Add Trajectory.from_messages/2 for building trajectories without a live chain

[brainlid]

Problem

LangChain.Trajectory.from_chain/1 is the only constructor that reads live run data, and it requires an %LLMChain{}. But the real work it does — tool-call extraction, token aggregation, and metadata — is a pure function of a [Message.t()] plus an optional llm.

Callers that only have a bare list of messages had no entry point: a persisted agent state, a workflow, or a sliced "last turn" of a conversation. There was no way to build a trajectory from those without reconstructing a chain.

Solution

Add Trajectory.from_messages/2, exposing the message list as a first-class seam:

# From a persisted agent state, analyzing the whole conversation:
trajectory = Trajectory.from_messages(state.messages, agent.model)

# From a pre-sliced "last turn" with no llm handy:
trajectory = Trajectory.from_messages(last_turn_messages)

llm is optional and defaults to nil, which yields empty metadata via the existing extract_metadata/1 fallback clause — no helper needed to change. Metadata does not survive a JSON roundtrip anyway, so omitting it is safe when you only have a stored message list.

from_chain/1 is refactored to delegate to from_messages/2, so there is a single implementation and the two entry points cannot drift apart.

Changes

• lib/trajectory.ex — Add from_messages/2 (with optional llm); refactor from_chain/1 to delegate to it
• lib/trajectory.ex (@moduledoc + @doc) — Document the new constructor and the "exchanged messages for an operation" framing
• test/trajectory_test.exs — Add a from_messages/2 describe block: bare-list construction with and without an llm (metadata present vs %{}), token aggregation over a hand-built list, matches?/3 on a no-llm trajectory, and a parity test asserting from_chain(chain) == from_messages(chain.exchanged_messages, chain.llm)

Testing

mix precommit passes clean: 31 doctests, 1773 tests, 0 failures (143 excluded live-API tests), plus a clean Sobelow scan. The 5 new tests are in test/trajectory_test.exs and run async with no external API calls. The parity test guards the delegation refactor by asserting full struct equality between the two entry points.

🤖 Generated with Claude Code