fix(chat): let allowlisted skill triggers beat the conversational filter (beat 20) (#256)

_try_skill ran _is_conversational() FIRST and bailed to the LLM for any
question-phrased message — so "what was I doing 1h ago?" never reached
observer_recall (allowlisted since #250); the LLM answered from memory and
fabricated. Now: if the message word-boundary-matches an allowlisted skill
trigger, honor it (explicit intent) even if it looks conversational; only fall
through to the LLM when no allowlisted skill matches. Verified: "what was I
doing 1h ago?" now routes to observer_recall.

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

Author: 3 people

routes/chat.py

@@ -442,13 +442,21 @@ def _enrich_messages(messages: list, config: dict, force_search: bool = False) -
 
 def _try_skill(user_text: str):
     """Check if user_text matches a skill. Returns (skill_name, result) or (None, None).
-    Skips skill matching for conversational messages to prevent false triggers."""
-    if _is_conversational(user_text):
-        return None, None
+
+    An explicit allowlisted-trigger match (e.g. "what was I doing 1h ago?" →
+    observer_recall) is honored even for question-phrased ("conversational")
+    messages: a word-boundary trigger match is a stronger intent signal than the
+    chatty-phrasing heuristic. Only when NO allowlisted skill matches do
+    conversational messages fall through to the LLM (the original behaviour).
+    Before this, questions like "what was I doing?" were skipped entirely and the
+    LLM answered from memory (and fabricated)."""
     try:
         from codec_dispatch import check_skill, run_skill
         skill = check_skill(user_text)
-        if skill and skill.get("name") in CHAT_SKILL_ALLOWLIST:
+        matched = bool(skill and skill.get("name") in CHAT_SKILL_ALLOWLIST)
+        if not matched and _is_conversational(user_text):
+            return None, None
+        if matched:
             # re-audit A2: destructive skills need explicit consent (reuses the
             # AskUserQuestion PWA panel; blocks this worker thread until answered).
             import codec_consent