Merge pull request #73 from ClaydeCode/fix/ntfy-ascii-title

fix(pebble): ASCII-coerce ntfy title; disambiguate STT against KB folders

Author: max-tet

src/clayde/webhook/notify.py

@@ -16,11 +16,36 @@
 log = logging.getLogger("clayde.webhook.notify")
 
 
+# ntfy header values are sent through httpx, which encodes headers as
+# latin-1. Anything outside that range raises UnicodeEncodeError before
+# the request goes out, so the user never sees the notification. We
+# normalise common typographic Unicode to ASCII and replace anything
+# left over with '?'.
+_UNICODE_TO_ASCII = str.maketrans({
+    "—": "-",   # em dash
+    "–": "-",   # en dash
+    "−": "-",   # minus sign
+    "‘": "'",   # left single quote
+    "’": "'",   # right single quote / apostrophe
+    "“": '"',   # left double quote
+    "”": '"',   # right double quote
+    "…": "...", # ellipsis
+    " ": " ",   # non-breaking space
+})
+
+
+def _to_ascii(text: str) -> str:
+    """Coerce arbitrary text to safe ASCII for use in HTTP headers."""
+    return text.translate(_UNICODE_TO_ASCII).encode("ascii", "replace").decode("ascii")
+
+
 class NotificationPayload(BaseModel):
     """Outcome of a Pebble run, as emitted by Claude in the JSON tail.
 
     Title is clamped to 40 chars and body to 300 chars at construction
     time so accidental over-long values never propagate to ntfy headers.
+    Title is additionally coerced to ASCII because it travels as an HTTP
+    header and httpx rejects non-latin-1 header values.
     """
 
     title: str
@@ -30,7 +55,9 @@ class NotificationPayload(BaseModel):
     @field_validator("title", mode="before")
     @classmethod
     def _clamp_title(cls, v):
-        return v[:40] if isinstance(v, str) else v
+        if not isinstance(v, str):
+            return v
+        return _to_ascii(v)[:40]
 
     @field_validator("body", mode="before")
     @classmethod

src/clayde/webhook/skills.py

@@ -53,6 +53,18 @@ def _parse_skill(path: Path) -> Skill:
 "log", or "capture", write a file there. No git operations — Syncthing
 handles sync.
 
+Disambiguate against the KB structure. Before acting on a phrase that
+seems nonsensical or oddly worded, list the top level of the knowledge
+base (e.g. `ls /home/clayde/knowledge_base`). Its top-level directories
+are stable nouns the user actually uses ("people", "specs", "inbox",
+"freeshard", ...). If a confusing token has a phonetic neighbour that
+matches one of those folders or a common verb pair ("add a", "note
+that", "capture"), prefer that reading. Worked example: "after people
+and tree for my brother-in-law" → "add a people entry for my
+brother-in-law", because "after" ≈ "add a" and "tree" ≈ "entry", and
+`people/` is a real folder. State the interpretation you picked in your
+narrative so the user can spot a wrong guess in the ntfy summary.
+
 {skill_section}
 
 Skills are suggestions, not constraints. Use as many as the command needs,

tests/test_webhook_notify.py

@@ -22,6 +22,36 @@ def test_notification_payload_accepts_short():
     assert p.success is True
 
 
+def test_notification_payload_em_dash_in_title_normalised():
+    # Real prod failure: em dash in title raised UnicodeEncodeError when
+    # httpx serialised the header as latin-1.
+    p = NotificationPayload(title="Thomas Stegger — plant prefs saved", body="ok", success=True)
+    assert "—" not in p.title
+    assert p.title == "Thomas Stegger - plant prefs saved"
+    # Must round-trip cleanly through latin-1 (the header codec httpx uses).
+    p.title.encode("latin-1")
+
+
+def test_notification_payload_smart_quotes_in_title_normalised():
+    p = NotificationPayload(title="“hi” ‘there’", body="ok", success=True)
+    assert p.title == '"hi" \'there\''
+
+
+def test_notification_payload_unknown_unicode_in_title_replaced():
+    p = NotificationPayload(title="emoji \U0001f600 tail", body="ok", success=True)
+    assert "\U0001f600" not in p.title
+    p.title.encode("ascii")
+
+
+def test_notification_payload_ascii_coercion_runs_before_clamp():
+    # "..." (3 chars) replaces "…" (1 char); clamp comes after, so a
+    # title that fit pre-replacement may not fit after — and that's fine.
+    long = "a" * 38 + "…"  # 39 chars in, 41 chars after replacement
+    p = NotificationPayload(title=long, body="ok", success=True)
+    assert len(p.title) == 40
+    p.title.encode("ascii")
+
+
 @pytest.mark.asyncio
 @respx.mock
 async def test_send_ntfy_success_headers():

tests/test_webhook_skills.py

@@ -160,6 +160,16 @@ def test_prompt_when_no_skills_still_invites_judgement():
     assert "judgement" in p.lower() or "judgment" in p.lower()
 
 
+def test_prompt_mentions_kb_structure_disambiguation():
+    from clayde.webhook.skills import build_system_prompt
+    p = build_system_prompt([])
+    # Tells Claude to inspect KB layout and prefer phonetic neighbours
+    # that match real folders ("after people and tree" → "add a people entry").
+    assert "ls /home/clayde/knowledge_base" in p
+    assert "phonetic" in p.lower()
+    assert "people" in p
+
+
 def test_discovers_builtin_alongside_host(tmp_path):
     from clayde.webhook.skills import discover_skills
     # Simulate the in-container layout: /skills/builtin + /skills/personal.