feat: implement filtering for harness-injected context in workflow an…

src/core/analyzer-workflows.test.ts

@@ -101,6 +101,21 @@ describe('WorkflowAnalyzer', () => {
       expect(result.clusters).toEqual([]);
     });
 
+    it('filters harness-injected AGENTS.md context as noise', () => {
+      const now = Date.now();
+      const injected = '# AGENTS.md instructions for /home/me/project\n\n<INSTRUCTIONS>\nfollow repo conventions\n</INSTRUCTIONS>';
+      const sessions = Array.from({ length: 5 }, (_, i) =>
+        makeSession({
+          sessionId: `inj-${i}`,
+          requests: [makeRequest({ messageText: injected, timestamp: now + i * 1000 })],
+          lastMessageDate: now + i * 1000,
+        })
+      );
+      const analyzer = createAnalyzer(sessions);
+      const result = analyzer.getWorkflowOptimization();
+      expect(result.clusters).toEqual([]);
+    });
+
     it('clusters repeated similar prompts', () => {
       const now = Date.now();
       // Need at least 3 occurrences with same fingerprint

src/core/analyzer-workflows.ts

@@ -7,7 +7,7 @@
 
 import { DateFilter, WorkflowCluster, WorkflowOptimizationData } from './types';
 import { AnalyzerBase } from './analyzer-base';
-import { toDateStr } from './helpers';
+import { isHarnessInjectedContext, toDateStr } from './helpers';
 
 /** Minimum prompt length to consider for clustering (skip trivial messages) */
 const MIN_PROMPT_LEN = 15;
@@ -53,6 +53,8 @@ function normalizePrompt(raw: string): string {
 /** Check if text is likely a system/bot message or noise rather than a user prompt */
 function isNoise(text: string): boolean {
   const t = text.trim();
+  // Harness-injected session-start context (e.g. Codex AGENTS.md / environment context)
+  if (isHarnessInjectedContext(t)) return true;
   // Decorative separators/borders
   if (/^[═─━=\-_*]{10,}/.test(t)) return true;
   // System auth/notification messages

src/core/helpers.ts

@@ -294,3 +294,25 @@ export function classifyWorkType(msg: string): WorkType {
   }
   return 'other';
 }
+
+/**
+ * Harness-injected session-start payloads that are recorded as `user` messages
+ * but are not real user prompts (e.g. Codex injects the repo `AGENTS.md` and
+ * environment/instruction context at session start). These should be treated as
+ * noise so they don't get clustered into workflow/skill recommendations.
+ *
+ * Matches a small set of known leading markers only — deliberately NOT a generic
+ * `#` markdown header or `<tag>` rule, both of which would suppress legitimate
+ * user prompts.
+ */
+const HARNESS_INJECTED_MARKERS: RegExp[] = [
+  /^# AGENTS\.md instructions\b/,
+  /^<environment_context\b/,
+  /^<INSTRUCTIONS\b/,
+  /^<user_instructions\b/,
+];
+
+export function isHarnessInjectedContext(text: string): boolean {
+  const t = text.trimStart();
+  return HARNESS_INJECTED_MARKERS.some(re => re.test(t));
+}

src/core/parser-codex.test.ts

@@ -186,6 +186,29 @@ describe('parseCodexSessions skillsUsed extraction', () => {
       expect(sessions[0].requests[0].skillsUsed).toEqual(['pdf']);
     });
   });
+
+  it('ignores harness-injected AGENTS.md context and captures the real prompt', () => {
+    withCodexFile([
+      { type: 'session_meta', payload: { id: 'sess-inject-1', cwd: '/Users/me/proj' } },
+      { type: 'turn_context', payload: { model: 'gpt-5.3-codex' } },
+      // Session-start injected context recorded as a user response_item.
+      { type: 'response_item', timestamp: '2025-06-15T10:00:00Z',
+        payload: { type: 'message', role: 'user', content: [
+          { type: 'input_text', text: '# AGENTS.md instructions for /Users/me/proj\n\nfollow repo conventions' },
+          { type: 'input_text', text: '<environment_context>\n  <cwd>/Users/me/proj</cwd>\n</environment_context>' },
+        ] } },
+      // The actual user prompt.
+      { type: 'event_msg', timestamp: '2025-06-15T10:00:05Z', payload: { type: 'user_message', message: 'what is this repo about?' } },
+      { type: 'event_msg', timestamp: '2025-06-15T10:00:06Z', payload: { type: 'assistant_message', content: 'a coach' } },
+    ], (sessionsDir) => {
+      const sessions = parseCodexSessions(sessionsDir);
+      expect(sessions).toHaveLength(1);
+      const texts = sessions[0].requests.map(r => r.messageText);
+      // The injected AGENTS.md / environment context must not be captured as a prompt.
+      expect(texts.some(t => t.startsWith('# AGENTS.md instructions'))).toBe(false);
+      expect(texts).toContain('what is this repo about?');
+    });
+  });
 });
 
 describe('findCodexDirs', () => {

src/core/parser-codex.ts

@@ -22,7 +22,7 @@ import * as path from 'path';
 import { StringDecoder } from 'string_decoder';
 import { ModelUsage, Session, SessionRequest } from './types';
 import { assertTrustedPath, createRequest, createSession, detectDevcontainerFromRequests, extractSkillNameFromPath, extractSkillPathsFromText } from './parser-shared';
-import { canonicalizeReasoningEffort, extractReasoningEffortFromModelId } from './helpers';
+import { canonicalizeReasoningEffort, extractReasoningEffortFromModelId, isHarnessInjectedContext } from './helpers';
 
 interface CodexLine {
   type: string;
@@ -268,6 +268,9 @@ function extractFilePath(args: Record<string, unknown> | null | undefined): stri
 
 function handleUserMessageEvent(payload: Record<string, unknown>, state: CodexParseState, ts: number | null, defaultModel: string): void {
   const newMessage = stringValue(payload.message) || stringValue(payload.text);
+  // Harness-injected session-start context is recorded as a user message but is
+  // not a real prompt; ignore it before flushing or mutating turn state.
+  if (isHarnessInjectedContext(newMessage)) return;
   if (state.currentUserMessage && state.currentUserMessage === newMessage && isTurnEmpty(state)) {
     if (state.turnStartTs == null) state.turnStartTs = ts;
     return;
@@ -342,7 +345,7 @@ function handleTurnContext(payload: Record<string, unknown>, state: CodexParseSt
 
 function handleUserResponseItem(payload: Record<string, unknown>, state: CodexParseState, ts: number | null, defaultModel: string): void {
   for (const item of extractContentItems(payload.content)) {
-    if (item.type !== 'input_text' || !item.text || item.text.startsWith('<')) continue;
+    if (item.type !== 'input_text' || !item.text || item.text.startsWith('<') || isHarnessInjectedContext(item.text)) continue;
     if (!state.currentUserMessage) {
       flushCodexTurn(state, defaultModel);
       state.currentUserMessage = item.text;