actiongate 0.2.0

Deterministic, pre-execution gating for semantic actions (e.g. tool calls) in agent systems.

ActionGate

Deterministic, pre-execution gating for semantic actions (e.g. tool calls) in agent systems.

Source of Truth

The canonical source is github.com/actiongate-oss/actiongate. PyPI distribution is a convenience mirror—verify against the repo if provenance matters to you.

Install

pip install actiongate            # PyPI
pip install actiongate[redis]     # with Redis support
pip install -e .                  # from source

Vendoring encouraged. This is a small, stable primitive. Copy it into your codebase, fork it, reimplement it in Rust/Go/whatever. See SEMANTICS.md for the behavioral contract if you reimplement.

---

Quick Start

from actiongate import Engine, Gate, Policy, Blocked

engine = Engine()

@engine.guard(
    Gate("api", "search", "user:123"),
    Policy(max_calls=5, window=60)
)
def search(query: str) -> list[str]:
    return api.search(query)

try:
    results = search("hello")
except Blocked as e:
    print(f"Rate limited: {e.decision.message}")

Core Concepts

Gate

Identifies what's being rate-limited:

Gate(namespace, action, principal)

Gate("billing", "refund", "user:123")    # per-user
Gate("support", "escalate", "agent:42")  # per-agent  
Gate("api", "search", "global")          # global limit

Policy

Policy(
    max_calls=5,        # allow N calls per window
    window=60,          # rolling window (seconds)
    cooldown=2,         # min seconds between calls
    mode=Mode.HARD,     # HARD raises, SOFT returns Result
    on_store_error=StoreErrorMode.FAIL_CLOSED
)

Two Decorator Styles

@engine.guard(gate, policy)        # returns T, raises Blocked
@engine.guard_result(gate, policy) # returns Result[T], never raises

---

Scope & Non-Goals

ActionGate does:

• Pre-execution rate limiting (allow N, block N+1)
• Cooldown enforcement (min time between calls)
• Atomic check-and-reserve for correctness under concurrency
• Deterministic decisions with full explainability

ActionGate does not:

• Make LLM calls or consume tokens
• Do semantic/intent analysis
• Manage cost, budgets, or billing
• Orchestrate multi-agent workflows
• Provide authentication or authorization
• Replace circuit breakers, retries, or backpressure

See SEMANTICS.md for the formal behavioral contract.

---

Distributed: Redis Backend

import redis
from actiongate import Engine, RedisStore

client = redis.Redis(host='localhost', port=6379, decode_responses=True)
engine = Engine(store=RedisStore(client))

• Atomic via Lua script
• Keys auto-expire at max(window, cooldown) × 1.5
• Member format {timestamp}:{nonce} prevents collision

---

Observability

engine.on_decision(lambda d: logger.info(f"{d.status}: {d.gate}"))

Every decision includes: status, reason, gate, policy, calls_in_window, time_since_last. See examples.py for Prometheus/StatsD patterns.

---

Benchmarks

python -m actiongate.bench
python -m actiongate.bench --redis localhost:6379

Results from local benchmark on Apple M1, Python 3.12. Run to reproduce on your hardware.

Store	p50	p95	p99	Throughput
MemoryStore	~2μs	~3μs	~5μs	~400k ops/s
RedisStore (localhost)	~200μs	~300μs	~500μs	~4k ops/s

---

API Reference

Type	Purpose
Engine	Core gating logic
Gate	Action identity tuple
Policy	Rate limit configuration
Decision	Evaluation result with full context
Result[T]	Wrapper for guard_result
Blocked	Exception from guard
MemoryStore	Single-process backend
RedisStore	Distributed backend

Enum	Values
Mode	HARD, SOFT
StoreErrorMode	FAIL_CLOSED, FAIL_OPEN
Status	ALLOW, BLOCK
BlockReason	RATE_LIMIT, COOLDOWN, STORE_ERROR

---

License

Apache License 2.0. See LICENSE for the full text.