tomi/hush
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

add configuration profiles with per-invocation --profile flag

Browse Source 
- Add SystemPrompt field to HushConfig (empty = built-in default)
- Refactor ConfigManager: extract ApplyTomlFields, add LoadWithProfile(),
  ListProfiles(), GetProfilePath(), EnsureProfilesDirExists(); remove
  HUSH_PROFILE env-var logic (profiles are now resolved by the CLI)
- Extend socket protocol: action commands (START/STOP/TOGGLE/ABORT) now
  carry a [4-byte LE length][optional HushConfig JSON] payload so the CLI
  can pass a per-invocation config override without restarting the daemon
- Add GENERATE_PROFILE (cmd 7) socket command: CLI sends a description,
  daemon calls the LLM and returns a generated system prompt
- Orchestrator: StopAndProcessAsync accepts optional HushConfig override;
  ProcessWithLlmAsync uses proper system/user chat roles and respects
  config.SystemPrompt; add GenerateProfilePromptAsync
- Split CompleteTextAsync signature to (systemPrompt, userMessage, model)
  across ITextStreamingProvider, GroqProvider, FireworksProvider
- Add --profile/-p flag to hush toggle and hush stop
- Add hush profiles subcommand: list, get, new (manual or AI-generated), edit
This commit is contained in:
Tomi Eckert
2026-03-23 00:38:29 +01:00
parent 70e784a1cc
commit eb0619dea2
14 changed files with 659 additions and 372 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Hush.Daemon/src/Orchestrator.cs
+52 -21
View File
@@ -18,6 +18,8 @@ public class Orchestrator
private bool _isRecording;
private readonly Lock _lock = new();
public Orchestrator(ConfigManager configManager)
{
_configManager = configManager;
@@ -61,7 +63,7 @@ public class Orchestrator
return _recorder.StartRecording(_recordingPath);
}
public async Task StopAndProcessAsync()
public async Task StopAndProcessAsync(HushConfig? overrideConfig = null)
{
string? recordingPath;
DateTime? recordingStartTime;
@@ -86,7 +88,7 @@ public class Orchestrator
try
{
var config = _configManager.Load();
var config = overrideConfig ?? _configManager.Load();
var recordingDuration = recordingStartTime.HasValue
? DateTime.UtcNow - recordingStartTime.Value
@@ -101,7 +103,7 @@ public class Orchestrator
var transcription = await TranscribeAsync(recordingPath, config);
var processedText = await ProcessWithLlmAsync(transcription, config);
await TypeAsync(processedText, config);
File.Delete(recordingPath);
@@ -138,32 +140,61 @@ public class Orchestrator
private async Task<string> TranscribeAsync(string path, HushConfig config)
{
var provider = GetAudioToTextProvider(config);
await using var stream = File.OpenRead(path);
return await provider.TranscribeAsync(stream, config.WhisperModel);
return await provider.TranscribeAsync(
stream,
config.WhisperModel,
language: string.IsNullOrEmpty(config.WhisperLanguage) ? null : config.WhisperLanguage);
}
private const string DefaultSystemPrompt =
"""
You are a transcription post-processor. Your task is to clean up raw speech-to-text output and return polished, ready-to-type text.
Rules:
- Detect the language of the transcription and process it entirely in that language — do not translate
- Fix grammar, spelling, and punctuation errors introduced by the speech recognizer, following the conventions of the detected language
- Capitalize sentences and proper nouns appropriately for the detected language
- Remove filler words and false starts appropriate to the detected language (e.g. "um", "uh", "like" in English; "euh", "bah" in French; "äh", "ähm" in German; "eh", "tipo" in Spanish/Italian)
- Preserve the speaker's original intent, vocabulary choices, and tone
- Do not add, remove, or reinterpret content beyond what was said
- Do not include any explanation, preamble, or metadata — output only the corrected text
- If the input is empty or unintelligible, return an empty string
""";
private async Task<string> ProcessWithLlmAsync(string text, HushConfig config)
{
var provider = GetTextProvider(config);
var prompt = $"""
You are a transcription post-processor. Your task is to clean up raw speech-to-text output and return polished, ready-to-type text.
var systemPrompt = string.IsNullOrWhiteSpace(config.SystemPrompt)
? DefaultSystemPrompt
: config.SystemPrompt;
Rules:
- Detect the language of the transcription and process it entirely in that language — do not translate
- Fix grammar, spelling, and punctuation errors introduced by the speech recognizer, following the conventions of the detected language
- Capitalize sentences and proper nouns appropriately for the detected language
- Remove filler words and false starts appropriate to the detected language (e.g. "um", "uh", "like" in English; "euh", "bah" in French; "äh", "ähm" in German; "eh", "tipo" in Spanish/Italian)
- Preserve the speaker's original intent, vocabulary choices, and tone
- Do not add, remove, or reinterpret content beyond what was said
- Do not include any explanation, preamble, or metadata — output only the corrected text
- If the input is empty or unintelligible, return an empty string
return await provider.CompleteTextAsync(systemPrompt, text, config.LlmModel);
}
Raw transcription: {text}
""";
return await provider.CompleteTextAsync(prompt, config.LlmModel);
public async Task<string> GenerateProfilePromptAsync(string description)
{
var config = _configManager.Load();
var provider = GetTextProvider(config);
const string systemPrompt =
"""
You are a configuration assistant for Hush, a Linux speech-to-text post-processor.
Hush records the user's voice, transcribes it with Whisper, then passes the transcription
to an LLM using a system prompt you will write.
Given the user's description of what they want the profile to do, write a precise, concise
system prompt that instructs the LLM how to transform the raw transcription.
Rules:
- Output only the system prompt text, nothing else
- Do not include meta-commentary, labels, or markdown formatting
- The prompt must be self-contained and unambiguous
- Always end with an instruction to output only the final result with no explanation
""";
return await provider.CompleteTextAsync(systemPrompt, description, config.LlmModel);
}
private async Task TypeAsync(string text, HushConfig config)