Context Isolation per BTreeRunner.run()

Decision

BTreeRunner.__enter__() captures the caller's ContextVar state with copy_context(). Each run() call executes the target inside an isolated copy of that snapshot. Mutations inside a run do not propagate back to the caller and do not accumulate across ticks.

Context

Behavior trees are tick-based: BTreeRunner.run() is called repeatedly, once per tick. A naive implementation using a persistent event loop (the original asyncio-based design) allowed ContextVar mutations inside one tick to survive into the next. This meant tick N's side effects silently affected tick N+1, making tree behavior non-deterministic and hard to test.

Why Isolated Copies

Ticks are independent. Each tick evaluates the tree from the same baseline state. Nodes that set ContextVar values communicate within a tick but not across ticks — matching standard BT semantics where each evaluation is a fresh read of the world.

Deterministic testing. A test can set up context state once and run multiple ticks without worrying about state leakage between calls. The xfail test test_runner_context_shared_across_runs documents the intentional non-sharing as a known tradeoff.

copy_context() is cheap. The copy is shallow — only the mapping of ContextVar → value is copied, not the values themselves. For typical usage (a handful of context vars) the overhead is negligible.

Trade-offs

Nodes cannot use ContextVar to accumulate state across ticks — they must use external mutable state (a class attribute, a closure variable, a database)
The snapshot is taken at __enter__, not at each run() call — mutations made between __enter__ and the first run() are included in the baseline