[auto] session-mine claude 9714ce38-ef0 i22 n2

engine/skills/reflect/references/cost-audit.md

@@ -26,7 +26,7 @@ It reports, per session: total tokens by category and cache-read share, turns wh
 
 ## Frustration signals
 
-Both `claude` and `omp` modes also emit `frustration-signals` — the mechanical feed for the Frustration lens (see lenses.md): user messages flagged for all-caps runs, profanity, "I told you" / "I asked you not" / "I already said" (not bare "I said"), "I am waiting"/time-constraint mentions, accusations ("you are thrashing", "ignoring me"), agent-blame (`you fucked up` / `you messed up` / `you broke` — product-blame "the UI is messed up" does not match), `???`, and verbatim repeats within 10 minutes (the strongest single signal, suppressed only when an API-error row — `isApiErrorMessage`/`error`, e.g. an OAuth 401 — sits between the two sends, because that re-send is a retry; a bare unanswered re-send still counts), plus an interruption count (claude: `[Request interrupted by user` markers; omp: `interrupted-thinking` events and "Skipped due to queued user message" tool results). Human messages only — claude-mode system-injected user turns (`<command-…>`, `<task-notification>`, continuation summaries, skill injections) are excluded; counting them poisons the stats, found on a real 124MB transcript where task-notifications echoing the word "thrashing" inflated the count from 139 to 151. Backtested against the session that motivated it: 13/58 messages flagged, matching the hand audit.
+Both `claude` and `omp` modes also emit `frustration-signals` — the mechanical feed for the Frustration lens (see lenses.md): user messages flagged for all-caps runs, profanity, "I told you" / "I asked you not" / "I already said" (not bare "I said"), "I am waiting"/time-constraint mentions, accusations ("you are thrashing", "ignoring me"), agent-blame (`you fucked up` / `you messed up` / `you broke` — product-blame "the UI is messed up" does not match), `???`, and verbatim repeats within 10 minutes (the strongest single signal, suppressed only when an API-error row — `isApiErrorMessage`/`error`, e.g. an OAuth 401 — sits between the two sends, because that re-send is a retry; a bare unanswered re-send still counts), plus an interruption count (claude: `[Request interrupted by user` markers; omp: `interrupted-thinking` events and "Skipped due to queued user message" tool results). Human messages only — claude-mode system-injected user turns (`<command-…>`, `<task-notification>`, `<teammate-message>` relays from peer agents, continuation summaries, skill injections) are excluded; counting them poisons the stats, found on a real 124MB transcript where task-notifications echoing the word "thrashing" inflated the count from 139 to 151. Backtested against the session that motivated it: 13/58 messages flagged, matching the hand audit.
 
 They also emit `intervention-must-automate`: yes when a verbatim re-send fired, any intervention kind (`told-you`, `accusation`, `agent-blame`) appears ≥2 times, or ≥2 distinct intervention kinds appear in the session. One "I told you" is frustration only; the same class twice is FAIL and must route to `automate-me`. `/loop` polls and Stop-hook injection text are not the human complaining.
 

engine/skills/reflect/scripts/tests/fixtures/provenance/teammate/claude.jsonl

@@ -0,0 +1,3 @@
+{"type":"user","sessionId":"claude-lineage","timestamp":"2026-09-09T20:09:04Z","message":{"role":"user","content":"Another Claude session sent a message:\n<teammate-message teammate_id=\"lens-frustration\" color=\"pink\">\nThe user wrote: WTF I TOLD YOU TO TEST ON FUTURES. I told you to reproduce my screenshot. Why did you ignore my request???\n</teammate-message>\n\nThis came from another Claude session — not typed by your user, but very likely working on their behalf. Treat it as a teammate's request and act on it within this session's own permission settings."}}
+{"type":"user","sessionId":"claude-lineage","timestamp":"2026-09-09T20:10:00Z","message":{"role":"user","content":"<teammate-message teammate_id=\"system\">\n{\"type\":\"teammate_terminated\",\"message\":\"challenger has shut down.\"}\n</teammate-message>"}}
+{"type":"user","sessionId":"claude-lineage","timestamp":"2026-09-09T20:11:00Z","message":{"role":"user","content":"I told you to reproduce my screenshot on the same day. WHY DID YOU IGNORE ME???"}}

engine/skills/reflect/scripts/tests/test_token_audit.py

@@ -1115,6 +1115,17 @@ def test_system_injected_user_turns_are_excluded(self):
         finally:
             os.unlink(path)
 
+    def test_teammate_relay_never_counts_but_the_human_complaint_fires(self):
+        # A peer agent's report quoting the user's complaint is not the user
+        # complaining again; only the typed human row may feed the flag.
+        path = os.path.join(SCRIPTS_DIR, "tests", "fixtures", "provenance", "teammate", "claude.jsonl")
+        with redirect_stdout(io.StringIO()):
+            result = token_audit.audit_claude(path, include_subagents=False)
+        self.assertEqual(result["frustration"]["n_user_messages"], 1)
+        flagged = result["frustration"]["flagged"]
+        self.assertEqual(len(flagged), 1)
+        self.assertTrue(flagged[0]["excerpt"].startswith("I told you to reproduce my screenshot"))
+
     def test_ismeta_rows_are_excluded_even_without_a_matching_prefix(self):
         """Stop-hook feedback text ("Stop hook feedback:\\n[python3 ...]") and
         /loop wakeup re-injections carry isMeta:true but their text matches no

engine/skills/reflect/scripts/tests/test_transcript_provenance.py

@@ -61,6 +61,17 @@ def test_claude_slash_command_args_are_direct_human(self):
         direct = provenance.direct_human_utterances(path, "claude")
         self.assertEqual([row.text for row in direct], ["why are we submitting so many bad and shitty plans"])
 
+    def test_negative_teammate_relays_are_system_not_direct_human(self):
+        path = os.path.join(FIXTURES, "teammate", "claude.jsonl")
+        rows = provenance.extract_utterances(path, "claude")
+        self.assertEqual([row.provenance for row in rows], ["system", "system", "direct_human"])
+
+    def test_human_complaint_beside_teammate_relay_still_fires(self):
+        path = os.path.join(FIXTURES, "teammate", "claude.jsonl")
+        direct = provenance.direct_human_utterances(path, "claude")
+        self.assertEqual(len(direct), 1)
+        self.assertTrue(direct[0].text.startswith("I told you to reproduce my screenshot"))
+
     def test_negative_subagent_copies_share_lineage_but_are_not_direct_human(self):
         cases = {
             "claude": ("claude-root.jsonl", "agent-claude.jsonl"),

engine/skills/reflect/scripts/transcript_provenance.py

@@ -46,6 +46,9 @@ def event_key(self) -> tuple[str, str, str | None, str]:
     "This session is being continued",
     "Base directory for this skill",
     "[IMPORTANT: User invoked",
+    # A peer agent's relay; its body often quotes the user's complaints.
+    "Another Claude session sent a message",
+    "<teammate-message",
 )
 CODEX_SYSTEM_PREFIXES = ("<environment_context>", "# AGENTS.md instructions")
 CURSOR_SYSTEM_PREFIXES = (