adventure: author Pete's dispatches in his voice with the local LLM

emitFact now runs the final fact through authorDispatch, which asks the local
Ollama model for a warm-reporter headline and lede and ships them on the Fact.
Pete guards and publishes them, falling back to its own templates on anything
it rejects — so authoring is best-effort by design: LLM off, a timeout, a
malformed generation, or an over-length pair all return an empty prose pair and
Pete templates the fact. The names allowed in the prose are the fact's Actors,
built from the post-opt-out subject/opponent, so what the model may say and what
Pete's guard permits are the same list. Synchronous like the holdem tip rewrite,
but on a tight 15s budget: news facts are infrequent and a template now beats a
voiced dispatch late. With no route for Pete to call back into this box, the
voice lives in the prompt here rather than in a Pete-owned inference endpoint.

internal/plugin/pete.go

@@ -183,6 +183,12 @@ func emitFact(f peteclient.Fact, subjectUser, opponentUser id.UserID) {
 		}
 	}
 	f.Actors = actors
+	// Author the prose in Pete's voice from the FINAL fact, so the names in the
+	// dispatch match the Actors allow-list Pete guards against. Best-effort: an
+	// empty pair (LLM off or authoring failed) just means Pete templates the
+	// fact. Synchronous, like the holdem tip rewrite — news facts are infrequent
+	// and the call is tightly bounded (dispatchLLMTimeout).
+	f.Headline, f.Lede = authorDispatch(f)
 	peteclient.Emit(f)
 }
 

internal/plugin/pete_dispatch_voice.go

@@ -0,0 +1,201 @@
+package plugin
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log/slog"
+	"net/http"
+	"os"
+	"strings"
+	"time"
+
+	"gogobee/internal/peteclient"
+)
+
+// LLM-authored adventure dispatches. gogobee owns the raw model compute; this is
+// where a structured fact becomes warm-reporter prose for Pete to publish. Pete
+// is still the editor and the safety boundary: it runs its own prose-guard over
+// whatever we send and falls back to its templates on anything it does not like,
+// so authoring here is best-effort by design — every failure path returns an
+// empty pair and Pete templates the fact.
+//
+// The voice must live somewhere, and with no route for Pete to call back into
+// this box (see roster.go in the Pete repo) it lives in the prompt below. Keep
+// it faithful to pete_adventure_news_voice.md; Pete's persona, not gogobee's.
+
+// dispatchLLMTimeout bounds the authoring call. emitFact runs on game-event
+// chokepoints (a party wipe fires one per member), so this is deliberately far
+// tighter than the interactive 120s tip budget: if the model cannot turn a
+// handful of facts into two sentences this fast, it is effectively down, and a
+// template dispatch now beats a voiced one late.
+const dispatchLLMTimeout = 15 * time.Second
+
+// Length ceilings, mirrored from Pete's proseGuard so we never ship prose Pete
+// will reject for length alone. Byte counts, matching Pete's len() check.
+const (
+	maxDispatchHeadline = 200
+	maxDispatchLede     = 800
+)
+
+var dispatchHTTP = &http.Client{Timeout: dispatchLLMTimeout}
+
+// authorDispatch turns a fact into a headline+lede in Pete's voice, or returns
+// two empty strings if the model is unconfigured, errors, times out, or produces
+// anything malformed. The fact must already have its FINAL Actors set (post
+// opt-out anonymisation) — that list is the only set of names the prose may use,
+// and it is what Pete's guard checks the output against.
+func authorDispatch(f peteclient.Fact) (headline, lede string) {
+	host := os.Getenv("OLLAMA_HOST")
+	model := os.Getenv("OLLAMA_MODEL")
+	if host == "" || model == "" {
+		return "", ""
+	}
+
+	prompt := buildDispatchPrompt(f)
+	raw, err := callOllamaDispatch(host, model, prompt)
+	if err != nil {
+		slog.Warn("pete dispatch: LLM authoring failed, Pete will template", "guid", f.GUID, "err", err)
+		return "", ""
+	}
+
+	h, l, ok := parseDispatch(raw)
+	if !ok {
+		slog.Warn("pete dispatch: unparseable LLM output, Pete will template", "guid", f.GUID)
+		return "", ""
+	}
+	// Ship only a complete, in-bounds pair. A half-authored dispatch or an
+	// over-long one is exactly what Pete would reject anyway; catch it here so a
+	// bad generation costs nothing on the wire.
+	if h == "" || l == "" || len(h) > maxDispatchHeadline || len(l) > maxDispatchLede {
+		slog.Warn("pete dispatch: LLM output empty or over length, Pete will template",
+			"guid", f.GUID, "headline_len", len(h), "lede_len", len(l))
+		return "", ""
+	}
+	return h, l
+}
+
+// buildDispatchPrompt renders the persona, the strict rules, and this fact's
+// structured facts into a single prompt. The facts block lists only the fields
+// that are set, each labelled, so the model has the who/what/where and no room
+// to invent the rest.
+func buildDispatchPrompt(f peteclient.Fact) string {
+	var facts strings.Builder
+	add := func(label, val string) {
+		if val != "" {
+			fmt.Fprintf(&facts, "- %s: %s\n", label, val)
+		}
+	}
+	addN := func(label string, n int) {
+		if n != 0 {
+			fmt.Fprintf(&facts, "- %s: %d\n", label, n)
+		}
+	}
+	add("event", f.EventType)
+	add("who this is about (the subject)", f.Subject)
+	add("the other person named", f.Opponent)
+	add("monster or boss", f.Boss)
+	add("dungeon or zone", f.Zone)
+	add("region", f.Region)
+	addN("character level", f.Level)
+	addN("count", f.Count)
+	add("outcome", f.Outcome)
+	add("stakes or item", f.Stakes)
+	add("class and race", f.ClassRace)
+	add("milestone", f.Milestone)
+
+	names := "(none — this is a realm-level event with no named adventurer)"
+	if len(f.Actors) > 0 {
+		names = strings.Join(f.Actors, ", ")
+	}
+
+	return fmt.Sprintf(`You are Pete, a warm, friendly local news reporter for a fantasy adventuring town. Think a beloved local newscaster who genuinely knows everyone and is glad to see them. You have journalistic bones — a clear headline and a who/what/where lede that gets the facts right — delivered with personable, first-person warmth. You root for the community, celebrate wins, mourn losses gently, welcome newcomers. Conversational, never snarky, never a caps-lock hype-man. Warmth carries the register, not exclamation marks.
+
+Write a short news dispatch about the event below.
+
+STRICT RULES — do not violate these:
+- Use ONLY these adventurer names, exactly as written: %s. Never invent a name, never use any other person's name, never use an @-handle.
+- Use ONLY the facts listed. Do not invent numbers, outcomes, items, or events that are not below.
+- The monster/boss, zone, region and item names are game names — you may use them as given.
+- Do not address the reader as "you" unless the event is Pete's own duel.
+- No markdown, no emoji, no quotation marks around the whole thing.
+
+Respond with ONLY a JSON object, no other text:
+{"headline": "one short sentence, a real headline", "lede": "one to three warm sentences with the who/what/where"}
+
+The event:
+%s`, names, facts.String())
+}
+
+// callOllamaDispatch posts a single non-streaming generation and returns the raw
+// completion (think-tags stripped). Its own bounded client, separate from the
+// interactive callOllama, because the game loop cannot wait 120s on the news.
+func callOllamaDispatch(host, model, prompt string) (string, error) {
+	payload := map[string]interface{}{
+		"model":  model,
+		"prompt": prompt,
+		"stream": false,
+		"think":  false,
+		"options": map[string]interface{}{
+			"num_ctx": 4096,
+		},
+	}
+	data, err := json.Marshal(payload)
+	if err != nil {
+		return "", fmt.Errorf("marshal payload: %w", err)
+	}
+	apiURL := strings.TrimRight(host, "/") + "/api/generate"
+	resp, err := dispatchHTTP.Post(apiURL, "application/json", bytes.NewReader(data))
+	if err != nil {
+		return "", fmt.Errorf("ollama request: %w", err)
+	}
+	defer resp.Body.Close()
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return "", fmt.Errorf("read response: %w", err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		return "", fmt.Errorf("ollama HTTP %d: %s", resp.StatusCode, string(body))
+	}
+	var result struct {
+		Response string `json:"response"`
+	}
+	if err := json.Unmarshal(body, &result); err != nil {
+		return "", fmt.Errorf("parse response: %w", err)
+	}
+	return result.Response, nil
+}
+
+// parseDispatch pulls {headline, lede} out of the model's completion, tolerating
+// the usual noise (think blocks, markdown fences, prose around the JSON). ok is
+// false when no JSON object with a headline can be recovered.
+func parseDispatch(raw string) (headline, lede string, ok bool) {
+	s := raw
+	// Drop a Qwen-style reasoning block if present.
+	if i := strings.Index(s, "<think>"); i != -1 {
+		if j := strings.Index(s, "</think>"); j != -1 {
+			s = s[:i] + s[j+len("</think>"):]
+		}
+	}
+	// Isolate the first {...} object so surrounding prose or fences don't break
+	// the decode.
+	start := strings.Index(s, "{")
+	end := strings.LastIndex(s, "}")
+	if start < 0 || end <= start {
+		return "", "", false
+	}
+	var out struct {
+		Headline string `json:"headline"`
+		Lede     string `json:"lede"`
+	}
+	if err := json.Unmarshal([]byte(s[start:end+1]), &out); err != nil {
+		return "", "", false
+	}
+	headline = strings.TrimSpace(out.Headline)
+	lede = strings.TrimSpace(out.Lede)
+	if headline == "" {
+		return "", "", false
+	}
+	return headline, lede, true
+}

internal/plugin/pete_dispatch_voice_test.go

@@ -0,0 +1,81 @@
+package plugin
+
+import (
+	"strings"
+	"testing"
+
+	"gogobee/internal/peteclient"
+)
+
+func TestParseDispatch(t *testing.T) {
+	cases := []struct {
+		name        string
+		raw         string
+		wantOK      bool
+		wantHeadPre string
+	}{
+		{
+			name:        "clean json",
+			raw:         `{"headline": "Josie cleared the Ossuary.", "lede": "Alone, no less."}`,
+			wantOK:      true,
+			wantHeadPre: "Josie cleared",
+		},
+		{
+			name:        "wrapped in prose and fences",
+			raw:         "Sure! Here you go:\n```json\n{\"headline\":\"A win.\",\"lede\":\"Nice one.\"}\n```",
+			wantOK:      true,
+			wantHeadPre: "A win.",
+		},
+		{
+			name:        "think block stripped",
+			raw:         "<think>let me consider the tone</think>\n{\"headline\":\"Held the line.\",\"lede\":\"Proud of you all.\"}",
+			wantOK:      true,
+			wantHeadPre: "Held the line.",
+		},
+		{name: "no json", raw: "I could not write that.", wantOK: false},
+		{name: "empty headline", raw: `{"headline":"","lede":"body"}`, wantOK: false},
+	}
+	for _, c := range cases {
+		t.Run(c.name, func(t *testing.T) {
+			h, l, ok := parseDispatch(c.raw)
+			if ok != c.wantOK {
+				t.Fatalf("ok = %v, want %v (h=%q l=%q)", ok, c.wantOK, h, l)
+			}
+			if ok && !strings.HasPrefix(h, c.wantHeadPre) {
+				t.Errorf("headline = %q, want prefix %q", h, c.wantHeadPre)
+			}
+		})
+	}
+}
+
+// TestBuildDispatchPrompt pins the two properties the prose-guard depends on:
+// the allowed names are stated verbatim, and only the set facts appear (no empty
+// labels for the model to fill in with invention).
+func TestBuildDispatchPrompt(t *testing.T) {
+	f := peteclient.Fact{
+		EventType: "boss_kill", Subject: "Josie", Boss: "the Bone Warden",
+		Zone: "the Ossuary", Level: 14, Actors: []string{"Josie"},
+	}
+	p := buildDispatchPrompt(f)
+
+	if !strings.Contains(p, "ONLY these adventurer names, exactly as written: Josie") {
+		t.Errorf("prompt does not constrain names to Actors:\n%s", p)
+	}
+	if !strings.Contains(p, "the Bone Warden") || !strings.Contains(p, "the Ossuary") {
+		t.Error("prompt dropped a supplied fact")
+	}
+	// Unset fields must not appear as empty labels.
+	if strings.Contains(p, "region:") || strings.Contains(p, "milestone:") {
+		t.Errorf("prompt lists an unset fact:\n%s", p)
+	}
+}
+
+// TestBuildDispatchPromptRealmEvent: a realm-level fact with no named adventurer
+// still produces a usable prompt that tells the model there is no name to use.
+func TestBuildDispatchPromptRealmEvent(t *testing.T) {
+	f := peteclient.Fact{EventType: "siege_start", Boss: "the Horde", Stakes: "the whole town"}
+	p := buildDispatchPrompt(f)
+	if !strings.Contains(p, "no named adventurer") {
+		t.Errorf("realm event prompt missing the no-name note:\n%s", p)
+	}
+}