TL;DR

Quit every coding agent and editor that launches MCP servers, then run uv cache prune. A running session holds the cache lock, so prune prints “Cache is currently in-use, waiting for other uv processes to finish” and, after uv’s lock timeout, gives up. Do not use --force to skip it. Expect much less back than prune reports: on this Mac it reported 111.1 GiB (about 119 GB) removed and the disk gained 27.5 GB, most likely because cloned files share blocks.

$ uv cache size -H
$ df -h /
$ diskutil info / | grep "Container Free Space"
$ lsof "$(uv cache dir)/.lock"
$ uv cache prune

My Mac was down to 38.4 GB free on a 1 TB disk. The largest directory on it was a 132.7 GiB uv cache that I had never thought about. The coding agents I use launch MCP servers with uvx, and that was enough to fill it.

uvx is uv’s command for running a Python tool from a temporary environment, without installing it. If you have never run uv by hand, an agent or editor ran it for you. This page is part of a series on cleaning up a developer Mac’s disk. Part one of the series, finding what’s filling the disk, is how to know the cache is the cause; you don’t need it to follow the steps here.

Check the size and location of the cache

The environments uvx builds go in uv’s cache. Ask uv for the size and the directory:

$ uv cache size -H

warning: `uv cache size` is experimental and may change without warning. Pass `--preview-features cache-size` to disable this warning.
132.7GiB

$ uv cache dir

/Users/<user>/.cache/uv

uv labels uv cache size as experimental, which is the reason for the warning. The number is still the one to start from. The --cache-dir option, the UV_CACHE_DIR and XDG_CACHE_HOME variables and the tool.uv.cache-dir setting move the directory, according to uv’s cache documentation (opens in a new tab). Every command below reads the directory from uv cache dir, so it follows a moved cache.

Find what fills the cache

Environments that uvx builds sit in the cache’s archive-v0 directory. Count its entries, then count the ones that are Python environments, which have a pyvenv.cfg at their root:

$ ls "$(uv cache dir)"/archive-v0 | wc -l
$ find "$(uv cache dir)"/archive-v0 -maxdepth 2 -name pyvenv.cfg | wc -l

On this Mac, archive-v0 held 4,005 entries, and 2,562 of them were environments. To see which MCP servers the environments belong to, count the executables with mcp in their names in the environments’ bin directories (a server named otherwise needs its own pattern):

$ ls -d "$(uv cache dir)"/archive-v0/*/bin/*mcp* | xargs -n1 basename | sort | uniq -c | sort -rn | head

Nearly all of the environments were copies of three MCP servers, with 1,150, 1,143 and 191 copies, plus 50 of a fourth. One server’s environments held only 158 distinct sets of installed packages, and environments built on the same day had the same installed packages.

Why each launch can add an environment

An MCP server configured as uvx <server>@latest asks uv to resolve the latest version on every launch. With uvx -v, the log shows uv fetching the latest version and revalidating the index. Launching one server did not explain more than a thousand copies, so I measured when a launch adds an environment:

When several sessions start together and the resolution isn’t cached yet, each launch builds its own environment. Once the resolution is cached, launches add nothing. That implies pinning a version, instead of @latest, would stop most of the copies. I didn’t test that. I also didn’t check whether a later release of uv reuses an environment that another process built moments earlier. All of this is uv 0.11.12.

Choose between uv cache prune and uv cache clean

uv cache prune removes the cache entries nothing uses, and all centralized project environments, which uv recreates as needed. uv cache clean removes every entry and empties the cache; uv cache clean <package> removes one package’s entries. Both are described in uv’s cache documentation (opens in a new tab). This page uses prune, which here took the cache from 132.7 GiB to 3.9 GiB and left 540 entries in place. I didn’t run clean.

Quit every session that holds the cache lock

Each running uvx process holds the cache lock, the .lock file in the cache directory, for as long as its server runs. Cache-modifying commands wait for the lock, so a prune started now waits too. List the holders:

$ lsof "$(uv cache dir)/.lock"

COMMAND   PID    USER   FD   TYPE DEVICE SIZE/OFF    NODE NAME
uv      <pid> <user>   9u   REG  <dev>        0 <inode> /Users/<user>/.cache/uv/.lock
uv      <pid> <user>   9u   REG  <dev>        0 <inode> /Users/<user>/.cache/uv/.lock

lsof prints nothing when the lock is free; empty output from a path that is not your cache directory is not a clear lock. Each row is one uvx process. To find the session that started it, look up the parent of each <pid>:

$ ps -o ppid= -p <pid>
$ ps -o comm= -p <ppid>

On this Mac, one holder was launched by a Claude Code session and another by the VS Code extension host. Quitting a session or an editor ends its uvx children, which release the lock:

1. Run lsof "$(uv cache dir)/.lock".
2. Look up the parent of each <pid>.
3. Quit that session or editor.
4. Run lsof again until it prints nothing.

Record free space before pruning

The comparison needs readings from before the prune, so take them now and save the output:

$ df -h /
$ diskutil info / | grep "Container Free Space"
$ uv cache size -H

Run the prune

This command deletes cached environments. With nothing running, it removed every environment in the cache, not only the ones that no process referenced. The next launch of each MCP server builds a fresh environment. Run it from a plain Terminal window, not from inside one of the sessions you quit:

$ uv cache prune

Pruning cache at: .cache/uv
Removed 7278379 files (111.1GiB)

It took 409 seconds. Do not interrupt it.

Check the space you got back

The free space went up by much less than 111.1 GiB. Run the same three commands again and compare them with the readings you saved:

The container gained 27.5 GB. uv reported 111.1 GiB removed, which is about 119 GB. uv prints binary units (GiB, 2^30 bytes) and diskutil prints decimal ones (GB), so do not compare the two figures as equals. Even in the same unit, the disk got back about a quarter of what uv reported.

The explanation that fits is the link mode; I didn’t test it directly. uv’s default on macOS is clone, per uv’s settings reference (opens in a new tab), which makes APFS copy-on-write clones. Environments share blocks with each other and with the cached wheels. uv 0.11.12 counts every cloned file at its full size, and deleting a clone frees nothing when another file still uses its blocks. uv’s cache documentation (opens in a new tab) describes a cache-physical-space preview feature that accounts for copy-on-write clones; I didn’t test it.

du -sh agreed with uv cache size at 3.9G after the prune. That is a drop of 128.8 GiB from 132.7GiB, while prune printed 111.1GiB removed; I did not find why the two differ.

A local snapshot could also hold space after a delete. On this Mac, tmutil listlocalsnapshots / listed none, and diskutil apfs listSnapshots showed none on the data volume, so no snapshot was holding any.

Confirm your MCP servers still start

The prune left archive-v0 with 540 entries and no environments, and emptied environments-v2. Start each agent or editor that uses MCP servers and open a session. Each server builds a fresh environment on its next launch; on this Mac, one appeared at the next session start. Check:

$ find "$(uv cache dir)"/archive-v0 -maxdepth 2 -name pyvenv.cfg | wc -l

A count above zero means at least one server built an environment. Check each agent’s MCP server list or status view for errors.

Do this with a coding agent

Paste this prompt into your own local coding agent. It measures and reports, and it proposes before anything is removed:

Investigate my uv cache on this Mac. Measure and report only; change nothing yet.

1. Report the cache directory (uv cache dir) and its size (uv cache size -H).
2. Report how many entries in archive-v0 are Python environments (a pyvenv.cfg
   at the root), and which tools they belong to (the executables in their bin
   directories).
3. Report free space two ways: df -h / and diskutil info / (Container Free
   Space).
4. List every process holding the cache lock (lsof on the .lock file in the
   cache directory), and for each one its parent process, so I can tell which
   agent or editor session to quit.
5. Separate what is safe from what is in use: running uv processes, open agent
   or editor sessions, and any project whose own virtual environment lives
   inside the cache. If this session is itself one of the lock holders, say so
   and stop; I will run the prune from a plain Terminal.

Then propose each removal as a separate item, with the command and the current
cache size as an upper bound. uv has no dry run, so do not state an expected
figure, and note that APFS clones make the real gain smaller. Remove nothing until I approve that item. Remove only
through uv's own command (uv cache prune). Never delete files under the cache
directory by hand and never use --force. Afterwards, report free space again
with the same two readings and say how far the figure uv printed is from the
space actually gained.

Common failures

Cache is currently in-use, waiting for other uv processes to finish

Cache is currently in-use, waiting for other uv processes to finish (use `--force` to override)

Another uv process holds the cache lock, and uv cache prune and uv cache clean wait for it. Run lsof "$(uv cache dir)/.lock", quit the session that owns each holder as described earlier, and run the command again. Nothing was changed while it waited: I stopped a prune that had waited eight seconds, and the cache was unchanged.

These uv cache commands give up after a timeout; uv’s cache documentation (opens in a new tab) gives its value and the UV_LOCK_TIMEOUT environment variable that changes it. Waiting longer does not help while a session still runs its server.

Do not use --force. The documentation reserves it for cases where no other uv process is reading or writing the cache, and a running uvx server is exactly such a process.

TL;DR

Record the free space, then narrow down one level at a time with du until a directory belongs to one tool. That tells you which cleanup you need. Clean it with that tool’s own command where it has one, never with rm on a cache a tool manages, and record the free space again; the cleanup’s own report is not what the Mac got back.

$ diskutil info / | grep "Container Free Space"
$ du -hd 1 ~ 2>/dev/null | sort -hr | head -15

My Mac had 37.7 GB free on a 1 TB disk. I expected Xcode to be the reason, since I keep a full set of iOS simulators, and Xcode did hold 79 GiB. The largest directory was a 132 GiB uv cache, which I had never thought about: most of the entries in it were environments uvx had built for MCP servers, covered in part 2. Which directory is largest on your Mac depends on what you build and which tools you run. The method for finding it is below, and the series covers each cause.

Record the free space before you start

On a modern Mac, df -h / reports the sealed system volume, so its Used column is misleading:

$ df -h /

Filesystem        Size    Used   Avail Capacity iused ifree %iused  Mounted on
/dev/disk3s1s1   926Gi    13Gi    40Gi    24%    484k  423M    0%   /

/ is the sealed system snapshot. Your files live on the Data volume in the same APFS container. Every volume in a container draws on one pool of free space, so Used for / counts the system alone, and Avail is the pool. The container’s own figure comes from the startup disk’s container only, so an external APFS drive cannot be mistaken for it:

$ diskutil info / | grep "Container Free Space"

   Container Free Space:      65.8 GB (65847853056 Bytes) (exactly 128609088 512-Byte-Units)

This reading was taken after the cleanup, so it does not match the df output above, which was taken partway through, or the 37.7 GB before any cleanup. df -h prints GiB and diskutil prints GB, so the same moment looks different in the two units. Use one of them for every reading, or the difference between the units looks like progress. A Mac keeps writing while you work, so treat a change of a gigabyte or two as noise.

Find what’s taking up space with du

du reads sizes and changes nothing. 2>/dev/null hides the permission errors that protected directories print.

1. List the largest directories in your home directory:

   Command

   $ du -hd 1 ~ 2>/dev/null | sort -hr | head -15

2. Skip the first line, which is the total for the directory you ran it on. Run the same command on the largest directory below it, and repeat until the directory belongs to one tool. On this Mac, that took three levels: ~/.cache, then ~/.cache/uv, then ~/.cache/uv/archive-v0.

3. Check the two places that hide the most, even if they didn’t top the first list:

   Command

   $ du -hd 1 ~/Library/Caches 2>/dev/null | sort -hr | head -15
   $ du -hd 1 ~/Library/Developer 2>/dev/null | sort -hr

4. For a package manager, ask the tool where its cache is rather than assuming a path. pip on macOS uses ~/Library/Caches/pip, not ~/.cache/pip:

   Command

   $ uv cache dir

   Output

   /Users/<user>/.cache/uv

   Command

   $ python3 -m pip cache dir

   Output

   /Users/<user>/Library/Caches/pip

A du total is an upper bound on what deleting frees. APFS clones share their blocks with the original until one of them changes, and du counts a cloned file at full size wherever it appears.

Match the directory to its cause

These are the directories that were large on this Mac, and the part of this series that covers each one:

For the uv cache, read part 2 before you prune: a running coding session holds a lock that makes the prune wait, and the prune reports far more than the disk gets back. The other parts are listed on the series page.

Some directories stay whatever their size. A Docker volume holds an app’s data, not cache. An app’s Application Support directory holds its state.

Clear Homebrew downloads and Ollama logs

Two directories on the list need no part of their own.

brew cleanup -s removes old versions and the cached downloads of anything no longer installed. It keeps the downloads of what is installed now, so on this Mac it freed about 270 MB of the 1.1 GiB:

$ brew cleanup -s

An app can keep its data somewhere else and still write its logs to the internal disk. My Ollama models live on an external SSD, linked into place:

$ ls -ld ~/.ollama/models

lrwxr-xr-x  1 <user>  staff  ... /Users/<user>/.ollama/models -> /Volumes/<drive>/ollama/models

The logs aren’t linked. List them first:

$ ls -lh ~/.ollama/logs

Here they had reached 947 MB in two files, server-1.log and app-1.log. Yours may have other names. Ollama has no cleanup command for them, so quit Ollama and delete the files that list showed, by name:

$ rm ~/.ollama/logs/server-1.log ~/.ollama/logs/app-1.log

Deleting a log leaves the models and the settings as they are.

Ollama wasn’t in Open at Login, but App Background Activity, under System Settings > General > Login Items, allowed it to run in the background. I turned that off. I didn’t check whether background runs are what filled the logs.

Record the free space again

Run the same command you started with:

$ diskutil info / | grep "Container Free Space"

Here it went from 37.7 GB before any cleanup to 65.8 GB after the uv prune. The cleanup commands had reported more than 140 GB between them. A cleanup command reports what it removed from its own storage; only the container’s figure says what the Mac got back.

Do this with a coding agent

Paste this prompt into your own local coding agent. It finds and reports, and it deletes nothing:

Find what is taking up space on this Mac's internal disk. Measure and report
only; delete, prune or clean nothing. Stay in my home directory, and don't
follow links to external drives.

1. Report free space with diskutil info / (Container Free Space) and
   df -h /, and say which units each prints.
2. Run du -hd 1 on my home directory, then on the largest subdirectory it
   lists (not the total line for the directory itself), and keep going down
   until each large directory belongs to one tool. Also check
   ~/Library/Caches and ~/Library/Developer. Ask package managers for their
   cache paths (for example uv cache dir, python3 -m pip cache dir) instead of
   assuming them.
3. For each directory over 1 GB, report its size, which tool owns it, what
   fills it, and the tool's own cleanup command if it has one.
4. Mark anything that is in use or holds work: running processes using the
   directory (lsof), open coding-agent or editor sessions, uncommitted or
   unpushed changes in any git checkout, Docker volumes, app data under
   Application Support. Say what would be lost if each were removed.
5. Note that du counts APFS clones at full size, so a total is an upper bound
   on what deleting frees.

Then stop and wait. I will choose what to clean.

Common failures

Free space barely moved after a cleanup

Docker reports space inside its VM disk, and du and uv cache size count cloned files at full size. Measure the container with diskutil info / before and after; that is the only figure that says what the Mac got back.

With Time Machine on, a local snapshot can keep a deleted file’s space in use. Time Machine keeps hourly snapshots for 24 hours and deletes them as they age or as space is needed, per Apple’s Time Machine guide (opens in a new tab). List them; on this Mac the list was empty:

$ tmutil listlocalsnapshots /

uv cache prune waits, then times out

uv cache prune waits while any other uv process is running, and every uvx server counts. After five minutes by default (UV_LOCK_TIMEOUT) it gives up; the uv cache documentation (opens in a new tab) covers the lock. Quit the coding sessions that started the servers, then run it again. To see what holds the lock:

$ lsof "$(uv cache dir)/.lock"

Part 2 shows how to find which session holds it.

npm cache clean left ~/.npm large

npm cache clean --force empties npm’s package cache, _cacache. The space is in _npx, where npx keeps the packages it runs; part 4 covers it.

I wanted every crossing of the I-90 floating bridge in 240 drive folders of dashcam footage, 1,958 GB, for a montage. I wasn’t going to watch the footage to find them, and my Rivian R1T’s GPS was too coarse to. What found them was the color of the lights in the Mount Baker tunnel, which sits at the Seattle end of the bridge. Brightness couldn’t see that tunnel. A red/blue ratio found it on 71 of the 79 bridge drives scanned, and a person looking at one frame per candidate confirmed 68 crossings. How to sort months of dashcam footage with Claude Code has the rest of the job this was one stage of.

Signals instead of pixels

The footage already describes itself if you read the right parts of it. Each 15-minute segment carries one GPS fix, at its start, in the container’s location tag. That sorts drives by area, but one fix per 15 minutes can’t place a bridge crossing to the minute, let alone the second.

The finer signal is in the keyframes. The truck writes one keyframe per second. Decoding only those, downscaled to 64 × 48, gives a per-second record of a drive and skips the 29 frames between them. Each second becomes four numbers: the mean change in green from the previous keyframe, which I called motion, and the mean red, green and blue. That record was computed once, cached as CSV, and every detector afterward read the cache instead of the video.

The first calibration read a motion of zero everywhere, parked or moving. The cause was ffmpeg duplicating the keyframes back up to 30 fps, so every frame was compared with a copy of itself. -fps_mode passthrough fixed it. The bug showed up only because the check ran on a known parked segment and a known moving one before it ran on anything else. A detector that reads zero on both cases is measuring itself.

Why brightness failed

A tunnel should be the easiest thing in a drive to find: the outside goes away and the picture gets darker. In this footage it doesn’t get darker. The camera’s auto-exposure raises the gain as the light drops and holds the frame’s brightness close to what it was in daylight. That’s what the exposure control is for, and it erases the one property a brightness threshold depends on.

Why color worked

Exposure scales brightness, and the hue of the lamps comes through. The Mount Baker tunnel is lit by sodium lamps, and sodium light is orange. Inside the tunnel, the front camera’s mean red rises above its mean blue. That ratio is the whole detector.

It runs with three guards:

• The red/blue ratio has to stay above 1.0 for 25–150 seconds, so a single orange frame or a long orange stretch doesn’t count.
• A rolling median smooths the ratio first, so one bright frame inside a tunnel doesn’t break the run in two.
• The first and last 3 minutes of every drive are ignored, because parking garages look the same to the ratio.

On the 79 bridge drives scanned, it found a tunnel on 71. Seven of the eight misses were after dark. My reading is that the ratio describes light, and once the light outside is artificial too, a tunnel stops standing out from the road around it.

Detector proposes, person confirms

Finding a tunnel isn’t the same as finding the right one. A drive on I-90 passes more than one stretch that reads as a tunnel, including the Mercer Island lid, and most drives had several candidates. The obvious rule, pick one tunnel by its position in the drive, was right about two times in three.

I could have spent the effort on a better rule. Instead the tool writes a guess for every drive and a contact sheet to check it against, then stops. Each row of the sheet is one drive, and each tile is a frame from the bridge side of one candidate. The bridge is unmistakable in a frame, so the right column is obvious, and the edit is one number in a picks file, which the cut turns into synced clips. Reviewing that way confirmed 68 crossings. Two night drives couldn’t be confirmed, because a dark frame of a dark bridge proves nothing, and one drive had no crossing among its candidates.

The layout of the sheet mattered as much as having one. The first version was a flat grid of every candidate frame, and it was easy to match a tile to the wrong drive. One row per drive makes that mistake visible rather than merely possible.

The sheet is also an instrument, and an instrument can be wrong. At one point I announced that the picks for some drives were the Mercer Island lid, not the Mount Baker tunnel. The sheets had been built with a seek that ffmpeg applied after tiling, so they showed a different window than the one they claimed. Extracting the same frame two ways exposed it, and the picks had been right all along. When a result contradicts work that was already verified, check the instrument before the conclusion.

The review step reads like an admission that the detector wasn’t good enough. I think it’s the reason the detector could stay simple. Take a rule that’s right two times in three, and add a sheet a person reads at a glance. That beats a rule that’s right more often and still needs someone to check it.

Signals that could not separate the cases

The same per-second record answered four other questions, and each answer was no for the same reason: the signal moved for the thing I wanted and for something I didn’t.

• Parked footage. Near-zero motion found parked time reliably, but there was little to find: 23 GB of the 1,760 GB scanned, because with Gear Guard turned off, the truck stops recording seconds after it parks.
• Hard stops. A sudden change in the picture marks braking, but turning changes the picture as much. Of 875 candidates, all 12 at the top were turns into parking lots or ordinary stops.
• Traffic jams. Sustained low motion should mark crawling traffic, but the neighboring cars keep moving, so the picture keeps changing and a jam looks like driving.
• Garage boundaries. A gray band at the top of the frame marks a concrete ceiling on a clear day, but an overcast sky is the same gray. The boundary was set by timing instead.

None of these is fixed by tuning a threshold. When a signal can’t tell two cases apart, a better number for the threshold still can’t.

Close passes did work once the frame’s overall brightness change was subtracted from the side cameras: 37 of the top 60 were real passes. A person still had to sort them.

Signals the camera leaves alone

A camera processes its own picture before anything reads it, and that processing exists to make the picture look normal. Auto-exposure made a tunnel look like daylight and left the lamp color intact. Before measuring anything in recorded video, ask which properties the camera flattens and which it passes through, and build on the second kind.

TL;DR

The camera didn’t stop recording. It restarted seconds later into a new drive folder that holds only the restarted cameras. Look for a folder named a few seconds after the broken one, with fewer than five cameras in it.

Don’t take the offset from the two folder names. Measure it by correlating the motion of two cameras that face the same way, such as the Gear Guard camera and the rear camera. Then re-encode the restarted cameras from the sibling folder at that offset, so that they start on the exact frame.

I cut a five-camera clip from a drive on August 7 and dropped it into a multicam timeline. Three angles played to the end. The rear camera and the side-left camera went black after 33 seconds. Every file was readable and nothing reported an error. The footage I thought was lost was in the next folder.

Rear or side camera stopped recording after a few seconds

The symptom has a few forms, depending on what you build from the drive:

• A camera’s only file in the drive folder is a few seconds long, while the other cameras have a full set of 15-minute segments.
• A multicam clip or a grid of the five cameras shows black tiles for one or two angles for most of the drive.
• A hyperlapse of one angle is much shorter than the others, or ends in black.

On the August 7 drive, the folder 08_07_26_140000 holds full segments for the front, Gear Guard and side-right cameras, and one 33-second file each for the rear and side-left cameras:

$ ffprobe -v error -show_entries format=duration -of csv=p=0 \
    08_07_26_140000/08_07_26_140000_rearCenter_001.mp4

33.000000

Cause: the camera restarted into a sibling folder

The truck writes one folder per drive, named MM_DD_YY_HHMMSS in local time, and one file per camera per 15-minute segment inside it. When a camera stops and restarts partway through a drive, its recording doesn’t continue in the drive’s folder. A new folder is created a few seconds later, and it holds only the cameras that restarted. The rest of the drive stays in the original folder.

The two folders together hold the whole drive. Nothing was lost, but the drive is now split across two clocks.

Find the sibling folder

Run this in the DriveStream folder of the truck’s drive, or of your copy of it. It prints each drive folder with the cameras it holds:

for d in */; do
  printf '%s ' "${d%/}"
  ls "$d" | grep -v '^\._' \
    | sed -E 's/^[0-9_]+_([A-Za-z]+)_[0-9]{3}\.mp4$/\1/' \
    | sort -u | tr '\n' ' '
  echo
done

The grep -v '^\._' drops the AppleDouble files that macOS leaves beside every clip on the truck’s exFAT drive. The August 7 folders look like this:

08_07_26_100000 frontCenter gearGuard rearCenter sideLeft sideRight
08_07_26_140000 frontCenter gearGuard rearCenter sideLeft sideRight
08_07_26_140034 rearCenter sideLeft
08_07_26_160000 frontCenter gearGuard rearCenter sideLeft sideRight

A sibling folder has all three of these properties:

• Its name is seconds, not minutes, after the previous folder’s name.
• It holds fewer than five cameras.
• Its cameras are the ones that stop early in the previous folder.

A short folder with one camera and no partner is a different case: a stray fragment, not half of a drive. Across 233 drive folders, one was a 21-second fragment from a single camera.

Folder name offset is wrong

The folder names differ by 34 seconds, so the obvious fix is to place the restarted cameras 34 seconds into the drive. Measured from the footage, the restarted cameras begin 32.7 seconds into the drive. The names are wrong by more than a second. The files’ creation_time tags are no better: the two folders' differ by 35 seconds, a third answer.

A folder name is a label, not a clock: nothing ties it to the first frame of the video inside. An error of more than a second is visible in a multicam edit. A car that passes the truck crosses the rear view and the side view at moments that no longer match. The truck’s files carry no timecode track to correct it, so the offset has to come from the pictures.

Measure the offset by motion correlation

Two cameras that see the same scene change at the same moments. When the truck hits a bump or a car passes behind it, both backward-facing pictures move at once. The script reduces each camera to a motion signal and slides one signal along the other until they match best:

1. Decode a window of each clip at a fixed sample rate, scaled to a tiny grayscale frame.
2. Take the mean absolute change between consecutive frames. That’s the motion signal: high when the picture changes, low when it holds still.
3. Standardize the reference, and each stretch of the other signal it’s compared with, so that a brighter or busier camera or stretch doesn’t outweigh the rest.
4. Try every lag within a search range around your guess, and keep the lag with the highest correlation.

It needs Python 3 and ffmpeg on your path, and nothing else:

"""Measure the time offset between two dashcam clips by correlating their motion.

Usage:
    python3 measure_offset.py REFERENCE OTHER --at SECONDS [--guess SECONDS]

Reads a window of REFERENCE starting at --at, and a slightly wider window of
OTHER around --at minus --guess. Prints the offset that lines the two up:
a moment at time t in REFERENCE is at time t - offset in OTHER.
"""
import argparse
import subprocess

WIDTH, HEIGHT = 32, 24
SAMPLES_PER_SECOND = 10


def motion(clip: str, start: float, seconds: float) -> list[float]:
    """Mean absolute change between consecutive tiny grayscale frames."""
    raw = subprocess.run(
        ["ffmpeg", "-nostdin", "-v", "error",
         "-ss", f"{start:.3f}", "-t", f"{seconds:.3f}", "-i", clip,
         "-vf", f"fps={SAMPLES_PER_SECOND},scale={WIDTH}:{HEIGHT},format=gray",
         "-f", "rawvideo", "-"],
        capture_output=True, check=True,
    ).stdout
    size = WIDTH * HEIGHT
    frames = [raw[i:i + size] for i in range(0, len(raw) - size + 1, size)]
    return [sum(abs(a - b) for a, b in zip(before, after)) / size
            for before, after in zip(frames, frames[1:])]


def standardize(values: list[float]) -> list[float]:
    mean = sum(values) / len(values)
    variance = sum((v - mean) ** 2 for v in values) / len(values)
    spread = variance ** 0.5 or 1.0
    return [(v - mean) / spread for v in values]


def best_lag(reference: list[float], other: list[float]) -> tuple[float, int]:
    """Slide reference along other; return the best correlation and its lag."""
    scores = []
    for lag in range(len(other) - len(reference) + 1):
        window = standardize(other[lag:lag + len(reference)])
        score = sum(a * b for a, b in zip(reference, window)) / len(reference)
        scores.append((score, lag))
    return max(scores)


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("reference")
    parser.add_argument("other")
    parser.add_argument("--at", type=float, required=True,
                        help="window start in REFERENCE, seconds")
    parser.add_argument("--guess", type=float, default=0.0,
                        help="expected offset, seconds")
    parser.add_argument("--length", type=float, default=120.0,
                        help="window length, seconds")
    parser.add_argument("--search", type=float, default=5.0,
                        help="how far either side of the guess to look")
    args = parser.parse_args()

    reference = standardize(motion(args.reference, args.at, args.length))
    # The other window starts --search seconds early, so lag 0 means guess + search.
    other_start = args.at - args.guess - args.search
    other_length = args.length + 2 * args.search
    other = motion(args.other, other_start, other_length)

    score, lag = best_lag(reference, other)
    offset = args.guess + args.search - lag / SAMPLES_PER_SECOND
    print(f"window at {args.at:g} s: correlation {score:.2f}, offset {offset:+.1f} s")


if __name__ == "__main__":
    main()

Give it a camera from the original folder as the reference, a camera from the sibling folder as the other, and the folder-name difference as the guess. Both folders number their segments from 001 and every segment but the last runs 900 seconds, so segment 002 of the sibling starts the same offset after segment 002 of the original. Compare files with the same segment number, and keep the window inside one segment.

In the DriveStream folder, run:

$ python3 measure_offset.py \
    08_07_26_140000/08_07_26_140000_gearGuard_001.mp4 \
    08_07_26_140034/08_07_26_140034_rearCenter_001.mp4 \
    --at 300 --guess 34

window at 300 s: correlation 0.84, offset +32.6 s

One window is one measurement. Repeat it in windows that don’t overlap, from different segments, and trust the offset only when they agree. Four windows on the August 7 drive gave this:

That’s 32.7 seconds, plus or minus 0.1, from four independent windows. A moment at time t in the original folder is at t − 32.7 in the sibling folder.

Compare cameras that face the same way

The front camera was my first choice for the reference, because it recorded the whole drive. Against the restarted rear camera it gave two answers:

The two windows disagree by 0.9 seconds, and each one looks plausible on its own. The front camera sees the road ahead and the rear camera sees the road behind. Their pictures change for different reasons most of the time, so the best match moves from one window to the next. The Gear Guard camera looks backward over the bed, the same direction as the rear camera, and the two agree.

Pick a reference camera that kept recording in the original folder and faces the same direction as one of the restarted cameras. The files in a sibling folder share one start time and the same segment lengths, so the offset you measure for one of its cameras applies to all of them.

Stitch the restarted cameras

A stream copy can only start on a keyframe, and the truck writes one keyframe per second. The stream-copied angles from the original folder start on their own keyframes. An offset of 32.7 seconds isn’t a whole number of seconds, so no keyframe in the sibling folder falls on the same instant. Re-encode only the restarted cameras, cut at the exact frame, and leave the rest as stream copies.

Suppose your clip covers START to END seconds of the original folder, and OFFSET is the measured offset. The window has to begin after the restart, so START is greater than OFFSET. In the sibling folder, the clip begins at START minus OFFSET and lasts END minus START seconds.

In the DriveStream folder, the following commands write a concat list of the restarted camera’s segments in the sibling folder, and then cut the clip from the joined segments, re-encoding to HEVC. They create two files and change nothing in the drive folders:

printf "file '%s'\n" "$PWD"/08_07_26_140034/08_07_26_140034_rearCenter_*.mp4 \
  > rearCenter.txt
ffmpeg -f concat -safe 0 -i rearCenter.txt -ss <sibling-start> -t <duration> \
  -c:v hevc_videotoolbox -q:v 80 -tag:v hvc1 -color_range pc -an \
  <clip-folder>/rearCenter.mp4

Run the same two commands for every other camera in the sibling folder.

-ss after -i decodes from the start of the list and discards frames up to the exact time, so the output starts on that instant rather than on the nearest keyframe. It’s slow for a window late in the drive, because ffmpeg decodes everything before the window first. Joining the segments in one concat pass, rather than cutting each file and joining the cuts, keeps the frames whose group of pictures runs across a segment boundary. Cut multi-camera dashcam clips with ffmpeg covers the stream-copied cameras and that join in full.

The encoder hevc_videotoolbox is the hardware HEVC encoder on macOS. On another system, use the encoder you’d use for the rest of the project. The option -color_range pc labels the re-encoded angle full range like the truck’s footage, so it matches the stream-copied ones; check it with ffprobe -show_entries stream=color_range. The truck records no audio, so -an drops nothing.

Every file now starts at time 0. The re-encoded angles start on the exact frame, and a stream-copied one starts early by the lead the cut script printed for it. In DaVinci Resolve, mark each lead as an In point, then create the multicam clip with angle sync set to the In point.

Confirm the stitch

Run the same script on the finished clip, with the guess left at zero. In the clip’s folder, compare the Gear Guard camera, which was stream-copied from the original folder, with the re-encoded rear camera:

$ python3 measure_offset.py 08_07_26_140000_gearGuard.mp4 \
    08_07_26_140000_rearCenter.mp4 --at 300 --length 90

window at 300 s: correlation 0.79, offset +0.1 s

An offset of +0.1 seconds is one sample step of SAMPLES_PER_SECOND, the resolution of the method. The angles are in sync. The same script reported 32.7 seconds on the unstitched folders, so it does notice an offset when one exists. A zero from a check that has never reported anything else proves nothing.

TL;DR

The GPS position is in the MP4 container’s location tag. Read it with ffprobe:

$ ffprobe -v error -show_entries format_tags=location -of default=nw=1:nk=1 \
    <clip>.mp4

47.589462-122.2655

The value is a signed latitude immediately followed by a signed longitude, in decimal degrees: 47.589462 north, -122.2655 west. That clip starts on the I-90 floating bridge. 0,0 means the truck had no fix.

There is one fix per file, taken where the 15-minute segment began. There is no per-second GPS track.

Spellings of the tag

The value has no separator between the two numbers. The sign of the longitude is the separator, so a parser has to split on the second + or -, not on a comma or a space. The truck never writes a leading plus sign. Across 475 front-camera segments, all from encoder tag Rivian Firmware ver. 2022-47, three spellings turned up, and 0,0 stands in when there is no fix:

Usual

47.589462-122.2655, on 439 of the 475 segments.

Trailing slash

the same form ending in /, on the other 36.

After ffmpeg copies the file

+47.5894-122.2655/, the ISO 6709 form, with a leading plus sign and four decimal places. This is ffmpeg’s rewrite, not a firmware variant.

No fix

0,0, which is a position in the Gulf of Guinea, not a reading.

ffprobe also prints a location-eng tag with the same value. Either one works; the commands here read location.

What one fix per segment rules out

The truck records one file per camera per 15-minute segment, and each file carries a single fix from the moment that segment started. Nothing inside the file records where the truck was a minute later. That rules out:

• A route drawn from the video. A segment gives you a point every 15 minutes, which is too coarse to follow a road.
• Speed. Two fixes 15 minutes apart give an average over the whole segment, not the speed at any moment.
• Finding a place by GPS alone. A drive across the floating bridge takes less than a segment, so the bridge usually falls between two fixes.

The five cameras of a segment each carry their own fix. In the clip earlier, the front and rear files read 47.589462-122.2655 and the two side cameras and the Gear Guard camera read 47.589474-122.2659. Treat any one of them as the segment’s position.

To place a moment inside a segment, read the picture instead of the container. Finding a bridge in dashcam video by the color of tunnel lights locates the I-90 floating bridge crossing from the picture, where the GPS can’t.

Parse the tag in Python

parse_location returns a (latitude, longitude) tuple, or None when the tag is missing, unreadable, or 0,0. read_location runs ffprobe on one file. The standard library is enough; Python 3.10 or later, for the | in the type hints.

import re
import subprocess
import sys

# A signed latitude immediately followed by a signed longitude. The leading
# plus sign and the trailing slash are optional because not every writer
# emits them.
LOCATION = re.compile(r"([+-]?\d+(?:\.\d+)?)([+-]\d+(?:\.\d+)?)")


def parse_location(tag: str) -> tuple[float, float] | None:
    match = LOCATION.search(tag)
    if not match:
        return None
    latitude, longitude = float(match[1]), float(match[2])
    # The truck writes 0,0 when it has no GPS fix.
    if latitude == 0 and longitude == 0:
        return None
    return latitude, longitude


def read_location(clip: str) -> tuple[float, float] | None:
    tag = subprocess.run(
        ["ffprobe", "-v", "error", "-show_entries", "format_tags=location",
         "-of", "default=nw=1:nk=1", clip],
        capture_output=True, text=True, check=True,
    ).stdout
    return parse_location(tag)


if __name__ == "__main__":
    for clip in sys.argv[1:]:
        point = read_location(clip)
        print(clip, "no fix" if point is None else f"{point[0]}, {point[1]}")

Pass it one or more clips:

$ python3 rivian_location.py <clip>.mp4

<clip>.mp4 47.589462, -122.2655

The tests cover every spelling in the list earlier, a southern and eastern position so both signs are exercised, and each way the tag can mean “no fix”:

import pytest

from rivian_location import parse_location


@pytest.mark.parametrize(("tag", "expected"), [
    ("47.589462-122.2655", (47.589462, -122.2655)),
    ("47.589462-122.2655/", (47.589462, -122.2655)),
    ("+47.5894-122.2655/", (47.5894, -122.2655)),
    ("-33.8688+151.2093/", (-33.8688, 151.2093)),
])
def test_reads_each_spelling_of_the_tag(tag, expected):
    assert parse_location(tag) == expected


@pytest.mark.parametrize("tag", ["", "\n", "garbage", "+0.0000+0.0000/", "0-0"])
def test_returns_none_for_no_fix(tag):
    assert parse_location(tag) is None

$ python3 -m pytest -q

.........                                                                [100%]
9 passed in 0.01s

Dropping the optional sign from the latitude’s half of the pattern fails the southern-hemisphere case, so the tests do notice a broken pattern.

Remove the GPS location from a clip before sharing it

The tag travels with the file. A raw clip posted online tells anyone who downloads it where it was recorded, and a folder of them maps where the truck goes every day. Trimming the clip doesn’t necessarily remove it: a stream copy with ffmpeg kept the tag, rewritten in the +…/ form, and exiftool reports it as a GPS position:

$ ffmpeg -i <clip>.mp4 -t 2 -c copy copy.mp4
$ exiftool -G1 -a "-gps*" "-location*" copy.mp4

[Composite]     GPS Altitude                    : 0 m
[Composite]     GPS Altitude Ref                : Above Sea Level
[Composite]     GPS Latitude                    : 47 deg 35' 22.02" N
[Composite]     GPS Longitude                   : 122 deg 15' 55.76" W
[Composite]     GPS Position                    : 47 deg 35' 22.02" N, 122 deg 15' 55.76" W
[UserData]      Location Information            : (none) Role=shooting Lat=47.58945 Lon=-122.26549 Alt=0.00 Body=earth Notes=

Other ways of writing a file behave differently. The multi-camera clips cut with ffmpeg’s concat demuxer came out with no location tag at all, because the demuxer drops file-level metadata. The hyperlapse tool I used kept it, because it copies the first source file’s metadata on purpose. Check each file rather than trusting the tool. To drop the container’s metadata while copying, add -map_metadata -1; the same exiftool command then prints nothing:

$ ffmpeg -i <clip>.mp4 -t 2 -c copy -map_metadata -1 strip.mp4
$ exiftool -G1 -a "-gps*" "-location*" strip.mp4

Empty output is the result you want. Run the check on the file you’re about to upload, not on its source: each tool between the two can keep the tag or drop it.

TL;DR

Give Claude Code one prompt per stage: survey, move, find, cut, publish. Each prompt states what the agent can’t touch and carries three demands: run on one drive first, break your own check on purpose and confirm that it fails, and list what you couldn’t verify. The agent reads the GPS tag and the keyframes instead of the pixels and proposes candidates. You confirm every pick on a contact sheet before anything is cut, moved, or deleted.

Driving from July to September filled the truck’s 2 TB dashcam drive, a SanDisk Extreme Portable SSD (opens in a new tab) (affiliate link): 240 drive folders, 1,958 GB, five cameras recording on every trip. I wanted to keep only what might make a video, and the backup drive couldn’t hold all of it. Watching it wasn’t an option. I directed Claude Code in plain language instead. It sorted the drives by where they went using GPS, found the I-90 floating bridge by the color of the tunnel lights, cut synced clips, and rendered a hyperlapse of every drive.

Almost every failure in that work was caught by a check: a size comparison, a frame count, a positive control, a contact sheet, or me. One failure was caught by nothing for hours. The prompts in this guide are general versions of the ones I typed, rewritten for your own footage, and each one carries the guardrail that the failure taught.

Signals already in the files

A Rivian writes one folder per drive and one MP4 per camera per 15-minute segment. Two things in those files do most of the work, and neither needs anyone to watch the video:

• GPS. The container’s location tag holds one fix per segment, taken at its start. That’s the only GPS in the file; there’s no per-second track.
• Keyframes. Every camera writes one keyframe per second. Decoding only the keyframes, downscaled to 64 × 48, gives a per-second record of motion and color without decoding the frames between them.

Read the tag from any clip:

$ ffprobe -v error -show_entries format_tags=location -of default=nw=1:nk=1 <clip>.mp4

It prints a signed latitude immediately followed by a signed longitude, such as 47.589462-122.2655. Some tags end in / and most don’t, and 0,0 means the truck had no fix. Read the GPS location from Rivian dashcam video with ffprobe covers every spelling. Other dashcams store GPS differently, which is why the first prompt asks the agent to find out rather than assume.

Tools and rules for every stage

You need:

• Claude Code, run in a terminal on the machine that has the footage mounted
• The ffmpeg and ffprobe commands on the PATH
• The exiftool command, for the privacy stage
• A second drive with room for whatever you keep

Every prompt carries three rules, worded for its stage. They’re repeated on purpose, so that a prompt pasted on its own still carries them:

• One drive first. A bug found on the first drive costs one drive’s time. The same bug found late in a batch costs every drive before it.
• Break the check on purpose. A check that has never failed proves nothing. The agent damages its own input or output, confirms the check catches it, and then restores it.
• Say what couldn’t be verified. Matching file sizes aren’t matching content, and a report that doesn’t say so reads as if they were.

Survey every drive without changing anything

In Claude Code, from the folder that holds the drive folders, paste:

The folder <footage-folder> holds video from my truck's dashcam: one subfolder per drive, and one MP4 per camera per segment. Survey it without changing anything.

For every drive folder, use ffprobe to read each segment's duration and GPS location tag. Report each drive's start time, length, GPS points, the distance between its two farthest points, and its size on disk. Ignore any ._ files macOS has left beside the clips on a FAT32 or exFAT drive.

Rules:
- Read only. Do not move, rename, re-encode, or delete anything.
- Run on one drive first and show me the result before you run the rest.
- The location tag is written in more than one format. Parse it permissively, and treat 0,0 as no fix, not as a place.
- Before you run the whole batch, feed your parser a malformed tag on purpose and confirm that it reports the drive instead of skipping it silently.
- Cache everything you read in one file, so that later steps don't read the drives again.
- Then propose a way to sort the drives into routine and unusual, and wait. I'll tell you my regular routes.
- Finish with a list of the drives you couldn't read and anything you couldn't verify.

What went wrong here. I asked for road trips, and the agent ranked drives by distance. The ranking itself showed there were none: no drive in the batch went far. Distance was the wrong question, and only I could supply the right one: the corridor across the I-90 bridges that most of the drives used. With that, the classification became a corridor test. A drive whose every GPS point falls inside a box around that corridor is a corridor drive; anything else is worth a look. The test found 168 corridor drives, 1,594 GB, and 14 off the corridor worth keeping, 171 GB. That’s why the prompt ends by waiting for you. Your routes are context the agent can’t get from the files.

The parser also failed on its first real run, on a tag that ended in /. The fix was a pattern that finds two signed numbers wherever they sit. Expect the same from your own camera’s metadata, and test both spellings.

Move the drives you keep

Decide which drives to keep from the survey, then paste:

Move these drives from <footage-folder> to <backup-drive>/<kept-folder>: <list of drive folders>.

Rules:
- Before you copy anything, compare the total size with the free space on <backup-drive>, and stop if it won't fit.
- Copy first. After each drive, compare every file's size on the destination with the source.
- Do not delete anything from the source until I say so. Show me the drives whose sizes all match, and wait.
- When I confirm, remove the source files with a tool that removes only what it has copied, such as rsync --remove-source-files. Do not rely on shell-specific glob features.
- Skip any folder that another process, such as Finder, is still writing to or moving.
- Report the empty folders that a move leaves behind. Do not delete them, and make every later step skip them.
- Move one drive first and show me the result.
- Before the batch, truncate a copy in a scratch folder and confirm your size check reports it.
- Say plainly that a size match isn't a content match, and list anything you couldn't verify.

What went wrong here. A queued move used a zsh glob qualifier that the shell Claude Code runs commands in doesn’t enable. It would have copied every file and deleted none. Inspecting the running process caught it, and the move switched to rsync --remove-source-files, which removes only what it copied. Later, a cleanup command of the form rm -f $B/*.jpg was blocked by Claude Code’s own guard, because an unset $B would expand the path to /*. The cleanup wasn’t needed, and the guard was right to stop it.

The moves also left empty drive folders behind. Nothing failed because of them until the hyperlapse run, in a later stage, which is why this prompt asks for them to be reported now.

Find a landmark from computed signals

GPS places a drive to within a 15-minute segment, which is too coarse to find a bridge crossing. The picture has to do it, and the cheap way to read the picture is to compute a few numbers per second once and reuse them. Paste:

I want the moment in each drive where the truck passes <landmark>. The GPS has one point per segment, which is too coarse, so find it from the picture.

1. From keyframes only, downscaled to 64x48, compute per-second signals for each camera: the mean change in green from the previous keyframe, and the mean red, green, and blue. Cache them per drive and per camera, so that no later step decodes video again.
2. Propose a signal that marks <landmark>, and explain why it should separate the landmark from things that look like it.
3. Propose candidates. Do not pick for me.

Rules:
- Read only on <footage-folder>. Write only under <work-folder>.
- When you extract keyframes as raw frames, make sure ffmpeg doesn't duplicate them back up to the source frame rate.
- Before you scan everything, calibrate the signal on a drive where I know the landmark is present and a drive where I know it isn't. If both read the same, suspect your extraction before the signal.
- Run on one drive first.
- Build a contact sheet with one row per drive, labeled with the drive's name, and one frame per candidate in time order. Write a picks file with your best guess for each drive, which I can edit. Then wait for me.
- Ignore the first and last few minutes of each drive, where the truck is parking or leaving.
- If a result contradicts something already confirmed, check your tool before you report the contradiction: extract the same frame two ways and compare them.
- List the drives with no candidate, and anything you couldn't verify.

For a tunnel, the signal turned out to be color. Auto-exposure keeps a tunnel’s brightness close to daylight’s, but sodium lighting tips the front camera’s red/blue ratio above 1.0. A run of 25–150 s above that ratio, away from the first and last 3 minutes of the drive, found a tunnel on 71 of the 79 bridge drives scanned. Seven of the eight misses were after dark. Finding a bridge in dashcam video by the color of tunnel lights is the full account, including what didn’t work.

What went wrong here. The first calibration read zero change between every pair of keyframes, parked or moving. A signal that’s flat on a known positive and a known negative is a broken instrument, not a finding. The cause was ffmpeg duplicating keyframes back up to 30 fps, so consecutive frames were identical. -fps_mode passthrough fixed it. That bug is why the prompt asks for a known positive and a known negative before the scan.

The automatic pick, one tunnel chosen by its position in the drive, was right about two times in three. The rest landed on other tunnels. The contact sheet is what turned an unreliable detector into a usable one. It showed one frame per candidate, taken from the bridge side of the tunnel. From it, I confirmed 68 crossings, couldn’t confirm two night drives, and found none on one. The first sheet was a flat grid of every candidate frame, and it was easy to match a frame to the wrong drive. One row per drive makes that mistake visible.

The worst wrong claim of the whole session came at this stage. Partway through, the agent reported that the picks for some drives were the wrong tunnel, the Mercer Island lid instead of the Mount Baker tunnel. The contact sheets had been built with an output-side -ss, which ffmpeg applied after tiling, so the sheets showed a different window from the one the picks named. Extracting one frame two ways exposed it. The picks had been right, and the claim was retracted.

Cut synced clips and render hyperlapses

With picks confirmed, paste:

For each confirmed pick in <picks-file>, cut a clip from every camera, from <seconds-before> seconds before the landmark to <seconds-after> seconds after it, without re-encoding. Then render a 30x hyperlapse of every drive for every camera, decoding keyframes only.

Rules:
- Read only on <footage-folder>. Write only under <work-folder>.
- A stream copy can start only on a keyframe. Measure each camera's keyframe timing first; do not assume the cameras share it.
- A camera's first frames in a segment file can depend on frames at the end of the previous file. Cut across segment files in one pass, and check the joins.
- After cutting, compare every camera's frame count, duration, and start_time. Same length isn't the same as in sync.
- Pad a camera that stopped early with black frames, per segment, so that every camera stays the same length as the front camera. Give the black the footage's pixel format and color range, and check the range of every joined file.
- Do not trust folder names or file timestamps for timing. If a camera restarted into another folder, measure its offset from cameras that see the same scene, in more than one window. If that offset doesn't land on one of the camera's keyframes, re-encode only that camera, cut at the exact frame, and tell me which files were re-encoded.
- Test on one drive first, and include a drive where a camera stopped early.
- Before the batch, shift one camera by a second on purpose and confirm that your sync check fails. Then restore it.
- In the batch, one bad drive must be logged and skipped, not stop the run. Write each output to a temporary name and rename it when it's complete, so that a rerun resumes.
- Time the test drive and estimate the batch from that.
- If the run will outlive this session, start a watcher that tells me when the run ends or stops making progress.
- List the drives that were skipped or padded, and anything you couldn't verify.

Each keyframe becomes one frame at 30 fps, so a hyperlapse plays at 30× speed. A 36-minute drive became 72 s and took 4.4 minutes to render. The hyperlapses of 218 drives, five cameras each, came to 107 GB. Cut multi-camera dashcam clips with ffmpeg and make a hyperlapse from keyframes with ffmpeg cover each step in full.

What went wrong here. On the first drive I cut, the side-left camera came out short. On other drives it’s a different camera, or none, so the prompt measures every camera. The frame-count comparison caught it twice:

1. Its GOPs run across segment files. Cutting each segment separately and joining them dropped the frames that reference the previous file: the camera came out 1.9 s short, with a jump at the join. One concat-demuxer pass over the source segments keeps them.
2. It was still 29 frames short, because its keyframes fall at n + 0.967 s, not on whole seconds like the other cameras’. Landing each camera’s in-point exactly on its own keyframe makes every file start at time 0. An offset camera starts a few frames early, and setting that clip’s In point that many frames in lines it up with the others.

A third check, on start_time, caught a version that had the right length but started 0.967 s late, which an editor might ignore.

The hyperlapse test drive caught two more. Two cameras came out at 1,303 frames against the front camera’s 2,170. The first guess was the black filler’s pixel format; the real cause was the rear camera stopping 33 s into a segment. The fix is to pad every segment to the front camera’s length. The filler has a trap of its own: it has to be full range like the footage, which takes -color_range pc; Apple’s hardware encoder ignores -pix_fmt yuvj420p. A mismatched part doesn’t make the stream-copy join fail. The joined file takes the first part’s label, so a camera whose file opens on filler was labeled limited range throughout: 11 of more than a thousand hyperlapse files, found and re-rendered later. Across 233 drive folders, six hyperlapse sets needed black-filled segments.

On that same drive, the rear and side-left cameras had restarted into a sibling folder whose name was 34 s later. The name was wrong by more than a second. Correlating motion between the Gear Guard camera and the restarted rear camera, which both face backward, measured 32.7 s ± 0.1 s across four independent windows. Front against rear disagreed by 0.9 s between windows, because those two cameras don’t see the same scene. Rivian dashcam camera missing from a drive covers the restart.

Then the batch failed, and no check caught it. Background jobs in the session stopped at a time limit, so the run was relaunched detached. A detached process sends no notice when it ends. It crashed at 10:11 on one of the empty folders the moves had left behind, and it sat dead until 16:48, when I asked what was next: more than six and a half hours. Two fixes came out of it, and both are in the prompt: an empty or bad drive is now a logged skip, and a long run gets a watcher.

Keep location out of anything you publish

Paste, for each file you mean to share:

Prepare <export-file> for publishing.

Rules:
- Treat GPS in any clip as private. List every metadata field that holds a location, with: exiftool -G1 -a "-gps*" "-location*" <file>
- First run that command on an original clip from the truck and confirm that it finds the location. Then run it on the export.
- If the export carries a location, write a copy without it, run the same command again, and show me the empty result.
- List the first and last frames of every clip, and flag any frame from either end of a drive, any license plate, and any face. Do not decide for me what to blur or cut.
- Also list creation dates, and any drive date or time in file names. Strip them with -map_metadata -1 and rename the files, unless I say otherwise.
- List what this check can't see, such as a street sign or a house number in the picture.

Run the check on every export rather than reasoning about which files carry a tag. In this work the answer differed by tool. The raw clips carry it. The hyperlapse tool I used copies the first source file’s metadata, GPS included, so the set stays traceable; the published hyperlapse script strips it. Clips cut with the concat demuxer came out without it, but a plain -c copy of a single file keeps it, and -map_metadata -1 removes it. Show only public landmarks on screen, and never frames from either end of a drive.

Detectors that didn’t work

Four ideas failed, and each failure says something about the signal rather than the threshold:

• Parked footage. The hypothesis was that recording while parked padded every drive. The measurement found 23 GB of 1,760 GB: with Gear Guard turned off, the truck stops recording seconds after parking. The scan still paid for itself, because its per-second signals fed every detector after it.
• Hard stops. The detector returned 875 candidates, and all 12 at the top were turns into parking lots or ordinary stops. Turning changes the picture as much as braking does.
• Traffic jams. Crawling traffic still changes the picture, because the neighboring cars move.
• Garage boundaries by color. An overcast sky is as gray as a concrete ceiling. It worked on a clear day and failed on a cloudy one, which is the case to test first.

One detector worked with a person in the loop. Close passes, read as a one-sided spike on a side camera while the front camera shows steady driving, were swamped by overpass shadows until the frame’s overall brightness change was subtracted. Of the top 60, 37 were real passes, and a contact sheet sorted them. A human pass was cheaper than a better detector.

Failures and the checks that caught them

Each stage in depth

Each later part of this series takes one stage and stands on its own:

2. Read the GPS location from Rivian dashcam video with ffprobe: the tag’s spellings, a parser, and how to strip it.
3. Rivian dashcam camera missing from a drive: it restarted in another folder: finding the sibling folder and measuring its real offset.
4. Cut multi-camera dashcam clips with ffmpeg without re-encoding, and keep them in sync: snapping each camera to its own keyframe.
5. Make a hyperlapse from keyframes with ffmpeg without decoding every frame: the 30× render, the padding, and the full-range filler.
6. Finding a bridge in dashcam video by the color of tunnel lights: the detector, the review sheet, and what didn’t work.

TL;DR

Blogger’s own export is one XML file. The v3 API gives you every published post as a record, with its labels, dates, URL and HTML body, and the script below writes each one to disk as JSON:

$ python export_blogger_posts.py --blog-url https://example.blogspot.com --write-html

The first run opens a browser to authorize a read-only scope and caches the token; every run after that is silent. It needs a Desktop OAuth client from Google Cloud, which takes the loopback redirect without any setup. A web-type client is what stops most first runs: it needs http://localhost:8080/ in its authorized redirect URIs, or the consent screen ends in redirect_uri_mismatch.

I started a blog on Blogger in 2001 and wrote there until 2012. Sixty-nine posts, mostly VBScript and SQL for Systems Management Server, and I wanted them as files before doing anything else with them. One record per post, the HTML body separate from the metadata, in a shape a script can read. Blogger’s Settings page offers a single Atom XML file of the whole blog. The API offers the posts.

1. Create an OAuth client

The Blogger API needs a Google Cloud project with the Blogger API enabled and an OAuth client of type Desktop. A web client works too, only if http://localhost:8080/ is in its authorized redirect URIs; a Desktop client allows loopback redirects without listing them. Download the client JSON.

Two settings on the consent screen matter for a personal tool. The app can stay in testing, and your own Google account has to be listed as a test user, or sign-in fails with access_denied before the consent screen. An app in testing gets short-lived refresh tokens (opens in a new tab), so a later run opens the browser again; the script treats a failed refresh as a first run.

2. Install the libraries

$ python3 -m venv .venv && source .venv/bin/activate
$ python -m pip install google-api-python-client google-auth-oauthlib python-dotenv

python-dotenv is optional; the script loads a .env when the module is present and reads the environment either way.

3. The Python script

#!/usr/bin/env python3
"""Export every published post from a Blogger blog as one JSON file per post."""

from __future__ import annotations

import argparse
import json
import os
import pathlib
import sys
from typing import Any

from google.auth.exceptions import RefreshError
from google.auth.transport.requests import Request
from google.oauth2.credentials import Credentials
from google_auth_oauthlib.flow import InstalledAppFlow
from googleapiclient.discovery import build
from googleapiclient.errors import HttpError

SCOPES = ["https://www.googleapis.com/auth/blogger.readonly"]


def parse_args() -> argparse.Namespace:
    try:
        from dotenv import load_dotenv
        load_dotenv()
    except ImportError:
        pass
    parser = argparse.ArgumentParser(description="Export Blogger posts to JSON files.")
    parser.add_argument(
        "--client-secret-file",
        default=os.getenv("BLOGGER_CLIENT_SECRET_FILE", "client_secret.json"),
    )
    parser.add_argument(
        "--token-file",
        default=os.getenv("BLOGGER_TOKEN_FILE", ".blogger_token.json"),
    )
    parser.add_argument("--blog-id", default=os.getenv("BLOGGER_BLOG_ID"))
    parser.add_argument("--blog-url", default=os.getenv("BLOGGER_BLOG_URL"))
    parser.add_argument(
        "--output-dir",
        default=os.getenv("BLOGGER_OUTPUT_DIR", "exported_posts"),
    )
    parser.add_argument(
        "--auth-port",
        type=int,
        default=int(os.getenv("BLOGGER_AUTH_PORT", "8080")),
        help="Loopback port for the OAuth redirect; fixed so it is authorized once.",
    )
    parser.add_argument(
        "--write-html",
        action="store_true",
        help="Also write each post's HTML body as a .html file.",
    )
    return parser.parse_args()


def get_credentials(
    client_secret: pathlib.Path, token_path: pathlib.Path, auth_port: int
) -> Credentials:
    creds = None
    if token_path.is_file():
        # Owner-only on macOS and Linux, including a token left by an older,
        # looser run. Windows ignores the mode; the file keeps its folder's.
        token_path.chmod(0o600)
        creds = Credentials.from_authorized_user_file(token_path, SCOPES)
    if creds and creds.expired and creds.refresh_token:
        try:
            creds.refresh(Request())
        except RefreshError:
            # An app in testing expires its refresh tokens; consent again.
            creds = None
    if not creds or not creds.valid:
        flow = InstalledAppFlow.from_client_secrets_file(client_secret, SCOPES)
        creds = flow.run_local_server(port=auth_port)
        # A new token file is created owner-only rather than tightened after.
        fd = os.open(token_path, os.O_WRONLY | os.O_CREAT | os.O_TRUNC, 0o600)
        with os.fdopen(fd, "w", encoding="utf-8") as token_file:
            token_file.write(creds.to_json())
    return creds


def resolve_blog_id(service: Any, blog_id: str | None, blog_url: str | None) -> str:
    if blog_id:
        return blog_id
    if not blog_url:
        raise ValueError("Either BLOGGER_BLOG_ID or BLOGGER_BLOG_URL must be provided.")
    return service.blogs().getByUrl(url=blog_url).execute()["id"]


def fetch_posts(service: Any, blog_id: str) -> list[dict[str, Any]]:
    posts: list[dict[str, Any]] = []
    params: dict[str, Any] = {
        "blogId": blog_id,
        "maxResults": 500,
        "fetchBodies": True,
        "status": ["LIVE"],
    }
    page_token = None
    while True:
        if page_token:
            params["pageToken"] = page_token
        response = service.posts().list(**params).execute()
        posts.extend(response.get("items", []))
        page_token = response.get("nextPageToken")
        if not page_token:
            return posts


def slugify(text: str) -> str:
    slug = "".join(ch.lower() if ch.isalnum() else "-" for ch in text)
    return "-".join(filter(None, slug.split("-"))) or "untitled"


def write_posts(
    posts: list[dict[str, Any]], output_dir: pathlib.Path, write_html: bool
) -> None:
    output_dir.mkdir(parents=True, exist_ok=True)
    for post in posts:
        title = post.get("title") or "untitled"
        published = (post.get("published") or "")[:10]
        parts = [published, slugify(title), post.get("id")]
        base_name = "-".join(part for part in parts if part)
        keys = (
            "id", "status", "title", "url", "published",
            "updated", "labels", "author", "content",
        )
        record = {key: post.get(key) for key in keys}
        (output_dir / f"{base_name}.json").write_text(
            json.dumps(record, indent=2, ensure_ascii=False), encoding="utf-8"
        )
        if write_html and post.get("content"):
            (output_dir / f"{base_name}.html").write_text(
                post["content"], encoding="utf-8"
            )


def main() -> None:
    args = parse_args()
    client_secret = pathlib.Path(args.client_secret_file)
    if not client_secret.exists():
        sys.exit(
            f"Client secret file not found at {client_secret}; "
            "set BLOGGER_CLIENT_SECRET_FILE."
        )
    try:
        creds = get_credentials(
            client_secret, pathlib.Path(args.token_file), args.auth_port
        )
        service = build("blogger", "v3", credentials=creds)
        blog_id = resolve_blog_id(service, args.blog_id, args.blog_url)
        posts = fetch_posts(service, blog_id)
        write_posts(posts, pathlib.Path(args.output_dir), args.write_html)
        print(f"Exported {len(posts)} posts to {args.output_dir}")
    except HttpError as err:
        sys.exit(f"Blogger API error: {err}")


if __name__ == "__main__":
    main()

Three decisions in it are worth knowing before you run it.

The redirect port is explicit. run_local_server listens on 8080 unless told otherwise, and a web-type client rejects a redirect it was not told about. The flag keeps the port in sight, so the URI you authorize on a web client and the one the script uses can’t drift apart. A Desktop client does not need the entry.

The scope is read-only. blogger.readonly is all an export needs. The token this caches cannot post, edit or delete, so a leaked .blogger_token.json is a smaller problem than it could be. On macOS and Linux the script also makes the file readable by you alone. On Windows it takes the permissions of the folder it’s in, because Python’s os.chmod (opens in a new tab) can set only the read-only flag there, so keep it in a folder only you can read.

The blog is found by URL. blogs().getByUrl turns the address you know into the numeric ID the rest of the API takes. Nothing numeric has to be copied out of the Blogger dashboard. Pass --blog-id if you have it already.

4. Run it

$ export BLOGGER_CLIENT_SECRET_FILE=client_secret.json
$ python export_blogger_posts.py --blog-url https://example.blogspot.com --write-html

Exported 69 posts to exported_posts

The first run opens a browser for consent and writes .blogger_token.json. Each post becomes YYYY-MM-DD-<title-slug>-<postId>.json, and with --write-html a .html beside it holding the body alone:

exported_posts/
  2001-08-15-sql-to-return-identity-when-inserting-2676633582946536328.json
  2001-08-15-sql-to-return-identity-when-inserting-2676633582946536328.html
  ...

Each record carries what the API returns for the post, with the body last:

{
  "id": "4298533488206221767",
  "status": null,
  "title": "Reading the SCCM Client GUID",
  "url": "http://example.blogspot.com/2012/03/reading-sccm-client-guid.html",
  "published": "2012-03-27T15:43:00-07:00",
  "updated": "2012-04-30T12:00:20-07:00",
  "labels": ["PowerShell", "SCCM", "VBScript", "WMI"],
  "author": { "displayName": "..." },
  "content": "<p>...</p>"
}

status came back null on every post in this export. The API sets it only for admin-level requests, and the script doesn’t make one.

Common failures

redirect_uri_mismatch on the consent screen

The client is web-type and http://localhost:8080/ isn’t in its authorized redirects; the trailing slash is part of the match. Add it, or recreate the client as Desktop.

access_denied before the consent screen

The app is in testing and your account isn’t a test user. Add it under the OAuth consent screen.

Fewer posts than expected

The script asks for 500 posts a page and follows nextPageToken until the API stops sending one. It asks for published posts only, so drafts and scheduled posts aren’t in the count.

Old files after a second export

The script writes into the output folder and never removes anything. A post renamed since the last run leaves its old file beside the new one, and an unpublished post keeps its file. Export into an empty folder, with --output-dir, when the folder has to match the blog.

Client secret file not found

The path is read from BLOGGER_CLIENT_SECRET_FILE or --client-secret-file and is relative to the directory you run from, not the script’s.