onoats 1.1.0__tar.gz

onoats-1.1.0/.githooks/pre-commit

@@ -0,0 +1,14 @@
+#!/bin/sh
+# onoats pre-commit hook. Enable once per clone with:
+#     git config core.hooksPath .githooks
+#
+# Validates that any staged dev-plan's reviewed contract section still matches
+# its review-marker hash (the same check CI runs on every PR). Catches an
+# above-marker edit that forgot to refresh the marker before it lands.
+set -e
+
+staged=$(git diff --cached --name-only --diff-filter=ACM | grep '^docs/dev_plans/.*\.md$' || true)
+if [ -n "$staged" ]; then
+    # Pass the staged paths explicitly; stdlib + git only, no uv/venv needed.
+    python3 scripts/check_review_markers.py $staged
+fi

onoats-1.1.0/.github/workflows/ci.yml

@@ -0,0 +1,44 @@
+name: CI
+
+on:
+  push:
+    branches: [main]
+  pull_request:
+
+jobs:
+  lint-and-test:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+
+      # Cheap, dependency-free gate (stdlib + git only) — fail fast before the
+      # heavy portaudio/uv install if a dev-plan's reviewed contract section was
+      # edited without refreshing its review marker.
+      - name: Dev-plan review-marker check
+        run: python3 scripts/check_review_markers.py
+
+      # pyaudio (via pipecat-ai[local]) builds a C extension against PortAudio,
+      # whose dev headers ship with macOS/Homebrew but not ubuntu-latest. This
+      # is the off-mac system dep the cross-platform baseline assumes.
+      - name: Install system deps (PortAudio for pyaudio)
+        run: sudo apt-get update && sudo apt-get install -y portaudio19-dev
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+        with:
+          python-version: "3.12"
+
+      - name: Sync dependencies
+        run: uv sync --dev
+
+      - name: Ruff format check
+        run: uv run ruff format --check src/onoats tests
+
+      - name: Ruff lint
+        run: uv run ruff check src/onoats tests
+
+      - name: Tests
+        run: uv run pytest -q
+
+      - name: CLI help smoke
+        run: uv run onoats --help

onoats-1.1.0/.github/workflows/release.yml

@@ -0,0 +1,78 @@
+name: Release
+
+# Publish to PyPI via trusted publishing (OIDC — no stored token) when a
+# version tag is pushed. The publish job is additionally gated behind the
+# `pypi` GitHub environment, and PyPI's trusted-publisher config pins
+# repo + this workflow filename + that environment, so no other workflow
+# (or fork) can mint a publishable token.
+on:
+  push:
+    tags: ["v*"]
+
+permissions:
+  contents: read
+
+jobs:
+  build:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/checkout@v5
+
+      - name: Guard — tag must match pyproject version
+        run: |
+          TAG="${GITHUB_REF_NAME#v}"
+          VERSION=$(python3 -c "import tomllib; print(tomllib.load(open('pyproject.toml','rb'))['project']['version'])")
+          if [ "$TAG" != "$VERSION" ]; then
+            echo "Tag v$TAG does not match pyproject version $VERSION" >&2
+            exit 1
+          fi
+
+      # Same baseline system dep as ci.yml — the release gate re-runs the
+      # suite so an artifact is never published from an untested tree.
+      - name: Install system deps (PortAudio for pyaudio)
+        run: sudo apt-get update && sudo apt-get install -y portaudio19-dev
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v7
+        with:
+          python-version: "3.12"
+
+      - name: Sync dependencies
+        run: uv sync --dev
+
+      - name: Tests
+        run: uv run pytest -q
+
+      - name: Build sdist + wheel
+        run: uv build
+
+      - name: Guard — no direct-URL dependencies in wheel metadata
+        run: |
+          if unzip -p dist/*.whl '*/METADATA' | grep -E '^Requires-Dist:.*@ (git\+|https?://)'; then
+            echo "Direct-URL dependency in metadata — PyPI will reject this" >&2
+            exit 1
+          fi
+
+      - uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+          if-no-files-found: error
+
+  publish:
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/p/onoats
+    permissions:
+      # OIDC token for PyPI trusted publishing — granted ONLY to this job.
+      id-token: write
+    steps:
+      - uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+      - name: Publish to PyPI
+        uses: pypa/gh-action-pypi-publish@release/v1

onoats-1.1.0/.gitignore

@@ -0,0 +1,17 @@
+# Python
+__pycache__/
+*.py[cod]
+.pytest_cache/
+.ruff_cache/
+*.egg-info/
+build/
+dist/
+
+# uv / venv
+.venv/
+
+# onoats runtime/data (XDG dirs are out-of-tree; guard local overrides)
+.conduct/
+*.db
+secrets.env
+.deep-review/

onoats-1.1.0/AGENTS.md

@@ -0,0 +1,226 @@
+# AGENTS.md — onoats maintainer & agent guide
+
+Conventions an agent (or human) needs before changing onoats. Scoped to the
+non-obvious parts; the code and `docs/` cover the rest.
+
+## Tooling
+
+- Package manager is **`uv`**. Install: `uv sync`. Run: `uv run <cmd>`.
+- Tests: `uv run pytest`. Lint/format before every push: `uv run ruff format`
+  **and** `uv run ruff check` (a PostToolUse hook formats on edit, but verify).
+- Prefer the **pipecat-context-hub MCP** tools for Pipecat framework questions
+  over reading `.venv` source.
+- **Dev-plan review markers are CI-gated.** `python3 scripts/check_review_markers.py`
+  (a CI step, stdlib + git only) fails if a reviewed plan's contract section —
+  everything above its `<!-- reviewed: YYYY-MM-DD @ <sha1> -->` marker — was edited
+  without refreshing the hash. The marker hashes only the above-marker bytes
+  (same convention as the skein `/review-plan` + `/conduct` tooling), so editing
+  the `## Progress` / `## Findings` workspace below it is free. Enable the same
+  check locally with `git config core.hooksPath .githooks`. To clear a failure:
+  re-run `/review-plan`, or recompute the hash for a purely administrative
+  above-marker edit (`head -n <marker_line-1> plan.md | git hash-object --stdin`).
+
+## Audio capture: PortAudio vs socket (`AUDIO_SOURCE`)
+
+`onoats bot` has two capture backends, selected by `AUDIO_SOURCE` (env /
+`config.toml [audio].source`), branched in exactly one place each:
+
+- **`portaudio`** (default) — today's `LocalAudioTransport` path, unchanged.
+- **`socket`** — reads framed PCM16 from two per-branch unix sockets via
+  `onoats.transports.socket_audio.UnixSocketAudioTransport`.
+
+Load-bearing invariants — do not break these without updating the tests that
+pin them (`tests/test_dual_socket_source.py`, `tests/test_socket_audio_transport.py`,
+`tests/test_socket_supervisor.py`, `tests/test_status_file.py`,
+`tests/test_native_contract_parity.py` — the last pins Swift literals to their
+Python contract constants):
+
+- **Never-mix.** One socket → one branch → one STT session → one `SourceTagger`.
+  Nothing fans a socket to both branches. `_build_socket_transports` refuses to
+  start if the two socket paths resolve (`Path.expanduser().resolve()`) to the
+  same file.
+- **No PortAudio on the socket path.** `AUDIO_SOURCE=socket` must neither import
+  nor invoke the PyAudio device-enumeration path (`select_dual_input_devices`).
+  `import onoats.dual` must succeed with **no native binary present** (CI is
+  native-free).
+- **EOF / read-idle is fatal, no self-reconnect.** The transport surfaces a
+  **fatal `ErrorFrame`** (`push_error(..., fatal=True)`, which goes *upstream* —
+  that's what cancels the pipeline) on EOF, a framing error, a read-idle timeout,
+  a BLOCK consumer-stall, or a staging-pump failure. The supervisor owns restart;
+  the transport never reconnects itself.
+- **Bounded staging + downstream gate.** The reader stages into a buffer bounded by
+  **both** a frame count (`max_buffered_frames`) and total bytes
+  (`max_buffered_bytes`); the drop policy (default `drop-oldest`, with a WARNING)
+  keeps memory capped under a faster-than-consumer writer. Because pipecat's
+  `_audio_in_queue` is unbounded, the pump also **gates on the base queue's depth**
+  (`_await_downstream_room`, high-water = `max_buffered_frames`) so a *stalled
+  consumer* can't grow it without limit — the frame cap bounds both queues (~2×
+  worst case). Don't reintroduce an unconditional `push_audio_frame` drain in the
+  pump; that's the bug Codex caught (re-verified via the integrated reader+pump
+  regression tests in `tests/test_socket_audio_transport.py`).
+
+Release metadata (canonical BSD-2-Clause LICENSE body, pyproject SPDX
+`license` field, `hatchling>=1.27` pin for PEP 639) is pinned by
+`tests/test_release_meta.py` — don't weaken those without touching the
+licensing decision in the release plan. The version lives on **three
+surfaces** that must agree — `pyproject.toml`, the `uv.lock` onoats entry
+(regenerate via `uv lock`), and the menu-bar Info.plist
+`CFBundleShortVersionString` — pinned by the same test file and gated at
+tag time by `scripts/release_check.sh vX.Y.Z <commit>` (run before pushing
+any release tag).
+
+Releasing to PyPI (v1.1.0+): pushing a `v*` tag triggers
+`.github/workflows/release.yml` — full suite + two guards (tag must equal
+the pyproject version; wheel metadata must carry no direct-URL deps), then
+publish via PyPI **trusted publishing** (OIDC, no stored token) gated behind
+the GitHub `pypi` environment. One-time prerequisites: a PyPI trusted
+publisher (project `onoats`, repo `vr000m/onoats-bot`, workflow
+`release.yml`, environment `pypi`) and the GitHub `pypi` environment itself.
+PyPI rejects direct-URL `Requires-Dist` entries — dependencies must come
+from PyPI (this is why `pipecat-local-stt-server` is a `>=0.1.2,<0.2`
+registry dep, not a git pin).
+
+## Supervisor ↔ capturer lifecycle (`cli._run_socket_supervisor`)
+
+- The **supervisor owns the capturer process.** It mints a private `0700` socket
+  dir + a fresh generation **nonce**, exports `ONOATS_MIC_SOCKET` /
+  `ONOATS_SYSTEM_SOCKET` / `ONOATS_CAPTURER_NONCE`, spawns `ONOATS_CAPTURER_BIN`
+  (paths + nonce via **both** argv and env), waits (bounded) for both sockets,
+  then runs the recorder. The capturer's tap preflight (release-plan Phase 7)
+  makes the TCC-prompting tap call **before** binding sockets, announced by
+  `ONOATS-EVENT waiting-for-permission`; if the base socket wait expires with
+  that event seen, the supervisor extends the wait once (+120 s) and surfaces
+  the pending prompt in the status file.
+- **Nonce gating end-to-end:** supervisor mints → `cfg.capturer_nonce` →
+  transport `expected_nonce`. A capturer presenting a missing/stale nonce on the
+  supervisor's paths is rejected at handshake. Ungated (None) when socket mode is
+  driven manually without the supervisor.
+- **Fail-loud is the contract.** Every failure path (missing/unspawnable capturer,
+  sockets that never appear, capturer death mid-session, STT preflight) yields a
+  non-zero exit + a WARNING/ERROR log, and a partial session still rotates into
+  `pending/` — never a hang.
+- **Capturer-exit-before-recorder is always fail-loud, even on `rc=0`** — the
+  recording is truncated regardless of exit code. Default-input-device changes
+  (e.g. AirPods removal) are the **capturer's** job to absorb by re-binding and
+  continuing to stream; see `docs/audio-socket-contract.md`.
+- Shared recorder arg handling lives in `dual._apply_recorder_args` — both
+  `dual.main` and the supervisor route through it, so interactive/category
+  handling can't drift.
+
+## Reviewing a subprocess / process-boundary change
+
+When a change spawns a child process (`create_subprocess_*` / `Popen` / `exec`)
+or otherwise crosses an OS boundary, the general review lenses tend to stay on
+the in-process logic and miss the boundary itself. The capturer supervisor's
+three post-review findings (one `[high]` no-ship) all came from this blind spot —
+the heavyweight gate stack passed; a single adversarial pass caught them. Run this
+checklist explicitly for any new spawn:
+
+- **Signals / session.** Does the child inherit the parent's controlling-terminal
+  signals (Ctrl+C / SIGTERM via the foreground process group)? If it must not,
+  spawn with `start_new_session=True`. (Without it, a graceful shutdown can be
+  mis-read as the child dying — the capturer's `[high]` finding.)
+- **Environment / secrets.** Never pass `dict(os.environ)` to a child. Build a
+  deny-by-default allowlist (see `cli._CAPTURER_ENV_POLICY`) so STT/application
+  secrets — and dylib-**injection** vars like `DYLD_INSERT_LIBRARIES` — never reach
+  a child that does not need them.
+- **Teardown reaches the whole group.** A session-leader child may spawn its own
+  helpers; signal the **process group** (`os.killpg(os.getpgid(pid), …)`), not just
+  the leader PID, on **both** the graceful and crash (leader-already-reaped) paths,
+  so nothing outlives the session holding a resource (e.g. the audio device).
+- **File descriptors.** Does the child inherit fds it should not (sockets, pipes,
+  log handles)? Pass only what is intended.
+- **Working dir + argv.** cwd is inherited; spawn via argv (no shell) so there is
+  no shell-injection surface — but verify `argv[0]` resolves to a trusted path.
+- **Failure is loud + bounded.** Spawn failure, child death, and a silent/hung
+  child each yield a non-zero exit + a WARNING/ERROR log + a bounded wait (no
+  hang) — see the fail-loud contract above.
+
+Each item should map to a regression test that fails against the pre-fix code
+(signal: spawn-kwarg + `rc`; env: a planted secret stripped from the child env;
+teardown: a spawned child PID is gone after stop). The supervisor's tests in
+`tests/test_socket_supervisor.py` are the worked example.
+
+## Wire-format contract
+
+`docs/audio-socket-contract.md` is the versioned (`v1`) capturer↔recorder
+contract and the source of truth for framing/handshake/constants. **Extending the
+wire format** means bumping `WIRE_VERSION` in `socket_audio.py` **and** the doc
+together — `tests/test_audio_socket_contract_parity.py` fails CI if the doc's
+constants table and the module drift.
+
+## Review Checklist
+
+Dismissed review findings (`won't-fix` / `analysis-error`) that future reviews
+should NOT re-flag go here, one per line:
+
+`- **[Category] disposition**: description (YYYY-MM-DD)`
+
+- **[Architecture] won't-fix**: `dual._apply_recorder_args` (and `_parse_args`)
+  keep their leading underscore despite being imported cross-module by `cli.py` —
+  this matches the established in-repo convention for shared-but-internal helpers
+  (`_parse_args`, `_build_socket_transports`); they are intentionally cross-module,
+  not a public API. (2026-06-09)
+- **[Architecture] won't-fix**: `max_buffered_bytes` is clamped to `max(1, …)`
+  only inside `UnixSocketAudioInputTransport`, not at the `UnixSocketAudioTransport`
+  facade — the inner clamp is the single point of use and mirrors how
+  `max_buffered_frames` is handled (`Queue(maxsize=max(1, …))`); forwarding the
+  raw value through the facade is intentional. (2026-06-09)
+- **[Logic] analysis-error**: `nonce[:8]` in the handshake log assumes a `str` —
+  unreachable: `parse_handshake` validates `nonce` is `str | None` and the log
+  guards on truthiness, so a non-string nonce can never reach the slice. (2026-06-09)
+- **[Architecture] won't-fix**: `native/spike/Info.plist` shares the production
+  `CFBundleIdentifier` — intentional: Spike 3's entire purpose was validating
+  TCC persistence on the PRODUCTION designated requirement (bundle id + cert),
+  so a distinct spike identity would invalidate the spike evidence. The spike
+  tree was deleted in Phase 6 of the 0.9→1.0 plan (preserved at the
+  `spike-archive` tag). (2026-06-10; resolved 2026-06-11)
+- **[Performance] won't-fix (scoped)**: the capturer IOProcs perform one bounded
+  heap copy (`Data(bytes:count:)`, ~10 ms chunk) and take an `NSCondition` lock
+  per callback (`MicCapture.enqueueChunk` / `SystemCapture.enqueueChunk`). A
+  textbook RT path would use a pre-allocated lock-free ring buffer; we keep the
+  copy+lock because (a) the worker holds the lock only for a queue pop —
+  microsecond contention window, (b) the pattern is hardware-verified across
+  the full Phase 4/5b smoke incl. hours-long real-call sessions with zero HAL
+  starvation, and (c) a ring-buffer rewrite would invalidate that evidence and
+  force a re-smoke for a latent, never-observed risk. What we DID fix
+  (2026-06-10): the drop-path `logLine` that ran on the realtime thread —
+  drops are now counted under the lock and reported from the worker thread.
+  Revisit only if a real session ever logs HAL silence/dropouts. (2026-06-10)
+- **[Logic] analysis-error**: `FrameChunker.append` back-extrapolating from the
+  total `pending.count` "over-counts leftover samples" — the math is exact while
+  capture is contiguous (leftovers are contiguous with the next buffer); only a
+  frame straddling a capture gap inherits a bounded <20 ms skew, governed by the
+  existing `lastEmittedEndNs` clamp. Comment added at the site. (2026-06-10)
+- **[Architecture] won't-fix**: `onoats bot --source` sets `AUDIO_SOURCE` in
+  `os.environ` rather than threading a parameter — deliberate: the env var is
+  the pre-existing public contract (config.toml/env already select the source),
+  the flag is a convenience alias onto that contract, and downstream re-parses
+  argv independently (documented at the site). Threading a parameter would
+  create a second, competing resolution path. (2026-06-11)
+- **[Security] won't-fix**: `make_cert.sh` passes the p12 transport password on
+  `security import -P` argv — `security import` has no file/stdin password
+  option. Residual is a one-shot random secret guarding a file that lives
+  seconds inside a 0700 tmpdir; documented at the site. (2026-06-11)
+- **[Architecture] won't-fix**: `LateBoundWriter` stays defined at file scope in
+  `main.swift` rather than moving to `Support.swift` — it is private plumbing of
+  main.swift's startup reorder (preflight-before-sockets), used nowhere else,
+  and the Swift sources just passed the full Phase 7 live smokes; relocating
+  code in a file with no Swift test runner would invalidate that evidence for
+  zero behavioural gain. (2026-06-11)
+- **[Architecture] analysis-error**: `permission_event` "is not documented in
+  the data-flow comment near `device_state`" — it is: the comment block at the
+  declaration site in `_supervise_socket_session` (directly under the
+  `device_state` comment) documents who sets it and who reads it. (2026-06-11)
+- **[Architecture] won't-fix**: the `auto`→`None` STT-language mapping lives in
+  `runtime._resolve_stt_language`, not in `OnoatsConfig.stt_language` —
+  deliberate: the property stays a plain `str` for parity with `stt_model`,
+  its docstring states the runtime does the mapping, and `_create_stt_service`
+  is the single consumer. Moving it would change the property's type contract
+  (`str | None`) for no caller. (2026-06-12)
+- **[Architecture] won't-fix**: `[stt].language` has no Swift menu-bar picker
+  or parity-test entry — intentional: the key is CLI/file-managed (like
+  `ws_socket`); the menu bar exposes config.toml via its open-config dropdown,
+  which is the supported edit path, and `ConfigStore` round-trips arbitrary
+  keys without code changes. A GUI picker would be a separate feature.
+  (2026-06-12)

onoats-1.1.0/CHANGELOG.md

@@ -0,0 +1,249 @@
+# Changelog
+
+All notable changes to onoats are documented here. The format is based on
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/), and the project
+adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+All listed versions — including the reconstructed pre-0.9 era — are licensed
+under BSD-2-Clause (see [LICENSE](LICENSE)).
+
+Versions before 0.9.0 were never published or tagged; they are reconstructed
+retrospectively from the merged-PR history (nothing was ever distributed, so
+no backdated tags exist). PR numbers `#1`–`#7` refer to this repository;
+older history predates the extraction and is cited by merge-commit SHA.
+Annotated tags exist from `v0.9.0` forward.
+
+## [1.1.0] - 2026-06-12
+
+First PyPI release (`pip install onoats` / `uv tool install onoats`).
+
+### Changed
+- `pipecat-local-stt-server` is now consumed from PyPI (`>=0.1.2,<0.2`)
+  instead of a git-URL pin — `0.1.2` is the release of the exact commit
+  previously pinned (`5062b98` == tag `v0.1.2`). This removes the
+  direct-reference metadata PyPI rejects, unblocking publication. Packaging
+  metadata (readme, authors, URLs, classifiers) added for the PyPI page.
+
+### Added
+- Tag-triggered PyPI publish via GitHub Actions trusted publishing
+  (`.github/workflows/release.yml`): pushing a `v*` tag runs the full suite
+  plus two guards (tag == pyproject version; no direct-URL deps in wheel
+  metadata), then publishes via OIDC behind the `pypi` GitHub environment —
+  no stored token.
+- `[stt] language` in `config.toml`: the STT decode language is now a
+  first-class config key (env `STT_LANGUAGE` > legacy alias `STT_WS_LANGUAGE`
+  > `[stt].language` > `en`),
+  shared by every launch path (CLI, menu-bar app, `onoats init`). `auto`
+  means auto-detect and maps to `None` at the backend boundary. The local
+  whisper/MLX branches now honour it too (they previously hardcoded `en`);
+  Deepgram does not consume it. `onoats init` prompts for it in the local
+  STT branch. (PR #22)
+
+### Fixed
+- Review fixes on the language key (PR #22): a whitespace-only
+  `[stt].language` now falls back to `en` instead of reaching the backend as
+  `language=""`; non-interactive `onoats init` re-runs carry an existing
+  `language` forward instead of silently erasing it; switching the wizard to
+  Deepgram preserves the key for a later switch back.
+
+## [1.0.0] - 2026-06-12
+
+First stable release. Closes out the 0.9.x series' 1.0.0 gates: pre-socket
+tap preflight (PR #17), BlackHole prose pruning + the `setup-cli` install
+path (PR #18), and the ConfigStore TOML-subset parity suite with its CRLF
+fix (PR #19).
+
+### Added
+- `make -C native setup-cli` (release-plan Phase 8): one-command CLI +
+  native-capture install (cert → capturer build/sign → CLI shim → `onoats
+  init`) that skips the menu-bar app bundle. README now documents the three
+  install paths side by side: menubar (`setup`), CLI + native capture
+  (`setup-cli`, macOS 14.4+), and CLI + PortAudio (toolchain-free, see
+  `docs/blackhole-fallback.md`).
+
+### Changed
+- BlackHole prose pruned to the conservative keep-list (release-plan
+  Phase 8): redundant README mentions and the `pyproject.toml` `[macos]`
+  comment now point at `docs/blackhole-fallback.md`; the `_LOOPBACK_HINTS`
+  auto-detection, no-loopback NOTE, backend help text, and all config-wiring
+  tests are unchanged.
+- Pre-socket tap preflight (release-plan Phase 7, PR #17): the capturer makes
+  the TCC-prompting tap call **before** binding its sockets, announced by
+  `ONOATS-EVENT waiting-for-permission`. A first Start with the Screen &
+  System Audio Recording dialog unanswered no longer dies at ~10 s — the
+  supervisor extends its socket wait once (+120 s) and surfaces "waiting for
+  the system-audio permission prompt" in the status file / menu bar.
+- rc=11 `exit_reason` renamed `system-audio-denied` → `system-audio-failed`:
+  a TCC denial never exits the capturer (denied taps deliver zeros and
+  surface as the zero-run warning); rc=11 fires only on genuine tap API
+  failure.
+
+### Fixed
+- Menu-bar settings edits no longer corrupt a `config.toml` with CRLF line
+  endings (release-plan Phase 9): ConfigStore's line scan trimmed with a
+  whitespace set that excludes `\r`, so CRLF section headers were never
+  matched and `writeValue` appended a duplicate section that fails TOML
+  parsing. Untouched lines keep their CRLF bytes verbatim; the edited line
+  is written with LF. Pinned by the new TOML-subset parity suite
+  (12 cases: comments, whitespace variants, absent section/key, duplicate
+  keys, non-string neighbours, byte-identity, CRLF).
+- Mic silence pacer now activates before the HAL bind, so a slow bind
+  (pending TCC dialog, Bluetooth device activation) can no longer trip the
+  recorder's 10 s read-idle and kill the session.
+- Latent start-timeout stamping bug: the prestart failure stamp read the
+  capturer's exit code after reaping it, so a hung-but-alive capturer was
+  mislabeled `capturer-start-failed`; `capturer-start-timeout` is now
+  reachable end-to-end.
+
+## [0.9.0] - 2026-06-11
+
+Milestone B: native macOS capture + menu-bar app (PR #5, `16da012`; docs
+ride-alongs PR #6 `6aa0025`, PR #7 `8db8840`).
+
+### Added
+- Native macOS system-audio capture via a Core Audio process tap
+  (`onoats-capturer` Swift binary); no loopback driver needed on macOS 14.4+.
+- SwiftUI menu-bar app (`Onoats.app`): Start/Flush/Stop, inline mic picker
+  (sets the macOS default input), STT service picker, data-dir chooser,
+  status display.
+- Per-chunk capture-generation stamping — each queued audio chunk carries its
+  generation's format and resampler, preventing stale-format races on device
+  switch.
+- Self-signed signing identity workflow: `make -C native cert` (refuses to
+  regenerate an existing cert) + `make -C native install` / `install-cli`.
+- Sustained all-zero-input detector (30 s) in the capturer — surfaces a
+  denied system-audio grant as a WARNING instead of silent empty recordings.
+- Kill-×3 tap/aggregate residue smoke (`native/residue_check.sh`) and
+  one-command wire smoke (`native/smoke_wire_check.sh`).
+
+### Changed
+- The native capturer is the default macOS capture story;
+  BlackHole/PortAudio demoted to the documented fallback (macOS 13.x–14.3 /
+  off-mac).
+- `ConfigStore` (menu bar) writes TOML with correct basic-string escaping;
+  data-dir handling made canonical and per-session.
+
+### Fixed
+- Pre-start capturer death now writes a status record — a denied grant no
+  longer surfaces as "failed: graceful".
+- Realtime-thread logging and state races; fail-loud flush path; IOProc
+  zero-guard.
+
+### Notes
+- The runtime dependency `pipecat-local-stt-server` is intentionally
+  git-pinned (`pyproject.toml`) — correct for the from-source install story;
+  revisit only if PyPI publishing is ever taken up.
+
+## [0.8.0] - 2026-06-09
+
+Milestone A: socket audio transport + supervisor (PR #4, `1f9dfdc`).
+
+### Added
+- `UnixSocketAudioTransport` — framed-PCM16 Unix-socket audio input pipeline
+  with bounded staging, downstream-queue gating, and fail-loud fatal errors.
+- `AUDIO_SOURCE=portaudio|socket` backend switch, branched in exactly one
+  place per layer.
+- CLI supervisor: private `0700` socket dir, per-session generation nonce
+  (handshake-enforced), bounded socket-appearance wait, process-group
+  teardown on both crash and graceful paths.
+- Audio-socket wire contract document (`docs/audio-socket-contract.md`) with
+  a parity test that fails CI when doc and code constants drift.
+- `AGENTS.md` contract notes; dev-plan review-marker CI gate
+  (`scripts/check_review_markers.py`).
+
+### Changed
+- Capturer environment built from a deny-by-default allowlist (blocks DYLD
+  injection); capturer spawned in an isolated session so terminal signals
+  don't trip the fail-loud path.
+
+### Fixed
+- Transport failures are fatal upstream errors, so the recorder terminates
+  instead of hanging; handshake reads bounded by the idle watchdog; tilde
+  expansion on socket paths; capturer group swept on the crash path.
+
+## [0.7.0] - 2026-06-08
+
+Packaging and extraction era: standalone `onoats` package (untagged;
+PR #1 `7f33e72`, PR #2 `645d90b`, PR #3 `8e6c165`, plus direct commits).
+
+### Added
+- `src/onoats/` installable package layout, extracted from the `bot/`
+  monolith; `onoats init` CLI; CI with ruff + PortAudio build steps.
+- Configurable `[storage].data_dir` (`config.toml` + `onoats init
+  --data-dir`); XDG data paths.
+- `STT_WS_LANGUAGE` env to control the websocket decoder language.
+
+### Changed
+- `config.toml` honoured in the recorder process for STT and device
+  settings; SQLite overlay removed; transcript date grouping uses local
+  date, not UTC.
+- STT server renamed `onoats-stt` → `pipecat-stt` (server v0.2.0): socket
+  paths and labels updated.
+
+### Fixed
+- STT RSS probe resolves `stt_server` from the config-layered env (PR #1).
+- Flush (SIGUSR1) verifies live process identity before signalling — a stale
+  PID file can no longer kill an unrelated process (PR #2).
+- Shutdown drains the final transcript segment before flush; Ctrl+C cancel
+  timeout capped (no more ~20 s hang); device provenance logged accurately
+  (PR #3).
+
+## [0.5.0] - 2026-05-21
+
+Dual-input diarization and STT-service era (untagged; pre-extraction
+history, cited by merge SHA).
+
+### Added
+- Standalone Whisper WebSocket transcription server with a Pipecat
+  `STTService` wrapper and launchd auto-restart (`c4f08d5`).
+- STT health preflight, status probe, and `./onoats stt` wrapper
+  (`e40f2de`).
+- Live passive transcript view with rotation and orphan guards (`2536687`).
+- SmartTurn V3 read-only shadow observer on the `me` branch + raw PCM dump
+  for offline A/B testing (`0e4ae64`).
+- Parakeet ASR backend with per-server multi-ASR selection (`21ebfae`).
+- Cron-driven post-processing worker decoupled from the bot harness; the
+  bot rotates session files instead of processing in-flight (`4d54b88`,
+  the `stt-extraction-base` tag's commit).
+
+### Changed
+- Dual-input pipeline became the default bot path — coarse Me/Them
+  diarization via two PortAudio branches (`196aef2`).
+- Post-processing queue rebuilt as an FSM; the bot harness no longer owns
+  worker state (`4d54b88`).
+
+### Fixed
+- Continuation-flush session-file swap race; speaker-label plumbing and
+  source resolution; STT reconnect backoff made exponential (0.5 → 8 s);
+  long turns chunked.
+
+## [0.3.0] - 2026-04-16
+
+Processing-pipeline era (untagged; pre-extraction history).
+
+### Added
+- Topic discovery, collation, and management pipeline (`6124d56`).
+- Multi-task LLM benchmark, LM Studio provider, per-task LLM routing, and
+  manual transcript flush via Ctrl+T / `./onoats flush` (`ced152a`).
+- "seminars" processing category + `--category` bot CLI flag (`81afbc0`).
+
+### Fixed
+- Timeline corruption: ownership-aware rollback + cross-midnight concat in
+  segmenter/classifier/merge (`7b97d7e`).
+- PID-file identity hardened; pipeline cancel interruptible on Ctrl+C
+  (`ced152a`).
+- `ONOATS_DATA_DIR` tilde expansion across all entry points.
+
+## [0.1.0] - 2026-04-04
+
+Initial recorder (untagged; root commit `d645650`, 2026-03-30).
+
+### Added
+- Pipecat-based meeting recorder: PortAudio input, silence detection,
+  transcript buffering, Ctrl+C graceful shutdown with the current session
+  processed before exit.
+- SQLite overlay, web intelligence layer, and transcript management
+  (`3d36ae3`).
+- Collation service: living documents aggregated from related idea
+  transcripts (`5cd908c`).
+- LLM transcript cleanup + dictionary/provenance plumbing (`9d974c1`).

onoats-1.1.0/LICENSE

@@ -0,0 +1,24 @@
+BSD 2-Clause License
+
+Copyright (c) 2025–2026 Varun Singh
+
+Redistribution and use in source and binary forms, with or without
+modification, are permitted provided that the following conditions are met:
+
+1. Redistributions of source code must retain the above copyright notice, this
+   list of conditions and the following disclaimer.
+
+2. Redistributions in binary form must reproduce the above copyright notice,
+   this list of conditions and the following disclaimer in the documentation
+   and/or other materials provided with the distribution.
+
+THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
+AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
+IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
+FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
+SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
+CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
+OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.