TokenGuard

Lightweight, decorator-first task routing for Python.

---

What It Does

TokenGuard routes your functions to dedicated CPU-pinned workers automatically. Decorate a function, call it, and it dispatches to the right worker without blocking the caller. No manual thread management, minimal boilerplate, no complexity creep between threaded and async contexts.

The routing layers account for task difficulty: 'heavy' tasks start loading front-bound on the first system core while 'light' and 'medium' bind toward the second and third. You get the performance characteristics of a well-tuned thread pool without having to build or maintain one.

---

How It Works

When you decorate a function with @task_token_guard, calling it no longer executes it directly. Instead, a TaskToken is created and handed to the coordinator's admission queue. The coordinator routes the token to a pinned mailbox worker based on its weight class and current core load, executes it there, and delivers the result. The caller keeps moving immediately — no blocking, no manual thread management.

The staggered position system ensures tokens are spread across workers in a predictable, thread-safe sequence. Each core tracks its own monotonic counter and position arithmetic naturally shuffles assignments across the worker slots when worker counts change. The stride stays globally consistent and uses a valid range which is determined by the current active worker formation.

---

Installation

pip install tg-guard

---

Quick Start

Give it a try (in your own code)

import asyncio
from tokenguard import task_token_guard, OperationsCoordinator

coordinator = OperationsCoordinator()  # Links the token decorator event bus
coordinator.start() # Starts the bus

@task_token_guard(  # Threads the callsite
    operation_type='resize_image', 
    tags={'weight': 'heavy'}
)
def resize_image(path, size):
    ...

# Caller is not blocked — resize_image runs on a worker
token = resize_image('photo.jpg', (1920, 1080))

# Result available when ready
result = token.get(timeout=30.0)

# Or awaitable
results  = await asyncio.gather(*tokens, return_exceptions=True)

# Always clean up the coordinator on shutdown or after the work is completed.
coordinator.stop()

---

Configuration

TokenGuard ships with sensible defaults. Use option and tg_option at module level to override them before starting the coordinator.

from tokenguard import option, tg_option

# Example Coordinator settings (I suggest testing these for yourself.)
option.enable_convergence(False)  # default: True
option.num_executors(12)          # default: 8
option.mailbox_max(500)           # default: 100 — max tokens per worker mailbox
option.recent_executions_max(50)  # default: 100  — history buffer size

# Logging
tg_option.enable('coordinator', 'worker')
tg_option.debug('coordinator', on=True)
tg_option.silence('convergence')
tg_option.silence_all()
tg_option.enable_all()

workers_per_core can also be adjusted but convergence must be disabled. (Never exceed 4 or go under 2 for workers.)

---

Decorator Reference

@task_token_guard(operation_type, tags)

Parameter	Type	Required	Description
operation_type	str	Yes	Stable label used for routing, metrics, and admin operations
tags	dict	No	Routing and policy hints — see Tag Reference below

The decorated function returns a TaskToken instead of executing. The caller is not blocked.

---

Tag Reference

Tag	Values	Effect
weight	'heavy' 'medium' 'light'	Routes token to a specific core range. Heavy → core 1+. Medium → core 2+. Light → core 3+. Defaults to medium
storage_speed	'FAST' 'SLOW' 'MODERATE' 'INSANE'	Wraps the function with storage throttling. Mutually exclusive with process_pool
process_pool	True	Routes to ProcessPoolExecutor instead of thread pool. Args must be picklable. Falls back to thread executor if pickling fails
sticky_anchor	any str	Pins all tokens sharing this key to the same core. Useful when a group of operations must stay cache-local
hash_policy	HashPolicy.STANDARD HashPolicy.FAST HashPolicy.NONE	Controls how args are hashed for external calls routing. See Domain Hashing
digest_policy	DigestPolicy.FULL DigestPolicy.MINIMAL DigestPolicy.FAST DigestPolicy.SHORT	Determines hash length and digest complexity, overlapping hash prevention.
external_calls	list[str]	Marks this token as a lead token that will dispatch child tokens. Opens a conductor seed domain — all children route to the same core automatically

Note: HashPolicy.NONE must be declared if you attempt to decorate an un-hashable type, the system will not not operate on these otherwise.

---

Working With Tokens

Decorated functions return a TaskToken. The most common pattern is fire-and-forget:

process_token_file(path)   # dispatched, caller continues

When you need the result:

@task_token_guard(
    operation_type='process_file', 
    tags={'weight': 'medium'}
)
def process_token_file(path):
    # ... do work ...
    return  # Happens instantly, caller continues!

# Context based resolution
result = token.get(timeout=30.0)

# Await in async context as well
result = await token

# Or gather a batch
results = await asyncio.gather(token_a, token_b, token_c)

Typing

TaskToken is generic over its result type. When you annotate concretely, the type checker knows what .get() returns and can verify downstream usage.

Annotating Tokens

@task_token_guard(operation_type='compute', tags={'weight': 'medium'})
def compute(n: int) -> int:
    return n * n

token: TaskToken[int] = compute(10)
result: int = token.get(timeout=30.0)  # type checker knows this is int

The type flows from the function's return annotation into the token. If the function returns str, the token is TaskToken[str], and so on.

Type Conversion Methods

The double underscored methods (int, float, bool, str, etc.) have concrete return types regardless of what T is:

token: TaskToken[int] = compute(10)

n: int   = int(token)    # ✓ — __int__ returns int
f: float = float(token)  # ✓ — __float__ returns float
b: bool  = bool(token)   # ✓ — __bool__ returns bool
s: str   = str(token)    # ✓ — __str__ returns str

Arithmetic & Operators

Arithmetic dunders (+, -, *, /, **, bitwise ops) and comparisons return Any because the result type depends on what T actually is at runtime — the type checker cannot verify this without knowing both operands:

token: TaskToken[int] = compute(10)

result = token + 5      # Any — works at runtime, opaque to type checker
result = token * 2.0    # Any — same

# For a concrete type, unwrap first:
result: int = int(token) + 5    # ✓ — int + int = int, checker is happy
result: int = token.get() + 5   # ✓ — same

Batches

tokens: list[TaskToken[int]] = [compute(i) for i in range(10)]
results: list[int] = [t.get() for t in tokens]

# Or gathered
results = await asyncio.gather(*tokens)

Mixed-type Batches

from typing import Any
from tokenguard import TaskToken

tokens: list[TaskToken[Any]] = [
    compute(10),          # TaskToken[int]
    process_file("x"),    # TaskToken[dict]
]

---

Token Lifecycle

CREATED → WAITING → ADMITTED → EXECUTING → COMPLETED
                                         → FAILED
    ↓         ↓         ↓          
KILLED / TIMEOUT (valid from any non-terminal state)

Terminal states are permanent unless failed. An EXECUTING, KILLED or COMPLETED token cannot be re-queued unless failed and re-admitted.

---

Domain Hashing and Sticky Routing

Two mechanisms prevent concurrent access to the same data from different cores.

Sticky anchor pins a group of tokens to one core by a shared key:

@task_token_guard(
    operation_type='write_user_record',
    tags={
        'weight': 'medium',
        'sticky_anchor': 'user_records',
    }
)
def write_user_record(user_id, data):
    ...

All tokens with sticky_anchor='user_records' land on the same core until the inflight token completes. The pin releases automatically on completion.

Conductor domains handle coordinated fan-out — a lead token dispatches children and all of them pin to the same core automatically:

from tokenguard import HashPolicy, DigestPolicy
@task_token_guard(
    operation_type='orchestrate_pipeline',
    tags={
        'weight': 'medium',
        'hash_policy': HashPolicy.FAST,
        'digest_policy': DigestPolicy.FAST,
        'external_calls': ['stage_a', 'stage_b', 'stage_c'],
    }
)
def orchestrate_pipeline(data):
    stage_a(data)
    stage_b(data)
    stage_c(data)

The lead token charges a conductor seed domain. Any token emitted inside that execution inherits the domain and is routed to the same core, regardless of weight class.

Hash policy controls how args are hashed for sticky key resolution:

Value	Behavior
HashPolicy.STANDARD	Default. Uses token args directly
HashPolicy.FAST	Converts args to hashable form before hashing
HashPolicy.NONE	Skips arg-based hashing — uses key name only
DigestPolicy.FULL	Uses the entire 64 char hash length to prevent overlap
DigestPolicy.MINIMAL	SHA-256 truncated to 8 chars (32-bit space)
DigestPolicy.SHORT	SHA-256 truncated to 16 chars (64-bit space)
DigestPolicy.FAST	Uses Blake2s 8 byte digest/w 64 char hex

---

OperationsCoordinator

One instance per process. Start it before any decorated functions are called.

from tokenguard import OperationsCoordinator

coordinator = OperationsCoordinator()
coordinator.start()

# ... application main runs ...

coordinator.stop() # Cleans up the coordinator.

Convergence Engine

When enable_convergence=True, TokenGuard monitors per-core utilization and adjusts active worker counts every 5 seconds. Leave these at their defaults unless you have a specific reason to change them:

Setting	Default	Description
utilization_high	80.0	Utilization % above which workers scale up
utilization_low	15.0	Utilization % below which workers scale down
queue_wait_threshold	4.0	Queue wait time (seconds) that triggers scaling
queue_depth_factor	3	Multiplier applied to queue depth pressure

These are set the same way using option:

option.utilization_high(75.0)
option.utilization_low(25.0)
option.queue_wait_threshold(5.0)
option.queue_depth_factor(4)

---

Admin API

Pool-level controls for interfaces, dashboards, or operational tooling. No WebSocket layer is included — these are the clean primitives to build on.

# Kill individual or grouped tokens
coordinator.kill_token(token_id)
coordinator.kill_all_by_operation('resize_image')

# Pause and resume admission (tokens still accept, just queue)
coordinator.pause_admission()
coordinator.resume_admission()

# Per-operation pause / resume
coordinator.pause_operation('resize_image')
coordinator.resume_operation('resize_image')

# Drain waiting tokens
coordinator.drain_pool()
coordinator.drain_operation('resize_image')

# Observability
stats = coordinator.get_stats()
coordinator.print_guard_house_dashboard()
coordinator.dump_execution_history('history.json')
coordinator.get_affinity_report()

get_stats() returns a composite snapshot covering topology, token pool state, admission gate, worker queue, and convergence status.

---

Logging

All TokenGuard output goes through tg_print. Every channel is off by default.

from tokenguard import tg_option

tg_option.enable('coordinator', 'worker')      # turn on specific channels
tg_option.debug('coordinator', on=True)        # enable debug level for a channel
tg_option.silence('convergence')               # turn off a channel
tg_option.silence_all()                        # quiet everything
tg_option.enable_all()                         # turn everything on

Available channels: gate pool token coordinator convergence worker storage guard overflow affinity sticky conductor

Note: Do not enable convergence in a REPL. It produces continuous output on a fast poll interval.

---

TokenGate vs TokenGuard

TokenGuard is a stable, focused branch. TokenGate is the experimental surface where new subsystems are developed and tested before being considered for a branch.

Feature	TokenGate	TokenGuard
@task_token_guard decorator	✓	✓
Pinned staggered queue	✓	✓
Convergence engine	✓	✓
Storage throttle	✓	✓
Process pool routing	✓	✓
Domain hashing / sticky routing	✓	✓
Admin API (pause/resume/drain/kill)	✓	✓
WebSocket orchestration	✓	—
Allocation optimizer	✓	—
Code inspector	✓	—

---

Stability & Releases

Versioning:
0  .  1  .  0  .  0
│     │     │     └── Patch      — bug fixes, typo corrections
│     │     └──────── Minor      — small improvements, non-breaking additions  
│     └────────────── Update     — meaningful feature additions or tuning
└──────────────────── Major      — architectural changes, breaking API shifts

TokenGuard follows a slow, deliberate release cadence by design. The core routing and execution model is stable — updates here are fixes and minor improvements, not architectural experiments.

New subsystems and experimental features are developed in TokenGate first. If something proves solid there, it may eventually be branched into TokenGuard. This means you can build on TokenGuard without worrying about unexpected API changes while you're still learning the performance characteristics and boundaries of the system.

Requirements

• Python *3.12+ (Earlier versions generally work but aren't tested in depth.)
• Windows, macOS, Linux

LICENSE