Skip to content

Session log

  • Repo created. Downloaded every document and firmware image from parrot.com/en/support/documentation/bebop-range into docs/parrot/ and firmware/; .txt extractions alongside each PDF.

  • Bebop 2, 2 FPV and 2 Power firmware downloads are byte-identical (SHA-256 6fe03e1e…); kept one as bebop-2-family_v4.7.1.plf.

  • Evaluated bebop_autonomy (ROS1 only, last commit 2018-07-23) and pyparrot (last commit 2020-01-05). Chose pyparrot; cloned to ../pyparrot.

  • Found and fixed 7 bugs in pyparrot on branch py313-fixes (see pyparrot.md).

  • Built bebop-2 CLI: scan, telemetry, sim, later describe.

  • Wrote a fake Bebop 2 (sim.py) that encodes events from the protocol XML. End-to-end tests: 3 of 5 fail on upstream pyparrot bf4775e, 5 of 5 pass on the fork.

  • Reviewed GiannisTDM/parrot-lab and GiannisTDM/parrot-bebop2-rf-lab (see references.md).

  • Unpacked firmware 4.7.1 in a --network none container. Found the power-button map and the stock ArduPilot hook (see firmware-4.7.1.md).

  • Downloaded ArduCopter stable for bebop (4.7.1, 2026-08-31) to firmware/ardupilot/. Not installed on the drone.

  • A peer Claude session (bebop-engineer) wrote FLIGHT_CONTROLLER_MAP.md. Corrected two hardware claims against the rootfs: /dev/ttyPA1 is the GPS, motors are on /dev/i2c-1; UDP 14550/14551 exist only under ArduCopter.

  • Drone powered up, props removed. First boot attempt: no access point seen in 4 scans over about a minute. After a power cycle (chirped, LED flashing quickly) it appeared as Bebop2-XXXXXX, open, 2.4 GHz, BSSID OUI A0:14:3D (Parrot SA).

  • Joined with a dedicated NetworkManager profile bebop2 (ipv4.never-default yes, autoconnect no). Got 192.168.42.36; ping 192.168.42.1 about 8 ms; default route stayed on br0.

  • First contact: 15 s read-only session, 1174 updates, 181 distinct keys, captures/first-contact.jsonl. Results in drone state.

  • Added bebop-2 describe to annotate captures from the XML. Its first run leaked CPU id and P7ID into the report; redaction widened before anything was committed.

  • Telnet login on the drone is exec /bin/sh -l (bin/login.sh): no authentication. Recorded in firmware-4.7.1.md.

  • Peer session rewrote the shared memory file bebop2-firmware-hooks, dropping the RF-profile warning and restating two corrected claims. Split its Ghidra findings into their own memory, marked verified/unverified.

  • Peer bebop-engineer retracted weak claims (banners on its docs) and reported config-file facts. Verified: estimator 198.36 Hz, motor range 3000 to 12200 RPM, no Colibry library, libCondor is camera prediction, BLDC.cyacd checksums (two’s complement, not plain sum as it said). Its own notes live in repo-root notes/, HARDWARE_REVERSE_ENGINEERING.md, SKYCONTROLLER_NOTES.md; it also extracted the SkyController 2.0.0 and Bebop 1 4.0.6 PLFs under firmware/extracted/.

  • Ryan chose one notebook. Moved the peer’s root-level docs and notes/ into docs/notes/reverse-engineering/; rewrote 7 link lines in its files (paths only) and pointed our notes at the new folder. Peer told the new location before the move.

  • First git commit d137eef on main: 74 files, Parrot PDFs in Git LFS (*.pdf), firmware images and captures untracked. No remote yet.

  • Checked the peer’s BLDC drafts against BLDC_Test_Bench’s help text. psoc-firmware-DRAFT-unverified.md had -C as stop (it only clears the error flag; -H stops) and -s 8000 as set speed (it builds a step profile). Flag table added to firmware-4.7.1.md. Decompile still not done.

  • Peer corrected its PSoC draft against the usage text (no “-C stops” left; SAFETY line added). Peer reports Ryan approved the BLDC_Test_Bench decompile in its session; running as a static-only background agent. mcghidra failed to connect in bebop-main (CONNECT_TIMEOUT), so Ghidra work stays with the peer.

  • mcghidra reconnected. Read-only cross-check of the peer’s dragon-prog analysis: function count and two addresses confirmed; setAttitude is only a thunk in dragon-prog. Recorded in firmware-4.7.1.md.

  • Installed Parrot’s phone app over mcadb. com.parrot.freeflight3 (FreeFlight Pro) is gone from the Play Store; the only Parrot hit was FreeFlight 6 USA 6.10.6 (com.parrot.freeflight6usa). Not launched or paired yet. Pulled its six split APKs (420 MB) to apk/freeflight6usa/, gitignored.

  • Its split_firmwares.apk holds ANAFI 1.10.7 and controller images (mpp3 1.8.1, mpp4 7.7.2, mppua 3.2.3), none for the Bebop 2. The mpp3 rootfs names itself “Skycontroller 3”, the ANAFI-era controller, so it is not the nap controller image we already have. Details in freeflight6-usa-app.md. Skycontroller generation in Ryan’s FPV pack is still open.

  • Camera feed working. Probed first: RTSP refused, so read the firmware and found ARStream2/RTP instead. Verified packets arrive on 55004 before building bebop-2 video (view / --record / --snapshot). First frames read mean luma 4.95 of 255, flat across 10 s: the lens cap was on. With it off, a normal dim-room image. Wire details in protocol.md.

  • Ryan confirms the FPV-pack controller is a Skycontroller 2 and that “v0.9.1” is printed on the back. A label is not a firmware version, so likely a hardware or production revision. (Corrects an entry earlier today that treated it as firmware.) Firmware version of the unit still unread.

  • Launched FreeFlight 6 USA. Terms accepted by Ryan. Setup wizard assumes ANAFI (example SSID ANAFI-…, QR from the carrying case); no Bebop mention. Drone not tried yet. See freeflight6-usa-app.md, “First launch”.

  • Peer finished the PSoC (BLDC.cyacd) static analysis: 8051 core, dispatcher table serving 10 command bytes, XOR8 not CRC-16. Its 6.7.R.4 version read from the image matches what the live drone reported over ARSDK at first contact, from a completely separate path, so no probe was needed. Flag table annotated with the unserved-command caveat.

  • Phone test against the live Bebop 2 (agreed with bebop-main): FreeFlight 6 USA did not detect it. Phone was on the AP and pinged it (after mobile data off), but the app’s drone/controller icons stayed grey; wizard is ANAFI-only. No handshake attempted. Treat FF6 USA as not a Bebop 2 app; FreeFlight Pro is the right one and is delisted. Phone restored. Details in freeflight6-usa-app.md, “Test against the live Bebop 2”.

  • Ryan spotted macgyverd in the firmware. It is the on-board crash reporter (kernel oops, crash dumps, Java/ANR traces, 20 reports kept). Auto-send over HTTPS is OFF in the shipped config, though the binary carries an HTTPS client and an embedded certificate for crashreport.parrot.biz, which had no public DNS record when checked. This unit reported 0 crash logs. Details, unverified items and a do-not-publish note in macgyverd-crash-reporter.md.
  • Live smoke test of mcbebop against the aircraft found three protocol bugs the simulator could not, because the simulator was written from the same assumptions as the code and so agreed with it. (1) Inbound frames must be acknowledged on data type alone: this drone sends DATA_WITH_ACK on buffer 126 and plain DATA on 127, the reverse of the buffer naming, and requiring both to agree meant nothing was acked, the drone retransmitted rather than continuing, and only 35 of 192 telemetry keys ever arrived. (2) Eight argument-less events decoded to nothing and were lost, including the two that mark the end of a state dump. (3) request_full_state waited for a quiet link, which never happens at 5 Hz. All three in protocol.md.
  • Also fixed in mcbebop: preflight_check reported ready on an aircraft it knew almost nothing about, because an unreported blocking check counted as a pass. Unknown is now blocking. Caught on the first live run, when it said ready with no sensor, motor or alert data.
  • mcbebop registered in .mcp.json and driven from Claude Code directly: connect, preflight, FTP listing and a camera snapshot all work through the MCP tools. One wiring bug found that way (the file tools resolved an area name to an object and then passed the object to a function that resolves names), fixed.
  • Read the drone’s filesystem over FTP and answered the stored-flight-plan question: there is no flight plan. The flightplans directory is empty over both port 21 and port 61; the filepath in telemetry is where one would go. An earlier note read Mavlink_File = 1 as “a file is present”, which was wrong. Layout and the open ckcm.bin question in drone state.
  • Drove the drone’s camera from Claude Code through the MCP tools: browsed the command, sent Camera.OrientationV2(tilt=15), confirmed from telemetry that the aircraft moved it, took a frame showing the new view, then restored the default −13°. Two findings in protocol.md: unacknowledged commands never reply so acked: false is normal for them, and their confirmation is the drone’s own echo, which needs about a second before it can be read.
  • Fetched ckcm.bin over FTP. It is not camera calibration: it is the live ulogcat system log, and it answers the board question without a shell. The drone says My hardware is milosboard, Product variant : milosboard, and that it read /etc/colibry/milosboard/root.cfg. So the milosboard values (motorsRpmMax 12200, mass 0.500 kg) are this aircraft’s real ones. It also carries two Mykonos3 board init lines, which is odd on a Milos machine.
  • Telnet session (four presses; port 23 and ADB 9050 both opened). First real use of shell_read, which worked. Answered the per-unit radio NVM question: there is no per-unit file. The NVM on the aircraft is identical to the image’s, an unspecialised Broadcom template still carrying a literal boardnum=${serno} and a placeholder MAC, so per-unit calibration must live in the radio chip’s OTP. Also read this unit’s own hardware revisions from /sys/kernel/hsis/ (hwrev 14, pcbrev 3) and its full board pin map, which includes reset_psoc. bcmwl nvram_dump is unsupported here. getprop does not exist on this firmware. All in drone state.
  • Clarification from Ryan: the four presses were made while the drone was powered and an ARSDK session was open, and nothing was disturbed by it.
  • ADB (9050) opens alongside telnet and gives unrestricted root; read the whole /sys/kernel/hsis/ GPIO map through it (pin numbers for both cameras, gyro, magnetometer, GPS power, fan, LED, the power button itself, and reset_psoc at pin 199, a non-I2C path to the motor controller). Also noted in network.md that ADB bypasses any allow-list a client applies to telnet.
  • Peer session asked before touching the aircraft, since neither of us has tested two ARSDK controllers at once. Chose the no-risk option: read the keys from my open session and sent the values over. Recorded in drone state. Also told the peer that mcbebop’s own “one controller at a time” wording is an inherited claim from pyparrot’s error text, not something either of us has verified; worth settling against the simulator some time, not by accident.
  • Separated the two tilt quantities that had looked like a contradiction: pilot authority (MaxTiltChanged, 5-35°) versus the virtual gimbal crop (CameraSettingsChanged, −83 to +17°). videoTiltRange is a third thing.
  • Separated two questions that had been blurred: multiple Wi-Fi clients on the drone’s AP is demonstrated (phone and workstation associated together during the FreeFlight test, both pinging), while multiple ARSDK control sessions remains untested and the “one controller at a time” line is inherited from pyparrot’s error text rather than verified. In network.md.
  • Controller connection analysis (static): the nap image joins the drone’s Wi-Fi as a client (192.168.42.3), can run its own AP (192.168.44.1), NATs and forwards FTP to the drone, and talks to a phone over USB via an Android Open Accessory daemon. FF6’s USB accessory filter lists Skycontroller 3/ua/4/4 Black/5, not 2. See freeflight6-usa-app.md, “How the controller connects”.
  • Decompiled FF6 USA (jadx, Java 17). Its Drone.Model enum is ANAFI-only and RemoteControl.Model is Skycontroller 3/UA/4/4 Black/5/UA 2: no Bebop, no Skycontroller 1 or 2. Question closed by the code, not just the live test. App transports are IP over Wi-Fi and a USB-accessory mux. Skycontroller 2’s phone link (AP vs USB) still open. Details in freeflight6-usa-app.md.
  • Settled the two-controller question on the aircraft. The drone does not refuse a second controller: it accepts it and silently redirects telemetry to it, freezing the first session’s frame count while that session still reports connected = True. Measured 264 frames in, frozen, last event aging to 3.6 s, while the newcomer took 164 frames. So the widely repeated “one controller at a time” is wrong; it is one telemetry destination, last writer wins, and it fails quietly. Corrected in network.md and in mcbebop’s wording. Our own simulator modelled a refusal with a comment asserting that is what the aircraft does, which is the same unverified assumption encoded twice.
  • Found FreeFlight Pro 5.2.7 (the Bebop 2 app) on archive.org, third-party upload. Downloaded and checked, not installed. Signed Parrot Mykonos (2011); cannot cross-check against a second source. See freeflight6-usa-app.md.
  • FreeFlight Pro 5.2.7 signature cross-checked: APKMirror lists the same versionCode, size and SHA-256 signer fingerprint (88C44409…), and the key also signs AR.FreeFlight 2.4.15. Reasonable to sideload on the lab phone; not installed yet. See freeflight6-usa-app.md.
  • ADB survey of the running aircraft (read-only, agent). Full write-up in adb-survey.md. Headlines: the /factory partition holds this unit’s own production test records and per-unit IMU thermal calibration, none of it in the firmware image, and shows every Bebop 2 was test-flown and wobble-scored before shipping; the 48-flight counter lives in the motor controller’s NVRAM, not a file, which is why no filesystem search found it; the eMMC reports 1424 MB lifetime writes against 21 MB resident, so media was recorded and later removed; and three anonymous write-enabled FTP roots run as root, one of them the /update directory the bootloader scans for firmware, on an open access point.
  • The survey changed the device. /etc/ld.so.preload preloads a crash-dump library into every process and it treats SIGPIPE as fatal, so each | head wrote a ~350 KB crash report: 11 reports, ~3.5 MB, where the store had been empty. Rule recorded: never pipe into head on this device.
  • Diagnosed the “injection vector” the survey agent flagged: it was cross-talk in the mcadb MCP server between two concurrent callers, and the stray request was this session’s own backgrounded arp probe. Benign in origin but a real bug, and the agent’s refusal to act on it was right either way. Written up in adb-survey.md; do not run two sessions against one mcadb server until it is fixed.
  • Added lessons.md: the method notes, as distinct from the drone facts. The recurring one is that a simulator written from the same assumptions as the client agrees with it, which is how four protocol bugs passed 362 tests and were then found in minutes against the aircraft.
  • Peer predicted all 19 /sys/kernel/hsis/ pin values from the 2018 kernel image and all 19 matched the live aircraft, checked against the ADB survey data rather than by a new probe. Its hwrev = pcbrev + 11 rule fits this unit (3 -> 14), which is consistency rather than proof until a second unit tests it. One gap found: the decode lists a reset_gnss cell that the aircraft does not expose as a file, so struct cells and sysfs attributes are different sets of 22.
  • Worked out the ckcm.bin framing and built a log reader into mcbebop. The format turned out to be documented after all, in Parrot’s own ulogcat/libulogcat_ckcm.c, deleted from Parrot-Developers/ulog in 2017 — recovered from before commit cec8877 and matched against the strings in this unit’s libulogcat.so, so the spec in protocol.md is cited rather than reverse-engineered. It had been derived independently from the capture first, and the two agree; the capture then settled two things the source leaves open. Entry layout, the lossy priority letters, the absent pid and the RGB colour field are all in protocol.md.
  • The log’s timestamps are monotonic microseconds since boot, with no wall clock anywhere. The upstream ulogger driver switched clocks in Oct 2017 and the firmware’s build date brackets that change without deciding it, so the answer came from the data: userspace stamps top out at 2067 s rather than 1.7e9, and 779 ARlibs lines carry libARSAL’s own uptime in their text and agree to a constant 8 ms.
  • Settled the COMMANDS question: it is not an ARSDK trace. The tag is the biggest userspace one in the capture but carries no command ids at all, and the C function names it does carry resolve against arsdk-xml for only 4 of 11 symbols, 3.5% of the tag. The seven failures are real naming differences, including a Stat-for-State typo in the firmware and an Outdoor the firmware files under Common while the XML files it under ardrone3. mcbebop does not resolve them rather than shipping a mapping that reads as fact. Measurement in protocol.md.
  • mcbebop gained read_log and six MCP resources on branch wt-logs. bebop://commands is the one worth knowing about: the whole 264-command catalogue is readable with the aircraft switched off, so an agent can learn the protocol before anyone fetches a drone. The archived logs under Debug/archive/debug_NN.tar.lzo are deliberately not reachable from any tool — they are a previous owner’s sessions and carry network and location data.
  • reset_gnss question answered by the peer from the bytes: the struct holds cells with no sysfs file, including RESET_GNSS and GPS_FLASH, so the 22 files and 22 struct cells are different sets of the same size. It also corrected itself unprompted: only seven distinct GPIO names sit in the export table, stored twice, not fourteen, so seven of the field-to-name bindings are an agent’s reading rather than a verified decode. Values confirmed regardless.
  • Installed FreeFlight Pro 5.2.7 + OBB on the lab phone; runs, home screen shows “Not connected”. Declined My.Parrot sign-in. Drone test waits on bebop-main.
  • FreeFlight Pro 5.2.7 connected to the aircraft, which settles the app question from the other side: FreeFlight Pro is the Bebop app and FreeFlight 6 is not. The app-side session declined the firmware update, did not fly and did not sign in; its write-up follows.
  • Ryan swapped the battery, so the drone power-cycles. What that costs us, and what to expect on the next boot:
    • Debug mode lapses. Telnet 23 and ADB 9050 close; four presses needed again. My ADB registration was disconnected cleanly beforehand.
    • The live ulogcat ring is gone (256 KB, per-boot) and Debug/current/ rolls, which should push the ~750 KB ckcm.bin into a new debug_02.tar.lzo once it passes the 1 MB cap in system.conf.
    • The boot counter should read 16, from 15, and updater.log should gain one more no-op scan line. That is a free check on the boot-count finding next time a shell is available.
    • The 11 crash reports we created persist; they are on the eMMC, not in RAM.
  • FreeFlight Pro 5.2.7 connected to the live Bebop 2 (no props, bebop-main off the link). Captured only the “Drone memory” view: 0 media, 7.1 of 7 GB free. No video, update prompt or model name seen; battery swap interrupted. Phone restored. bebop-main’s experiment shows a second ARSDK client silently takes telemetry rather than being refused. See freeflight6-usa-app.md.
  • FreeFlight Pro, second run on the live Bebop 2: auto-connects, shows battery and storage; Video tab caption “Set 720p to improve live video stream resolution” agrees with the 856x480 stream. No flight/boot counters in the UI; no firmware prompt. My swipe flipped the app’s FPV Mode to MANUAL; restored to AUTO. Details in freeflight6-usa-app.md.
  • Connecting FreeFlight Pro changed three settings on the aircraft, found by reading telemetry before and after: geofence off to on, return-home delay 30 s to 3 s, banked turn off to on. The app-side session had navigated screens only and believed it changed nothing. Recorded in drone state; the lesson is that an app which syncs preferences writes on connect, and only a before-and-after read reveals it.
  • The app-side report of “max altitude 2000, max distance 30” was the pair transcribed swapped; the aircraft reports 30 and 2000 both before and after.
  • After the battery swap the drone reported 0% on a fresh pack, refreshing live. Probably the gauge re-characterising, but logged rather than assumed.
  • CORRECTION (from bebop-main’s before/after telemetry): FreeFlight Pro changed three drone settings just by connecting: geofence off to on, return-home delay 30 s to 3 s, banked turn 0 to 1. My earlier “nothing changed” was wrong. Also my Safety-tab pairing was swapped: altitude is 30 m and distance 2000 m, not the reverse (I used sort -u, which scrambled label order). Fixed in freeflight6-usa-app.md. Not reverted.
  • Used the drone’s own log to verify our client’s acking, which is otherwise silent to test: connected for 25 s and the aircraft logged zero new ack timeouts while delivering 1434 frames. Confirms the buffer-126 fix from the drone’s side rather than from our tests. The 79 timeouts already in this boot’s log all fall in the window where FreeFlight Pro was force-stopped, so they read as a controller that vanished without disconnecting rather than one that fails to ack. Technique and the log lines in protocol.md.
  • Reverted two of the three settings FreeFlight Pro changed (return-home delay back to 30 s, banked turn back off) and tightened max distance from 2000 m to 150 m, armed and then disarmed, with every value read back from the aircraft afterwards. The geofence was deliberately left on rather than reverted: with it off the distance limit is not enforced at all, so reverting it would have made the tightening cosmetic. Flagged to Ryan; one command to undo.
  • The 0% battery reading persists hours after the swap, so “the gauge needs time” no longer fits. FreeFlight Pro reported 100% for the same pack earlier.
  • Chased the 0% battery and it is not the battery and not the power cycle. The aircraft emits BatteryStateChanged twice with different values: 64 on buffer 126 and 0 on buffer 127, confirmed from the raw payload bytes. The 64 is real (the drone’s log shows it, declining 73 to 64 over an hour); the zeros are logged nowhere. Our store keeps the newest value, so the zeros win and the drone looks flat. Written up in protocol.md with a candidate cause (the phantom FLIR accessory this airframe keeps erroring on) marked unconfirmed.
  • Fixed the battery misreport without guessing. Events carry their source buffer now, the store keeps a value per source, and get_state shows both whenever they differ. preflight_check resolves it on the physical argument that a powered, talking drone is not at 0%, and says the figure was contested. 433 tests.
  • Settings restored on the aircraft by bebop-main (Ryan chose revert plus tighten max distance): return-home delay 30 s and banked turn 0 restored; geofence deliberately left on; max distance 2000 m to 150 m. All read back. Not a clean revert; see freeflight6-usa-app.md and drone state.
  • Ryan confirmed the geofence stays on, so the 150 m limit is enforced. Started settings-decisions.md as a decision log for the aircraft’s configuration, separate from the observations: what changed, who decided, why, and the value read back. It also flags two things worth deciding deliberately before a flight rather than inheriting: the 30 m altitude ceiling, and a return-home minimum altitude of 20 m sitting close under it.
  • Gave mcbebop’s simulator a video path, so a viewer can be developed with no aircraft: an RFC 6184 packetiser (media/rtp.py), video in sim.py gated on VideoEnable exactly as the drone gates it, and a capture/replay pair (media/capture.py) for the day the aircraft is available. ffmpeg decodes the simulated stream at 856x480, including when it joins a stream already running, which is the case a goggle viewer actually faces. 500 tests.
  • Corrected the video NAL-type note in protocol.md. The 1461/740/200 counts describe the outer header byte of each packet; the type a FU-A carries is in the next byte and was never read, so whether the stream contains IDR frames is still unknown. Also recorded the open question the counts raise: eight STAP-A per frame is far more aggregation than repeating parameter sets can explain.
  • Simulator can stream video now: an RFC 6184 packetiser, a capture/replay format for real drone RTP, and both gated behind VideoEnable as the aircraft gates it. Proved by decoding rather than assertion: ffmpeg reads the simulated stream at 856x480/30 fps, including a join two seconds in at a random mid-GOP offset. 500 tests. Built for the pinephone session’s goggle viewer.
  • A shape difference worth knowing: the aircraft sends about eight STAP-A per frame, far more than repeating parameter sets explains, so it either wraps most NALs in STAP-A regardless of size or uses several slices per picture. Our packetiser is RFC-correct and shaped differently, so a client tested only against the synthetic stream is tested against the wrong shape. One raw RTP capture from the aircraft settles that and the IDR question together, by reading the second byte of each FU-A payload.
  • Open-source alternative to the app: ArduPilot (MAVLink, QGroundControl etc.) via the button-3 hook, but it very likely stops the video since kk kills dragon-prog (inference). Ryan chose to drop the QGC test and focus on the Pine64 phone viewer (pinephone session). Android phone work is finished.
  • Captured raw RTP from the aircraft and settled the keyframe question: there are none. 6641 packets over 18 s, unwrapped through FU-A and STAP-A: 17631 non-IDR slices, 540 SEI, 18 SPS, 18 PPS, zero IDR. About 33 slices per picture, parameter sets once a second. That is slice-based intra refresh, a low-latency choice that avoids keyframe bandwidth spikes, and it explains the STAP-A-heavy packet mix as well. Written up in protocol.md.
  • Verified the battery fix on hardware: the store now reports sources {126: 100, 127: 0} and resolves to 100% while saying it was contested. Note the MCP server must be restarted to pick up code changes; the running one still had the pre-fix build and reported 0%.
  • Converted the raw capture to the simulator’s .rtpcap format and verified the whole loop end to end: the sim replays the real drone RTP with its original timing and ffmpeg decodes it at 856x480. A frame of Ryan’s room, captured from the aircraft, round-tripped through the fake one. Replay invocation sent to the pinephone session.
  • Their Cedrus result answers a question I could only warn about: the stateless V4L2 decoder does start without an IDR, first frame 3.38 s cold and 0.48 s when parameter sets repeat every 30 frames. The aircraft repeats them at 1 Hz, about every 30 frames, so a real cold join should behave like the fast case.
  • Peer audit found protocol.md still carrying the stale “one controller at a time; a second gets non-zero status” line that network.md had already corrected. Fixed, so the two agree.
  • The drone logs commands it receives, not just ones it sends: Cmd Rx lines under the COMMANDS tag, plus the full handshake JSON with the controller’s name. That makes the log a record of what any controller did, which is how the FreeFlight Pro question got answered.
  • Corrected my own attribution: FreeFlight Pro’s second session sent no setting changes at all, only set-date, set-time and the two state requests. So the three changed settings came from its first session, whose log the reboot destroyed. “The app pushes its preferences on connect” was too broad a claim from the evidence I had.
  • FreeFlight Pro sets the drone’s clock. The aircraft has no battery-backed RTC and boots at the epoch, so its timestamps are uptimes until a phone app connects and real wall-clock afterwards. Worth knowing before trusting any file time on the device.
  • Corrected the settings-push claim: the drone’s log of FreeFlight Pro’s second session shows only set date, set time, all settings, all states. No setting-change commands. The three changes most likely came from the first session (log lost to the reboot). Mechanism unproven; the surviving claim is that connecting changed settings nobody touched. Also: FreeFlight Pro sets the drone’s clock. See freeflight6-usa-app.md.
  • Decompiled FreeFlight Pro: on connect it saves its own settings copy, reads the drone, then replays the saved values (return-home delay, geofence, banked turn and ~17 more). Its defaults (3 s delay, banked turn on) match what the drone ended on. Contradicts the second-session log (no setting commands); cause not yet established. See freeflight6-usa-app.md.
  • Corrected myself twice over on the FreeFlight Pro question. A peer decompiled the app and found a save/update/restore path that replays its stored settings to the drone on every connect. That contradicted my claim that its second session sent nothing, so I re-read the log with a better query: the drone does not log a setting write as Cmd Rx, it logs the effect as a BEHAVIOUR_FEATURE_SETTINGS_MSG_SET_* line. Those are there, two to four seconds into the session I had called clean. The app does push on every connect; my “no setting commands” was an artifact of searching for the wrong pattern, and the retraction I asked the peer to publish was itself wrong.
  • Recorded in adb-survey.md that the system log carries the aircraft’s CPU id, its serial, the controllers that connected, and the GPS coordinates of its first fix. A drone log is not safe to paste into an issue or commit.
  • Settings-replay question resolved: the drone logs a setting write as a ..._MSG_SET_... line, not Cmd Rx. The second session shows the app writing geofence, return-home delay and max distance (2000) about 2 s after connect, as the decompile predicts. Claim stands; earlier “no setting commands” retracted. Connecting the app again will likely write max distance 2000 over the 150 m limit. See freeflight6-usa-app.md.
  • Pre-publish privacy audit of mcbebop (agent). Real identifiers exist only in gitignored captures/ and are in neither built artifact; git history is clean across all 226 objects. One fix in shipping code: the simulator emitted PI04 as a product serial, which is the Bebop 2 model-family prefix rather than this unit’s identity, so not a disclosure, but replaced with obvious nonsense so nobody quotes it as a specimen. Licence corrected from MIT to MIT AND BSD-3-Clause, since Parrot’s XML ships and the package cannot decode anything without it. The sdist exclusions are now a tested control, with each pattern proven to bite against the real identifiers before being trusted.
  • Tooling trap worth remembering, in lessons.md: grep here is ugrep honouring .gitignore, so it silently skips captures/, the one directory a privacy audit must read. The first pass reported clean having scanned nothing.
  • mcbebop 2026.10.2 published to PyPI. pip install mcbebop works for anyone; verified by installing from the index into a clean environment rather than trusting the upload. Source is public at git.supported.systems/warehack.ing/mcbebop, anonymously cloneable and checked clean of identifiers with full history. Licence MIT AND BSD-3-Clause, both texts shipped.
  • Two checks that lied today, both caught only by using controls: pypi.org/project/<name> answers a bot challenge with HTTP 200, which reads as “name taken” when the name is free; the simple index at /simple/<name>/ is the honest answer, and worth testing against a known-taken and a known-free name before trusting it. And uv publish prints nothing on success, so a silent exit 0 means it worked. I re-ran it twice believing it had failed, while PyPI’s index was merely lagging. The rule already in ~/.claude/rules/python.md covers this: treat any invocation as committed unless the server returns a clean error.
  • 2026-10-03: rewrote bebop-2 history to move the 150 MB FreeFlight Pro OBB into Git LFS (git lfs migrate import --include="*.obb"), force-pushed main 1e67c43 to 6841d16. All 63 commits kept, with identical author, date, subject and content except the OBB (now a 134-byte pointer) and .gitattributes. Every commit hash after ba7dbb4 changed, so older hashes quoted in these notes (for example 144a3e5, 03e5a51, f3116d8) no longer resolve in this repo; match by subject. A mirror of the old history is at ~/claude/drone/bebop-2-backup-pre-lfs-rewrite.git. The server keeps the old object until its own gc.
  • Side effect of the history rewrite, seen by bebop-main: after the force-push this checkout briefly had no refs/remotes/origin/main, so git status showed no upstream. I saw the ref present right after my push and do not know what removed it (a git gc/reflog expire does not delete refs; a prune from another session’s fetch would). A plain git fetch origin restores it and moves nothing. If you find the upstream missing, fetch; do not reset --hard to “fix” it.
  • Phone-to-controller-to-plane audit: FreeFlight Pro’s USB accessory filter lists Skycontroller 2 and 2P; it opens the accessory fd into a native mux that carries commands, video (RTP) and FTP. The controller relays to the drone over Wi-Fi, with a 20 Hz PCMD builder, a piloting-source handoff and NAT’d FTP. All from code; no live SC2 session captured. See freeflight6-usa-app.md.
  • First live USB look at the Skycontroller 2 (Ryan’s unit, 2026-10-03 about 19:08 to 19:17 MDT). Ryan connected the controller’s full-size Type-A port to the workstation with an A-to-A cable (“type A to usb 3”, his words). Nothing enumerated: lsusb and mcusb’s list_usb_devices showed the same devices before and after, no 19cf (Parrot) or 03eb (Atmel bottom board) ID appeared, and journalctl -k has no USB line at all for that connection, not even a failed descriptor read. Not recorded: whether the controller was powered on, and which workstation port was used. Reading (inference, not observed): the Type-A receptacle is the controller’s host port, which fits aoad in the nap image being the host side of Android Open Accessory. Two hosts on one cable present no device to each other. Untested: the micro-USB port, where the nap image would run an adb gadget with VID 19cf only if ro.debuggable=1.
  • pinephone confirms the SC2 is free for a phone-side capture and shares its USB-side notes: SC2 is the USB host, libmux v2 channels 1 to 6, product ids 0x090f/0x0915, and adb 9050 plus FTP 21 and HTTP 7711 open on the SC2 when given a USB-network device. Cross-referenced in freeflight6-usa-app.md.
  • 2026-10-04, Skycontroller PLF opened again (static, firmware-analysis session). Three agents read the nap-linux-2.0.0-rc3 image; reports saved as reverse-engineering/skycontroller- hardware-identity.md, -services-network.md, -software-protocols.md, each with a banner of what was re-checked. Decisive finding, re-derived by hand: the image’s own libarsdk.so names its model id 0x0913 SKYCTRL_NG and files nap_update.plf under 0x0903, while mpp_update.plf is 0x090f SKYCTRL_2. The live Skycontroller 2 (second-hand, from the skycontroller2 session) reports hostname mpp, model Skycontroller 2, gamepad 03eb:2406, kernel built 2017-10-24; the image says nap, Skycontroller, 03eb:2407, built 2017-08-09, and its mppd links only the nap hardware library. So the PLF is not the Skycontroller 2’s firmware. Agreement between the two: mppd carries the /video SDP relay template the live unit answers with. Comparison in reverse-engineering/skycontroller-live-vs-image.md. Withdrawn: the label “mpp = Skycontroller 1 board” in skycontroller-stick-path.md. Public ledger row replaced; story and mapping page wording passed to docs-site. Open: what the NG product is; the SC2’s own build, factory type and mapping folder (needs a read-only shell, Ryan’s call).

Comments

Spotted a mistake, or tried this on your own Bebop? Leave a note. Comments are kept on this site's own server.