devlaunch

A streamlined CLI for devpod with intuitive autocomplete and a built-in fuzzy selector.

dl owner/repo@branch is the whole interface: it clones the repo if it has to, builds or starts the devcontainer, and drops you into a shell inside it. Every branch gets its own container, and launching one again attaches to what is already there.

Start here: Features · Installation · Usage · Workspace Commands · Global Commands · aid · A terminal beside the agent · GitHub auth · Tools in every workspace · Shell completion

Reference: Options · Workspace IDs · Cleaning up · pixi cache · Worktree backend · Launch timing · Development

Features

• Fuzzy Selection: When called without arguments — or with a verb and no workspace — opens a built-in fuzzy selector; nothing to install, and no fzf on PATH to find
• Smart Completion: Tab completion for workspaces, GitHub repos (owner/repo format), and paths
• GitHub Shorthand: Use owner/repo instead of full URLs - automatically expands to github.com/owner/repo
• Branch Support: Specify branches with owner/repo@branch syntax
• Fast Autocomplete: Completion cache for ~3ms response time (vs ~700ms without cache)
• One Round-Trip Per Question: every devpod call costs ~0.45s, far more than dl itself, so a command reads the workspace list at most once — and everything a container needs on the way in (naming it, then the tools probe) rides one setup pass, so an interactive dl <ws> and a one-shot dl <ws> -- <cmd> cost the same trips

Installation

Pixi (Recommended)

pixi global install --channel conda-forge --channel https://prefix.dev/blooop devlaunch

This installs devlaunch along with devpod and all dependencies automatically.

Pip

pip install devlaunch

dl and aid are compiled binaries: the wheel is linux-64 and needs a glibc no older than 2.28 (Ubuntu 20.04, RHEL 8), and once installed it needs no Python at all — pip is only how it gets onto PATH. The conda package above has no such floor, so prefer it on an older distribution.

Note: When using pip, you must install devpod separately. If devpod is not on PATH, every command that needs it prints a single install hint on stderr and exits 127 (the shell's "command not found" code). dl --help and dl --version keep working without it.

A devpod that is installed but cannot answer is a different failure and gets a different exit code. If devpod list exits non-zero, or prints something that is not a --output json workspace listing, dl quotes what devpod said on stderr and exits 1 rather than reporting that you have no workspaces — so dl --purge stops instead of deleting caches it never checked. Shell completion is the deliberate exception: dl --install, dl --refresh and dl --completion-data log the failure and carry on with the repos and branches they can still discover on local disk, so an unreachable devpod costs you workspace-name completion and nothing more.

Shell Completions

After installation, set up shell completions for dl and aid:

dl --install
source ~/.bashrc  # or restart your terminal

Usage

dl                               # Interactive workspace selector
dl <user/repo>                   # Start workspace and attach shell
dl <user/repo> <cmd>             # Run workspace command (stop, code, etc.)
dl <cmd> <user/repo>             # The same, verb first
dl <cmd>                         # The verb, on a workspace you pick interactively
dl <user/repo> -- <command>      # Run shell command in workspace

The selector is built in — no fzf on PATH and no iterfzf, which is why there is nothing to install for it and why dl with its input redirected away from a terminal simply declines to open one. A reserved verb wins over a workspace name of the same spelling: dl stop opens the selector to stop something, it does not look for a workspace called stop.

Examples

dl                               # Select a workspace interactively
dl myproject                     # Open an existing workspace by name
dl loft-sh/devpod                # Create from a GitHub repo
dl blooop/devlaunch@main         # Create from a specific branch
dl ./my-project                  # Create from a local path
dl blooop/devlaunch code         # Open in VS Code
dl blooop/devlaunch -- make test # Run a command in the workspace
dl blooop/devlaunch stop         # Stop the workspace

A name that is not owner/repo, a URL or a path is looked up among the workspaces you already have.

Commands that need a terminal

dl <ws> -- <command> gives the command a terminal whenever dl itself has one, so interactive programs — a coding agent, htop, git rebase -i, a REPL — start and stay up instead of exiting immediately. Redirect the output and the terminal goes away again, so dl <ws> -- ls > files.txt stays free of escape sequences.

This needs the ssh host alias devpod up writes to ~/.ssh/config. If a workspace has none, dl says so and falls back to the plain devpod ssh transport, which has no terminal; dl <ws> restart republishes the alias. Set DEVLAUNCH_NO_TTY=1 to force the fallback everywhere.

Workspace Commands

Every verb in that table also takes the workspace second — dl stop <user/repo> — and with no workspace at all it opens the selector and applies itself to what you pick. stop and rm answer to --stop and --rm as well, since the flag spellings were documented long before they worked.

--stop and --rm can be appended to a line that says something else

The flag spellings do one thing the words cannot: they may be typed at the end of a line that already asked for something, and they win over it.

aid kinisi/repo@fix/x 'review this pr'      # work happens
aid kinisi/repo@fix/x 'review this pr' --rm --force   # ↑, with `--rm --force` appended

Both dl and aid accept it, and a leading verb word on the recalled line is not mistaken for the workspace, so dl prune <ws> --force recalled with --rm appended still removes <ws>. This exists because a shell makes appending to the previous line cheap and rewriting the front of it expensive — deleting the workspace you were just working in should not cost an edit in the middle of a long prompt.

What the suffix beat is named on stderr before anything is removed:

--rm overrode the rest of the line: 'review this pr' was not acted on.

That notice is the price of the convenience and is not optional: the line is now allowed to carry an instruction it will not carry out, so a deliberate --rm and a slip have to be told apart. Note that a -- command tail cannot be overridden this way — everything after -- belongs to the workspace's command, so a --rm typed there is an argument to that command and not a verb.

prune is no longer a spelling of the rm verb

dl <ws> prune used to delete one workspace and dl --prune removes clone directories and no workspace at all — one word, two unrelated commands, told apart by two dashes. Reach for the wrong one and you either lose a workspace you meant to keep or get refused for a reason the message could not explain (--prune takes no workspace: it is not a workspace command.). So the verb spelling is gone, and typing it says what to use instead:

$ dl <ws> prune
'prune' is no longer a workspace verb. Use 'dl <workspace> rm' to delete a workspace,
or 'dl --prune' to remove the clone directories no workspace opens any more.

dl --prune is unchanged. The word is still recognised rather than forgotten, so it is never read as a workspace name: a dl prune <ws> --force line recalled with --rm appended still removes <ws>, and a workspace that really is called prune is still reachable as dl stop prune. Use dl <ws> rm from now on.

Global Commands

$ dl --version
dl 0.1.0

The version and nothing else. aid --version reports the same version under its own name — the two binaries are built from one cargo package and cannot disagree about it. (Through 0.0.29 an editable install appended (dev, editable from <tree>), read out of the installed package's PEP 610 metadata. A compiled binary has no such metadata and no editable installs, so there is nothing left to say.)

aid: start a coding agent in a workspace

aid is dl with a coding agent started for you:

aid <user/repo>[@branch] [prompt...]   # Open the workspace, start the agent

It is a shortcut, not a second launcher. aid rewrites its command line into a dl one and hands it to dl itself, so

aid blooop/devlaunch@fix/42 fix the flaky test

is exactly

dl blooop/devlaunch@fix/42 -- IS_SANDBOX=1 claude --dangerously-skip-permissions 'fix the flaky test'

That means an aid workspace is the dl workspace: same clone, same workspace id, same container — started if stopped, attached to if already running, and never rebuilt just because aid asked for it. Anything dl learns, aid gets.

claude is started with --dangerously-skip-permissions. The agent is already inside a disposable container holding only this repo, so the per-tool prompts it would ask on the host protect nothing here and would stall an unattended run. IS_SANDBOX=1 rides along because claude otherwise refuses that flag outright under uid 0, and devcontainers that run as root are ordinary. The variable is scoped to the agent process, not exported into your shell.

The trade is worth stating plainly: an agent started this way edits, runs and deletes inside the container without asking. It cannot reach your host, but it can rewrite the checkout it is in, so review an aid workspace before pushing rather than treating it as a sandbox that will stop it for you. --codex and --gemini are unaffected, and dl <ws> -- claude still runs exactly what you typed.

Everything after the workspace is the prompt, flags and all, so it never needs quoting to survive aid's own parsing. Managing workspaces — listing, stopping, deleting, VS Code — stays with dl.

The agent's CLI has to be installed in the container; aid runs it there, it does not install it.

A terminal beside the agent

Every workspace dl opens also has zellij on PATH, which buys one thing the other tools do not: an agent running in a container can open a second terminal next to itself, in the same container, and you can attach to it from anywhere to watch or to type.

Nothing has to cooperate for this. It does not come from your dotfiles, it does not need an edit to any repo's devcontainer.json, and it works in images dl has never seen — the same argument the rest of "Tools in every workspace" makes, for the same reason: dl launches arbitrary repos.

Opening a pane from inside a session

From anywhere inside the container — including from a completely non-interactive command, with no terminal attached to anything:

zellij -s devlaunch action new-pane -- htop

-s <name> is the form to use and the only one worth depending on. Bare zellij action new-pane happens to work by falling back to the single running session, which stops being a single session the moment there are two of them.

devlaunch is the session name dl creates and the one to name here.

Switching the wrap on

The session an agent opens panes into has to exist first, and creating it is off by default:

DEVLAUNCH_ZELLIJ=1 dl someone/repo -- claude -p "do the thing"

With it off, no invocation changes meaning at all — that is what off means here, and it is why the switch exists rather than the behaviour simply being on.

The command runs beside the session, not inside a pane of it. That is deliberate. Putting the command in a pane would hand its stdin, stdout and exit status to zellij, and all three are things dl promises to leave alone: dl <ws> -- cmd > file has to put the command's own output in the file, and a failing command has to come back with its own status. Since zellij -s <name> action new-pane works perfectly well from a command that is in no session at all, running beside the session costs nothing and delivers the same pane.

The interactive session of a bare dl <workspace> is untouched, switched on or off. An interactive attach sends no command for the wrap to attach to — that is exactly what gets it a terminal from devpod — and giving it one would cost either the terminal or a round trip in front of every shell. You land in an ordinary login shell with zellij on PATH, so zellij attach -c devlaunch gets you the session, and any panes an agent has opened in it, whenever you want them.

There is one exception, and it is a pleasant one: if you also run with DEVLAUNCH_DOTFILES_ON_ATTACH=1, that refresh is a command, so it gets wrapped like any other and the session is already there when the shell arrives.

Existing workspaces

zellij arrives on the setup pass, which runs on every devpod up. So a workspace that predates this picks it up on its next dl <workspace> restart — a full dl <workspace> recreate also works but is not needed, because nothing here is a bind mount and mounts are the thing that only lands at container creation.

Attaching to a workspace that is already running skips devpod up and so skips this too, which is what makes the restart necessary rather than automatic.

What it costs

Almost nothing, and that was measured rather than assumed. zellij is a conda-forge package installed by pixi into the container, so it lands in the shared package cache above and every container after the first extracts rather than downloads:

It can never fail a launch. Provisioning zellij is a stage of the setup pass, so a container with no network, no pixi and no way to get either reports the stage as failed, by name, and then opens exactly as it would have. A container that ends up without zellij still works; with the wrap on, the command still runs, because the session setup is allowed to fail and the command runs regardless.

DEVLAUNCH_NO_TOOLS=1 turns this off along with the rest of tool provisioning — installing zellij is tool provisioning, where naming a container is not.

GitHub Authentication

Every workspace dl opens inherits the host's GitHub login, so gh is already authenticated inside the container and the devcontainer.json does not have to arrange anything for it. devpod forwards the ssh agent and git credentials on its own, but nothing else carries gh.

devlaunch takes the token from GH_TOKEN, GITHUB_TOKEN, or gh auth token, whichever answers first, and hands it to the container as GH_TOKEN. That reaches any image and any container user, unlike a bind-mount of ~/.config/gh, and it works whether the host keeps its token in hosts.yml or in a keyring. The token is passed to devpod through a private file and through devpod's own environment, never on a command line, so it does not appear in ps. dl installs gh itself (see Tools in every workspace), so the login has something to be spent on whatever the image ships. Check a workspace with:

dl <workspace> -- gh auth status

If no token can be found, dl warns on stderr and opens the workspace anyway rather than failing — and the warning names the config directory gh consulted, because the usual cause is a shell that scoped XDG_CONFIG_HOME somewhere gh has no login, not a host that is actually logged out.

Who gets the token

Everything running in the container does — including a postCreateCommand from a repo you did not write. dl someone/repo builds and runs that project's devcontainer with your GitHub token in its environment, and a gh auth login token usually carries repo, workflow, gist and read:org scopes. devpod already forwards the ssh agent to every workspace, so this is not a new trust boundary, but it is a wider one. Skip it for a repo you have not read:

DEVLAUNCH_NO_GH_TOKEN=1 dl someone/repo

When the token changes

dl refreshes the token on every start, so rotating it on the host is enough for any workspace that gets started or restarted afterwards. Attaching to a workspace that is already running skips that step, and the token it was given at startup stays in place — including one it was given before you set DEVLAUNCH_NO_GH_TOKEN. Run dl <workspace> restart to replace it.

Tools in every workspace

gh and claude are available in every workspace dl opens, in every kind of session — an interactive dl <workspace>, a one-shot dl <workspace> -- <command>, and aid. The repo's devcontainer.json does not have to provide them, and most do not: dl launches arbitrary repos, so a guarantee that depended on the image would not be a guarantee.

How they get there

On devpod up, at most three round trips, each one earning the next.

1. The setup pass — the only trip a ready workspace ever pays. One trip carries everything the host wants done on the way into a running container: the stages first, then the probe. Naming the container — the hostname your shell prompt shows — is the one stage today, and it costs nothing extra because the probe was paying for the trip anyway. Each stage reports ok, failed with its exit status, or not reached; one that fails stops neither the stages behind it nor the probe, and dl says which one it was.

The probe is the tail of that trip. The container reports what only it can know: whether both tools answer at all, where its claude resolves to, and where ~/.local/share/claude/versions in its own home resolves to. It reports those and names no verdict; the host reads them, so "a real claude" is defined in exactly one place. The reading is one of three:

• provisioned — gh answers on the login PATH and claude resolves to a binary the official installer put in the versions directory. Nothing else happens.
• lendable — both names answer, but that claude is a shim or a wrapper.
• absent — a tool is genuinely missing.

2. A lend, for lendable and absent. dl streams its own gh and claude into the container as a tar over the devpod ssh channel it already holds — a local pipe, no network and no download. Nothing lands outside a staging directory until both binaries have been run there once, so a container that cannot execute them (a different libc, a different architecture) is left exactly as it was.

3. The network install, for absent only. When the host had nothing to lend, or the lend was refused, pixi global installs both tools — and pixi itself first if the image has none. A lendable container never reaches this trip: it stops after the lend, or — when the host had nothing to lend — after the probe itself. A claude already answers there, and this install decides what to do with the same command -v that a shim satisfies, so the trip would install nothing.

Tools reach the PATH of a login shell through whichever of ~/.bash_profile, ~/.bash_login or ~/.profile bash actually reads — it sources only the first of those that exists, so an image shipping a ~/.bash_profile never reads ~/.profile.

An install that fails costs the workspace its tools, not its launch: dl logs a warning and hands you the session anyway.

What to bake so a launch does no work at all

To make every dl launch of an image stop at trip 1. The probe asks a login shell to resolve each name, so every bullet here is about what a login shell can find:

• gh anywhere on the login PATH.
• claude in the layout its official installer creates — the binary at ~/.local/share/claude/versions/<version>, a direct child of that directory named for the version, with ~/.local/bin/claude symlinked to it. Nested any deeper — versions/<version>/bin/claude, the shape a downloader parked there would take — is read as somebody else's tree that merely starts with the official path, and does not count.
• ~/.local/bin on the login PATH. The symlink above is how claude answers at all; a login shell that cannot find that directory reads the image as absent however carefully the rest was baked, and it pays the full lend. Ubuntu's stock ~/.profile prepends ~/.local/bin itself — but an image shipping a ~/.bash_profile never reads ~/.profile (above), and then nothing does.

Nothing else counts as a claude, and that is the point. A shim — a small launcher that downloads the real binary the first time it is called — answers command -v claude exactly as the real thing does, while the workspace still owes a multi-hundred-megabyte download at the least convenient moment. So dl resolves the name rather than running it (running a shim is the download), reads a shim as lendable, and sends the host's real binary. The lend prepends ~/.local/bin to the login PATH, which is what puts the lent binary in front of the shim from then on — intended, and the reason the next launch probes provisioned and the transfer is paid once rather than forever.

This repo's own devcontainer feature bakes a shim. .devcontainer/claude-code/install.sh installs claude-shim, so an image built from it does not meet the contract by itself: its first dl launch is lent a real claude, and only launches after that do nothing. Build the official layout into the image if you want the first launch free too.

What this deliberately does not do

• No per-tool transfer. The lend is all-or-nothing — an image with a real gh but a shimmed claude is sent both. Splitting the payload would save part of one transfer, paid once per workspace, in exchange for a matrix of half-lent states every later step would have to reason about. (The network install is already per tool: each install guards itself with its own command -v.)
• No version sync. A real claude already in the container is left alone whatever its version. dl lends what is missing; it is not a package manager, and keeping versions in step would mean deciding what to do when the container is the newer one. The official binary self-updates in a long-lived workspace, and rebuilding one re-provisions it from scratch. The single upgrade dl does perform is replacing a shim with a real binary.

Turning it off

DEVLAUNCH_NO_TOOLS=1 dl someone/repo

Attaching to a workspace that is already running skips devpod up, and so skips this too. A workspace started by something other than dl — or created before this existed — picks the tools up on its next dl <workspace> restart.

Shell Completion

After running dl --install, you get intelligent tab completion:

• Workspace names from your devpod list
• Known GitHub owners and repositories from your workspaces
• File/directory paths when starting with ./, /, or ~
• All global flags (--ls, --install, etc.) and workspace commands

How the completion cache stays current

The data behind completions lives in ~/.cache/devlaunch/completions.json, and building it means a git ls-remote per known repo — seconds of work. So it is rebuilt in the background at most once an hour (the same interval the worktree backend's background fetch sweep uses), and at most once per dl invocation. Commands that change your workspaces (starting, stopping or deleting one) rebuild it as soon as they finish, regardless of when it was last built. Commands with no use for it — dl --help, dl --version — do not touch it at all.

A branch created on a remote in the last hour may therefore not be offered yet. dl --refresh rebuilds the cache immediately and ignores the interval.

Everything above is what it takes to use dl. What follows is reference — flags, ids, disk accounting, and the internals worth knowing when something surprises you.

Options

Projects with demanding devcontainers — several variants, compose sidecars, or a host-side initializeCommand that has to tell branch workspaces apart — are covered in docs/devcontainer-projects.md.

Refreshing dotfiles on attach

devpod applies dotfiles when it provisions a workspace, so a workspace that has been up for a fortnight still has the dotfiles it was born with. dl <ws> dotfiles fixes that when you think of it; DEVLAUNCH_DOTFILES_ON_ATTACH=1 makes dl think of it for you, running the same refresh just before it hands you the shell.

DEVLAUNCH_DOTFILES_ON_ATTACH=1 dl someone/repo

It is off unless you set it, and that is the point rather than caution. The refresh is a devpod ssh round-trip — measured at ~1.7s, almost all of it connection setup — with a git pull behind it, and it would otherwise be charged to every attach on every machine to close a gap most people do not have.

Two things it deliberately does not do:

• It never runs for dl <ws> -- <command>. A one-shot command renders no prompt and sources no interactive shell, so refreshing in front of it would buy that command nothing and cost it the round-trip. That path is the one agent launchers use, and it stays exactly as fast as it was.
• It never holds the shell hostage. The refresh gets 60 seconds; an unreachable dotfiles remote, or one that wants a password nobody is there to type, means a pause and then your shell, not a hang. Failure is a warning — you get the workspace either way.

Refreshes run every time you attach, with no cooldown, because you asked for them. If that is too often, unset the variable and use dl <ws> dotfiles.

Workspace IDs

dl user/repo@branch derives one id that names both the devpod workspace (what you see in dl --ls) and the clone directory under ~/.cache/devlaunch/repos/:

<repo-slug>-<branch-slug>-<syllables>      at most 47 characters

blooop/devlaunch@main                             -> devlaunch-main-zovomobo
blooop/devlaunch@feature/auth                     -> devlaunch-feature-auth-poliseno
blooop/devlaunch@feature-auth                     -> devlaunch-feature-auth-nesatabe
blooop/test_renv@nb4                              -> test-renv-nb4-polenita
kinisi-robotics/kinisi_ros@ags-devcontainer-tooling-support
                                                  -> kinisi-ros-ags-devcontainer-tooling-su-lenevere
blooop/devlaunch@dependabot/github_actions/codecov/codecov-action-6
                                                  -> devlaunch-dependabot-codecov-action-6-sifivasa

The eight-character syllable suffix is a hash of the full (owner, repo, branch) triple. It is what makes the id unique: the readable part is shortened to fit the length limit, and shortening it does not affect whether two branches share an id. Long branch names drop whole /-separated middle segments before losing characters, so the part that identifies the branch survives. Note the third and fourth lines above: feature/auth and feature-auth read the same once slugged but are different branches, and they get different ids.

Owner and repo are matched case-insensitively, the way GitHub treats them, so dl NVIDIA/cuda-samples@main and dl nvidia/cuda-samples@main are the same workspace. Branch names are case-sensitive, because git refs are.

URL specs (dl github.com/owner/repo) get an id in the same shape, with the suffix hashed over the URL.

The id is also the container hostname. 47 characters is one inside devpod's own hard ceiling of 48 — a 49-character id is refused outright, not truncated — and well inside the 64-byte hostname limit on its own, but tools that stack their own prefixes onto the container name have about 17 characters to work with, so a tool that wants more is the one that has to shorten.

Branch names must be safe as both git refs and directory names — a name with a space or a leading dash is rejected rather than quietly rewritten.

Upgrading from an older devlaunch

This id format is new, and the directories and containers on your machine were named by the previous scheme. The first dl user/repo… command after upgrading migrates the cache once, and leaves what it did behind in the cache directory (the two listings named below). dl --help, dl --version, dl --ls and opening an existing workspace by name do not trigger it.

Your clone directories are renamed. What was ~/.cache/devlaunch/repos/blooop/devlaunch/main becomes ~/.cache/devlaunch/repos/blooop/devlaunch/devlaunch-main-zovomobo. A workspace is a git clone whose origin points at the .bare cache next to it, and .bare does not move, so this is a plain rename: branches, history and uncommitted changes all survive — only the folder name changes. metadata.json is updated in the same pass, so nothing is left pointing at the old name.

Your existing devpod containers keep their old ids and are orphaned — and they can often be repaired rather than replaced. An orphaned container is sourced at the path this migration just renamed, with the real clone sitting next to it under the new name, which is precisely what dl --reconcile is for: it re-points devpod's record at the renamed clone, and dl <workspace> recreate finishes the repair. That gives you back the clone association and the workspace's identity — not state that lived only inside the old container, which nothing can bring back. The repair is order-dependent: relaunching the branch claims the renamed clone for a fresh container, and reconcile never re-points a clone a live container holds — so reconcile first, then relaunch. Left alone, the next dl user/repo@branch simply builds a fresh container under the new id, and deleting the old one is all that remains for it.

dl does not delete containers for you — deleting by id is how a running sidecar got destroyed the last time something tried (kinisi_ros#9766) — so it writes the old ids to ~/.cache/devlaunch/orphaned-workspaces.txt. For the workspaces you are finished with, the disposal command reads from that listing:

xargs -r -n1 devpod delete < ~/.cache/devlaunch/orphaned-workspaces.txt

A clone directory with no metadata record is left alone. Nothing records which branch it was cloned for, and the old directory name cannot be turned back into one — feature/auth and feature-auth both became feature-auth — so a guessed name would be worse than no rename. Those directories stay exactly where they are and are listed in ~/.cache/devlaunch/unmigrated-clones.txt.

Running dl again changes nothing: the migration is keyed on the version field in metadata.json, not on directory names, so a branch that happens to look like a new-scheme id is never mistaken for one. If a migration is interrupted, the next run finishes it — the version is written last, in the same atomic save as the new paths, so it never claims more than the filesystem has actually done. A rename the filesystem refuses (a read-only mount, tightened permissions) is treated the same way: the version stays put and every later run retries the refused directories and repeats the notice until the underlying refusal is fixed by hand.

Cleaning up: purge, prune, reconcile

The three global commands that touch what is already on disk, in full. --prune takes clone directories and no workspaces, --purge takes both but only what devlaunch made, and --reconcile removes nothing — it repairs records that stopped matching the disk. Each one prints its plan and asks before acting, and -y is what skips the question.

One exception, and it is not a released build: a binary compiled from somebody's working tree appends -dev (dl <version>-dev). That comes from the dev-build cargo feature, which ./dev.sh builds with and nothing that ships enables, and it is what tells dl-next apart from dl when both are on PATH — see "Two installs" in AGENTS.md.

What purge deletes

devpod's workspace list is shared. A workspace you made with devpod up, or that another tool made, sits in the same list as the ones dl made, and dl --purge has no business destroying it. So it deletes only the workspaces devlaunch created — the clones it made under its own cache directory ($XDG_CACHE_HOME or ~/.cache, then devlaunch/repos/<owner>/<repo>/<id>), which is exactly the directory the purge is about to remove anyway. Everything else keeps working afterwards, because nothing a purge touches backs it.

Anything it is leaving is named before it asks:

$ dl --purge
This will remove all devlaunch data:
  - 4 DevPod workspace(s)
  - /home/you/.cache/devlaunch/ (workspace clones, repo caches, the shared pixi cache, completions)

Leaving 2 workspace(s) devlaunch did not create:
  - pythontemplate
  - my-hand-made-workspace

Are you sure? [y/N]

Three things dl does create are in that second list rather than the first. dl ./some/path and dl <git-url> open a source dl did not clone, so it cannot tell them from a workspace you made by hand — and a config.toml that points repos_dir outside the cache puts the clones somewhere --purge does not remove either, so those are left too. Delete any of them with dl <workspace> rm. Erring this way is deliberate — a purge that skips one of your own workspaces costs you a command, and the other kind of mistake costs you work you cannot get back.

When part of the cache will not go

A container writes into its clone as its own user — vscode, uid 1000, in the standard devcontainer base image. Where your host user is uid 1000 too, nothing here comes up. Where it is not — CI, a shared machine, a container running as root, or devlaunch developed inside its own devcontainer — the directories the container made cannot be emptied by you, and the purge cannot remove them.

It removes everything else anyway, and names what is left:

$ dl --purge -y
Removed what was permitted under /home/you/.cache/devlaunch. These refused:
  - /home/you/.cache/devlaunch/repos/blooop/bencher/bencher-main-kivagede: Permission denied

Usually this means a container wrote them as a different user, and:
  sudo rm -rf '/home/you/.cache/devlaunch'
clears them. Check the reasons above first -- it does not fix all of them.
devlaunch does not manage Docker images or volumes: the containers these workspaces used may still hold disk, and `docker system df` shows what Docker is holding.

That last line ends every purge, including one that found nothing to purge and one you answered n to — dl --prune ends on the same one, in the same words. See the disk neither command frees.

Exit status is 1, because a clone you were told would go is still on disk. It used to be 1 with the whole cache still standing: the first refusal stopped the purge, so the completion caches, metadata.json and every other clone survived on account of one directory.

When none of it goes — nothing under the cache came away at all, which is what a symlinked cache root gives you, or one that cannot even be looked at, or one whose every entry refused — the headline says that instead of claiming a partial success:

$ dl --purge -y
Removed nothing under /home/you/.cache/devlaunch. These refused:
  - /home/you/.cache/devlaunch: Permission denied

The report underneath is the same one, and so is the exit status: 0 means the cache is gone and nothing else does, which is the only distinction a script can act on. Removed everything, removed what it was permitted to and removed nothing are three outcomes rather than two, and the sentence is where the third one lives — because it is the one that decides whether you still have clones to go and look for.

What is listed is the directory, once — not the hundreds of files inside it. Unlinking needs write permission on the directory rather than on the file, so every entry in that clone refuses separately and they are all the same fact. Two separately unwritable directories on one path are two lines, though, because clearing the inner one would leave the outer one just as stuck.

Each line carries what the system actually said. A container running as another user is the common cause, but a read-only mount, chattr +i and a busy mountpoint all land here too — and sudo rm -rf does not fix those, which is why the report offers the cause rather than asserting it.

If you have moved your cache by making ~/.cache/devlaunch a symlink, a purge refuses it and names the target rather than following it. Remove the real directory yourself if you meant to: following the link would empty a directory you never named, and removing just the link would report a clean sweep while your clones sat on the other volume.

Pruning the clones nothing opens

A workspace per branch means clone directories accumulate under the cache, and until now nothing removed them: measured on one host, 52 clone directories for 17 live devpod workspaces — 37 of them attached to nothing, 4.00 GB, against 7.86 GB still in use. --purge is the wrong tool for that, being all-or-nothing: the only way to get the 4 GB back was to destroy the 7.86 GB too, and every bare cache with it.

dl --prune removes exactly the clone directories no live workspace opens. It never deletes a devpod workspace, a container, an image or a volume, never touches a repo's .bare cache (0.08 GB for seven repos, and it is what makes the next clone of a repo fast), and never looks outside <cache>/devlaunch/repos. Every directory it finds is one of three things:

• a live workspace opens it — kept, and named with the workspace that has it. "Opens" means at or under: a workspace opened on a subdirectory of a clone still needs the clone;
• nothing opens it — removed, unless it holds work that exists nowhere else, or git would not say what it holds. A clone a container wrote as another user is unreadable rather than empty, and "cannot tell" is kept, not removed;
• dl's records and devpod's disagree about it — kept, always. This is #88's shape. On that ticket's host, 36 devpod workspaces out of 39 recorded a source folder that was gone or was a config-only stub, while the real checkout sat beside it under a newer naming scheme — so a perfectly healthy clone was opened by nobody, and the stub was the only thing anything pointed at. --prune will not guess which clone such a workspace needs: it keeps every clone of that repository and names the record to go and fix. --force does not move any of them. dl --reconcile is what fixes them.

Note that every directory two levels under <cache>/devlaunch/repos is a candidate — a stray directory somebody left there is looked at like any other. The cache is dl's to manage; things that are not clones do not belong in it. But git cannot say what a directory that is not a repository holds, and "cannot say" is kept rather than removed, so clearing junk out of the cache takes --force. That is the same refusal a clone with a half-written .git gets, and deliberately so: telling the two apart would mean --prune forming its own opinion about a directory dl <workspace> rm already refuses on.

$ dl --prune
Clone directories under /home/you/.cache/devlaunch/repos:

Removing 2 that nothing references -- 1.4 GiB:
  - /home/you/.cache/devlaunch/repos/blooop/bencher/bencher-test1-pipagito (1.1 GiB)
  - /home/you/.cache/devlaunch/repos/blooop/devlaunch/devlaunch-t1-vebilote (317.0 MiB)

Leaving 3:
  - /home/you/.cache/devlaunch/repos/blooop/devlaunch/devlaunch-main-zovomobo: workspace devlaunch-main-zovomobo still opens it
  - /home/you/.cache/devlaunch/repos/blooop/wayfinder/wayfinder-devlaunch-kilarabo: holds 2 unpushed commit(s) -- add --force to remove it anyway
  - /home/you/.cache/devlaunch/repos/blooop/rockerc/rockerc-main-ludomane: devpod lists workspace rockerc-main-ludomane and sources it at /home/you/.cache/devlaunch/repos/blooop/rockerc/main; see devlaunch#88

Dropping 12 record(s) of directories already gone.

Are you sure? [y/N]

-y skips the question. A clone holding uncommitted or unpushed work is kept and named, in the same words dl <workspace> rm refuses in — 13 of those 37 stale clones did, two of them with real unpushed commits, so this is load-bearing rather than a formality. --force promotes that one case and nothing else. Erring this way costs you a flag; erring the other way costs work that cannot be recovered.

The sizes are the same exclusive bytes dl --ls --size reports, and they mean the same thing: what removing that directory would actually free, not what du would print. Where a walk could not read something the figure reads ≥ and so does the total, because a floor printed as a total is a cleanup tool telling you a directory is small when it is not.

Directories that will not come away are named the same way a purge names them, the rest still go, and the exit status is 1.

Nothing here runs on its own. A full scan measured 1017 ms on that host — about two warm launches — and it gets slower exactly as the cache gets fuller, so it is never on a launch path and never folded into dl --ls. Answering n is the read-only view; there is no separate flag for it. It costs one devpod list to build the plan and no devpod status at all, because whether a workspace is running has no bearing on whether a directory is opened by one. A run you say yes to pays a second devpod list before it removes anything, and classifies every directory again: a launch that finishes while the report is on screen registers a workspace for one of the directories in the plan, and that is the one thing the plan cannot be re-checked against from disk. The set you approved can shrink between the report and the act. It can never grow.

It also drops the metadata.json records of directories that are already gone. That file was append-only in practice — 49 records for 17 live workspaces on the same host — and this is the first thing that prunes it.

The disk neither command frees

Both commands end on the same line, in the same words:

$ dl --prune -y
...
Removed 2 clone director(ies) -- 1.4 GiB.
devlaunch does not manage Docker images or volumes: the containers these workspaces used may still hold disk, and `docker system df` shows what Docker is holding.

The gigabytes a cleanup reports are usually not the ones you are looking for. On the host this was measured, --prune had 4.00 GB of stale clones to give back while docker system df read 86.5 GB of reclaimable images, 43.18 GB of volumes and 13.88 GB of build cache — an order of magnitude more, sitting behind a command that had just said "Removed". Saying nothing is what makes a freed figure read as all of it, so both commands say this instead, whether they removed 40 clones, found nothing to remove, or were answered n at the confirmation. The report you get for saying n is a reason to print it, not an exception: that is where somebody is deciding what is worth deleting.

It is a sentence, not a measurement. dl runs no docker command to print it, so there is nothing to be slow and nothing to fail where Docker is absent, stopped, or reachable only as another user. The figures above are this README's, from the host it was measured on, not from your machine — docker system df is where yours are.

And it points rather than offers. There is deliberately no dl flag that removes an image, and no list of image ids here to paste into docker image rm. Images devpod builds carry no devlaunch or devpod label, so any list dl printed would be a guess at which of them belong to these workspaces, and docker image prune -a is not scoped to devlaunch at all — it would take images built by everything else on the machine. Deleting them is a decision with your own containers on the other side of it, and docker system df is the tool that shows you what it costs.

Reconciling records that disagree

dl keeps its own record of every workspace, and devpod keeps one too. They agree until the naming that connects them moves — and it did move once, when workspace ids and clone-directory names gained a hashed suffix. dl's records were migrated to the new naming; devpod's were not, because nothing knew to touch them. On the host that reported it, 36 of 39 devpod workspaces recorded a source folder that was missing, or was a stub with no .git in it, while the real checkout sat next to it under the new name. Nothing was deleted and nothing was corrupted: dl was simply asking devpod about workspaces devpod had never been given, and devpod was answering correctly that there were none.

Two things fix that, and they are different jobs. dl now writes the devpod workspace id down when it creates a workspace, so the naming can move again without taking anything with it — that is automatic and needs no command. It does nothing for the records that already disagree, because they were written before there was a field to write it in. dl --reconcile is for those:

$ dl --reconcile
devpod workspaces sourced under /home/you/.cache/devlaunch/repos at something that is not a clone:

Re-pointing 2:
  - devlaunch-main: .../blooop/devlaunch/main -> .../blooop/devlaunch/devlaunch-main-zovomobo
  - bencher-test1: .../blooop/bencher/test1 -> .../blooop/bencher/bencher-test1-pipagito

Each of these needs `dl <workspace> recreate` afterwards: the container
still has the old source bind-mounted, and no record change moves a mount.

Leaving 1, which dl will not guess at:
  - rockerc-main (.../blooop/rockerc/main): no clone of that repository answers to this name

Nothing here is deleted. `dl <workspace> rm` is how one goes, if it should.

It matches the two sides by path, never by id — the id is the thing that changed, so it connects nothing, while the source folder devpod kept still names the owner and the repository exactly, and its last component still names the branch in one of the three ways dl has named a clone directory. Where that match is not unique it is refused rather than guessed: a clone a live workspace already opens — at it, or anywhere under it — is never taken from it, a clone two dead records both match is claimed by neither, and a name that two clones answer to (the old flattened spelling turned feature/auth, feature auth and feature:auth all into feature-auth) adopts neither of them. If a live workspace's source cannot be followed at all, the whole command stops the way dl --prune does, because such a workspace could be holding any of the clones on offer. Nothing is ever deleted. A workspace dl cannot match is named and left exactly where it is, because whether a workspace is finished with is not something dl can know, and the two mistakes are not the same size.

Run it as often as you like — a repaired workspace is no longer sourced at a non-checkout, so a second run finds nothing to do.

A re-pointed workspace still needs rebuilding. Its container was built with the dead path bind-mounted into it, and changing a record does not move a mount. dl <workspace> recreate is what finishes the repair, and it is the step that needs Docker.

Do not point an old dl at a reconciled cache. A dl from before the naming changed derives the old directory name, does not find it, and treats the launch as a cold one: it clones a second directory under the old name, registers a second devpod workspace, and rewrites that branch's record with the old naming and an empty workspace id — undoing the repair for that one workspace, and leaving you two clones of the branch. It is not destructive and the next dl --reconcile sorts it out, but a machine that runs both builds against one cache will keep re-breaking. Upgrade the old one, or give it its own XDG_CACHE_HOME.

Cleaning up workspaces

One workspace per branch means workspaces accumulate, and --purge is the wrong tool for tidying: it is all-or-nothing and takes the caches with it.

devlaunch does not decide which workspaces are finished. Whether a piece of work is over is a fact about a ticket, a review, or somebody's intent, and dl knows about clones and containers. Inferring it from the branch — merged into the default, or deleted from the remote — was tried and dropped: it reads like a git fact but is a guess at intent, and it cannot tell a squash-merged branch from an abandoned one. So dl supplies the two halves a tool that does know needs, and that tool drives the cleanup:

dl --ls --json          # what exists, and what each workspace holds
dl --ls --json --size   # ...and what removing each one would free
dl <workspace> rm       # remove one

The JSON reports, per workspace: id, devlaunch (did dl create it), repo, branch (what the workspace was made for), checkedOut (what its clone is on now, which can differ), path, state, lastUsed, and — the field a cleanup tool must not ignore — unsaved:

{
  "id": "devlaunch-wayfinder-devlaunch-80-ladepomi",
  "devlaunch": true,
  "repo": "blooop/devlaunch",
  "branch": "wayfinder/devlaunch-80",
  "state": "Stopped",
  "unsaved": {
    "wouldLose": "2 uncommitted change(s) (pixi.lock, notes.md) and 1 unpushed commit(s)"
  }
}

unsaved is an object with exactly one key, and the key says which of three answers it is:

The changed paths are named, not just counted, and that matters more than it looks: a devcontainer that runs a package install in its postCreateCommand can leave a tracked lockfile modified in every workspace it builds — this repo's own did, until that install became pixi install --frozen — and as a bare count that is indistinguishable from an hour of unsaved work. A cleanup tool believing the count would then never clean anything. Named, it is judgeable. A workspace dl did not create reports devlaunch: false and no unsaved — unsaved is null exactly where devlaunch is false, and nowhere else: there is no clone of dl's to protect, and it has no business inspecting your checkout. (repo and branch are a weaker test and not the same set: they come from dl's metadata record, and a clone dl owns can have lost its record while the clone and the work in it are still on disk. That clone is inspected and reported like any other.)

dl <workspace> rm refuses to delete a clone it would lose work from — when the recorded clone holds unsaved work, and when it cannot tell what that clone holds, so a caller that forgets to read the field is still caught. (Recorded, because that is the directory the guard reads; the case with no record is neither, and is described below.)

$ dl blooop/repo@feature rm
error: devlaunch-repo-feature-xyz holds 1 unpushed commit(s).
       Push or commit it, or run: dl blooop/repo@feature rm --force

$ dl blooop/repo@feature rm
error: devlaunch-repo-feature-xyz: git could not read /home/…/repo/feature:
       fatal: not a git repository. devlaunch will not delete a clone it cannot
       check. Look at it, or run: dl blooop/repo@feature rm --force

That refusal is the only judgement dl makes here, and it is not about finished work — it is dl declining to destroy the only copy of something, including when it cannot prove there is another copy. Say --force if you mean it.

--force changes one more answer: an already-absent workspace counts as deleted, like rm -f. Unforced, rm reports devpod's refusal to delete a workspace it does not have; forced, the contract is the state afterwards, not that a delete happened — which is what lets the cold benchmark's per-run reset run before the first launch, when there is nothing to remove yet.

The guard reads dl's metadata record, so the recorded directory is the one it asks about. (The delete does not always remove that same directory: when the recorded path is not on disk it falls back to a derived one. That divergence is older than this guard and is tracked as devlaunch#174.) One case is therefore neither a refusal nor a delete: a clone under dl's cache that has no record — a metadata write that failed, a record pruned, a cache restored without one. The listing still reports what that clone holds, so unsaved is the field to read; but rm removes the devpod workspace, exits 0 without asking for --force, and leaves the clone on disk, because there is no recorded directory for it to remove either. Nothing is destroyed, and nothing then points at the clone: it is yours to keep or to rm -rf by hand.

wf is the caller this was built for: it names its branches after its tickets, so it knows which workspaces belong to finished work and removes those.

How much disk a workspace costs

dl --ls --size adds a SIZE column, and dl --ls --json --size adds a disk object beside the other per-workspace facts:

$ dl --ls --size
WORKSPACE                      TYPE   SOURCE                                              SIZE  LAST USED
kinisi-ros-main-lubadaha       local  /home/…/repos/kinisi-robotics/kinisi_ros/main    64.9 MiB  2026-08-08 11:43:27
my-own-checkout                local  /home/…/projects/scratch                                -  2026-08-01 09:12:04

The number is what deleting that workspace would give back, not what du prints. Those differ, and the gap is the point of the design. A repo is cloned once into a bare cache and every workspace clone hardlinks its git objects out of that one copy, so the objects exist once on disk however many workspaces share them. A size that walked each workspace on its own — which is what du does when you point it at one directory, counting the blocks every file in it occupies — bills each workspace for the whole shared pool.

The measurement the row above comes from, taken with the shipped code on one machine (Ubuntu 24.04, ext4, warm page cache) on a real clone of that repo made by git clone from the bare in dl's own cache:

du bills that workspace 5.2x what deleting it would actually free. The difference is a single 270,823,424-byte pack file with one link in the clone and one in the bare, so removing either end frees none of it.

That sharing is a promise, not a coincidence, and a test holds it to that. git clone <path> <path> hardlinks pack files by default, and the default is all that was ever keeping it true — a file:// URL, an intermediate copy, or an explicit --no-hardlinks would each forfeit it with nothing failing and no warning printed. Measured on this repo — du -sc over the cache and each clone's .git, ext4, git 2.55.0 — that is 2400 KB for the cache plus one workspace against 4472 KB unshared, and 196 KB rather than 2268 KB of .git for every workspace after the first. So an integration test asserts the pack files are the cache's — same inode, more than one link — and that assertion goes red on all three. No clone flag is used to guard it: --local is already the default and does not even reject a file:// source, and --shared/--reference were measured to leave a workspace that fails git fsck once the cache has fetched and gc'd, for a 2 KB saving.

Sharing does erode, in one measured way that is a safety property rather than a fault: when the cache repacks, an existing workspace's pack loses its second link and becomes that workspace's own complete copy, still passing git fsck. The workspace stops being cheap and never stops being valid — which is the trade --shared and --reference get wrong, and the reason they are not used.

Large files are shared the same way, but nothing about git clone does it for you. git-lfs objects are not git objects: the clone does not carry them at all, so a workspace of an LFS repo used to download the entire payload from the forge and keep a private copy of it in .git/lfs/objects — every workspace, every time, on top of the worktree copy. dl now makes the bare cache the repo's LFS store as well: the payload is fetched once into <repo>/.bare/lfs for the branch being launched, and each workspace materializes out of that, which git-lfs does by hardlinking. Measured with git-lfs 3.7.1 on ext4: the workspace's object file is the same (st_dev, st_ino) as the cache's, so its store costs nothing, and the materialization succeeds with the remote deleted from disk — the second workspace of an LFS repo touches the network for its large files not at all. What remains per workspace is the worktree copy, which is real bytes and cannot be shared: a container build has to be able to read them. If the cache cannot supply an object — a first launch offline, a payload the branch alone introduces — the old download from origin still runs, and a workspace left holding pointer files is retried on the next launch rather than written off.

Nothing about that is written into the workspace's .git/config, and that restraint is load-bearing rather than tidy: dl bind-mounts the clone directory into the devcontainer and .bare is a sibling that is not mounted, so an lfs.storage entry or an added remote naming a host path would break every git checkout of an LFS repo inside the container while working perfectly on the host. A test asserts the clone keeps exactly one remote, still pointing at the forge, and no lfs.storage at all.

So dl counts a file only when every one of its hardlinks lies inside the workspace being measured. Two consequences, both deliberate:

• The sizes do not add up to the size of the cache. Bytes shared between workspaces belong to none of them, because deleting any one frees none of them. They become the last workspace's the moment it is the last one — which is exactly when deleting it would free them. In the table above that is the last two rows read against each other: 68,702,208 reported bytes against 353,882,112 held.
• A workspace's size can change without the workspace changing, when a sibling that was sharing with it goes away. That is the truth about shared storage.

A workspace dl did not create reads - (null in JSON): there is no clone of dl's there to measure, and walking your own project directory is not dl's to do. The table and the JSON decide that from the same rule — is the clone one dl put in its own cache, the same question --purge deletes by — so the two always name the same set of workspaces as measurable. Where a walk hits a directory it cannot read — a container writes into its clone as its own user, so this happens — the answer is a floor rather than a total: ≥2.0 MiB in the table, and {"atLeastBytes": …, "unreadable": 1} in JSON instead of {"exclusiveBytes": …}. A partial measurement never comes back looking like a complete one.

It is opt-in because it walks the whole clone. Plain dl --ls is one devpod round-trip and no filesystem work at all, and the walk is O(files) with no ceiling. Measured with the shipped code on one machine — Ubuntu 24.04, ext4, warm page cache, five runs after a warm-up, the machine otherwise busy — a real 8,309-entry clone walked in 24–28 ms, this repo's own tree with its built environment inside it (9,124 entries) in 17–21 ms, and a 114,817-entry tree in 232–239 ms. No cold-cache figure is quoted because none was taken: dropping the page cache needs root on that machine. Those are one machine's numbers on warm cache and yours will differ, but the shape is the point — it grows with the file count, and a devcontainer that builds its environment inside the clone (this repo's own does) is most of that count. That is not a bill a listing should present unasked.

Docker images and named volumes are not counted: dl did not create the layer store and does not manage volumes. docker system df is the tool that knows — the same boundary --prune and --purge name when they finish.

The shared pixi package cache

Every container dl creates gets one host directory bound into it, and PIXI_CACHE_DIR pointed at it, so that dotfiles which provision their tools with pixi global sync download each package once per machine instead of once per container:

Measured on the profile this was built for — 23 pixi-global environments — a container with a cold cache spends 62–113 s and downloads 1.2 GB; one that finds the packages already there finishes in 18–28 s and fetches nothing. Two containers syncing against it at the same time is fine: the downloads are content-addressed and rattler takes a lock per package.

Deleting it is always safe, at any moment, including while containers are running. It holds nothing but downloaded package archives — every one of them re-fetchable from the network, and none of them referenced by a path anything inside a container has stored. The worst a deletion costs is the next container's download.

rm -rf ~/.cache/devlaunch/pixi

dl --purge takes it away with the rest of ~/.cache/devlaunch/, for the same reason.

Two things it deliberately is not. It is not the host's own ~/.cache/rattler/cache: containers write into it as their own remote user, whose uid only happens to match yours, and a pixi clean cache you run for your own reasons must not be able to pull packages out from under a running container. And it is not a shared PIXI_HOME — the installed environments and their trampolines are baked with absolute paths, and two containers sharing one environment tree is pixi#5476. Only the download cache is shared, which is the part that is safe to share.

If the directory cannot be created, or is not there when the launch reaches it — a full disk, a read-only cache home, a cache swept between the two — the launch goes ahead without the mount and the container downloads its own packages, exactly as it did before this existed.

Sharing requires the container's user to be able to write the directory, which in practice means its uid matches yours or it is root. The mount carries host ownership through unchanged, and pixi does not degrade to reading a cache it cannot write: pointing PIXI_CACHE_DIR at a directory owned by another uid fails the install outright (Permission denied on the repodata, exit 1) even when every package it wants is already in there. So an image whose remote user is neither root nor your uid does not merely lose the sharing — its pixi global sync fails, and its tools do not get provisioned.

dl cannot see the container's uid before it launches, so it cannot decide this for you. In practice the common case is safe: every mainstream base declares a remote user at uid 1000, which is the first human user on a Linux host. If you hit the failure, the fixes available to you are to run that image as your own uid, or to take the cache out of play for it (rm -rf ~/.cache/devlaunch/pixi recovers a directory an earlier container left owned by someone else).

The case that is not a developer's machine is CI. What makes the common case safe is that uid 1000 is both the base image's remote user and the first human user on a Linux host — and on a hosted runner it is only the first of those. A GitHub runner's own user is somebody else, so a launch there hits this on its first container and every container after it. This repo's own launch benchmark did exactly that for twenty consecutive merges to main: failed to create directory /var/tmp/devlaunch-pixi/pkgs: Permission denied, from the benched repo's pixi install, before anything was timed.

Where you know the uid you are handing the directory to, there is a third fix the list above does not offer, and it is what .github/workflows/bench.yml now does — create the directory yourself and widen it, before the first launch:

mkdir -p ~/.cache/devlaunch/pixi && chmod 1777 ~/.cache/devlaunch/pixi

dl's own mkdir does not re-mode a directory it finds, so the mode survives every launch after it. 1777 is what /var/tmp carries at the other end of the same mount, and it makes the same trade: every uid can write, and the sticky bit means none of them can unlink another's entries.

Where the tools themselves land

The cache above is shared; the environments devlaunch installs into are not, and they are not the container's ~/.pixi either. gh, claude and zellij go into ~/.devlaunch/pixi, a pixi home of devlaunch's own, because pixi global install is not only an install — it is an edit to $PIXI_HOME/manifests/pixi-global.toml, a declarative file that in a container already has an owner. Writing there made devlaunch a second author, and cost something in both directions:

• pixi global sync removes every environment the manifest does not list, so a dotfiles apply that rewrites the manifest and syncs uninstalls the zellij devlaunch just installed — and the next launch reinstalls it, forever.
• The manifest is not always a file. A devcontainer is free to symlink ~/.pixi/manifests/pixi-global.toml onto a tracked file inside the checkout, and one does; the append then landed in the work tree and every git status in the workspace came up dirty.

Neither is expressible against a home devlaunch created: nothing syncs that manifest, and no repo state can sit under that path. It costs a duplicate extracted prefix in the one case where a tool is installed but unreachable from a login shell — disk only, since PIXI_HOME does not move the download cache — and that is the case where the old behaviour reinstalled on every launch anyway.

Not ~/.local/share/devlaunch/pixi, which is the conventional path and the wrong one here: containers bind-mount ~/.cache, ~/.config and ~/.local/share straight from the host, so a prefix tree under one of them would be shared by every container on the machine and written into your own home — pixi#5476 again, the hazard the cache mount is careful to keep PIXI_HOME away from.

PIXI_HOME is set only for devlaunch's own install scripts, never exported into the login profile, so your own pixi global install in a workspace still goes to your own ~/.pixi. Only the bin directory goes on PATH.

Existing containers, and what a recreate is for

A mount lands only when a container is created. devpod re-applies --workspace-env on every up, but it will not add a bind mount to a container that already exists — passing --mount there is a silent no-op. So a container built before this feature, or before a change to where the mount lands, keeps whatever it was created with until dl <workspace> recreate, and only then picks the current arrangement up.

In between, PIXI_CACHE_DIR points at /var/tmp/devlaunch-pixi with nothing mounted on it. That is a working private cache, not a failure — /var/tmp is world-writable in every image, so pixi creates the directory and fills it. The container re-warms itself and simply never shares, abandoning whatever pixi had already warmed in its default location. This is the reason the container-side path is under /var/tmp rather than somewhere tidier like /var/cache: a target whose parent is root-owned is a hard pixi global sync failure on every container that predates it, not a lost optimisation.

One older breakage needs the recreate rather than a restart. Devlaunch briefly mounted this cache inside ~/.cache, which left that directory root-owned in any image that ships no ~/.cache of its own. $HOME lives on the container's own layer, so dl <workspace> stop and a fresh up keep the root-owned directory; dl <workspace> recreate gets a new layer where ~/.cache is the user's own again.

Worktree Backend

For git repositories, devlaunch uses an efficient worktree backend by default:

• Efficient Storage: Repos are cloned once to ~/.cache/devlaunch/repos/owner/repo/, then git worktrees are created for each branch
• Shared Git Objects: All branches share git objects, saving disk space
• Targeted Fetch: A launch fetches only the one branch it is launching, so no launch waits on a repo-wide refresh

What you get when you push and immediately launch

• Attaching to a workspace devpod already knows: no git at all. The workspace is exactly as you left it; freshness inside it is your own git pull.
• A cold launch (first time this branch is launched on this machine, or a clone devpod has forgotten): one targeted fetch of that branch, every time. Push upstream and immediately dl the branch and you get the pushed tip.
• A branch that does not exist yet: created from the default branch's freshly fetched tip.
• Offline: a warning, and the launch proceeds from whatever the cache holds. It only fails when there is nothing cached to launch from.
• Everything else (other branches, tags, prunes) is refreshed by the background updater within the configured interval (default: 1 hour), which never blocks a launch.

Container Sharing Mode

Use --shared to share a single container across multiple branches of the same repo:

dl --shared owner/repo@branch1  # Creates container "owner-repo"
dl --shared owner/repo@branch2  # Reuses "owner-repo" container

Pre-warming

Use --warm to prepare a workspace without attaching a shell:

dl --warm owner/repo@branch  # Creates container in background

Measuring launch time

Set DEVLAUNCH_TIMING=1 and a dl command ends with one summary on stderr, naming each subprocess round trip and the total. Unset (or 0) records nothing and prints nothing.

Captured from a real warm launch (the launch's own output elided):

$ DEVLAUNCH_TIMING=1 dl-next blooop/mcp-devtasks -- true
...
dl-timing: devpod status 0.454s
dl-timing: gh auth token 0.036s
dl-timing: devpod ssh 1.952s
dl-timing: total 2.444s (in-process, excluding interpreter startup)

The same launch, machine-readable

DEVLAUNCH_TIMING=json swaps that prose for one document on a single dl-timing-json: line, so a trend job can read a launch without scraping prose. It decomposes the launch into five ownership-boundary stages — one per party that could actually make it faster — with the round trips nested inside the stage that paid for them:

Two rules are worth knowing before reading one:

• A stage that never ran is absent, not zero. A warm launch reports no host-prep at all, because it did none. A stage that failed is present, timed up to the failure, and marked failed.
• A stage totals over its whole arm, not just over its round trips, so the host-side work between two spawns is attributed rather than lost. Stages never double-count each other: tools runs inside the launch devpod-up brackets, and those seconds are charged to tools alone. The in-process stages therefore add up to the total — measured on a real cold launch below, they came to 20.834s against a 20.834s total.

handoff is the exception to that sum, and the only one: it ends where total begins, so it is time the process could not have measured from inside itself. A consumer adding stages up against the total leaves it out.

Two optional environment variables let whatever launches dl — a shell function, an agent front-end — close the loop on the time before dl existed. Both are Unix epoch seconds, which is what date +%s.%N prints:

With the prewarm stamp set, the document also reports what that prewarm was worth: the head start it bought, and which shape the launch then took — hit (the workspace was already up), partial (this launch queued behind a prewarm still running) or miss (this launch ran the up itself). dl decides that, not the firer: a prewarm is fired and forgotten, so only the launch that followed can see whether it helped. A stamp that is missing, unreadable, or ahead of this clock reports nothing rather than a zero — an absent handoff and an instantaneous one are different facts, and a trend cannot tell them apart once one is written as the other.

Captured from a real warm launch with both stamps set (one line, wrapped and elided here for reading):

$ DEVLAUNCH_TIMING=json DEVLAUNCH_HANDOFF_T0=$(date +%s.%N) \
    DEVLAUNCH_PREWARM_FIRED_AT=... dl-next blooop/mcp-devtasks -- true
...
dl-timing-json: {"total": 2.210768, "total_epoch": "in-process, excluding interpreter startup",
  "stages": [{"stage": "handoff",   "seconds": 0.130542, "outcome": "ok", "spans": []},
             {"stage": "host-prep", "seconds": 0.027799, "outcome": "ok",
              "spans": [{"label": "gh auth token", "seconds": 0.02747}]},
             {"stage": "devpod-up", "seconds": 0.455188, "outcome": "ok",
              "spans": [{"label": "devpod status", "seconds": 0.455158}]},
             {"stage": "attach",    "seconds": 1.726961, "outcome": "ok",
              "spans": [{"label": "devpod ssh", "seconds": 1.72661}]}],
  "prewarm": {"head_start_seconds": 42.489243, "shape": "hit"}}

Stages appear in the order the launch first entered them, and only the ones it reached appear at all — this warm launch built nothing and lent nothing, so there is no tools stage and no devpod up inside devpod-up. That handoff: 0.131s is the exec and the interpreter start, which nothing else measures.

The cold launch of the same repo, same host and session, decomposed as: host-prep 2.257s (git clone --bare 1.602 + git fetch 0.427 + workspace git clone 0.065 + LFS probe 0.002 + token 0.034), devpod-up 5.848s (devpod up 5.113 of it), tools 10.224s (probe trip 1.584 + tools tar 0.111 + transfer 8.445), attach 2.505s — 20.834s of stages against a 20.834s total.

For before/after numbers, scripts/bench_launch.py runs a command N times and reports the median — one command per side of a change:

python3 scripts/bench_launch.py -n 5 -- dl-next owner/repo -- true   # warm launch

(pixi run bench -n 5 -- ... in the devcontainer.) It reports no median if any run fails, so a broken launch cannot pass as a fast one. See bench_launch.py --help for --before — the per-run reset that makes a cold median cold and whose rm --force also succeeds on the first run, when there is nothing to remove yet — and for why its wall clock and dl-timing: total are not the same quantity. For scale, on the host the session above was captured on, the warm median over 5 runs was 2.176s. Running the cold recipe exactly as the epilog writes it — -n 5, container recreated per run — gave a median of 15.899s (runs: 15.9, 20.0, 15.2, 15.7, 17.8). Read that as the cost of recreating a container, not of a first-ever launch: the reset removes the workspace but leaves the docker image layers and the bare clone cache, so every run after the first starts from both. A machine that must also pull or build the image pays more, by an amount this recipe does not measure — but the gap is large: an earlier 3-run median on this same host, reported as its first real launch, was 33.204s.

Every number in the two paragraphs above was copied into this prose by hand. --record is how that stops: it writes the same invocation as one JSON object a trend job can upload without anyone reading it.

python3 scripts/bench_launch.py -n 5 --record warm.json --shape warm \
    -- dl-next owner/repo -- true

The record holds the command, the run count, each run's wall time and per-stage seconds, and the medians of those — and nothing else, because a CI job stamps its own commit, clock and host better than this script can. Four things about it are load-bearing:

• The median is the point, the runs are its evidence. A trend compares a point against the immediately previous one, so N runs published as N points would read ordinary spread as a regression.
• A stage no run reported is absent, not zero. A warm launch legitimately has no cold-path stages; a zero would claim the work happened instantly rather than not at all. A stage only some runs reported is a median over those runs, carrying a count of how many — a median of two and a median of five are not the same claim.
• Recording asks the launch for its stages (DEVLAUNCH_TIMING=json), so a run that reports no timing document is an error and no record — same discipline as no median over a failed run.
• --shape labels the trend line (warm, cold-recreate). It is the caller's to say: the same command benches either shape depending on --before, and a wrong label is worse in a trend than a missing one.

The trend on main

Every push to main runs .github/workflows/bench.yml, which benches both shapes on the runner and publishes one point per stage to https://blooop.github.io/devlaunch/dev/bench/. It can also be dispatched by hand. Reading it needs nothing but the chart; what follows is for changing it.

scripts/bench_points.py is the step between the two formats — bench records in, one flat array of trend cases out:

python3 scripts/bench_points.py warm.json cold-recreate.json --out bench.json \
    --require-stages-on cold-recreate

(pixi run bench-points ... in the devcontainer.) One case per stage the shape reported, plus that shape's own total: warm / host-prep, cold-recreate / devpod-up, warm / total. Six properties of it are load-bearing:

• The published value is the median, and the case name is a key. The trend compares a point against the immediately previous point and nothing older, so the de-noising has to have happened before publishing. Renaming a stage starts a new, empty series beside a frozen old one.
• The spread rides along as the point's error bar, and the outside stopwatch (wall=) as part of its extra. Evidence beside the number, rather than a second trend line for the same launch that disagrees with the first by a constant.
• An absent stage is absent. A warm launch lends nothing, so there is no warm / tools case at all — never a zero, which would claim an instantaneous lend and would drag the line down exactly where a regression should show.
• --require-stages-on fails the job rather than the trend. The way this decomposition is expected to break is an absence, not a wrong number: a stage stops being emitted and the total keeps working. So the cold-recreate shape, where every stage is known to be present, asserts them all, and a run that lost one publishes nothing and goes red. Naming a shape that was not benched fails too — an assertion that covers nothing reads exactly like one that passed.
• A record that is missing or unreadable refuses the same way. The step that writes the records can exit 0 without having written one, so the absence arrives here — and it prints bench_points: <reason> naming the file and writes nothing, like every other refusal, rather than a traceback.
• A regression alerts; it never gates. The workflow is deliberately not a job in ci.yml: a job there would join the CI gate's needs by house convention and turn a noisy wall-clock measurement into a merge gate. A point above the threshold leaves a commit comment and a red mark on the chart, and the build stays green.

Going back to the Python build

0.1.0 is the Rust rewrite: same commands, same cache, same metadata.json, different implementation (the whole of it is in docs/rust-rewrite-plan.md, whose divergence table is the complete list of what changed on purpose). The last Python release is 0.0.29, and nothing in the cache stops you going back to it:

pixi global install --channel conda-forge --channel https://prefix.dev/blooop "devlaunch<0.1"
pip install "devlaunch<0.1"

Both builds read and write the same metadata.json at schema 2 under the same locks, so a downgrade keeps every workspace you already have and needs no cleanup. Please open an issue for whatever sent you back.

Development

dl and aid are Rust. rust/ is what ships: one cargo workspace, built and tested with cargo, and it is where the tests of dl's own behaviour live.

cd rust
cargo build --release                          # target/release/{dl,aid}
cargo test
cargo clippy --all-targets -- -D warnings
cargo fmt --check

Those cargo commands need no toolchain of your own inside this repository's devcontainer: rust 1.97.1 is in the default pixi environment, so pixi run cargo test --workspace works the moment a container comes up. pixi run keeps the directory it was called from, so cd rust first, exactly as above. rust/rust-toolchain.toml remains the pin of record — a conda rust does not read it, so the version is named again in pyproject.toml and moves by hand with ci.yml's two toolchain steps.

The Python that remains is the acceptance harness and the repo's own documentation guards. It judges the shipped binaries from outside — spawning them against a real devpod, or against the fake one in test/fixtures/devpod_shim.py — and imports nothing that ships:

pixi run ci                                    # ruff, pylint, ty, and the harness under coverage
pixi run test-e2e                              # the real-devpod tier: builds real containers

Both build rust/target/release/{dl,aid} first, because that is what the harness runs. DEVLAUNCH_DL_CMD / DEVLAUNCH_AID_CMD are the seams that point it somewhere else — a debug build, an installed release, the wheel's binary:

# a bare pytest, so the release build the `test` task depends on is skipped too
DEVLAUNCH_DL_CMD='cargo run -q --manifest-path rust/Cargo.toml -p dl --bin dl --' pixi run pytest

Inside this repository's devcontainer, pixi run dl and pixi run aid are cargo run over the working tree; on a host, ./dev.sh installs it as dl-next/aid-next beside the released pair. Both print a -dev version so a working-tree build is never mistaken for a released one. See AGENTS.md.

Until 0.1.0 this repository held a second, Python implementation of dl, and a parity harness that ran both against the same fixtures and compared them. Both retired once the binaries shipped; the history is in docs/rust-rewrite-plan.md, whose divergence table still records every deliberate behavioural difference from that build.

The two published artifacts are built from the same cargo package, both taking their version from rust/Cargo.toml — the only place it is written down:

cd packaging/wheel && maturin build --release --locked -o dist    # the PyPI wheel: two binaries
rattler-build build --experimental --recipe conda.recipe/recipe.yaml   # the conda package

.github/workflows/publish.yml and conda-publish.yml do exactly that on a version bump, in that order, off one tag; ci.yml's packaging job builds the wheel and renders the recipe on every pull request, so a broken release is a red tick rather than a surprise.

The Python half uses pixi for environment management.

# Run tests
pixi run test

# Run the e2e suite: real devpod, real containers
pixi run test-e2e

# Run full CI suite
pixi run ci

# Format and lint
pixi run style

pixi run test skips the e2e tests, which need devpod and a Docker daemon and build real containers. CI runs pixi run test-e2e in a job of its own, outside the Python matrix, on a throwaway runner — on every push to main and on every pull request, whatever branch that pull request targets. Stacked chains, where each link targets its predecessor rather than main, get the same CI as anything else.

Alongside the matrix and e2e there is a gate job that does nothing but fail unless every other job in that workflow succeeded. It exists so that a branch ruleset has one stable name to require rather than a list: requiring the jobs one by one means literal strings in a repository setting, which nobody reviews and which goes stale the moment a job is added or renamed — and a required check that no longer exists does not turn a merge red, it stops gating it. Adding a job means adding it to gate's needs, in the same pull request, where it can be seen. It reaches only as far as its own workflow file, so the prek lint job is not behind it and has to be required alongside it.

Running it yourself is a different proposition. This repo's devcontainer carries a Docker daemon of its own, through the docker-in-docker feature, and pins the same devpod a host installs, so pixi run test-e2e from inside it builds its containers in there rather than on your Docker. You can also run it on a machine you do not mind it writing to — an ephemeral CI runner, say. It is skipped by default rather than gated on a container, because what it needs is a daemon, not nesting. Either way the suite exercises dl --purge, so it gives itself a private devpod namespace before collection begins — but the containers it builds are real ones, and it wants several minutes and a 1.25 GB image pull the first time.

Its skips mean one thing only. A test that opts out does so through fixtures.e2e_guard.opt_out, and any other skip is reported as a failure, because a run that could not reach a registry used to be indistinguishable from a healthy one. Every run also prints what it actually built:

--------------------------------- e2e session ---------------------------------
22 e2e tests attempted, 5 workspaces created: e2e-test-create, e2e-test-lifecycle, e2e-test-git, e2e-purge-devlaunchs, e2e-purge-hand-made

A run whose workspace-building tests built nothing does not pass: the shortfall is counted into the last line of the run, so 4 passed, 18 skipped becomes 1 failed, 4 passed, 18 skipped. A run with no workspace-building tests in it — pytest -m e2e -k TestDLCommandsE2E, say — has nothing to answer for and says so instead.

The nested daemon is also why the devcontainer does not join the host's network namespace: a nested daemon needs a namespace of its own, or it co-manages the host's docker0 bridge and writes its NAT rules into the host's netfilter tables.

The prebuilt dev container image

Opening this repository's devcontainer used to build it: base image, pixi, the local claude-code feature and docker-in-docker, several minutes of it, once per branch. CI publishes that image now, so opening a workspace pulls it instead.

ghcr.io/blooop/devlaunch-devcontainer

Nothing has to be configured to use it. .devcontainer/devcontainer.json names the repository under customizations.devpod.prebuildRepository, and devpod up — which is every dl launch — checks it before building anything. The check is for one exact tag: devpod hashes the build config together with the build context and asks for <repository>:devpod-<hash>. A hit is used directly, features included; a miss falls through to a local build, silently and without failing. So the prebuild is a speed-up that cannot break a launch, and the question to ask when a container open is slow is whether the tag matched.

Two consequences worth knowing:

• The build context is .devcontainer, not the repository root, and that is what makes the tag usable. The Dockerfile copies nothing out of the context, so the root was never needed — but it was hashed, which meant a different tag on every commit to any file and a prebuilt image that never matched one. Scoped to .devcontainer, the tag moves when .devcontainer/** moves, the Dockerfile and the local feature's scripts included.
• A commit whose .devcontainer/ differs from the last prebuild builds locally. That is the correct answer rather than a gap: the alternative is a container built from something other than what the branch asks for. The pull comes back once the change is on main. What the tag does not promise is the converse — that one .devcontainer/ tree always yields one image; see "What the prebuild tag does not promise" below.

.github/workflows/devcontainer-prebuild.yml publishes it, on pushes to main that touch .devcontainer/** and on manual dispatch. Its path filter is exactly the set of inputs to the hash, so a commit it skips is one that could not have moved the tag. To publish from a branch by hand:

docker login ghcr.io                  # a PAT with write:packages, or `gh auth token`
pixi run devcontainer-prebuild        # devpod build . --tag latest

Both are idempotent: an existing prebuild is found and returned rather than rebuilt and repushed. By hand publishes for the architecture of the machine you run it on and no other — see "Two architectures, two tags" below — and takes the moving alias as an argument (pixi run devcontainer-prebuild latest-arm64) if that machine is not amd64.

The package came up public on its own, and needed no manual step. Measured on the first run (d05e4ce): an anonymous pull of both tags returns 200, with nobody having touched a visibility setting. GHCR gave the package the visibility of the public repository whose workflow published it — the org.opencontainers.image.source label in the Dockerfile is what links the two.

This is worth checking rather than trusting, because the failure is silent: a private package makes the lookup return DENIED, devpod reads that as a cache miss, and every launch quietly builds locally — the behaviour from before any of this existed.

curl -s -o /dev/null -w '%{http_code}\n' \
  -H "Authorization: Bearer $(curl -s \
    'https://ghcr.io/token?scope=repository:blooop/devlaunch-devcontainer:pull&service=ghcr.io' \
    | python3 -c 'import sys,json;print(json.load(sys.stdin)["token"])')" \
  https://ghcr.io/v2/blooop/devlaunch-devcontainer/manifests/latest
# 200 => public, anonymous pulls work
# 401/403 => private; make it public as below

If it ever is private — a package created some other way, or a visibility that gets changed — the fix is a one-time setting, and one no workflow can make: there is no REST endpoint for container visibility and no gh subcommand. https://github.com/blooop?tab=packages → devlaunch-devcontainer → Package settings → Danger Zone → Change visibility → Public. Nothing to configure before the first publish creates the package, so those pages 404 until then.

The :latest tag that command also pushes is not what devpod looks for. It is the moving alias build.cacheFrom points at, a best-effort layer cache for builders that know nothing about devpod prebuilds — VS Code's "Reopen in Container", a plain devcontainer up. Those still run a build; what they save is whatever layers the cache can serve them.

Two architectures, two tags. The target architecture is hashed along with the build config and the context, so amd64 and arm64 ask for different tags, and the workflow publishes both — a matrix over ubuntu-latest and ubuntu-24.04-arm, GitHub's hosted arm64 runner, which is free without limit on a public repository. Nothing merges the two into a multi-arch manifest, because devpod never looks one up: it asks for one exact tag and pulls the variant for the architecture it is running on. An architecture whose tag is missing is not a failure, only the same silent local build as any other miss.

Neither leg passes --platform, and that is the point of using a native runner rather than emulation. The architecture in the hash is the one the driver reports, which for docker is the runtime.GOARCH of the devpod binary doing the build — so the runner decides it, exactly as the launching machine decides it on the lookup side, where devpod up passes no platform at all. --platform linux/arm64 would hash to the same string on an arm64 runner, which is the argument against it: a second source of truth for something already settled, and one that can disagree without saying so.

The legs are fail-fast: false. Each publishes a tag nothing else reads, so a run where one architecture fails leaves the other pulling, which is better than the two local builds that cancelling the survivor would leave behind.

latest is amd64's, and arm64 publishes latest-arm64. devpod arch-qualifies nothing, so two legs passing the same alias would race for it and the winner would be whichever finished last; and the one thing that reads latest is build.cacheFrom, a layer cache which serves nothing at all across architectures. So latest keeps meaning what it has always meant, instead of becoming a cache that works or does not per run for reasons nobody could see. What dl reads is neither alias: it is the devpod-<hash> tag for its own architecture.

Making the arm64 leg possible at all needed linux-aarch64 in the pixi workspace (platforms in pyproject.toml). A pixi workspace refuses outright on a platform it does not declare, so without it the leg fails at pixi install before it can build anything — and so does an arm64 container's own postCreateCommand, which would have made an arm64 prebuild an image that pulls fast and then cannot come up. It says nothing about the release: the wheel and the conda package are still linux-64.

The arm64 leg went in unexercised, since this repository is developed on x86: every piece the image needs was checked to exist for linux-aarch64 — the multi-arch base, the aarch64 pixi and claude-shim builds, devpod itself — and none of it was checked to build. If it turns out not to, the symptom on an arm64 host is the local build that host was already doing.

postCreateCommand is not in the image and cannot be: pixi install and the provider registration run at container create, after the image exists. The <workspace>-pixi volume is what makes them cheap the second time.

That install is pixi install --frozen, which is the same resolution CI uses (frozen: true in ci.yml) and is load-bearing for two reasons beyond matching it. A bare pixi install treats a lock it cannot read as a missing one: it warns, exits 0, solves a fresh environment, and rewrites the tracked pixi.lock on its way past — so the container that is supposed to reproduce the committed environment quietly stops being it, and says nothing. And the solve it does instead reaches the network: resolving pypi dependencies alongside conda ones needs a conda-pypi name mapping that pixi fetches remotely (the compressed mapping is served out of prefix-dev/parselmouth on raw.githubusercontent.com), and a create whose fetch fails dies in postCreateCommand with the workspace never opening.

--frozen installs the committed resolution, so there is no solve and no mapping to fetch. That is measured rather than reasoned: with the mapping cache deleted, pixi install --frozen --offline installs the default environment and never recreates the cache, while a solving install recreates it on the spot. And a lock the container's pixi cannot read now fails immediately and says which pixi it would need, instead of succeeding into the wrong environment.

What the prebuild tag does not promise

The tag is a hash of the build recipe — the config and the context — and not of the image the recipe produces. Three of the recipe's inputs float, so the same .devcontainer/ tree published on two different days is two different images: the base mcr.microsoft.com/devcontainers/base:ubuntu-24.04 is a moving tag, ghcr.io/devcontainers/features/docker-in-docker:2 is a moving major, and the local feature's pixi global install … claude-shim names no version. (The Dockerfile's ARG PIXI_VERSION is the one that is pinned.) A prebuild pulled today and a local build of the identical tree tomorrow can therefore differ, and the direction they differ in is forwards: a republish picks up whatever those three point at now.

That is written down rather than fixed, and the shim is the case that shows why. claude-shim is deliberately unversioned, and pinning it was considered and rejected on four grounds:

• The package holds no claude. It is 21 KB of bash — claude, cld, cldr — whose first run downloads the current stable binary from Anthropic's GCS bucket into ~/.claude/cache, and which re-checks stable hourly after that. The image cannot serve a stale claude, because it contains none; a pin would freeze the fetcher and nothing it fetches.
• On the launch path this container is opened by, the baked shim never runs. dl's probe reads a shim-provided claude as lendable and the lend puts the host's real binary in front of it on the PATH — "What to bake so a launch does no work at all", above. The shim is baked so that command -v claude answers at all.
• A pin freezes it harder than no pin does. Unversioned, the shim is refreshed by every republish, which is every commit to .devcontainer/**. Pinned, it stops at whatever was current the day the pin was written.
• One package would then have two policies. The shipping provisioner (rust/devlaunch-core/src/flows/provision.rs) installs the same package, unversioned, into every workspace dl opens — the copy that reaches users rather than us. test/unit/test_devcontainer_manifest.py asserts that spec and the feature installer's still match, so pinning one side alone fails rather than drifting, while pinning both deliberately passes.

Measured 2026-08-20: the published prebuild carries claude-shim 0.7.0, the newest of the channel's 14 releases and unmoved since 2026-04-06, so the drift a pin would have prevented is currently zero. A publish is also the gate on a broken shim release, not a channel for it: the feature installer checks the trampoline exists and returns non-zero if it does not, which fails devpod build and leaves the previous prebuild as the newest published tag.

If image contents ever do need pinning, the base image tag is the input that carries the packages; pinning the shim alone would be precision about the smallest of the three.

Disk cost of the dev container

Opening a devcontainer for a branch costs about 3.3 GB on the host before you do anything in it: ~600 MB of image layers unique to this image, a ~680 MB container writable layer, and a ~2.0 GB <workspace>-pixi volume.

That volume was ~520 MB until the rust feature went into the default environment — measured either side of that change, the pixi environment is 521 MB without it and 2044 MB with it, and only about a quarter of the difference is the compiler itself: conda's rust links with a gcc toolchain, a sysroot and binutils.

That 1.5 GB per branch is bought deliberately, and what it buys is isolation. The toolchain that builds the crates is pinned in the project environment beside the pinned devpod, so a workspace compiles what it is editing with no host toolchain, no rustup, and no global install anywhere in the picture — and it is the same version on every machine and in CI. Paying for that once per branch workspace is the trade; a leaner environment that sent whoever is working in the container back to a host cargo would be the wrong end of it.

The container carries its own Docker daemon, and that daemon's /var/lib/docker lives on a second named volume. One pixi run test-e2e plus a couple of nested workspaces puts ~2.3 GB in there, and nothing garbage-collects it — the inner daemon reports ~45% of its images reclaimable with no reclaimer. Nested daemons share no layers with the host or with each other, so this is paid once per branch.

Budget ~5.5 GB per branch you are actively developing and e2e-testing — about 17 GB for three concurrent branches.

The time cost is cold pulls in a fresh nested daemon: the first devpod up inside a new container takes ~25s, ~16s of which is pulling a base image the host already has. Workspaces after that reuse it and take ~8s.

These volumes are not reclaimed automatically. devpod delete removes the container with docker rm and never touches volumes, and Docker never garbage-collects a named volume — so <workspace>-pixi and dind-var-lib-docker-* outlive the workspace that created them. To see what has piled up:

docker system df -v      # under Local Volumes, LINKS 0 means no container uses it

Cross-check a name against devpod list before removing it with docker volume rm: a volume belonging to a live workspace shows LINKS 1.