Add skip_knowledge flag to filter irrelevant sessions from knowledge base

src/background.ts

@@ -12,6 +12,7 @@ import { SESSION_DIR } from "./paths.js";
 import { indexSession } from "./qmd.js";
 import { redactSecrets } from "./redact.js";
 import {
+  addFrontmatterField,
   addTags,
   appendDecision,
   appendMistake,
@@ -94,79 +95,90 @@ try {
     // 2. Extract tags, decisions, mistakes
     const sections = extractSections(summary);
 
+    // Check if AI flagged this session as not relevant for knowledge
+    if (sections.skipKnowledge) {
+      log("Session flagged as skip_knowledge by AI — skipping knowledge ingestion");
+      // Mark in frontmatter so downstream consumers can filter
+      const currentContent = readFileSync(sessionPath, "utf8");
+      const updated = addFrontmatterField(currentContent, "skip_knowledge", true);
+      writeFileSync(sessionPath, updated);
+    }
+
     if (sections.tags.length > 0) {
       addTags(repoRoot, sessionId, sections.tags);
       log(`Tagged: ${sections.tags.join(", ")}`);
     }
 
-    // Read session data for context
-    const sessionContent = readFileSync(sessionPath, "utf8");
-    const { frontmatter } = parseFrontmatter(sessionContent);
-    const modifiedFiles = extractModifiedFiles(sessionContent);
-    const commitSha = (frontmatter.base_commit as string) || "unknown";
-    const sessionDate = sessionId.slice(0, 10);
-
-    for (const decision of sections.decisions) {
-      const { title, description } = parseTitleDescription(decision.text);
-      if (isJunkEntry(title)) continue;
-      const files = decision.files.length > 0 ? decision.files : modifiedFiles.slice(0, 5);
-      appendDecision(repoRoot, {
-        title,
-        description,
-        sessionId,
-        commitSha,
-        files,
-        area: deriveArea(files),
-        date: sessionDate,
-        tried: decision.tried,
-        rule: decision.rule,
-      });
-    }
-    if (sections.decisions.length > 0) {
-      log(`Logged ${sections.decisions.length} decision(s)`);
-    }
-
-    for (const mistake of sections.mistakes) {
-      const { title, description } = parseTitleDescription(mistake.text);
-      if (isJunkEntry(title)) continue;
-      const files = mistake.files.length > 0 ? mistake.files : modifiedFiles.slice(0, 5);
-      appendMistake(repoRoot, {
-        title,
-        description,
-        sessionId,
-        commitSha,
-        files,
-        area: deriveArea(files),
-        date: sessionDate,
-        tried: mistake.tried,
-        rule: mistake.rule,
-      });
-    }
-    if (sections.mistakes.length > 0) {
-      log(`Logged ${sections.mistakes.length} mistake(s)`);
-    }
+    if (!sections.skipKnowledge) {
+      // Read session data for context
+      const sessionContent = readFileSync(sessionPath, "utf8");
+      const { frontmatter } = parseFrontmatter(sessionContent);
+      const modifiedFiles = extractModifiedFiles(sessionContent);
+      const commitSha = (frontmatter.base_commit as string) || "unknown";
+      const sessionDate = sessionId.slice(0, 10);
+
+      for (const decision of sections.decisions) {
+        const { title, description } = parseTitleDescription(decision.text);
+        if (isJunkEntry(title)) continue;
+        const files = decision.files.length > 0 ? decision.files : modifiedFiles.slice(0, 5);
+        appendDecision(repoRoot, {
+          title,
+          description,
+          sessionId,
+          commitSha,
+          files,
+          area: deriveArea(files),
+          date: sessionDate,
+          tried: decision.tried,
+          rule: decision.rule,
+        });
+      }
+      if (sections.decisions.length > 0) {
+        log(`Logged ${sections.decisions.length} decision(s)`);
+      }
 
-    // Auto-detect corrections: files modified in consecutive turns
-    const corrections = detectCorrections(sessionContent);
-    if (corrections.length > 0) {
-      const fileCounts: Record<string, number> = {};
-      for (const c of corrections) {
-        fileCounts[c.file] = (fileCounts[c.file] || 0) + 1;
+      for (const mistake of sections.mistakes) {
+        const { title, description } = parseTitleDescription(mistake.text);
+        if (isJunkEntry(title)) continue;
+        const files = mistake.files.length > 0 ? mistake.files : modifiedFiles.slice(0, 5);
+        appendMistake(repoRoot, {
+          title,
+          description,
+          sessionId,
+          commitSha,
+          files,
+          area: deriveArea(files),
+          date: sessionDate,
+          tried: mistake.tried,
+          rule: mistake.rule,
+        });
       }
-      for (const [file, count] of Object.entries(fileCounts)) {
-        if (count >= 2) {
-          appendMistake(repoRoot, {
-            title: `Repeated modifications to ${file}`,
-            description: `File was modified across multiple consecutive turns — may indicate the AI struggled with this file. Review session ${sessionId} for the correct approach.`,
-            sessionId,
-            commitSha,
-            files: [file],
-            area: deriveArea([file]),
-            date: sessionDate,
-            tried: [],
-            rule: "",
-          });
-          log(`Auto-detected correction pattern for ${file}`);
+      if (sections.mistakes.length > 0) {
+        log(`Logged ${sections.mistakes.length} mistake(s)`);
+      }
+
+      // Auto-detect corrections: files modified in consecutive turns
+      const corrections = detectCorrections(sessionContent);
+      if (corrections.length > 0) {
+        const fileCounts: Record<string, number> = {};
+        for (const c of corrections) {
+          fileCounts[c.file] = (fileCounts[c.file] || 0) + 1;
+        }
+        for (const [file, count] of Object.entries(fileCounts)) {
+          if (count >= 2) {
+            appendMistake(repoRoot, {
+              title: `Repeated modifications to ${file}`,
+              description: `File was modified across multiple consecutive turns — may indicate the AI struggled with this file. Review session ${sessionId} for the correct approach.`,
+              sessionId,
+              commitSha,
+              files: [file],
+              area: deriveArea([file]),
+              date: sessionDate,
+              tried: [],
+              rule: "",
+            });
+            log(`Auto-detected correction pattern for ${file}`);
+          }
         }
       }
     }

src/index.ts

@@ -327,7 +327,8 @@ if (import.meta.main) {
           const { frontmatter } = parseFrontmatter(content);
           const tags = (frontmatter.tags as string[]) || [];
           const tagStr = tags.length > 0 ? ` ${c.dim}[${tags.join(", ")}]${c.reset}` : "";
-          console.log(`${c.cyan}${id}${c.reset} ${c.dim}${frontmatter.branch || ""}${c.reset}${tagStr}`);
+          const skipStr = frontmatter.skip_knowledge ? ` ${c.dim}(skipped)${c.reset}` : "";
+          console.log(`${c.cyan}${id}${c.reset} ${c.dim}${frontmatter.branch || ""}${c.reset}${tagStr}${skipStr}`);
         }
         if (sessions.length > count) {
           console.log(`${c.dim}... and ${sessions.length - count} more${c.reset}`);

src/knowledge.ts

@@ -4,7 +4,14 @@ import { checkClaude } from "./deps.js";
 import { completedDir, decisionsPath, knowledgePath, mistakesPath, SESSION_DIR } from "./paths.js";
 import { searchSessions } from "./qmd.js";
 import type { KnowledgeEntry } from "./session.js";
-import { appendDecision, appendMistake, deriveArea, listCompletedSessions, parseKnowledgeEntries } from "./session.js";
+import {
+  appendDecision,
+  appendMistake,
+  deriveArea,
+  listCompletedSessions,
+  parseFrontmatter,
+  parseKnowledgeEntries,
+} from "./session.js";
 
 // =============================================================================
 // Claude CLI Check
@@ -66,17 +73,26 @@ export async function buildKnowledge(repoRoot: string): Promise<void> {
     return;
   }
 
-  // Gather all session summaries
+  // Gather all session summaries (skip sessions flagged as not relevant)
   const summaries: string[] = [];
+  let skippedCount = 0;
   for (const id of sessions) {
     const path = join(completedDir(repoRoot), `${id}.md`);
     if (!existsSync(path)) continue;
     const content = readFileSync(path, "utf8");
+    const { frontmatter } = parseFrontmatter(content);
+    if (frontmatter.skip_knowledge) {
+      skippedCount++;
+      continue;
+    }
     const summaryMatch = content.match(/## Summary\n([\s\S]*?)$/);
     if (summaryMatch) {
       summaries.push(`### Session ${id}\n${summaryMatch[1]!.trim()}`);
     }
   }
+  if (skippedCount > 0) {
+    console.log(`Skipped ${skippedCount} session(s) flagged as not relevant.`);
+  }
 
   if (summaries.length === 0) {
     console.log("No session summaries found. Run sessions first or wait for AI summarization.");

src/session.ts

@@ -1027,7 +1027,7 @@ export function parseFrontmatter(content: string): { frontmatter: Record<string,
 }
 
 /** Add a field to existing frontmatter */
-function addFrontmatterField(content: string, key: string, value: unknown): string {
+export function addFrontmatterField(content: string, key: string, value: unknown): string {
   const parsed = parseFrontmatter(content);
   parsed.frontmatter[key] = value;
   return `---\n${YAML.stringify(parsed.frontmatter).trim()}\n---\n${parsed.body}`;

src/summarize.test.ts

@@ -193,6 +193,85 @@ Rule: WHEN adding cache layers ALWAYS use Redis`;
   });
 });
 
+describe("extractSections — skipKnowledge", () => {
+  test("returns skipKnowledge=true when Relevance is 'skip'", () => {
+    const summary = `## Intent
+Testing the AI tool.
+
+## Changes
+- No real changes.
+
+## Decisions
+None
+
+## Mistakes
+None
+
+## Open Items
+None
+
+## Relevance
+skip
+
+## Tags
+test`;
+
+    const sections = extractSections(summary);
+    expect(sections.skipKnowledge).toBe(true);
+  });
+
+  test("returns skipKnowledge=false when Relevance is 'keep'", () => {
+    const summary = `## Intent
+Migrate cart system.
+
+## Changes
+- Updated fees.
+
+## Decisions
+None
+
+## Mistakes
+None
+
+## Open Items
+None
+
+## Relevance
+keep
+
+## Tags
+area:cart`;
+
+    const sections = extractSections(summary);
+    expect(sections.skipKnowledge).toBe(false);
+  });
+
+  test("returns skipKnowledge=false when Relevance section is missing", () => {
+    const summary = `## Intent
+Quick fix.
+
+## Tags
+fix`;
+
+    const sections = extractSections(summary);
+    expect(sections.skipKnowledge).toBe(false);
+  });
+
+  test("handles case-insensitive 'Skip'", () => {
+    const summary = `## Intent
+Test chat.
+
+## Relevance
+Skip
+
+## Tags
+test`;
+
+    const sections = extractSections(summary);
+    expect(sections.skipKnowledge).toBe(true);
+  });
+});
+
 describe("extractMistakeEntries — junk filtering", () => {
   test("filters 'None' variations", () => {
     expect(extractMistakeEntries("## Mistakes\nNone")).toEqual([]);

src/summarize.ts

@@ -41,6 +41,14 @@ The Tried:, Files:, and Rule: lines are optional — omit if not applicable.
 ## Open Items
 Anything left unfinished or flagged for follow-up
 
+## Relevance
+Assess whether this session contains knowledge worth preserving for future project context.
+Write "skip" on its own line if the session is NOT relevant — e.g. test/demo chats, conversations about
+the AI itself, troubleshooting the tooling rather than the project, trivial or throwaway interactions,
+or anything that would add noise rather than signal to the project knowledge base.
+Write "keep" on its own line if the session contains useful project context (decisions, patterns, mistakes, etc.).
+Default to "keep" if uncertain.
+
 ## Tags
 Comma-separated topic tags inferred from the session content.
 Use namespace:value format where appropriate (e.g. area:cart, type:bug-fix).`;
@@ -87,6 +95,7 @@ export interface ExtractedSections {
   intent: string;
   changes: string;
   openItems: string;
+  skipKnowledge: boolean;
 }
 
 /** Extract structured sections from an AI summary */
@@ -98,9 +107,17 @@ export function extractSections(summary: string): ExtractedSections {
     intent: extractNamedSection(summary, "Intent"),
     changes: extractNamedSection(summary, "Changes"),
     openItems: extractNamedSection(summary, "Open Items"),
+    skipKnowledge: extractSkipKnowledge(summary),
   };
 }
 
+/** Check if the AI flagged this session as not worth preserving */
+function extractSkipKnowledge(summary: string): boolean {
+  const section = extractNamedSection(summary, "Relevance");
+  if (!section) return false;
+  return /^\s*skip\s*$/im.test(section);
+}
+
 /** Extract tags from the Tags section */
 function extractTags(summary: string): string[] {
   const section = extractNamedSection(summary, "Tags");