aes67-ESP32-P4/CLAUDE.md

aes67-ESP32-P4

AES67 sender on a Waveshare ESP32-P4-ETH (PoE). Sources: Spotify Connect (cspot) and HLS/m3u8, with failover. Riedel-style web UI and a reusable AES67 core (PTP, RTP TX, SDP/SAP, VLAN, syslog, health) meant for later AES67 projects.

Repo: https://gitea.apointless.space/bsncubed/aes67-ESP32-P4

Read first

• docs/aes67-core-base.md: reusable core. Config schema, API, PTP (roles, Riedel defaults, status fields), AES67 TX/SDP/SAP, VLAN, syslog, temperatures, component layout.
• docs/hardware-and-design-notes.md: board pinout, chip revision caveat, project pipeline (cspot/HLS), failover, player API.
• web/index.html: finished web UI (single file). It is the API contract; firmware must match the JSON it reads and writes.

Working rules

• One change at a time ("one fuckup at a time"). Change or test one variable per step; never stack several fixes or hypotheses. If something breaks, go back to the last known-good state.
• Follow the build order below. Finish and verify a step on hardware before starting the next.
• After every change: idf.py build, flash, and check the serial monitor output. Show the relevant log lines.
• Don't mark a step done until it has been verified on the board.
• Keep core and project separate: components/aes67_* must not depend on main/ (sources/player). Project code plugs in through the config, status and route registries.
• Licences: core (components/aes67_*) and web/ are MIT, the rest GPL-3.0-or-later (see THIRD_PARTY.md). Never pull GPL code (cspot, bell, squeezelite) into the core. New third-party code: check its licence first and add it to THIRD_PARTY.md.
• web/index.html: keep the CORE / PROJECT markers. If the firmware needs an API change, change the doc and the page together.
• Small, focused commits per step.

Toolchain

• ESP-IDF 5.5 or later (needed for the P4 and for examples/network/vlan_support). Target esp32p4.
• idf.py set-target esp32p4, idf.py build, idf.py -p <PORT> flash monitor
• Before first build: run esptool.py chip_id and esptool.py flash_id.
  • Chip revision < v3.0 (engineering silicon, seen on some of these boards) needs CONFIG_ESP32P4_SELECTS_REV_LESS_V3=y. Our board: v1.3 (set in sdkconfig.defaults, min rev v1.0).
  • Our board: 32 MB flash (GigaDevice c8/4019). App slots must stay below 16 MB (cache mapping above 16 MB is experimental in IDF).
  • Rev < 3 also limits Espressif's prebuilt audio libraries: esp_audio_codec must stay < 2.6 and esp_audio_effects < 1.4 (newer versions use P4 assembly that needs rev >= 3; the build fails with a message saying so). Check this for any new Espressif binary component.
• Embed web/index.html via EMBED_TXTFILES in aes67_web.
• cspot (Spotify, external/cspot git submodule, philippe44 fork, pinned): after cloning run git submodule update --init external/cspot && git -C external/cspot submodule update --init cspot/bell. Its nanopb code generator needs, in the IDF Python env: python -m pip install protobuf grpcio-tools 'setuptools<81' (after . export.sh). Our fixes to cspot/bell live in components/spotify/patches/{cspot,bell}/*.patch and are applied automatically at configure time; don't edit the submodule directly, add a patch.
• Firmware version: <version.txt>-<git short hash>, e.g. 0.0.1-1d613b1. Bump version.txt for a release; the hash is the commit that was built (commit first, then build, so it matches). No dirty flag.
• Flash over the network (normal way since step 2b; keep USB for recovery): curl -f --data-binary @build/aes67_p4.bin -H 'Content-Type: application/octet-stream' http://p4-aes67/api/ota The board reboots into the new image on trial; check GET /api/ota shows the new version with pending_verify: false.
• Another board, over USB: tools/flash_board.py (checks chip revision and flash size, optional erase, flashes, finds the IP in the boot log, sets hostname / stream name / multicast so boards don't clash). Needs a build and esptool (IDF env or pip install esptool).
• Serial: opening /dev/ttyACM0 can reset the board. Don't open it while an OTA image is on trial (a reset then counts as a failed boot and rolls back).
• Rollback test build (self-test always fails), in its own build dir: idf.py -B build-selftest-fail -DSDKCONFIG=build-selftest-fail/sdkconfig -DSDKCONFIG_DEFAULTS="sdkconfig.defaults;sdkconfig.selftest_fail" build

Board quick reference (full details in docs)

• Ethernet: IP101GRI, RMII, PHY addr 1, ref clock from PHY into GPIO50 (EMAC_CLK_EXT_IN). MDC 31, MDIO 52, PHY reset/power 51, TX_EN 49, TXD0 34, TXD1 35, CRS_DV 28, RXD0 29, RXD1 30.
• No Wi-Fi/BT on this board.
• On-die temperature sensor only (driver/temperature_sensor.h).

Build order / status

• 0. Check chip revision and flash size; create IDF project and empty component stubs (layout in aes67-core-base.md). Set up the OTA partition table (two app slots, no factory) and PROJECT_VER now, so the layout never changes later. (PROJECT_VER = version.txt + git short hash.)
• 1. Ethernet: IP101 up, DHCP, IP logged. Ping works.
• 2. Web server + config store (cJSON in NVS) + embedded index.html; /api/config, /api/status (stub values), /api/reboot.
• 2a. Finding the device: mDNS (hostname.local + _http._tcp), then LLDP (switch shows name + IP).
• 2b. Firmware update: /api/ota upload + rollback self-test. Test: update to a new build, then deliberately flash a build that fails its self-test and confirm it rolls back. After this, flash over the network; keep USB for recovery.
• 3. PTP TimeReceiver: lock to an existing GM (Riedel), fill status.ptp. Confirm EMAC hardware timestamps work. First check whether the installed ESP-IDF has a PTP example/component for the P4 before writing one.
• 4. AES67 TX with a 1 kHz test tone, PTP-paced; /stream.sdp. Verify: import SDP on a Riedel Artist 4-wire AES67 port, and check packets/timestamps in Wireshark. Done and checked with a receiver script (all ptimes, L16/L24, tone phase-locked to PTP); still open: the Riedel import and a Wireshark capture.
• 5. SAP discovery, then syslog, then health/temperatures (one at a time). VLAN split moved to phase 2.
• 6. PTP TimeTransmitter: BMCA roles (slave/auto; the separate "master" role was dropped later, auto with a low priority1 does the same), hybrid mode.
• 7. Sources: HLS player, then cspot (Spotify Connect), then failover + /api/player.
  • 7.1-7.4 HLS plays on AES67 (PSRAM ring, TS demux + AAC, 44.1 -> 48 kHz). Tested with Triple J Hottest (TS, AAC-LC 44.1k).
  • Come back to HLS (open items):
    • Clock drift: the station's encoder clock vs our PTP clock is not corrected. Starts 3 segments (~30 s) behind live, so it shows after hours/days (skip when falling out of the live window, or buffering). Fix: steer the 44.1 -> 48 kHz ratio by a few ppm from the distance to the live edge / ring level.
    • Download speed ~1.4 Mbit/s over TLS (fine for ~250 kbit/s; tune buffer sizes / per-chunk overhead).
    • Not yet tested: HE-AAC variant (140k), other stations, fMP4/ADTS-only playlists, discontinuities (#EXT-X-DISCONTINUITY), network loss and recovery, long runs.
    • Audio starts only after PTP lock (~20 s after boot): intended, TX needs PTP.
  • cspot (Spotify Connect): login, audio, pause/skip/seek, app volume, live mode switching work; stays connected across track changes and past 6 min (7.5b). Fixed so far: logger NULL crash, Vorbis symbol clash, volume starting at 0, reconnect race (PR #3 patch), 6-minute AP resets (delayed Pong patch), app dropping the device (notifyAudioReachedPlayback). Open:
    • OTA with a session: now ends the session first; 2 of 2 uploads with a running session were clean since. Flash writes still stall TX > 20 ms during uploads (stream stutters while updating).
    • First connects sometimes fail ("Can't connect to spotify servers"), a retry works.
    • Internal heap drops from ~365 KB to ~231 KB with a session; check what can move to PSRAM.
    • Track display in the app can switch ~3-5 s early.
    • Mute at 0 % volume not yet confirmed.
    • Consider offering the delayed-Pong fix upstream (philippe44/cspot).
    • An old, long-idle session can get out of sync with the app: it sends empty Load frames ("No tracks in frame") instead of Pause/Play, so controls do nothing. Quitting and reopening Spotify on the Mac fixes it.
  • failover (auto mode): no session or (with failover_on_pause) paused for failover_delay_s -> HLS; Spotify playing -> Spotify within ~1 s; 30 ms fades; HLS suspended while Spotify plays; cspot is held back (not drained) while it isn't Spotify's turn, so it resumes where it paused. Verified: no session -> HLS, play -> Spotify, pause stays (on_pause off), pause -> HLS after 5 s (on_pause on), play -> Spotify at the paused position.
    • Open: aes67_tx resync (TX 20-27 ms late) ~11 s after each switch to Spotify (seen 3 times); clicks at switches not yet checked in a recording; HLS start sometimes hits CDN read timeouts.
  • /api/player: GET, transport, seek, volume (the app follows), source override and url (runtime, not saved); web UI progress bar.
• 8. Mono sum, gain, polish.
  • Mono sum: stereo source mapped to a 1-channel stream in the core ((L+R)/2, 64 bit). Checked: 156 B packets, SDP L24/48000/1.
  • Gain: source.gain_db as a trim on top of the volume, saturating. Checked: -12 dB = -12 dB peak, +12 dB clips at 0 dBFS.
  • Pink noise source (-18 dBFS RMS, octaves flat within 0.4 dB 63 Hz-4 kHz).
  • 48 kHz only (96 kHz dropped: all sources are 48 kHz).
  • Network: status mask/gw/dns, UI shows the lease; static IP (reboot on change, validation).
  • Web UI: sticky save bar, dark default + light toggle, progress bar, PTP panel (2 min average offset, details), uptime d/h/m/s, VLAN option hidden, no hw_ts / preferred-TimeTransmitter options.
  • Versioning <version.txt>-<git hash>; tools/flash_board.py (tested on this board); docs/user-guide.md.
  • Mute at 0 % volume: confirm on the board.

Phase 2 (parked)

• Replace the Espressif binary audio libraries (GPL-3 conflict, see THIRD_PARTY.md): AAC via opencore-aacdec (Apache-2.0, in bell/external), 44.1 -> 48 kHz via the Speex resampler (BSD). Do together with / right before the clock-drift correction (Speex can steer the ratio).
• Clock-drift correction (do before AirPlay): steer the 44.1 -> 48 kHz converter by a few ppm from the ring level / distance to the live edge, so push sources that run on the sender's clock don't slowly over- or underrun. Closes the HLS clock-drift item in step 7; AirPlay needs the same.
• AirPlay 1 (RAOP) as a source: _raop._tcp via our mDNS, RTSP control, AES (mbedTLS), ALAC decode, RTP audio + timing, 44.1 kHz into the existing converter, uses the clock-drift correction. Start from philippe44's RAOP code in squeezelite-esp32 (components/raop, checked 2026-09-27 at commit 1d542bd):
  • Take raop.c, util.c (MIT, philippe44) and rtp.c (MIT, James Laird/shairport); keep their headers. Write our own glue to the player (their raop_sink.c has no licence header and the repo has no LICENSE file; it's specific to squeezelite anyway).
  • ALAC: they link a prebuilt libalac.a; build Apple's ALAC from source instead (Apache-2.0, already in external/cspot/cspot/bell/external/alac).
  • dmap_parser.c (track metadata) has no header: confirm its origin (likely the MIT dmap-parser project) or leave it out.
  • raop.c contains the AirPort Express RSA private key (reverse-engineered in 2011, in every AirPlay 1 receiver): known grey area.
  • Code is written for squeezelite-esp32's platform (pthreads, their logging, OpenSSL/mbedTLS switches, NVS): needs porting like cspot. rtp.c hands each frame to the sink with its play time (NTP-synced to the sender); following the sender's clock is the sink's job, i.e. our clock-drift correction. Covers iPhone/iPad/Mac (incl. Tidal and Apple Music apps) and Windows via AirPlay senders (TuneBlade, Airfoil). Not AirPlay 2 (HomeKit pairing, its own PTP on ports 319/320 clashes with ours). Then: auto mode as a priority list (whichever source plays wins, HLS as fallback). Check internal heap with Spotify + AirPlay sessions (buffers to PSRAM).
• VLAN split: AES67 untagged + internet tagged (inet.vlan_id/pcp), per aes67-core-base.md "Network". Needs a tagged VLAN with DHCP on the switch port. Config group inet and the UI fields exist already; nothing is applied yet. The "VLAN split" checkbox is hidden in web/index.html (<label hidden>); unhide it when this is done.
• NTP (SNTP): servers as hostnames or IPs (e.g. pool.ntp.org, several allowed; names resolved via DNS, re-resolved on failure). Uses: seed the PTP clock with real time before becoming GM (today it starts at 1970), syslog timestamps. Needs a time config group + UI fields (change doc and page together).

Phase 3 (parked)

• Remote access via VPN / Tailscale. There is no official Tailscale client for ESP32; options to evaluate first: (a) no firmware change: a Tailscale subnet router on the LAN (e.g. the dev PC or a Pi) advertising the device's subnet; (b) WireGuard on the device (e.g. the esp_wireguard component) to a WireGuard server or a Tailscale/Headscale-compatible peer. Before exposing the web/API remotely: add authentication (bearer token) as noted in the OTA/security section of aes67-core-base.md. Keep AES67/PTP traffic off the tunnel.