Sync from GitHub 86a5d9c

app.py

@@ -42,9 +42,11 @@ Module layout:
 import base64
 import io
 import json
+import logging
 import os
 import shutil
 import time
+import warnings
 
 import gradio as gr
 from fastapi.responses import FileResponse, HTMLResponse, Response
@@ -61,6 +63,19 @@ from core.constants import (
     VISUAL_SUBDIR,
 )
 
+# gradio 6.17.3 (pinned — see README frontmatter) still uses starlette's old
+# 422 constant, so every queue join emits a StarletteDeprecationWarning. Not
+# ours to fix; silence it so the Space logs stay readable.
+warnings.filterwarnings(
+    "ignore", message=r"'HTTP_422_UNPROCESSABLE_ENTITY' is deprecated"
+)
+
+logging.basicConfig(
+    level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s | %(message)s"
+)
+logging.getLogger("httpx").setLevel(logging.WARNING)  # logs every hub request
+log = logging.getLogger("repairguy")
+
 
 def _build_libraries() -> dict:
     """The {approach: (store, pipeline)} map, constructed once at startup.
@@ -105,8 +120,9 @@ def sync_library() -> None:
         snapshot_download(
             LIBRARY_DATASET_ID, repo_type="dataset", local_dir=PREINDEXED_DIR
         )
+        log.info("library synced from %s", LIBRARY_DATASET_ID)
     except Exception as e:
-        print(f"Library dataset not synced ({LIBRARY_DATASET_ID}): {e}")
+        log.warning("library dataset not synced (%s): %s", LIBRARY_DATASET_ID, e)
         return
     # PREINDEXED_DIR mirrors the dataset (one dir per method); prune top-level
     # leftovers from the pre-method-prefix layout, which snapshot_download
@@ -243,10 +259,13 @@ def api_ask(
         }
         return
     start = time.monotonic()
+    hist = _clean_history(history)
+    log.info(
+        "ask: manual=%s approach=%s k=%s history=%d q=%r",
+        manual, approach, k, len(hist), question[:200],
+    )
     try:
-        events = pipeline.run_agent(
-            store, question, _clean_history(history), [manual], int(k)
-        )
+        events = pipeline.run_agent(store, question, hist, [manual], int(k))
         for ev in events:
             if ev.get("type") == "tool_result" and "gallery" in ev:
                 pages = [p for _, p in ev["page_refs"]]
@@ -260,16 +279,21 @@ def api_ask(
                     ],
                 }
             elif ev.get("type") == "answer":
-                yield {
-                    **ev,
-                    "elapsed": round(time.monotonic() - start, 1),
-                    "approach": approach,
-                    "k": int(k),
-                }
+                elapsed = round(time.monotonic() - start, 1)
+                log.info(
+                    "ask: answered in %.1fs (%d chars, history=%d, summarized=%s)",
+                    elapsed, len(ev.get("answer") or ""),
+                    len(ev.get("history") or []), ev.get("summarized"),
+                )
+                yield {**ev, "elapsed": elapsed, "approach": approach, "k": int(k)}
             else:
                 yield ev
     except ValueError as e:
+        log.warning("ask: rejected — %s", e)
         yield {"type": "error", "error": f"⚠️ {e}"}
+    except Exception as e:
+        log.exception("ask: failed after %.1fs", time.monotonic() - start)
+        yield {"type": "error", "error": f"⚠️ Something went wrong: {e}"}
 
 
 # --- custom FastAPI routes: serve the SPA and the source PDFs ---------------

frontend/index.html

@@ -172,8 +172,11 @@
     <!-- RIGHT: PDF + cited pages -->
     <section class="min-h-0 flex flex-col bg-white rounded-2xl shadow-panel ring-1 ring-navy/10 overflow-hidden">
       <div class="flex-1 min-h-0 relative bg-brand-50/40">
-        <template x-if="pdfDocId">
-          <iframe :src="pdfUrl" class="absolute inset-0 w-full h-full"></iframe>
+        <!-- keyed on the full URL: PDF viewers only honor the #page fragment at
+             load time and ignore hash-only src changes, so the iframe must be
+             recreated to jump pages (same-origin reload, served from cache) -->
+        <template x-for="u in (pdfDocId ? [pdfUrl] : [])" :key="u">
+          <iframe :src="u" class="absolute inset-0 w-full h-full"></iframe>
         </template>
         <div x-show="!pdfDocId" class="chat-grid absolute inset-0 grid place-items-center text-center px-8">
           <div>

pipelines/agent_ask.py

@@ -31,6 +31,8 @@ cannot live in history under chat()'s 16384-token input cap).
 
 from __future__ import annotations
 
+import logging
+
 import spaces
 
 from core.constants import (
@@ -42,6 +44,8 @@ from core.constants import (
 from core.pdf import render_page
 from models import minicpm
 
+log = logging.getLogger("repairguy.agent")
+
 GIVE_UP_ANSWER = (
     "I couldn't put together a grounded answer within my tool budget — try "
     "rephrasing the question or pointing me at a section of the manual."
@@ -77,8 +81,12 @@ def agent_events(
     summarized = False
     if (
         len(history) > HISTORY_KEEP_MESSAGES
-        and minicpm.history_tokens(history) > HISTORY_TOKEN_BUDGET
+        and (tokens := minicpm.history_tokens(history)) > HISTORY_TOKEN_BUDGET
     ):
+        log.info(
+            "history %d msgs / %d tokens over budget %d — summarizing",
+            len(history), tokens, HISTORY_TOKEN_BUDGET,
+        )
         yield {
             "type": "status",
             "kind": "summarizing",
@@ -86,16 +94,26 @@ def agent_events(
         }
         history = minicpm.summarize_history(history)
         summarized = True
+        log.info(
+            "summarized to %d msgs / %d tokens",
+            len(history), minicpm.history_tokens(history),
+        )
 
     msgs = [{"role": m["role"], "content": [m["content"]]} for m in history]
     msgs.append({"role": "user", "content": [question]})
 
     trace: list[str] = []  # text record of tool calls for the durable history
     answer = None
-    for _ in range(AGENT_MAX_STEPS):
+    for step in range(1, AGENT_MAX_STEPS + 1):
         yield {"type": "status", "text": "Thinking…"}
         reply = minicpm.generate_step(msgs, manual)
         call = minicpm.parse_tool_call(reply)
+        log.info(
+            "step %d/%d: %s | reply=%r",
+            step, AGENT_MAX_STEPS,
+            f"tool_call {call['tool']} {call['args']}" if call else "final answer",
+            reply[:200],
+        )
         if call is None:
             answer = reply
             break
@@ -105,6 +123,10 @@ def agent_events(
             query = str(call["args"].get("query") or "").strip() or question
             yield {"type": "tool_call", "tool": "search_docs", "args": {"query": query}}
             refs = search(query, store, doc_ids, top_k)
+            log.info(
+                "search_docs(%r) → %s",
+                query, [(d, p, round(s, 3)) for d, p, s in refs],
+            )
             pages = [
                 (f"{names[doc_id]} — p.{page}", render_page(store.pdf_path(doc_id), page))
                 for doc_id, page, _ in refs
@@ -157,6 +179,10 @@ def agent_events(
             trace.append(f"[displayed page {page} in the viewer]")
 
     if answer is None:  # tool budget exhausted: force a plain answer
+        log.warning(
+            "tool budget exhausted after %d steps — forcing a plain answer",
+            AGENT_MAX_STEPS,
+        )
         yield {"type": "status", "text": "Wrapping up…"}
         msgs.append(
             {
@@ -169,6 +195,7 @@ def agent_events(
         )
         answer = minicpm.generate_step(msgs, manual)
         if minicpm.parse_tool_call(answer) is not None:
+            log.warning("forced answer was still a tool call — using give-up text")
             answer = GIVE_UP_ANSWER
 
     durable = ("\n".join(trace) + "\n\n" if trace else "") + answer