fix: give the model actions for the calendar and the inbox

[luca-71]

Asked "what's my schedule today?", JARVIS took a screenshot and described the monitor.

The cause

Visible in the raw completions, once I stopped guessing and printed them:

[ACTION:SCREEN]

Good evening, sir — let me pull up your calendar and see what the day holds.

The intent was right and the vocabulary was wrong. The calendar and mail lookups already existed and worked — but only the hardcoded English phrase table in detect_fast_action could reach them. They were never offered to the model as action tags. [ACTION:SCREEN] was the only inspection action it had, so it substituted, and a photograph of the monitor came back instead of a schedule.

The fix

Add [ACTION:CALENDAR] and [ACTION:MAIL], wired to the same _lookup_and_report machinery the phrase table already used, and narrow SCREEN's description to the display itself — it photographs a monitor, so it can say nothing about an inbox.

This is not an i18n fix

That was my first assumption and it was wrong. English was equally broken; it was merely masked, because the phrase table intercepts those exact wordings before the model is ever consulted. Anything unlisted fell straight through — "When is my next appointment?" matches no entry, went to SCREEN, and now routes correctly.

Measuring the model directly (calling generate_response and reading the tag, so a misroute costs a token round-trip rather than a screenshot) over 12 phrasings × 3 runs per language:

	before	after
English	6/21 wrong	0/36
Italian	9/21 wrong	0/36

The phrasings deliberately vary — "Do I have meetings tomorrow?", "When is my next appointment?", "Check my inbox." — rather than repeating the two sentences the fix was developed against, plus conversational controls that must produce no tag at all.

Verified end to end too: "Qual è il mio prossimo appuntamento?" → [ACTION:CALENDAR] → real Apple Calendar read → "Your schedule is clear today, sir."

One more thing

The two actions get their own acknowledgement line. The model tends to write prose after the tag, which extract_action parses as the target and strips, so without this the spoken response degraded to a bare "Right away, sir."

Scope

The phrase table is left alone. It is a latency optimisation — it answers without an LLM round-trip — and it is still correct for the wordings it lists. This change fixes what happens to everything else.

Independent of #21 and #22; cut from main.

Testing

pytest tests/ gives 35 passed, 8 failed — identical to main, so no regressions. The 8 pre-existing failures are unrelated: 7 need playwright install, and test_browse_action_keywords imports ACTION_KEYWORDS, a symbol that no longer exists in server.py.

🤖 Generated with Claude Code