@bacnh85/pi-model-tools 0.3.1 → 0.3.2

package/extensions/index.ts

@@ -345,8 +345,12 @@ export default function (pi: ExtensionAPI) {
       if (readHint) systemPrompt = `${systemPrompt}\n\n${readHint}`;
     }
 
-    // apply_patch preference — DeepSeek V4 Flash only (GLM is better with exact-match
-    // edit; apply_patch context anchors confuse it and it lacks agentic fallback).
+    // apply_patch preference — all DeepSeek V4 (flash+pro); GLM excluded per eval.
+    // Eval (2026-07-29, 15 trials, 3 targets) showed all models use edit with zero
+    // edit_mismatch errors. DeepSeek keeps guidance as a safety net for real-world
+    // multi-file/frontmatter edits beyond the eval's scope; GLM excluded because it
+    // doesn't receive the suite of DeepSeek-specific steering (Super Power, selection
+    // guidance, semantic-miss blocking) and thus doesn't need the companion hint.
     if (activeFamily === "deepseek-v4") {
       const patchHint = applyPatchPreferenceGuidance(activeForHint);
       if (patchHint) systemPrompt = `${systemPrompt}\n\n${patchHint}`;

package/extensions/lib/guidance.ts

@@ -96,18 +96,21 @@ export function deepSeekSelectionGuidance(activeTools: readonly string[]): strin
 // ── apply_patch preference hint (DeepSeek/GLM) ──
 
 /**
- * When apply_patch is available, tell weaker models to prefer it over edit for
- * anything beyond a single tiny replacement. Returns undefined if the tool is
- * not active so it never misdirects. Gated to DeepSeek/GLM by the caller (see
- * index.ts before_agent_start), since Claude/OpenAI are already reliable with edit.
+ * When apply_patch is available, tell the model to prefer it over edit for
+ * anything beyond a small change (>3 lines, YAML frontmatter, or
+ * multi-hunk/multi-file). Returns undefined if the tool is not active so it
+ * never misdirects. Gated per family by the caller (see index.ts
+ * before_agent_start).
  */
 export function applyPatchPreferenceGuidance(activeTools: readonly string[]): string | undefined {
   if (!activeTools.includes("apply_patch")) return undefined;
   return [
     "apply_patch (preferred for non-trivial edits):",
-    "  • For multi-line, multi-hunk, or multi-file edits, emit a small V4D diff (context + -/+ lines) via apply_patch instead of reproducing large verbatim oldText blocks — far fewer match failures.",
+    "  • For multi-line (>3 lines), multi-hunk, or multi-file edits, emit a small V4D diff (context + -/+ lines) via apply_patch instead of reproducing large verbatim oldText blocks — far fewer match failures.",
     "  • Each Update hunk must anchor on enough unchanged context that the context+removed block matches UNIQUELY in the file.",
-    "  • Keep using edit for a single tiny one-line exact replacement.",
+    "  • Keep using edit for a single tiny one-line exact replacement (≤3 lines).",
+    "  • Touching frontmatter (YAML) — always use apply_patch. YAML is whitespace-sensitive; edit\'s exact-match oldText rarely succeeds on the first try.",
+    "  • One-strike rule: if edit fails once, stop — switch to apply_patch, do not retry edit.",
   ].join("\n");
 }
 

package/package.json

@@ -1,6 +1,6 @@
 {
     "name": "@bacnh85/pi-model-tools",
-    "version": "0.3.1",
+    "version": "0.3.2",
     "description": "Unified tool-wrapping, argument repair, reasoning management, DeepSeek V4 guidance + Super Power Mode, defensive leak-cleaning, edit mismatch repair (read-notice stripping + trim-tolerant retry), and a Codex-style apply_patch diff tool for DeepSeek V4 and GLM model families from any provider.",
     "type": "module",
     "license": "MIT",