!explain via ❓ reaction: thread-reply with a 3-bullet TL;DR

When a user reacts ❓ on one of Pete's posts, fetch the article body,
ask Ollama for a 3-bullet summary, and post it as a threaded reply
rooted at the original story event. Per-process cooldown of 5min per
story keeps repeated reactions from re-summarizing.

- ingestion.FetchArticleBody: visible <p> text capped at 8000 chars
- classifier.OllamaClient.GenerateText: non-JSON variant
- storage.GetStoryByGUID: full row lookup
- matrix.PostThreadedReply: m.thread + m.in_reply_to fallback
- poster.SetReactionCallback: optional hook fired after recording
- New package: internal/explainer

internal/classifier/ollama.go

@@ -51,20 +51,28 @@ func NewOllamaClient(cfg config.OllamaConfig) *OllamaClient {
 
 // Generate sends a prompt to Ollama via the chat API and returns the raw response text.
 func (o *OllamaClient) Generate(prompt string) (string, error) {
+	return o.call(
+		"You are a JSON-only news classification API. Respond with valid JSON only. No preamble, no explanation, no markdown.",
+		prompt,
+		"json",
+	)
+}
+
+// GenerateText sends a system + user prompt to Ollama and returns the raw
+// text response. Unlike Generate, no JSON mode is forced.
+func (o *OllamaClient) GenerateText(systemPrompt, userPrompt string) (string, error) {
+	return o.call(systemPrompt, userPrompt, "")
+}
+
+func (o *OllamaClient) call(systemPrompt, userPrompt, format string) (string, error) {
 	reqBody := chatRequest{
 		Model: o.model,
 		Messages: []chatMessage{
-			{
-				Role:    "system",
-				Content: "You are a JSON-only news classification API. Respond with valid JSON only. No preamble, no explanation, no markdown.",
-			},
-			{
-				Role:    "user",
-				Content: prompt,
-			},
+			{Role: "system", Content: systemPrompt},
+			{Role: "user", Content: userPrompt},
 		},
 		Stream: false,
-		Format: "json",
+		Format: format,
 	}
 
 	body, err := json.Marshal(reqBody)

internal/explainer/explainer.go

@@ -0,0 +1,169 @@
+// Package explainer summarizes a posted story on demand: when a user reacts ❓
+// on one of Pete's posts, Pete fetches the article body, asks Ollama for a
+// 3-bullet TL;DR, and replies in a thread rooted at the original post.
+package explainer
+
+import (
+	"fmt"
+	"html"
+	"log/slog"
+	"strings"
+	"sync"
+	"time"
+
+	"pete/internal/classifier"
+	"pete/internal/ingestion"
+	"pete/internal/matrix"
+	"pete/internal/storage"
+
+	"maunium.net/go/mautrix/id"
+)
+
+// QuestionEmoji is the reaction key that triggers a summary. Matrix delivers
+// the reaction key verbatim — Element sends "❓" (U+2753).
+const QuestionEmoji = "❓"
+
+// cooldown for repeat-explain on the same story (per process).
+const explainCooldown = 5 * time.Minute
+
+// Explainer holds the dependencies needed to produce a story summary.
+type Explainer struct {
+	mx     *matrix.Client
+	ollama *classifier.OllamaClient
+
+	mu       sync.Mutex
+	lastSeen map[string]time.Time // guid -> last explanation time
+}
+
+// New builds an Explainer wired to the Matrix client and Ollama backend.
+func New(mx *matrix.Client, ollama *classifier.OllamaClient) *Explainer {
+	return &Explainer{
+		mx:       mx,
+		ollama:   ollama,
+		lastSeen: make(map[string]time.Time),
+	}
+}
+
+// Handle is the entry point used as a callback from poster.HandleReaction.
+// It runs synchronously — callers should invoke it in a goroutine.
+func (e *Explainer) Handle(roomID id.RoomID, rootEventID id.EventID, guid, channel, emoji string) {
+	if emoji != QuestionEmoji {
+		return
+	}
+	if !e.acquire(guid) {
+		slog.Debug("explain: skipping recent duplicate", "guid", guid)
+		return
+	}
+
+	story, err := storage.GetStoryByGUID(guid)
+	if err != nil || story == nil {
+		slog.Warn("explain: story lookup failed", "guid", guid, "err", err)
+		return
+	}
+
+	body := ingestion.FetchArticleBody(story.ArticleURL)
+	if body == "" {
+		// Couldn't get the body — fall back to lede so the user at least gets something.
+		body = story.Lede
+	}
+	if body == "" {
+		slog.Warn("explain: no body or lede available", "guid", guid, "url", story.ArticleURL)
+		_ = e.mx.PostThreadedReply(channel, rootEventID,
+			"Couldn't fetch the article body to summarize.",
+			"<em>Couldn't fetch the article body to summarize.</em>")
+		return
+	}
+
+	summary, err := e.summarize(story.Headline, body)
+	if err != nil {
+		slog.Error("explain: ollama failed", "guid", guid, "err", err)
+		_ = e.mx.PostThreadedReply(channel, rootEventID,
+			"Sorry, I couldn't generate a summary right now.",
+			"<em>Sorry, I couldn't generate a summary right now.</em>")
+		return
+	}
+
+	plain, htmlBody := formatSummary(summary)
+	if err := e.mx.PostThreadedReply(channel, rootEventID, plain, htmlBody); err != nil {
+		slog.Error("explain: send reply failed", "guid", guid, "err", err)
+		return
+	}
+	slog.Info("explain: summary posted", "guid", guid, "channel", channel)
+}
+
+// acquire returns true if this guid hasn't been explained within the cooldown.
+// Updates the timestamp on success.
+func (e *Explainer) acquire(guid string) bool {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	if t, ok := e.lastSeen[guid]; ok && time.Since(t) < explainCooldown {
+		return false
+	}
+	e.lastSeen[guid] = time.Now()
+	return true
+}
+
+const summarySystem = `You summarize news articles in exactly 3 short bullet points for a chat reader.
+Rules:
+- Output ONLY the 3 bullets, each on its own line, prefixed with "- ".
+- No preamble, no headings, no closing remarks.
+- Each bullet is one concise sentence (under 25 words).
+- Stick strictly to facts in the article. Do not speculate.`
+
+func (e *Explainer) summarize(headline, body string) (string, error) {
+	prompt := fmt.Sprintf("Headline: %s\n\nArticle body:\n%s", headline, body)
+	out, err := e.ollama.GenerateText(summarySystem, prompt)
+	if err != nil {
+		return "", err
+	}
+	return strings.TrimSpace(out), nil
+}
+
+// formatSummary converts the LLM's "- bullet\n- bullet" output into a Matrix
+// plain + HTML pair. Any stray prefixes ("Summary:", "•", "*") are normalized
+// to "-" bullets.
+func formatSummary(raw string) (plain, htmlBody string) {
+	lines := strings.Split(raw, "\n")
+	var bullets []string
+	for _, line := range lines {
+		line = strings.TrimSpace(line)
+		if line == "" {
+			continue
+		}
+		// Normalize common bullet markers
+		for _, prefix := range []string{"- ", "* ", "• ", "·"} {
+			if strings.HasPrefix(line, prefix) {
+				line = strings.TrimSpace(line[len(prefix):])
+				break
+			}
+		}
+		// Skip header-ish lines the model sometimes adds
+		low := strings.ToLower(line)
+		if strings.HasPrefix(low, "summary:") || strings.HasPrefix(low, "tl;dr") {
+			continue
+		}
+		if line == "" {
+			continue
+		}
+		bullets = append(bullets, line)
+	}
+	if len(bullets) == 0 {
+		// Model produced something but we couldn't parse bullets — pass through.
+		return raw, "<pre>" + html.EscapeString(raw) + "</pre>"
+	}
+
+	var pb, hb strings.Builder
+	hb.WriteString("<ul>")
+	for i, b := range bullets {
+		if i > 0 {
+			pb.WriteByte('\n')
+		}
+		pb.WriteString("• ")
+		pb.WriteString(b)
+		hb.WriteString("<li>")
+		hb.WriteString(html.EscapeString(b))
+		hb.WriteString("</li>")
+	}
+	hb.WriteString("</ul>")
+	return pb.String(), hb.String()
+}

internal/ingestion/article.go

@@ -101,6 +101,69 @@ func FetchOGImage(articleURL string) string {
 	return FetchArticleMeta(articleURL).ImageURL
 }
 
+// MaxBodyChars is the cap on body text returned by FetchArticleBody. Keeps
+// LLM prompts bounded; most news articles fit well under this.
+const MaxBodyChars = 8000
+
+// FetchArticleBody fetches the article and returns the concatenated visible
+// body text (<article>/<main> <p> tags, falling back to all <p>), trimmed
+// and capped at MaxBodyChars. Returns "" on any fetch failure.
+func FetchArticleBody(articleURL string) string {
+	if articleURL == "" {
+		return ""
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 12*time.Second)
+	defer cancel()
+
+	req, err := http.NewRequestWithContext(ctx, "GET", articleURL, nil)
+	if err != nil {
+		return ""
+	}
+	req.Header.Set("User-Agent", userAgent)
+	req.Header.Set("Accept", "text/html,application/xhtml+xml")
+
+	resp, err := articleClient.Do(req)
+	if err != nil {
+		return ""
+	}
+	defer resp.Body.Close()
+	if resp.StatusCode != http.StatusOK {
+		return ""
+	}
+
+	doc, err := goquery.NewDocumentFromReader(resp.Body)
+	if err != nil {
+		return ""
+	}
+	return extractBodyText(doc)
+}
+
+func extractBodyText(doc *goquery.Document) string {
+	sel := doc.Find("article p, main p")
+	if sel.Length() == 0 {
+		sel = doc.Find("p")
+	}
+	var b strings.Builder
+	sel.Each(func(_ int, s *goquery.Selection) {
+		t := strings.TrimSpace(s.Text())
+		if t == "" {
+			return
+		}
+		if b.Len() > 0 {
+			b.WriteString("\n\n")
+		}
+		b.WriteString(t)
+		if b.Len() >= MaxBodyChars {
+			return
+		}
+	})
+	out := strings.TrimSpace(b.String())
+	if len(out) > MaxBodyChars {
+		out = out[:MaxBodyChars]
+	}
+	return out
+}
+
 func extractOGImage(doc *goquery.Document, base string) string {
 	selectors := []string{
 		`meta[property="og:image:secure_url"]`,

internal/matrix/client.go

@@ -294,6 +294,35 @@ func (c *Client) Stop() {
 	}
 }
 
+// PostThreadedReply sends a plain/HTML message into the thread rooted at
+// rootEventID. Resolves channel name → room ID like PostStory. The is_falling_back
+// flag plus m.in_reply_to gives non-thread-aware clients a sensible quoted reply.
+func (c *Client) PostThreadedReply(channel string, rootEventID id.EventID, plain, htmlBody string) error {
+	roomID, ok := c.channels[channel]
+	if !ok {
+		return fmt.Errorf("unknown channel: %s", channel)
+	}
+	ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
+	defer cancel()
+
+	content := &event.MessageEventContent{
+		MsgType:       event.MsgText,
+		Body:          plain,
+		Format:        event.FormatHTML,
+		FormattedBody: htmlBody,
+		RelatesTo: &event.RelatesTo{
+			Type:    event.RelThread,
+			EventID: rootEventID,
+			InReplyTo: &event.InReplyTo{
+				EventID: rootEventID,
+			},
+			IsFallingBack: true,
+		},
+	}
+	_, err := c.mx.SendMessageEvent(ctx, roomID, event.EventMessage, content)
+	return err
+}
+
 // PostStory sends a story to a channel. Returns the text event ID for reaction tracking.
 // If the story has a validated image, it's uploaded and sent as a separate m.image event first.
 func (c *Client) PostStory(channel string, story *PostableStory) (id.EventID, error) {

internal/poster/tracker.go

@@ -2,6 +2,7 @@ package poster
 
 import (
 	"log/slog"
+	"sync"
 	"time"
 
 	"pete/internal/storage"
@@ -9,6 +10,24 @@ import (
 	"maunium.net/go/mautrix/id"
 )
 
+// ReactionCallback is an optional hook fired after a reaction is recorded.
+// Used to wire reaction-driven features (e.g. ❓ → summary) without coupling
+// poster to those packages.
+type ReactionCallback func(roomID id.RoomID, targetEventID id.EventID, guid, channel, emoji string)
+
+var (
+	reactionCallbackMu sync.RWMutex
+	reactionCallback   ReactionCallback
+)
+
+// SetReactionCallback installs a hook called after every recorded reaction.
+// Pass nil to clear. Safe to call before or after the matrix client starts.
+func SetReactionCallback(fn ReactionCallback) {
+	reactionCallbackMu.Lock()
+	reactionCallback = fn
+	reactionCallbackMu.Unlock()
+}
+
 // HandleReaction processes a reaction event by mapping it back to the story GUID.
 func HandleReaction(roomID id.RoomID, eventID id.EventID, targetEventID id.EventID, emoji string, userID id.UserID) {
 	guid, channel, found := storage.LookupPostGUID(string(targetEventID))
@@ -24,4 +43,11 @@ func HandleReaction(roomID id.RoomID, eventID id.EventID, targetEventID id.Event
 		"emoji", emoji,
 		"user", userID,
 	)
+
+	reactionCallbackMu.RLock()
+	cb := reactionCallback
+	reactionCallbackMu.RUnlock()
+	if cb != nil {
+		go cb(roomID, targetEventID, guid, channel, emoji)
+	}
 }

internal/storage/queries.go

@@ -118,6 +118,18 @@ func GetUnclassifiedStories(source string) ([]Story, error) {
 	return stories, rows.Err()
 }
 
+// GetStoryByGUID returns the full story record for a GUID, or nil if not found.
+func GetStoryByGUID(guid string) (*Story, error) {
+	row := Get().QueryRow(
+		`SELECT guid, headline, lede, image_url, article_url, source, feed_hint, platforms, channel, seen_at
+		 FROM stories WHERE guid = ?`, guid)
+	var s Story
+	if err := row.Scan(&s.GUID, &s.Headline, &s.Lede, &s.ImageURL, &s.ArticleURL, &s.Source, &s.FeedHint, &s.Platforms, &s.Channel, &s.SeenAt); err != nil {
+		return nil, err
+	}
+	return &s, nil
+}
+
 // InsertRecentHeadline adds a headline to the dedup context window.
 func InsertRecentHeadline(guid, headline, source string) {
 	exec("insert recent_headline",

main.go

@@ -11,6 +11,7 @@ import (
 
 	"pete/internal/classifier"
 	"pete/internal/config"
+	"pete/internal/explainer"
 	"pete/internal/ingestion"
 	"pete/internal/matrix"
 	"pete/internal/poster"
@@ -73,8 +74,10 @@ func main() {
 	// Create post queue
 	queue := poster.NewQueue(cfg.Posting, mx)
 
-	// Wire reaction handler
+	// Wire reaction handler + ❓ → summary hook
 	mx.SetReactionHandler(poster.HandleReaction)
+	exp := explainer.New(mx, ollamaClient)
+	poster.SetReactionCallback(exp.Handle)
 
 	// Set up graceful shutdown
 	ctx, cancel := context.WithCancel(context.Background())