CI overview

build-and-test.yml runs a plan job that reads the event, changed paths, and labels and emits a run_* flag per bucket; every other job just gates on those flags. Nothing else inspects paths/labels/events.

Workflow

• Open a PR and push freely. Each push runs only the bucket(s) whose paths it touched — cheap, fast feedback. (No draft/non-draft distinction.)
• Per-push CI is intentionally partial — it's not complete or thorough, so green on a PR push does not mean the whole project passes. For the full matrix, add the full-ci label or comment /full-ci.
• Merge via the merge queue. It runs full CI against the merged tree — that's the real gate, so nothing lands on main unvalidated even though PR runs are lean. A push to main afterwards only deploys docs.

Buckets

bucket	what runs	path trigger	always on
lint	rustfmt + clippy	any	every event
regen	uniffi/flutter bindings regen-drift	core/, uniffi/, grammar/, Cargo.*, */generated/, binding config	tag, merge queue
rust_core	nix flake check	core/	tag, merge queue
python	wheels + pytest + pip-install + multimodal (+ always-on static checks: ruff/ty/stubs)	python/	tag, merge queue
python_models	6-model tool-calling matrix (linux wheel only)	core/	tag, merge queue
godot	godot build (linux/win/macos/android)	godot/	tag, merge queue
flutter	flutter build + multimodal tests + xcframework	flutter/	tag, merge queue
swift	uniffi Apple build + xcframework + tests	swift/, uniffi/	tag, merge queue
kotlin	uniffi build + JVM/Android tests	kotlin/, uniffi/	tag, merge queue
react_native	uniffi build + RN xcframework	react-native/, uniffi/	tag, merge queue
docs	docusaurus build + Cloudflare Pages deploy	—	main only
release	publish PyPI / pub.dev / npm / Maven / Swift	—	release tag

Cross-bucket: core/** → rust_core + python_models; uniffi/** → swift/kotlin/react_native. Otherwise a bucket runs only on its own path. Cargo.lock and .github/workflows/** do not auto-trigger full CI — use full-ci or the merge queue.

Triggers

trigger	runs
PR push	buckets touched by that push (+ uniffi → swift/kotlin/RN)
full-ci label / /full-ci / workflow_dispatch (full_ci)	everything
merge_group (merge queue)	everything, against the merged tree
tag nobodywho-*	everything + release
push main	docs deploy + always-on floor only
[skip ci]	nothing

Always-on floor (every event): lint + flutter doctest-drift. Concurrency: PR runs cancel on a new push; merge_group/main/tags/dispatch run to completion.

macOS granularity

cargo-build-macos is the priciest job, so matrix-gen emits an explicit {integration, target} list — only the combos a triggered consumer downloads (all release):

consumer	macOS targets
godot / flutter	macOS (x86_64+arm64) + iOS device/sim
swift	uniffi: macOS + iOS device/sim, + visionOS/watchOS (full runs only)
react-native	uniffi: iOS device/sim
kotlin	none (CI tests use the Linux lib)

visionOS/watchOS compile ONNX Runtime from source and are swift-only, so they build only on full runs (plan's is_full → build.yml's apple_extended). Partial swift PRs package the xcframework from whatever slices exist; swift-ci tests the macOS slice, so it passes. The merge queue (full) verifies all 7 platforms before merge.

required aggregator

build-and-test.yml has one required job depending on all others; it passes when each result is success or skipped, fails on failure/cancelled. Branch protection should require exactly required — matrix renames never break it.

Workflow files

plan.yml            Source of truth: paths/labels/event → run_* flags.
build-and-test.yml  Entry point: calls plan, gates children, defines `required`.
linting.yml         Always-on rustfmt + clippy.
regen-checks.yml    Bindings regen-drift checks (gated by run_regen).
build.yml           Per-platform cargo builds; matrix-gen computes integration + macOS matrix.
test.yml            nix flake check (run_rust_core) + flutter tests (run_flutter) + always-on doctest-drift.
python-ci.yml       Static checks always; wheels/tests by run_python; model matrix by run_python_models.
swift-ci.yml        Swift tests. kotlin-ci.yml  Kotlin/Android tests. (both gated upstream)
docs.yml            Docusaurus deploy (main only).
release.yml         Package publish (release tag).
full-ci-comment.yml `/full-ci` comment → label + dispatch full CI.
ai-review.yml       `/ai-review` comment. (independent, not plan-gated)
rust-install-test.yml / test-npm-publish.yml  Standalone smoke tests (dispatch only).

Adding a bucket

1. Add a run_<bucket> output in plan.yml and its trigger in the decide step.
2. In build-and-test.yml, add the call with needs: plan + if: needs.plan.outputs.run_<bucket> == 'true', and to required's needs.
3. Update the table above.