fix(observer_recall): answer conversationally, not as a machine dump (beat 20) (#263)

Recall replied like a log printer:
  "Here's the last 20 min (from CODEC's observer):
   You were in: Claude.
   Timeline:
     12:20  Claude
   Files touched: best-noise-cancelling-headphones-2026.md."

Accurate but robotic — CODEC should answer like someone who was watching over
your shoulder. Now:
  "Over the last 9 minutes, you were in Claude the whole time.
   You touched 4 files: …"

Prose adapts to what actually happened (one app / two apps / moved around a fair
bit), mentions the last thing on screen, and pluralises files properly. Kept
deterministic — no second LLM call — so it stays instant and, more importantly,
can't invent anything (the whole point of the earlier anti-fabrication work).
The honest out-of-range answer for clock-time asks is unchanged.

Manifest regenerated. 52 observer tests pass, ruff clean.

Co-authored-by: Mickael Farina <farina.mickael@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>

Author: 3 people

skills/.manifest.json

@@ -62,7 +62,7 @@
     "network_info.py": "bd776b619cf7c18d67fe03cb0f0456cf9c4f9bf71475740a233a9ca1e6672fcd",
     "notes.py": "7d50d1544ea955f59917a1f0e7902d115e9dbccd3188b641057a55b6a5b2803a",
     "notification_reader.py": "681208ae4253dfe549512cce4c722c76ca85f2bbbdb8737e37c026ec9444c972",
-    "observer_recall.py": "f7084d11c069603d815e648dd765628bfed6d31c1df4ac59f0dac35259aa6fe2",
+    "observer_recall.py": "5234ab88d151003d8acb00b0e72139d668cfbcd5f091cdda52f247682698a9ff",
     "password_generator.py": "f11a917299e14cbd2560111da0bb748cd08792cf715cc4098c64eb62da8c54e3",
     "philips_hue.py": "fa831712c39dc6327c84199d8f0aeb09a169a264ce3d936cc337a5c5d6632f7e",
     "pilot.py": "f9967890f138bc7a48ae4abc8b4170dca13961b81659ef4e530dc619c1cf90d4",

skills/observer_recall.py

@@ -105,13 +105,30 @@ def _parse_ts(entry: dict) -> datetime | None:
         return None
 
 
+def _fmt_span(entries: list) -> str:
+    """Natural opener for how long the snapshots span, e.g. "Over the last
+    25 minutes" / "In the last couple of minutes". "" if it can't be derived."""
+    if len(entries) < 2:
+        return ""
+    first, last = _parse_ts(entries[0]), _parse_ts(entries[-1])
+    if not first or not last:
+        return ""
+    mins = int((last - first).total_seconds() // 60)
+    if mins <= 1:
+        return "In the last minute or so"
+    if mins < 5:
+        return "Over the last couple of minutes"
+    return f"Over the last {mins} minutes"
+
+
 def _summarise(entries: list) -> str:
     """A compact, human timeline of the entries (oldest→newest)."""
     if not entries:
         return "nothing recorded in that window."
 
     lines: list[str] = []
     last_app = None
+    last_title_seen = ""
     apps_seen: list[str] = []
     files_seen: set[str] = set()
     ocr_bits: list[str] = []
@@ -122,6 +139,8 @@ def _summarise(entries: list) -> str:
         win = e.get("active_window") or {}
         app = win.get("app")
         title = (win.get("title") or "").strip()
+        if title:
+            last_title_seen = title
         if app and app != last_app:
             label = app + (f" — {title[:60]}" if title else "")
             lines.append(f"  {stamp}  {label}")
@@ -136,15 +155,35 @@ def _summarise(entries: list) -> str:
         if ocr and len(ocr) > 20:
             ocr_bits.append(ocr[:120])
 
+    # Conversational prose, not a machine dump. The old format printed
+    # "You were in: X. / Timeline: / 12:20 X / Files touched: y" — accurate but
+    # robotic. CODEC should answer like a person who was watching over your
+    # shoulder. Deterministic templating (no second LLM call) keeps it instant
+    # and, crucially, keeps it from inventing anything.
     out = []
     if apps_seen:
-        out.append("You were in: " + ", ".join(apps_seen[:6]) + ".")
-    if lines:
-        out.append("Timeline:\n" + "\n".join(lines[:12]))
+        span = _fmt_span(entries)
+        if len(apps_seen) == 1:
+            body = f"you were in {apps_seen[0]} the whole time"
+        elif len(apps_seen) == 2:
+            body = f"you went back and forth between {apps_seen[0]} and {apps_seen[1]}"
+        else:
+            shown = apps_seen[:4]
+            body = ("you moved around a fair bit — "
+                    + ", ".join(shown[:-1]) + f" and {shown[-1]}")
+        opener = f"{span}, {body}" if span else body[0].upper() + body[1:]
+        if last_title_seen:
+            opener += f". Last thing on screen was \"{last_title_seen[:70]}\""
+        out.append(opener + ".")
     if files_seen:
-        out.append("Files touched: " + ", ".join(sorted(files_seen)[:8]) + ".")
+        fl = sorted(files_seen)
+        if len(fl) == 1:
+            out.append(f"You touched one file: {fl[0]}.")
+        else:
+            out.append(f"You touched {len(fl)} files: " + ", ".join(fl[:6])
+                       + ("…" if len(fl) > 6 else "") + ".")
     if ocr_bits:
-        out.append('On screen (excerpt): "' + ocr_bits[-1] + '"')
+        out.append(f"Text on screen included: \"{ocr_bits[-1].strip()}\"")
     if not out:
         # Entries exist but carry no window/title/OCR/file content — the observer
         # is polling but capturing nothing. This is almost always a macOS
@@ -204,4 +243,6 @@ def run(task: str, context: str = "") -> str:
         return (f"Nothing in {span} — the observer keeps roughly the last 10 "
                 f"minutes, so I can't see that far back.{depth}")
 
-    return f"Here's {span} (from CODEC's observer):\n{_summarise(kept)}"
+    # _summarise already opens conversationally ("Over the last 25 minutes, you
+    # were in Claude…"), so no robotic "Here's X (from CODEC's observer):" header.
+    return _summarise(kept)