[eric] workflows: a test that runs long hands back what it has instead of telling the agent to poll; memory stops re-deriving facts it already stored

backend/apps/agents/schedule_mcp_server.py

@@ -546,7 +546,10 @@ def handle_delete_step(args: dict) -> dict:
 
 # How long a synchronous test may hold the turn. Long enough for a real multi-step workflow, short
 # enough that a wedged test returns an honest "still running" instead of hanging the conversation.
-TEST_WAIT_S = 240
+# A real workflow test does several searches and page fetches per step. The first budget was 240s
+# and a live three-step digest took 243, so it missed by three seconds and the agent had to make a
+# second hop anyway, which is the exact re-pinging this tool exists to remove.
+TEST_WAIT_S = 600
 TEST_POLL_S = 3
 
 
@@ -574,10 +577,21 @@ def handle_test_workflow(args: dict) -> dict:
             continue
         transcript = t.get("transcript") or "(empty transcript)"
         return _ok(f"Test finished (status: {last_status}). Transcript:\n\n{transcript}")
-    return _ok(
-        f"Test Agent (session {sid[:8]}) is still running after {TEST_WAIT_S}s, so it is a long one. "
-        f"Last status: {last_status}. Call ReadTestTranscript to pick up the result."
+    # Hand back whatever the run has actually produced. Returning only "call ReadTestTranscript" made
+    # the model guess when to poll, which is the same dead end as not waiting at all; a partial
+    # transcript is something it can reason about right now.
+    partial = ""
+    try:
+        t = _call("GET", f"/{wid}/test-transcript")
+        if "_error" not in t:
+            partial = (t.get("transcript") or "").strip()
+    except Exception:
+        partial = ""
+    head = (
+        f"Test Agent (session {sid[:8]}) has not finished after {TEST_WAIT_S}s (status: {last_status}). "
+        "Everything it has produced so far follows; call ReadTestTranscript for the rest once it lands."
     )
+    return _ok(f"{head}\n\n{partial}" if partial else head)
 
 
 def handle_read_test_transcript(args: dict) -> dict:

backend/apps/memory/distill.py

@@ -11,6 +11,7 @@ from backend.apps.agents.core.aux_llm import aux_max_tokens_for
 from backend.apps.agents.core.models import AgentSession
 from backend.apps.agents.manager.predict_followups import conversation_tail
 from backend.apps.agents.manager.session.history_compaction import get_branch_messages
+from backend.apps.memory import store
 from backend.apps.memory.store import add_fact
 
 logger = logging.getLogger(__name__)
@@ -65,7 +66,17 @@ async def distill_session_memory(session: AgentSession) -> List[str]:
             f"Return at most {MAX_FACTS_PER_DISTILL} facts, one per line, no numbering, no quotes. "
             "If the conversation reveals nothing durable, return the single word NOTHING."
         )
-        user_turn = "Conversation:\n<transcript>\n" + tail + "\n</transcript>\n\nExtract the facts."
+        # Show it what is already known. Without this the model re-derives facts it recorded weeks
+        # ago, phrased differently every time, and the store's token-overlap guard cannot catch a
+        # paraphrase: the four duplicate pairs found in a real memory scored 0.11 to 0.33 against a
+        # 0.60 threshold. The generator is semantic, so the dedupe has to be too.
+        known = "\n".join(f"- {f.text}" for f in store.list_facts())
+        already = (
+            "Facts already stored (do NOT repeat these, in any wording):\n" + known + "\n\n"
+            "Return a fact ONLY if it is genuinely new, or is strictly more specific than one above "
+            "(in which case return the sharper version and it will replace the old one).\n\n"
+        ) if known else ""
+        user_turn = already + "Conversation:\n<transcript>\n" + tail + "\n</transcript>\n\nExtract the facts."
         chunks: List[str] = []
         async with client.messages.stream(
             model=aux_model,