llm: route every LLM caller through a shared backend client

Replaces the per-plugin Ollama HTTP calls with internal/llm, which picks a
backend from the environment (vLLM or Ollama) behind one Chat interface, plus
internal/plugin/llm_client.go as the plugin-facing wrapper.

Startup now logs llm_backend/llm_endpoint/llm_model instead of the two
OLLAMA_* vars, which no longer describe where inference actually goes.

These files were already running in prod from the vLLM migration but had never
been committed; this is that live state, byte-for-byte.

internal/plugin/botinfo.go

@@ -1,12 +1,9 @@
 package plugin
 
 import (
-	"encoding/json"
+	"context"
 	"fmt"
-	"io"
 	"log/slog"
-	"net/http"
-	"os"
 	"strings"
 	"sync/atomic"
 	"time"
@@ -121,10 +118,8 @@ func (p *BotInfoPlugin) handleBotInfo(ctx MessageContext) error {
 	sb.WriteString(fmt.Sprintf("Active reminders: %d\n", activeReminders))
 
 	// LLM status
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	if ollamaHost != "" {
-		llmStatus := p.checkLLMStatus(ollamaHost)
-		sb.WriteString(fmt.Sprintf("LLM status: %s\n", llmStatus))
+	if llmConfigured() {
+		sb.WriteString(fmt.Sprintf("LLM status: %s\n", p.checkLLMStatus()))
 	} else {
 		sb.WriteString("LLM status: not configured\n")
 	}
@@ -159,42 +154,16 @@ func (p *BotInfoPlugin) handleBotInfo(ctx MessageContext) error {
 	return p.SendReply(ctx.RoomID, ctx.EventID, sb.String())
 }
 
-func (p *BotInfoPlugin) checkLLMStatus(ollamaHost string) string {
-	client := &http.Client{Timeout: 5 * time.Second}
-	apiURL := strings.TrimRight(ollamaHost, "/") + "/api/tags"
-
-	resp, err := client.Get(apiURL)
+// checkLLMStatus reports backend liveness for /botinfo. The endpoint it probes
+// differs per backend, which the llm package hides behind Ping.
+func (p *BotInfoPlugin) checkLLMStatus() string {
+	c := llmClient()
+	models, err := c.Ping(context.Background())
 	if err != nil {
-		return fmt.Sprintf("offline (%s)", err.Error())
+		return fmt.Sprintf("offline (%s: %s)", c.Backend(), err.Error())
 	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != 200 {
-		return fmt.Sprintf("error (HTTP %d)", resp.StatusCode)
+	if len(models) == 0 {
+		return fmt.Sprintf("online (%s, no models loaded)", c.Backend())
 	}
-
-	body, err := io.ReadAll(resp.Body)
-	if err != nil {
-		return "online (could not read response)"
-	}
-
-	var result struct {
-		Models []struct {
-			Name string `json:"name"`
-		} `json:"models"`
-	}
-	if err := json.Unmarshal(body, &result); err != nil {
-		return "online (could not parse response)"
-	}
-
-	modelNames := make([]string, 0, len(result.Models))
-	for _, m := range result.Models {
-		modelNames = append(modelNames, m.Name)
-	}
-
-	if len(modelNames) == 0 {
-		return "online (no models loaded)"
-	}
-
-	return fmt.Sprintf("online (%d models: %s)", len(modelNames), strings.Join(modelNames, ", "))
+	return fmt.Sprintf("online (%s, %d models: %s)", c.Backend(), len(models), strings.Join(models, ", "))
 }

internal/plugin/hangman.go

@@ -856,9 +856,7 @@ func (p *HangmanPlugin) handleSubmit(ctx MessageContext, phrase string) error {
 	}
 
 	// LLM screening
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	ollamaModel := os.Getenv("OLLAMA_MODEL")
-	if ollamaHost == "" || ollamaModel == "" {
+	if !llmConfigured() {
 		// No LLM available — add directly
 		if err := p.addPhrase(phrase); err != nil {
 			if err.Error() == "duplicate phrase" {
@@ -879,7 +877,7 @@ or
 
 Phrase: %s`, phrase)
 
-	result, err := callOllama(ollamaHost, ollamaModel, prompt)
+	result, err := callLLM(prompt)
 	if err != nil {
 		slog.Error("hangman: LLM screening failed", "err", err)
 		// Fail open — add it

internal/plugin/holdem_tips.go

@@ -1,24 +1,24 @@
 package plugin
 
 import (
-	"bytes"
-	"encoding/json"
+	"context"
 	"fmt"
-	"io"
 	"log/slog"
-	"net/http"
-	"os"
 	"regexp"
 	"strings"
 	"time"
 
 	"gogobee/internal/db"
+	"gogobee/internal/llm"
 
 	"github.com/chehsunliu/poker"
 	"maunium.net/go/mautrix/id"
 )
 
-var holdemTipsClient = &http.Client{Timeout: 60 * time.Second}
+// holdemTipTimeout preserves the 60s budget the tip rewriter's own http.Client
+// enforced. Tips are delivered as private messages during a hand, so this sits
+// between the passive 30s paths and the interactive 120s default.
+const holdemTipTimeout = 60 * time.Second
 
 // loadTipsPref loads a user's tip preference from the database.
 func loadTipsPref(userID id.UserID) bool {
@@ -491,11 +491,8 @@ func cardSuitIndex(c poker.Card) int {
 func generateTip(ctx holdemTipContext) string {
 	base := generateRulesTip(ctx)
 
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-
-	if host != "" && model != "" {
-		rewritten, err := rewriteTipWithLLM(host, model, ctx, base)
+	if llmConfigured() {
+		rewritten, err := rewriteTipWithLLM(ctx, base)
 		if err != nil {
 			slog.Warn("holdem: LLM tip rewrite failed, using rules tip", "err", err)
 		} else if rewritten != "" {
@@ -626,40 +623,19 @@ func buildTipUserPrompt(ctx holdemTipContext) string {
 // variety. The rules tip is the source of truth — if the rewrite diverges
 // (empty, action vocabulary changed, etc.) we reject it and the caller falls
 // back to the original.
-func rewriteTipWithLLM(host, model string, ctx holdemTipContext, base string) (string, error) {
+func rewriteTipWithLLM(ctx holdemTipContext, base string) (string, error) {
 	userMsg := buildTipUserPrompt(ctx) + "\nTIP:\n" + base + "\n"
-	req := ollamaChatRequest{
-		Model: model,
-		Messages: []chatMessage{
-			{Role: "system", Content: buildTipSystemPrompt()},
-			{Role: "user", Content: userMsg},
-		},
-		Stream: false,
-	}
 
-	body, err := json.Marshal(req)
+	raw, err := llmGenerate(context.Background(), llm.Request{
+		System:  buildTipSystemPrompt(),
+		Prompt:  userMsg,
+		Timeout: holdemTipTimeout,
+	})
 	if err != nil {
-		return "", fmt.Errorf("marshal: %w", err)
+		return "", err
 	}
 
-	url := strings.TrimRight(host, "/") + "/api/chat"
-	resp, err := holdemTipsClient.Post(url, "application/json", bytes.NewReader(body))
-	if err != nil {
-		return "", fmt.Errorf("request: %w", err)
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != http.StatusOK {
-		respBody, _ := io.ReadAll(resp.Body)
-		return "", fmt.Errorf("status %d: %s", resp.StatusCode, string(respBody))
-	}
-
-	var ollamaResp ollamaChatResponse
-	if err := json.NewDecoder(resp.Body).Decode(&ollamaResp); err != nil {
-		return "", fmt.Errorf("decode: %w", err)
-	}
-
-	tip := extractTipFromResponse(ollamaResp.Message.Content)
+	tip := extractTipFromResponse(raw)
 	if tip == "" {
 		return "", fmt.Errorf("empty response")
 	}

internal/plugin/howami.go

@@ -1,18 +1,15 @@
 package plugin
 
 import (
-	"bytes"
+	"context"
 	"database/sql"
-	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
-	"net/http"
 	"os"
 	"strings"
-	"time"
 
 	"gogobee/internal/db"
+	"gogobee/internal/llm"
 
 	"maunium.net/go/mautrix"
 	"maunium.net/go/mautrix/id"
@@ -47,9 +44,7 @@ func (p *HowAmIPlugin) OnMessage(ctx MessageContext) error {
 		return nil
 	}
 
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	ollamaModel := os.Getenv("OLLAMA_MODEL")
-	if ollamaHost == "" || ollamaModel == "" {
+	if !llmConfigured() {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "LLM is not configured.")
 	}
 
@@ -84,9 +79,9 @@ Write the roast now. Do not include any preamble or explanation, just the roast
 			botName, string(target), profile,
 		)
 
-		response, err := callOllama(ollamaHost, ollamaModel, prompt)
+		response, err := callLLM(prompt)
 		if err != nil {
-			slog.Error("howami: ollama call", "err", err)
+			slog.Error("howami: llm call", "err", err)
 			p.SendReply(ctx.RoomID, ctx.EventID, "Couldn't generate the profile. Thanks, Ollama.")
 			return
 		}
@@ -181,55 +176,13 @@ func (p *HowAmIPlugin) gatherProfile(userID id.UserID) string {
 	return sb.String()
 }
 
-// callOllama sends a prompt to the Ollama generate endpoint and returns the response.
-func callOllama(host, model, prompt string) (string, error) {
-	apiURL := strings.TrimRight(host, "/") + "/api/generate"
-
-	payload := map[string]interface{}{
-		"model":  model,
-		"prompt": prompt,
-		"stream": false,
-		"think":  false,
-		"options": map[string]interface{}{
-			"num_ctx": 8192,
-		},
-	}
-
-	data, err := json.Marshal(payload)
-	if err != nil {
-		return "", fmt.Errorf("marshal payload: %w", err)
-	}
-
-	client := &http.Client{Timeout: 120 * time.Second}
-	resp, err := client.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 != 200 {
-		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)
-	}
-
-	response := result.Response
-	// Strip <think>...</think> blocks (Qwen 3.5 reasoning)
-	if i := strings.Index(response, "<think>"); i != -1 {
-		if j := strings.Index(response, "</think>"); j != -1 {
-			response = response[:i] + response[j+len("</think>"):]
-		}
-	}
-
-	return strings.TrimSpace(response), nil
+// callLLM sends a prompt to whichever backend is configured and returns the
+// completion. Reasoning blocks are stripped by the client. The 8192 context
+// hint is what this path has always asked Ollama for; vLLM ignores it and uses
+// the window fixed at server launch.
+func callLLM(prompt string) (string, error) {
+	return llmGenerate(context.Background(), llm.Request{
+		Prompt: prompt,
+		NumCtx: 8192,
+	})
 }

internal/plugin/llm_client.go

@@ -0,0 +1,45 @@
+package plugin
+
+import (
+	"context"
+	"sync"
+
+	"gogobee/internal/llm"
+)
+
+// The inference backend is process-wide: one endpoint, one model, selected by
+// env at startup. Plugins share a single client rather than each rebuilding one
+// per invocation, and read config through llmConfigured/llmGenerate so that
+// swapping Ollama for vLLM is a config change rather than a code change.
+var (
+	llmOnce   sync.Once
+	llmShared llm.Client
+	llmCfg    llm.Config
+)
+
+func llmInit() {
+	llmOnce.Do(func() {
+		llmCfg = llm.ConfigFromEnv()
+		llmShared = llm.New(llmCfg)
+	})
+}
+
+// llmConfigured reports whether an endpoint and model are set. Plugins call
+// this to stay dormant instead of erroring on every invocation — the same role
+// the old `if ollamaHost == "" || ollamaModel == ""` guards played.
+func llmConfigured() bool {
+	llmInit()
+	return llmCfg.Configured()
+}
+
+// llmClient returns the shared backend client.
+func llmClient() llm.Client {
+	llmInit()
+	return llmShared
+}
+
+// llmGenerate is the one-line path for the common case: a raw prompt in,
+// visible completion out, reasoning blocks already stripped.
+func llmGenerate(ctx context.Context, req llm.Request) (string, error) {
+	return llmClient().Generate(ctx, req)
+}

internal/plugin/llm_passive.go

@@ -1,13 +1,11 @@
 package plugin
 
 import (
-	"bytes"
+	"context"
 	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
 	"math/rand"
-	"net/http"
 	"os"
 	"regexp"
 	"strconv"
@@ -18,6 +16,7 @@ import (
 
 	"gogobee/internal/db"
 	"gogobee/internal/dreamclient"
+	"gogobee/internal/llm"
 
 	"maunium.net/go/mautrix"
 	"maunium.net/go/mautrix/id"
@@ -54,29 +53,31 @@ type queueItem struct {
 	FormattedBody string
 }
 
-// LLMPassivePlugin classifies messages using Ollama and reacts accordingly.
+// classifyTimeout is the per-message budget for passive classification. It
+// preserves the 30s cap the plugin's own http.Client used to enforce, which is
+// deliberately tighter than the interactive default: classification runs on
+// sampled traffic and must never back up the queue.
+const classifyTimeout = 30 * time.Second
+
+// LLMPassivePlugin classifies messages using the configured LLM backend and
+// reacts accordingly.
 type LLMPassivePlugin struct {
 	Base
-	xp          *XPPlugin
-	dict        *dreamclient.Client
-	ollamaHost  string
-	ollamaModel string
-	sampleRate  float64
-	enabled     bool
+	xp         *XPPlugin
+	dict       *dreamclient.Client
+	sampleRate float64
+	enabled    bool
 
 	mu      sync.Mutex
 	queue   []queueItem
 	backoff time.Duration
 
-	httpClient *http.Client
-	stopCh     chan struct{}
+	stopCh chan struct{}
 }
 
 // NewLLMPassivePlugin creates a new LLM passive classification plugin.
 func NewLLMPassivePlugin(client *mautrix.Client, xp *XPPlugin, dict *dreamclient.Client) *LLMPassivePlugin {
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	enabled := host != "" && model != ""
+	enabled := llmConfigured()
 
 	sampleRate := 0.15
 	if v := os.Getenv("LLM_SAMPLE_RATE"); v != "" {
@@ -86,16 +87,13 @@ func NewLLMPassivePlugin(client *mautrix.Client, xp *XPPlugin, dict *dreamclient
 	}
 
 	p := &LLMPassivePlugin{
-		Base:        NewBase(client),
-		xp:          xp,
-		dict:        dict,
-		ollamaHost:  host,
-		ollamaModel: model,
-		sampleRate:  sampleRate,
-		enabled:     enabled,
-		backoff:     5 * time.Second,
-		httpClient:  &http.Client{Timeout: 30 * time.Second},
-		stopCh:      make(chan struct{}),
+		Base:       NewBase(client),
+		xp:         xp,
+		dict:       dict,
+		sampleRate: sampleRate,
+		enabled:    enabled,
+		backoff:    5 * time.Second,
+		stopCh:     make(chan struct{}),
 	}
 
 	return p
@@ -116,11 +114,11 @@ func (p *LLMPassivePlugin) Commands() []CommandDef {
 
 func (p *LLMPassivePlugin) Init() error {
 	if p.enabled {
-		slog.Info("llm_passive: enabled", "host", p.ollamaHost, "model", p.ollamaModel, "sample_rate", p.sampleRate)
+		slog.Info("llm_passive: enabled", "backend", llmClient().Backend(),
+			"model", llmClient().Model(), "sample_rate", p.sampleRate)
 		go p.processQueue()
 	} else {
-		slog.Warn("llm_passive: disabled (OLLAMA_HOST or OLLAMA_MODEL not set)",
-			"host", p.ollamaHost, "model", p.ollamaModel)
+		slog.Warn("llm_passive: disabled (LLM endpoint or model not set)")
 	}
 	return nil
 }
@@ -329,9 +327,9 @@ func (p *LLMPassivePlugin) classifyAndProcess(item queueItem) error {
 	var todayWOTD string
 	db.Get().QueryRow(`SELECT word FROM wotd_log WHERE date = ?`, today).Scan(&todayWOTD)
 
-	result, err := p.callOllama(item.Body+mentionHint, todayWOTD)
+	result, err := p.classify(item.Body+mentionHint, todayWOTD)
 	if err != nil {
-		return fmt.Errorf("ollama call: %w", err)
+		return fmt.Errorf("llm call: %w", err)
 	}
 
 	// Resolve any display names in LLM targets back to MXIDs
@@ -464,21 +462,9 @@ func (p *LLMPassivePlugin) classifyAndProcess(item queueItem) error {
 	return nil
 }
 
-// ollamaRequest is the request body for the Ollama API.
-type ollamaRequest struct {
-	Model  string `json:"model"`
-	Prompt string `json:"prompt"`
-	Stream bool   `json:"stream"`
-	Think  bool   `json:"think"`
-}
-
-// ollamaResponse is the response from the Ollama API.
-type ollamaResponse struct {
-	Response string `json:"response"`
-}
-
-// callOllama sends a classification prompt to Ollama and parses the JSON result.
-func (p *LLMPassivePlugin) callOllama(messageText, wotd string) (*classificationResult, error) {
+// classify sends a classification prompt to the configured backend and parses
+// the JSON result.
+func (p *LLMPassivePlugin) classify(messageText, wotd string) (*classificationResult, error) {
 	wotdInstruction := `"wotd_used": false`
 	if wotd != "" {
 		wotdInstruction = fmt.Sprintf(`"wotd_used": true | false (whether the message uses the word "%s" correctly and meaningfully — not just mentioning or quoting it)`, wotd)
@@ -500,36 +486,18 @@ JSON schema:
 
 Message: %s`, wotdInstruction, messageText)
 
-	reqBody := ollamaRequest{
-		Model:  p.ollamaModel,
-		Prompt: prompt,
-		Stream: false,
-	}
-
-	body, err := json.Marshal(reqBody)
+	slog.Debug("llm_passive: calling backend", "backend", llmClient().Backend(), "model", llmClient().Model())
+	// Classification rides the passive path on every sampled message, so it keeps
+	// the tighter budget it always had rather than the interactive default.
+	raw, err := llmGenerate(context.Background(), llm.Request{
+		Prompt:  prompt,
+		Timeout: classifyTimeout,
+	})
 	if err != nil {
-		return nil, fmt.Errorf("marshal request: %w", err)
+		return nil, err
 	}
 
-	url := strings.TrimRight(p.ollamaHost, "/") + "/api/generate"
-	slog.Debug("llm_passive: calling ollama", "url", url, "model", p.ollamaModel)
-	resp, err := p.httpClient.Post(url, "application/json", bytes.NewReader(body))
-	if err != nil {
-		return nil, fmt.Errorf("ollama request: %w", err)
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != http.StatusOK {
-		respBody, _ := io.ReadAll(resp.Body)
-		return nil, fmt.Errorf("ollama status %d: %s", resp.StatusCode, string(respBody))
-	}
-
-	var ollamaResp ollamaResponse
-	if err := json.NewDecoder(resp.Body).Decode(&ollamaResp); err != nil {
-		return nil, fmt.Errorf("decode ollama response: %w", err)
-	}
-
-	result, err := parseClassification(ollamaResp.Response)
+	result, err := parseClassification(raw)
 	if err != nil {
 		return nil, fmt.Errorf("parse classification: %w", err)
 	}

internal/plugin/market.go

@@ -1,10 +1,9 @@
 package plugin
 
 import (
-	"bytes"
+	"context"
 	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
 	"math"
 	"net/http"
@@ -15,6 +14,7 @@ import (
 	"time"
 
 	"gogobee/internal/db"
+	"gogobee/internal/llm"
 
 	"maunium.net/go/mautrix"
 	"maunium.net/go/mautrix/id"
@@ -311,9 +311,7 @@ Do not use em dashes. Do not use exclamation marks. Do not offer financial advic
 If markets are closed or data is stale, note it briefly and move on.`
 
 func (p *MarketPlugin) generateDailySummary(date string) string {
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	if host == "" || model == "" {
+	if !llmConfigured() {
 		return ""
 	}
 
@@ -349,7 +347,7 @@ func (p *MarketPlugin) generateDailySummary(date string) string {
 	}
 	prompt.WriteString("\nWrite a 2-3 sentence summary.")
 
-	result, err := p.callOllamaChat(host, model, marketSystemPrompt, prompt.String())
+	result, err := p.chatLLM(marketSystemPrompt, prompt.String())
 	if err != nil {
 		slog.Error("market: ollama summary failed", "err", err)
 		return ""
@@ -358,9 +356,7 @@ func (p *MarketPlugin) generateDailySummary(date string) string {
 }
 
 func (p *MarketPlugin) generateReportSummary(snapsByDate map[string][]marketSnapshot, dateRange string) string {
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	if host == "" || model == "" {
+	if !llmConfigured() {
 		return ""
 	}
 
@@ -395,7 +391,7 @@ func (p *MarketPlugin) generateReportSummary(snapsByDate map[string][]marketSnap
 	}
 	prompt.WriteString("\nDescribe the trend in 2-3 sentences. Note any significant moves or divergences between indices.\nBe sardonic but accurate. Reference the VIX trajectory when relevant.")
 
-	result, err := p.callOllamaChat(host, model, marketSystemPrompt, prompt.String())
+	result, err := p.chatLLM(marketSystemPrompt, prompt.String())
 	if err != nil {
 		slog.Warn("market: ollama report summary failed", "err", err)
 		return ""
@@ -403,49 +399,14 @@ func (p *MarketPlugin) generateReportSummary(snapsByDate map[string][]marketSnap
 	return result
 }
 
-// callOllamaChat calls the Ollama /api/chat endpoint with a system and user message.
-// Uses the types already defined in holdem_tips.go (same package).
-func (p *MarketPlugin) callOllamaChat(host, model, systemPrompt, userPrompt string) (string, error) {
-	req := ollamaChatRequest{
-		Model: model,
-		Messages: []chatMessage{
-			{Role: "system", Content: systemPrompt},
-			{Role: "user", Content: userPrompt},
-		},
-		Stream: false,
-	}
-
-	body, err := json.Marshal(req)
-	if err != nil {
-		return "", fmt.Errorf("marshal: %w", err)
-	}
-
-	url := strings.TrimRight(host, "/") + "/api/chat"
-	resp, err := p.httpClient.Post(url, "application/json", bytes.NewReader(body))
-	if err != nil {
-		return "", fmt.Errorf("request: %w", err)
-	}
-	defer resp.Body.Close()
-
-	if resp.StatusCode != http.StatusOK {
-		respBody, _ := io.ReadAll(resp.Body)
-		return "", fmt.Errorf("status %d: %s", resp.StatusCode, string(respBody))
-	}
-
-	var ollamaResp ollamaChatResponse
-	if err := json.NewDecoder(resp.Body).Decode(&ollamaResp); err != nil {
-		return "", fmt.Errorf("decode: %w", err)
-	}
-
-	text := ollamaResp.Message.Content
-	// Strip <think>...</think> blocks (reasoning models)
-	if i := strings.Index(text, "<think>"); i != -1 {
-		if j := strings.Index(text, "</think>"); j != -1 {
-			text = text[:i] + text[j+len("</think>"):]
-		}
-	}
-
-	return strings.TrimSpace(text), nil
+// chatLLM sends a system+user pair to the configured backend. Both backends
+// map this onto their own chat shape, so the market summaries read the same
+// whichever one is serving.
+func (p *MarketPlugin) chatLLM(systemPrompt, userPrompt string) (string, error) {
+	return llmGenerate(context.Background(), llm.Request{
+		System: systemPrompt,
+		Prompt: userPrompt,
+	})
 }
 
 // ── DB Helpers ───────────────────────────────────────────────────────────────
@@ -916,12 +877,10 @@ func (p *MarketPlugin) handleVixReport(ctx MessageContext) error {
 	}
 
 	var summary string
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	if host != "" && model != "" {
+	if llmConfigured() {
 		prompt := fmt.Sprintf("VIX (fear index) data over %d days (%s to %s):\n%s\n\nDescribe the fear/greed trajectory in 2-3 sentences. Be sardonic but accurate.",
 			len(entries), entries[0].Date, entries[len(entries)-1].Date, strings.Join(prices, ", "))
-		summary, _ = p.callOllamaChat(host, model, marketSystemPrompt, prompt)
+		summary, _ = p.chatLLM(marketSystemPrompt, prompt)
 	}
 
 	var sb strings.Builder
@@ -1020,12 +979,10 @@ func (p *MarketPlugin) handleCompare(ctx MessageContext, args []string) error {
 	for _, e := range entries {
 		prices = append(prices, fmt.Sprintf("%.2f", e.Price))
 	}
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	if host != "" && model != "" {
+	if llmConfigured() {
 		prompt := fmt.Sprintf("%s over %d days (%s to %s):\n%s\n\nDescribe the trend in 2-3 sentences. Be sardonic but accurate.",
 			idx.DisplayName, len(entries), entries[0].Date, entries[len(entries)-1].Date, strings.Join(prices, ", "))
-		if summary, err := p.callOllamaChat(host, model, marketSystemPrompt, prompt); err == nil && summary != "" {
+		if summary, err := p.chatLLM(marketSystemPrompt, prompt); err == nil && summary != "" {
 			sb.WriteString(summary)
 			sb.WriteString("\n\n")
 		}

internal/plugin/pete_dispatch_voice.go

@@ -1,16 +1,14 @@
 package plugin
 
 import (
-	"bytes"
+	"context"
 	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
-	"net/http"
-	"os"
 	"strings"
 	"time"
 
+	"gogobee/internal/llm"
 	"gogobee/internal/peteclient"
 )
 
@@ -39,22 +37,18 @@ const (
 	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 == "" {
+	if !llmConfigured() {
 		return "", ""
 	}
 
 	prompt := buildDispatchPrompt(f)
-	raw, err := callOllamaDispatch(dispatchHTTP, host, model, prompt)
+	raw, err := callLLMDispatch(dispatchLLMTimeout, prompt)
 	if err != nil {
 		slog.Warn("pete dispatch: LLM authoring failed, Pete will template", "guid", f.GUID, "err", err)
 		return "", ""
@@ -128,45 +122,17 @@ The event:
 %s`, names, facts.String())
 }
 
-// callOllamaDispatch posts a single non-streaming generation and returns the raw
-// completion (think-tags stripped). The client is a parameter because the two
-// callers have genuinely different patience: a dispatch is authored on a game
-// chokepoint and must not stall it, while a run summary rides a background
-// ticker and can afford to wait for a bigger model. See runSummaryHTTP.
-func callOllamaDispatch(client *http.Client, 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 := client.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
+// callLLMDispatch posts a single non-streaming generation and returns the raw
+// completion. The timeout is a parameter because the two callers have genuinely
+// different patience: a dispatch is authored on a game chokepoint and must not
+// stall it, while a run summary rides a background ticker and can afford to wait
+// for a bigger model. See runSummaryTimeout.
+func callLLMDispatch(timeout time.Duration, prompt string) (string, error) {
+	return llmGenerate(context.Background(), llm.Request{
+		Prompt:  prompt,
+		NumCtx:  4096,
+		Timeout: timeout,
+	})
 }
 
 // parseDispatch pulls {headline, lede} out of the model's completion, tolerating

internal/plugin/pete_run_summary.go

@@ -4,8 +4,6 @@ import (
 	"encoding/json"
 	"fmt"
 	"log/slog"
-	"net/http"
-	"os"
 	"sort"
 	"strings"
 	"sync/atomic"
@@ -70,8 +68,6 @@ const maxRunSummary = 1200
 // for load-then-generate, and a timeout here really does mean the box is down.
 const runSummaryTimeout = 5 * time.Minute
 
-var runSummaryHTTP = &http.Client{Timeout: runSummaryTimeout}
-
 // runSummaryBusy is the whole concurrency story: at most one sweep in flight,
 // ever. The ticker starts one and moves on, so a cold model loading for minutes
 // costs the board nothing, and the ticks that fire meanwhile find the flag set
@@ -105,7 +101,7 @@ func (p *AdventurePlugin) sweepRunSummaries() {
 	if !peteclient.Enabled() || !newsEmissionOn() {
 		return
 	}
-	if os.Getenv("OLLAMA_HOST") == "" || os.Getenv("OLLAMA_MODEL") == "" {
+	if !llmConfigured() {
 		return // no model, no summary, no wasted queries asking which run needs one
 	}
 	runID := nextRunNeedingSummary()
@@ -174,8 +170,7 @@ func authorRunSummary(runID string) (summary, name string) {
 		return "", ""
 	}
 
-	raw, err := callOllamaDispatch(runSummaryHTTP, os.Getenv("OLLAMA_HOST"), os.Getenv("OLLAMA_MODEL"),
-		buildRunSummaryPrompt(name, beats))
+	raw, err := callLLMDispatch(runSummaryTimeout, buildRunSummaryPrompt(name, beats))
 	if err != nil {
 		slog.Warn("run summary: LLM authoring failed", "run", runID, "err", err)
 		return "", name

internal/plugin/tarot.go

@@ -4,7 +4,6 @@ import (
 	"fmt"
 	"log/slog"
 	"math/rand"
-	"os"
 	"strings"
 
 	"gogobee/internal/db"
@@ -116,9 +115,7 @@ func (p *TarotPlugin) handleTarot(ctx MessageContext) error {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "You've used up your readings for today. The cards need rest, even if you don't.")
 	}
 
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	ollamaModel := os.Getenv("OLLAMA_MODEL")
-	if ollamaHost == "" || ollamaModel == "" {
+	if !llmConfigured() {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "Tarot reader is on a union-mandated vacation and will return when morale improves.")
 	}
 
@@ -151,7 +148,7 @@ func (p *TarotPlugin) handleTarot(ctx MessageContext) error {
 	}
 	prompt := fmt.Sprintf("%s\n%s\n\n%s\n\nCard drawn: %s%s\nGive the reading.", tarotBasePrompt, tarotSingleSuffix, tarotFewShot, card, extraLines)
 
-	response, err := callOllama(ollamaHost, ollamaModel, prompt)
+	response, err := callLLM(prompt)
 	if err != nil {
 		slog.Error("tarot: ollama call", "err", err)
 		return p.SendReply(ctx.RoomID, ctx.EventID, "Tarot reader is on a union-mandated vacation and will return when morale improves.")
@@ -166,9 +163,7 @@ func (p *TarotPlugin) handleSpread(ctx MessageContext) error {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "You've used up your readings for today. The cards need rest, even if you don't.")
 	}
 
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	ollamaModel := os.Getenv("OLLAMA_MODEL")
-	if ollamaHost == "" || ollamaModel == "" {
+	if !llmConfigured() {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "Tarot reader is on a union-mandated vacation and will return when morale improves.")
 	}
 
@@ -202,7 +197,7 @@ func (p *TarotPlugin) handleSpread(ctx MessageContext) error {
 	prompt := fmt.Sprintf("%s\n%s\n\n%s\n\nCards drawn:\n- Past: %s\n- Present: %s\n- Future: %s%s\nGive the reading.",
 		tarotBasePrompt, tarotSpreadSuffix, tarotFewShot, cards[0], cards[1], cards[2], extraLines)
 
-	response, err := callOllama(ollamaHost, ollamaModel, prompt)
+	response, err := callLLM(prompt)
 	if err != nil {
 		slog.Error("tarot: ollama call", "err", err)
 		return p.SendReply(ctx.RoomID, ctx.EventID, "Tarot reader is on a union-mandated vacation and will return when morale improves.")

internal/plugin/vibe.go

@@ -120,9 +120,7 @@ func (p *VibePlugin) resetCooldown(roomID id.RoomID) {
 }
 
 func (p *VibePlugin) handleVibe(ctx MessageContext) error {
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	ollamaModel := os.Getenv("OLLAMA_MODEL")
-	if ollamaHost == "" || ollamaModel == "" {
+	if !llmConfigured() {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "LLM is not configured.")
 	}
 
@@ -154,7 +152,7 @@ Describe the room's current vibe:`, botName, transcript)
 		slog.Error("vibe: send thinking", "err", err)
 	}
 
-	response, err := callOllama(ollamaHost, ollamaModel, prompt)
+	response, err := callLLM(prompt)
 	if err != nil {
 		slog.Error("vibe: ollama call", "err", err)
 		p.resetCooldown(ctx.RoomID) // Don't consume cooldown on failure
@@ -165,9 +163,7 @@ Describe the room's current vibe:`, botName, transcript)
 }
 
 func (p *VibePlugin) handleTLDR(ctx MessageContext) error {
-	ollamaHost := os.Getenv("OLLAMA_HOST")
-	ollamaModel := os.Getenv("OLLAMA_MODEL")
-	if ollamaHost == "" || ollamaModel == "" {
+	if !llmConfigured() {
 		return p.SendReply(ctx.RoomID, ctx.EventID, "LLM is not configured.")
 	}
 
@@ -199,7 +195,7 @@ Summary:`, tldrBotName, transcript)
 		slog.Error("vibe: send thinking", "err", err)
 	}
 
-	response, err := callOllama(ollamaHost, ollamaModel, prompt)
+	response, err := callLLM(prompt)
 	if err != nil {
 		slog.Error("vibe: ollama call", "err", err)
 		p.resetCooldown(ctx.RoomID) // Don't consume cooldown on failure

internal/plugin/wotd.go

@@ -1,19 +1,17 @@
 package plugin
 
 import (
-	"bytes"
+	"context"
 	"database/sql"
 	"encoding/json"
 	"fmt"
-	"io"
 	"log/slog"
-	"net/http"
-	"os"
 	"strings"
 	"time"
 
 	"gogobee/internal/db"
 	"gogobee/internal/dreamclient"
+	"gogobee/internal/llm"
 
 	"maunium.net/go/mautrix"
 	"maunium.net/go/mautrix/id"
@@ -533,9 +531,7 @@ func (p *WOTDPlugin) trackUsage(ctx MessageContext) {
 // verifyUsage asks the LLM whether the word was used correctly in context.
 // Returns false if LLM is not configured or on any error.
 func (p *WOTDPlugin) verifyUsage(word, message string) bool {
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	if host == "" || model == "" {
+	if !llmConfigured() {
 		return false
 	}
 
@@ -545,58 +541,30 @@ The Word of the Day is "%s". Was this word used correctly and meaningfully in th
 
 Respond with ONLY "yes" or "no".`, message, word)
 
-	payload := map[string]interface{}{
-		"model":  model,
-		"prompt": prompt,
-		"stream": false,
-		"think":  false,
-	}
-	data, err := json.Marshal(payload)
-	if err != nil {
-		return false
-	}
-
-	apiURL := strings.TrimRight(host, "/") + "/api/generate"
-	slog.Debug("wotd: sending LLM verification request", "url", apiURL, "word", word)
-	client := &http.Client{Timeout: 30 * time.Second}
-	resp, err := client.Post(apiURL, "application/json", bytes.NewReader(data))
+	response, err := llmGenerate(context.Background(), llm.Request{
+		Prompt:  prompt,
+		Timeout: wotdLLMTimeout,
+	})
 	if err != nil {
 		slog.Error("wotd: LLM verify request failed", "err", err)
 		return false
 	}
-	defer resp.Body.Close()
 
-	body, err := io.ReadAll(resp.Body)
-	if err != nil || resp.StatusCode != http.StatusOK {
-		return false
-	}
-
-	var result struct {
-		Response string `json:"response"`
-	}
-	if json.Unmarshal(body, &result) != nil {
-		return false
-	}
-
-	response := result.Response
-	// Strip <think>...</think> blocks (Qwen 3.5 reasoning)
-	if i := strings.Index(response, "<think>"); i != -1 {
-		if j := strings.Index(response, "</think>"); j != -1 {
-			response = response[:i] + response[j+len("</think>"):]
-		}
-	}
 	answer := strings.ToLower(strings.TrimSpace(response))
 	accepted := strings.HasPrefix(answer, "yes")
 	slog.Debug("wotd: LLM verification", "word", word, "answer", answer, "accepted", accepted)
 	return accepted
 }
 
+// wotdLLMTimeout preserves the 30s cap both WOTD paths enforced with their own
+// http.Client. These run on message traffic, so they stay well under the
+// interactive default.
+const wotdLLMTimeout = 30 * time.Second
+
 // llmTranslate asks the LLM for a brief English translation of a foreign word.
 // Returns empty string on failure.
 func (p *WOTDPlugin) llmTranslate(word, lang string) string {
-	host := os.Getenv("OLLAMA_HOST")
-	model := os.Getenv("OLLAMA_MODEL")
-	if host == "" || model == "" {
+	if !llmConfigured() {
 		return ""
 	}
 
@@ -612,44 +580,15 @@ func (p *WOTDPlugin) llmTranslate(word, lang string) string {
 		`Translate the %s word "%s" into English. Reply with ONLY the English translation — one or two words, no explanation, no punctuation.`,
 		langName, word)
 
-	payload := map[string]interface{}{
-		"model":  model,
-		"prompt": prompt,
-		"stream": false,
-		"think":  false,
-	}
-	data, err := json.Marshal(payload)
-	if err != nil {
-		return ""
-	}
-
-	apiURL := strings.TrimRight(host, "/") + "/api/generate"
-	client := &http.Client{Timeout: 30 * time.Second}
-	resp, err := client.Post(apiURL, "application/json", bytes.NewReader(data))
+	response, err := llmGenerate(context.Background(), llm.Request{
+		Prompt:  prompt,
+		Timeout: wotdLLMTimeout,
+	})
 	if err != nil {
 		slog.Error("wotd: LLM translate request failed", "err", err)
 		return ""
 	}
-	defer resp.Body.Close()
 
-	body, err := io.ReadAll(resp.Body)
-	if err != nil || resp.StatusCode != http.StatusOK {
-		return ""
-	}
-
-	var result struct {
-		Response string `json:"response"`
-	}
-	if json.Unmarshal(body, &result) != nil {
-		return ""
-	}
-
-	response := result.Response
-	if i := strings.Index(response, "<think>"); i != -1 {
-		if j := strings.Index(response, "</think>"); j != -1 {
-			response = response[:i] + response[j+len("</think>"):]
-		}
-	}
 	translation := strings.TrimSpace(response)
 	if translation == "" || len(translation) > 50 {
 		return ""