add the greeting msg

go-api/internal/runtime/smalltalk.go

@@ -0,0 +1,126 @@
+package runtime
+
+import "strings"
+
+// isSmalltalk reports a message that cannot be answered any better by looking
+// something up.
+//
+// "Hi" used to cost a full operational turn. Retrieval is gated on
+// configuration and never on the question (see retrieve), so a greeting arrived
+// at the model wrapped in eight policy chunks, with the whole tool catalogue
+// attached and the evidence placed BEFORE the question. A model handed that
+// reasonably concludes it was asked for an operational briefing, and answers
+// with one — screening backlog, uncovered shifts, citations and all. Measured
+// against production: 6,174 tokens over two model calls, for the word "hi".
+//
+// The cost is the smaller half. The real damage is that the product appears not
+// to understand being greeted, which is the first thing anybody tries.
+//
+// This is deliberately NOT a per-agent rule, and not an `if agent_key == ...`
+// — §13 lists that as the anti-pattern it is. It is a property of the MESSAGE,
+// applied identically to every spec, so adding an agent still requires no
+// runtime change (I6).
+//
+// Conservative by construction: the normalised message must match a phrase in
+// the set EXACTLY. Nothing substring-matches, so "hi, which shifts are
+// uncovered?" is an operational question and keeps its tools and its evidence.
+// A false negative costs a few thousand tokens; a false positive answers a real
+// question with a greeting, so the set only holds phrases that carry no request
+// at all.
+func isSmalltalk(q string) bool {
+	n := normaliseSmalltalk(q)
+	if n == "" {
+		return false
+	}
+	_, ok := smalltalkPhrases[n]
+	return ok
+}
+
+// normaliseSmalltalk reduces a message to lowercase letters and single spaces.
+//
+// Punctuation and emoji are dropped rather than enumerated, so "Hi!", "hi :)"
+// and "HI 👋" all arrive as "hi" without the set needing a row for each. Digits
+// are NOT letters and so are dropped too, which is harmless here: no phrase in
+// the set contains one, and a message that does — "shift 12?" — fails the exact
+// match either way.
+func normaliseSmalltalk(q string) string {
+	var b strings.Builder
+	b.Grow(len(q))
+	space := false
+	for _, r := range strings.ToLower(strings.TrimSpace(q)) {
+		switch {
+		case r >= 'a' && r <= 'z':
+			if space && b.Len() > 0 {
+				b.WriteByte(' ')
+			}
+			space = false
+			b.WriteRune(r)
+		case r == '\'' || r == '’':
+			// Dropped outright rather than treated as a separator, so "how's"
+			// stays one word. Both the ASCII quote and the curly one a phone
+			// keyboard substitutes — the same character to whoever typed it,
+			// and not to the first version of this function.
+		default:
+			// Any other run of non-letters is one separator, so "thank-you"
+			// and "thank  you" normalise alike.
+			space = true
+		}
+	}
+	return b.String()
+}
+
+// smalltalkPhrases is the whole rule, as data.
+//
+// Greetings, thanks and farewells only. Each is a complete message that asks
+// for nothing, which is what makes skipping retrieval and tools safe rather
+// than merely cheap. Acknowledgements like "ok" and "cool" are deliberately
+// absent: they are plausible smalltalk but also plausible answers to a
+// question the agent just asked, and the cost of being wrong is higher than
+// the tokens being saved.
+var smalltalkPhrases = map[string]struct{}{
+	"hi": {}, "hii": {}, "hiya": {}, "hello": {}, "helo": {}, "hey": {},
+	"yo": {}, "howdy": {}, "greetings": {}, "hi there": {},
+	"hello there": {}, "hey there": {}, "hi owliver": {},
+	"hello owliver": {}, "hey owliver": {},
+
+	"good morning": {}, "good afternoon": {}, "good evening": {},
+	"good day": {}, "morning": {}, "afternoon": {}, "evening": {},
+	"gm": {}, "ge": {},
+
+	"how are you": {}, "how are you doing": {}, "hows it going": {},
+	"how is it going": {}, "you there": {}, "are you there": {},
+
+	"thanks": {}, "thank you": {}, "thanks a lot": {},
+	"thank you very much": {}, "thanks very much": {}, "many thanks": {},
+	"ty": {}, "cheers": {}, "thank u": {},
+
+	"bye": {}, "goodbye": {}, "good bye": {}, "see you": {},
+	"see ya": {}, "good night": {}, "goodnight": {}, "later": {},
+}
+
+// smalltalkDirective is appended to the system prompt for a smalltalk turn.
+//
+// Needed because the agent's own instructions describe an operational analyst,
+// and an operational analyst greeted with "hi" and given no tools will still
+// reach for the longest answer it can justify. Removing the evidence removes
+// the citations; it does not by itself shorten the reply.
+//
+// Appended to the SYSTEM prompt rather than wrapped around the user's message:
+// it is a standing instruction from the platform, not something the person
+// said, and putting words in their mouth is how a transcript stops matching
+// what was typed. I7 is untouched — this is the runtime's own text, not
+// retrieved content, and nothing retrieved can reach here because retrieval did
+// not run.
+const smalltalkDirective = "\n\nThe person has greeted you or said something " +
+	"conversational. Reply in one or two short sentences: greet them back and " +
+	"offer to help. Do not summarise data, do not list findings or next steps, " +
+	"and do not cite sources — you have not looked anything up."
+
+// smalltalkMaxOutputTokens caps a greeting's reply.
+//
+// A ceiling the model is not told about truncates mid-sentence rather than
+// winding down, so this sits well above any sane greeting (a sentence or two is
+// well under 100 tokens) and acts only as a backstop for a model that ignores
+// the directive above. The directive does the shortening; this bounds the bill
+// when it does not.
+const smalltalkMaxOutputTokens = 256