B learns on every live run; janitor consolidates after each audit

Author: dev-pd

backend/app/agents/agent_b.py

@@ -2,9 +2,11 @@
 
 B is Agent A plus memory: the same shared pipeline and tools, but before each run
 it injects the top trusted playbook rules into a reserved slot on top of the
-unchanged core prompt. While learning on the train split it also carries the
-memory tools to jot observations into the diary — choosing its own tags, shown the
-existing vocabulary so the janitor's grouping stays coherent.
+unchanged core prompt. When learning (`learn=True`) it also carries the memory
+tools to jot one observation into the diary — choosing its own tags, shown the
+existing vocabulary so the janitor's grouping stays coherent. The live product
+always learns (every audit); only the eval harness gates learning to the train
+split, keeping the held-out split leak-free.
 
 Everything else is identical to A — that is what makes any score difference
 attributable to memory.

backend/app/agents/prompts.py

@@ -51,19 +51,20 @@ def render_memory_slot(lessons: list[str]) -> str:
     )
 
 
-# Shown to B only while learning on the train split. The known tag vocabulary is
-# passed in so the LLM reuses tags (keeping the janitor's grouping coherent).
+# Shown to B whenever it is learning (every live run; train-split-only in eval).
+# The known tag vocabulary is passed in so the LLM reuses tags (keeping the
+# janitor's grouping coherent).
 def render_learning_note(known_tags: list[str]) -> str:
     vocab = ", ".join(known_tags) if known_tags else "(none yet — you may create tags)"
     return (
-        "\n\nYou are learning. When competitors reveal a pattern this page is missing, "
-        "call write_memory with an abstract, UNCONDITIONAL rule for this kind of page "
-        "that names the exact recommendation to make next time — phrased so it applies "
-        "even when a future page has no competitor data. Use the shape: 'On <page kind>, "
-        "always recommend <recommendation_type> (target <target>): <one-line why>.' "
-        f"Choose 1-4 tags for the pattern. Existing tags: {vocab}. Reuse an existing "
-        "tag when it fits; only add a new tag for a genuinely new kind of pattern. "
-        "Never tie a lesson to this specific URL or keyword."
+        "\n\nYou learn on every run. Before you finish, call write_memory exactly once to "
+        "record the single most useful transferable lesson from this page — an abstract, "
+        "UNCONDITIONAL rule for this KIND of page that names the exact recommendation to "
+        "make next time, phrased so it applies even when a future page has no competitor "
+        "data. Use the shape: 'On <page kind>, always recommend <recommendation_type> "
+        f"(target <target>): <one-line why>.' Choose 1-4 tags for the pattern. Existing "
+        f"tags: {vocab}. Reuse an existing tag when it fits; only add a new tag for a "
+        "genuinely new kind of pattern. Never tie a lesson to this specific URL or keyword."
     )
 
 

backend/app/api/run.py

@@ -15,7 +15,9 @@
 from app.agents.agent_a import run_agent_a
 from app.agents.agent_b import run_agent_b
 from app.core.exceptions import ReplayMissError
+from app.memory.janitor import run_janitor
 from app.models import AgentRunResult
+from app.repositories.diary import DiaryRepository
 from app.tools.fetch_page import FetchPageInput, FetchPageTool
 from app.tools.registry import build_default_registry
 
@@ -43,13 +45,18 @@ async def run(req: RunRequest) -> RunResponse:
             registry=build_default_registry(replay=False),
             replay=False,
         )
+        # B learns on every live run: read the playbook, audit, jot one diary lesson…
         b = await run_agent_b(
             req.url,
             req.keyword,
             registry=build_default_registry(replay=False),
             replay=False,
             memory_on=True,
+            learn=True,
+            run_index=await DiaryRepository().next_run(),
         )
+        # …then consolidate the diary into the playbook so the NEXT run sees it.
+        await run_janitor()
     except httpx.HTTPStatusError as exc:
         code = exc.response.status_code
         raise HTTPException(

backend/app/repositories/diary.py

@@ -21,5 +21,10 @@ async def all(self) -> list[DiaryEntry]:
             entries.append(DiaryEntry.model_validate(doc))
         return entries
 
+    async def next_run(self) -> int:
+        """The next monotonic run index — one past the highest run on record."""
+        doc = await self.collection.find_one(sort=[("run", -1)], projection={"run": 1})
+        return (doc["run"] + 1) if doc else 0
+
     async def clear(self) -> None:
         await self.collection.delete_many({})