Declare `depends_on "ffmpeg"` in the formula skeleton so `brew install
aai` yields a working CLI without a separate `brew install ffmpeg`.
ffmpeg is a runtime executable that transcribe/stream shell out to for
decoding non-WAV/URL audio (not a build/link dep of any Python package),
so it belongs alongside portaudio in the runtime depends_on list.

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

…l docs

Implements the in-repo Homebrew tap (Formula/aai.rb) from the tap plan:
a Language::Python::Virtualenv formula that builds aai from the v0.1.0
GitHub tag tarball and installs its 51-package runtime closure from
pinned resource stanzas. Native deps (portaudio, ffmpeg, openssl@3,
python@3.13, rust/pkgconf for builds) are declared via depends_on so
brew install yields a fully working CLI.

The resources were generated offline from uv.lock via the new
scripts/generate_brew_resources.py (plan Appendix A) since brew isn't
available in this environment; the two Linux-only keyring deps
(jeepney, secretstorage) are scoped to on_linux and the Windows-only
pywin32-ctypes is dropped. README documents the brew tap install path.

Pending before the tap goes live: cut the v0.1.0 release/tag on main to
fill the placeholder source sha256, then brew style/audit/install/test
on macOS to validate the build (neither possible in this Linux env).

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

CI only exercises Python 3.12, so the previous 3.11 floor (and the stale
3.10-3.13 note) over-claimed support. Bump the minimum to 3.12 across the
board: pyproject requires-python + classifiers, mypy/pyright/ruff target
versions, the install.sh runtime gate and its tests, the doctor check,
the README badge/prose, and AGENTS.md. 3.13 stays supported (the Homebrew
formula pins python@3.13 and audioop-lts backfills audioop there).

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

Two changes from a quality pass:
- Regenerate Formula/aai.rb against the post-3.12 uv.lock. Bumping
  requires-python to >=3.12 dropped three Python <=3.11 backports from
  the runtime closure (backports-tarfile, importlib-metadata, zipp), but
  the formula still carried them as phantom resources. Now 48 resources,
  matching the lock.
- Consolidate the per-line out.append() calls in the resource generator
  into one resource-block string (byte-identical output).

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

Add a lint-formula job that runs brew style on Formula/aai.rb via the
SHA-pinned Homebrew/actions/setup-homebrew action, so the formula gets
real Homebrew linting on every PR instead of only a ruby -c syntax check.

Scoped to brew style (offline) deliberately: the stricter
brew audit --strict --online and a brew install/test build need the real
source sha256, which stays a placeholder until the v0.1.0 tag is cut, so
those belong in a release-time job (tap plan Task 4/6).

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

The 3.12 bump set ruff target-version=py312, which activated PEP 695
modernization lints across the tree (I'd only linted changed files):
- UP047: use def f[T](...) generics in auth/flow.py and output.py
- UP040: use the 'type' keyword for the alias in streaming/macos.py

The new brew style job also flagged two real formula offenses:
- DependencyOrder: sort the runtime depends_on (ffmpeg before openssl@3/portaudio)
- ComponentsOrder: move the on_linux block above the resource stanzas

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

CI previously tested only 3.12 while the Homebrew formula ships
python@3.13 and the classifiers claim 3.13 — an untested shipped runtime.
Matrix the check job over [3.12, 3.13] (fail-fast off) and pin UV_PYTHON
to the matrix version so check.sh's uv run/build actually exercise each
interpreter. Verified the full suite passes on both locally.

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

…aths-ci-Uhfly

# Conflicts:
#	aai_cli/output.py

Add 'synchronize' to the pull_request trigger types so a PR's checks
re-run on every new commit, instead of going stale at the run from when
the PR was opened/reopened.

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

- Scope push to main: PRs are now covered by the pull_request trigger
  (incl. synchronize), so pushing a PR commit no longer fires both a push
  and a pull_request run for the same SHA.
- Add a concurrency group that cancels superseded PR/branch runs on a new
  push, while never cancelling a main run.
- Add timeout-minutes: 15 to every job so a hung step fails fast instead
  of burning toward GitHub's 6-hour default.

Trade-off: branches without an open PR no longer auto-run CI; they run
once a PR is opened (which also fixes fork-PR coverage).

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq

…ection

The py3.12/py3.13 matrix renamed the check contexts to '... (py3.12)' and
'... (py3.13)', so the branch-protection rule requiring the un-suffixed
'lint + typecheck + tests' could never be satisfied and blocked the PR.

Add a check-result gate job that re-publishes that exact name, green only
when every matrix cell passed (if: always() + explicit result check, so a
failed/skipped/cancelled matrix can't satisfy it). Pointing branch
protection at this one stable name also makes future matrix changes safe.

https://claude.ai/code/session_01Y6P9zAX4orQUNAJDc4zLdq