feat: microphone domain — list/inspect/gain/array/param/stream/record (0.9.0) - #6
Merged
Conversation
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Add EXIT_BUSY_ERROR=3 to cli/_errors.py and a new microphone_cli/access.py
(cited from webcam-cli webcam_cli/access.py) with AccessState
{ok,absent,forbidden,busy}, check_access/require_access, a best-effort
/proc/*/fd holder lookup, and kind-specific remediations: 'audio' group
membership for /dev/snd nodes, a udev SUBSYSTEM=="usb" rule for
/dev/bus/usb nodes. tests/test_access.py covers all four states against a
fake /proc tree, hardware-free (os.open is monkeypatched throughout; no
real /dev or /proc access).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Cites webcam-cli/webcam_cli/activation.py:26-129, renamed to MICROPHONE_ACTIVATION_LOG and adapted to microphone-cli's field set (verb, device, params, started_at, ended_at). Appends one JSON line per --apply action to XDG_STATE_HOME/microphone-cli/activation.jsonl (or the env override), never overwriting prior lines. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Ports the audio subset of webcam-cli's engine.py (Capability/AudioFormat, detect(), require_engine()) for microphone-cli's own element set — alsasrc/audioconvert/audioresample/queue core, opusenc/matroskamux/wavenc/ rtpopuspay/rtpL16pay/udpsink optional — plus require_elements() and two pure argv builders: build_audio_stream_argv() (RTP/UDP passthrough or opus) and build_audio_record_argv() (Matroska/Opus or WAV, with an alsasrc num-buffers+latency-time bound for duration_s). No PyGObject/gi import; detect() is the only function that shells out. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Add microphone_cli/devices.py: enumerate USB audio capture cards from /proc/asound/cards plus the sysfs USB parent walk, build stable ids from the USB descriptors (udev's escaping rules; sysfs-path fallback when the firmware ships no serial), flag XVF3800 arrays by idVendor:idProduct (38fb:1001, 2886:001a), report the capture channel count from /proc/asound/cardN/stream0, and resolve() a selector to exactly one device — refusing raw hw:N / plughw:N / bare-integer card numbers and refusing ambiguity, both as CliError(EXIT_USER_ERROR) naming the stable ids to use instead. Cited from webcam_cli/devices.py with the video half dropped; provenance is in the module docstring. Four synthetic fixture trees under tests/fixtures (text files and relative symlinks only, no real captures): host-baseline, host-renumbered, two-arrays, respeaker. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Adds the hardware-free protocol core for microphone-cli:
* microphone_cli/usbctl.py — ctypes mirror of struct usbdevfs_ctrltransfer,
USBDEVFS_CONTROL derived as _IOWR('U', 0, sizeof(struct)) (0xC0185500 on
64-bit), control_transfer() over fcntl.ioctl behind a module-level _ioctl
seam, sysfs device enumeration to /dev/bus/usb/BBB/DDD node paths, and
open_device() mapping EACCES/ENOENT onto CliError with a udev remediation.
* microphone_cli/xvf3800.py — the vendor protocol ported from reachy_mini's
media/audio_control_utils.py (Apache-2.0, cited in the module docstring),
with the PARAMETERS table vendored verbatim, typed pack/unpack, the
status-64 retry loop, PERSISTENT/KNOWN_IDS, and param_info().
Tests inject a fake ioctl; no test opens /dev or touches real hardware.
Runtime dependencies stay empty — stdlib only.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
- list_devices.py: enumerate_devices() + per-device audio_access probe
via access.check_access on the ALSA capture PCM node
- inspect.py: parses <root>/proc/asound/cardN/stream0 for formats/rates/
channels; on an XVF3800 array reads firmware.{version,build,host,
repo_hash} via usbctl.find_devices + Xvf3800.firmware_info(), reporting
{"error": ...} rather than hard-failing on a lookup or permission problem
- Cites ../webcam-cli/webcam_cli/cli/_commands/list_devices.py for the
"compose, never hard-fail" shape
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
`array` noun group: bare `array` and `array overview` describe the group; `array doa <device>` reads DOA_VALUE_RADIANS as raw firmware radians (no coordinate transform), single-shot or `--watch` JSON Lines with `--interval`/`--count`, Ctrl-C exits 0 and a lost device exits 2; `array aec get|set` reports converged/bypass/hpf/echo/num_mics/geometry and flips --echo/--bypass/--hpf, dry-run unless --apply (which writes inside an activation_scope). Every nested add_subparsers passes parser_class=type(p) so parse errors keep the structured error contract. Hardware is reached only through the module-level _open_array / _sleep seams; tests serve every transfer from an in-memory fake and never open a device node. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Adds microphone_cli/cli/_commands/param.py: `param overview`, `param list`,
`param get <device> <NAME>`, `param set <device> <NAME> <values...>
[--apply] [--allow-persistent]`. Name/access/count validation (via
xvf3800.param_info and a local per-type value parser) happens before any
device is opened, so an unknown name, a ro name, or a wrong value count
never issues a transfer, whether or not --apply is passed. The persistent
tier (xvf3800.PERSISTENT, already added by t3) is checked before opening the
device or writing when --apply is set: a persistent/destructive name
without --allow-persistent is refused with a hint naming the flag and
explaining the volatile-vs-persistent distinction. Applied writes run
inside activation_scope("param set", ...) and include a readback when the
parameter is also readable (rw).
Adds tests/test_param.py, which builds a standalone parser via
param.register() (the noun isn't wired into cli/__init__.py yet — that's a
separate task) and monkeypatches param.resolve / param._open_array so no
test opens /dev.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Adds microphone_cli/mixer.py (amixer -c <card> contents/cset wrapper,
hardware-free via a run= seam) and microphone_cli/cli/_commands/gain.py
exposing `gain overview|get|set`. `gain get` reports the ALSA
capture-volume control and, on array devices, AUDIO_MGR_MIC_GAIN read
over the XVF3800 vendor protocol (firmware read failures degrade to
firmware: {error} rather than failing the whole command). `gain set`
is a dry run by default (plans the amixer argv/firmware write without
issuing them) and only applies with --apply, wrapped in exactly one
activation_scope("gain set", ...) log line.
Not yet wired into cli/__init__.py (a separate task owns that file);
tests/test_gain.py drives gain.register() through a standalone
_CliArgumentParser instead of microphone_cli.cli.main.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Two capture verbs, both cited (audio paths only) from webcam-cli's stream.py/record.py and both keeping the three-level hardware split: * `stream audio <device>` — RTP/UDP attachment point via udpsink, with --port/--host/--rate/--channels/--format/--encode. Default is a dry run that prints the exact gst-launch-1.0 argv and touches nothing; --probe detects the engine and checks the capture node without spawning; --apply requires the engine and the encode's elements, enforces access (busy is exit 3), spawns through the module-level `_spawn` seam inside an activation_scope, and returns the pid plus per-codec receive pipelines. * `record <device> <output>` — bounded capture to .mka/.wav (container from the extension), with --duration (default 30, 0 < d <= 3600) and --max-bytes (default 256 MiB, <= 4 GiB) enforced both by the argv's own num-buffers bound and by a poll loop over the growing artifact through the `_sleep`/`_monotonic` seams; stopped_reason reports which bound won. Refuses a missing parent dir, an unknown extension, and an existing file without --overwrite. Payload key names follow the cited modules, minus the video-only ones. Tests are hardware-free by construction: subprocess.Popen/run are booby-trapped autouse, devices resolve under tests/fixtures/host-baseline, the capture node is root-joined so it is absent under a fixture root, and the activation log is redirected into tmp_path. Not wired into cli/__init__.py, learn/overview/catalog — those files belong to the wiring task; the tests build the same parser shape by hand. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Register every wave-0/1 command module in `_build_parser` (list, inspect,
gain, array, param, stream, record) alongside the introspection verbs, and
set `prog="microphone"` to match `[project.scripts]` — the binary is
`microphone`, so nothing the CLI prints should tell an agent to type
`microphone-cli`.
Rewrite the three hand-maintained duplicates of the surface for the
microphone domain:
* `learn.py` — purpose, the three-level hardware-touch split first
(dry-run / --probe / --apply) plus the read-only vendor control transfers
that still open the USB node, the full command map, the stable-id selector
rule, exit codes 0/1/2/3, the persistent tier, raw firmware radians,
consent and the activation-log pointer, and an audience section naming
media-cli as the composing consumer. `_as_json_payload` gains `command`,
`import_package`, `audience`, `hardware_activation`, `device_selector`,
`bounds`, `persistent_tier`, `raw_firmware_values`, `consent`.
* `overview.py` — `_VERBS` covering all 29 registered paths, plus
`_HARDWARE`, `_CONTRACTS` and a consent section; `cli_sections` now reuses
`_VERBS` verbatim instead of re-declaring it.
* `explain/catalog.py` — an entry per registered path, with
`("microphone-cli",)` kept as a resolvable legacy alias.
Ported from webcam-cli's tests/test_cli.py:190-327: `_registered_paths()`
walks the live argparse tree, and the parity tests require the catalog,
`overview._VERBS` and learn's command map to agree with it. Two prose gates
join them — no user-facing string may present `microphone-cli` as a typable
command, and none may carry the scaffold's clonable/template wording.
`doctor`'s checks and JSON shape are untouched; only its text header and
docstring changed. Three wave-1 modules needed their overview *subject*
strings retitled (`microphone-cli array` -> `microphone array`, same for
`param` and `stream`) so the dead-command gate can pass; their flags and
payload keys are unchanged, and the affected assertions in
tests/test_array.py, test_param.py and test_stream.py moved with them.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
README.md and CLAUDE.md replace the scaffold-state narrative from 5f9b1bd with the domain state: 13 top-level verbs (list, inspect, gain get/set, array doa, array aec get/set, param list/get/set, stream audio, record, plus the six agent-first verbs), 276 tests, 92% coverage. README follows webcam-cli's heading shape (Status, Scope, Quickstart, CLI table, What comes out, What touches the hardware, Why device identity is the hard part, What this repo carries, License). CLAUDE.md gets a module map, the three-level hardware split, the testing seams (root=, _open_array, _ioctl, _spawn, _sleep, run=), the fixture trees, the parity tests, and the issue #3 hardware-acceptance status; the console-script note is marked RESOLVED now that prog and the script both read `microphone`. Adds docs/xvf3800-parameters.md: attribution for the vendored XVF3800 parameter table (Pollen Robotics' reachy_mini, Apache-2.0) and a guide to its resid groups and persistent tier. docs/skill-sources.md gains rows for the recall/remember skills (first-party to eidetic-cli, not guildmaster), which were vendored but never entered into the provenance ledger. Version bumped 0.8.2 -> 0.9.0 (MINOR) via the version-bump skill. No Python source, tests, or .github/workflows touched. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
…ot-joined capture node Both found during on-device acceptance against a ReSpeaker XVF3800 (2886:001a): the permission hint told the user to add a rule for 38fb:1001, and 'list --root <fixture>' probed the host's real /dev/snd/pcmC1D0c because the fixture's card 1 collided with the live card 1. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
…rmware Found on hardware (ReSpeaker XVF3800, Seeed USB firmware v2.1.0): the vendored reachy_mini table is the 38fb:1001 map. Seeed's 2886 firmware has no DOA_VALUE_RADIANS, its DOA_VALUE is two uint16 (degrees, speech flag), and it adds LED_RING_COLOR and the AIC3104 output levels. Adds the uint16 codec, FIRMWARE_OVERLAYS keyed by USB vendor id, Xvf3800(vendor=...), param resolution after device resolution, 'param list --vendor', and azimuth_deg alongside azimuth_rad in 'array doa' (unit conversion only). Plan deviation d2. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
- mixer: accept ',index=N' on the amixer numid line so a second same-named control no longer overwrites the first one's values (gain readback was stale) - engine: passthrough stream converts to S16BE before rtpL16pay (the payloader refused S16LE: 'could not link queue0 to rtpl16pay0') - stream/record: --rate/--channels/--format default to what the device advertises in /proc/asound stream0 instead of a fixed 48 kHz mono, which the exact caps filter could never negotiate on a 16 kHz stereo array; the payload records each field's source (explicit/advertised/default) Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
PR Summary by QodoAdd complete microphone control, streaming, and recording domain
AI Description
Diagram
High-Level Assessment
Files changed (115)
|
Code Review by Qodo
1.
|
![qodo-code-review[bot]](https://avatars.githubusercontent.com/in/484649?s=60&v=4){kind=small}
…e ranges, unvalidated gain - cli/__init__.py: wrap parser construction/parse_args in main() so any non-SystemExit, non-KeyboardInterrupt exception raised before _dispatch's boundary (e.g. an import or registration failure) becomes a structured CliError instead of a raw traceback. - xvf3800.py _decode: the uint8 branch now checks the reply length against info.count and raises the same short-reply CliError the wide/half types raise. - xvf3800.py _encode: uint8 and uint16 values are range-checked (0..255 / 0..65535) before packing, so an out-of-range value raises CliError(1) instead of being silently masked/truncated (-1 -> 255, 256 -> 0) and sent to firmware. - gain.py cmd_gain_set: validates the gain value is finite and in 0.0..1.0 before any planning or writing, for every --target, instead of relying on _map_to_alsa's clamp (which never reached the firmware write). Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
…in validation (PR #6 threads 1, 6, 10, 11)
- record_activation() looped os.write() once and ignored its return value, so a short write could truncate an activation line. Loop until the full payload is written (or raise OSError on a 0-byte/failed write), keeping the single-open O_APPEND semantics that make lines under PIPE_BUF atomic. - activation_scope() only opened the log after running the protected action, so a permission/ENOSPC problem surfaced as a generic failure after hardware had already changed, inviting a retry of a non-idempotent action. It now resolves the path, creates parent directories, and opens the log for append (then closes it) before yielding, raising CliError(code=2) up front on failure. If the final write still fails after the body ran, it now raises CliError(code=2) stating the action was applied and only the audit write failed, instead of a bare OSError. Tests: test_record_activation_survives_short_writes, test_record_activation_raises_on_zero_byte_write, test_activation_scope_raises_before_body_when_log_dir_unwritable, test_activation_scope_reports_applied_but_not_logged_after_body_ran, test_activation_scope_reports_applied_but_not_logged_on_late_write_failure_after_raise. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
…y (PR #6 threads 7, 8)
Four review findings on the --apply paths of `record` and `stream audio`, all fixed test-first. 1. Recordings can outlive their bounds (record.py `_stop`/`_run_bounded`). A child that ignored SIGTERM was only warned about, and the bound was reported as if it had stopped the recording. `_stop` now escalates to SIGKILL, waits for it, and raises CliError(code=2) if the process survives even that — it never returns while the child is alive. 2. Recordings can exceed their size cap (record.py `_run_bounded`/`_apply`). The process-exit check returned "eos" before the artifact was compared with --max-bytes, so a pipeline that blew the cap and exited inside one poll interval looked like a clean bounded recording. The finished artifact is now stat'ed whatever the stopped_reason, and an oversized file is a typed exit-2 error naming the bytes written and the cap. The file is deliberately kept at its path — deleting it is the caller's call — and the activation record still carries the outcome. 3. Broken streams are reported as live (stream.py `cmd_stream_audio`). A successful Popen was treated as a running stream and the child's stderr was discarded. The child's stderr now goes to a temp file, the command waits STARTUP_SETTLE_S (0.5s, via the new `_sleep` seam) and polls; a child that has already exited is CliError(code=2, "stream pipeline exited <rc> during startup") whose remediation quotes the pipeline's own first ERROR/WARNING/ not-negotiated lines — never a traceback. 4. Live streams look finished in the audit (stream.py + activation). An unbounded stream got ended_at at launch. It is now logged honestly with ended_at=None and params carrying pid and lifetime="unbounded" (via record_activation with an explicit Activation, not activation_scope); a startup failure still logs a closed line with params.error. The payload's lifetime/consent text says the log line is open-ended. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
…ection, open-ended stream audit (PR #6 threads 3, 4, 5, 9)
…n unstopped over-cap artifact is an error Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Hoist setup calls out of pytest.raises blocks so each contains exactly one call that can raise (S5778), split composite `and` assertions into separate asserts (S9073), and drop empty parentheses from bare @pytest.fixture decorators (S9083). Behaviour and assertions are unchanged; test count stays at 304 passed. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Behaviour-preserving refactor of the four command modules Sonar flagged: - record._payload took 14 parameters (S107): the container, format, bounds and output path now travel together in a frozen `_Plan` dataclass, built once in cmd_record and also passed to _apply. Payload keys and values are unchanged. - record's `mode` selection was a nested conditional expression (S3358); it is now a plain if/elif/else. - cmd_record, cmd_stream_audio and cmd_param_list always returned 0 (S3516). Per the dispatch contract (_dispatch maps None to exit 0) these handlers — and their unconditionally-zero neighbours cmd_param_get, cmd_param_set, cmd_param_overview, cmd_gain_get, cmd_gain_set, cmd_gain_overview and both _no_verb wrappers — now return None. - stream.advertised_format had cognitive complexity 16 (S3776); the per-field fallback logic moved into _resolved_rate / _resolved_channels / _resolved_sample_format, each returning (value, source). - "Emit structured JSON." was duplicated in param.py (5x) and gain.py (4x) (S1192). It is now JSON_FLAG_HELP in microphone_cli/cli/_commands/__init__, used by param, gain, stream and record. 304 tests pass unchanged; teken cli doctor --strict stays 26/26. Every --json payload was diffed against the pre-refactor build and is identical apart from its own resolved_at timestamp. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
…nt continue, type mismatch - mixer.py: _NUMID_RE's lazy `.*?` name group (ambiguous with the closing quote char) replaced with `[^']*` (S8786); list_controls' line-kind branching split into _header_fields/_apply_attr_line/_apply_value_line/ _control_from_fields helpers, dropping cognitive complexity from 19 and removing the now-unreachable trailing `continue` (S3776, S3626). - inspect.py: _RATES_RE's `.+` (ambiguous with the preceding `\s*`, both matching whitespace) replaced with `\S.*` (S8786); _parse_capture_block split into _section_header/_append_format/_append_channels/_append_rates/ _parse_capture_line helpers, dropping cognitive complexity from 27 (S3776). - usbctl.py: find_devices' per-device attribute read and the vendor/ product/serial filter extracted into _read_device_attrs/_matches_filters, dropping cognitive complexity from 17 (S3776). - activation.py: explicitly annotate activation_scope's `target` local as Path so its type is unambiguous at both _record_or_report_applied call sites (S5655). All regex rewrites verified against the existing amixer/stream0 fixtures, including the ",index=1" duplicate-control case. 304 tests pass unchanged; black/isort/flake8/bandit clean. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
… Activation explicitly Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
…g regex Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za
|
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
The microphone domain, built from a converged devague spec and plan and accepted on real hardware.
list,inspect,gain get|set,array doa(single-shot or--watchJSON Lines) andarray aec get|set,param list|get|setover the full XVF3800 table with a persistent tier behind--allow-persistent,stream audio(RTP/UDP passthrough or Opus),record(bounded WAV/Matroska), plus the agent-first baseline.devices(stable ids from USB serial),access(ok/absent/forbidden/busy → exit 0/1/2/3),usbctl(stdlibusbdevfscontrol transfers, no pyusb),xvf3800(vendored parameter table with per-firmware overlays, typed codec, status-64 retry),mixer(amixer),engine(GStreamer argv),activation(append-only log of every--apply).--probechecks without opening;--applytouches hardware and is logged.progis nowmicrophone, matching the console script; template prose is gone and a parity test keeps catalog/learn/overview in sync with the parser.teken cli doctor --strict26/26, workflows untouched,dependencies = [].How it was built
/scope→/think→/challenge→/spec-to-plan→/assign-to-workforce: 13 tasks in 5 waves, one agent per task in isolated worktrees, TDD-gated merges. Artifacts:docs/specs/2026-09-06-microphone-domain.md,docs/plans/2026-09-06-microphone-domain.md(+ split),.devague/.Two approved plan deviations: d1 acceptance ran on a Seeed ReSpeaker XVF3800 (
2886:001a) instead of the Reachy Mini Lite the plan named (the robot on hand was a Reachy Mini whose array lives on its own Pi); d2 the parameter map turned out to be firmware-specific, soxvf3800.FIRMWARE_OVERLAYSwas added.On-device acceptance
scripts/acceptance/run.sh --writes --mediapassed 10/10 on the array (Seeed USB firmware 2.1.0): DoA matched Seeed's own reference reader exactly, all volatile writes round-tripped and were restored, record produced a real 16 kHz stereo WAV, and a blind consumer attached to the stream from the JSON payload alone. Evidence and the seven hardware-only defects it surfaced (all fixed here) are indocs/acceptance-microphone-domain.md.Follow-ups
On-device acceptance: XVF3800 DoA/AEC/gain against a Reachy Mini Lite (USB) #3 stays open for the Reachy Mini Lite (
38fb:1001) path.Support XVF3800 firmware variants and DFU reflash (XIAO board ships I2S firmware, silent on USB) #4 CLI support for the I2S→USB firmware bring-up (safe mode, dfu-util, udev).
Expose firmware voice activity detection: array vad get/watch/set #5 voice-activity exposure (
array vad).microphone-cli (Claude)
🤖 Generated with Claude Code
https://claude.ai/code/session_01Jei6EarsNVUDHMNxb1U1Za