Skip to content

Session Management#

LIT manages sessions programmatically — context is structured, queryable, and controllable, not a fragile conversation thread you have to manage manually.

What Is a Session?#

A session is a bounded conversation with an LLM: a start point, a message thread, a model configuration, and an end point. Sessions are first-class objects in LIT — they have IDs, timestamps, and are stored persistently.

Unlike most AI tools where a "conversation" is an opaque browser tab, LIT sessions are structured data you can read, analyze, and act on programmatically.

Programmatic Session Control#

from lit import Session

# Start a new session with explicit configuration
session = Session.create(
    channel="volatility-model",
    model="claude-opus-4-6",
    system_prompt="You are a deep learning research assistant...",
)

# Send a message and get a response
response = session.send("What architectures haven't we tried yet?")

# Session is automatically persisted — pick it up later
session_id = session.id

Context Management#

Sessions handle context automatically. As conversations grow, LIT manages the context window — summarizing prior history, preserving critical artifacts, and ensuring the model always has the most relevant context without blowing the token limit.

Unlike manual session management (copy-pasting summaries, starting new chats), LIT's context management is deterministic and auditable.

Safe Mode#

In safe mode, an agent pauses before any state-modifying action and requests human confirmation. The pause is logged, the confirmation is logged, and the subsequent action is logged. Safe mode is appropriate when:

  • Agents have access to production systems
  • Actions are hard to reverse (deleting data, sending external API calls)
  • You want a human in the loop for a specific workflow

Safe mode can be set at the session level, the channel level, or the agent configuration level.

Session History#

Every channel maintains a full session history. You can browse prior sessions, search across them, and see exactly how a project evolved over time. This is the "context that compounds" — a new collaborator (human or AI) can onboard by reading the channel history.