ctrlcode 0.4.0

Adaptive coding harness with differential fuzzing - transforms AI slop into production-ready code

ctrl+code

AI coding harness that transforms raw LLM output into production-ready code. Middleware-driven session management, pluggable backends, proactive context management, a composable skills system, and a multi-agent workflow engine.

NOTE: This is beta software, there are rough edges. Please report anything you find

Installation

pip install ctrlcode

Usage

Start the TUI (auto-launches server):

ctrlcode

Or start server separately:

ctrlcode-server

Configuration

ctrl+code follows platform conventions for config and data storage:

Platform	Config	Data	Cache
Linux	~/.config/ctrlcode/	~/.local/share/ctrlcode/	~/.cache/ctrlcode/
macOS	~/Library/Application Support/ctrlcode/	~/Library/Application Support/ctrlcode/	~/Library/Caches/ctrlcode/
Windows	%APPDATA%\ctrlcode\	%LOCALAPPDATA%\ctrlcode\	%LOCALAPPDATA%\ctrlcode\Cache\

Environment Variables

Override default directories:

• CTRLCODE_CONFIG_DIR: Config file location
• CTRLCODE_DATA_DIR: Session logs and persistent data
• CTRLCODE_CACHE_DIR: Conversation storage and temp files

Configuration File

Copy config.example.toml to your config directory as config.toml and fill in your API keys.

[providers.anthropic]
api_key = "sk-ant-..."
model = "claude-sonnet-4-6"

[context]
prune_protect = 40000

[agents]
max_delegation_depth = 2   # Max recursive sub-agent depth (0 = disable delegation)

[trajectories]
enabled = false            # Opt-in: record turns as training data
dir = ".ctrlcode/trajectories"

[permissions]
safe_commands = ["pytest", "ruff"]   # Commands that don't require approval

Agent Instructions (AGENT.md)

Customize agent behavior with AGENT.md files, loaded hierarchically:

1. Global (~/.config/ctrlcode/AGENT.md) — Your personal defaults across all projects
2. Project (<workspace>/AGENT.md) — Project-specific instructions

Example project AGENT.md:

# MyProject Instructions

## Architecture
- Frontend: React + TypeScript
- Backend: FastAPI + PostgreSQL

## Style
- Use async/await for all I/O
- Prefer functional components

Instructions are injected into the system prompt, giving the agent context about your preferences and project structure.

Skills

Skills are reusable workflows invoked with /name in the TUI, or applied automatically when the agent recognizes the intent.

Built-in Skills

Skill	Description
/commit	Create a semantic git commit (analyzes diff, stages, writes message)
/refactor	Refactor a function or module; provide the target as an argument
/test	Run test suite, analyze failures, suggest missing coverage
/review	Review current git diff for quality, bugs, and security issues
/docs	Document a class/function; provide the target as an argument
/debug-session	Structured debug workflow: reproduce → isolate → fix → verify
/plan	Write a tasks/todo.md plan with checkboxes before implementing
/pr	Create a GitHub pull request: summarize commits, push branch, open PR

Custom Skills

Create a folder with a SKILL.md file. Skills are discovered from multiple locations (later overrides earlier):

1. Built-in (src/ctrlcode/skills/builtin/) — lowest priority
2. Global (~/.config/ctrlcode/skills/)
3. Project-local (.ctrlcode/skills/) — auto-detected per workspace
4. Config override (config.skills_directory) — highest priority

Example .ctrlcode/skills/deploy/SKILL.md:

---
name: deploy
description: Deploy to staging and run smoke tests
license: project
requires_args: false
---

# Deploy

## Instructions

1. Run `make build`
2. Push to staging with `./scripts/deploy.sh staging`
3. Run `./scripts/smoke_test.sh` and report results

Context Management

ctrl+code manages the LLM context window automatically:

• Proactive monitoring: Checks usage before each turn; compacts automatically at 85% of the model's limit
• Agent-triggered compaction: The agent can call compact_conversation at phase boundaries — last 6 messages are always preserved verbatim, older messages are summarized
• History archive: All compacted messages are archived to .ctrlcode/conversation_history/<session_id>.md
• Model limits: Claude family = 200k tokens; GPT-4o = 128k tokens; fallback = 170k

Multi-Agent Workflows

For complex tasks ctrl+code can decompose work across a team of specialized agents orchestrated by WorkflowOrchestrator:

Agent	Role
planner	Analyzes intent, produces a task graph with parallel groups and checkpoints
coder	Implements tasks; reads, writes, and runs code
reviewer	Reviews completed work and requests changes if needed
executor	Runs tests, fetches URLs, validates that changes work

The workflow is: plan → execute (parallel where safe) → review → re-execute if needed → validate.

Recursive Delegation

The coder agent can spawn sub-agents via the spawn_agent tool to parallelize independent subtasks. Depth is bounded by max_delegation_depth in config (default: 2). reviewer and executor agents cannot delegate.

Session Tracing

Every spawned agent is recorded in .ctrlcode/agents/spawn_index.jsonl with its agent_id, parent_agent_id, depth, and conv_id. The parent session conversation records each spawn_agent call and result including the child agent_id, so the full delegation tree is reconstructable from the main session log.

Sub-agent conversations are stored at .ctrlcode/agents/{agent_id}/.

Architecture Overview

┌─────────────────────────────────────────────────┐
│                    TUI Client                    │
│         (Textual-based terminal UI)             │
└────────────┬────────────────────────────────────┘
             │ JSON-RPC over HTTP
┌────────────▼────────────────────────────────────┐
│              Server (aiohttp)                   │
│  ┌──────────────────────────────────────────┐   │
│  │         Session Manager                  │   │
│  │  ┌───────────────────────────────────┐   │   │
│  │  │       Middleware Stack            │   │   │
│  │  │  Memory → Workspace → Skill →     │   │   │
│  │  │  Baseline → ContextAssembly →     │   │   │
│  │  │  Permission → Progress            │   │   │
│  │  └───────────────────────────────────┘   │   │
│  └──────────────────────────────────────────┘   │
│                                                  │
│  ┌───────────────┐  ┌──────────────────────────┐ │
│  │ AgentReActLoop│  │  Tool Registry &         │ │
│  │               │  │  Executor                │ │
│  └───────────────┘  └──────────────────────────┘ │
│                                                  │
│  ┌───────────────────────────────────────────┐   │
│  │  WorkflowOrchestrator                     │   │
│  │    AgentCoordinator                       │   │
│  │      planner / coder / reviewer /         │   │
│  │      executor  (each: AgentReActLoop)     │   │
│  └───────────────────────────────────────────┘   │
│                                                  │
│  ┌───────────────┐  ┌──────────────────────────┐ │
│  │ Backend Layer │  │  Rootset DB              │ │
│  │ (Pluggable)   │  │  (.ctrlcode/rootset.db)  │ │
│  └───────────────┘  └──────────────────────────┘ │
└──────────────────────────────────────────────────┘

Key subsystems:

• Middleware Stack: Composable BaseMiddleware objects with four hook points (before_turn, pre_tool, after_tool_call, after_turn). Schema migrations for rootset.db run automatically on startup.
• Pluggable Backends: BackendProtocol abstracts all tool-layer I/O — swap FilesystemBackend for StateBackend (in-memory) in tests, or use CompositeBackend to route by path prefix
• Rootset DB: SQLite artifact store at .ctrlcode/rootset.db with versioned schema migrations (auto-applied on startup)
• Context Window Monitor: Per-session 85%-threshold monitor with LLM-powered summarization
• Multi-Agent Engine: WorkflowOrchestrator drives a planner→coder→reviewer→executor pipeline with parallel task groups, checkpoints, and configurable recursive delegation

Developer Docs

• Architecture — system design and patterns
• Features Guide — complete feature reference
• Getting Started — setup guide
• Configuration — all config options
• Security — security model
• Agent Docs — documentation map for agents working in this codebase