pi-tool-repair 0.1.6 → 0.1.8

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-tool-repair",
-  "version": "0.1.6",
+  "version": "0.1.8",
   "description": "Validate-then-repair extension for pi — fixes common LLM tool-call mistakes (null fields, stringified arrays, wrong field names, anchor bleed) before tools execute",
   "type": "module",
   "author": "Tom X Nguyen",

package/src/grammar-repair.ts

@@ -156,6 +156,21 @@ export function repairAssistantMessageGrammarLeaks(
     changed = true;
     for (const candidate of candidates) {
       if (candidate.stripOnly) continue;
+
+      // Safety: don't recover tool calls whose grammar parsed to an empty
+      // argument object. These are almost always malformed GLM/XML fragments
+      // (e.g. `<tool_call>write</tool_call>`) that would otherwise crash as a
+      // real tool invocation with missing required properties.
+      if (Object.keys(candidate.arguments).length === 0) {
+        if (config.debug) {
+          process.stderr.write(
+            `[pi-tool-repair] grammar-repair skipping ${candidate.grammar}:${candidate.name} ` +
+            "(empty arguments, likely malformed grammar)\n",
+          );
+        }
+        continue;
+      }
+
       recoveredCalls.push({
         name: candidate.name,
         arguments: candidate.arguments,

package/src/index.ts

@@ -12,6 +12,7 @@ import type { MinimalAssistantMessage } from "./grammar-repair.js";
 export interface RepairConfig {
   debug: boolean;
   anchorBleedModels: RegExp[];
+  grammarLeakModels: RegExp[];
   fieldAliases: Record<string, Record<string, string[]>>;
   stringArgTools: Record<string, { field: string; shape: "string" | "array" }>;
 }
@@ -23,6 +24,9 @@ export const DEFAULT_CONFIG: RepairConfig = {
     /minimax/i,
     /glm/i,
   ],
+  grammarLeakModels: [
+    /glm/i,
+  ],
   fieldAliases: {
     read: {
       path: [
@@ -80,6 +84,11 @@ export function hasAnchorBleedBug(model: { id?: string } | null | undefined): bo
   return DEFAULT_CONFIG.anchorBleedModels.some((re) => re.test(model.id!));
 }
 
+export function hasGrammarLeakBug(model: { id?: string } | null | undefined): boolean {
+  if (!model || !model.id) return false;
+  return DEFAULT_CONFIG.grammarLeakModels.some((re) => re.test(model.id!));
+}
+
 export function sanitizePattern(pattern: string): string | undefined {
   if (pattern.includes("|") && (pattern.includes("^") || pattern.includes("$"))) {
     return undefined;

package/tool-repair.ts

@@ -43,6 +43,7 @@ import type { ExtensionAPI } from "@earendil-works/pi-coding-agent";
 import {
   DEFAULT_CONFIG,
   hasAnchorBleedBug,
+  hasGrammarLeakBug,
   sanitizeSchemaAnchors,
   stripAnchorBleedInPlace,
   stripGrammarTokenLeaksInPlace,
@@ -240,9 +241,10 @@ export default function (pi: ExtensionAPI) {
       }
     }
 
-    // Defense-in-depth: strip leaked grammar tokens (e.g. GLM <arg_key>) from
-    // parsed tool-call keys/values before the schema sees them.
-    if (input && typeof input === "object") {
+    // Strip leaked grammar tokens (e.g. GLM <arg_key>) from parsed tool-call
+    // keys/values before the schema sees them. Only for models known to produce
+    // grammar token leaks.
+    if (input && typeof input === "object" && model && hasGrammarLeakBug(model)) {
       stripGrammarTokenLeaksInPlace(input as Record<string, unknown>);
     }