docs update (wip)

This commit is contained in:
2026-03-08 01:46:34 -05:00
parent d9a06fd2fe
commit d34c35941f
14 changed files with 1213 additions and 105 deletions

View File

@@ -1,12 +1,18 @@
# Architecture
[Top](../Readme.md) | [Tools & IPC](guide_tools.md) | [MMA Orchestration](guide_mma.md) | [Simulations](guide_simulations.md)
[Top](../README.md) | [Tools & IPC](guide_tools.md) | [MMA Orchestration](guide_mma.md) | [Simulations](guide_simulations.md)
---
## Philosophy: The Decoupled State Machine
Manual Slop solves a single tension: **AI reasoning is high-latency and non-deterministic; GUI interaction must be low-latency and responsive.** The engine enforces strict decoupling between three thread domains so that multi-second LLM calls never block the render loop, and every AI-generated payload passes through a human-auditable gate before execution.
Manual Slop solves a single tension: **AI reasoning is high-latency and non-deterministic; GUI interaction must be low-latency and responsive.** The engine enforces strict decoupling between four thread domains so that multi-second LLM calls never block the render loop, and every AI-generated payload passes through a human-auditable gate before execution.
The architectural philosophy follows data-oriented design principles:
- The GUI (`gui_2.py`, `app_controller.py`) remains a pure visualization of application state
- State mutations occur only through lock-guarded queues consumed on the main render thread
- Background threads never write GUI state directly — they serialize task dicts for later consumption
- All cross-thread communication uses explicit synchronization primitives (Locks, Conditions, Events)
## Project Structure
@@ -36,17 +42,17 @@ manual_slop/
Four distinct thread domains operate concurrently:
| Domain | Created By | Purpose | Lifecycle |
|---|---|---|---|
| **Main / GUI** | `immapp.run()` | Dear ImGui retained-mode render loop; sole writer of GUI state | App lifetime |
| **Asyncio Worker** | `App.__init__` via `threading.Thread(daemon=True)` | Event queue processing, AI client calls | Daemon (dies with process) |
| **HookServer** | `api_hooks.HookServer.start()` | HTTP API on `:8999` for external automation and IPC | Daemon thread |
| **Ad-hoc** | Transient `threading.Thread` calls | Model-fetching, legacy send paths | Short-lived |
| Domain | Created By | Purpose | Lifecycle | Key Synchronization Primitives |
|---|---|---|---|---|
| **Main / GUI** | `immapp.run()` | Dear ImGui retained-mode render loop; sole writer of GUI state | App lifetime | None (consumer of queues) |
| **Asyncio Worker** | `App.__init__` via `threading.Thread(daemon=True)` | Event queue processing, AI client calls | Daemon (dies with process) | `AsyncEventQueue`, `threading.Lock` |
| **HookServer** | `api_hooks.HookServer.start()` | HTTP API on `:8999` for external automation and IPC | Daemon thread | `threading.Lock`, `threading.Event` |
| **Ad-hoc** | Transient `threading.Thread` calls | Model-fetching, legacy send paths, log pruning | Short-lived | Task-specific locks |
The asyncio worker is **not** the main thread's event loop. It runs a dedicated `asyncio.new_event_loop()` on its own daemon thread:
```python
# App.__init__:
# AppController.__init__:
self._loop = asyncio.new_event_loop()
self._loop_thread = threading.Thread(target=self._run_event_loop, daemon=True)
self._loop_thread.start()
@@ -60,6 +66,25 @@ def _run_event_loop(self) -> None:
The GUI thread uses `asyncio.run_coroutine_threadsafe(coro, self._loop)` to push work into this loop.
### Thread-Local Context Isolation
For concurrent multi-agent execution, the application uses `threading.local()` to manage per-thread context:
```python
# ai_client.py
_local_storage = threading.local()
def get_current_tier() -> Optional[str]:
"""Returns the current tier from thread-local storage."""
return getattr(_local_storage, "current_tier", None)
def set_current_tier(tier: Optional[str]) -> None:
"""Sets the current tier in thread-local storage."""
_local_storage.current_tier = tier
```
This ensures that comms log entries and tool calls are correctly tagged with their source tier even when multiple workers execute concurrently.
---
## Cross-Thread Data Structures
@@ -553,12 +578,247 @@ Every interaction is designed to be auditable:
- **CLI Call Logs**: Subprocess execution details (command, stdin, stdout, stderr, latency) to `clicalls.log` as JSON-L.
- **Performance Monitor**: Real-time FPS, Frame Time, CPU, Input Lag tracked and queryable via Hook API.
### Telemetry Data Structures
```python
# Comms log entry (JSON-L)
{
"ts": "14:32:05",
"direction": "OUT",
"kind": "tool_call",
"provider": "gemini",
"model": "gemini-2.5-flash-lite",
"payload": {
"name": "run_powershell",
"id": "call_abc123",
"script": "Get-ChildItem"
},
"source_tier": "Tier 3",
"local_ts": 1709875925.123
}
# Performance metrics (via get_metrics())
{
"fps": 60.0,
"fps_avg": 58.5,
"last_frame_time_ms": 16.67,
"frame_time_ms_avg": 17.1,
"cpu_percent": 12.5,
"cpu_percent_avg": 15.2,
"input_lag_ms": 2.3,
"input_lag_ms_avg": 3.1,
"time_render_mma_dashboard_ms": 5.2,
"time_render_mma_dashboard_ms_avg": 4.8
}
```
---
## MMA Engine Architecture
### WorkerPool: Concurrent Worker Management
The `WorkerPool` class in `multi_agent_conductor.py` manages a bounded pool of worker threads:
```python
class WorkerPool:
def __init__(self, max_workers: int = 4):
self.max_workers = max_workers
self._active: dict[str, threading.Thread] = {}
self._lock = threading.Lock()
self._semaphore = threading.Semaphore(max_workers)
def spawn(self, ticket_id: str, target: Callable, args: tuple) -> Optional[threading.Thread]:
with self._lock:
if len(self._active) >= self.max_workers:
return None
def wrapper(*a, **kw):
try:
with self._semaphore:
target(*a, **kw)
finally:
with self._lock:
self._active.pop(ticket_id, None)
t = threading.Thread(target=wrapper, args=args, daemon=True)
with self._lock:
self._active[ticket_id] = t
t.start()
return t
```
**Key behaviors**:
- **Bounded concurrency**: `max_workers` (default 4) limits parallel ticket execution
- **Semaphore gating**: Ensures no more than `max_workers` can execute simultaneously
- **Automatic cleanup**: Thread removes itself from `_active` dict on completion
- **Non-blocking spawn**: Returns `None` if pool is full, allowing the engine to defer
### ConductorEngine: Orchestration Loop
The `ConductorEngine` orchestrates ticket execution within a track:
```python
class ConductorEngine:
def __init__(self, track: Track, event_queue: Optional[SyncEventQueue] = None,
auto_queue: bool = False) -> None:
self.track = track
self.event_queue = event_queue
self.dag = TrackDAG(self.track.tickets)
self.engine = ExecutionEngine(self.dag, auto_queue=auto_queue)
self.pool = WorkerPool(max_workers=4)
self._abort_events: dict[str, threading.Event] = {}
self._pause_event = threading.Event()
self._tier_usage_lock = threading.Lock()
self.tier_usage = {
"Tier 1": {"input": 0, "output": 0, "model": "gemini-3.1-pro-preview"},
"Tier 2": {"input": 0, "output": 0, "model": "gemini-3-flash-preview"},
"Tier 3": {"input": 0, "output": 0, "model": "gemini-2.5-flash-lite"},
"Tier 4": {"input": 0, "output": 0, "model": "gemini-2.5-flash-lite"},
}
```
**Main execution loop** (`run` method):
1. **Pause check**: If `_pause_event` is set, sleep and broadcast "paused" status
2. **DAG tick**: Call `engine.tick()` to get ready tasks
3. **Completion check**: If no ready tasks and all completed, break with "done" status
4. **Wait for workers**: If tasks in-progress or pool active, sleep and continue
5. **Blockage detection**: If no ready, no in-progress, and not all done, break with "blocked" status
6. **Spawn workers**: For each ready task, spawn a worker via `pool.spawn()`
7. **Model escalation**: Workers use `models_list[min(retry_count, 2)]` for capability upgrade on retries
### Abort Event Propagation
Each ticket has an associated `threading.Event` for abort signaling:
```python
# Before spawning worker
self._abort_events[ticket.id] = threading.Event()
# Worker checks abort at three points:
# 1. Before major work
if abort_event.is_set():
ticket.status = "killed"
return "ABORTED"
# 2. Before tool execution (in clutch_callback)
if abort_event.is_set():
return False # Reject tool
# 3. After blocking send() returns
if abort_event.is_set():
ticket.status = "killed"
return "ABORTED"
```
---
## Architectural Invariants
1. **Single-writer principle**: All GUI state mutations happen on the main thread via `_process_pending_gui_tasks`. Background threads never write GUI state directly.
2. **Copy-and-clear lock pattern**: `_process_pending_gui_tasks` snapshots and clears the task list under the lock, then processes outside the lock.
3. **Context Amnesia**: Each MMA Tier 3 Worker starts with `ai_client.reset_session()`. No conversational bleed between tickets.
4. **Send serialization**: `_send_lock` ensures only one provider call is in-flight at a time across all threads.
5. **Dual-Flush persistence**: On exit, state is committed to both project-level and global-level config files.
6. **No cross-thread GUI mutation**: Background threads must push tasks to `_pending_gui_tasks` rather than calling GUI methods directly.
7. **Abort-before-execution**: Workers check abort events before major work phases, enabling clean cancellation.
8. **Bounded worker pool**: `WorkerPool` enforces `max_workers` limit to prevent resource exhaustion.
---
## Error Classification & Recovery
### ProviderError Taxonomy
The `ProviderError` class provides structured error classification:
```python
class ProviderError(Exception):
def __init__(self, kind: str, provider: str, original: Exception):
self.kind = kind # "quota" | "rate_limit" | "auth" | "balance" | "network" | "unknown"
self.provider = provider
self.original = original
def ui_message(self) -> str:
labels = {
"quota": "QUOTA EXHAUSTED",
"rate_limit": "RATE LIMITED",
"auth": "AUTH / API KEY ERROR",
"balance": "BALANCE / BILLING ERROR",
"network": "NETWORK / CONNECTION ERROR",
"unknown": "API ERROR",
}
return f"[{self.provider.upper()} {labels.get(self.kind, 'API ERROR')}]\n\n{self.original}"
```
### Error Recovery Patterns
| Error Kind | Recovery Strategy |
|---|---|
| `quota` | Display in UI, await user intervention |
| `rate_limit` | Exponential backoff (not yet implemented) |
| `auth` | Prompt for credential verification |
| `balance` | Display billing alert |
| `network` | Auto-retry with timeout |
| `unknown` | Log full traceback, display in UI |
---
## Memory Management
### History Trimming Strategies
**Gemini (40% threshold)**:
```python
if total_in > _GEMINI_MAX_INPUT_TOKENS * 0.4:
while len(hist) > 4 and total_in > _GEMINI_MAX_INPUT_TOKENS * 0.3:
# Drop oldest message pairs
hist.pop(0) # Assistant
hist.pop(0) # User
```
**Anthropic (180K limit)**:
```python
def _trim_anthropic_history(system_blocks, history):
est = _estimate_prompt_tokens(system_blocks, history)
while len(history) > 3 and est > _ANTHROPIC_MAX_PROMPT_TOKENS:
# Drop turn pairs, preserving tool_result chains
...
```
### Tool Output Budget
```python
_MAX_TOOL_OUTPUT_BYTES: int = 500_000 # 500KB cumulative
if _cumulative_tool_bytes > _MAX_TOOL_OUTPUT_BYTES:
# Inject warning, force final answer
parts.append("SYSTEM WARNING: Cumulative tool output exceeded 500KB budget.")
```
### AST Cache (file_cache.py)
```python
_ast_cache: Dict[str, Tuple[float, tree_sitter.Tree]] = {}
def get_cached_tree(self, path: Optional[str], code: str) -> tree_sitter.Tree:
mtime = p.stat().st_mtime if p.exists() else 0.0
if path in _ast_cache:
cached_mtime, tree = _ast_cache[path]
if cached_mtime == mtime:
return tree
# Parse and cache with simple LRU (max 10 entries)
if len(_ast_cache) >= 10:
del _ast_cache[next(iter(_ast_cache))]
tree = self.parse(code)
_ast_cache[path] = (mtime, tree)
return tree
```