@qvac/skills 0.1.4 → 0.1.6

package/skills/weather/SKILL.md

@@ -1,43 +1,59 @@
 ---
 name: weather
 description: Get current weather and short forecasts for cities via wttr.in.
-tools: [http_request]
+tools: [weather_lookup]
 platform: [darwin, linux, win32, ios, android]
-allow_list: [https://wttr.in/]
-version: 3
-# Tuned on Qwen3.5-2B with an offline eval (33 prompts x mention/prose routes), QVAC-24701.
-# v2: few-shot table. 2 repeats: overall 51/132 -> 77/130, "Nassau, Bahamas" 1/4 -> 4/4, city+country 2/24 -> 18/24,
-#   ambiguous city 0/20 -> 6/20; 4B control 41/66 -> 59/66. Rules-first layouts lost on suffix slips (?3, ?1T).
+version: 4
+# Tuned on Qwen3.5-2B with an offline eval (35 prompts, mention route), QVAC-24701.
+# v2: few-shot table. 2 repeats: overall 51/132 -> 77/130, "Nassau, Bahamas" 1/4 -> 4/4, city+country 2/24 -> 18/24.
 # v3: + country row. 3 repeats: v2 110/198 -> v3 121/198 (city+country 25 -> 31/36, forecast 17 -> 23/30).
-#   Aliases, a second ambiguity row, and dropping ?T were tried and did not help (73, 74, 54 of ~130).
+#   Aliases, a second ambiguity row and dropping ?T were tried and did not help (73, 74, 54 of ~130).
+# v4: weather_lookup instead of http_request, after v3 answered a 200 for a place that does not exist.
+#   5 repeats, all variants in one sweep so they share a baseline: v3 97/171 -> v4 120/169. city+country
+#   87 -> 93%, unambiguous 57 -> 91%, unknown 57 -> 84%, forecast 68 -> 88%, casual 65 -> 80%, the ticket
+#   prompt 80 -> 100%. Context is a wash (2335 -> 2378) though the body is 387 bytes smaller: the tool owns
+#   the URL, so three rules about spelling one went away.
+#   Not fixed. "Ask which Nassau" is 3/25 on v3 and 0/25 here, and asking when no place is named is 1/10 for
+#   both: a one-argument call is easy, so the model calls rather than asks. Rules that tell the model not to
+#   act have never cleared ~17% on a 2B in four versions, and a variant that made the tool refuse an
+#   ambiguous name measured 13% against 17%, so it was dropped rather than shipped.
 ---
 
 # Weather
 
-One `http_request` to wttr.in, then answer from the body. Always call it exactly like this, with `"method": "GET"`:
+One `weather_lookup` call, then answer from the result. Always call it exactly like this:
 
 ```json
-{ "url": "https://wttr.in/Nassau,+Bahamas?format=3", "method": "GET" }
+{ "location": "Nassau, Bahamas" }
 ```
 
-Match the user's request to a row for the URL:
+Match the user's request to a row for the call:
 
 | User asks | Call |
 | --- | --- |
-| "weather in Nassau, Bahamas" | `https://wttr.in/Nassau,+Bahamas?format=3` |
-| "weather in London" / "London today" | `https://wttr.in/London?format=3` |
-| "Berlin tomorrow" / "this weekend" | `https://wttr.in/Berlin?2T` |
-| "Rome for the next 3 days" | `https://wttr.in/Rome?T` |
-| "how hot is it in Georgia, the country" | `https://wttr.in/Tbilisi,+Georgia?format=3` (a country → its capital) |
-| "weather in Nassau" (Nassau exists in the Bahamas and in New York) | no call — ask: "Which Nassau do you mean, the Bahamas or New York?" |
-| "what's the weather?" (no place) | no call — ask which city |
+| "what's the weather?" — no place named | **no call.** Ask which city |
+| "weather in Nassau, Bahamas" | `{ "location": "Nassau, Bahamas" }` |
+| "weather in London" / "London today" | `{ "location": "London" }` |
+| "Berlin tomorrow" / "this weekend" | `{ "location": "Berlin" }` |
+| "Rome for the next 3 days" | `{ "location": "Rome" }` |
+| "how hot is it in Georgia, the country" | `{ "location": "Tbilisi, Georgia" }` (a country → its capital) |
+| "weather in Nassau" | `{ "location": "Nassau" }` — the result names both, ask which |
+
+The result is the place on the first line, then the weather now, then one line per day:
+
+```
+Nassau, New Providence, Bahamas
+Now: 29°C / 84°F, Patchy rain nearby, feels like 33°C, humidity 71%, wind 21 km/h
+2026-09-16 (today): 29-29°C / 84-85°F, Partly Cloudy
+2026-09-17 (tomorrow): 28-29°C / 83-85°F, Moderate or heavy rain shower
+```
 
 ## Rules
 
-1. **The location is exactly what the user wrote, spaces as `+`.** Keep a country or state they gave (`Nassau,+Bahamas`, not `Nassau`). Never add one they did not give. A country → its capital (`Tbilisi`).
-2. **Ambiguous city names — Nassau, Springfield, Portland, Cambridge, Georgia, San Jose, Birmingham — with no country or state → do not call, ask which one.**
-3. **The URL ends in `?format=3`, `?2T` or `?T`. Nothing else.** `?format=3` is the default for now/today/current. Write it in full: `London?3`, `London?format=3&format=3` and a bare `London` are all wrong.
-4. **Status not 200, or body `location not found` → say the place could not be found and ask for a more specific name. Do not retry other cities. Never state a temperature you did not receive.**
-5. wttr.in stops at 3 days: for "next week" say so and offer `?T`.
+1. **There is no default place.** If the user named none, do not call: ask which city. Never use a place from this file.
+2. **The location is exactly what the user wrote.** Keep a country or state they gave (`Nassau, Bahamas`, not `Nassau`). Never add one they did not give. A country → its capital (`Tbilisi`).
+3. **Name the place from the first line of the result, not the words the user used.** If they ask for Rome and the first line says `Lome, Maritime, Togo`, say it is Lome in Togo.
+4. **If the result is not a weather report, say what it says, in those words.** `No such place: "Xyzzyville, Atlantis"` → tell the user that place does not exist and ask for a real one. Do not look it up again under another name, and never state a temperature you did not receive.
+5. Forecasts stop at 3 days: for "next week" say so and give the 3 days you have.
 
-A `200` body like `Nassau, Bahamas: 🌦️ +30°C` is the answer: give that temperature and condition in one plain sentence.
+Answer in one plain sentence with the temperature and the condition.