The dashboard showed "No Cameras Connected" despite the API returning the
camera:
- middleware.Timeout + http.Server.WriteTimeout (10s) cancelled the
long-lived /api/v1/events/stream every 10s, before any 30s status event
could arrive — so the SSE-fed store never populated. Drop the global
request timeout and set WriteTimeout=0 (closed-LAN kiosk).
- The SPA never seeded from GET /api/v1/cameras (SSE only pushes on change).
Fetch the list once on mount and setCameras().
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The SPA doesn't send X-API-Key, so a non-empty api_key made the dashboard
401 and show no cameras. Default api_key to "" (no auth) for the closed
travel-router network, consistent with anonymous MQTT. Document the kiosk
decision, the GoPro Hero 3 protocol, and the gotcha that the pull updater
deploys only the binary (config.yaml must be changed on the Pi).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Validated against a Hero 3 Silver:
- fetchStatus() now GETs the status endpoint /camera/se (was /bacpac/SH?p=%01,
which *started recording* every poll), at the correct host 10.5.5.9.
- Read the response from the stream, not getString(): the blob is binary and
starts with 0x00, which truncated the Arduino String to empty.
- Offsets: recording = byte 29 (confirmed by not-recording vs recording diff),
battery_raw = byte 19 (drains with charge; calibrate on the hub),
video_remaining = bytes 25-26 (provisional).
- Default config set to this camera (goprosilver-1 / 10.5.5.9); per-camera
values can still be overridden at runtime via set_config.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Camera rig-86d978 registers + lists in the API/dashboard with status
ingested. Add decisions for modernc/sqlite datetime scanning and legacy
camera-id migration.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
modernc/sqlite returns a COALESCE() expression as a raw string (no column
type affinity), which can't scan into *time.Time. Drop the COALESCE on the
timestamp and scan the plain DATETIME column (which modernc returns as
time.Time) through sql.NullTime, so a camera with no status row yet lists
with a zero time instead of erroring out the whole list.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Three bugs surfaced once the camera reported in:
- ListCameras LEFT JOIN returns NULL status columns for a camera with no
status rows yet, which failed scanning into non-nullable int/time fields
(recording_state, online, recorded_at) and emptied the whole list.
COALESCE them (recorded_at falls back to the camera's created_at).
- handleHeartbeat rejected every heartbeat ("cannot unmarshal number into
string") because the node sends a numeric millis() timestamp. The handler
doesn't use it, so drop the Timestamp field and let it be ignored.
- handleAnnounce kept a stale cam-NNN row registered by MAC under the old
(pre-self-id) scheme, so self-id status inserts hit a FOREIGN KEY error.
When a MAC is known under a different id than the node's self-id, migrate:
drop the old row and re-register under the self-id.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Living context + decision log for humans and LLMs: architecture, network,
repo workflow, hardware pin map, firmware behavior, pull-based CI/CD,
key decisions/gotchas, current status, and handy commands. Cross-links the
deeper docs/ references.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The publish step died with "fetch failed: ECONNRESET" mid-run, leaving a
half-created release (no version.txt asset → the Pi got 404s). Wrap the
Gitea API calls in a small retry (rfetch) so a flaky connection doesn't
leave the rolling release incomplete.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Auto-registration never completed: the firmware announced on the wrong
topic, the hub never replied, and an unregistered node couldn't receive a
reply anyway. Switch to self-assigned IDs:
firmware (esp32-mqtt-bridge.cpp):
- camera_id defaults to the device id (clientID, e.g. rig-86d978)
- always subscribe to <id>/command; announce on the contract topic
remoterig/cameras/<id>/announce (was the unmatched announce-<id> form)
- drop the bogus numeric timestamp from status (node has no clock)
hub (subscriber.go):
- handleAnnounce registers new cameras under the node's self-assigned id
(no cam-NNN, no registered reply)
- handleStatus tolerates an empty/invalid timestamp and stamps server-side
(previously rejected the status outright)
docs/MQTT_CONTRACT.md updated to match.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
cp over /opt/remoterig/remoterig fails with "Text file busy" once the
service is running. Copy to a .new file and rename over the target
(works on a live binary), in both the deploy and rollback paths.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
A release tag named "dev" collides with the dev branch, making refs
ambiguous ("refname 'dev' is ambiguous") and breaking git push/checkout.
Publish the rolling build to tag "dev-latest" instead; pi-update.sh pulls
from there.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The service unit hard-defaulted to User=pi, but not every Pi has a 'pi'
user (e.g. this hub uses 'overseer') — systemd then fails with 217/USER.
Default SERVICE_USER to ${SUDO_USER:-pi} so the service + /opt/remoterig
ownership match the actual operator. Override with --service-user.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The 3-job ci.yaml re-cloned actions/checkout from github.com per job, and
those clones intermittently fail with connection resets (build job died
there even though lint/typecheck/test passed). Collapse to a single job:
one checkout, then lint -> typecheck -> test -> build. Fewer github.com
clones, faster, less flaky.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
lint/typecheck/test pass; the build job failed only on
actions/upload-artifact@v4, which Gitea Actions doesn't support. Drop the
artifact upload and the placeholder production deploy job (the real build +
deploy is build-dev.yaml's pull-based release). Keep lint, typecheck, test,
and a build compile-check.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
build-dev failed with "go: command not found" — the go-react image ships
Node but no Go. Restore setup-go (its static Go binary runs on this runner,
unlike setup-node's dynamic Node), and keep Node from the image.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On the go-react runner image, npm/node/go already work (npm ci succeeds),
but the setup-go/setup-node actions install tool-cache binaries that can't
execute on this runner (node/22.22.3/x64: "cannot execute"), failing the
jobs in their post steps. Drop those actions and use the image's built-in
toolchains; also reduces flaky github.com action clones.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The workflows used runs-on: ubuntu-latest, which mapped to
docker.gitea.com/runner-images:ubuntu-latest — an image whose Node from
setup-node won't execute (exit 127) and which lacks curl/jq. The runner
already advertises a purpose-built "go-react" CI image; point the
workflows at it instead.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The build-dev publish step failed with exit 127 — the act runner image is
minimal (no curl, jq, or sudo; runs as root). Node is always present
(setup-node), so do the release publish in Node using built-in fetch/crypto
and FormData/Blob for the asset upload. No external tools needed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
cmd/server/main.go has //go:embed all:src/dist (relative to cmd/server/),
but Vite built to repo-root dist/, so cmd/server/src/dist never existed and
every `go build` failed with "pattern all:src/dist: no matching files found".
The hub binary has never built in CI as a result.
- vite.config.ts: outDir -> cmd/server/src/dist (emptyOutDir)
- commit cmd/server/src/dist/index.html placeholder so the embed always has
a file (real build overwrites it)
- .gitignore: scope dist ignore to /dist; ignore cmd/server/src/dist/* but
keep the index.html placeholder (the prior !src/dist/index.html rule
pointed at the wrong path)
- ci.yaml: upload artifact from the new output path
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Pi is on a closed travel-router LAN, so push-based deploy from a
runner can't reach it. Switch to pull: the runner builds + publishes,
the Pi fetches.
- build-dev.yaml: after the arm64 build, publish the binary + sha256 +
version.txt to a rolling "dev" Gitea release (replaces the
upload-artifact + repository_dispatch -> deploy-dev hop)
- remove deploy-dev.yaml (push/scp-based deploy no longer used)
- scripts/pi-update.sh: poll the dev release, verify sha256, install via
deploy.sh (backup/restart/rollback); only updates when version changes
- scripts/remoterig-update.{service,timer}: run the updater every 5 min
- setup-pi.sh: install deploy.sh + pi-update.sh + update.env template +
the updater timer; summary now reflects the pull flow
- README: document the pull-based CI/CD; fix stale GOARM=6 (Zero 2 W is
arm64 on 64-bit OS / arm GOARM=7 on 32-bit)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Per maintainer decision: this is a private repo and the closed
travel-router Wi-Fi password is low-sensitivity, so keep the real
value in the tracked config for reproducible uploadfs provisioning.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The project was designed around a 10.60.1.0/24 travel-router network,
but the actual RemoteRig router uses 192.168.8.0/24 (the C6 associates
and gets 192.168.8.x; hub confirmed at 192.168.8.56). Replace the
network prefix everywhere (last octet preserved; GoPro 10.5.5.1 left
alone).
- scripts/setup-pi.sh: static IP 192.168.8.56/24, gateway 192.168.8.1,
deploy/health command examples updated
- esp32-mqtt-bridge.cpp: default mqtt_broker -> 192.168.8.56
- firmware/data/config.json: broker -> 192.168.8.56 (wifi_password kept
blank in git; real value flashed to the device only)
- docs (CONTEXT, MQTT_CONTRACT, READMEs, wireframes): gateway/hub/DHCP
and example IPs re-addressed for consistency
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The C6 never loaded its /config.json, so it fell back to defaults
(SSID RemoteRig, empty password) and couldn't join Wi-Fi. Two bugs:
- Data file was named esp32-config.json but the firmware reads
/config.json → renamed to config.json.
- Firmware used SPIFFS while pioarduino's uploadfs builds a LittleFS
image; the SPIFFS mount then reformatted it empty. Switch the C6 to
LittleFS (matches the toolchain default and the ESP-01S).
Also:
- log loaded ssid/broker/camera_id on config load (not the password)
- platformio.ini: land the ESP-01S env retarget (board d1_mini ->
esp01_1m, dout, upload_speed 115200) that was missed in 403e1d9
- committed config.json keeps wifi_password blank; the real value is
flashed to the device, not stored in git
Verified: C6 loads config and associates (got a DHCP lease). MQTT to
the broker is a separate network issue (hub IP / subnet).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Updating the buried ESP-01S currently means a USB-UART adapter and a
GPIO0 jumper. Add a path to change its settings without reflashing, and
lay the groundwork for full firmware updates over the existing UART.
set_config (no reflash for settings):
- ESP-01S: add saveConfig() + a set_config command — updates GoPro
SSID/password/IP and poll interval, persists to LittleFS, acks, and
re-associates Wi-Fi if creds changed
- XIAO: forward an MQTT set_camera_config down to the ESP-01S over UART
(hub -> MQTT -> XIAO -> UART -> ESP-01S/LittleFS)
UART-OTA groundwork ("XIAO as flasher"):
- reserve XIAO GPIOs ESP01_RST_PIN=D8, ESP01_PGM_PIN=D10 for driving the
ESP-01S serial bootloader (not driven yet)
- docs/design/esp01s-uart-ota.md: full design (why Wi-Fi OTA doesn't fit
the 1MB ESP-01S on the GoPro AP, bootloader entry, ROM flash protocol,
HTTP-pull delivery, scope)
- hardware/README.md: fix stale ESP32-C3 -> XIAO ESP32-C6 wiring, add the
two control lines (Notion wiring diagram updated to match)
Both firmwares build clean and are flashed; set_config round-trip needs
the broker to exercise end-to-end.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Battery calibration:
- two-point linear cal (bat_raw_min->0%, bat_raw_max->100%) of the
GoPro offset-57 raw byte, persisted in SPIFFS config
- publish battery_pct in MQTT status only when calibrated (omit
otherwise, per MQTT_CONTRACT); OLED shows % when calibrated, raw
until then
- set_battery_cal MQTT command: explicit {raw_min,raw_max} or
capture-current {point:"full"|"empty"} for field calibration
RGB STAT LED:
- drive D0/D1/D2 (R/G/B) with health colors instead of the single
green channel: red=offline, magenta=wifi-but-no-hub,
yellow=hub-but-no-camera, green=healthy; blue during boot
- RGB_COMMON_ANODE polarity flag; this module is common-anode
Verified on hardware: boots, OLED ok, RGB shows correct colors
(blue->red on the bench).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bring up the 1.3" SH1106 128x64 I2C OLED on the XIAO ESP32-C6
(D4/SDA, D5/SCL @ 0x3C) per the Notion wiring diagram.
- add U8g2 dependency to the seeed_xiao_esp32c6 env
- I2C bus scan at boot (logs responders to serial)
- boot splash + live status screen: camera id, IDLE/REC + session
timer, battery (raw until calibrated) + video-remaining, hub link
state (MQTT/wifi/offline), and camera reachability
- refresh runs at the top of loop() so the panel stays live even
when WiFi/MQTT are down
Verified on hardware: I2C scan finds 0x3C, U8g2 begin ok, panel
shows clean readable text.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The MCU changed from ESP32 Dev Board to a Seeed Studio XIAO ESP32-C6,
but the firmware still targeted esp32dev. Retarget it and fix the
build so it compiles and flashes for the C6.
platformio.ini:
- env esp32-mqtt -> seeed_xiao_esp32c6 on the pioarduino platform fork
(mainline espressif32 lags the Arduino-core 3.x the C6 needs)
- add ARDUINO_USB_MODE=1 / ARDUINO_USB_CDC_ON_BOOT=1 for Serial over
the C6 native USB
- fix build_src_filter ordering in BOTH envs: -<*.cpp> ran last and
re-excluded the target, leaving setup()/loop() undefined at link
esp32-mqtt-bridge.cpp:
- UART Serial2 RX16/TX17 -> Serial1 RX=D7/TX=D6 (XIAO C6 mapping)
- status LED GPIO2 -> D1 (green channel of the RGB STAT LED)
- fix pre-existing ArduinoJson v7 / PubSubClient build errors
(.c_str() on a const char*, String topic where const char* required)
Verified: builds clean and boots on hardware (native-USB serial banner
confirmed).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
- Add migration 002: UNIQUE index on status_logs(camera_id, recorded_at)
- Upgrade migration system to version-tracked (schema_version table)
- Prevents race-condition double-inserts that application-level COUNT(*) check cannot guard against
- Complements existing application-level dedup from CUB-230
- Add request validation: Content-Type check, body size limit (64KB)
- Add field length validation (camera_id: 64, friendly_name: 128, mode: 32, resolution: 32)
- Add FPS range validation (0-240)
- Add battery_pct range validation (0-100)
- Replace ad-hoc map[string]string errors with structured APIError {error, code, details}
- Fix isUniqueConstraintErr to catch both camera_id and mac_address constraint violations
- Fix MacAddress model field from string to *string for NULL handling
- Fix splitSQL to strip -- line comments before splitting (was causing migration failures with modernc.org/sqlite)
- Add 30 integration tests covering all endpoints
- All tests pass: ok github.com/cubecraft/remoterig/internal/api
Joshua King
overseer
Sketch
Hermes