Keep internal storage and the separately mounted recording drive distinct.
Gate added-drive telemetry and both views to C3, recheck filesystem identity,
and report available bytes without counting reserved space as free.
Cover absent drives, mount transitions, read failures, zero free space,
C3X/C4 exclusion, and actual Vue rendering. Refresh the inventory receipt
for JavaScript test-only protected changes; native artifacts are unchanged.
Jetlink v0.9.0 keeps its own index of comma's big models and the ONNX
each was built from, checked hourly. Active Big's choice, by the ONNX's
hash, is sized from it rather than from comma's LFS, so JetlinkPointers
is no longer read and is dropped.
JetlinkPointers leaves params_keys.h: the native params libraries need
their usual rebuild.
- A NetworkManager restart left the dedicated lease on stale device paths,
and the rest of the drive fell back to wlan0. A non-timeout aa0 failure
now rebuilds aa0 once before falling back.
- DedicatedLease lacked `dedicated`, so SharedRadioWatch (and the ACK grace
for home Wi-Fi changes) never ran.
- When wlan0 loses its network mid-projection, pause its autoconnect (and
stop a join under way) until the session ends. Each retry took the shared
radio off the car's channel and stalled video past the ACK timeout.
AGNOS 20.x no longer includes hostapd (agnos-builder 0cf5062), so every
start failed and the worker created and deleted galaxy0 every 15 s. On
2026-11-04 that churn crashed NetworkManager mid-drive and dropped StarPort.
- Ship Ubuntu 24.04's hostapd 2.10 and hostapd_cli (arm64, checksums in
bin/hostapd.sha256) and run those.
- Check for hostapd before galaxy0 exists.
- Back off repeated failures from 15 s up to 5 min (DFS refusals exempt).
- Write a runtime udev rule (NM_UNMANAGED) before creating galaxy0, and wait
for wpa_supplicant to let go of it before starting hostapd. Its release
stopped the AP that hostapd had just started.
- Steady the curve-speed glow and the gas-override blink on the on-road screen
- Stop the Theme maker preview flapping to "unavailable" and allow for Galaxy tunnel latency
- Carry the widget sprite sheet in a header so it survives the tunnel
- Retry a saved sign-in when the first attempt at boot fails, for example before the clock has synced
- Let the supervisor restart a player that never logged in
- Add connect tests for the retry
- Add lane guidance to the on-road navigation banner, fed by new route engine fields in the custom cereal schema
- Calm map rotation and steady the navigation layout, including the Galaxy layout preview
- Add route runtime, on-road map and navigation tests
- Use StarPort's default widget layer order in the car projection when no widgetOrder is saved, so the map no longer draws over the PiP
- Make the layout editor and the car screen agree on that order
- Cover the default order in the projection layout and widget layer tests
Hold the mock background upload until in-flight assertions finish and join it on every exit. Preserve strict completion and payload checks.
Type the dynamically populated startup fixture namespace as object and assert its main entry is callable before invocation. Preserve startup ordering assertions and refresh the canonical protected inventory receipt.
Reduce leading JSON indentation only, preserving every token, value, key and list order, reviewed reference binary and provenance pin. Keep the unchanged 120 KiB lint limit and all validation gates.
Refresh the canonical protected inventory receipt for the formatting-only file change.
Pin seven independently reviewed replay outputs to the unchanged recorded inputs and baseline provenance. Correct Subaru test helper cleanup without changing native admission rules. Declare the exact generated pandad test executable in the build-output inventory.
Run the unchanged whole-target RELEASE workload in sixteen mandatory shards, retaining all cases, assertions, skip and provenance gates, and existing deadlines. Refresh the canonical protected inventory receipt.
Native qualification completed 10,211 passed and 115 existing skips across all sixteen bins; fresh source-bound hosted CI remains required.
Select the real headless backend for hosted AOL and RELEASE tests. Isolate SpacePilot scripted clocks from unrelated UI readers. Update camera and replay fixtures for existing native APIs, pin the reviewed schema addition, and supply declared CAN batch and throttle inputs.
Preserve production guards, acceptance assertions, selected cases, capability skips and receipt gating. Refresh the canonical protected inventory receipt.
Preserve the exact published Toyota source blob and restore its prior100755 executable mode. Refresh the canonical protected inventory receipt. No source semantics or test expectations change.
Re-audit only the GPU/map source and shared fixture fingerprints following the native binding migration. Preserve the exact capability guard, reason, selected nodes and unavailable disposition. Add a real-source fingerprint regression so mocked gate fixtures cannot hide stale source pins.
Discover exact OS/DHCP/PTR identities independently of request Host. Redirect arbitrary aliases only at the direct-local inert landing document, preserving Host, Origin, session and remote boundaries. Refresh the protected-inventory receipt for the documentation-only protected change; native inputs are unchanged.
- When the comma's Wi-Fi joins or roams while the car is on aa0, wait up to 8 s for video acks instead of 2.5 s; reconnect at once if the car takes its screen back.
- After a drop it causes, the car screen shows "Reconnected: comma Wi-Fi Joined <network>" (the network's SSID) for 8 s; Galaxy and bug reports name it too.
- Logs ack_grace, ack_grace_recovered and other_wifi_changed events.
- Link the bridge and discovery units with --no-reload and reload once (was four reloads, 1-3 s each on a comma four).
- Wait (up to 45 s after boot) while starpilot-bluetooth-radio is still activating, so the reload can't stall btattach or bluetoothd.
- Failure handling unchanged; on a comma four this removed the ~20 s Bluetooth boots (now 16.4-16.6 s with stock AGNOS).
The selected adapter or app segment is solid purple with a soft glow instead
of grey, the selected connection option gets a purple-tinted row and a filled
icon tile, the current step and the primary button glow, and a faint purple
wash sits behind the step track.
Choosing a USB cable or a tablet app switches the connection at once, which
counts as setup done, so the card closed under the person's finger and left
no way to change the choice. A choice made in the card now holds it open
until Done. In the guided setup the home card's button reads Change how your
car connects, so the way back is visible.
Done steps and checks use the theme's success color, the current step and
selection use the primary purple, text uses Galaxy's Inter and weight tokens,
and the connection options use the outline icons the rest of the page uses
(car-front, usb-symbol, usb-c, tablet) on a single primary-tinted tile
instead of Apple's multicolour gradients.
The guided card's lists, paired-car rows, checklist rows and handshake
disclosure now use the page's own gx-starport-list, gx-starport-row,
gx-starport-row__text and gx-starport-wiki classes, restyled once for the
card, instead of parallel guided-only copies. Radii, spacing and motion use
Galaxy's tokens. Only the step track, icon tiles, radio rows, segmented
control and checklist marks remain new, and the block is about a fifth
shorter.
The guided card is rebuilt in an Apple-style: a three-step track (support,
car, connect) that fills as setup moves along, a large title, inset grouped
lists with coloured icon tiles for how the car connects, a segmented control
for the adapter model and the tablet app, iOS-style checklist rows for the
connection, and one full-width primary button. It uses Galaxy's theme colors
so light and dark both work, and reads cleanly down to a phone.
The classic step 2 and the home card's dropdowns now render only in the
classic setup. The guided card replaces them and hosts the page's own
pairing, cable and app instructions and the Connect button, so nothing is
lost. A finished setup can be reopened from Setup and Pairing, and Done
closes it. Paired cars the page's own list does not show yet are offered in
the card; the rest stay in the pairing step.
The logs bundle, and so the Discord send, now carries report.json: the same
tags and per-session detail a Bugsink report has (setup, offered modes, reach
class, ladder, SDP, strategy) for the newest five sessions, sanitized and
size-capped. It is built best-effort: if it cannot be, the bundle is
unaffected.
The adapter setup's 'Connect your paired car in Galaxy Bluetooth first'
refusal (27 Bugsink reports) now shows a step with a link to Bluetooth, next
to the usual question while setup is unfinished and alone once it is done.
The comma can stay powered between drives, so the three passes belong to an
ignition (an onroad start), not to a boot; a reboot still resets them. When a
learned strategy is dropped after two handshake failures the ladder carries
on from the next rung instead of starting at the one that just failed, which
the dongle flip also avoided.
Choosing the only paired Android Auto car from housekeeping runs with nobody
at the screen, and selecting a receiver unselects that device as the
Bluetooth audio output. It now skips a car that is the selected audio device,
or when that cannot be read, leaves it to the setup screen's Use button. It
reports problems once instead of every 20 s and gives the projection role
back, keeping only the standby gateway auto-connect needs.
The 686 pre-video failures Bugsink showed (54 distinct texts) are kept as a
fixture. The test checks the intended class for each: Host is down and page
timeouts are a sleeping car, refused and unanswered handshakes are not ready,
and anything after the link is up is not a reachability problem. Reports tag
the reply style and Wi-Fi band the working strategy used.
Reports carry the SDP outcome (where the RFCOMM channel came from and how
many retries it took). The Discord post names the handshake options tried
and how each ended. A car that once needed 70 s to be ready now gets that
many stage timeouts to answer the Wi-Fi handshake, never fewer than the
standard wait and capped at nine.
A card above the classic steps asks one question (how does the car connect:
built-in wireless, adapter, USB cable or app), shows only the instructions
for that answer, and lists paired cars that offer Android Auto to choose
between when there is more than one. While connecting it shows a live
checklist of the stages (Bluetooth, the car's Wi-Fi, securing the connection,
the dashboard), a list of the handshake options being tried once one has
failed or one is remembered, and a notice that the first connection can take
a few minutes or more than one drive. Every action is one the page already had.
The card is on by default; Use classic setup turns the switch off, and Try
the guided setup turns it back on. Returning drivers see the card only while
a connection is in progress. The state function is pure and tested without a
browser.
The setup response carries newSetup, and a new_setup control action turns the
guided setup off or on. It is handled before the StarPort-enabled check and
never calls the daemon, so it works as the way back to the classic screens
when the service is off or broken. It needs a signed-in session and park.
One module now decides how to answer a car's handshake, and it is also the
learned profile: the strategy that last worked is the one tried first.
The first pass is today's behavior. Later rungs are knobs that already
existed: the other version reply (the flip the dongle work added, now owned
here), the phone-style reply with both bands and with 2.4 GHz only, a longer
wait for a car that keeps saying not ready, and, for adapters only, the
handshake variations.
A failure advances the ladder only when the car answered the version
exchange, then went silent or refused the Wi-Fi setup, before any video was
acknowledged in the session. A sleeping car, this comma's own Bluetooth
trouble and anything after video (a car shutting off) never move it. Three
passes per boot at most, resuming at the first untried rung on the next
ignition; a learned strategy is dropped after two handshake failures; a
session that streams for a minute makes its strategy the learned one.
State is a small per-car file beside the config (setup_state.json, newest 16
cars, damaged or unreadable means a fresh start, never blocks a connection).
The status exposes the ladder for the setup screen, and reports carry what
was tried, how each rung ended and how long the car took to be ready.
Allow H.265 video (Developer > StarPort, off by default) lets a car that
offers only H.265 get H.265. A car with an H.264 mode is chosen exactly as
before whatever the setting; this is replay-tested over several offers.
The hardware encoder library is ABI 4: an HEVC option selects H.265 Main
profile at high-tier level 5 and allows up to 1920x1080 (H.264 stays at
1280x720). The Python side keeps H.265 apart from the H.264 code: its own
codec-config and wire-frame functions for the two-byte NAL header, its own
access unit check, and an encoder subclass. If the car is told H.265 and the
hardware encoder cannot open, the session fails with the reason; there is no
fallback to another codec.
Measured on a parked comma with a scratch copy of the library: H.265 1080p
costs about 13 ms per frame on the NV12 path (30 ms from RGBA), mid-stream
keyframes carry VPS/SPS/PPS, and the H.264 session is unchanged. The
encoder splits every H.265 picture into two slices, which the check allows.
A car that is paired and offers wireless Android Auto but was never chosen
(the pairing window closed first, or it was paired in Bluetooth settings)
is adopted by housekeeping when it is the only candidate. This reads the
cached BlueZ service list, so it never pages a device, and it does not run
during a pairing window or session, for adapter setups, or once a car is
chosen. With several candidates nothing is guessed; they are listed in the
status as car_candidates for the setup screen to offer.
The first version waited on any state below disconnected, which also caught
the unknown state and broke the lease tests. Wait on states 10 and 20 only.
Nineteen Bugsink joins were refused with 'device is strictly unmanaged'.
On a shared interface the join now waits up to 8 s for NetworkManager to
manage it (dedicated mode already waited) and otherwise reports that the
interface is unmanaged and for how long, which the diagnosis already maps
to wifi_not_ready.
Stream losses keep the projection_interrupted category but now carry a kind
(stalled, silent, disconnected) with matching next steps, reported as
aa.loss_kind, so a car shutting off is no longer read as link loss.
The Developer page gets a heading row, StarPort, that groups the StarPort
settings: the New pairing experience switch (on unless turned off), Stay on
Wi-Fi (moved under the heading, same key, label and default) and Allow
H.265 video (off by default, daemon setting allow_hevc).
The heading is a row whose key starts with heading:, projected with
heading=true. It has no choices, is never available and cannot be changed,
and Galaxy draws it as a label. The new pairing switch is a plain file next
to the other StarPort data rather than a Params key, so it reads and changes
with the StarPort service down and needs no device rebuild; absent or
unreadable means on.
Most failed drives in Bugsink never reached video and stopped at Bluetooth,
but one backoff treated a sleeping radio, a service that was not up yet and
this comma's own busy stack alike.
reachability.classify() separates a car that never answered a page (Host is
down, page timeout), a car that answered but refused or reset, and a local
Bluetooth failure. A sleeping car or busy controller now waits at least as
long as before and often longer; other failures wait exactly as before. A
sleeping car no longer makes the comma forget the cached RFCOMM channel.
The class is logged on attempt_failed and reported as aa.reach_class.
The 141 'unknown' failures at the discovering stage now get a diagnosis:
car_asleep, car_not_ready or bluetooth_stack_busy, with next steps.
The Discord post now names the setup type, profile and whether it was
chosen or detected, from the newest session log that recorded one, next to
the car and head unit. The zip already carries the same data in each
session log's session_start and discovery events.
Reports could not tell a built-in unit from a dongle, and offers and
handshakes were only visible in failure text, so there was nothing to
compare working pairings against.
session_start now records the setup type (builtin/adapter/usb/app), the
resolved profile and whether it was chosen or detected. Session reports
carry every video mode the head unit offered, whether the comma's address
reached the car in the pairing request, and, for streaming sessions, the
strategy that worked (window, reply style, Wi-Fi channel type). The drive
report gains low-cardinality tags (aa.setup_type, aa.adapter_model,
aa.profile_source, aa.phone_address_sent); long values stay in contexts.
Old logs without these fields still report. aa.schema is now 3.
LeaseAwarePhone (used for every wireless session) had no adapter() method,
so Supervisor._adapter_address() hit AttributeError, swallowed it, and
returned "". The Bluetooth-service pairing request was therefore never sent
(bluetooth_opened always logged phone_address: "").
A 2020 myChevrolet (head unit 2.46.1.6) through an AAWireless Two+ then
accepted and acknowledged video but never showed it. A real phone sends the
pairing request about 1 s into the session; the car shows Android Auto once
its hands-free link to that address is up.
Pass adapter() through to the wrapped BluezPhone. It is a read-only property
lookup, so it needs no phone-role lease, like devices() and device().
_adapter_address() now logs adapter_address_unavailable instead of failing
silently.
Tested on the car (session-000052): phone_address sent, pairing response
received, picture displayed, touch input worked, 20-29 fps fully acked.
Co-authored-by: droobah <34214975+droobah@users.noreply.github.com>
Adds a Send logs card to the StarPort Logs page so a user can report a
problem with all of their reportable logs: an optional note of up to 60
characters, and the same zip as Download all logs, sent from the device to
the developer's log inbox. Each device keeps its own thread, titled with the
car fingerprint and a hashed device tag. Manual sends only; no automatic
failure reports. If the zip is over the 10 MB size limit, it is resent
without the device snapshot.
Delegate the isolated Ioniq5PE, Toyota shell and GM pitch mock default lookups to real Params metadata. Preserve existing assertions, no-setting-write checks and runtime behavior.
Delegate mock default lookup to Params metadata and supply the existing healthy UI fixtures with their post-start receive frame and onroad state. Preserve all runtime code, assertions and negative authority coverage.
Feed exact emitted steering payloads and their accepted/rejected host-bus receipts on the next parser cycle. Preserve all assertions and runtime behavior.
Keep existing assertions and runtime behavior unchanged. Model healthy transport by returning only actually emitted steering commands; retain missing, stale, wrong-bus and rejected receipt coverage.
Preserve SLC validation behavior with explicit None narrowing, use existing Mock test doubles, and restore controlsd executable metadata without content changes. Keep exact torque gain expectations while matching the existing list representation.
Validated 1095 related tests, targeted ty and Ruff; prior baseline comparison and 4236 guard-equivalence cases preserve existing rejection semantics. Refresh the inventory receipt only for these validated changes.
Use one saved schedule with editable zero defaults. Preserve current schedules and round displayed offsets and edits to whole units without conversion drift.
Re-vendored at the v0.2.1 release (2b9dc6c), which adds StarPilot as a host.
Since the last copy: the matrix tests only re-key SAFETY_PARAM when the host's
stock base defines it (IQ.Pilot's iqdbc doesn't; StarPilot's does, so no change
here), and a FORKS.md note.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The previous commit let the camera's actuator through again as soon as one host frame was rejected. Mid-steer
that put the camera's inactive frame (its own counter) on the bus in place of the blocked one, and
the F-150's PSCM ramped out and reported lateral control faulty for several seconds, from a single
blocked frame. Without AOL, static blocking makes a rejected frame a missed one, which the PSCM rides
through.
Block the camera's copy while the host is sending, accepted or not; only a host that has stopped
sending for the AOL host timeout gives the actuator back to the camera.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
With AOL, the camera's actuator (0x3D3/0x3D6/0x3CA) is forwarded unless AOL owns steering, but the
host sends its own every frame, active or not. When AOL isn't steering, the PSCM hears both, with
unrelated counters. On a 2021 F-150 this flipped LatCtlSte_D_Stat between Unavailable and Available
every 50-100 ms for seconds at a time, worst when the two counters ran 1-2 apart: steerFaultTemporary
alerts, the panda's LKA-ready check denying lateral, and AOL dropping its latch on re-engagement.
BluePilot's angle-mode release during a manual turn made it easy to hit.
Block the camera's actuator while the host's last frame was accepted and is within the AOL host
timeout, as static blocking does without AOL. A rejected or stale host frame lets the camera's
copy through again. Permission and display ownership are unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
On the Steering page, for every Ford that can run BluePilot lateral (not the Transit), in Galaxy and the
native settings alike:
- BluePilot Lateral (off by default), Primary Control (curvature/angle) and Use Pinion Yaw Sensor (platforms
with a pinion geometry row) change while parked; the switch and the pinion sensor apply at the next vehicle
startup, and Safe Mode keeps BluePilot lateral off.
- the tuning rows of the current mode change any time, since the library reads them every frame; rows that
depend on Lane Positioning or the custom profile wait for it.
Ranges come from the library's manifest and defaults from params_keys.h, so the rows can't drift from what the
controller reads. A change is checked against the row it was made on, so a request from before its switch
changed is refused.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- params_keys.h: every setting in the library's manifest (python -m opendbc.bluepilot_lateral.params.manifest),
without BACKUP, which StarPilot's params don't have. disable_BP_lat_UI defaults to "1": BluePilot lateral
is off until the user turns it on.
- car init (VehicleStartupPreferences.prepare): init_ford_safety_param_starpilot ORs the library's startup
bits into the Ford param. They stay off without the openpilot toggle, in Safe Mode, or when a startup
setting can't be read.
- card (configure_controller): the Ford CarController gets a BluePilotInputs, which fills the library's
LateralInputs from modelV2, vehicleParameters and lateralDelay each frame and re-reads the per-frame
settings every 3 s.
- test_bluepilot_lateral.py: the init path's bits, the real CarController and libsafety with the same
param, for a CAN and a CAN FD Ford, off/on/on+pinion/pinion-only, curvature and angle mode: every
steering and LKA frame passes, and only angle mode with BluePilot on steers through the path angle.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- safety/modes/ford.h: includes the library's safety with StarPilot's lateral permission
(lateral_controls_allowed(), which includes AOL) as its lateral binding. ford_init hands the library
bits 9-14 of the Ford param (never on the Transit); with the bit set, LateralMotionControl/2 go through
the library's 4-signal checks instead of the stock ones, Lane_Assist_Data1 carries the angle-mode state,
and the pinion source replaces yaw and joins the selected RX list (stock, stock-switch or AOL).
- car/ford/carcontroller.py: one BluePilotLateral per Ford; while it's active it sends the steering and
LKA messages, otherwise the stock path runs unchanged. StarPilot's openpilot side attaches the per-frame
inputs and settings as bluepilot_inputs.
- safety tests: test.sh runs the library's safety tests by directory and gates the library headers'
coverage in a run of their own; test_ford_bluepilot_hooks.py covers where the bits ride, the Transit,
the pinion RX check on every list and the full-list guard. The qualification and mutation runners
include the library's safety tests.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The CAN FD display fixtures and the stock-switch class left alternative_experience at 32. Now that the
stock words are AOL words, a later Ford test could run with AOL configured; they reset it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
opendbc_repo/opendbc/bluepilot_lateral is BluePilotDev/bluepilot-lateral as released, made by the
library's tools/vendor.sh and checked by its tools/check_vendored.sh (VERSION names the release and
commit). Nothing inside the folder is edited here: changes go to the library, get released, and are
vendored again.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
StarPilot's angle-only bluepilot_lateral adaptation and the BluePilot-derived curvature strategies (Mach-E,
classic, Explorer, generic CAN FD, new-port curvature, manual-turn latch, curvature preview, pinion
measurement) are removed, with their extended safety profiles 16/32/64 and the interim angle word 128. Ford
lateral is upstream's stock curvature path until the BluePilot lateral library is mounted.
StarPilot's own Ford features stay, re-keyed from the extended words to the stock ones:
- AOL: classic 32 -> 0, Mach-E 18 and the CAN FD trucks 66 -> 2 (+1 with openpilot long); the CAN FD
retained-display rules apply to word 2. Edge/Mondeo/Transit and the manual bit (256) are unchanged.
- release CAN FD long: word 3 (was 19/67); 19 and 67 are now retired and denied.
- stock cruise switch, CAN FD camera-lead radar handling, hands-free cluster: keyed on the Ford family.
- classic stock ACC (0) keeps the no-ACCDATA TX list the extended classic profile had.
Bits 9-14 of the Ford param are reserved for the BluePilot lateral library's startup bits
(safety/modes/ford_param.h); AOL and pandad mask them out of the profile word.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Retain real composition and input-routing methods with typed drawing boundaries.
Preserve existing assertions and validate optional settings requests explicitly.
Normalize a same-type local copy instead of modifying the input parameter.
Existing safety logic and all 34 paired firmware artifacts remain byte-identical.
Validated 67 debug and 67 release native cases; refresh source inventory receipt.
With Jetlink running the big model (on, supported, no Chestnut), Active Big
in Galaxy and on the device is Jetlink's choice: the model manager marks a
big model selectable when its catalog entry names the ONNX it was built from
(source_sha256), downloaded or not, since Jetlink's host fetches the ONNX
itself. The choice is the new JetlinkModel param, {sha256, displayName}, the
adapter's big-model key; Chestnut's own Active Big is left as it was. None
leaves the model to the Jetlink app, and the manager words it.
A newer choice made in the Jetlink app is written back into JetlinkModel by
jetlink, so Active Big shows the model that drives; one StarPilot does not
list is named as the app's.
JetlinkModel is a new key in params_keys.h: the native params libraries need
their usual rebuild.
At ignition the head unit pages the comma for hands-free at the same moment
the supervisor paged the car. The crossed connections can leave the
controller holding a link the 4.9 kernel no longer knows ("ACL packet for
unknown connection handle 7"), after which every SDP/RFCOMM connect to the
car fails with EALREADY and the session retries the same sequence until the
car gives up.
- Ignition-triggered sessions wait up to 8 s for the car's own hands-free
link before paging it.
- EALREADY on SDP or the cached RFCOMM channel fails the attempt at once,
keeps the cached channel, drops the stray handle with a raw HCI Disconnect
(only when BlueZ reports the car disconnected and the kernel does not list
that handle), and lets the car connect first on the retry.
The 2022+ Kia Stinger has no radar DBC: the radar's object list is sent on
bus 1 as 10 slots x 3 frames (0x238..0x255, 8 bytes, 50 Hz). Decode it by
hand and publish the objects as radar tracks, gated to KIA_STINGER_2022.
- Stinger2022ObjectDecoder pairs the three frames of a slot (same cycle
counter and object id in all three, completed within 20 ms, the last
frame consumed so it cannot be reused) and decode_stinger_2022_triplet
extracts distance, lateral offset, absolute speed and type.
- Ego speed is the WHL_SPD11 4-wheel average on bus 0 times
wheelSpeedFactor; relative speed is the radar's absolute speed minus it.
- yRel = -lateral_raw (left positive in the car frame). Tracks not seen in
a cycle are dropped. canError is reported while the parser is invalid or
no wheel speed has been seen, and objects decoded in such a cycle are
discarded. The cycle is emitted on the last address, 0x255.
- interface.py: radarUnavailable is False only when all 30 addresses are
present in the bus 1 fingerprint with length 8.
Two measured corrections to the raw speed:
- A lead braking to a stop is reported at about -0.3 to -0.6 m/s for 1-3 s
before settling at 0, while the stock SCC11 lead stays near -0.1 (median
-0.13 over 13 stock-ACC stops, versus -0.45 from the object frames).
radard's KF turns the recovery into a positive aLeadK for a parked car,
which the MPC extrapolates as the lead pulling away. Negative speeds
above -0.7 m/s are reported as 0 for objects within 2 m of centre;
faster reversals, adjacent-lane and oncoming traffic stay as measured.
- Absolute speed minus the wheel-speed average reads ~0.14 m/s more closing
than the stock SCC11 lead (ACC_ObjRelSpd) whenever the car is moving, and
matches it at standstill: the radar's own ego speed runs below the wheel
speeds. The 0.14 m/s is subtracted from the ego speed, ramped in over the
first 1 m/s so standstill is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Clear prior recovery wait state only after the existing healthy unlatched baseline is accepted. Preserve native acknowledgment, fault, identity and expiry gates; cover actual Card and native permission transitions.
Preserve every node, grouping rule, shard count and timeout. Calibrate material module estimates from complete hosted outcomes with conservative scheduling margin.
Reuse the existing original proof deadline for qualified Mach-E safety19 only. Require fresh independent authority for successors and retain fault, cancel, identity and expiry checks.
Withdraw native-denied LONG immediately while retaining only an already proven LAT grant. Preserve fault checks, original proof deadlines and ordinary mismatch counter.
Classify only the recorded firmware-identified Elantra2021 CAN1 IRQ
bit as diagnostic. Preserve raw reports and native permission, request,
receive health, relay, heartbeat and freshness checks. Alpha-long and
all other configurations remain excluded.
Retain expanded caller/mixed-fault/configuration regression coverage.
Validated 257 focused Elantra and GM/Bolt tests; Ruff clean.
Revert the Bolt-specific admission change from 4b754298 while retaining
native authority, receive-state, relay, heartbeat and actuator checks.
Keep raw diagnostic reports and all other profile policies unchanged.
Add actual Card coverage and restore exact diagnostic policy regressions.
Rebuild all distributed Panda variants and bootstubs from 4cbffe53 with verified Arm GNU12.3.Rel1; preserve development signing convention. Native debug and release builds passed. Refresh canonical protected source/binary inventory receipt. Default-off Ford option is init-only; no vehicle or road execution claimed.
Build nested typed evidence rows directly while retaining every original field and validation. Complete serialized bytes match the prior writer on recorded d8 evidence and regression fixtures.
Fresh-head combined qualification: 459 tests passed. Cloud performance measurements do not claim device loop-rate recovery or a successful learning trial.
Merge bounded control/output FIFO heads using original publisher headers. Preserve original age, native, source, overflow and malformed-input checks; retain producer order after joins.
Validated with actual pinned cloud dependencies and fresh-head regression tests.
Restore the existing Dom lateral-active dashboard icon behavior only for
qualified Elantra 2021 classic-long AOL profiles. Preserve stock-profile
handling, disengagement blinking, warning fields and all actuator/SCC/FCA
packets. Add native-permission and packet-equivalence regression coverage.
Refresh the canonical protected-source receipt.
Scope live VIN and firmware queries to bus0 only when the passive pedal
layout and configured/effective GM query topology agree. Preserve normal
CAN fingerprinting, control configuration, and all Panda fault vetoes.
Regression covers recorded Bolt startup layout, real ISO-TP discovery,
negative topology cases, and unchanged pedal safety configuration.
Symmetric native probes could qualify a coordinate even when only the opposite direction changed its encoded command. Carry each direction’s measured visibility and physical effect into selection and interrupted-study recovery so a quantized direction cannot borrow the other direction’s evidence.
Keep existing trial bounds, slew, deadlines and independent acceptance criteria. Validate the previous selector failure, full FLM suite and recorded d5 producer/learner replay. Real-drive improvement requires post-update recorded evidence.
The QR code only linked to the removed APK download, so /api/companion/qr.svg is gone and /api/galaxy/qr.svg answers only when Galaxy is paired. Hide the Enable phone internet button, its restart checkbox and the sharing blurb on the Galaxy page.
Android users join the Play beta group from Galaxy, so remove the bundled installer, its version file and the workflow that verified them. The comma now reports no shipped app version, so it no longer offers an offline update.
Serve the Galaxy app key and its QR code from /api/galaxy/session to direct local connections so the Android app can import the tunnel key on its own. Let the app and the iOS app register for Sentry push alerts.
The d5 Ioniq 6 logs repeatedly reset source tracking when Card repeats a command that already arrived on time. Discard only a recent unchanged repeat after the existing native CAN verifier succeeds; add no sample, application time or permission.
Use independently admitted pending CarState brackets to complete original poses without waiting for that command to reach Card. Classify stale control source clocks as transient stale evidence, preserving the original bounded study deadline. Replace Python character loops with equivalent strict text predicates.
Validation: focused regression suite and paced d5 replay with original Card command references; actual controller/CAN trial application and worsening-trial rollback. No timing/acceptance threshold changes or claimed road improvement.
The output used to move after the small model's warmup had linked the
warp to its address, which needed the warp to link again by dropping a
private tinygrad cached property. Had that ever stopped working, the
small model's own warp would have kept writing the freed memory.
modeld now asks the adapter's coherent_warp() (Jetlink's make_coherent
with Jetlink on, off legacy9d) and load_verified_model moves the small
model's warp output before warmup(), as Jetlink's own warp moves its
output before its first call. CatalogWarp only reads the answer
(warp_copied): sent from where it is, or copied out, chosen once.
tinygrad maps the warp's QCOM output write-combined, which the CPU reads
uncached: CatalogWarp spent ~2.9 ms a frame copying the frame out, out
of the 46 ms a host has to answer in. It now moves the output to
IO-coherent write-back memory with Jetlink's warp.make_coherent, as
Jetlink's own warp does, and the link sends straight from it (0.22 ms
measured for Jetlink's warp). The warp has run by then, and tinygrad
bakes a held buffer's address in at link time, so its program links
again on the next call. If the kernel refuses the memory, the frame is
copied out as before, and the adapter's log says which.
TinyJit's per-call input checks are left alone: 0.045 ms a call on a
Mac under this tinygrad, not the ~1 ms they cost under Jetlink's.
warp.make_coherent is public, for an adapter's own warp: it moves a QCOM
output to memory the CPU reads cached, as Jetlink's warp does.
CREDITS.md and JETLINK.md name the new revision.
CatalogModelState.warp_frames is the warp branch of run(), which run()
and CatalogWarp now both call, so the link enqueues exactly what the JIT
was captured with and reads npy and input_queues live rather than as
they were when it was built. CatalogWarp binds the output once (the
captured JIT returns the same buffer every call), reuses one pair of
warm-up frames for the probe and frees them, and drops their blob
cache entries. model_warp lets a failure raise without logging it
too: Jetlink logs it as it keeps the small model.
Jetlink's stock driving warp is a legacy9d pickle, which a native02e
modeld cannot load, so the link never took over there. The adapter's
model_warp (Jetlink's optional hook since 0605e34) now hands Jetlink
the small model's own camera warp (CatalogWarp) on any other runtime:
the same JIT, inputs and blob cache, read into host memory after each
frame, with a probe at load that refuses an output not read live.
Under legacy9d it returns None and Jetlink loads the stock pickle.
From #253, without its edit to jetlink_repo or the hold tolerance
override.
The adapter may now supply the warp the link sends (ModelSide.model_warp,
optional; what it returns is checked against interface.FrameWarp before
the frame loop). Without one, or when it returns None, Jetlink loads the
stock driving warp pickle as before. API stays 3.
CREDITS.md and JETLINK.md name the new revision.
Discard a preempted clock pair and try up to three fresh pairs while
retaining the 1 ms accepted span and all existing freshness/resume gates.
Yield after receive-only decoder construction so normal UI collection
refreshes before full physical admission, Offroad request or input.
Recorded d5 clock failure; reproduced timing regressions; 250 focused
native cases, lint/type checks and independent source/evidence review.
No vehicle commands, device deployment or physical cure claim.
Clarify connection stages across wireless, adapter, USB, and tablet options with regression coverage for progress states and responsive layouts.
(cherry picked from commit 489dad0e7733127c7a9654c9a7491d1ab8a554a4)
Count distinct touch releases by their original timestamps, retain taps
during valid Park dwell, and let preparation earn its own admission.
Cancel only the pending C4 startup camera redirect for a freshly parked
open game. All session, physical, alert, control, noOutput, ignition and
input authority checks remain unchanged.
231 focused native cases, original before regressions, current successor
entry/decoder qualification, canonical lint/type checks and independent
review. Host functional evidence; no device changes or headroom claim.
A hard override on the Civic puts the EPS in NO_TORQUE_ALERT_1, which
reads as a temporary steering fault. The armed-AOL warning added
steerTempUnavailableSilent for it every time, even though the driver was
holding the wheel. CarSpecificEvents already stays quiet in that case.
Route 00000060: all 23 status-2 windows cleared before the driver let go.
The warning still shows if the fault is present while the wheel is free.
(cherry picked from commit f9167e2a24007c0822b4eae8870a24100a709325)
Renew the existing activity timer only during active offroad game ticks.
Observe only the exact game-owned Offroad request and restore Auto after
continuous fresh all-Panda ignition-off, closed game, stopped driving
services and healthy noOutput through the existing owner transaction.
Ordinary exit, stale data, reconnect and external mode changes never
restore Auto. UI restart does not adopt a previous game request.
142 source-bound focused cases, original timeout regression, lint/type
checks and independent source/evidence review. Synthetic host lifecycle
qualification only; no device changes or new physical/headroom claim.
Use the supported Offroad lifecycle with fresh Park, receive-only CAN,
stopped-process and noOutput checks. Add offroad touch/controller play
and the native Big UI five-version-tap entry. Exiting never restores
driving mode automatically within the existing manager session.
Qualified by210 unique focused passing cases, lint/type checks, native
host rendering/input/resource checks and independent lifecycle review.
Actual vehicle lifecycle, hardware inputs and device headroom remain
unqualified; no deployment is part of this change.
Add LAN and Bluetooth connections, secure pairing, device-matched Galaxy page
caching, phone location and CarPlay navigation. Include connection recovery,
diagnostics, regression tests and a simulator CI build. Add shared companion
docs and host tests, and remove obsolete bundled-web lint exceptions.
Connect over LAN or Bluetooth with network discovery, build-matched Galaxy
pages, phone GPS and a background uplink. Guide pairing with readiness retries,
cancellable progress and bond cleanup. Include app updates, regression tests,
the current APK, and a workflow to test, build and sign releases.
Add companion installation and updates, Spotify setup and playback, destination
sharing, and Android Auto controls. Adapt Galaxy pages to phone transports and
device type, improve Bluetooth pairing feedback, and cover the new web flows.
Expose pairing readiness and coordinate adapter hand-offs with Android Auto.
Show app pairings separately, support individual removal, and improve pairing
controls on both comma screens. Add Galaxy companion endpoints and web asset
manifests, preserve the plain camera look, and fix headless speed-limit shapes.
Made card headers, icons, spacing, and controls more consistent across Galaxy pages.
Improved vehicle selection and Bluetooth device layouts.
Refined Model Manager filters and moved running model status into its four-item summary.
Moved Jetlink above Controls.
Updated UI tests to match the revised layouts.
Improved consistency across Galaxy pages
Reworked the Model Manager layout, filters, downloads, randomizer, and Jetlink controls.
Improved vehicle selection and the Galaxy pairing and installation flow.
Fixed settings-row rendering, Android Auto upload timeouts, and Sentry event ID validation.
Updated tests for current UI behavior - all 102 JavaScript and 797 Python Galaxy tests pass.
Use GPS transport services in navtilesd and MapFeed, preserving phone GPS through starpilotNavigation. Add startup subscription regressions.
Source-only correction; no prebuilt receipt or native artifacts changed.
Bundle the verified VASM model, enable it at startup, and split VASM and vision speed-sign inference across little cores when both are active.
Restore VASM warnings in PiP activation, red outlines, and onroad blind-spot borders, including expanded Android Auto views.
Improve cabin snapshots with longer warm-up, adjusted capture timeouts, a fresh preview frame, and clearer capture buttons and save prompts.
Polish Galaxy settings with dimmed gated controls and icons, clearer vehicle fingerprinting and steering availability descriptions, rounded speed-offset ranges, and immediate refresh after saving.
Improve updater startup messages and preserve update/reconnect screens across service interruptions and restarts.
Add regression tests for the updated behavior.
jetlink's own kit checks the adapter module and modeld's hook order. Tests cover in_control with AOL armed and paused, the driving warp jetlink runs, and the catalog runner's history reset on tinygrad keeping its buffers. JETLINK.md describes the stay-engaged hand-back and the cable rule.
The Jetlink card's state, activity and running model come from jetlink's Status (phase, active, running, running_sha256, running_checkpoint), so StarPilot no longer derives them. A build in progress reads preparing, with its line below.
A Jetlink hand-back while engaged soft-disabled with Big Model Failed and
"Restart the car to retry", although the small model drives on and Jetlink
retries in seconds. Now nothing disengages: a Big Model Lost warning shows
for 5 s. A native chestnut's loss still raises Big Model Failed.
jetlink's swap events also refuse engaging for the second after a swap
(bigModelLoading) and hold the comm and localizer errors the swap's dropped
frame raises, as a chestnut load does. In control is the rule the swap
itself waits on (openpilot, or AOL armed or paused), and a gap in modeld's
runtime record holds the last state instead of reading as a switch.
The warning is a selfdrived alert, like aolAvailability: no new
OnroadEvent, so cereal and schema-policy.json are unchanged.
- jetlinkd is ancillary: a dead owner sends the drive back to the small
model and no longer blocks engaging
- a crashed owner now restarts, as porting.md asks
- Offroad_JetlinkUnavailable says why a Jetlink the user turned on cannot
run (jetlink's reason())
- JetlinkBigModel, JetlinkCatalog and JetlinkRuntime are no longer used
params_keys.h changed: libparams_c.so, libcommon.a and the prebuilt
manifest need regenerating with the prebuilt build. Until then the
tracked params library does not know Offroad_JetlinkUnavailable, so
hardwared skips that alert (logged once) and it stays inert.
jetlink runs the driving_warp pickles modeld already ships, so its own warp build, the four jetlink_warp pickles and their QCOM import and validation go.
cable_connected() is true for any host on the USB-C port. With a Chestnut pick downloaded and Jetlink on, every drive waited 10 s for a chestnut, raised Big Model Failed and left Jetlink unattached. The cable now counts only with Jetlink off.
StarPilotAdapter subclasses jetlink's Stock: StarPilot's param names, the
AOL engagement rule, the model root and a CatalogModelState history reset
that keeps its JIT buffers. jetlinkd runs jetlink's owner as it is, modeld
calls the hooks, and Galaxy and analytics read jetlink's runtime record in
place of JetlinkRuntime.
- no big-model slot or catalog keys: jetlink provisions its default model
- jetlink restores the owner's tuning itself; the port draw stays cut with
a host plugged in
- the Jetson is asked to power off on DoShutdown only, capped at 30 s
Vendor only the comma payload jetlink commits to (jetlink/, LICENSE,
pyproject.toml, README.md) with an include list, so the exclude list goes
and upstream's README stays as published; StarPilot's notes live in
JETLINK.md. The root helper now ships inside the package.
The adapter moves to API 3 in the next commit; until then Jetlink stays
off.
d24328d stands in for the jetlink release tag commit: re-run
tools/vendor/sync.py jetlink_repo <tag commit> --apply before merging and
update the revision in CREDITS.md and JETLINK.md with it.
Install the pinned H.264 test encoder in the unit-test environment and align the UI contract test with existing wired support. Preserve runtime code, timeouts, and ownership gates.
Preserve existing USB behavior while typing test doubles and the diagnostic log proxy. Correct prose spelling and the existing CLI executable mode. Add proxy forwarding coverage and refresh the canonical source receipt.
Adapt PR #226 stationary lead release and breakaway to the global Faster Takeoffs toggle. Preserve strict freshness, committed-stop clearance, selected acceleration profiles and planner bounds. PR #218 launch admission is already implemented through the same optional path.
Interpolate the measured Ioniq 6 steering delay instead of rounding delayed model requests to whole frames. Bound unsaturated friction growth without exceeding the current requested component or retaining its old direction; preserve full breakaway response, saved tunes and native limits. Cover actual Controls, replay state and controller timing.
Keep bounded automatic learning collection through valid live calibration refinements. Retire generations on resets or invalid calibration, preserve live physical transforms and existing controller tunes, and validate fitted coefficient bounds before publication. Exercise the current controller and historical wire paths.
Wire longitudinal feature selection and inputs into planner hosts after their owners are available.
Co-authored-by: Danny <dannyirwin31@yahoo.com>
Co-authored-by: whoisdomi <dcantu83@gmail.com>
Provide alert playback and volume control with a native audio callback that does not depend on the Python GIL.
Co-authored-by: Daniel Koepping <elkoled@gmail.com>
Select qualified speed-limit sources, preserve acceptance and override decisions, and publish distinct control and display coordinates. Carry accepted and pending source identity with each displayed decision.
Co-authored-by: firestarsdog <229254897+firestarsdog@users.noreply.github.com>
Own model discovery, compilation and registry selection through explicit model metadata and geometry contracts.
Co-authored-by: whoisdomi <dcantu83@gmail.com>
Implement Bolt Pedal command generation and mode selection with dedicated native safety coverage. Keep pedal actuation separate from the factory cruise families.
Define GM platform groups, startup preferences and CAN messages before their control families are registered.
Co-authored-by: firestarsdog <229254897+firestarsdog@users.noreply.github.com>
Serve fresh camera snapshots with explicit image responses and expose live crop controls without placing file or camera startup work on the UI thread.
Co-authored-by: Prabhaav Pillai <prabhaav@msu.edu>
Co-authored-by: dirwin31 <83434411+dirwin31@users.noreply.github.com>
Coordinate finite camera leases with manager-owned processes and reject frames that predate a request. Preserve the native frame timestamp contract across camera variants.
Resolve cruise speed increments through a shared interval owner for consistent button behavior.
Co-authored-by: Harald Schäfer <harald.the.engineer@gmail.com>
Extend native safety and health contracts with explicit AOL state and bounded message handling. Keep vehicle-specific admissions outside the generic layer.
Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>
Introduce shared feature settings and presentation state so both UI layouts can consume the same resolved values.
Co-authored-by: firestarsdog <229254897+firestarsdog@users.noreply.github.com>
* pigeond: aid startup with last GPS position
* pigeond: aid position after receiver setup
* pigeond: fixed 100km position accuracy, skip stale positions
* store LastGPSPosition as capnp bytes
* ubloxd: save LastGPSPosition
* pigeond: send time and position once system time is valid
* pigeond: use datetime for saved position age
* pigeond: remove position accuracy comment
* pigeond: send position assistance during initialization
* Update ubloxd.py
---------
Co-authored-by: haraschax <6804392+haraschax@users.noreply.github.com>
(cherry picked from commit 1623441a9f080138570e1d37c9a261ee3511357e)
* pigeond: fix stale AssistNow time assistance
* pigeond: save GNSS state after stopping receiver
* pigeond: preserve fractional UTC assistance
* Revert "pigeond: save GNSS state after stopping receiver"
This reverts commit 5bf5ea84d8a995f7561a58bdf6d071cc78759703.
* pigeond: send current time after AssistNow assistance
* pigeond: simplify AssistNow initialization filter
* pigeond: send host time only with AssistNow
* pigeond: check host time validity inside send_time
* Update pigeond.py
---------
Co-authored-by: haraschax <6804392+haraschax@users.noreply.github.com>
(cherry picked from commit 4c535577640a1dc8d91b017584de2c9b010f491b)
* test_onroad: chestnut usb unplug fallback
* test_onroad: fix chestnut unplug test
* test_onroad: chestnut fallback
* check that small model is running
* fix small model check
(cherry picked from commit 043759da3627803304956abbab4d10341cee8328)
* force metered on comma prime webbing sim
* rearrange
---------
Co-authored-by: Daniel Koepping <elkoled@gmail.com>
(cherry picked from commit 9b585d58f54d0129a1dfa434b59506a29c011336)
* ui: mici longitudinal icons
* ui: rm orange state
* ui: lead policy indicator
* fix
* show on op long
* make own widget
* sync long indicator with dmoji
* fast fade in with dmoji
Original-upstream-commit: 035a447d1f0f0ef5abf2ed9724e33d9265399019
Upstream: https://github.com/commaai/openpilot/commit/035a447d1f0f0ef5abf2ed9724e33d9265399019
* Add gas override boost, mici boost bar, and Rivian maximum set speed
* Refactor acceleration boost and gate accumulation on model limit
* Fade acceleration boost with model acceleration
* Limit boost buildup per gas override
* Latch boost eligibility before gas override
* Move model limiting check into AccelBoost
* Use DT_MDL directly in AccelBoost
* Clarify acceleration boost state names
* Scale acceleration boost by speed instead of decaying it
* Enable Experimental Mode by default
* Scale boost buildup with the application speed ramp
* Fix boost serialization and model-limit eligibility
* Remove boost UI and Experimental Mode default changes
* Combine boost update and application
Original-upstream-commit: ab8c84c3989bab34969d83e78c5fad3ee5b95112
Upstream: https://github.com/commaai/openpilot/commit/ab8c84c3989bab34969d83e78c5fad3ee5b95112
Keep official openpilot ancestry and dependency licenses. Store resources directly in Git, remove LFS setup requirements, and leave the downloadable big model outside the source tree.
ui: match the comma 3X road camera speed gates to the comma 4
Experimental Mode switches the road view to the wide camera below 5 m/s (11 mph) and back to the narrow camera above 10 m/s (22 mph) on both devices. Also corrects the comma 4's mph comments (were 10 and 25 mph).
* Bump opendbc and panda for Rivian alpha longitudinal
* Bump opendbc to include Tesla reverts
---------
Co-authored-by: Shane Smiskol <shane@smiskol.com>
* cabana: allow heatmap resizing without clipping
* cabana: keep outer grid lines inside the heatmap
* cabana: add a visible heatmap resize handle
* cabana: use the standard pane divider for resizing
* cabana: refine signal heatmap grid
* cabana: restore direct signal color blending
* cabana: strengthen heatmap signal saturation
* Revert "cabana: strengthen heatmap signal saturation"
This reverts commit 07ee9664d533f260b5bbfaed1afc4391c795603c.
* cabana: remove signal cell gradients
* cabana: centralize accessible theme and data colors
* cabana: remove color contrast tests
* cabana: simplify theme color helpers
* cabana: soften inactive message rows in dark mode
* cabana: retain original signal tooltip behavior
* Show pairing provider and Prime status in device settings
* remove unnecessary
* simplify
* bespoke human code
* clean
* readme
* prime layout
* sepearte qr code stuff
* move icons
* move qr code into its own widget
* clean
* Fix prime menu lint
* Update widget leak test for prime menu
* fix lint
* update to new theme
* change names
* fix font size
* get rid of subtitle
* adjust commacare icon
* drop manage, its cleaner
* edit text
* clean
* for github and apple always return generic
* pairing notification
* update copy and remove duplicates
* don't write too often
* remove
* make sure it doesn't show up on 3X
* reorg
* remove redundancies
* Show pairing provider and Prime status in device settings
* remove unnecessary
* simplify
* bespoke human code
* clean
* readme
* prime layout
* sepearte qr code stuff
* move icons
* move qr code into its own widget
* clean
* Fix prime menu lint
* Update widget leak test for prime menu
* fix lint
* update to new theme
* change names
* fix font size
* get rid of subtitle
* adjust commacare icon
* drop manage, its cleaner
* edit text
* clean
* for github and apple always return generic
* updates
* go back to original
* noice
* better overwrite
* close on unpairing if in management section
* missing 20 px when icon is there
* clean diff
* fix lint
* Update tinygrad to latest upstream with minimal openpilot patches
* Rebase tinygrad on latest upstream with one remaining patch
* Use retargetable model and warp artifacts
* Pack compiled model weights to reduce startup time
* Memory-map out-of-band model weights
* Use unmodified upstream tinygrad master
* Keep locally built model artifacts native
* Remove model chunking and use LFS for Chestnut releases
* Check Git blob sizes once during release packaging
* Download the big model for Chestnut nightly builds
* Reuse published LFS models when pushing releases
* Preserve LFS configuration when copying Chestnut releases
* Detect release builds independently of LFS attributes
* not worth 1m
* saves 24s
* save 13s
* parallelize model compilation
* explain compiler affinity
* full cpufreq
* rm that for now
* cmt
---------
Co-authored-by: elkoled <elkoled@gmail.com>
* Ship precompiled eGPU model and camera warps
Compile f78ed37d-afad-4dbc-8050-40ea885eedde/12864 through xx/ml_tools/openpilot_compile using the pinned tinygrad version.
* Precompile the existing master driving model
Use the unchanged master ONNX (SHA-256 6fee5937923c74848df4a63f6239eb6331c6274dd4bdb7a5d6ec0388a8b543d5) instead of updating the trained model.
* Compile camera warps on device
* Remove obsolete ONNX chunking and big model build check
* Chunk model artifacts only during release packaging
* Require model and camera warps for Chestnut readiness
* Recompile precompiled CPU helpers for the runtime host
* Ship the eGPU model with an ARM submission helper
* Exempt model pickles from the build product size limit
* Use tinygrad generic ONNX compiler artifacts
* Keep compiler temporary model buffers off device tmpfs
* Keep recurrent model state updates on the GPU
* Use compiled output buffers and swap recurrent state inputs
* Note when ONNX chunk reassembly can be removed
* Bump tinygrad to rebased compilers and AMD queue fix
* Keep recurrent buffers fixed and use faster USB argument updates
* Bump tinygrad for faster model startup
* modeld: use tinygrad ONNX and warp compilers for driving and DM
* Bump tinygrad with current openpilot model compile tests
* modeld: prioritize tinygrad compiler imports in build environment
* Use compiler branch based on openpilot's pinned tinygrad
* modeld: invoke tinygrad compilers without firmware patch
* modeld: move history queues into ONNX
* Move ONNX history export into xx exporter
* Rename ONNX image input to new_img
* Combine ONNX model history state
* Revert model history README addition
* Keep float32 input packing from master
* Rename model input specs to input shapes
* cabana: show persistent UI FPS
* cabana: render whole fps in monospace
* cabana: keep fps aligned beside centered progress
* cabana: move fps to top right
* Revert "cabana: move fps to top right"
This reverts commit 51a5334aec79807f8102b9a6f1f902f22814d5c7.
* cabana: add browser sign-in for remote routes
* cabana: use shared dialog styling for browser sign-in
* cabana: support spacious centered action buttons
* cabana: vertically center sign-in states
* cabana: keep sign-in layout stable between states
* cabana: center sign-in error text
* cabana: avoid hint flash when retrying sign-in
* replay: retry failed segment loads
* replay: rename test helper for spellcheck
* replay: include portable temporary directory declaration
* replay: name attempt limit and shorten test delays
* replay: drop timing-dependent retry tests
* cabana: align center tabs and keep signal panel docked
* cabana: prevent panels from docking into the tabless center
* cabana: use native docking for the Signals panel
* cabana: migrate dock tabs when loading the layout
* Cabana: standardize floating dropdown menus
* Cabana: use uniform items in menus and combo lists
* Cabana: remove dropdown test and explanatory comments
* cabana: standardize button spacing and control alignment
* cabana: center shared step controls and inset chart content
* cabana: use uniform per-chart content insets
* cabana: redraw floating panes during native resizing
* cabana: keep rendering while native resize is held
* cabana: use ImGui decorations to avoid native resize loops
* cabana: draw docking previews only on the target viewport
selfdrived: don't alert on a message that was never received
posenetInvalid, locationdTemporaryError and paramsdTemporaryError are raised
off the capnp defaults of deviceMotion and vehicleParameters before either
message has ever arrived. When modeld is down, locationd and paramsd never
publish, the defaults read as a nan posenet speed, and that NoEntryAlert can
hide the process not running alert that names modeld.
Same guard as the big model alert in #38500.
* Move LFS asset transfers to xet-core
* Limit migration to Git LFS hosting on Hugging Face
---------
Co-authored-by: Harald Schäfer <6804392+haraschax@users.noreply.github.com>
* esim: MICI profile download UI
* Show temporary feedback for non-LPA QR codes
* Scan eSIM QR codes every 250 ms
* Shrink eSIM error text and log displayed errors
* Activate downloaded eSIM through the profile tap flow
* Let the eSIM connectivity error wrap naturally
* Release add-profile button feedback immediately
* Check internet before opening the eSIM QR scanner
* Use UI network state and standard connectivity dialog for eSIM
* Drop LPA prime settings change from eSIM download UI
* Scope eSIM UI constants to their owning classes
* Remove unused ot codespell exception
* esim: simplify profile download operation and UI state
* esim: share activation code validation with QR scanner
* esim: match QR scanner video to cabin preview
* esim: show errors in a scrollable text dialog
* esim: require nonblank profile nicknames
* qrcode: add QR decoder
Decodes a QR code from a grayscale image or a module matrix: adaptive
binarization, finder pattern search by run ratios, perspective sampling
with alignment pattern refinement, format info, unmasking, block
de-interleaving, Reed-Solomon correction, and numeric, alphanumeric,
byte (with ECI), and Kanji segments.
Fixtures are packed matrices from python-qrcode 8.2 covering every
version and level, plus Segno 1.6.6 matrices for the ECI cases.
* qrcode: tune decoder for the device
Measured on a comma mici with real cabin frames downsampled to 672x380,
the size the eSIM screen scans at. Per frame: 30 ms -> 8 ms with no
code in view, 42 ms -> 13 ms with a code.
- binarize: fixed two-radius fill from one integral image instead of an
iterative outward fill (a dark cabin is almost all flat tiles), tile
stats on a contiguous layout, flat tiles decided whole so sensor noise
cannot become speckle, uint8 threshold compare
- finder search: scan every 4th row, build column runs only at
candidate columns
- alignment search starts at a 2-module radius
- Reed-Solomon syndromes through table lookups, placement order cached
* qrcode: simplify decoder
Fold the function-module mask into _data_coords and the format
coordinates into _read_format, evaluate masks directly on the data
coordinates, compact the ECI, numeric, and alphanumeric parsing, and
flatten the retry loop in decode. No behavior change.
Replace the level-L-only Reed-Solomon parameters, the single hardcoded
format word, and the bitwise GF(256) multiply with the full EC table,
the 32 format words, and log/antilog tables. Block splitting and
interleaving move into _block_lengths and _interleaved. This is the
structure a decoder needs to read any version, level, and mask; the
encoder's module matrices are unchanged and test_encoder pins them.
The closed-form step between alignment patterns rounds the wrong way
for version 32 (28 instead of 26). Use the form that is right for every
version. The encoder only emits versions 1-20, so its output is
unchanged.
* modeld: build DM warp and model JITs for all camera configs
In #38684, camera_configs on comma_arm64 was restricted to only the host device
running SCons (selecting either _os_fisheye or _ar_ox_fisheye).
However, prebuilt release and nightly bundles are built on TICI/tizi runners
in Jenkins and distributed to both Comma 3/3X (TICI) and Comma 4 (MICI) devices.
Because the build machine was TICI, dm_warp_1344x760_tinygrad.pkl was omitted
from the packaged bundle, causing dmonitoringmodeld to fail with FileNotFoundError
on Comma 4 and leaving driver-monitoring calibration stuck at 0%.
Restore CAMERA_CONFIGS so both camera resolutions (1928x1208 and 1344x760)
are built for all platforms.
Fixes#38762
* simplify
---------
Co-authored-by: ZwX1616 <zwx1616@gmail.com>
* esim: MICI eSIM profile management UI
* esim: align rename button position regardless of delete button
* esim: skip profile UI when SIM is not an eUICC
Probe is_euicc() on the first profile poll and cache it; non-eUICC SIMs
avoid list_profiles/process_notifications (which hang on a plain SIM) and
the eSIM button shows ICCID/MCC-MNC metadata instead of opening the
profile management screen.
* esim: satisfy ruff E731 in action_pressed helper
* esim: get modem info from modem.py state instead of shelling out
* esim: simplify switch lifecycle, drop active flag and settle window
* esim: only show 'switching...' on the target profile
* esim: read cell strength directly from HARDWARE
* esim: hide checkmark and dim cell icon during switch; keep rename available
* esim: disable active profile button (rename still clickable as overlay)
* esim: stop rename/delete buttons and labels from flashing during operations
* esim: show 'comma prime' on network button for comma profile
* esim: move display_name and is_comma onto Profile dataclass
* esim: anchor rename button to rightmost slot to prevent shift
* esim: include iccid prefix check in Profile.is_comma
* esim: drop process_notifications from cellular manager
* esim: disable network button when SIM isn't an eUICC
* esim: rename ESim* classes to Esim*
* esim: sort imports
* esim: default to 'loading...' instead of 'no active profile'
* esim: rename PROFILE_POLL_INTERVAL to PROFILE_POLL_INTERVAL_S
* esim: slim EsimNetworkButton
* esim: use DEFAULT_TEXT_COLOR; load delete dialog texture in __init__
* esim: revert cell-icon index trick, name each NetworkStrength explicitly
* esim: tighten delete/rename spacing so 'switch' label fits on one line
* esim: shrink delete/rename buttons by 25%
* esim: keep rename button at full size, only delete shrinks
* Revert "disable modem.py for now"
This reverts commit 1eeba86ec1.
* Revert "modem.py is disabled"
This reverts commit d238a1ccc4.
* lpa: inline comma iccid prefix
* ui/cellular_manager: clear switching state when LPA returns
* ui/cellular_manager: lock callback queue, drop poll log noise
* esim: sort NetworkStrength/NetworkType imports
* esim: drop switching_iccid concept
* esim: show 'switching...' on the clicked profile button
* esim: optimistic switch, skip post-switch list_profiles to avoid flicker
* esim: only style profile button from local op flags, not global busy
* esim: remove deleting/switching state, collapse profile button branches
* esim: show GSM settings on full prime when non-comma profile is active
* esim: expose CellularManager.active_profile, dedup callers
* esim: apply comma prime defaults on switch (roaming, metered, no APN)
* esim: restore re_sort on show_event to avoid first-frame jank
* lpa: apply comma prime defaults inside TiciLPA.switch_profile
* lpa: lazy-import Params to break circular import
* lpa: treat +CME ERROR 13 (SIM failure) as non-eUICC in is_euicc
* esim: show 'obtaining IP...' on non-eUICC path while connecting
* Revert "lpa: treat +CME ERROR 13 (SIM failure) as non-eUICC in is_euicc"
This reverts commit 475ca448ea914ff8f9892b4cbf4525d961d7618b.
* esim: poll profiles every 5s
* lpa: tighten Profile.is_comma to require both Webbing provider and comma BIN
* lpa: fall back to '<unnamed>' instead of iccid prefix in display_name
* esim: re-probe is_euicc each poll for runtime SIM swaps
* esim_ui: rename EsimUIMici to EsimUI
* esim: simplify cellular manager and profile UI
- optimistic profile switch at click time so the active button no longer bounces back
- unify LPA worker threads, refresh_profiles resets the poll timer
- deterministic profile ordering on show
- reuse wifi ForgetButton for delete, drop DeleteButton
- construct CellularManager inside NetworkLayoutMici
- drop comma prime param defaults from LPA.switch_profile (moved to a separate PR)
* esim: drop eUICC state change log
* ui: gate cellular settings on prime subscription
* ui: disable eSIM management for full prime
* ui: keep eSIM management disabled for full prime
* ui: unify eSIM profile action buttons
* ui: tighten eSIM profile action spacing
* ui: place rename before delete in eSIM actions
* esim: restrict individual profiles and process switch notifications
* esim: defer poll after operations and confirm eUICC loss before clearing profiles
* ui: allow cellular settings for unregistered and unpaired devices
* ui: disable all profile switching for full prime
* ui: simplify cellular access to not full prime
* ui: make profile poll interval a CellularManager constant
* ui: read modem state on the profile poll cadence
* ui: scope eSIM profile colors to their class
* ui: inline modem state read in cellular polling
* ui: rely on hardware modem state fallback
* cabana: connect palette theme, rounded cards, floating page switch
Move the jotpluggler port's connect palette into a shared theme module
(ui/theme.h/.cc) and apply it to all of cabana: Palette struct for dark
and light, applyTheme fills ImGui and ImPlot styles, readableCurveColor
and sectionTitle for reuse by the analysis workspace. Drop the old
DarkTheme constants and per-theme color branches in widgets.
Docked panels are borderless with content in rounded bordered cards;
custom-drawn rects use the style radii. The detail widget's bottom tab
bar is replaced by a floating Messages/Logs pill.
* cabana: scrollbar and menu contrast, page switch below the page
The scrollbar grab uses the border color on the window-colored track.
Selection colors are stronger so an open menu reads as blue in both
themes and the hovered item is visible in light mode. The Messages/Logs
pill sits in a strip below the page instead of covering its last rows.
* cabana: even spacing around the page switch, steady top menu highlight
* cabana: inner spacing before the clear button of a clearable input
* cabana: center the page switch below the page, space the signal row buttons
* cabana: logs table cells padded and bordered like the messages table
* cabana: recess the signals card in the window color
* cabana: one button vocabulary
Three button shapes, all one frame height with the style padding and
rounding: ImGui::Button for framed text, iconButton for a framed square
icon, toolButton for a flat square icon (or flat text). Adjacent buttons
are ItemInnerSpacing apart, groups ItemSpacing apart. Tool bars drop
their private padding and 1 px spacing, the signal row, chart header,
detail header, logs export and signal selector buttons all use the
shared sizes, and the page switch is built from the same style values.
* cabana: center the glyph ink in square icon buttons
* cabana: framed tool buttons and plain icon glyphs
toolButton is now iconButton with an optional text, so every button in
the app is framed. Boxed glyphs (plus-square, x-square, dash-square,
arrow-*-square, window-stack) are replaced by their plain variants
(plus-lg, trash, arrows-collapse, box-arrow-*, window-plus), the light
x/plus/dash by their -lg forms, the chart menu by three-dots-vertical
and the toolbar overflow by chevron-double-right.
* cabana: draw square button glyphs at 80% so full bleed icons keep a margin
* cabana: signal values in the mono font
* cabana: center square button icons on their atlas ink, snapped to framebuffer pixels
* cabana: draw square button glyph quads directly, AddText truncates the position
* cabana: right align the messages and signals tool bars to the table edge, full size collapse button
* cabana: detail header and tab scroll buttons flush with the right edge
* cabana: header and footer strips, gap above the panels, video splitter as a separator
The menu bar and the status bar are strips in the surface color with a
separator line towards the panels, and the dockspace sits one item
spacing below the menu bar, matching the spacing above the status bar.
The video/charts splitter draws the same 1 px separator line as the
dock splitters, centered in a gap as tall as the side padding, and a
double click snaps the video back to its natural height.
* cabana: top menus flush with the menu bar, 2 px video splitter without double click
* cabana: dropdowns continue the menu bar, splitter in the border color, row button gap
Top level menus draw no top border and square top corners, and the menu
bar leaves its separator out under the open dropdown. The video splitter
uses the border color like the dock splitters. The signal row buttons
keep one inner spacing before the scrollbar.
* cabana: dropdowns share the bar's 1 px border line, messages toolbar spacing matches the video card
* cabana: tool bar groups, framed dropdown buttons, chart content flush with the tool bar
Tool bar items are ItemSpacing apart and the items of a group (a button
pair, a separator and its neighbor) ItemInnerSpacing apart, instead of
a private 1 px spacing. The dropdown buttons are framed like the other
buttons. The video time is a plain mono readout that toggles on click.
Chart cards have no horizontal margin so their header and plot line up
with the tool bar above.
* cabana: chart list spacing: gap above the scroll area, charts clear of the scrollbar, even legend gaps
* cabana: loop and speed buttons grouped like the playback buttons
* cabana: survive narrow panels and small windows
Stress tested by dragging the column and video splitters to their
limits and shrinking the window to 640x480. Fixes: the messages tool
bar and the message detail header overflow into the >> menu instead of
clipping (the view button and the name stay); the binary view keeps a
minimum cell width and scrolls sideways; the Messages/Logs switch drops
to icons when the strip is too narrow; the camera keeps its last frame
across reconnects, so a video restored from a collapsed splitter while
paused is no longer black.
* cabana: range slider spans the charts >> menu, remove all and float as menu entries
* cabana: the charts card keeps its tool bar row or collapses, like the video at the other end
* cabana: uniform >> menus: dropdowns become submenus, the time readout an action entry
* cabana: video/charts splitter gap matches the column splitters
* cabana: drop unused lambda capture in chart type menu
Claude-Session: https://claude.ai/code/session_01XTFaJpf235LeHBmyFRgeJ1
* cabana: the menu bar and the status bar stay in full screen
* cabana: simplify theme helpers and toolbar actions
* cabana: preserve panel width for overflow edit dialog
* cabana: keep message editor usable in narrow panels
* cabana: group heatmap modes in a dropdown
* cabana: trim redundant restyle comments
* cabana: fix review findings from the restyle
- set the caret and other unset colors so the light theme has no white caret
- tolerate float error in the toolbar fit check to stop spurious overflow
- restore the full screen chrome hiding
- clear stale camera frames when the server changes, keep them across a collapse
- cache icon glyph ink bounds instead of scanning the atlas every frame
- drop dead plot_bg and ImPlot globals, add a badge palette role
- add toolbarMenu and use toolbarAction for the remove message item
- derive the signal view button size seed from the font and style
* cabana: remove remaining redundant comments
On Wayland a vsynced glfwSwapBuffers blocks until the compositor sends a
frame callback, which it stops doing while the window is on another
workspace or otherwise off-screen. The main thread never gets back to
glfwPollEvents, so Hyprland reports cabana as not responding.
Run with swap interval 0 and pace the loop to the monitor refresh rate.
git-lfs is vendored in the venv, but `git lfs install` writes filters and hooks that invoke a bare `git-lfs`, which is only on PATH inside the venv. With filter.lfs.required set, any checkout from a plain shell fails.
git runs filters and hooks from the worktree root, so a relative path into the venv resolves from anywhere.
* cabana: imgui frontend base infra (ui/)
* cabana: line-for-line imgui ports of the Qt widgets and dialogs into ui/
* cabana ui: fixes from the side by side parity test
* cabana ui: deqt.md status after the line-for-line ports
* cabana ui: fixes from the line-for-line review
* cabana ui: parity fixes from the end to end matrix, event driven frame pacing
* cabana ui: preserve F1 help formatting
* cabana ui: parity fixes from the end to end matrix, event driven frame pacing
* cabana ui: deqt.md status
* cabana: fix macOS build of imgui frontend
- link Security.framework (libusb's darwin backend needs
SecTaskCreateFromSelf/SecTaskCopyValueForEntitlement)
- drop unused loop counter in SignalView::updateChartState
- avoid -Wshadow on Sparkline::size and SignalView::highlight
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: parity round 3 fixes, thread pool and one malloc arena for flat memory
* cabana ui: toolbar overflow budget, menu button carets, help overlay colors, signal editor validation
* cabana ui: deqt.md status
* cabana: let vsync pace the frame loop
The loop threw away frames with glfwWaitEventsTimeout() to hit a 30fps
(10fps idle) target while glfwSwapBuffers() was already blocking on
vsync, so two clocks beat against each other. Input events also reset
frameInterval() to 0 for a few frames, so a click swung the cadence from
~33ms to ~13ms and back -- the camera view redraws on the UI's schedule,
so that showed up as a visible hitch on every message selection.
Poll and draw every iteration and let glfwSwapInterval(1) do the pacing.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana: fix macOS build of thread pool / malloc arena changes
- malloc.h and mallopt(M_ARENA_MAX) are glibc only; guard on __GLIBC__.
macOS has a single allocator zone, so there is nothing to cap.
- capturing structured bindings in a lambda is a C++20 extension and
clang rejects it under -Werror; bind the event range to plain locals.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: chart axis label precision, signal list click on empty space clears the selection
* cabana ui: codespell
* cabana ui: taller signal rows, system theme detection, full dark palette
- signal rows are 1.5x a frame tall with the sparkline sized to the row,
so the graphs are readable. The expanded sub-rows keep the Qt height.
- "Automatic" resolved to the light theme regardless of the system: Qt
gets this from QStyle::standardPalette(), which follows the macOS
appearance, so read AppleInterfaceStyle and match it. Other platforms
stay light, like standardPalette() does there.
- widgets that tested settings.theme == DARK_THEME themselves went light
under AUTO_THEME while the style went dark; they now ask isDarkTheme()
for the theme applyTheme() actually resolved.
- the dark branch only set 10 ImGuiCol_ entries, so buttons, scrollbars,
tabs, title bars, separators, tables and grips still came from imgui's
stock dark theme. Derive the whole palette from DarkTheme instead.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: drop the dead y axis alignment chain and chart cache stub
ImPlot::BeginAlignedPlots aligns the chart y axes, so the hand-rolled
alignment protocol ported from Qt no longer feeds anything: align_to was
write-only. Remove the whole chain (y_label_width and its measuring loop
in draw(), axisYLabelWidthChanged, align_timer, alignCharts), which
reduces updatePlotArea(int, bool) to updatePlotArea().
resetChartCache() has been an empty stub since the port (imgui redraws
every frame); drop it and its four call sites. Also replace the empty
move_icon_rect branch in mousePressEvent with an early return.
No behavior change. updateAxisY's "|| y_label_width == 0" guard only
fired after a unit change, and y_precision/y_tick_count depend solely on
min_y/max_y/tick_count, so the recompute it triggered was a no-op.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: stop the main window from scrolling
The host window that holds the dockspace uses zero WindowPadding but the
default ItemSpacing, so the dockspace and the status bar below it summed
to ItemSpacing.y (5px) more than the viewport, leaving the whole UI
scrollable by a few pixels.
Reserve the spacing along with the status bar height, and mark the host
window NoScrollbar/NoScrollWithMouse: it is pinned to the viewport work
area and should never scroll, as QMainWindow does not. Full screen is
unaffected, it draws no status bar and reserves no height.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: inset the status bar so its ends are not clipped
A borderless BeginChild gets WindowPadding forced to zero, so the status
bar text sat flush against both edges and the first glyph of "For Help,
Press F1" was cut off. Inset both ends by WindowPadding.x, which lines
the text up with the content of the docked panels above it.
Right align the cached minutes / FPS label to the captured child width
rather than calling GetContentRegionAvail() again after the message
text, which only returned the full width because the text had already
wrapped the cursor to a new line.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: drop the leftover imgui object census log
The block printed an imgui/implot object count to stderr every 10s,
which was instrumentation for tracking down a leak, not something the
port needs. It was the only user of implot_internal.h in mainwin.cc, so
drop that include with it.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: keep the selected row highlighted while hovered
Selectable() picks its background as HeaderActive when held, else
HeaderHovered when hovered, else Header. The hovered test comes before
the selected one, so a selected row under the cursor takes the hover
color. Both tables push a transparent HeaderHovered to suppress the hover
highlight that QTreeView/QTableView do not have, which also blanked the
selection: clicking a row walked press (HeaderActive, visible), release
(HeaderHovered, transparent, looks unselected), then mouse off the row
(Header, visible), so the selection appeared to flicker until the cursor
moved away.
Only make HeaderHovered transparent for rows that are not selected.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: draw the binary view hex column in bold and in the text color
The Qt delegate never sets a pen for the hex column, so it inherited the
black QPainter default and the two halves of a row did not match. Use
paletteText(is_message_active), the same color the bit columns use, so
both halves read as one row and dim together when the message goes
inactive.
JetBrains Mono ships no bold variant and imgui has no embolden option, so
give drawStaticText a bold flag that redraws the glyphs a fraction of a
pixel to the right. That keeps the monospace advance, which switching to
the proportional bold face would not.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: highlight the whole row in the speed menu
MenuItem spans from the cursor to the right edge, so the Indent() that
made room for the radio bullet also pushed the highlight in by one
FontSize and left the bullet column outside it. QMenu highlights the
whole item, indicator included.
Draw each row as one full width Selectable with the bullet and the label
inside it. Every row declares the same width, so the popup is as wide as
the widest entry and the highlights reach both edges; a label-less
Selectable sized zero would declare no layout width and collapse the
auto-sizing popup. Selectable auto-closes its parent popup like MenuItem
does, so clicking still works in the dropdown and in the Speed submenu of
the toolbar overflow menu, which shares this function.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: fix help menu overlay, Ctrl+F11, time label clipping, toolbar overflow and tool table parity
* cabana ui: click to edit signal cells, light theme selection, logs cell selection and filter fixes
* cabana ui: chart x label margin, zoom context menu items, video splitter collapse, signal editor close
* cabana ui: signal editor lifetime and close, logs cell hover, tool table headers
* cabana ui: branch indicator clicks, editor swallows shortcut and button clicks, selection colors
* cabana ui: float dock windows into real OS windows
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: toolbar menu arrow spacing, only open the editor on editable cells
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* op: add cabana2 to run the imgui cabana
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana: cabana2 launcher script builds _cabana_ui before running it
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: only float the dock panels into their own OS windows
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: open the video pane at the camera's natural aspect ratio
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana: default fps to 30
* cabana ui: fusion palette, shared fusion slider, chart grid colors
* cabana ui: smaller checkbox indicator, closer to the Fusion size
* cabana ui: video toolbar height and item centering, separator extent and menu arrow like Fusion
* cabana ui: message and signal row heights match Qt
* cabana ui: timeline slider height, groove and stacking match the Qt video layout
* cabana ui: fix crash opening a live stream, the video pane default height read the missing camera widget
* cabana: drop the FPS setting, the streams update the UI at a fixed 30 fps
* cabana ui: video and charts pane run edge to edge with a uniform toolbar margin like the Qt frames
* cabana ui: binary view M/L marker size and timeline event fills match Qt
* cabana ui: signal view keeps its sparklines while filtering, number label size matches Qt
* cabana ui: no modal dim fade, QDialog does not animate in
* cabana ui: collapse all is an auto-raise tool button with the Qt icon size
* cabana ui: dialogs and tool windows open as real OS windows like QDialog
* cabana ui: hold back a focus loss while a mouse button is down so tearing a panel off does not abort the drag and create its window twice
* cabana ui: speed menu rows pad the label on the right like QMenu
* cabana ui: raise the dark theme contrast, the Darcula grays all sat on top of the background
the chart y axis guides are opaque and 2 px in dark mode, and text, outlines,
separators, table borders, scrollbar and slider grabs move away from the window
and base grays. light theme is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: simplify comments, drop the Qt references and the ones that only restate the code
* cabana ui: keep the sparkline antialiased, thin peaks sparkled as the window slid
Qt dropped antialiasing above 500 points because it rasterized into a pixmap. drawing
straight into the frame, an aliased 1 px segment between two columns rounds into one of
them and the rounding flips as the window advances, so the peaks of the dense signals
popped between columns. updateState runs at STREAM_UPDATE_FPS, which at a 30s range moves
the curve well under a pixel per update, so the flip is all that is visible.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: anchor the sparkline to the time window, it jittered sideways as samples expired
x was measured from the first sample still inside the window, which changes whenever the
oldest one falls out, so every update slid the whole curve sideways by the sample spacing.
measuring from the start of the window pins the right edge and lets the samples flow left
with the clock. the density heuristic wanted the data span, which is no longer the x of the
last point, and the flat line started at x 0 instead of at its first sample.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: scroll the sparkline at the frame rate, it stepped at the message rate
the window ended at the last message of the id and updateState only runs when one arrives,
so a slow message held the sparkline still for several frames and then jumped it. the window
now ends at the playback clock, and draw() slides the rendered polyline by the time that has
passed since it was built, clipped to the cell.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: keep the signal value on the right of the column before there is a sparkline
it was drawn left aligned from the start of the column, so the value sat next to the name
until the first samples arrived and then jumped across the row.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: read a little past the sparkline window so the oldest points slide out of it
a sample used to disappear the moment it aged out, which on a fixed rate signal is a point
popping off the left edge. the query reaches back a bit further and the clip rect hides it
instead. the lead in must not move the scale, so it is left out of the min and max.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: use the macOS fullscreen space, taking over the monitor rendered at the old size
glfwSetWindowMonitor switched the display mode and returned a drawable at that mode, so the
ui was rendered at the windowed size and stretched over the screen. NSWindow toggleFullScreen
keeps the backing scale. the green button goes through the same space, so the state is read
back from the style mask rather than tracked.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: let the signal value column shrink again, it held the widest value ever seen
max_value_width only ever grew, so one long value left the sparklines ending
well short of the text for the rest of the message. Track the current widest
value instead, growing on demand and giving room back once it is a couple of
characters too wide, which keeps the width from jittering per sample.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: make the signal rows a quarter taller so the sparklines read better
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: lift the binary view heatmap alphas in dark mode, the tints washed out
The alphas were tuned against a white background; over the dark base the same
values left the cells barely colored. The floor comes up and the ramp gets a
gamma in dark mode only, which keeps the flip counts in the same order.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: saturate the sparkline color in light mode, the pastel lines vanished
Signal colors are low saturation and high value, which is aimed at dark fills;
a 1 px line of one of them on white is hard to see. Draw the sparkline with
twice the saturation and less value when the theme is light.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: give the docked panels a minimum width, the dividers ran over content
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* replay: capture the downloader's stderr so the progress lines reach the handler
file_downloader.py writes PROGRESS:<cur>:<total> to stderr while it streams a
file, but runPython() only piped stdout and left stderr attached to the parent,
so the progress handler only ever fired on failure and cabana downloaded a
segment with no progress bar. Pipe stderr as well, turn the PROGRESS lines into
handler calls, and pass everything else through to the parent's stderr.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* replay: read the downloader's stderr on a thread, the two fd select was overkill
getline() handles the line splitting, partial lines and EOF, so the manual
splitter, the second drain loop and the multiplexing bookkeeping go away and
the stdout loop goes back to what it was.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: one toolButton helper in imgui_util.h
* cabana ui: one toImU32/toImVec4/paletteBrightText in imgui_util.h
* cabana ui: string helpers live in utils/strings.h
* cabana ui: share the fusion slider handle from imgui_util.h
* cabana ui: style.cc declarations move to imgui_util.h
* cabana ui: validatedInput and the qt validators move to imgui_util.h
* cabana ui: viewSelectable helper for the item view rows
* cabana ui: drop the comboBox pass-through
* cabana ui: tabbar moves to its own file, ready for the charts port
* cabana ui: find signal searches on the main thread again, deferred one frame like QTimer::singleShot(0)
* cabana ui: find signal computes its button and label state from the model instead of caching it
* cabana ui: share the queued popup owner protocol between the message box and the file dialog
* cabana ui: one beginDialog helper for the three centered modals, with the reopen recovery
* cabana ui: tool dialogs share the title, begin/end and escape handling, and own their connections
* cabana ui: comboBox helper replaces the hand rolled zero separated item buffers
* cabana ui: drop the unused dialog isOpen accessors
* cabana ui: one validatedText helper for the ip, double and int line edits
* cabana ui: route info reads the route segments instead of caching a row copy
* cabana ui: settings drops the phantom fps label, the constant spin flags and the raw log path buffer
* cabana ui: find similar bits drops the no-op find button disable and the address round trip
* cabana ui: message box warning overload without the empty detailed text
* cabana ui: help overlay only collects rects while it is up, skips torn off panels, and the fingerprint compare is ordered
* cabana ui: drop the glfw callbacks that only forward to the imgui backend, which already installs them
* cabana ui: elided label renders with RenderTextEllipsis instead of a paint time cache
* cabana ui: one save loop, inline cached minutes label, and drop the dead help parser branches
* cabana ui: signal cell bands are built from the two corner notches instead of a region sweep
* cabana ui: binary view items go through itemAt and updateItem assigns unconditionally
* cabana ui: message bytes read the cell padding inline and colored bytes keep the qt default pen
* cabana ui: detail widget drops createToolBar, the welcome widget flag and the unused whatsThis
* cabana ui: messages widget owns its view, header and model directly, fixed size filter editors, from_chars range parsing
* cabana ui: colored hex bytes keep the row pen, the qt delegate does not reset it per cell
* cabana ui: closeEditor drops the queued commit, its item is about to be deleted
* cabana ui: rowsInserted carries the insert position so the selected row follows its message
* cabana ui: the log filter and the message name editors use the shared validators
* cabana ui: the log header sizing and painting are free functions, no HeaderView object
* cabana ui: the visible row range is the change detector, no mirrored scroll position
* cabana ui: the line editor takes the focus flag as an argument and lets InputText revert on escape
* cabana ui: signal view drops the unreachable column parameters, unused item values and repeated sparkline tests
* cabana ui: rows outside the tree viewport are measured but not painted
* cabana ui: one toolbar item table, playback state read in draw()
* cabana ui: draw the stored thumbnail scaled by AddImage, no cpu prescaled copy
* cabana ui: one radioMenuItem helper for the speed and series type menus
* cabana ui: drop dead camera members, hoist the jpeg colour space branch, decode qlog thumbnails on the thread pool
* cabana ui: stop the vipc thread when the splitter collapses the video pane
* cabana ui: the toolbar strings and enabled state are computed in drawToolBar
* cabana ui: chartswidget uses the shared tabbar widget
* cabana ui: the chart header layout is recomputed in resizeEvent, no updateTitle
* cabana ui: updateLayout has no force flag
* cabana ui: one condition hides the value tip
* cabana ui: signal selector uses the shared dialog buttons and row indexes
* cabana ui: the chart tile geometry is one Layout struct saved and restored by drawGhost
* cabana ui: drop the empty theme branch, the unused size hint and the charts container indirections
* cabana ui: the chart holds its tip label by value
* cabana ui: one toolbar item list for both toolbar groups
* cabana ui: the sparkline polyline is emitted through the draw list path
* cabana ui: the tab bar latches a programmatic selection and defers the close request past EndTabBar
* cabana ui: validatedText refuses an invalid edit inside the imgui buffer and the ip field keeps its char filter
* cabana ui: drop the ambiguous three argument warning overload
* cabana ui: combobox reports a real change, menu rows span the popup, one popup protocol for the dialogs
* cabana ui: the startup stream is owned by the window until the first frame opens it
* cabana ui: the slider toolbar index has an out of range sentinel
* cabana ui: the find signal search runs on the frame after the pending state is painted
* cabana ui: the log clears its selection when the rows are dropped and cells are identified by their message
* cabana ui: the open editor survives a scroll, collapse all commits it and escape in the size editor does not
* cabana ui: the header display order is filled with iota
* cabana ui: qlog thumbnails decode on their own threads, the hover thumbnail is mipmapped and the time tooltip appears on the first toggle
* cabana ui: a collapsed video dock stops the vipc thread
* cabana ui: one TOOLBAR_ITEM_SPACING, the signal view used the QCommonStyle 4 instead of Fusion's 1
* cabana ui: sparkline includes imgui_util.h for isDarkTheme
* cabana ui: imgui_util.h becomes util.h and util.cc
* cabana: delete deqt.md
* cabana ui: panel min width, matching video panel padding, centered tabs
The dock panels stop shrinking at the width where the signal view tool bar
squishes instead of at a hardcoded 440. The video/charts panel keeps the
standard window padding so it lines up with the other panels, and the
Msg/Logs tabs are centered in their bar. The cached minutes label is gone
from the status bar.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: persist imgui state in cabana.json and migrate the Qt frontend's layout blobs
* cabana ui: 2px sparkline stroke
a 1px antialiased line is all fringe and no solid core, so the curve
faded and shimmered as the window scrolled.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana: drop the Automatic color theme
it only did anything on macOS: the Qt frontend fell through to
standardPalette() for both Automatic and Light, so the two were
identical everywhere else.
LIGHT_THEME and DARK_THEME keep their values so a persisted dark theme
still reads as dark; the old Automatic 0 falls back to light on load.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: window color between the binary and signal views, fixed split
the binary view sits at its size hint and no longer resizes, and the gap
below it takes the window color instead of leaving a white seam.
also border the video and chart panels like the signal view.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: bigger tab bar scroll buttons
imgui's built in scroll arrows are fixed at FontSize - 2 wide with the
glyph drawn in a FontSize box, so they are small and sit off center.
draw chevron buttons at the frame height instead, spaced and disabled at
the ends of the scroll range like the rest of the tool buttons.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
* cabana ui: 1.5px sparkline stroke
* cabana ui: stable sparkline scrolling, one sample per column, aliased spikes
Scroll the polyline by whole physical pixels with each sample's subpixel
phase fixed to its timestamp, thin to one sample per pixel column, and
draw near-vertical segments aliased so spikes stay crisp while smooth
curves keep antialiasing.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: fix spelling
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: size the value column for the widest value a signal can produce
Sizing it to the values in the last message moved the sparklines every
time a value changed length.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: center the tab bar scroll chevrons and give the buttons a frame
The icon font glyph sits off center in its padded advance, so the
chevron is drawn in the button rect. The buttons take the theme's
button background and border like the tool buttons next to them.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: darker grid lines between table cells
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: the gap between the video and the charts matches the side padding
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: plus-square icon for the new chart button, like the remove all button
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: video toolbar buttons as tall as the charts toolbar buttons
The buttons carried 10 px of vertical padding, so the hover rect was
much taller than the glyph and the row looked off center.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: the time display in the mono font at the ui font size
With proportional digits the time changed width as it ticked and the
items after it moved.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: closing a floated side panel docks it back, no hide tab bar button, the charts container never scrolls
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: the message tabs use the TabBar widget, chevron scroll buttons for both tab bars
The chevron scroll buttons move into a scrollable tab bar widget that
both use. The wheel scrolls the tabs, both tab bars get Close Other
Tabs, and the Msg tab is labeled Messages.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: deliver a focus loss right away on macOS, holding it back swallowed the first click in a popup
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: draw the remote routes dialog at the modal's level, opened inside the tab's child window it never appeared
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cabana ui: --no-cache turns off the local route file cache, same as replay
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cabana: parse merged segments on replay's merge thread, the parse on the main thread dropped frames on every merge while downloading
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cabana ui: Esc leaves full screen on all platforms
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* cabana ui: an app bundle on macOS so the menu bar and the dock show Cabana
A bare binary shows its file name there. The build wraps _cabana_ui in
a minimal Cabana.app with an Info.plist and the launcher runs that.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* Revert "cabana ui: an app bundle on macOS so the menu bar and the dock show Cabana"
This reverts commit 1a45540125847ca46fcd4d7ac0522918e315057d.
* cabana ui: the app menu on macOS says Cabana
A bare binary gets an info dictionary with its file name in it, which
glfw reads for the app menu. The name is set there before glfw brings
up cocoa.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: macapp moves into ui/util
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: setMacAppName lives in ui/util.cc
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: apply the code review
Bugs: an empty decode batch dereferenced vals.front(), the chart x comparator
took a float, reserve used capacity(), the bit text went black while resizing
a signal, byte cells under-sized non-multiple-of-8 payloads, the CAN speed
lookup read the data speed table, the export tooltip never showed while
disabled, and the thumbnail parser and the signal search joined ad-hoc
threads on the render thread. Both now run off the pool; the search fans out
over it from its own thread.
Shared helpers in ui/util: the tool bar with its overflow menu, the menu
button, modal dialog setup, Escape inside dialogButtons, drawText,
drawElidedText, color markers, alignRight, clearableInput,
disabledItemTooltip, inputTextMultiline, tableHeadersRow with the right
click, withAlpha, boldFont, the Cocoa full screen glue, the messages panel
id. utils gets hexByte, guarded and nonEmptyDBCFiles; threadpool gets
parallelFor.
The help overlay leaves mainwin.cc for its own file, the stream teardown is
one releaseStream(), the charts own their views through unique_ptr, the
find-similar-bits message list follows the source bus, the clipboard copies
every open DBC, and the Qt vestiges (event names, model API, protected
sections, restating comments, stale TODOs, mixed member naming) are gone.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: collapse the model/view/delegate split
The Qt-era model, view and delegate classes were ported one to one, but
imgui draws a table in one function. MessageView, MessageViewHeader and
MessageListModel's string API fold into MessagesWidget over a plain
MessageList; HistoryLogModel folds into LogsWidget; BinaryViewModel and
BinaryItemDelegate fold into BinaryView; SignalItemDelegate folds into
SignalView, and SignalModel keeps only the item tree with three predicates
instead of the flags bitmask; FindSignalModel becomes SignalSearch. The
byte cell painter is a set of free functions shared by the two tables.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
* cabana ui: load the route behind the window
The route file listing is an HTTPS round trip to the comma API and took
the whole first second of startup before the window existed. The replay
branch of main() now hands run() a loader that a worker executes after
the first frame; the window maps in ~0.1 s instead of ~1 s. On a failed
load the app stays open in the No Stream state with the error box
instead of exiting before a window ever appeared.
* cabana ui: fix the live stream video controls layout
* cabana ui: only copy dbc to clipboard with a single file
* cabana ui: report route loading failures instead of hanging
* cabana: launch the imgui frontend by default, --legacy for Qt
---------
Co-authored-by: Trey Moen <trey@Treys-MacBook-Pro.local>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
E.118 specifies decimal digits, but many real SIMs carry hex characters
in EF_ICCID (e.g. China Mobile's 898600B5... range, some MVNO/IoT SIMs).
AT+QCCID returns them verbatim, and the strict isdigit() check blanked
the ICCID, leaving the modem daemon stuck in INITIALIZING forever and
cellular dead. ModemManager parses ICCID as hex for the same reason.
Verified on a comma four with a China Mobile SIM (EG916Q-GL): previously
stuck retrying 'identity read incomplete', now dials and passes traffic.
The destructor built its cleanup commands as "rm %s -rf", with the flags
after the operand. GNU rm permutes arguments so this works on device and
in CI, but BSD rm on macOS stops option parsing at the first operand and
treats "-rf" as a second filename:
$ mkdir -p /tmp/rmtest/sub && rm /tmp/rmtest -rf
rm: /tmp/rmtest: is a directory
rm: -rf: No such file or directory
exit=1
So nothing is removed, and each of the four calls prints two errors plus
"system command failed (256)" from check_system. Every run of a tool that
owns an OpenpilotPrefix (replay, cabana) leaks its params dir, its
comma_home and its /tmp/msgq_ dir; 33 of each had accumulated on my
machine.
Pass the flags first.
* system: add chestnut offroad alerts
* system: refine chestnut offroad alerts
* system: refine chestnut power alerts
* system: confirm chestnut power recovery from PCIe
* system: detect missing chestnut power from INA voltage
* Revert "ui: fix text and icon overlap on button (#38628)"
This reverts commit d9c4120f89.
* simple
* can do this
* fix eliding
* Revert "fix eliding"
This reverts commit b271a350182ad87f9942d7363383ee8ec72d0e36.
* clean up
* clean up
* ui: fix inverted DM head tracking on RHD
b29d0a17af switched the mici driver state renderer from computing pose
from driverStateV2 itself to consuming driverMonitoringState.visionPolicyState.pose.
That published yaw is mirrored for RHD (policy.py:269) so the asymmetric yaw
offsets/thresholds can be tuned in a single frame, so the UI's unconditional
sign flip left RHD head tracking backwards. LHD was unaffected, which is why
the refactor looked like a no-op.
Undo the mirror on the UI side.
* Apply suggestions from code review
* fold the sign into the RHD multiplier
* dmvision c703670b-6a4a-4743-8f79-69d5486cadf7 (a62f9948 with new sleep head)
* dmvision c703670b-6a4a-4743-8f79-69d5486cadf7 (a62f9948 with new sleep head and same batchnorms)
* add h264 profile option and different profiles for comma3X and comma4
* update to high profile
* resize to be baseline 3.1
* fix
* gop 15 and clean
* revert gop 15
* fix white space
* simplify teleoprtc
* bump teleop
* Process eSIM notifications after profile operations
* Clarify sleep purpose after operation in execute_and_process_notifications
Added comment to clarify the purpose of the sleep.
* chestnut: production flasher with ROM bootloader recovery
* hardwared: flash chestnut firmware offroad
* comments
* single FW version
* hardwared: fix chestnut flasher path
* chestnut: update firmware to 1d368808
* hardwared: log the flasher output
* comment
* hardwared: flash in a thread
* match panda flashing style
* fix
* chestnut: defer signals and harden the flasher
* chestnut: detect bootloader mode and retry
* comments
* chestnut: name the config dir for its content
* chestnut: name the ROM bootloader consistently
* chestnut: clearer name for the device lookup
* chestnut: spell out the low level names
* chestnut: claim the interface, not the device
* exp mode confirm dialog
* full
* base
* clean up
* too diff
* fix?
* rmnl
* only toggle if exp mode is confirmed
---------
Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>
Two macOS-only cabana fixes:
- signalview.h: rename the SignalView::setModel parameter that shadowed
the `model` member. Newer clang (Apple clang 21) folds
"parameter shadows a field" into plain -Wshadow, so with -Werror this
breaks the macOS build. Older clang (CI's image) only flags it under
-Wshadow-all, which is why CI stays green. The rename is clean under
any -Wshadow level and gcc.
- devicestream.cc: the --zmq bridge path had a stale `openpilot/`
segment (`../../openpilot/cereal/messaging/bridge`), a monorepo-layout
leftover that resolved to a non-existent triple-`openpilot` path, so
the bridge failed to exec and live streaming was broken. Drop it to
`../../cereal/messaging/bridge`.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* add support for multi-track with wideRoad single track as default
* remove stale debug mode
* multi track support
* make cameras list optional
* remove multi camera support from athena, make it only local
* update to teleop master
* replace PIL in emoji
* replace PIL in test
* fix uv.lock
* fix uv lock
* revert emoji removal from this P R
* revert merge artifact
* remove more merge artifacts
* fix lock
* red diff
* red diff +
* try cffi params
* ctypes
* lil more
* rm cython
* lil more
* that was fine
* lil more
* Drop unrelated Params concurrency changes
* Clean up ctypes Params build integration
* just c
* Clarify Params exception translation
* Expand Params C wrappers
* lil more
* power draw first
* speedup encoder
* speed up hardware test suites
* refresh IRQ actions after hardware setup
* skip unused big model in device checkout
* move LFS pruning off device checkout path
* wtf is that for?
* reverts
* locality
* rm that
* ci: test op setup with empty submodules
* ci: init submodules via op setup instead of checkout
* ci: keep submodules for build release
* ci: drop explicit submodules: false
* ci: remove checkout comment
* ci: update GitHub Actions to latest versions
Bump all workflow actions to their latest releases, including major
updates for checkout (v7), upload-artifact (v7), github-script (v9),
action-download-artifact (v21), and thollander comment action (v3).
* ci: pin third-party actions to commit SHAs
Keep official GitHub actions on version tags, but pin community
actions to immutable commit hashes for supply-chain safety.
* Reapply "usbgpu: never materialize model weights in RAM" (#38271)
This reverts commit 8196b743af.
* release oob buffers as we load them
* don't leak fds
* lint
* bigger big
* don't roundtrip the features to qcom
* Revert "don't roundtrip the features to qcom"
This reverts commit 0d2ca73d7a71bd32c46beda425957339f1dbbcf8.
* tmp bump allowed lag
* faster
* Revert "tmp bump allowed lag"
This reverts commit b407702a0242074eb12abdff14e61789fc05109a.
* Reapply "usbgpu: never materialize model weights in RAM" (#38271)
This reverts commit 8196b743af.
* release oob buffers as we load them
* don't leak fds
* lint
* 10s faster loading time for big model
* bigger big
* don't load big model in ram
* patch tg to allow out of band buffers
* dump/load oob
* roundtrip pkl to disk
* streaming chunking
* to_disk
* fixup
* linter
* Revert "bigger big"
This reverts commit 7a8147dd7fab88db963d512ab4008de940714025.
* athenad and webrtcd updates
* remove feature stream services from webrtcd split
* stream encoder thread
* reduce diff
* wire webrtc to livestream camera encoder
* request livestream camera switch service
* remove camera list in favour of init camera field
* remove cors
* clean
* remove unused
* remove extra try except
* add back exception trace
* add stream road camera info to stream cameras
* fix
* clean diff
* clean diff
* add testJoystick only on body
* fix camera list
* remove reference to future service
* encode all cameras and swap in video track in webrtc
* clean
* explicitly gate bridge send
* clean leftover
* initial idea
* add a watchdog to kill the process after disconnect or onroad
* start camerad and stream encoderd as well
* add message handler even without bridge
* remove carState when offroad
* transition to onroad works for body, kill for car not working
* turn off stream processes on started car
* cereal messaging sub field allow list, carparamspersistent, ping webrtcd to see when awake
* fix imports
* fix and increase max retries
* reduce max retries, increase timeout
* timeout on wait for connection
* remove cereal sub message and update process config to run on body ignition
* debounce 5s on teardown
* update error messages
* fix tear down crash
* clean
* clean
* fix lint
* split out and refactor
* simplify
* rm extra line
* use threadpoolexecutor to handle exceptions
* simplify
* improve
---------
Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>
* ui: ensure zoom accommodates full view area in AugmentedRoadView
* fix: use _content_rect
---------
Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>
* add param that requests keyframe
* set keyframe request to false on cleanup
* timing and errors
* video_enabled option and ability to add video later
* remove loggerd changes from this PR
* video enabled default to on, clean identifier
* request keyframe on start regardless of video enabled or not
* move error handling additions to different PR
* add param that requests keyframe
* set keyframe request to false on cleanup
* clean
* timing and errors
* remove other changes
* clean diff
* clean diff
* clean up live adaptation params and functions
* better function names
* fix syntax
* remove unused var
* clean diff
* improve lat maneuver report
* rm torque jerk
* add jitter maneuver
* add to 30mph
* add jerk
* color
* more state
* override
* abort reason
* abort
* rm text
* pad for small text
- patch `aioice` to only get host address of active network interface (wlan0 when on wifi, ppp0 when not)
- use common webrtc session between `startStream` and `addIceCandidates`
- first time connection speed ups
- move imports to top of file to move delay to webrtcd server start up not first connection
- warm up webrtc stack by creating a webrtc connection object to import all submodule classes
op.sh: don't eval CMD, run argv directly so special chars work
Previously op_run_command joined "$@" into a single string and eval'd
it, which broke on inputs containing shell metacharacters. For example,
'op esim --download "LPA:1$[rsp.truphone.com]($url)$QRF-SPEEDTEST" name'
would fail with a syntax error on the unquoted parens. Run "$@" instead
and only use $* for the printed display string. Also quote the unquoted
$@ in the other op_run_command callers.
* modeld: RL driving model with 3-file split
Split the driving model into vision + off_policy + on_policy ONNX
files and wire up the RL policy:
- 3-file model split (vision / off_policy / on_policy), replacing the
combined big_driving_policy/vision models
- compiler updates for the split models
- actually consume the policy action in modeld
- add desire state to the driving model
- model iterations (smoothness, off/on-policy weight updates)
* modeld: update driving model
* 1e72cf5a-785f-45ea-888f-28cdb14785de/100
* tinygrad hack
* fix parsing
* looser timing
* big
* Remove unnecessary modeld rebase changes
* Tighten modeld split cleanup
---------
Co-authored-by: Comma Device <device@comma.ai>
Co-authored-by: Armandpl <adpl33@gmail.com>
* bitrate controller
* ability to set the quality so a certain level, make message handler a switch case
* distinct bitrate levels and exponential backoff of raising bitrate
* add med level threshold
* fix
* fix
* simplify
* clean
* remove CarState.brake refs
Intending to deprecate this field since it's set incorrectly by
most car implementations, and (essentially) unused in openpilot.
Everything should be using CarState.brakePressed.
See https://github.com/commaai/opendbc/pull/3338.
* fix test_models.py
* hardware: read network info without NetworkManager DBus
* hardware: simplify wpa_cli SSID escape decoding
* hardware: restore cellular block in get_network_type to match master
* hardware: factor wpa_cli helper for key=value parsing
* hardware: comment SSID byte conversion for keyfile match
* hardware: comment NM metered enum values
* hardware: use check_output for ip route and wpa_cli helpers
* hardware: read default route iface from /proc/net/route
* hardware: simplify default route iface parsing
* hardware: only check for metered == 1
* hardware: also look for *.nmconnection in /data/etc/NetworkManager/system-connections
* hardware: use nmcli for runtime metered guess on wifi
* socket
* cleanup
* poor
* lil more
* mv that
---------
Co-authored-by: Adeeb Shihadeh <adeebshihadeh@gmail.com>
* initial codex
* remove ISP delay from encode ms
* try something simpler
* not allowing increase bug
* decrease bitrate increase rate
* upsample only on consecutive good checks
* rearrange
* clean
* Revert frame timing processing time change
* clean
* remove bitrate option and just do it on --stream
* make set_bitrate a lot more simple
* clean default impl
* further cleaning
* struct for bitrate to pass ci
* refactor a bit
* add comment
* small clean
* add init()
* increase bitrate
* update ffmpeg message
* remove unnecessary include
* modify set_bitrate
* change to Param
* remove getInt helper; add back in diff pr
* reword unclear error message
* move webrtcd change to new branch
* Reapply "modeld: split warp" (#38085)
This reverts commit d489dd8909.
* don't time make_random_inputs
* depend on chunk targets
* also depend on compile_modeld's dependencies
* athenad and webrtcd updates
* remove feature stream services from webrtcd split
* stream encoder thread
* reduce diff
* wire webrtc to livestream camera encoder
* request livestream camera switch service
* add frame timing headers
* rfctr
* diff
* clean
* remove camera list in favour of init camera field
* remove cors
* clean
* remove unused
* remove extra try except
* remove video_tracks unused
* add back exception trace
* add stream road camera info to stream cameras
* fix
* add back assert
* clean diff
* clean diff
* add testJoystick only on body
* fix camera list
* remove reference to future service
* video_tracks list add back
* encode all cameras and swap in video track in webrtc
* clean
* explicitly gate bridge send
* clean leftover
* rearrange video.py
* fix lint
* athenad and webrtcd updates
* remove feature stream services from webrtcd split
* stream encoder thread
* reduce diff
* wire webrtc to livestream camera encoder
* request livestream camera switch service
* remove camera list in favour of init camera field
* remove cors
* clean
* remove unused
* remove extra try except
* add back exception trace
* add stream road camera info to stream cameras
* fix
* clean diff
* clean diff
* add testJoystick only on body
* fix camera list
* remove reference to future service
* encode all cameras and swap in video track in webrtc
* clean
* explicitly gate bridge send
* clean leftover
* make local bodyteleop work still
* athenad and webrtcd updates
* remove feature stream services from webrtcd split
* remove cors
* clean
* remove unused
* remove extra try except
* add back exception trace
* clean diff
* clean diff
* add testJoystick only on body
* fix camera list
* remove reference to future service
* zero ll patched big model
* probe in a subprocess so usbgpu lock gets released
* compiles
* runs
* num_jobs gets overwritten, use side effect
* poll tg devices
* make sure build crashes on missing gpu
* fine not to rely on Device.default
* seperate tg env for each model runner
* comment
* Revert "seperate tg env for each model runner"
This reverts commit f6470cc4258eaeb3e8e37907ef370871c9af5aa4.
* env is shared, gate on flag
* no fallback warp dev must be set
* build for current device only, unless pc/release
* comment
* list
* listen for plug in
* add icon to status bar, read params on every frame (?)
* log available devices
* try copy out when loading?
* Revert "log available devices"
This reverts commit e8c52a5d59456d4820ecb13b99a6c46ea1386a20.
* Revert "try copy out when loading?"
This reverts commit 518f403aa03faeda1950fe3dbce0d9e4c1584455.
* don't trigger device probe/caching on modeld prepare
* re-export with ll and road edges
* dont cache devices in manager process
* get USBGPU from params
* no usbgpu env
* missed one
* sconscript don't poll
* unconditional env
* always explicitely set devices on input tensors
* set DEV so amd uses right compiler and iface??
* fix flag
* bump tg
* rm xdg_cache_home
* tg don't bump all the way
* missing gmmu=0 at compile time
* dm set dev
* tg backend
* update gitignore
* missing import
* unused imports
* rely on Device.DEFAULT at compile time (already the case bc onnxrunner)
* comments
* dm warp needs DEV set too
* build both smol and big
* misc typos
* set dev at compile time
* don't need
* DEV=CPU when getting metadata, ensure we don't grab gpu lock
* this would also grab lock
* put bool
* warp compile always prepare only
* missed one
* poll ui
* missing here
* don't force usbgpu at build time
* tmp patch fetch_fw
* catch all, follow hardwared patterns
* simpler
* compile make input queues
* revert this
* group this more readable
* rm empty line
* make dummy frame using numpy
* revert compile make input queues
* no compiler at runtime
* cleanup
* fine to rebuild all on change to device node for now
* fix usbgpu_present
* fix sconscript
* no size in header stream decompress
* DEBUG=2
* minimal viable feedback
* egpu gray
* oops
* gotta do this actually
* modeld build only depends on modeld devices
* don't ship onnx to release? or chunk
* don't need
* can only set compiler on dev=
* none device works, will use default
* make linter happy
* chunk agnostic onnx input to compile_modeld
* chunk big onnx
* +x chunker
* fix #!
* and don't ship chunked onnx to release
* firmware now in correct location
* better err on missing onnx/chunk
* SConscript also need to accept chunked onnx
* metadata also need to load maybe chunked
* dedupe cmd
* this needs to be on cpu
* devices are set in the tgflags, we already depend on them
* rebuilding on changed order is fine
* read file chunked can already load either chunked or not
* chunk all big onnx
* less confusing
* unused import
* python device to load onnx bytes
* default device for runners, python for metadata
* why not
* chunked to shm
process_replay/migration: drop transitive dep on long_mpc/acados
migrate_longitudinalPlan only needs get_accel_from_plan + CONTROL_N_T_IDX,
both available from drive_helpers + ModelConstants directly. Importing
from longitudinal_planner instead pulled in long_mpc -> libacados.so as
an eager side-effect, which downstream log-migration consumers (e.g.
commaai/xx pipeline) shouldn't pay for.
* serverless code sync
* rm stuff
* we have ls-files for this
* rm
* no debounce
fix
* fix
* no del
* no init
* clean up
* clean up
* smol
* fixup
* rm
* clean up
* any change
* rm arg
* rm
* rm
* speedup slowest
* less setup
* rm som reset logging
* simpliy a lil more
* lil more
* down to 36s
* sleeeeeeep
* finishing touches
* oopsie
* ty fix
---------
Co-authored-by: Comma Device <device@comma.ai>
* filter lead prob
* rename
* try correcting model bias
* Revert "try correcting model bias"
This reverts commit b5e9b7147e58f200ca2e02ccea8adf88be99e206.
* fast gain slow lose
* cmt
* deb
* rename
* rename
* end
ui: nonblocking writes for ExperimentalMode + DriverView toggles + cycle-restart
All four put calls fire from the main render thread on user interaction
and block on disk fsync, causing visible UI frame spikes.
Each consumer is safe under nonblocking:
- onboarding inactivity_callback: write-and-forget (~25 ms saved)
- home long-press exp toggle: ui_state.experimental_mode owns visual state (~10 ms)
- onroad exp_button: $held_mode + selfdriveState owns visual state
- restart_needed_callback (OnroadCycleRequested): cross-process signal,
consumer is selfdrived which polls the param
BigParamControl-driven toggles in settings (developer.py, toggles.py)
are intentionally left blocking — those widgets refresh visual state
from disk every frame to mirror external changes, which would race a
nonblocking write.
* snap logic in scroll panel
* match previous snap speed
* velocity lookahead makes scrolling on main layout super easy
* turn off
* remove velocity lookahead
* cmts
* remove handle out of bounds
* cmts
* old style
* back
* clean up
* re-use
* revert
* finish snap when settings is clicked while moving
* rm cmt
* log raylib fps
* log fps from frame time
* whitespace
* or just log frame time?
* init pubmaster in init window
* yield timings
* bump ordinal
* dont log on screen off
* UInt
* lint
* /0
* oops
* oops2
* more precise raylib frame time, can get fps with 1/ft
* don't crash on screen off
* NL
* no _
---------
Co-authored-by: Shane Smiskol <shane@smiskol.com>
modem.py: retry identity read until valid; stay in INIT on partial reads
Under rapid SIM hot-swap stress, AT identity reads (CGSN/QCCID/CIMI/GMR) can return
empty values or echo the command itself. Validate each field before accepting it,
retry the whole sequence in a loop, and stay in INITIALIZING (rather than progressing
to SEARCHING) until imei and iccid are populated. Prevents stale/empty identity from
being published to /dev/shm/modem after a glitchy reinit.
* ui: speed up `mici/AugmentedRoadView` by optimizing _calc_frame_matrix caching (#36669)
speed up AugmentedRoadView by optimizing _calc_frame_matrix caching
* ui: apply rect.x/y as a 2D screen offset post-projection
Removes the parent rect's screen position from the cached
video_transform passed to ModelRenderer. Instead, ModelRenderer
applies (rect.x, rect.y) as a 2D offset to projected_points at draw
time.
Why this works: the rect.x/y term in video_transform gets multiplied
by P_calib[2] before the perspective divide, then divided by the same
value, which cancels out to a simple additive shift on the final
screen coordinate. So adding x to video_transform[0,2] is equivalent
to adding x to screen_x post-projection.
Net effect: the cache key in _calc_frame_matrix no longer needs to
include rect.x/y. Cache stays hot under translation (e.g. swiping
between layouts), and the model overlay tracks the camera at 60Hz
because the offset is updated cheaply each frame.
This addresses the original revert reason for #36669 (model overlay
visually desyncs from camera during a home<->onroad swipe).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* ui: update model screen offset every frame, not just on cache miss
set_screen_offset() was called inside the cache-miss path of
_calc_frame_matrix, so the offset only updated at ~20Hz (calib publish
rate). The model overlay visibly lagged the camera during a swipe.
Move it out of the cached path so it updates each frame.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fully clean up
* shorter cmts
* make non
* clean up
* fix ty
---------
Co-authored-by: Dean Lee <deanlee3@gmail.com>
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* ui: keep arc_bar_pts cache hot when bar is translated
The LRU cache key included (cx, cy), which change every frame when
the parent widget's rect translates (e.g. during a scroll animation
in MainLayout). That made every cache lookup miss, recomputing the
arc geometry twice per frame and dropping fps.
Compute the shape at origin (so the cache key only depends on the
geometry: radius, thickness, angles), then translate to (cx, cy)
after. Cache stays hot under translation.
Measured on comma four (engaged, scrolling): ~5fps recovered while
moving between layouts.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fully clean up
* fully clean up
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
* fix frame under text
* claude always taking the hard path
* cmt
* it still shows on start
* revert
* fix
* wait to show
* already a bug, don't fix here
* cmt
* cmt
* update blob
* fixed
* didnt catch this
* add back
* needs BLC built in
* for real
* attempt2
* clean up override
* this should keep ox as was
* disable for OX
* update descs
---------
Co-authored-by: Comma Device <device@comma.ai>
* spellings
* unused
* no roll
* lil more
* lil more
* one more
* policy enum
* better trans
* set_timer -> set_policy
* set_timer -> set_policy
* no yaonet
* del redundant code
---------
Co-authored-by: ZwX1616 <zwx1616@gmail.com>
* c4 body ui
* clean up diff
* clean up
* default bodyview debug with True
* remove battery indicator and fix close settings bug
* organize debug file
* clean
* clean up frame index
* remove unneccessary is body check
* update bodyview
* remove joystick_debug_mode based events
* remove debug script
* apply suggestions
* clean diff
* clean diff
* move ignition message
* save a line
* remove visibility set and fix tici body face when sidebar open
* remove explicit textColor offroad message
* remove unused imports
* revert pairing dialog
* apply suggestions
* add body specific icon
* add body icon for homescreen
* set mode for state on tici after on body changed
* tiny
* tweak
* v
* tweaks
* icon ratio was wrong!
* same order
* rm
* apply suggestions
* remove commented lines in animation
* onroad click callback was on home bug and fix setup widget settings call back hack
* fix body changed
* one liner
* clean up
* formatting
* if false
* revert to master + reimplement
* close sidebar on body home tici
* make ignition message bigger on c3
* flip eye direction when turning
---------
Co-authored-by: Nick <nickorie@gmail.com>
Co-authored-by: Shane Smiskol <shane@smiskol.com>
* compile_modeld.py
* update estimates
* missing image=2?
* Revert "missing image=2?"
This reverts commit 2f5952eb63ba1e3f24cbf5769e6b5e9170d7f0a6.
* Revert "update estimates"
This reverts commit 1f72feef2ffdec6126e3c941e899b46ace7b4b65.
* Revert "compile_modeld.py"
This reverts commit f10541502efca02725f368deda2a21d1f786f57d.
* load warp in ModelState init
* dead code
* prep
* compile modeld
* update SConscript
* tmp save plot locally
* Revert "tmp save plot locally"
This reverts commit ec22f15161ad3b0241a097546b35860f989219f5.
* openpilot hacks?
* no float16
* tmp more chunks
* Revert "tmp more chunks"
This reverts commit 9e1d9b4d0dc36ff530d2a70b565fbfabd7afb00d.
* Revert "no float16"
This reverts commit 6204956e98e3c0818ed1985ede8eeccb810f63e3.
* realize boundaries
* Revert "realize boundaries"
This reverts commit ffaa19259eba70944e7793e8f51a0f87089531b3.
* prune=False?
* Reapply "tmp more chunks"
This reverts commit 2599c41cea93b4a6b4e946cdffc6a617663a7d23.
* tg bug?
* load first?
* Revert "load first?"
This reverts commit f643d082d76a424b23295e254179eb111e936e61.
* revert
* Reapply "tmp save plot locally"
This reverts commit 1b95b82ee58654bd908b1cb04ab0ddbcd1a5955d.
* 0 tol pc
* warp -> modeld
* rename
* bypass chunking?
* dont chunk
* Revert "dont chunk"
This reverts commit cc97fc67b3203456e123f02babe5c83b87c7e264.
* dont chunk
* debug
* Revert "debug"
This reverts commit b3c2f2e7a095fd32f8d8562a68fd1cca42357eac.
* Revert "dont chunk"
This reverts commit 42bd9b6f6ad0722c50348ba11ba7e2a64fdf997d.
* Revert "bypass chunking?"
This reverts commit ad5422a93483ffd8a59ba62e5fb72ced3b5d04d0.
* corrupt model outputs
* Revert "corrupt model outputs"
This reverts commit 245feb94480e02f83a20b65a9488652bcbfc88b0.
* image=0 for warp, match master
* dedupe enqueue
* pass traffic convention
* tg buffer for desire
* dedupe buffer creation
* compile_modeld: nuke stale cached pkl before compiling
The UNSAFE CI checkout keeps gitignored files (.pkl, .sconsign.dblite),
so stale pkl files from previous commits can persist and be reused
instead of being recompiled. Delete them explicitly before compiling.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* test vs compile
* all outputs need to be different on different inputs
* randomize numpy inputs
* randomize on every step
* SConscript: nuke stale pkl+chunks before compile_modeld
Move the stale artifact cleanup from compile_modeld.py into the
SConscript build command. This ensures stale gitignored pkl and chunk
files are deleted even if scons decides to skip the compile step
(due to a stale .sconsign.dblite from UNSAFE CI checkout).
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* compile_modeld: restore Context(IMAGE=0) for warp
The warp operations must run under IMAGE=0 to avoid QCOM image texture
optimizations that corrupt the output buffer after ~33 frames.
This was accidentally commented out in a855173.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* modeld: create SubMaster before model loading
Move PubMaster/SubMaster creation before the model loading step.
During model loading (3.5s+), process_replay may send liveCalibration.
If SubMaster doesn't exist yet, the message is dropped and the warp
transform stays as zeros, producing garbage warped images.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* Revert "modeld: create SubMaster before model loading"
This reverts commit 968c987c2fbb3fce141c4e345d10ddea559b6c50.
* stale metadata?
* claude debug
* Revert "claude debug"
This reverts commit 49e754c6affa45a8ea8834588a00227b8090b17a.
* Revert "stale metadata?"
This reverts commit 870388513c0d4a67dcf970cd277b6db56cb2b478.
* modeld: realize jit outputs before parsing
* Update modeld.py
* modeld: fix NameError by removing redundant MODELS_DIR definition
* test buffers in test vs. compile
* 2x inputs before running
* fixup 2x inputs test
* realize onnx weights?
* Revert "realize onnx weights?"
This reverts commit 49c8b9a505db38ff22f342db011a3a6b6526d398.
* move openpilot_hacks flag to sconscript
* stricter test vs compile
* correct timings
* more run more fail?
* Revert "more run more fail?"
This reverts commit 9e94bb63940751ec29e81b634c42449113e1f2e5.
* numpy shenanigans
* correct shapes
* dont assert timings for now
* Revert "correct shapes"
This reverts commit 5b9ff6c84c0022327d21801d179e9e51c39e8f78.
* Revert "numpy shenanigans"
This reverts commit b4f6fb3078d7e9b09698895b88728fd8eea8c8a8.
* no need to nuke
* comment unused
* don't use NPY device
* copy instead of from_blob
* to device before jit
* Revert "to device before jit"
This reverts commit 7a59ed9b1ac88657b5a3917986b6ff92e59a2ee3.
* Revert "copy instead of from_blob"
This reverts commit 196c4892a06ffba89ef631876372cecf137cc1b4.
* Revert "don't use NPY device"
This reverts commit 18abf43bbac46ad47a60c03dd8d1ef40b3f59227.
* 3 runs is enough
* no_memory_planner=1
* lint
* restore model_replay.py
* on policy -> policy
* unused
* prepare only enqueues full images
* warp with image=2?
* unused args
* test vs compile, check different inputs different outputs
* avoid uop cache collision
* dont need realize here
* misc
* input queues diverged
* strict zip
* monkey patch for now
* memory planner
* prev desire correct order
* dedupe pkl paths / compile targets
* don't change behavior, warp and enqueue frames when skipping model eval
* actually prepare only
* warm up warp jit
* correct path
* oops
* explicit warmup
* need continuous + can't have dupplicate jit inputs
* whitespace
* bufs -> input_queues
* master tg
* /N_RUNS
* bump tg, remove uop cache patch
* more readable
* Revert "bump tg, remove uop cache patch"
This reverts commit 499acca2591becd389de4025943f9e776a5b337c.
* missing dep
---------
Co-authored-by: Bruce Wayne <harald.the.engineer@gmail.com>
Co-authored-by: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
* turn off ir leds on notCar
* reduce startup time on notCar
* fix: check notCar once after onroad is true
* save a disk read by using existing is_onroad
* Revert "OP model 7 (#37760)"
This reverts commit 052692b25d.
* Revert "OP model (#37740)"
This reverts commit cb32793300.
* dead
* parse_model_outputs: drop extra space
The plural `git/refs/heads/{bucket}` endpoint does prefix matching
and returns a list when multiple refs share the prefix, which makes
`r.json()['object']` raise TypeError. Switch to the singular
`git/ref/heads/{bucket}` endpoint so we only match the exact bucket
and get a clean 404 otherwise.
2026-04-12 15:30:29 -04:00
10518 changed files with 2831007 additions and 165653 deletions
The comma-side Jetlink client and USB transport in `jetlink_repo` are by Zeph Leggett and the [Jetlink contributors](https://github.com/zoompilot/jetlink). They are vendored from revision `9cfde98eec1d7cc714d800504944cf1a3944513d` under the MIT license, retained in [jetlink_repo/LICENSE](jetlink_repo/LICENSE). StarPilot's adapter (`openpilot/starpilot/models/jetlink_adapter.py`) builds on Jetlink's openpilot layer and owns its param names, Always On Lateral rule, model runner reset, settings and alerts. Companion applications and inference servers remain distributed by Jetlink.
## MRR35 Hyundai angle and radar
The four MRR35 angle platforms (2026 Genesis GV70, second-generation electrified GV70, 2025 Genesis GV80, and Hyundai Ioniq 9) adapt Hyundai CAN FD angle-control design and the MRR35 radar DBC generator from the locally frozen StarPilot source snapshot `678af783`. That snapshot traces its Hyundai extension work to sunnypilot/opendbc `hkg-angle-steering-2025` at `cc4b08625a98e94b318cab15e45e05dad58042bd`; its historical import revision is not established. The relevant upstream contributors include Haibin (Jason) Wen, Shane Smiskol, and DevTekVE. Their authorship remains attributable to the upstream history; they do not maintain or endorse this port.
The copied MRR35 generator source and DBC definition are at `opendbc_repo/opendbc/dbc/generator/hyundai/hyundai_mrr35_radar.py` and `.dbc`. The host angle adapter and native safety profile are adaptations, not copies of a generic DBC or of unrelated vehicle support. Applicable upstream notices are preserved in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
## Ford support adapted from BluePilot
Ford lateral control vendors [BluePilot lateral](https://github.com/BluePilotDev/bluepilot-lateral) v0.2.1 (`2b9dc6cc784faac47a1885d0bda0561f7b76f74d`), developed by Alan Polk (`alan-polk`) and the BluePilot contributors. Alan provided the [StarPilot integration in PR #269](https://github.com/StarPilotHQ/StarPilot/pull/269); his commits and authorship are retained. BluePilot maintains the library and its integration; StarPilot keeps its settings, AOL policy, packaging and integration tests separate. Complete contributor and license notices remain in [the library credits](opendbc_repo/opendbc/bluepilot_lateral/CREDITS.md) and [license](opendbc_repo/opendbc/bluepilot_lateral/LICENSE).
The Ford CAN FD camera-lead radar filtering is adapted from StarPilot `2efe6017bfe14cee4807833e05f9a7f6ba255cdb`. Its lineage includes BluePilot's `bp-7.0` Ford work, principally by Alan Polk, with contributions from John Christman, Jacob Neulight, tonesto7, Haibin Wen and other sunnypilot and BluePilot contributors. The reference snapshot is [`e1d051d7ba270261b4455068bd68f1a58db15a4a`](https://github.com/BluePilotDev/bluepilot/commit/e1d051d7ba270261b4455068bd68f1a58db15a4a); the exact checkout used for the earlier StarPilot import was not recorded.
The radar interface preserves that attribution while adapting to the current interfaces. Upstream contributors do not maintain or endorse this adaptation. Applicable notices are retained in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
## RAV4 Continental ARS510 receiver
The optional ACC-target and fused-track receivers and DBC preserve firestar5683's Dom integration from `b220a37f4bd119e402225a7aa16a7cd84878d0af` (Sea Turtles, 2026-10-13) and its ARS510 source attribution to ds-sebastian/ars510-radar `fb5b29c2b5999f34afeb6f7ac494a73949d66c67`, copyright 2026 ARS510 radar contributors. The complete MIT permission and warranty notice is retained in THIRD_PARTY_NOTICES.md. Startup/settings ownership is adapted to the current StarPilot architecture; no original firmware binaries or safety permissions are copied.
## Spotify Bluetooth output
Spotify's Bluetooth output binds the comma's BlueALSA A2DP source to the selected speaker through a process-local ALSA configuration, leaving the system ALSA default, soundd and the warning speaker untouched. That approach is adapted from firestar5683's RoadScore Bluetooth output (`roadscore/prototype/bluetooth_output.py`, firestar5683/StarPilot `RoadScore` at `249eb88b3a397db847e4a595d7e23b7e85d329e5`). The Soloist player itself is never included: each user downloads it from Spotify, and the PulseAudio server that carries its audio is vendored unmodified from Ubuntu (see THIRD_PARTY_NOTICES.md).
## Tesla Pre-AP pedal control
The Pre-AP configuration, pedal codec/feedback, VirtualDAS feedforward table, grade compensation, torque-zero estimator and vehicle-specific gains preserve firestar5683's intentional Dom work, including `7837a3c1aab`, `aed0b7fdb9e9` and pinned `b220a37f4bd119e402225a7aa16a7cd84878d0af`. Original author attribution: firestar5683 <168790843+firestar5683@users.noreply.github.com>. Startup/calibration ownership, current upstream button/PID interfaces and native command bounds are adapted to the current StarPilot architecture; the old permissive native gateway is not transplanted. Existing source/license notices remain intact.
## Subaru angle steering from Jacob Waller's fork
Subaru angle support includes implementation maintained and integrated by **Jacob Waller (jacobw)** in [jacobwaller/openpilot, jul-angle-based](https://github.com/jacobwaller/openpilot/tree/542816196a18dba5f6f0af093c1ece4f415b5a22), with its [opendbc implementation at 8acbeccdab45a47aa7af170d2d17b61f0eb5e4f0](https://github.com/jacobwaller/opendbc/tree/8acbeccdab45a47aa7af170d2d17b61f0eb5e4f0). Jacob authored the VehicleModel angle-limit type correction and integrated the Crosstrek support. The branch also retains the original angle implementation and follow-up work by **lukasloetkolben and elkoled**, alongside comma contributors; their authorship remains theirs.
The inherited source covers Subaru angle platform/firmware definitions, `ES_Status` cruise state, angle controller/CAN integration, vehicle-model limits and native angle enforcement. StarPilot has additional platform-specific limits, driver handoff, discovery and AOL safety guards; this attribution does not claim identical tuning or complete parity with the reference fork. Original source commits, author names, emails and titles are recorded in [the Subaru origin history](opendbc_repo/opendbc/car/subaru/ORIGIN_HISTORY.tsv).
The explicit StarPilot [jacobw Subaru adapter](opendbc_repo/opendbc/car/subaru/JACOBW.md) additionally integrates that branch's low-speed filter, target-first limiting and Crosstrek EPS envelope. Its preserved stricter guards and platform scope are documented alongside the code.
## StarPilot Ford demand tuning
The previous optional StarPilot Controller, formerly in `openpilot/starpilot/car/ford/starpilot_tuning.py`, restored demand tuning from StarPilot `8da0c9d916^`, descended from BluePilot bp-7.0 `e1d051d7ba270261b4455068bd68f1a58db15a4a` through StarPilot `2efe6017bfe14cee4807833e05f9a7f6ba255cdb`. Contributors to that lineage include Alan Polk, John Christman, Jacob Neulight, tonesto7 and Haibin Wen. The current protocol adapter is maintained by StarPilot separately from the pinned BluePilot library; this attribution does not imply BluePilot maintains or endorses the adapted tune.
The current StarPilot Controller replaces that historical curvature-primary tune with a separately maintained adaptation of BluePilot v0.2.1's angle strategy (`2b9dc6cc784faac47a1885d0bda0561f7b76f74d`, `core/lateral_angle_ext.py`). Alan Polk is the principal controller author; John Christman contributed angle-mode anti-stall work, Jacob Neulight the pinion/shadow measurement, Nathan Ingraham and Praeuner high-speed damping, and other contributors are retained in the pinned library's `CREDITS.md`. StarPilot adds bounded small-signal conditioning ahead of the unchanged angle limit chain. This adaptation is not maintained or endorsed by BluePilot; the pinned library and its imported hooks are unchanged.
1.**Supported Device:** a comma four, available at [comma.ai/shop/comma-four](https://www.comma.ai/shop/comma-four).
2.**Software:** The setup procedure for the comma four allows users to enter a URL for custom software. Use the URL `openpilot.comma.ai` to install the release version.
3.**Supported Car:** Ensure that you have one of [the 275+ supported cars](docs/CARS.md).
4.**Car Harness:** You will also need a [car harness](https://comma.ai/shop/car-harness) to connect your comma four to your car.
**StarPilot** is a custom fork of [comma.ai's openpilot](https://comma.ai/openpilot),
an open source driver assistance system.
We have detailed instructions for [how to install the harness and device in a car](https://comma.ai/setup). Note that it's possible to run openpilot on [other hardware](https://blog.comma.ai/self-driving-car-for-free/), although it's not plug-and-play.
**This is the StarPilot 7.0 development branch (`SecretGoodStarPilot`). Feature migration,
vehicle qualification, and Auto Learn validation are still in progress.**
Openpilot provides
* Automated Lane Centering
* Adaptive Cruise Control
* Lane Change Assist
* Driver Monitoring *without wheel nags*
Ford-specific lateral-control and vehicle-support work includes substantial adaptations from
[BluePilot](https://github.com/BluePilotDev/bluepilot/tree/bp-7.0), principally developed by
[Alan Polk](https://github.com/alan-polk) with additional BluePilot contributors. StarPilot's
implementation has since diverged, but that does not erase its lineage. See [CREDITS.md](CREDITS.md)
for the code-level provenance and [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md) for applicable
upstream notices and terms. BluePilot and its contributors do not maintain or endorse StarPilot;
please direct support requests for this adaptation to the StarPilot project.
Hyundai, Kia, and Genesis angle steering and related vehicle support include substantial adaptations
from [sunnypilot](https://github.com/sunnypilot/sunnypilot/tree/hkg-angle-steering-2025) and its
| `release-mici` | `release-tizi` | openpilot.comma.ai | This is openpilot's release branch. |
| `release-mici-staging` | `release-tizi-staging` | openpilot-test.comma.ai | This is the staging branch for releases. Use it to get new releases slightly early. |
| `nightly` | `nightly` | openpilot-nightly.comma.ai | This is the bleeding edge development branch. Do not expect this to be stable. |
| `nightly-dev` | `nightly-dev` | installer.comma.ai/commaai/nightly-dev | Same as nightly, but includes experimental development features for some cars. |
## Development
To start developing openpilot
------
Dependencies are included as ordinary source folders. No recursive submodule checkout is needed. [upstream-sync.json](upstream-sync.json) records the openpilot baseline and the source revisions used for vendored dependencies.
openpilot is developed by [comma](https://comma.ai/) and by users like you. We welcome both pull requests and issues on [GitHub](http://github.com/commaai/openpilot).
The familiar development entry points are retained:
* Join the [community Discord](https://discord.comma.ai)
* Check out [the contributing docs](docs/CONTRIBUTING.md)
* Check out the [openpilot tools](tools/)
* Code documentation lives at https://docs.comma.ai
* Information about running openpilot lives on the [community wiki](https://github.com/commaai/openpilot/wiki)
| Command | Purpose |
| --- | --- |
| `./build` | Build the device software using the matching AGNOS environment |
| `./build --panda` | Build the supported Panda firmware targets |
| `./c3` | Open the comma 3 interface on a development host |
| `./c4` | Open the comma four interface on a development host |
| `./onroad` | Open the onroad development view |
| `./dev` | Use the isolated host development environment |
| `./test` | Run local pre-push regression checks without changing device artifacts |
Want to get paid to work on openpilot? [comma is hiring](https://comma.ai/jobs#open-positions) and offers lots of [bounties](https://comma.ai/bounties) for external contributors.
See the [build guide](docs/how-to/laptop-device-build.md) for setup, prerequisites, and target-specific behavior. The [UI guide](openpilot/starpilot/ui/README.md) describes desktop previews and their limits.
Safety and Testing
----
## Testing and vehicle support
* openpilot observes [ISO26262](https://en.wikipedia.org/wiki/ISO_26262) guidelines, see [SAFETY.md](docs/SAFETY.md) for more details.
* openpilot has software-in-the-loop [tests](.github/workflows/tests.yaml) that run on every commit.
* The code enforcing the safety model lives in panda and is written in C, see [code rigor](https://github.com/commaai/panda#code-rigor) for more details.
* panda has software-in-the-loop [safety tests](https://github.com/commaai/panda/tree/master/tests/safety).
* Internally, we have a hardware-in-the-loop Jenkins test suite that builds and unit tests the various processes.
* panda has additional hardware-in-the-loop [tests](https://github.com/commaai/panda/blob/master/Jenkinsfile).
* We run the latest openpilot in a testing closet containing 10 comma devices continuously replaying routes.
StarPilot retains upstream vehicle and Panda safety interfaces and adds regression coverage for its custom behavior. The [safety workflow](.github/workflows/safety.yaml) and feature tests cover distinct parts of the system. A passing interface test, model comparison, or bench run does not establish driving support for every configuration.
<details>
<summary>MIT Licensed</summary>
Support depends on the vehicle configuration and enabled feature. An existing vehicle identifier or settings page does not by itself establish driving support. Model Laboratory currently exposes management controls, but paired driving inference remains unavailable pending integration and Chestnut validation.
openpilot is released under the MIT license. Some parts of the software are released under other licenses as specified.
## Third-Party Notices
Any user of this software shall indemnify and hold harmless Comma.ai, Inc. and its directors, officers, employees, agents, stockholders, affiliates, subcontractors and customers from and against all allegations, claims, actions, suits, demands, damages, liabilities, obligations, losses, settlements, judgments, costs and expenses (including without limitation attorneys’ fees and costs) which arise out of, relate to or result from any use of this software by user.
* Portions of this software include modified versions of the Material Design Icons provided by Google under the Apache License 2.0. A copy of the license is included in the `LICENSE-MDI` file.
* Ford support includes software adapted from BluePilot's `bp-7.0` branch. The source repository contains both a standard MIT notice and a separate custom SUNNYPILOT LLC notice; StarPilot preserves both in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
* Hyundai, Kia, and Genesis support includes software adapted directly from sunnypilot's HKG angle-steering branch and sunnypilot/opendbc. StarPilot preserves the applicable notices in [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
**THIS IS ALPHA QUALITY SOFTWARE FOR RESEARCH PURPOSES ONLY. THIS IS NOT A PRODUCT.
YOU ARE RESPONSIBLE FOR COMPLYING WITH LOCAL LAWS AND REGULATIONS.
NO WARRANTY EXPRESSED OR IMPLIED.**
</details>
<details>
<summary>User Data and comma Account</summary>
By default, openpilot uploads the driving data to our servers. You can also access your data through [comma connect](https://connect.comma.ai/). We use your data to train better models and improve openpilot for everyone.
openpilot is open source software: the user is free to disable data collection if they wish to do so.
openpilot logs the road-facing cameras, CAN, GPS, IMU, magnetometer, thermal sensors, crashes, and operating system logs.
The driver-facing camera and microphone are only logged if you explicitly opt-in in settings.
By using openpilot, you agree to [our Privacy Policy](https://comma.ai/privacy). You understand that use of this software or its related services will generate certain types of user data, which may be logged and stored at the sole discretion of comma. By accepting this agreement, you grant an irrevocable, perpetual, worldwide right to comma for the use of this data.
</details>
> This project uses software from Haibin Wen and SUNNYPILOT LLC and is licensed under a custom license requiring permission for use.
The openpilot rebase is in progress. Domathon is not yet a completed StarPilot release, and feature or vehicle presence alone does not establish full parity with Dom.
See the [project overview](README.md) for the current development entry points and the [history guide](docs/DEVELOPMENT_HISTORY.md) for the upstream boundary and feature organization. Release-specific support and validation will be recorded when a release is prepared.
# Upstream openpilot release history
The entries below are retained upstream release notes. They describe openpilot releases, not completed StarPilot 7.0 milestones.
Version 0.11.3 (2026-XX-XX)
========================
* Autospeed
* On supported cars, openpilot will intelligently decide what speed to drive based on the scene, including:
* stopping at red lights, and going at green lights
* slowing down for curves
* adjusting speed based on surrounding traffic
* This behavior was previously available with openpilot's "Experimental mode", but is now out of beta and always enabled
* With openpilot deciding which speed to drive, the car's cruise speed is treated as a maximum speed openpilot will not exceed
Version 0.11.2 (2026-10-14)
=======================
* New driving model
* Big model with 880M parameters
* Support for big models running on an external GPU
* Live stream cameras from comma connect
* Generate dashcam clips from comma connect
* Remote comma body control from comma connect
* New alert sounds
* CUPRA Born 2021-2023 support thanks to DaHansi!
* Volkswagen ID.4 2021-2025 support thanks to DaHansi!
Version 0.11.1 (2026-05-18)
========================
* New driver monitoring model
* Improved image processing pipeline for driver camera
* Improved thermal policy for comma four
* Acura MDX 2022-24 support thanks to mvl-boston!
* Rivian R1S and R1T 2025 support thanks to lukasloetkolben!
Version 0.11.0 (2026-03-17)
@@ -765,7 +799,7 @@ Version 0.5.10 (2019-03-19)
* New Driver Monitoring Model
* Support QR codes for login using comma connect
* Refactor comma pedal FW and use CRC-8 checksum algorithm for safety. Reflashing pedal is required.
Please see `#hw-pedal` on [discord](discord.comma.ai) for assistance updating comma pedal.
Please see `#hw-pedal` on [discord](https://discord.comma.ai) for assistance updating comma pedal.
* Additional speed limit rules for Germany thanks to arne182
* Allow negative speed limit offsets
@@ -1098,7 +1132,7 @@ Version 0.2.1 (2016-12-14)
Version 0.2 (2016-12-12)
=========================
* Car/Radar abstraction layers have shipped, see cereal/car.capnp
* Car/Radar abstraction layers have shipped, see openpilot/cereal/car.capnp
Jetlink's comma client is copyright (c) 2026 Zeph Leggett and distributed under the MIT License. Its complete license and permission notice are retained in [jetlink_repo/LICENSE](jetlink_repo/LICENSE); source and revision are recorded in [CREDITS.md](CREDITS.md#jetlink-external-inference) and `upstream-sync.json`.
## MRR35 Hyundai angle and radar
The MRR35 radar generator and related Hyundai angle integration derive from the frozen StarPilot snapshot `678af783`, with relevant sunnypilot/opendbc angle-branch ancestry at `cc4b08625a98e94b318cab15e45e05dad58042bd`. Both upstream `LICENSE` and `LICENSE.md` were published for that source. Their scope is not unambiguous, so both applicable notices and the extension-file attribution are retained here without choosing between them. See [CREDITS.md](CREDITS.md) for the narrow source map.
### Upstream `opendbc` `LICENSE` notice (MIT)
Copyright (c) 2020, Comma.ai, Inc.
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and
associated documentation files (the "Software"), to deal in the Software without restriction,
including without limitation the rights to use, copy, modify, merge, publish, distribute,
sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or
substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT
NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM,
DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT
OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
### Upstream extension file notice
> Copyright (c) 2021-, Haibin Wen, sunnypilot, and a number of other contributors.
>
> This file is part of sunnypilot and is licensed under the MIT License.
> See the LICENSE.md file in the root directory for more details.
### Upstream `LICENSE.md` notice (published as “Custom MIT License”)
The following notice is reproduced verbatim from the reference branch:
> # Custom MIT License
>
> Copyright (c) 2024, Haibin Wen, SUNNYPILOT LLC
>
> Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to view and modify the Software, subject to the following conditions:
>
> 1. **Permission Required**: Permission Required for Commercial, For-Profit, or Closed Source Use: Use of the Software, in whole or in part, for any commercial purposes, for-profit projects, or in closed source projects requires explicit written permission from the original author(s).
>
> 2. **Redistribution**: Any redistribution of the Software, modified or unmodified, must retain this license notice and the following acknowledgment:
> "This software is licensed under a custom license requiring permission for use."
>
> 3. **Visibility**: Any project that uses the Software must visibly mention the following acknowledgment:
> "This project uses software from Haibin Wen and SUNNYPILOT LLC and is licensed under a custom license requiring permission for use."
>
> 4. **No Warranty**: THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
>
> Contact sunnypilot Support <support@sunnypilot.ai> for permission requests.
>
> ---
>
> Haibin Wen, SUNNYPILOT LLC
Required upstream acknowledgments:
> This software is licensed under a custom license requiring permission for use.
>
> This project uses software from Haibin Wen and SUNNYPILOT LLC and is licensed under a custom license requiring permission for use.
For commercial, for-profit, or closed-source use, consult the upstream notice and obtain any
permission it requires. The upstream repository's simultaneous publication of two differently
scoped notices should be clarified with the relevant copyright holders before relying on one to
the exclusion of the other.
## BluePilot Ford support
The Ford CAN FD camera-radar filtering includes work adapted from BluePilot's `bp-7.0` source, through StarPilot `2efe6017bfe14cee4807833e05f9a7f6ba255cdb`. See [CREDITS.md](CREDITS.md#ford-support-adapted-from-bluepilot) for contributors and source revisions. The upstream extension-file and `LICENSE.md` notices reproduced above also apply to that source; they are retained without resolving the difference between their published license descriptions.
## Continental ARS510 radar
The optional RAV4 radar decoders and `toyota_tss2_ars510.dbc` are adapted from
<https://github.com/ds-sebastian/ars510-radar> at
`fb5b29c2b5999f34afeb6f7ac494a73949d66c67`.
MIT License
Copyright (c) 2026 ARS510 radar contributors
Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all
copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.
## PulseAudio server for Spotify
`openpilot/starpilot/soloist/pulseaudio/` contains unmodified AArch64 binaries from Ubuntu 24.04 packages: PulseAudio 16.1 (LGPL-2.1-or-later), libltdl (LGPL-2.1-or-later), libtdb (LGPL-3.0-or-later), libsoxr (LGPL-2.1-or-later), liborc and speexdsp (BSD-style). The complete copyright and license notices are in that directory's `licenses/`, package hashes in `PACKAGES.sha512`, and the corresponding source is available from Ubuntu's archive for the exact versions listed in its README. The libraries are dynamically linked and may be replaced with `update.sh`.
## Galaxy Android companion tunnel
`openpilot/starpilot/galaxy/bin/hev-socks5-tunnel_linux_arm64` is the upstream
static Linux ARM64 release, under the MIT license. Its notice is retained in
`openpilot/starpilot/galaxy/bin/LICENSE.hev-socks5-tunnel`; the adjacent SHA-256
file pins the downloaded artifact. The Android companion uses the Kotlin
standard library (Apache-2.0) and the Gradle wrapper (Apache-2.0); no third-party
UI framework is bundled.
## Galaxy Bluetooth companion source snapshot
The Galaxy Bluetooth phone bridge and iOS app from https://github.com/StarPilotHQ/StarPilot
(natehuby76, PRs #224/#225) are included under `openpilot/starpilot/galaxy_bluetooth/`
and `openpilot/starpilot/companion/ios/`, with their license files (`UPSTREAM-LICENSE`, `openpilot/starpilot/companion/ios/UPSTREAM-GALAXY-LICENSE`).
The bridge runs on openpilot's Python with PyCryptodome 3.23.0 (BSD/public-domain
components, from opendbc). Its other dependencies are vendored unmodified, as
pure-Python sources, in `openpilot/starpilot/galaxy_bluetooth/vendor/`:
dbus-next 0.2.3 (MIT), ifaddr 0.2.0 (MIT) and Zeroconf 0.151.3
(LGPL-2.1-or-later). Their license texts are in `vendor/licenses/`, and
`vendor/VENDOR.json` records each PyPI source file and its SHA-256. Zeroconf is
kept as separate, unmodified source files that can be replaced with another
compatible version, as the LGPL requires.
## Subaru angle support from jacobw
The Subaru angle implementation includes work from Jacob Waller's `jul-angle-based` openpilot/opendbc branches and their upstream contributors. See [CREDITS.md](CREDITS.md#subaru-angle-steering-from-jacob-wallers-fork) and `opendbc_repo/opendbc/car/subaru/ORIGIN_HISTORY.tsv` for pinned sources, original authors and titles. The source repositories publish the comma MIT license; the complete applicable notice remains in the project `LICENSE` and `opendbc_repo/LICENSE`. StarPilot's adaptations preserve additional native and caller safety guards.
## StarPilot Ford demand tuning
The historical tuning formerly in `openpilot/starpilot/car/ford/starpilot_tuning.py` included the BluePilot bp-7.0-derived work identified in [CREDITS.md](CREDITS.md#starpilot-ford-demand-tuning). The published upstream license notices retained above and in the pinned BluePilot library remain applicable to the adapted source. StarPilot's separate protocol adapter does not modify those notices or resolve their scope.
The angle-primary adaptation in `openpilot/starpilot/car/ford/starpilot_controller.py` is derived from BluePilot v0.2.1. It retains the complete MIT copyright and permission notice in `opendbc_repo/opendbc/bluepilot_lateral/LICENSE`, including Comma.ai, BluePilot and StarPilot contributors. See the contributor attribution in `CREDITS.md`.
cereal is the messaging system for openpilot. It uses [msgq](https://github.com/commaai/msgq) as a pub/sub backend, and [Cap'n proto](https://capnproto.org/capnp-tool.html) for serialization of the structs.
## Messaging Spec
You'll find the message types in [log.capnp](log.capnp). It uses [Cap'n proto](https://capnproto.org/capnp-tool.html) and defines one struct called `Event`.
All `Events` have a `logMonoTime` and a `valid`. Then a big union defines the packet type.
### Best Practices
- **All fields must describe quantities in SI units**, unless otherwise specified in the field name.
- In the context of the message they are in, field names should be completely unambiguous.
- All values should be easy to plot and be human-readable with minimal parsing.
### Maintaining backwards-compatibility
When making changes to the messaging spec you want to maintain backwards-compatibility, such that old logs can
be parsed with a new version of cereal. Adding structs and adding members to structs is generally safe, most other
things are not. Read more details [here](https://capnproto.org/language.html).
### Custom forks
Forks of [openpilot](https://github.com/commaai/openpilot) might want to add things to the messaging
spec, however this could conflict with future changes made in mainline cereal/openpilot. Rebasing against mainline openpilot
then means breaking backwards-compatibility with all old logs of your fork. So we added reserved events in
[custom.capnp](custom.capnp) that we will leave empty in mainline cereal/openpilot. **If you only modify those, you can ensure your
fork will remain backwards-compatible with all versions of mainline openpilot and your fork.**
<!--- AUTOGENERATED FROM selfdrive/car/CARS_template.md, DO NOT EDIT. --->
<!--- AUTOGENERATED FROM openpilot/selfdrive/car/CARS_template.md, DO NOT EDIT. --->
# Supported Cars
A supported vehicle is one that just works when you install a comma device. All supported cars provide a better experience than any stock system. Supported vehicles reference the US market unless otherwise specified.
@@ -135,17 +137,17 @@ A supported vehicle is one that just works when you install a comma device. All
|Hyundai|Kona 2020|Smart Cruise Control (SCC)|openpilot available[<sup>1</sup>](#footnotes)|6 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 Hyundai B connector<br>- 1 OBD-C cable (2 ft)<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Hyundai Kona 2020">Buy Here</a></sub></details>|||
|Hyundai|Kona Electric 2018-21|Smart Cruise Control (SCC)|openpilot available[<sup>1</sup>](#footnotes)|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 Hyundai G connector<br>- 1 OBD-C cable (2 ft)<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Hyundai Kona Electric 2018-21">Buy Here</a></sub></details>|||
|Hyundai|Kona Electric 2022-23|Smart Cruise Control (SCC)|openpilot available[<sup>1</sup>](#footnotes)|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 Hyundai O connector<br>- 1 OBD-C cable (2 ft)<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Hyundai Kona Electric 2022-23">Buy Here</a></sub></details>|||
|Hyundai|Kona Electric (with HDA II, Korea only) 2023|Smart Cruise Control (SCC)|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 Hyundai R connector<br>- 1 OBD-C cable (2 ft)<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Hyundai Kona Electric (with HDA II, Korea only) 2023">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=U2fOCmcQ8hw" target="_blank"><img height="18px" src="assets/icon-youtube.svg"></img></a>||
|Hyundai|Kona Electric (with HDA II, Korea only) 2023|Smart Cruise Control (SCC)|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 Hyundai R connector<br>- 1 OBD-C cable (2 ft)<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Hyundai Kona Electric (with HDA II, Korea only) 2023">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=U2fOCmcQ8hw" target="_blank"><img height="18px" src="assets/icon-youtube.svg" /></a>||
|Hyundai|Kona Hybrid 2020|Smart Cruise Control (SCC)|openpilot available[<sup>1</sup>](#footnotes)|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 Hyundai I connector<br>- 1 OBD-C cable (2 ft)<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Hyundai Kona Hybrid 2020">Buy Here</a></sub></details>|||
|Toyota|Mirai 2021|All|openpilot|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Mirai 2021">Buy Here</a></sub></details>|||
|Toyota|Prius 2016|Toyota Safety Sense P|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius 2016">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=8zopPJI8XQ0" target="_blank"><img height="18px" src="assets/icon-youtube.svg"></img></a>||
|Toyota|Prius 2017-20|All|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius 2017-20">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=8zopPJI8XQ0" target="_blank"><img height="18px" src="assets/icon-youtube.svg"></img></a>||
|Toyota|Prius 2021-22|All|openpilot|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius 2021-22">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=J58TvCpUd4U" target="_blank"><img height="18px" src="assets/icon-youtube.svg"></img></a>||
|Toyota|Prius Prime 2017-20|All|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius Prime 2017-20">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=8zopPJI8XQ0" target="_blank"><img height="18px" src="assets/icon-youtube.svg"></img></a>||
|Toyota|Prius Prime 2021-22|All|openpilot|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius Prime 2021-22">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=J58TvCpUd4U" target="_blank"><img height="18px" src="assets/icon-youtube.svg"></img></a>||
|Toyota|Prius 2016|Toyota Safety Sense P|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius 2016">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=8zopPJI8XQ0" target="_blank"><img height="18px" src="assets/icon-youtube.svg" /></a>||
|Toyota|Prius 2017-20|All|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius 2017-20">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=8zopPJI8XQ0" target="_blank"><img height="18px" src="assets/icon-youtube.svg" /></a>||
|Toyota|Prius 2021-22|All|openpilot|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius 2021-22">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=J58TvCpUd4U" target="_blank"><img height="18px" src="assets/icon-youtube.svg" /></a>||
|Toyota|Prius Prime 2017-20|All|Stock|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius Prime 2017-20">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=8zopPJI8XQ0" target="_blank"><img height="18px" src="assets/icon-youtube.svg" /></a>||
|Toyota|Prius Prime 2021-22|All|openpilot|0 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius Prime 2021-22">Buy Here</a></sub></details>|<a href="https://www.youtube.com/watch?v=J58TvCpUd4U" target="_blank"><img height="18px" src="assets/icon-youtube.svg" /></a>||
|Toyota|Prius v 2017|Toyota Safety Sense P|Stock|19 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota Prius v 2017">Buy Here</a></sub></details>|||
|Toyota|RAV4 2016|Toyota Safety Sense P|Stock|19 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota RAV4 2016">Buy Here</a></sub></details>|||
|Toyota|RAV4 2017-18|All|Stock|19 mph|0 mph|[](##)|[](##)|<details><summary>Parts</summary><sub>- 1 OBD-C cable (2 ft)<br>- 1 Toyota A connector<br>- 1 comma four<br>- 1 comma power v3<br>- 1 harness box<br>- 1 mount<br><a href="https://comma.ai/shop/comma-3x?harness=Toyota RAV4 2017-18">Buy Here</a></sub></details>|||
<sup>1</sup>openpilot Longitudinal Control (Alpha) is available behind a toggle; the toggle is only available in non-release branches such as `nightly-dev`. <br />
<sup>2</sup>Refers only to the Focus Mk4 (C519) available in Europe/China/Taiwan/Australasia, not the Focus Mk3 (C346) in North and South America/Southeast Asia. <br />
<sup>3</sup>See more setup details for <a href="https://github.com/commaai/openpilot/wiki/gm" target="_blank">GM</a>. <br />
<sup>4</sup>2019 Honda Civic 1.6L Diesel Sedan does not have ALC below 12mph. <br />
<sup>5</sup>See more setup details for <a href="https://github.com/commaai/openpilot/wiki/nissan" target="_blank">Nissan</a>. <br />
<sup>6</sup>In the non-US market, openpilot requires the car to come equipped with EyeSight with Lane Keep Assistance. <br />
<sup>7</sup>Enabling longitudinal control (alpha) will disable all EyeSight functionality, including AEB, LDW, and RAB. <br />
<sup>8</sup>Some 2023 model years have HW4. To check which hardware type your vehicle has, look for <b>Autopilot computer</b> under <b>Software -> Additional Vehicle Information</b> on your vehicle's touchscreen. See <a href="https://www.notateslaapp.com/news/2173/how-to-check-if-your-tesla-has-hardware-4-ai4-or-hardware-3">this page</a> for more information. <br />
<sup>9</sup>See more setup details for <a href="https://github.com/commaai/openpilot/wiki/tesla" target="_blank">Tesla</a>. <br />
<sup>10</sup>openpilot operates above 28mph for Camry 4CYL L, 4CYL LE and 4CYL SE which don't have Full-Speed Range Dynamic Radar Cruise Control. <br />
<sup>11</sup>The J533 harness plugs in at the CAN gateway under the dashboard, just above the steering column. More information can be found at <a href="https://docs.howtocomma.com/docs/j533-harness-install" target="_blank">this guide</a>. <br />
<sup>12</sup>Not including the China market Kamiq, which is based on the (currently) unsupported PQ34 platform. <br />
<sup>13</sup>Refers only to the MQB-based European B8 Passat, not the NMS Passat in the USA/China/Mideast markets. <br />
<sup>14</sup>Some Škoda vehicles are equipped with heated windshields, which are known to block GPS signal needed for some comma four functionality. <br />
<sup>15</sup>Only available for vehicles using a gateway (J533) harness. At this time, vehicles using a camera harness are limited to using stock ACC. <br />
<sup>16</sup>Model-years 2022 and beyond may have a combined CAN gateway and BCM, which is supported by openpilot in software, but doesn't yet have a harness available from the comma store. <br />
<sup>5</sup>Enabling longitudinal control (alpha) will disable all CMBS functionality, including AEB and FCW. <br />
<sup>6</sup>See more setup details for <a href="https://github.com/commaai/openpilot/wiki/nissan" target="_blank">Nissan</a>. <br />
<sup>7</sup>In the non-US market, openpilot requires the car to come equipped with EyeSight with Lane Keep Assistance. <br />
<sup>8</sup>Enabling longitudinal control (alpha) will disable all EyeSight functionality, including AEB, LDW, and RAB. <br />
<sup>9</sup>Model years 2023 and 2024 can have either hardware type, depending on build date and factory. To check which hardware type your vehicle has, look for <b>Autopilot computer</b> under <b>Software -> Additional Vehicle Information</b> on your vehicle's touchscreen. See <a href="https://www.notateslaapp.com/news/2173/how-to-check-if-your-tesla-has-hardware-4-ai4-or-hardware-3">this page</a> for more information. <br />
<sup>10</sup>See more setup details for <a href="https://github.com/commaai/openpilot/wiki/tesla" target="_blank">Tesla</a>. <br />
<sup>11</sup>openpilot operates above 28mph for Camry 4CYL L, 4CYL LE and 4CYL SE which don't have Full-Speed Range Dynamic Radar Cruise Control. <br />
<sup>12</sup>The J533 harness plugs in at the CAN gateway under the dashboard, just above the steering column. More information can be found at <a href="https://docs.howtocomma.com/docs/j533-harness-install" target="_blank">this guide</a>. <br />
<sup>13</sup>Not including the China market Kamiq, which is based on the (currently) unsupported PQ34 platform. <br />
<sup>14</sup>Refers only to the MQB-based European B8 Passat, not the NMS Passat in the USA/China/Mideast markets. <br />
<sup>15</sup>Some Škoda vehicles are equipped with heated windshields, which are known to block GPS signal needed for some comma four functionality. <br />
<sup>16</sup>Only available for vehicles using a gateway (J533) harness. At this time, vehicles using a camera harness are limited to using stock ACC. <br />
<sup>17</sup>Model-years 2022 and beyond may have a combined CAN gateway and BCM, which is supported by openpilot in software, but doesn't yet have a harness available from the comma store. <br />
## Community Maintained Cars
Although they're not upstream, the community has openpilot running on other makes and models. See the 'Community Supported Models' section of each make [on our wiki](https://wiki.comma.ai/).
Our software is open source so you can solve your own problems without needing help from others. And if you solve a problem and are so kind, you can upstream it for the rest of the world to use. Check out our [post about externalization](https://blog.comma.ai/a-2020-theme-externalization/).
Start with the [project overview](../README.md), [build guide](how-to/laptop-device-build.md), and [developer commands](../tools/STARPILOT_DEVELOPMENT.md).
Development is coordinated through [Discord](https://discord.comma.ai) and GitHub.
## Scope and behavior
### Getting Started
Keep each change focused on one feature or defect. Describe the behavior before and after the change, the affected vehicle configurations or interfaces, and how it was verified. Preserve intentional StarPilot behavior and familiar controls unless the change explicitly proposes a different experience.
* Set up your [development environment](/tools/)
* Join our [Discord](https://discord.comma.ai)
* Docs are at https://docs.comma.ai and https://blog.comma.ai
Prefer small integration points around upstream interfaces. Keep feature settings, state, and side effects with their owning feature. Shared code should express a real common contract, with regression coverage for the features that use it. Comments should explain an invariant, protocol requirement, or non-obvious decision.
## What contributions are we looking for?
## Interfaces and dependencies
**openpilot's priorities are [safety](SAFETY.md), stability, quality, and features, in that order.**
openpilot is part of comma's mission to *solve self-driving cars while delivering shippable intermediaries*, and all development is towards that goal.
Follow cereal's [custom-fork guidance](../openpilot/cereal/README.md#custom-forks). Preserve upstream message definitions and field meanings; put fork-specific messaging in the custom schema. Vehicle support outside an upstream platform's contract needs a distinct platform identity and corresponding interface and safety evidence.
### What gets merged?
Dependencies are tracked source folders. Use the [vendoring workflow](../tools/vendor/README.md), retain provenance and licenses, and review local changes when updating an upstream revision.
The probability of a pull request being merged is a function of its value to the project and the effort it will take us to get it merged.
If a PR offers *some* value but will take lots of time to get merged, it will be closed.
Simple, well-tested bug fixes are the easiest to merge, and new features are the hardest to get merged.
* simple car model port: https://github.com/commaai/openpilot/pull/30245
* car brand port: https://github.com/commaai/openpilot/pull/23331
Run checks appropriate to the change and describe the result in the pull request:
### What doesn't get merged?
- For a bug fix, demonstrate the failure and the corrected behavior.
- For driving behavior, compare against the intended source behavior or recorded evidence and cover engagement, override, reset, and relevant limits.
- For vehicle or safety changes, identify the exact configuration and run the affected interface and native safety checks.
- For UI changes, inspect the affected interface at its actual size and preserve the intended interactions.
- For performance changes, provide a reproducible comparison and distinguish a component benchmark from whole-system performance.
* **style changes**: code is art, and it's up to the author to make it beautiful
* **500+ line PRs**: clean it up, break it up into smaller PRs, or both
* **PRs without a clear goal**: every PR must have a singular and clear goal
* **UI design**: we do not have a good review process for this yet
* **New features**: We believe openpilot is mostly feature-complete, and the rest is a matter of refinement and fixing bugs. As a result of this, most feature PRs will be immediately closed, however the beauty of open source is that forks can and do offer features that upstream openpilot doesn't.
* **Negative expected value**: This is a class of PRs that makes an improvement, but the risk or validation costs more than the improvement. The risk can be mitigated by first getting a failing test merged.
Keep automated tests, bench observations, and driving results separate. State what remains unverified rather than broadening a narrow result into a fleet-wide claim.
### First contribution
## Commit history
[Projects / openpilot bounties](https://github.com/orgs/commaai/projects/26/views/1?pane=info) is the best place toget started and goes in-depth on what's expected when working on a bounty.
There are a lot of bounties that don't require a comma four or a car.
Use concise feature or behavior names. Keep related changes together, including their tests and documentation. Preserve ordinary follow-up commits so contributors can follow changes.
## Pull Requests
Pull requests should be against the master branch.
A good pull request has all of the following:
* a clearly stated purpose
* every line changed directly contributes to the stated purpose
* verification, i.e. how did you test your PR?
* justification
* if you've optimized something, post benchmarks to prove it's better
* if you've improved your car's tuning, post before and after plots
* passes the CI tests
## Contributing without Code
* Report bugs in GitHub issues.
* Report driving issues in the `#driving-feedback` Discord channel.
* Consider opting into driver camera uploads to improve the driver monitoring model.
* Connect your device to Wi-Fi regularly, so that we can pull data for training better driving models.
* Run the `nightly` branch and report issues. This branch is like `master` but it's built just like a release.
* Annotate images in the [comma10k dataset](https://github.com/commaai/comma10k).
## Contributing Training Data
### A guide for forks
In order for your fork's data to be eligible for the training set:
* **Your cereal messaging structs must be [compatible](../cereal#custom-forks)**
* **The definitions of all the stock messaging structs must not change**: Do not change how any of the fields are set, including everything from `selfdriveState.enabled` to `carState.steeringAngleDeg`. Instead, create your own structs and set them however you'd like.
* **Do not include cars that are not supported in upstream platforms**: Instead, create new opendbc platforms for cars that you'd like to support outside of upstream, even if it's just a trim-level difference.
The original [openpilot contribution guide](https://github.com/commaai/openpilot/blob/521db4c825d37eb5f29acf955daa88da003e4433/docs/CONTRIBUTING.md) documents upstream's process. Changes intended for comma's openpilot should follow that project's contribution requirements.
@@ -5,6 +5,8 @@ openpilot ALC and openpilot LDW do not automatically drive the vehicle or reduce
While changing lanes, openpilot is not capable of looking next to you or checking your blind spot. Only nudge the wheel to initiate a lane change after you have confirmed it's safe to do so.
Unexpected steering may occur. openpilot ALC may not steer to avoid stopped or slow-moving vehicles, cross traffic, construction barriers or cones, motorcycles, children, pedestrians, animals, or other objects on the road.
Many factors can impact the performance of openpilot ALC and openpilot LDW, causing them to be unable to function as intended. These include, but are not limited to:
* Poor visibility (heavy rain, snow, fog, etc.) or weather conditions that may interfere with sensor operation.
@@ -13,6 +15,14 @@ Many factors can impact the performance of openpilot ALC and openpilot LDW, caus
* The device is mounted incorrectly.
* When in sharp curves, like on-off ramps, intersections etc...; openpilot is designed to be limited in the amount of steering torque it can produce.
* In the presence of restricted lanes or construction zones.
* When the lane markings are unclear, worn, missing, or they cross each other.
* Tar marks, shadows, road cracks, temporary or construction lane markings, or other road imperfections.
* When lanes merge or split.
* When entering a highway on-ramp or exiting an off-ramp.
* On roads with intersections or roundabouts.
* On slippery roads.
* When towing a trailer.
* When entering or exiting a tunnel.
* When driving on highly banked roads or in presence of strong cross-wind.
* Extremely hot or cold temperatures.
* Bright light (due to oncoming headlights, direct sunlight, etc.).
@@ -24,6 +34,8 @@ The list above does not represent an exhaustive list of situations that may inte
openpilot ACC and openpilot FCW are not systems that allow careless or inattentive driving. It is still necessary for the driver to pay close attention to the vehicle’s surroundings and to be ready to re-take control of the gas and the brake at all times.
openpilot ACC may not stop your vehicle in time to avoid a collision with a vehicle that quickly stops in front of you. openpilot ACC may occasionally cause your vehicle to brake when not required or when you are not expecting it.
Many factors can impact the performance of openpilot ACC and openpilot FCW, causing them to be unable to function as intended. These include, but are not limited to:
* Poor visibility (heavy rain, snow, fog, etc.) or weather conditions that may interfere with sensor operation.
@@ -35,9 +47,20 @@ Many factors can impact the performance of openpilot ACC and openpilot FCW, caus
* In presence of traffic signs or stop lights, which are not detected by openpilot at this time.
* When the posted speed limit is below the user selected set speed. openpilot does not detect speed limits at this time.
* In presence of vehicles in the same lane that are not moving.
* When a vehicle you are following moves out of your driving path and a stationary or slow-moving vehicle or object is in front of you.
* Narrow vehicles, such as bicycles or motorcycles.
* Vehicles or objects only partially in the driving lane.
* On curves, openpilot ACC may not detect a vehicle in the same lane, and may accelerate to the set speed.
* Cross traffic.
* Oncoming vehicles.
* Animals or objects in the road.
* Rare objects such as trailers, ramps, cargo, open doors, etc. protruding from vehicles.
* When abrupt braking maneuvers are required. openpilot is designed to be limited in the amount of deceleration and acceleration that it can produce.
* When surrounding vehicles perform close cut-ins from neighbor lanes.
* Driving on hills, narrow, or winding roads.
* On slippery roads.
* When towing a trailer.
* When entering or exiting a tunnel.
* Extremely hot or cold temperatures.
* Bright light (due to oncoming headlights, direct sunlight, etc.).
* Interference from other equipment that generates radar waves.
@@ -46,13 +69,15 @@ The list above does not represent an exhaustive list of situations that may inte
## Limitations of openpilot DM
openpilot DM should not be considered an exact measurement of the alertness of the driver.
openpilot DM displays a series of escalating warnings. Disregard of these warnings may result in openpilot being unavailable for a period of time.
openpilot DM should not be considered an exact measurement of the alertness of the driver. openpilot DM cannot always detect if the driver is tired or inattentive.
Many factors can impact the performance of openpilot DM, causing it to be unable to function as intended. These include, but are not limited to:
* Low light conditions, such as driving at night or in dark tunnels.
* Bright light (due to oncoming headlights, direct sunlight, etc.).
* The driver's face is partially or completely outside field of view of the driver facing camera.
* The driver facing camera is obstructed, covered, or damaged.
* The driver's face is partially or completely outside field of view of the cabin camera.
* The cabin camera is obstructed, covered, or damaged.
The list above does not represent an exhaustive list of situations that may interfere with proper operation of openpilot components. A driver should not rely on openpilot DM to assess their level of attention.
@@ -4,8 +4,14 @@ openpilot is an Adaptive Cruise Control (ACC) and Automated Lane Centering (ALC)
Like other ACC and ALC systems, openpilot is a failsafe passive system and it requires the
driver to be alert and to pay attention at all times.
In order to enforce driver alertness, openpilot includes a driver monitoring feature
that alerts the driver when distracted.
openpilot is not an autonomous driving system. It is not a collision avoidance system or a
substitute for safe and attentive driving. The driver is always responsible for the safe
operation of the vehicle and must be ready to take control at any time. The driver can override
openpilot at any time by braking, accelerating, or steering the vehicle. openpilot may
disengage without warning. Read the [limitations](https://github.com/commaai/openpilot/blob/master/docs/LIMITATIONS.md) carefully before using openpilot.
To assist the driver in maintaining alertness, openpilot includes a driver monitoring feature
that alerts when it detects driver distraction.
However, even with an attentive driver, we must make further efforts for the system to be
safe. We repeat, **driver alertness is necessary, but not sufficient, for openpilot to be
@@ -35,8 +41,8 @@ For additional safety implementation details, refer to [panda safety model](http
### Forks of openpilot
* Do not disable or nerf [driver monitoring](https://github.com/commaai/openpilot/tree/master/selfdrive/monitoring)
* Do not disable or nerf [excessive actuation checks](https://github.com/commaai/openpilot/tree/master/selfdrive/selfdrived/helpers.py)
* Do not disable or nerf [driver monitoring](https://github.com/commaai/openpilot/tree/master/openpilot/selfdrive/monitoring)
* Do not disable or nerf [excessive actuation checks](https://github.com/commaai/openpilot/tree/master/openpilot/selfdrive/selfdrived/helpers.py)
* If your fork modifies any of the code in `opendbc/safety/`:
* your fork cannot use the openpilot trademark
* your fork must preserve the full [safety test suite](https://github.com/commaai/opendbc/tree/master/opendbc/safety/tests) and all tests must pass, including any new coverage required by the fork's changes
A model port is a port of openpilot to a new car model within an already supported brand. Model ports are easier than brand ports because the car's existing APIs are already known.
Here's an example of one: https://github.com/commaai/openpilot/pull/30672/.
A car port enables openpilot support on a particular car. Each car model openpilot supports needs to be individually ported. The complexity of a car port varies depending on many factors including:
* existing openpilot support for similar cars
* architecture and APIs available in the car
# Structure of a car port
Virtually all car-specific code is contained in two other repositories: [opendbc](https://github.com/commaai/opendbc) and [panda](https://github.com/commaai/panda).
## opendbc
Each car brand is supported by a standard interface structure in `opendbc/car/[brand]`:
* `interface.py`: Interface for the car, defines the CarInterface class
* `carstate.py`: Reads CAN messages from the car and builds openpilot CarState messages
* `carcontroller.py`: Control logic for executing openpilot CarControl actions on the car
* `[brand]can.py`: Composes CAN messages for carcontroller to send
* `values.py`: Limits for actuation, general constants for cars, and supported car documentation
* `radar_interface.py`: Interface for parsing radar points from the car, if applicable
* `opendbc_repo/opendbc/safety/tests/test_[brand].py`: Brand-specific safety CI tests
## openpilot
For historical reasons, openpilot still contains a small amount of car-specific logic. This will eventually be migrated to opendbc or otherwise removed.
* **onroad**: openpilot's system state while ignition is on
* **offroad**: openpilot's system state while ignition is off
* **route**: a route is a recording of an onroad session
* **segment**: routes are split into one minute chunks called segments.
* **comma connect**: the web viewer for all your routes; check it out at [connect.comma.ai](https://connect.comma.ai).
* **panda**: this is the secondary processor on the device that implements the functional safety and directly talks to the car over CAN. See the [panda repo](https://github.com/commaai/panda).
* **comma four**: the latest hardware by comma.ai for running openpilot. more info at [comma.ai/shop/comma-four](https://www.comma.ai/shop/comma-four).
openpilot records routes in one minute chunks called segments. A route starts on the rising edge of ignition and ends on the falling edge.
Check out our [Python library](https://github.com/commaai/openpilot/blob/master/tools/lib/logreader.py) for reading openpilot logs. Also checkout our [tools](https://github.com/commaai/openpilot/tree/master/tools) to replay and view your data. These are the same tools we use to debug and develop openpilot.
Check out our [Python library](https://github.com/commaai/openpilot/blob/master/openpilot/tools/lib/logreader.py) for reading openpilot logs. Also checkout our [tools](https://github.com/commaai/openpilot/tree/master/openpilot/tools) to replay and view your data. These are the same tools we use to debug and develop openpilot.
For each segment, openpilot records the following log types:
## rlog.bz2
## rlog.zst
rlogs contain all the messages passed amongst openpilot's processes. See [cereal/services.py](https://github.com/commaai/cereal/blob/master/services.py) for a list of all the logged services. They're a bzip2 archive of the serialized capnproto messages.
rlogs contain all the messages passed amongst openpilot's processes. See [openpilot/cereal/services.py](https://github.com/commaai/openpilot/blob/master/openpilot/cereal/services.py) for a list of all the logged services. They're a zstd archive of the serialized [Cap’n Proto](https://capnproto.org/) messages.
## {f,e,d}camera.hevc
## camera video files
Each camera stream is H.265 encoded and written to its respective file.
* `fcamera.hevc` is the road camera
* `fcamera.hevc` is the narrow road camera (the main forward camera)
* `ecamera.hevc` is the wide road camera
* `dcamera.hevc` is the driver camera
* `dcamera.hevc` is the cabin camera
## qlog.bz2 & qcamera.ts
qlogs are a decimated subset of the rlogs. Check out [cereal/services.py](https://github.com/commaai/cereal/blob/master/services.py) for the decimation.
## qlog.zst & qcamera.ts
qlogs are a decimated subset of the rlogs. Check out [openpilot/cereal/services.py](https://github.com/commaai/openpilot/blob/master/openpilot/cereal/services.py) for the decimation.
qcameras are H.264 encoded, lower res versions of the fcamera.hevc. The video shown in [comma connect](https://connect.comma.ai/) is from the qcameras.
qlogs and qcameras are designed to be small enough to upload instantly on slow internet and store forever, yet useful enough for most analysis and debugging.
qlogs and qcameras are designed to be small enough to upload instantly on slow internet, yet useful enough for most analysis and debugging.
pointer-events:auto;/* Allow interaction when visible */
}
.tooltip-content.tooltip-glossary-link{
display:inline-block;
margin-top:8px;
font-size:12px;
color:#007bff;
text-decoration:none;
}
.tooltip-content.tooltip-glossary-link:hover{
color:#0056b3;
text-decoration:underline;
}
Some files were not shown because too many files have changed in this diff
Show More
Reference in New Issue
Block a user
Blocking a user prevents them from interacting with repositories, such as opening or commenting on pull requests or issues. Learn more about blocking a user.
firestar5683
