gitmem-mcp 1.6.5 → 1.6.6

package/CHANGELOG.md

@@ -7,6 +7,11 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
+## [1.6.6] - 2026-06-27
+
+### Changed
+- **`recall` stubs low-confidence scars to cut wasted tokens**: scars scoring below the `0.55` similarity threshold (already flagged `[low confidence]`, ~66% N/A rate in that band) now render as a one-line stub — title, severity, score, short id, and the `[low confidence]` tag — instead of hydrating their full body (description, counter-arguments, applies-when, why-this-matters, action-protocol, self-check, related triples). High-confidence scars (≥ 0.55) are unchanged, and blocking-verification scars always render in full regardless of score. The `0.55` cutoff is now a single named constant (`LOW_CONFIDENCE_THRESHOLD`) shared by the tag and the stub so they can't drift apart. The scar's id stays visible, so an agent can still pull detail on demand. (GIT-49)
+
 ## [1.6.5] - 2026-06-27
 
 ### Fixed

package/dist/tools/recall.d.ts

@@ -15,6 +15,16 @@
 import type { KnowledgeTriple } from "../services/supabase-client.js";
 import { type ScarWithVariant } from "../services/variant-assignment.js";
 import type { Project, PerformanceData } from "../types/index.js";
+/**
+ * Confidence cutoff for scar rendering.
+ *
+ * Matches below this similarity are flagged `[low confidence]` and, per the
+ * 1.5.0 UX-audit calibration, are N/A ~66% of the time. They pass the pro-tier
+ * 0.45 inclusion filter but are rendered as stubs (header only) rather than full
+ * bodies — the tag boundary and the stub boundary share this constant so they
+ * can never drift apart.
+ */
+export declare const LOW_CONFIDENCE_THRESHOLD = 0.55;
 /**
  * Parameters for recall tool
  */

package/dist/tools/recall.js

@@ -27,6 +27,16 @@ import { getSessionPath } from "../services/gitmem-dir.js";
 import { wrapDisplay, SEV, dimText, ANSI } from "../services/display-protocol.js";
 import { formatNudgeHeader } from "../services/nudge-variants.js";
 import { fetchDismissalCounts } from "../services/behavioral-decay.js";
+/**
+ * Confidence cutoff for scar rendering.
+ *
+ * Matches below this similarity are flagged `[low confidence]` and, per the
+ * 1.5.0 UX-audit calibration, are N/A ~66% of the time. They pass the pro-tier
+ * 0.45 inclusion filter but are rendered as stubs (header only) rather than full
+ * bodies — the tag boundary and the stub boundary share this constant so they
+ * can never drift apart.
+ */
+export const LOW_CONFIDENCE_THRESHOLD = 0.55;
 /**
  * Format scars into a readable response for Claude
  */
@@ -80,8 +90,8 @@ No past lessons match this plan closely enough. Scars accumulate as you work —
     for (const scar of scars) {
         const sev = SEV[scar.severity] || "[?]";
         const starterTag = scar.is_starter ? ` ${dimText("[starter]")}` : "";
-        // Confidence tier: marginal matches (< 0.55) get flagged — 66% N/A rate in this range
-        const confidenceTag = scar.similarity < 0.55 ? ` ${dimText("[low confidence]")}` : "";
+        // Confidence tier: marginal matches get flagged — 66% N/A rate below the threshold
+        const confidenceTag = scar.similarity < LOW_CONFIDENCE_THRESHOLD ? ` ${dimText("[low confidence]")}` : "";
         // Pro: decay tag for scars with reduced behavioral relevance
         const decayTag = hasProInsights() && scar.decay_multiplier !== undefined && scar.decay_multiplier < 0.8
             ? ` ${dimText(`[decay: ${Math.round(scar.decay_multiplier * 100)}%]`)}`
@@ -94,55 +104,61 @@ No past lessons match this plan closely enough. Scars accumulate as you work —
                 lines.push(`  _[dismissed ${counts.dismissed}/${counts.surfaced} times — re-evaluate whether this still applies]_`);
             }
         }
-        // Use variant enforcement text if available (blind to variant name)
-        if (scar.variant_info?.has_variants && scar.variant_info.variant) {
-            const variantText = formatVariantEnforcement(scar.variant_info.variant, scar.title);
-            lines.push(variantText);
-        }
-        else {
-            // Legacy path: use original scar description
-            lines.push(scar.description);
-        }
-        if (scar.counter_arguments.length > 0) {
-            lines.push("");
-            lines.push("*You might think:*");
-            for (const counter of scar.counter_arguments.slice(0, 2)) {
-                lines.push(`  - ${counter}`);
+        // Stub low-confidence scars: render header only, skip the heavy body.
+        // ~66% of sub-threshold matches are N/A, so their full bodies are wasted
+        // tokens. Blocking-verification scars always render full regardless of score.
+        const isStub = scar.similarity < LOW_CONFIDENCE_THRESHOLD && !scar.required_verification?.blocking;
+        if (!isStub) {
+            // Use variant enforcement text if available (blind to variant name)
+            if (scar.variant_info?.has_variants && scar.variant_info.variant) {
+                const variantText = formatVariantEnforcement(scar.variant_info.variant, scar.title);
+                lines.push(variantText);
             }
-        }
-        if (scar.applies_when.length > 0) {
-            lines.push("");
-            lines.push("*Applies when:* " + scar.applies_when.slice(0, 3).join(", "));
-        }
-        // Render LLM-cooperative enforcement fields
-        if (scar.why_this_matters) {
-            lines.push("");
-            lines.push(`**Why this matters:** ${scar.why_this_matters}`);
-        }
-        if (scar.action_protocol && scar.action_protocol.length > 0) {
-            lines.push("");
-            lines.push("**Action Protocol:**");
-            scar.action_protocol.forEach((step, i) => {
-                lines.push(`  ${i + 1}. ${step}`);
-            });
-        }
-        if (scar.self_check_criteria && scar.self_check_criteria.length > 0) {
-            lines.push("");
-            lines.push("**Self-Check:**");
-            for (const criterion of scar.self_check_criteria) {
-                lines.push(`  - [ ] ${criterion}`);
+            else {
+                // Legacy path: use original scar description
+                lines.push(scar.description);
             }
-        }
-        // Render related knowledge triples
-        if (scar.related_triples && scar.related_triples.length > 0) {
-            lines.push("");
-            lines.push("*Related knowledge:*");
-            for (const triple of scar.related_triples) {
-                lines.push(`  - ${triple.subject} **${triple.predicate}** ${triple.object}`);
+            if (scar.counter_arguments.length > 0) {
+                lines.push("");
+                lines.push("*You might think:*");
+                for (const counter of scar.counter_arguments.slice(0, 2)) {
+                    lines.push(`  - ${counter}`);
+                }
+            }
+            if (scar.applies_when.length > 0) {
+                lines.push("");
+                lines.push("*Applies when:* " + scar.applies_when.slice(0, 3).join(", "));
+            }
+            // Render LLM-cooperative enforcement fields
+            if (scar.why_this_matters) {
+                lines.push("");
+                lines.push(`**Why this matters:** ${scar.why_this_matters}`);
+            }
+            if (scar.action_protocol && scar.action_protocol.length > 0) {
+                lines.push("");
+                lines.push("**Action Protocol:**");
+                scar.action_protocol.forEach((step, i) => {
+                    lines.push(`  ${i + 1}. ${step}`);
+                });
+            }
+            if (scar.self_check_criteria && scar.self_check_criteria.length > 0) {
+                lines.push("");
+                lines.push("**Self-Check:**");
+                for (const criterion of scar.self_check_criteria) {
+                    lines.push(`  - [ ] ${criterion}`);
+                }
+            }
+            // Render related knowledge triples
+            if (scar.related_triples && scar.related_triples.length > 0) {
+                lines.push("");
+                lines.push("*Related knowledge:*");
+                for (const triple of scar.related_triples) {
+                    lines.push(`  - ${triple.subject} **${triple.predicate}** ${triple.object}`);
+                }
+            }
+            if (scar.source_issue) {
+                lines.push(`*Source:* ${scar.source_issue}`);
             }
-        }
-        if (scar.source_issue) {
-            lines.push(`*Source:* ${scar.source_issue}`);
         }
         lines.push("");
         lines.push("---");

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "gitmem-mcp",
-  "version": "1.6.5",
+  "version": "1.6.6",
   "mcpName": "io.github.gitmem-dev/gitmem",
   "description": "Persistent learning memory for AI coding agents. Memory that compounds.",
   "type": "module",