LLMIFY THIS MOFO

document push-to-talk keybinding patterns

README.md

@@ -83,6 +83,8 @@ systemctl --user enable --now hyprvoice.service
 bind = SUPER, R, exec, hyprvoice toggle
 ```
 
+See [Hyprland Keybindings](#hyprland-keybindings) for push-to-talk and other patterns.
+
 4. Test voice input:
 
 ```bash
@@ -91,6 +93,38 @@ hyprvoice toggle
 
 Run `hyprvoice configure` anytime for advanced settings.
 
+## Hyprland Keybindings
+
+### Simple toggle
+
+```bash
+# ~/.config/hypr/hyprland.conf
+bind = SUPER, R, exec, hyprvoice toggle
+```
+
+Each press toggles between recording and idle.
+
+### Push-to-talk (hold-to-record)
+
+Combine both bind types to get hold-to-record behavior — press to start, release to stop:
+
+```bash
+# ~/.config/hypr/hyprland.conf
+bind  = SUPER, R, exec, hyprvoice toggle   # key down → start recording
+bindr = SUPER, R, exec, hyprvoice toggle    # key up   → stop and transcribe
+```
+
+This gives a walkie-talkie feel: hold the key while speaking, release when done. The daemon receives two `toggle` commands — the first starts recording, the second stops it and triggers transcription.
+
+### `bind` vs `bindr`
+
+| Keyword | Fires on |
+|---------|----------|
+| `bind` | Key **press** (down) |
+| `bindr` | Key **release** (up) |
+
+With `bindr`, modifier keys (SUPER, CTRL, etc.) are fully released before the command executes. This can prevent modifiers from interfering with text injection.
+
 ## Commands
 
 ### Core CLI

docs/config.md

@@ -21,6 +21,7 @@ Configuration is stored in `~/.config/hyprvoice/config.toml` and changes are app
 
 ## Table of Contents
 
+- [Hyprland Keybinding Patterns](#hyprland-keybinding-patterns)
 - [Unified Provider System](#unified-provider-system)
 - [Transcription Providers](#transcription-providers)
   - [Cloud Providers](#cloud-providers)
@@ -36,6 +37,40 @@ Configuration is stored in `~/.config/hyprvoice/config.toml` and changes are app
 - [Example Configurations](#example-configurations)
 - [Legacy Configs](#legacy-configs)
 
+## Hyprland Keybinding Patterns
+
+Hyprvoice is triggered via Hyprland keybindings. The bind type you choose affects reliability and workflow.
+
+### Simple toggle
+
+```bash
+# ~/.config/hypr/hyprland.conf
+bind = SUPER, R, exec, hyprvoice toggle
+```
+
+Each press toggles between recording and idle.
+
+### Push-to-talk (hold-to-record)
+
+Pair `bind` (press) with `bindr` (release) on the same key for hold-to-record:
+
+```bash
+# ~/.config/hypr/hyprland.conf
+bind  = SUPER, R, exec, hyprvoice toggle   # key down → start recording
+bindr = SUPER, R, exec, hyprvoice toggle    # key up   → stop and transcribe
+```
+
+Hold the key while speaking, release when done. Both lines send `toggle` to the daemon — the first starts the pipeline, the second stops it. Because the stop fires on release, modifiers are clean for injection.
+
+### `bind` vs `bindr`
+
+| Keyword | Fires on |
+|---------|----------|
+| `bind` | Key **press** (down) |
+| `bindr` | Key **release** (up) |
+
+With `bindr`, modifier keys (SUPER, CTRL, etc.) are fully released before the command executes. This can prevent modifiers from interfering with text injection.
+
 ## Unified Provider System
 
 Hyprvoice uses a unified provider system where API keys are configured once and shared between transcription and LLM features:
@@ -56,12 +91,16 @@ Hyprvoice uses a unified provider system where API keys are configured once and
 
 [providers.deepgram]
   api_key = "..."              # Or set DEEPGRAM_API_KEY env var
+
+[providers.llama-swap]
+  api_key = "..."              # Or set LLAMA_SWAP_API_KEY
+  base_url = "http://llama-swap.example:8080" # Do not include /v1
 ```
 
 **API key resolution order:**
 
 1. `[providers.X]` section in config
-2. Environment variable (`OPENAI_API_KEY`, `GROQ_API_KEY`, etc.)
+2. Environment variable (`OPENAI_API_KEY`, `GROQ_API_KEY`, `LLAMA_SWAP_API_KEY`, etc.)
 
 ## Transcription Providers
 
@@ -69,6 +108,29 @@ Hyprvoice supports multiple transcription backends. See [docs/providers.md](./pr
 
 ### Cloud Providers
 
+### LlamaSwap (network-hosted OpenAI-compatible models)
+
+LlamaSwap proxies OpenAI-compatible endpoints, including `/v1/audio/transcriptions` and `/v1/chat/completions`. Configure its host once, then set the model IDs exactly as they appear in LlamaSwap's `/v1/models` response:
+
+```toml
+[providers.llama-swap]
+api_key = "your-llama-swap-api-key"
+base_url = "http://192.168.1.50:8080" # No trailing /v1
+
+[transcription]
+provider = "llama-swap"
+model = "whisper-large-v3-turbo"
+language = ""
+streaming = false
+
+[llm]
+enabled = true
+provider = "llama-swap"
+model = "your-chat-model-id"
+```
+
+The transcription and chat model IDs are deliberately not restricted by Hyprvoice: LlamaSwap selects from the models configured on your server. Hyprvoice submits transcription after recording ends, so set `streaming = false`.
+
 ### OpenAI Whisper API
 
 Cloud-based transcription using OpenAI's Whisper API:
@@ -420,7 +482,7 @@ Configurable text injection with multiple backends:
 
 ```toml
 [injection]
-backends = ["ydotool", "wtype", "clipboard"]  # Ordered fallback chain
+backends = ["clipboard-paste", "ydotool", "wtype", "clipboard"]  # Ordered fallback chain
 ydotool_timeout = "5s"
 wtype_timeout = "5s"
 clipboard_timeout = "3s"
@@ -429,6 +491,7 @@ clipboard_timeout = "3s"
 ### Injection Backends
 
 - **`ydotool`**: Uses ydotool (requires `ydotoold` daemon for ydotool v1.0.0+). Most compatible with Chromium/Electron apps.
+- **`clipboard-paste`**: Copies text to the primary selection then sends Shift+Insert through ydotool. Fast, layout-independent, and preserves the regular clipboard; recommended for Dvorak/Colemak users.
 - **`wtype`**: Uses wtype for Wayland. May have issues with some Chromium-based apps (known upstream bug).
 - **`clipboard`**: Copies text to clipboard only. Most reliable, but requires manual paste.
 

internal/bus/bus.go

@@ -8,6 +8,7 @@ import (
 	"os"
 	"path/filepath"
 	"strconv"
+	"strings"
 	"syscall"
 )
 
@@ -99,6 +100,19 @@ func (pm *pidManager) isProcessAlive(pid int) bool {
 		return false
 	}
 
+	// Verify the process is actually hyprvoice and not a recycled PID
+	cmdline, err := os.ReadFile(fmt.Sprintf("/proc/%d/cmdline", pid))
+	if err != nil {
+		log.Printf("Process %d alive but cannot read cmdline, assuming stale: %v", pid, err)
+		return false
+	}
+
+	exe := string(cmdline)
+	if len(exe) == 0 || !strings.Contains(exe, "hyprvoice") {
+		log.Printf("Process %d is alive but is not hyprvoice (cmdline: %q), stale PID file", pid, exe)
+		return false
+	}
+
 	return true
 }
 

internal/config/config_test.go

@@ -696,6 +696,31 @@ func TestConfig_ConversionMethods(t *testing.T) {
 	})
 }
 
+func TestConfig_LlamaSwap(t *testing.T) {
+	cfg := createTestConfig()
+	cfg.Transcription.Provider = "llama-swap"
+	cfg.Transcription.Model = "whisper-large-v3-turbo"
+	cfg.Providers = map[string]ProviderConfig{
+		"llama-swap": {APIKey: "test-key", BaseURL: "http://192.168.1.50:8080"},
+	}
+	cfg.LLM = LLMConfig{Enabled: true, Provider: "llama-swap", Model: "qwen3"}
+
+	if err := cfg.Validate(); err != nil {
+		t.Fatalf("Validate() error = %v", err)
+	}
+	if got := cfg.ToTranscriberConfig().BaseURL; got != "http://192.168.1.50:8080" {
+		t.Errorf("transcriber BaseURL = %q", got)
+	}
+	if got := cfg.ToLLMConfig().BaseURL; got != "http://192.168.1.50:8080" {
+		t.Errorf("LLM BaseURL = %q", got)
+	}
+
+	cfg.Providers["llama-swap"] = ProviderConfig{APIKey: "test-key", BaseURL: "http://192.168.1.50:8080/v1"}
+	if err := cfg.Validate(); err == nil || !strings.Contains(err.Error(), "omit the /v1 suffix") {
+		t.Errorf("Validate() error = %v, want /v1 validation error", err)
+	}
+}
+
 func TestValidateModelLanguageCompatibility(t *testing.T) {
 	tests := []struct {
 		name        string

internal/config/convert.go

@@ -32,6 +32,9 @@ func (c *Config) ToTranscriberConfig() transcriber.Config {
 	}
 
 	config.APIKey = c.resolveAPIKeyForProvider(c.Transcription.Provider)
+	if c.Transcription.Provider == provider.ProviderLlamaSwap {
+		config.BaseURL = c.Providers[provider.ProviderLlamaSwap].BaseURL
+	}
 
 	return config
 }
@@ -74,6 +77,9 @@ func (c *Config) ToLLMConfig() LLMAdapterConfig {
 	if c.LLM.Provider != "" {
 		config.APIKey = c.resolveAPIKeyForLLMProvider(c.LLM.Provider)
 	}
+	if c.LLM.Provider == provider.ProviderLlamaSwap {
+		config.BaseURL = c.Providers[provider.ProviderLlamaSwap].BaseURL
+	}
 
 	if c.LLM.CustomPrompt.Enabled && c.LLM.CustomPrompt.Prompt != "" {
 		config.CustomPrompt = c.LLM.CustomPrompt.Prompt

internal/config/save.go

@@ -47,6 +47,9 @@ func Save(cfg *Config) error {
 		for name, pc := range cfg.Providers {
 			sb.WriteString(fmt.Sprintf("[providers.%s]\n", name))
 			sb.WriteString(fmt.Sprintf("  api_key = %q\n", pc.APIKey))
+			if pc.BaseURL != "" {
+				sb.WriteString(fmt.Sprintf("  base_url = %q\n", pc.BaseURL))
+			}
 			sb.WriteString("\n")
 		}
 	}
@@ -221,6 +224,9 @@ keywords = []
 #   api_key = ""               # ElevenLabs API key (or set ELEVENLABS_API_KEY env var)
 # [providers.deepgram]
 #   api_key = ""               # Deepgram API key (or set DEEPGRAM_API_KEY env var)
+# [providers.llama-swap]
+#   api_key = ""               # Or set LLAMA_SWAP_API_KEY
+#   base_url = "http://llama-swap.example:8080" # No /v1 suffix
 
 # ─────────────────────────────────────────────────────────────────────────────
 # Audio Recording
@@ -241,7 +247,7 @@ keywords = []
 # ─────────────────────────────────────────────────────────────────────────────
 
 [transcription]
-  provider = "openai"          # "openai", "groq-transcription", "mistral-transcription", "elevenlabs", "whisper-cpp"
+  provider = "openai"          # Also "llama-swap" for a remote OpenAI-compatible LlamaSwap server
   model = "whisper-1"          # Model: OpenAI="whisper-1", Groq="whisper-large-v3", Mistral="voxtral-mini-latest", ElevenLabs="scribe_v1"
   language = ""                # ISO 639-1 code (e.g., en, es, de). Empty for auto-detect.
   threads = 0                  # CPU threads for local transcription (0 = auto: uses NumCPU-1)
@@ -253,7 +259,7 @@ keywords = []
 
 [llm]
   enabled = true               # Enable LLM post-processing (highly recommended)
-  provider = "openai"          # "openai" or "groq" (must have API key configured above)
+  provider = "openai"          # "openai", "groq", or "llama-swap" (must have API key configured above)
   model = "gpt-4o-mini"        # OpenAI: "gpt-4o-mini", Groq: "llama-3.3-70b-versatile"
 
 [llm.post_processing]
@@ -272,7 +278,7 @@ keywords = []
 # ─────────────────────────────────────────────────────────────────────────────
 
 [injection]
-  backends = ["ydotool", "wtype", "clipboard"]  # Ordered fallback chain (tries each until one succeeds)
+  backends = ["clipboard-paste", "ydotool", "wtype", "clipboard"]  # Ordered fallback chain (tries each until one succeeds)
   ydotool_timeout = "5s"       # Timeout for ydotool commands
   wtype_timeout = "5s"         # Timeout for wtype commands
   clipboard_timeout = "3s"     # Timeout for clipboard operations
@@ -322,12 +328,15 @@ keywords = []
 # - "groq-transcription": Groq Whisper API (very fast, models: whisper-large-v3, whisper-large-v3-turbo)
 # - "mistral-transcription": Mistral Voxtral API (excellent for European languages, model: voxtral-mini-latest)
 # - "elevenlabs": ElevenLabs Scribe API (99 languages, models: scribe_v1, scribe_v2, scribe_v2_realtime)
+# - "llama-swap": Remote OpenAI-compatible LlamaSwap (any configured transcription model)
 #
 # LLM providers (for post-processing):
 # - "openai": GPT models (gpt-4o-mini recommended for cost/quality balance)
 # - "groq": Fast inference (llama-3.3-70b-versatile recommended)
+# - "llama-swap": Any chat model configured in your LlamaSwap server
 #
 # Injection backends:
+# - "clipboard-paste": Uses primary selection + ydotool Shift+Insert; fast, layout-independent, and preserves clipboard.
 # - "ydotool": Uses ydotool (requires ydotoold daemon). Best for Chromium/Electron apps.
 # - "wtype": Uses wtype for Wayland. May have issues with some Chromium apps.
 # - "clipboard": Copies to clipboard only (most reliable, requires manual paste).

internal/config/types.go

@@ -26,6 +26,7 @@ type Config struct {
 // ProviderConfig holds API key for a provider
 type ProviderConfig struct {
 	APIKey  string `toml:"api_key"`
+	BaseURL string `toml:"base_url"` // OpenAI-compatible base URL, without /v1
 }
 
 // LLMConfig configures the LLM post-processing phase
@@ -139,4 +140,5 @@ type LLMAdapterConfig struct {
 	RemoveFillerWords bool
 	CustomPrompt      string
 	Keywords          []string
+	BaseURL           string
 }

internal/config/validate.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"fmt"
+	"net/url"
 	"strings"
 
 	"github.com/leonardotrapani/hyprvoice/internal/provider"
@@ -34,6 +35,8 @@ func envVarForProvider(registryName string) string {
 		return "ELEVENLABS_API_KEY"
 	case "deepgram":
 		return "DEEPGRAM_API_KEY"
+	case "llama-swap":
+		return "LLAMA_SWAP_API_KEY"
 	default:
 		return ""
 	}
@@ -82,12 +85,24 @@ func (c *Config) Validate() error {
 				strings.Title(registryName), registryName, envVar)
 		}
 	}
+	if registryName == provider.ProviderLlamaSwap {
+		if err := validateLlamaSwapBaseURL(c.Providers[provider.ProviderLlamaSwap].BaseURL); err != nil {
+			return err
+		}
+	}
 
 	// validate model exists
 	if c.Transcription.Model == "" {
 		return fmt.Errorf("invalid transcription.model: empty")
 	}
 
+	// LlamaSwap is an OpenAI-compatible router: its model IDs are defined by the
+	// remote server, so they cannot be validated against Hyprvoice's static registry.
+	if registryName == provider.ProviderLlamaSwap {
+		if c.Transcription.Streaming {
+			return fmt.Errorf("llama-swap transcription supports batch mode only (set transcription.streaming = false)")
+		}
+	} else {
 		// validate model exists in provider
 		_, err := provider.GetModel(registryName, c.Transcription.Model)
 		if err != nil {
@@ -104,6 +119,7 @@ func (c *Config) Validate() error {
 		if err := ValidateModelLanguageCompatibility(registryName, c.Transcription.Model, effectiveLanguage); err != nil {
 			return err
 		}
+	}
 
 	// LLM validation
 	if c.LLM.Enabled {
@@ -121,6 +137,11 @@ func (c *Config) Validate() error {
 			return fmt.Errorf("invalid llm.provider: %s (available: %s)", c.LLM.Provider, strings.Join(providers, ", "))
 		}
 
+		if c.LLM.Provider == provider.ProviderLlamaSwap {
+			if err := validateLlamaSwapBaseURL(c.Providers[provider.ProviderLlamaSwap].BaseURL); err != nil {
+				return err
+			}
+		} else {
 			// validate LLM model exists
 			llmModel, err := provider.GetModel(c.LLM.Provider, c.LLM.Model)
 			if err != nil {
@@ -136,6 +157,7 @@ func (c *Config) Validate() error {
 			if llmModel.Type != provider.LLM {
 				return fmt.Errorf("invalid llm.model: %s is not an LLM model", c.LLM.Model)
 			}
+		}
 
 		// validate LLM API key
 		if llmProvider.RequiresAPIKey() {
@@ -151,10 +173,10 @@ func (c *Config) Validate() error {
 	if len(c.Injection.Backends) == 0 {
 		return fmt.Errorf("invalid injection.backends: empty (must have at least one backend)")
 	}
-	validBackends := map[string]bool{"ydotool": true, "wtype": true, "clipboard": true}
+	validBackends := map[string]bool{"clipboard-paste": true, "ydotool": true, "wtype": true, "clipboard": true}
 	for _, backend := range c.Injection.Backends {
 		if !validBackends[backend] {
-			return fmt.Errorf("invalid injection.backends: unknown backend %q (must be ydotool, wtype, or clipboard)", backend)
+			return fmt.Errorf("invalid injection.backends: unknown backend %q (must be clipboard-paste, ydotool, wtype, or clipboard)", backend)
 		}
 	}
 	if c.Injection.YdotoolTimeout <= 0 {
@@ -175,6 +197,20 @@ func (c *Config) Validate() error {
 	return nil
 }
 
+func validateLlamaSwapBaseURL(baseURL string) error {
+	if baseURL == "" {
+		return fmt.Errorf("llama-swap base_url required in providers.llama-swap.base_url")
+	}
+	u, err := url.Parse(baseURL)
+	if err != nil || u.Scheme == "" || u.Host == "" {
+		return fmt.Errorf("invalid llama-swap base_url: %q (expected http://host:port, without /v1)", baseURL)
+	}
+	if strings.TrimRight(u.Path, "/") == "/v1" {
+		return fmt.Errorf("invalid llama-swap base_url: omit the /v1 suffix")
+	}
+	return nil
+}
+
 // ValidateModelLanguageCompatibility validates that a model supports the given language.
 // Returns error if the language is not supported, nil if supported or if langCode is empty (auto).
 func ValidateModelLanguageCompatibility(registryProvider, modelID, langCode string) error {

internal/injection/clipboard_paste.go

@@ -0,0 +1,76 @@
+package injection
+
+import (
+	"context"
+	"fmt"
+	"os"
+	"os/exec"
+	"strings"
+	"time"
+)
+
+// clipboardPasteBackend copies text to the primary selection and pastes it
+// with ydotool's layout-independent Shift+Insert key codes.
+//
+// This preserves the regular clipboard, is independent of the active keyboard
+// layout, and delivers the whole string at once.
+type clipboardPasteBackend struct {
+	clipboard *clipboardBackend
+	ydotool   *ydotoolBackend
+}
+
+func NewClipboardPasteBackend() Backend {
+	return &clipboardPasteBackend{
+		clipboard: NewClipboardBackend().(*clipboardBackend),
+		ydotool:   NewYdotoolBackend().(*ydotoolBackend),
+	}
+}
+
+func (c *clipboardPasteBackend) Name() string {
+	return "clipboard-paste"
+}
+
+func (c *clipboardPasteBackend) Available() error {
+	if err := c.clipboard.Available(); err != nil {
+		return err
+	}
+	if err := c.ydotool.Available(); err != nil {
+		return err
+	}
+	return nil
+}
+
+func (c *clipboardPasteBackend) Inject(ctx context.Context, text string, timeout time.Duration) error {
+	ctx, cancel := context.WithTimeout(ctx, timeout)
+	defer cancel()
+
+	if err := c.Available(); err != nil {
+		return err
+	}
+	// Shift+Insert pastes the primary selection, leaving the user's regular
+	// clipboard untouched.
+	primaryCopy := exec.CommandContext(ctx, "wl-copy", "--primary")
+	primaryCopy.Stdin = strings.NewReader(text)
+	if err := primaryCopy.Run(); err != nil {
+		return fmt.Errorf("copy transcription to primary selection: %w", err)
+	}
+	// wl-copy returns before every client has observed the new selection. Give
+	// the compositor a short chance to publish it; this is still perceived as
+	// an immediate paste and avoids inserting the previous clipboard contents.
+	select {
+	case <-time.After(150 * time.Millisecond):
+	case <-ctx.Done():
+		return ctx.Err()
+	}
+
+	// Linux input event codes: KEY_LEFTSHIFT=42, KEY_INSERT=110. These keys do
+	// not depend on the alphabetic layout, and Shift+Insert is widely supported.
+	cmd := exec.CommandContext(ctx, "ydotool", "key", "42:1", "110:1", "110:0", "42:0")
+	if socketPath := c.ydotool.getSocketPath(); socketPath != "" {
+		cmd.Env = append(os.Environ(), "YDOTOOL_SOCKET="+socketPath)
+	}
+	if err := cmd.Run(); err != nil {
+		return fmt.Errorf("paste primary selection with ydotool: %w", err)
+	}
+	return nil
+}

internal/injection/injection.go

@@ -12,7 +12,7 @@ type Injector interface {
 }
 
 type Config struct {
-	Backends         []string      // Ordered list: "ydotool", "wtype", "clipboard"
+	Backends         []string      // Ordered list: "clipboard-paste", "ydotool", "wtype", "clipboard"
 	YdotoolTimeout   time.Duration // Timeout for ydotool commands
 	WtypeTimeout     time.Duration // Timeout for wtype commands
 	ClipboardTimeout time.Duration // Timeout for clipboard operations
@@ -34,6 +34,8 @@ func NewInjector(config Config) Injector {
 			backends = append(backends, NewWtypeBackend())
 		case "clipboard":
 			backends = append(backends, NewClipboardBackend())
+		case "clipboard-paste":
+			backends = append(backends, NewClipboardPasteBackend())
 		default:
 			log.Printf("Injection: unknown backend %q, skipping", name)
 		}

internal/injection/injection_test.go

@@ -66,9 +66,10 @@ func TestNewInjector_IgnoresUnknownBackends(t *testing.T) {
 }
 
 func TestInjector_Inject(t *testing.T) {
-	// Skip integration tests in CI environments
-	if os.Getenv("CI") == "true" {
-		t.Skip("Skipping integration test in CI environment")
+	// This test invokes real Wayland input tools and must run in a graphical
+	// session, not merely outside CI.
+	if os.Getenv("CI") == "true" || os.Getenv("WAYLAND_DISPLAY") == "" {
+		t.Skip("Skipping integration test outside a Wayland session")
 	}
 
 	tests := []struct {
@@ -202,6 +203,13 @@ func TestClipboardBackend(t *testing.T) {
 	t.Logf("clipboard is available")
 }
 
+func TestClipboardPasteBackend(t *testing.T) {
+	backend := NewClipboardPasteBackend()
+	if backend.Name() != "clipboard-paste" {
+		t.Errorf("Name() = %s, want clipboard-paste", backend.Name())
+	}
+}
+
 // TestInjector_ClipboardMode tests clipboard-only injection
 func TestInjector_ClipboardMode(t *testing.T) {
 	config := Config{

internal/injection/ydotool.go

@@ -85,9 +85,13 @@ func (y *ydotoolBackend) Inject(ctx context.Context, text string, timeout time.D
 	if err := y.Available(); err != nil {
 		return err
 	}
+	socketPath := y.getSocketPath()
 
 	// ydotool type -- "text"
 	cmd := exec.CommandContext(ctx, "ydotool", "type", "--", text)
+	if socketPath != "" {
+		cmd.Env = append(os.Environ(), "YDOTOOL_SOCKET="+socketPath)
+	}
 	if err := cmd.Run(); err != nil {
 		return fmt.Errorf("ydotool failed: %w", err)
 	}

internal/llm/adapter_openai.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"fmt"
 	"log"
+	"strings"
 	"time"
 
 	"github.com/sashabaranov/go-openai"
@@ -17,6 +18,11 @@ type OpenAIAdapter struct {
 
 // NewOpenAIAdapter creates a new OpenAI LLM adapter
 func NewOpenAIAdapter(cfg Config) *OpenAIAdapter {
+	if cfg.BaseURL != "" {
+		clientConfig := openai.DefaultConfig(cfg.APIKey)
+		clientConfig.BaseURL = strings.TrimRight(cfg.BaseURL, "/") + "/v1"
+		return &OpenAIAdapter{client: openai.NewClientWithConfig(clientConfig), config: cfg}
+	}
 	return &OpenAIAdapter{
 		client: openai.NewClient(cfg.APIKey),
 		config: cfg,

internal/llm/llm.go

@@ -21,6 +21,7 @@ type Config struct {
 	RemoveFillerWords bool
 	CustomPrompt      string
 	Keywords          []string
+	BaseURL           string
 }
 
 // NewAdapter creates an LLM adapter based on the provider
@@ -36,6 +37,14 @@ func NewAdapter(cfg Config) (Adapter, error) {
 			return nil, fmt.Errorf("Groq API key required")
 		}
 		return NewGroqAdapter(cfg), nil
+	case "llama-swap":
+		if cfg.APIKey == "" {
+			return nil, fmt.Errorf("LlamaSwap API key required")
+		}
+		if cfg.BaseURL == "" {
+			return nil, fmt.Errorf("LlamaSwap base_url required")
+		}
+		return NewOpenAIAdapter(cfg), nil
 	default:
 		return nil, fmt.Errorf("unsupported LLM provider: %s", cfg.Provider)
 	}

internal/pipeline/pipeline.go

@@ -3,6 +3,7 @@ package pipeline
 import (
 	"context"
 	"log"
+	"strings"
 	"sync"
 	"sync/atomic"
 
@@ -316,6 +317,7 @@ func (p *pipeline) handleInjectAction(ctx context.Context, recorder recording.Re
 			RemoveFillerWords: llmCfg.RemoveFillerWords,
 			CustomPrompt:      llmCfg.CustomPrompt,
 			Keywords:          llmCfg.Keywords,
+			BaseURL:           llmCfg.BaseURL,
 		})
 		if err != nil {
 			log.Printf("Pipeline: Failed to create LLM adapter: %v, using raw transcription", err)
@@ -331,6 +333,17 @@ func (p *pipeline) handleInjectAction(ctx context.Context, recorder recording.Re
 		p.setStatus(Injecting)
 	}
 
+	// Sanitize: replace line-terminating characters with spaces to prevent
+	// unintended Enter keypresses during injection, which can submit forms mid-sentence.
+	// Covers ASCII controls (\r, \n, \v, \f), Unicode NEL (U+0085),
+	// LINE SEPARATOR (U+2028), and PARAGRAPH SEPARATOR (U+2029).
+	textToInject = strings.Map(func(r rune) rune {
+		switch r {
+		case '\r', '\n', '\v', '\f', '\u0085', '\u2028', '\u2029':
+			return ' '
+		}
+		return r
+	}, textToInject)
 	injector := p.injectorFactory(p.config.ToInjectionConfig())
 
 	if err := injector.Inject(ctx, textToInject); err != nil {

internal/provider/llama_swap.go

@@ -0,0 +1,14 @@
+package provider
+
+// LlamaSwapProvider represents LlamaSwap's authenticated OpenAI-compatible
+// proxy. The concrete model IDs are supplied in config because LlamaSwap can
+// route to any model configured by its operator.
+type LlamaSwapProvider struct{}
+
+func (p *LlamaSwapProvider) Name() string                   { return ProviderLlamaSwap }
+func (p *LlamaSwapProvider) RequiresAPIKey() bool           { return true }
+func (p *LlamaSwapProvider) ValidateAPIKey(key string) bool { return key != "" }
+func (p *LlamaSwapProvider) APIKeyURL() string              { return "" }
+func (p *LlamaSwapProvider) IsLocal() bool                  { return false }
+func (p *LlamaSwapProvider) Models() []Model                { return nil }
+func (p *LlamaSwapProvider) DefaultModel(ModelType) string  { return "" }

internal/provider/names.go

@@ -8,6 +8,7 @@ const (
 	ProviderElevenLabs = "elevenlabs"
 	ProviderDeepgram   = "deepgram"
 	ProviderWhisperCpp = "whisper-cpp"
+	ProviderLlamaSwap  = "llama-swap"
 )
 
 // Config provider names (used in config file transcription.provider)
@@ -27,6 +28,7 @@ const (
 	EnvMistralKey    = "MISTRAL_API_KEY"
 	EnvElevenLabsKey = "ELEVENLABS_API_KEY"
 	EnvDeepgramKey   = "DEEPGRAM_API_KEY"
+	EnvLlamaSwapKey  = "LLAMA_SWAP_API_KEY"
 )
 
 // Adapter type constants for transcription backends
@@ -66,6 +68,8 @@ func EnvVarForProvider(provider string) string {
 		return EnvElevenLabsKey
 	case ProviderDeepgram:
 		return EnvDeepgramKey
+	case ProviderLlamaSwap:
+		return EnvLlamaSwapKey
 	default:
 		return ""
 	}

internal/provider/provider.go

@@ -31,6 +31,7 @@ func init() {
 	Register(&ElevenLabsProvider{})
 	Register(&WhisperCppProvider{})
 	Register(&DeepgramProvider{})
+	Register(&LlamaSwapProvider{})
 }
 
 // Register adds a provider to the registry

internal/transcriber/adapter_openai.go

@@ -35,7 +35,7 @@ func NewOpenAIAdapter(endpoint *provider.EndpointConfig, apiKey, model, lang str
 	if endpoint != nil && endpoint.BaseURL != "" {
 		// use custom endpoint
 		clientConfig := openai.DefaultConfig(apiKey)
-		clientConfig.BaseURL = endpoint.BaseURL + "/v1"
+		clientConfig.BaseURL = strings.TrimRight(endpoint.BaseURL, "/") + "/v1"
 		client = openai.NewClientWithConfig(clientConfig)
 	} else {
 		// default to OpenAI

internal/transcriber/transcriber.go

@@ -33,6 +33,7 @@ type Config struct {
 	Keywords  []string
 	Threads   int    // CPU threads for local transcription (0 = auto)
 	Streaming bool   // use streaming mode if model supports it
+	BaseURL   string // OpenAI-compatible base URL, without /v1
 }
 
 // NewTranscriber creates a new transcriber based on model metadata
@@ -55,6 +56,22 @@ func NewTranscriber(config Config) (Transcriber, error) {
 		return nil, fmt.Errorf("%s API key required", cases.Title(language.English).String(registryProvider))
 	}
 
+	// llama-swap proxies arbitrary OpenAI-compatible model IDs, so models are
+	// intentionally configured by the user rather than limited to this registry.
+	if registryProvider == provider.ProviderLlamaSwap {
+		if config.Model == "" {
+			return nil, fmt.Errorf("model is required for llama-swap")
+		}
+		if config.BaseURL == "" {
+			return nil, fmt.Errorf("llama-swap base_url required")
+		}
+		if config.Streaming {
+			return nil, fmt.Errorf("llama-swap transcription currently supports batch mode only (set streaming = false)")
+		}
+		endpoint := &provider.EndpointConfig{BaseURL: config.BaseURL}
+		return NewSimpleTranscriber(config, NewOpenAIAdapter(endpoint, config.APIKey, config.Model, config.Language, config.Keywords, registryProvider)), nil
+	}
+
 	// lookup model from provider
 	model, err := provider.GetModel(registryProvider, config.Model)
 	if err != nil {

packaging/PKGBUILD

@@ -1,6 +1,6 @@
 # Maintainer: Leonardo Trapani <leo@trapani.sh>
 pkgname=hyprvoice-bin
-pkgver=v1.0.1
+pkgver=v1.0.2
 pkgrel=1
 pkgdesc="Voice-powered typing for Wayland/Hyprland"
 arch=('x86_64')
@@ -26,7 +26,7 @@ source=(
     "hyprvoice-${pkgver}::https://github.com/leonardotrapani/hyprvoice/releases/download/${pkgver}/hyprvoice-linux-x86_64"
     "hyprvoice.service"
 )
-sha256sums=('d0d1dc952fe374917e7598b36a7c228a4a5222dd1a68f4c859a46e87a1ee1613'
+sha256sums=('d055b01352b0cb4e50ac2e729e0e1baac259468559bb86a792b2f20f30bb6cc4'
             '5631c957777882870e61934176d2142fab6c553bb650ffd8b3eb74b8dce955bd')
 install=hyprvoice.install