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/llm/llm.go

@@ -0,0 +1,161 @@
+// Package llm wraps the local inference endpoint behind a backend-agnostic
+// interface. Two concrete backends — Ollama (native /api/generate) and vLLM
+// (OpenAI-compatible /v1/chat/completions) — implement Client; plugin code
+// calls the interface only and never knows which one is active.
+//
+// Deliberately not routed through internal/safehttp: that client blocks
+// RFC1918 and loopback destinations to defend against SSRF from feed-supplied
+// URLs, and the inference endpoint is precisely such a destination. The URL
+// here comes from our own config, never from user input.
+package llm
+
+import (
+	"context"
+	"os"
+	"strings"
+	"time"
+)
+
+// Request is the backend-neutral generation request. Zero-valued fields fall
+// back to backend defaults.
+type Request struct {
+	// Prompt is a single raw instruction. The vLLM backend wraps it as one
+	// user message so the model's chat template still applies; sending it to
+	// /v1/completions instead would bypass the template and degrade an
+	// instruction-tuned model badly.
+	Prompt string
+	// System is an optional system message. Empty means none, which keeps the
+	// single-message shape the majority of callers use.
+	System string
+	// NumCtx is the per-request context window. Ollama honours it directly;
+	// vLLM fixes the window server-side at launch (--max-model-len), so this
+	// is ignored there rather than silently misapplied.
+	NumCtx int
+	// MaxTokens caps the completion length. 0 means the backend default.
+	MaxTokens int
+	// Temperature is passed through when non-zero.
+	Temperature float64
+	// Timeout overrides the client's default per-request budget.
+	Timeout time.Duration
+}
+
+// Client is the single surface plugin code depends on.
+type Client interface {
+	// Generate returns the full completion in one shot. Reasoning blocks are
+	// stripped before returning (see StripThink) — every caller in this repo
+	// wants the visible answer, not the chain of thought.
+	Generate(ctx context.Context, req Request) (string, error)
+	// Model reports the configured model id, for logging and /botinfo.
+	Model() string
+	// Ping reports the model ids the backend is currently serving. Used by
+	// /botinfo for a liveness line; the two backends expose this on different
+	// paths (/api/tags vs /v1/models), which is exactly the sort of difference
+	// this interface exists to hide.
+	Ping(ctx context.Context) ([]string, error)
+	// Backend reports "ollama" or "vllm", for logging and /botinfo.
+	Backend() string
+}
+
+// Config selects and configures a backend.
+type Config struct {
+	Backend  string // "ollama" | "vllm"
+	Endpoint string
+	Model    string
+	Timeout  time.Duration
+}
+
+// DefaultTimeout matches the budget the pre-refactor callOllama used.
+const DefaultTimeout = 120 * time.Second
+
+// ConfigFromEnv reads backend settings, preferring the new LLM_* names and
+// falling back to the legacy OLLAMA_* pair so an existing deployment keeps
+// working untouched after this refactor.
+func ConfigFromEnv() Config {
+	backend := strings.ToLower(strings.TrimSpace(os.Getenv("LLM_BACKEND")))
+	if backend == "" {
+		backend = "ollama"
+	}
+
+	endpoint := firstNonEmpty(os.Getenv("LLM_ENDPOINT"), os.Getenv("OLLAMA_HOST"))
+	model := firstNonEmpty(os.Getenv("LLM_MODEL"), os.Getenv("OLLAMA_MODEL"))
+
+	timeout := DefaultTimeout
+	if d, err := time.ParseDuration(os.Getenv("LLM_TIMEOUT")); err == nil && d > 0 {
+		timeout = d
+	}
+
+	return Config{Backend: backend, Endpoint: endpoint, Model: model, Timeout: timeout}
+}
+
+// Configured reports whether enough config is present to talk to a backend.
+// Plugins check this to stay dormant rather than erroring on every invocation,
+// which is what the old `if ollamaHost == "" || ollamaModel == ""` guards did.
+func (c Config) Configured() bool {
+	return c.Endpoint != "" && c.Model != ""
+}
+
+// New builds the client for cfg.Backend. An unrecognised backend falls back to
+// Ollama, which is what every existing deployment runs.
+func New(cfg Config) Client {
+	if cfg.Timeout <= 0 {
+		cfg.Timeout = DefaultTimeout
+	}
+	base := backend{
+		endpoint: strings.TrimRight(cfg.Endpoint, "/"),
+		model:    cfg.Model,
+		timeout:  cfg.Timeout,
+	}
+	switch cfg.Backend {
+	case "vllm":
+		return &VLLMClient{base}
+	default:
+		return &OllamaClient{base}
+	}
+}
+
+// backend holds the fields shared by both concrete clients.
+type backend struct {
+	endpoint string
+	model    string
+	timeout  time.Duration
+}
+
+func (b backend) Model() string { return b.model }
+
+// timeoutFor lets a single call widen or narrow the client default. The two
+// dispatch-voice callers rely on this: a dispatch is authored on a game
+// chokepoint and must not stall it, while a run summary rides a background
+// ticker and can afford a bigger model.
+func (b backend) timeoutFor(req Request) time.Duration {
+	if req.Timeout > 0 {
+		return req.Timeout
+	}
+	return b.timeout
+}
+
+// StripThink removes a leading <think>...</think> reasoning block, which Qwen
+// models emit even when thinking is disabled by some backends. Callers that
+// parse JSON out of the completion depend on this running first.
+func StripThink(s string) string {
+	for {
+		i := strings.Index(s, "<think>")
+		if i < 0 {
+			break
+		}
+		j := strings.Index(s, "</think>")
+		if j < 0 || j < i {
+			break
+		}
+		s = s[:i] + s[j+len("</think>"):]
+	}
+	return strings.TrimSpace(s)
+}
+
+func firstNonEmpty(vals ...string) string {
+	for _, v := range vals {
+		if v = strings.TrimSpace(v); v != "" {
+			return v
+		}
+	}
+	return ""
+}