and it
+ * keeps paging blind against content it never saw. These reads are honored in
+ * full for the turn that made them; the post-turn pass still truncates them
+ * at turn end, after the model has consumed the content. This exemption is
+ * also what lets `READ_CHAR_BUDGET` exceed `THRESHOLD_CHARS` without a fresh
+ * read being stubbed before the model sees it.
  */
-export const RESULT_TIME_SPOOL_EXEMPT_TOOLS = new Set<string>(["web_fetch"]);
+export const RESULT_TIME_SPOOL_EXEMPT_TOOLS = new Set<string>([
+  "web_fetch",
+  "file_read",
+  "host_file_read",
+]);
 
 /**
  * Whether a tool result is eligible for the result-time spool/stub pass: the
@@ -39,12 +46,12 @@ export const RESULT_TIME_SPOOL_EXEMPT_TOOLS = new Set<string>(["web_fetch"]);
  * The spooled-read exemption is what keeps paging possible: a file read is the
  * model's only way to pull spooled content back into context, so stubbing a
  * read of a `.tool-results/` file would write a fresh copy and hand back
- * another stub, putting that content out of reach for good. It is scoped by
- * target path rather than by tool name, because a file read aimed anywhere
- * else is ordinary oversized output with no such circularity, and exempting it
- * keeps a whole file inline on every LLM call for the rest of the turn.
- * Explicit self-sized reads are honored in full; the post-turn pass still
- * truncates them at turn end, after the model has consumed the content.
+ * another stub, putting that content out of reach for good. The file-read
+ * tools are also name-exempt as explicit self-sized reads, but this
+ * path-scoped guarantee stands on its own so spooled content stays reachable
+ * regardless of how the exempt-tools set evolves. Explicit self-sized reads
+ * are honored in full; the post-turn pass still truncates them at turn end,
+ * after the model has consumed the content.
  */
 function isSpoolEligible(
   tr: ToolResultContent,

package/src/tools/filesystem/read.ts

@@ -45,7 +45,7 @@ export const fileReadInputSchema = z.looseObject({
   max_chars: z
     .number()
     .describe(
-      "Maximum number of characters to read. Defaults to 20000, which is also the ceiling. Text files only.",
+      "Maximum number of characters to read. Defaults to 100000, which is also the ceiling. Text files only.",
     )
     .optional()
     .catch(undefined),
@@ -61,7 +61,7 @@ export const fileReadInputSchema = z.looseObject({
 export const fileReadTool = {
   name: "file_read",
   description:
-    "Read the contents of a file on your own machine. Text reads return the first 20000 characters unless you pass `max_chars`; when a read stops short the result says so, and `start_index` pages on from there. To find where something is in a large file, code_search is cheaper than paging through it. For image files (JPEG, PNG, GIF, WebP), returns the image for visual analysis. For audio files (MP3, WAV, OGG, FLAC, AAC, M4A), returns the audio for listening. Use host_file_read for files on your guardian's device instead.",
+    "Read the contents of a file on your own machine. Text reads return the first 100000 characters unless you pass `max_chars`; when a read stops short the result says so, and `start_index` pages on from there. To find where something is in a large file, code_search is cheaper than paging through it. For image files (JPEG, PNG, GIF, WebP), returns the image for visual analysis. For audio files (MP3, WAV, OGG, FLAC, AAC, M4A), returns the audio for listening. Use host_file_read for files on your guardian's device instead.",
   category: "filesystem",
   executionTarget: "sandbox",
   defaultRiskLevel: RiskLevel.Low,

package/src/tools/host-filesystem/read.ts

@@ -53,7 +53,7 @@ export const hostFileReadInputSchema = z.looseObject({
   max_chars: z
     .number()
     .describe(
-      "Maximum number of characters to read. Defaults to 20000, which is also the ceiling. Text files only.",
+      "Maximum number of characters to read. Defaults to 100000, which is also the ceiling. Text files only.",
     )
     .optional()
     .catch(undefined),
@@ -69,7 +69,7 @@ export const hostFileReadInputSchema = z.looseObject({
 export const hostFileReadTool = {
   name: "host_file_read",
   description:
-    "Read the contents of a file on your guardian's device, including images (JPEG, PNG, GIF, WebP) and audio (MP3, WAV, OGG, FLAC, AAC, M4A). Text reads return the first 20000 characters unless you pass `max_chars`; when a read stops short the result says so, and `start_index` pages on from there. For files on your own machine, use file_read instead.",
+    "Read the contents of a file on your guardian's device, including images (JPEG, PNG, GIF, WebP) and audio (MP3, WAV, OGG, FLAC, AAC, M4A). Text reads return the first 100000 characters unless you pass `max_chars`; when a read stops short the result says so, and `start_index` pages on from there. For files on your own machine, use file_read instead.",
   category: "host-filesystem",
   executionTarget: "host",
   defaultRiskLevel: RiskLevel.Medium,

package/src/tools/shared/filesystem/file-ops-service.ts

@@ -78,12 +78,15 @@ function pathError(
 }
 
 /**
- * Characters returned by a read that names no `max_chars`. Stays under
- * `THRESHOLD_CHARS` in `context/post-turn-tool-result-truncation.ts`, which
- * spools any larger tool result to disk and replaces it inline with a short
- * stub, so a default read returns content rather than a stub.
+ * Characters returned by a read that names no `max_chars`, and the ceiling for
+ * one that does. File reads are exempt from the result-time spool pass
+ * (`RESULT_TIME_SPOOL_EXEMPT_TOOLS` in `context/tool-result-spool.ts`), so the
+ * full window reaches the model for the turn that read it; the post-turn pass
+ * compacts oversized results to recoverable stubs at turn end, and the
+ * `tool-result-truncate` plugin bounds any single result against the model's
+ * context window.
  */
-export const READ_CHAR_BUDGET = 20_000;
+export const READ_CHAR_BUDGET = 100_000;
 
 /**
  * Trailing marker appended when a read stops short of the end of the file. A
@@ -169,8 +172,8 @@ export class FileSystemOps {
     try {
       const raw = await readFile(filePath, "utf-8");
 
-      // A ceiling, not just a default: a larger window would be spooled to
-      // disk and replaced with a stub, returning less than this.
+      // A ceiling, not just a default: an uncapped `max_chars` would put an
+      // arbitrarily large file inline in a single result.
       const maxChars = Math.min(
         READ_CHAR_BUDGET,
         Math.max(0, input.maxChars ?? READ_CHAR_BUDGET),