Crate claudette 

Claudette — local-first AI personal secretary, powered by Ollama.

This crate bundles the agent-loop/session/compaction/permissions kernel (src/runtime/*.rs) and the Claudette secretary layer (tools, REPL, TUI, Codet sidecar, agents, Telegram bot). Single-crate — no path dependencies.

Runtime modules are mounted at the crate root via #[path = "runtime/..."] attributes so their internal use crate::session::X / use crate::compact::X paths resolve without rewriting.

Re-exports

pub use api::probe_ollama;

pub use api::resolve_ollama_url;

pub use api::OllamaApiClient;

pub use executor::SecretaryToolExecutor;

pub use memory::default_memory_path;

pub use memory::try_load_memory;

pub use memory::try_load_memory_at;

pub use memory::MAX_MEMORY_CHARS;

pub use prompt::forge_system_prompt;

pub use prompt::secretary_system_prompt;

pub use prompt::secretary_system_prompt_with_memory;

pub use run::default_session_path;

pub use run::run_forge_mission;

pub use run::run_secretary;

pub use run::run_secretary_repl;

pub use run::save_session;

pub use run::save_session_at;

pub use run::try_load_session;

pub use run::try_load_session_at;

pub use run::SessionOptions;

pub use tools::secretary_tools_json;

pub use tools::workspace_startup_diagnostics;

pub use compact::compact_session;

pub use compact::estimate_session_tokens;

pub use compact::CompactionConfig;

pub use conversation::ApiClient;

pub use conversation::ApiRequest;

pub use conversation::AssistantEvent;

pub use conversation::ConversationRuntime;

pub use conversation::RuntimeError;

pub use conversation::ToolError;

pub use conversation::ToolExecutor;

pub use conversation::TurnSummary;

pub use permissions::PermissionMode;

pub use permissions::PermissionOutcome;

pub use permissions::PermissionPolicy;

pub use permissions::PermissionPromptDecision;

pub use permissions::PermissionPrompter;

pub use permissions::PermissionRequest;

pub use prompt_runtime::ProjectContext;

pub use session::ContentBlock;

pub use session::ConversationMessage;

pub use session::MessageRole;

pub use session::Session;

pub use usage::TokenUsage;

Modules

agents

Multi-agent team for the claudette secretary.

antipatterns

Antipattern prompt overlay — injects “don’t repeat past failures” rules into the forge Coder system prompt.

api

ApiClient implementation that talks to a local Ollama instance.

brain_selector

Sprint 14 — tiered brain fallback.

briefing

Morning-briefing prompt (Phase 3 of the life-agent sprint).

clock

Time abstraction for the scheduler (AD-5 from docs/life_agent.md).

codet

Codet — the code-validator sidecar.

commands

Slash-command dispatcher for the claudette REPL.

compact

config

conversation

cto

CTO chat agent — the strategic layer above the forge fix-loop.

doctor

claudette --doctor — diagnostic probe of every external dependency.

egress

Offline mode — the enforced backing of claudette’s “air-gapped by design” claim. Until now “air-gapped” was a posture: claudette only talks to a local model by default, but nothing stopped a tool (or a prompt-injected model) from reaching the open internet. Offline mode turns that posture into a guarantee.

executor

ToolExecutor implementation that dispatches secretary tools.

firstrun

First-run remediation — when the brain probe fails at startup, offer to fix it instead of dead-ending with an error message.

forge

Forge — dormant plumbing for forge-mode (NL mission → planner → coder → PR).

google_auth

Google OAuth 2.0 for installed / desktop apps — loopback flow.

hooks

hw

Hardware detection — GPU VRAM probing + the VRAM→brain recommendation used by --doctor’s “pick a brain” section and the TUI HW tab.

image_attach

Shared helpers for turning user input into image attachments — used by both the TUI (Alt+V clipboard, bracketed-paste drop) and the REPL (drag-drop path detection on submit).

json

memory

Optional user-supplied “memory file” loader.

missions

Brownfield missions — session-scoped active workspace pointing at a cloned external repo under ~/.claudette/missions/<slug>/.

model_config

Sprint 14 — tiered brain configuration for claudette.

permissions

prompt

System prompt for the claudette secretary agent.

prompt_runtime

recall

Cross-session semantic recall — long-term memory the agent can query across sessions.

run

Top-level entry points — single-shot and REPL.

scheduler

Persistent scheduler (AD-4 from docs/life_agent.md).

secrets

File-backed secret storage with env-var override.

security_review

Opt-in security-review stage for forge-mode.

session

telegram_mode

Telegram bot mode — Claudette as a long-running Telegram bot.

test_runner

Subprocess wrappers for running code validation: syntax checks and unit tests. Sync try_wait poll loop for timeouts plus pytest output parsing.

theme

Colour + emoji theme for the claudette REPL.

tool_groups

On-demand tool group registry.

tools

Tool definitions for the secretary, in Ollama’s native tool-call schema.

transcript

Action transcript + trash — recoverability for destructive operations.

tts

Text-to-speech via Microsoft Edge TTS (free neural voices, no account needed).

tui

Ratatui TUI entry point for Claudette.

tui_events

TUI event types — messages from the worker thread to the render loop, and user inputs from the render loop back to the worker thread.

tui_executor

TuiToolExecutor — wraps SecretaryToolExecutor and fires TuiEvents before and after every tool call so the TUI can show live tool activity.

tui_worker

Worker thread for the TUI.

usage

voice

Voice transcription via whisper.cpp (local STT).