v0.3.0

• The planner is written in Rust. Searches run on background threads, off the main thread; agents, goals and actions are still written in GDScript. The addon now needs its native library, so a release ships GdPlanningAI-X.Y.Z.zip with Linux, Windows and macOS (universal) binaries in addons/GdPlanningAI/bin. GitHub's own source archives hold Git LFS pointers instead of binaries, so install from the zip. docs/ALGORITHM.md explains how a plan is found.
• macOS support. The universal macOS library is new; before this release no macOS binary was shipped.
• Library version check. The Rust library reports its version (GdPAIPlanScheduler.get_engine_version), and the addon warns at startup when it differs from plugin.cfg, which means a stale or mismatched binary.
• Planning is cheaper and steadier. Verified steps are reused across branches that share a chain prefix, branch snapshots are shared instead of cloned, each planning cycle has a time slice (time_slice_ms), and finished engines are freed off the main thread. A failed plan backs off before the agent plans again (failure_backoff), and the backoff clears on success.
• Plans are checked more carefully. Verification re-checks custom action preconditions and the goal after the last action, an action that only serves the goal is placed at the end of the chain, plans with open requirements are never chosen as the best plan, and goals are processed by priority. Provisions and bindings chain through the plan, so plan costs use the binding that execution uses, and the example wander action plans as a step instead of a simulated position.
• Debugging. GdPAIPlanScheduler.get_debug_tree(agent) prints the agent's last planning search, including which precondition failed verification. Log verbosity is one level in plugin.cfg (log_level, 0 to 3), and each plan logs a one-line RESULT summary. The interactive debugger tab from 0.2.0 is gone and has not been brought back.
• Breaking: custom preconditions are created inline. Build one with Precondition.custom(func(blackboard, world_state): ...) directly in an append() call, instead of making an object and setting eval_func first. Existing code that sets up preconditions the old way needs updating.
• New 3D examples and a rewritten tutorial. campfire_3d, campfire_large_3d (agents tend fires in pairs, with navmesh carving around campfires and wood piles) and agent_showcase_3d, alongside the 2D scenes. The docs are rewritten: the authoring guide with the timing contract of every GDScript hook, the examples as a tutorial with a reading order, and the planning algorithm.
• Release archive. Each release attaches GdPlanningAI-X.Y.Z.zip, the addon with its binaries for all three platforms.

For contributors

• CI. Every PR runs the pre-commit hooks, Rust format, clippy and tests, and the GUT suite on Linux, Windows and macOS against binaries built in the job. One ci check gates merging into main.
• Releases are one command. scripts/release.sh X.Y.Z bumps plugin.cfg, Cargo.toml and Cargo.lock, opens and merges the release PR, tags it, and the tag builds and publishes the release.
• Quiet and verbose tests. make test, test-rust and test-godot keep their short output; CI runs test-rust-verbose and test-godot-verbose for full logs.
• Rust toolchain pinned in addons/GdPlanningAI/rust/rust-toolchain.toml, and the Godot version in global.json.