tokencircuit 8.0.0

Detect infinite loops in LLM agentic workflows and interrupt them

TokenCircuit

The Pre-Model Intervention Engine for Autonomous Agents.

---

Most guardrails are blunt instruments. They wait for your agent to burn $50 in API credits, hit a hard recursion_limit, and crash—wiping the state and returning an error to your user.

TokenCircuit is a surgical, pre-model intervention layer. Using zero-dependency semantic shingling, it detects paraphrased loops before the next LLM call. Instead of just killing the run, it uses a Progressive Intervention Protocol (Nudge → Override → Hard Stop) to safely rewrite the transcript and force the agent to pivot strategies.

< 20µs overhead per turn. Zero network calls. 100% local.

Key Features

• Progressive Intervention Protocol: Escalates from a Nudge (soft coaching injection) to an Override (surgical transcript compaction) before a Hard Stop (clean termination with state preserved).
• Zero-Dependency Semantic Loop Detection: Shingle-based fingerprinting catches paraphrased loops without embedding models, tokenizers, network calls, or external APIs.
• Atomic Transcript Surgery: Removes orphaned tool-call transactions to prevent LLM API validation errors (400 Bad Request) before they happen.
• Zero-Trust Privacy: Every detection runs in your process. No telemetry, no prompts leave your RAM.

Quick Start

TokenCircuit natively integrates with LangGraph v1.0.8+ through official extension points—no monkey-patching required.

from langgraph.prebuilt import ToolNode, create_react_agent
from tokencircuit import InterventionEngine
from tokencircuit.adapters.langgraph import tc_pre_model_hook, tc_wrap_tool_call

engine = InterventionEngine()
safe_tools = ToolNode(tools, wrap_tool_call=tc_wrap_tool_call(engine.get_thread_ledger))
agent = create_react_agent(
    model,
    tools=safe_tools,
    pre_model_hook=lambda s: tc_pre_model_hook(s, engine=engine),
)

The hook intercepts the agent before each LLM call, runs the intervention pipeline (~20µs), and returns ephemeral message mutations. No graph state is modified—the original checkpoint remains clean.

Manual graph with named hooks

from tokencircuit import InterventionConfig, InterventionEngine
from tokencircuit.adapters.langgraph import tc_pre_model_hook

engine = InterventionEngine(config=InterventionConfig(...))
builder.add_node(
    "agent", call_model,
    pre_model_hook=lambda s: tc_pre_model_hook(s, engine=engine, node_name="agent"),
)

Installation

pip install tokencircuit                    # Core engine
pip install "tokencircuit[langgraph]"       # + LangGraph adapter
pip install "tokencircuit[crewai]"          # + CrewAI adapter
pip install "tokencircuit[cli]"             # + TUI Dashboard

Performance

Benchmarked on the full intervention pipeline:

Scenario	Latency
decide() hot path (PASS)	~1.4 µs
decide() with NUDGE	~3.5 µs
process() full pipeline	~20 µs
process() with tool calls	~33 µs

Zero external embedding dependencies. All detection is local shingle-based fingerprinting.

Production Audit & Mathematical Guarantees

TokenCircuit was built for zero-trust, high-throughput enterprise environments. It undergoes ruthless structural auditing to ensure it never crashes an agent or corrupts state.

• Zero PII Telemetry: Telemetry is strictly isolated to integers and enum metadata. A mathematical proof in the payload compiler ensures that raw user prompts, tool arguments, and tool results are structurally excluded and cannot leak.
• Bounded Algorithm Complexity: The semantic loop detector uses stdlib tokenization and guarantees a worst-case shingle comparison complexity of O(N). A hard cap on the sliding window prevents CPU algorithmic blowups even on massive multi-turn infinite loops.
• Atomic Transcript Surgery: TokenCircuit adheres strictly to LangGraph v1.0.8+ API boundaries. Stray, duplicate, or dangling tool calls resulting from an intervention are surgically stripped from the LLM prompt, definitively eliminating OpenAI/Anthropic 400 Bad Request API errors.
• Async & Event-Loop Safety: All metrics fire asynchronously in daemon threads. Core pre-model hooks and interceptors run perfectly synchronously, ensuring they never starve or block the primary application event loop.

Supported Frameworks

• LangGraph: Native pre_model_hook and ToolNode(wrap_tool_call=...) integration (v1.0.8+).
• CrewAI: Execution hook support (crewai>=0.60) for proactive intervention.

---

Built for the 2026 Agentic Economy.