# Manual Slop

![img](./gallery/splash.png)

A high-density GUI orchestrator for local LLM-driven coding sessions. Manual Slop bridges high-latency AI reasoning with a low-latency ImGui render loop via a thread-safe asynchronous pipeline, ensuring every AI-generated payload passes through a human-auditable gate before execution.

**Design Philosophy**: Full manual control over vendor API metrics, agent capabilities, and context memory usage. High information density, tactile interactions, and explicit confirmation for destructive actions.

**Tech Stack**: Python 3.11+, Dear PyGui / ImGui Bundle, FastAPI, Uvicorn, tree-sitter
**Providers**: Gemini API, Anthropic API, DeepSeek, Gemini CLI (headless), MiniMax
**Platform**: Windows (PowerShell) — single developer, local use

![img](./gallery/python_2026-03-07_14-32-50.png)

---

## Key Features

### Multi-Provider Integration
- **Gemini SDK**: Server-side context caching with TTL management, automatic cache rebuilding at 90% TTL
- **Anthropic**: Ephemeral prompt caching with 4-breakpoint system, automatic history truncation at 180K tokens
- **DeepSeek**: Dedicated SDK for code-optimized reasoning
- **Gemini CLI**: Headless adapter with full functional parity, synchronous HITL bridge
- **MiniMax**: Alternative provider support

### 4-Tier MMA Orchestration
Hierarchical task decomposition with specialized models and strict token firewalling:
- **Tier 1 (Orchestrator)**: Product alignment, epic → tracks
- **Tier 2 (Tech Lead)**: Track → tickets (DAG), persistent context
- **Tier 3 (Worker)**: Stateless TDD implementation, context amnesia
- **Tier 4 (QA)**: Stateless error analysis, no fixes

### Strict Human-in-the-Loop (HITL)
- **Execution Clutch**: All destructive actions suspend on `threading.Condition` pending GUI approval
- **Three Dialog Types**: ConfirmDialog (scripts), MMAApprovalDialog (steps), MMASpawnApprovalDialog (workers)
- **Editable Payloads**: Review, modify, or reject any AI-generated content before execution

### 26 MCP Tools with Sandboxing
Three-layer security model: Allowlist Construction → Path Validation → Resolution Gate
- **File I/O**: read, list, search, slice, edit, tree
- **AST-Based (Python)**: skeleton, outline, definition, signature, class summary, docstring
- **Analysis**: summary, git diff, find usages, imports, syntax check, hierarchy
- **Network**: web search, URL fetch
- **Runtime**: UI performance metrics

### Parallel Tool Execution
Multiple independent tool calls within a single AI turn execute concurrently via `asyncio.gather`, significantly reducing latency.

### AST-Based Context Management
- **Skeleton View**: Signatures + docstrings, bodies replaced with `...`
- **Curated View**: Preserves `@core_logic` decorated functions and `[HOT]` comment blocks
- **Targeted View**: Extracts only specified symbols and their dependencies
- **Heuristic Summaries**: Token-efficient structural descriptions without AI calls

---

## Architecture at a Glance

Four thread domains operate concurrently: the ImGui main loop, an asyncio worker for AI calls, a `HookServer` (HTTP on `:8999`) for external automation, and transient threads for model fetching. Background threads never write GUI state directly — they serialize task dicts into lock-guarded lists that the main thread drains once per frame ([details](./docs/guide_architecture.md#the-task-pipeline-producer-consumer-synchronization)).

The **Execution Clutch** suspends the AI execution thread on a `threading.Condition` when a destructive action (PowerShell script, sub-agent spawn) is requested. The GUI renders a modal where the user can read, edit, or reject the payload. On approval, the condition is signaled and execution resumes ([details](./docs/guide_architecture.md#the-execution-clutch-human-in-the-loop)).

The **MMA (Multi-Model Agent)** system decomposes epics into tracks, tracks into DAG-ordered tickets, and executes each ticket with a stateless Tier 3 worker that starts from `ai_client.reset_session()` — no conversational bleed between tickets ([details](./docs/guide_mma.md)).

---

## Documentation

| Guide | Scope |
|---|---|
| [Readme](./docs/Readme.md) | Documentation index, GUI panel reference, configuration files, environment variables |
| [Architecture](./docs/guide_architecture.md) | Threading model, event system, AI client multi-provider architecture, HITL mechanism, comms logging |
| [Tools & IPC](./docs/guide_tools.md) | MCP Bridge 3-layer security, 26 tool inventory, Hook API endpoints, ApiHookClient reference, shell runner |
| [MMA Orchestration](./docs/guide_mma.md) | 4-tier hierarchy, Ticket/Track data structures, DAG engine, ConductorEngine, worker lifecycle, abort propagation |
| [Simulations](./docs/guide_simulations.md) | `live_gui` fixture, Puppeteer pattern, mock provider, visual verification, ASTParser / summarizer |
| [Meta-Boundary](./docs/guide_meta_boundary.md) | Application vs Meta-Tooling domains, inter-domain bridges, safety model separation |

---

## Setup

### Prerequisites

- Python 3.11+
- [`uv`](https://github.com/astral-sh/uv) for package management

### Installation

```powershell
git clone <repo>
cd manual_slop
uv sync
```

### Credentials

Configure in `credentials.toml`:

```toml
[gemini]
api_key = "YOUR_KEY"

[anthropic]
api_key = "YOUR_KEY"

[deepseek]
api_key = "YOUR_KEY"
```

### Running

```powershell
uv run sloppy.py                        # Normal mode
uv run sloppy.py --enable-test-hooks    # With Hook API on :8999
```

### Running Tests

```powershell
uv run pytest tests/ -v
```

> **Note:** See the [Structural Testing Contract](./docs/guide_simulations.md#structural-testing-contract) for rules regarding mock patching, `live_gui` standard usage, and artifact isolation (logs are generated in `tests/logs/` and `tests/artifacts/`).

---

## MMA 4-Tier Architecture

The Multi-Model Agent system uses hierarchical task decomposition with specialized models at each tier:

| Tier | Role | Model | Responsibility |
|------|------|-------|----------------|
| **Tier 1** | Orchestrator | `gemini-3.1-pro-preview` | Product alignment, epic → tracks, track initialization |
| **Tier 2** | Tech Lead | `gemini-3-flash-preview` | Track → tickets (DAG), architectural oversight, persistent context |
| **Tier 3** | Worker | `gemini-2.5-flash-lite` / `deepseek-v3` | Stateless TDD implementation per ticket, context amnesia |
| **Tier 4** | QA | `gemini-2.5-flash-lite` / `deepseek-v3` | Stateless error analysis, diagnostics only (no fixes) |

**Key Principles:**
- **Context Amnesia**: Tier 3/4 workers start with `ai_client.reset_session()` — no history bleed
- **Token Firewalling**: Each tier receives only the context it needs
- **Model Escalation**: Failed tickets automatically retry with more capable models
- **WorkerPool**: Bounded concurrency (default: 4 workers) with semaphore gating

---

## Module by Domain

### src/ — Core implementation

| File | Role |
|---|---|
| `src/gui_2.py` | Primary ImGui interface — App class, frame-sync, HITL dialogs, event system |
| `src/ai_client.py` | Multi-provider LLM abstraction (Gemini, Anthropic, DeepSeek, MiniMax) |
| `src/mcp_client.py` |       26 MCP tools with filesystem sandboxing and tool dispatch |
| `src/api_hooks.py`  |          HookServer — REST API on `127.0.0.1:8999 for external automation |
| `src/api_hook_client.py` |       Python client for the Hook API (used by tests and external tooling) |
| `src/multi_agent_conductor.py` |   ConductorEngine — Tier 2 orchestration loop with DAG execution  |
| `src/conductor_tech_lead.py`  |   Tier 2 ticket generation from track briefs |
| `src/dag_engine.py`  |       TrackDAG (dependency graph) + ExecutionEngine (tick-based state machine) |
| `src/models.py`  |       Ticket, Track, WorkerContext, Metadata, Track state |
| `src/events.py`  |           EventEmitter, AsyncEventQueue, UserRequestEvent |
| `src/project_manager.py`  |       TOML config persistence, discussion management, track state |
| `src/session_logger.py`  |       JSON-L + markdown audit trails (comms, tools, CLI, hooks) |
| `src/shell_runner.py`  |       PowerShell execution with timeout, env config, QA callback |
| `src/file_cache.py`  |       ASTParser (tree-sitter) — skeleton, curated, and targeted views |
| `src/summarize.py`  |       Heuristic file summaries (imports, classes, functions) |
| `src/outline_tool.py`  |       Hierarchical code outline via stdlib `ast` |
| `src/performance_monitor.py`  |       FPS, frame time, CPU, input lag tracking |
| `src/log_registry.py`  |       Session metadata persistence |
| `src/log_pruner.py`  |       Automated log cleanup based on age and whitelist |
| `src/paths.py`  |       Centralized path resolution with environment variable overrides |
| `src/cost_tracker.py`  |       Token cost estimation for API calls |
| `src/gemini_cli_adapter.py`  |       CLI subprocess adapter with session management |
| `src/mma_prompts.py`  |       Tier-specific system prompts for MMA orchestration |
| `src/theme_*.py` |        UI theming (dark, light modes) |

Simulation modules in `simulation/`:
| File | Role |
|---|--- |
| `simulation/sim_base.py` |       BaseSimulation class with setup/teardown lifecycle |
| `simulation/workflow_sim.py` |       WorkflowSimulator — high-level GUI automation |
| `simulation/user_agent.py` |        UserSimAgent — simulated user behavior (reading time, thinking delays) |

---

## Setup
The MCP Bridge implements a three-layer security model in `mcp_client.py`:

Every tool accessing the filesystem passes through `_resolve_and_check(path)` before any I/O.

### Layer 1: Allowlist Construction (`configure`)
Called by `ai_client` before each send cycle:
1. Resets `_allowed_paths` and `_base_dirs` to empty sets
2. Sets `_primary_base_dir` from `extra_base_dirs[0]`
3. Iterates `file_items`, resolving paths, adding to allowlist
4. Blacklist check: `history.toml`, `*_history.toml`, `config.toml`, `credentials.toml` are NEVER allowed

### Layer 2: Path Validation (`_is_allowed`)
Checks run in order:
1. **Blacklist**: `history.toml`, `*_history.toml` → hard deny
2. **Explicit allowlist**: Path in `_allowed_paths` → allow
3. **CWD fallback**: If no base dirs, allow `cwd()` subpaths
4. **Base containment**: Must be subpath of `_base_dirs`
5. **Default deny**: All other paths rejected

### Layer 3: Resolution Gate (`_resolve_and_check`)
1. Convert raw path string to `Path`
2. If not absolute, prepend `_primary_base_dir`
3. Resolve to absolute (follows symlinks)
4. Call `_is_allowed()`
5. Return `(resolved_path, "")` on success or `(None, error_message)` on failure

All paths are resolved (following symlinks) before comparison, preventing symlink-based traversal attacks.

### Security Model

The MCP Bridge implements a three-layer security model in `mcp_client.py`. Every tool accessing the filesystem passes through `_resolve_and_check(path)` before any I/O.

### Layer 1: Allowlist Construction (`configure`)
Called by `ai_client` before each send cycle:
1. Resets `_allowed_paths` and `_base_dirs` to empty sets.
2. Sets `_primary_base_dir` from `extra_base_dirs[0]` (resolved) or falls back to cwd().
3. Iterates `file_items`, resolving each path to an absolute path, adding to `_allowed_paths`; its parent directory is added to `_base_dirs`.
4. Any entries in `extra_base_dirs` that are valid directories are also added to `_base_dirs`.

### Layer 2: Path Validation (`_is_allowed`)
Checks run in this exact order:
1. **Blacklist**: `history.toml`, `*_history.toml`, `config`, `credentials` → hard deny
2. **Explicit allowlist**: Path in `_allowed_paths` → allow
7. **CWD fallback**: If no base dirs, any under `cwd()` is allowed (fail-safe for projects without explicit base dirs)
8. **Base containment**: Must be a subpath of at least one entry in `_base_dirs` (via `relative_to()`)
9. **Default deny**: All other paths rejected
All paths are resolved (following symlinks) before comparison, preventing symlink-based traversal attacks.

### Layer 3: Resolution Gate (`_resolve_and_check`)
Every tool call passes through this:
1. Convert raw path string to `Path`.
2. If not absolute, prepend `_primary_base_dir`.
3. Resolve to absolute.
4. Call `_is_allowed()`.
5. Return `(resolved_path, "")` on success, `(None, error_message)` on failure
All paths are resolved (following symlinks) before comparison, preventing symlink-based traversal attacks.

---

## Conductor SystemThe project uses a spec-driven track system in `conductor/` for structured development:

```
conductor/
├── workflow.md           # Task lifecycle, TDD protocol, phase verification
├── tech-stack.md         # Technology constraints and patterns
├── product.md            # Product vision and guidelines
├── product-guidelines.md # Code standards, UX principles
└── tracks/
    └── <track_name>_<YYYYMMDD>/
        ├── spec.md       # Track specification
        ├── plan.md       # Implementation plan with checkbox tasks
        ├── metadata.json # Track metadata
        └── state.toml    # Structured state with task list
```

**Key Concepts:**
- **Tracks**: Self-contained implementation units with spec, plan, and state
- **TDD Protocol**: Red (failing tests) → Green (pass) → Refactor
- **Phase Checkpoints**: Verification gates with git notes for audit trails
- **MMA Delegation**: Tracks are executed via the 4-tier agent hierarchy

See `conductor/workflow.md` for the full development workflow.

---

## Project Configuration

Projects are stored as `<name>.toml` files. The discussion history is split into a sibling `<name>_history.toml` to keep the main config lean.

```toml
[project]
name = "my_project"
git_dir = "./my_repo"
system_prompt = ""

[files]
base_dir = "./my_repo"
paths = ["src/**/*.py", "README.md"]

[screenshots]
base_dir = "./my_repo"
paths = []

[output]
output_dir = "./md_gen"

[gemini_cli]
binary_path = "gemini"

[agent.tools]
run_powershell = true
read_file = true
# ... 26 tool flags
```

---

## Quick Reference

### Hook API Endpoints (port 8999)

| Endpoint | Method | Description |
|----------|--------|-------------|
| `/status` | GET | Health check |
| `/api/project` | GET/POST | Project config |
| `/api/session` | GET/POST | Discussion entries |
| `/api/gui` | POST | GUI task queue |
| `/api/gui/mma_status` | GET | Full MMA state |
| `/api/gui/value/<tag>` | GET | Read GUI field |
| `/api/ask` | POST | Blocking HITL dialog |

### MCP Tool Categories

| Category | Tools |
|----------|-------|
| **File I/O** | `read_file`, `list_directory`, `search_files`, `get_tree`, `get_file_slice`, `set_file_slice`, `edit_file` |
| **AST (Python)** | `py_get_skeleton`, `py_get_code_outline`, `py_get_definition`, `py_update_definition`, `py_get_signature`, `py_set_signature`, `py_get_class_summary`, `py_get_var_declaration`, `py_set_var_declaration`, `py_get_docstring` |
| **Analysis** | `get_file_summary`, `get_git_diff`, `py_find_usages`, `py_get_imports`, `py_check_syntax`, `py_get_hierarchy` |
| **Network** | `web_search`, `fetch_url` |
| **Runtime** | `get_ui_performance` |

---